Troutmal CalculusOpControl

Undergraduate Texts in Mathematics
Editors
S. Axler
F.W. Gehring
P.R. Halmos
Springer
New York
Berlin
Heidelberg
Barcelona
Budapest
Hong Kong
London
Milan
Paris
Santa Clara
Singapore
Tokyo
Anglin: Mathematics: A Concise History Elaydi: Introduction to Difference

and Philosophy. Equations.
Readings in Mathematics. Fischer: Intermediate Real Analysis.
Anglin/Lambek: The Heritage of Thales. Flanigan/Kazdan: Calculus Two: Linear
Readings in Mathematics. and Nonlinear Functions. Second
Apostol: Introduction to Analytic edition.
Number Theory. Second edition. Fleming: Functions of Several Variables.
Armstrong: Basic Topology. Second edition.
Armstrong: Groups and Symmetry. Foulds: Combinatorial Optimization for
BaklNewman: Complex Analysis. Undergraduates.
BanchofflWermer: Linear Algebra Foulds: Optimization Techniques: An
Through Geometry. Second edition. Introduction.
Berberian: A First Course in Real Franklin: Methods of Mathematical
Analysis. Economics.
Bremaud: An Introduction to HairerlWanner: Analysis by Its History.
Probabilistic Modeling. Readings in Mathematics.
Bressoud: Factorization and Primality HaImos: Finite-Dimensional Vector
Testing. Spaces. Second edition.
Bressoud: Second Year Calculus. Halmos: Naive Set Theory.
Readings in Mathematics. HiimmerlinIHoffmann: Numerical
Brickman: Mathematical Introduction Mathematics.
to Linear Programming and Game Readings in Mathematics.
Theory. Iooss/Joseph: Elementary Stability and
Browder: Mathematical Analysis: Bifurcation Theory. Second edition.
An Introduction. Isaac: The Pleasures of Probability.
Cederberg: A Course in Modern Readings in Mathematics.
Geometries. James: Topological and Uniform Spaces.
Childs: A Concrete Introduction to Jiinich: Linear Algebra.
Higher Algebra. Second edition. Jiinich: Topology.
Chung: Elementary Probability Theory Kemeny/Snell: Finite Markov Chains.
with Stochastic Processes. Third Kinsey: Topology of Surfaces.
edition. Klambauer: Aspects of Calculus.
CoxlLittlelO'Shea: Ideals, Varieties, Lang: A First Course in Calculus. Fifth
and Algorithms. edition.
Croom: Basic Concepts of Algebraic Lang: Calculus of Several Variables.
Topology. Third edition.
Curtis: Linear Algebra: An Introductory Lang: Introduction to Linear Algebra.
Approach. Fourth edition. Second edition.
Devlin: The Joy of Sets: Fundamentals Lang: Linear Algebra. Third edition.
of Contemporary Set Theory. Second Lang: Undergraduate Algebra. Second
edition. edition.
Dixmier: General Topology. Lang: Undergraduate Analysis.
Driver: Why Math? LaxIBursteinlLax: Calculus with
Ebbinghaus/FIum/Thomas: Applications and Computing.
Mathematical Logic. Second edition. Volume 1.
Edgar: Measure, Topology, and Fractal LeCuyer: College Mathematics with
Geometry. APL.
(continued after index)
John L. Troutman
Variational Calculus and

Optimal Control
Optimization with
Elementary Convexity
Second Edition
With 87 Illustrations
Springer
John L. Troutman
Department of Mathematics
Syracuse University
Syracuse, NY 13210
USA
Editorial Board:
Sheldon Axler F.W. Gehring
Department of Mathematics Department of Mathematics
Michigan State University University of Michigan
East Lansing, MI 48824 Ann Arbor, MI 48109
USA USA
Paul R. Halmos
Department of Mathematics
Santa Clara University
Santa Clara, CA 95053
USA
Mathematics Subject Classifications (1991): 49-01
Library of Congress Cataloging-in-Publication Data

Troutman, John L.
Variational calculus and optimal control: Optimization with elementary convexity
/ John L. Troutman. 2nd edition.
p. cm.-(Undergraduate texts in mathematics. Readings in
mathematics.)
Includes bibliographical references and index.
ISBN 0-387-94511-3 (hardcover: a1k. paper)
1. Calculus of variations. 2. Control theory. 3. Mathematical
optimization. 4. Convex functions. I. Title. II. Series.
QA315.T724 1995
515'.64-dc20 95-12918
Printed on acid-free paper.
© 1996, 1983 Springer-Verlag New York, Inc.

All rights reserved. This work may not be translated or copied in whole or in part without the
written permission of the publisher (Springer-Verlag New York, Inc., 175 Fifth Avenue, New
York, NY 10010, USA), except for brief excerpts in connection with reviews or scholarly
analysis. Use in connection with any form of information storage and retrieval, electronic
adaptation, computer software, or by similar or dissimilar methodology now known or hereafter
developed is forbidden.
The use of general descriptive names, trade names, trademarks, etc., in this publication, even if
the former are not especially identified, is not to be taken as a sign that such names, as
understood by the Trade Marks and Merchandise Marks Act, may accordingly be used freely by
anyone.
Production coordinated by Brian Howe and managed by Henry Krell; manufacturing super-
vised by Jeffrey Taub.
Typeset by Asco Trade Typesetting Ltd., Hong Kong.
Printed and bound by R.R. Donnelley & Sons, Harrisonburg, VA.
Printed in the United States of America.
9 8 7 6 5 432 1
ISBN 0-387-94511-3 Springer-Verlag New York Berlin Heidelberg

This book is dedicated to my parents
and to my teachers
Preface
Although the calculus of variations has ancient origins in questions of Ar-

istotle and Zenodoros, its mathematical principles first emerged in the post-
calculus investigations of Newton, the Bernoullis, Euler, and Lagrange. Its
results now supply fundamental tools of exploration to both mathematicians
and those in the applied sciences. (Indeed, the macroscopic statements ob-
tained through variational principles may provide the only valid mathemati-
cal formulations of many physical laws.) Because of its classical origins,
variational calculus retains the spirit of natural philosophy common to most
mathematical investigations prior to this century. The original applications,
including the Bernoulli problem of finding the brachistochrone, require opti-
mizing (maximizing or minimizing) the mass, force, time, or energy of some
physical system under various constraints. The solutions to these problems
satisfy related differential equations discovered by Euler and Lagrange, and
the variational principles of mechanics (especially that of Hamilton from the
last century) show the importance of also considering solutions that just
provide stationary behavior for some measure of performance of the system.
However, many recent applications do involve optimization, in particular,
those concerned with problems in optimal control.
Optimal control is the rapidly expanding field developed during the last
half-century to analyze optimal behavior of a constrained process that
evolves in time according to prescribed laws. Its applications now embrace a
variety of new disciplines, including economics and production planning.l In
1 Even the perennial question of how a falling cat rights itself in midair can be cast as a control
problem in geometric robotics! See Dynamics and Control of Mechanical Systems: The Falling
Cat and Related Problems, by Michael Enos, Ed. American Mathematical Society, 1993.
vii
viii Preface
this text we will view optimal control as a special form of variational calculus,
although with proper interpretation, these distinctions can be reversed.
In either field, most initial work consisted of finding (necessary) conditions
that characterize an optimal solution tacitly assumed to exist. These condi-
tions were not easy to justify mathematically, and the subsequent theories
that gave (sufficient) conditions guaranteeing that a candidate solution does
optimize were usually substantially harder to implement. (Conditions that
ensure existence of an optimizing solution were-and are-far more difficult
to investigate, and they cannot be considered at the introductory level of this
text. See [Cel) Now, in any of these directions, the statements of most later
theoretical results incorporate some form of convexity in the defining func-
tions (at times in a disguised form). Of course, convexity was to be expected
in view of its importance in characterizing extrema of functions in ordinary
calculus, and it is natural to employ this central theme as the basis for an
introductory treatment.
The present book is both a refinement and an extension of the author's
earlier text, Variational Calculus with Elementary Convexity (Springer-Verlag,
1983) and its supplement, Optimal Control with Elementary Convexity (1986).
It is addressed to the same audience of junior to first-year graduate students
in the sciences who have some background in multidimensional calculus and
differential equations. The goal remains to solve problems completely (and
exactly) whenever possible at the mathematical level required to formulate
them. To help achieve this, the book incorporates a sliding scale-of-difficulty
that allows its user to become gradually more sophisticated, both technically
and theoretically. The few starred (*) sections, examples, and problems out-
side this scheme can usually be overlooked or treated lightly on first reading.
For our purposes, a convex function is a differentiable real-valued func-
tion whose graph lies above its tangent planes. In application, it may be
enough that a function of several variables have this behavior only in some
of the variables, and such "elementary" convexity can often be inferred
through pattern recognition. Moreover, with proper formulation, many more
problems possess this convexity than is popularly supposed. In fact, using
only standard calculus results, we can solve most of the problems that moti-
vated development of the variational calculus, as well as many problems of
interest in optimal control.
The paradigm for our treatment is as follows: Elementary convexity sug-
gests simple sufficiency conditions that can often lead to direct solution, and
they in turn inform the search for necessary conditions that hold whether or
not such convexity is present. For problems that can be formulated on a fixed
interval (or set) this statement remains valid even when fixed-endpoint condi-
tions are relaxed, or certain constraints (isoperimetric or Lagrangian) are
imposed. Moreover, sufficiency arguments involving elementary convexity
are so natural that even multidimensional generalizations readily suggest
themselves.
Preface IX
In Part I, we provide the standard results of variational calculus in the

context of linear function spaces, together with those in Chapter 3 that use
elementary convexity to establish sufficiency. In Part II, we extend this devel-
opment into more sophisticated areas, including Weierstrass-Hilbert field
theory of sufficiency (Chapter 9). We also give an introduction to Hamil-
tonian mechanics and use it in §8.8 to motivate a different means for recog-
nizing convexity, that leads to new elementary solutions of some classical
problems (including that of the brachistochrone). Throughout these parts, we
derive and solve many optimization problems of physical significance in-
cluding some involving optimal controls. But we postpone our discussion of
control theory until Part III, where we use elementary convexity to suggest
sufficiency of the Pontjragin principle before establishing its necessity in the
concluding chapter.
Most of this material has been class-tested, and in particular, that of Part
I has been used at Syracuse University over 15 years as the text for one
semester of a year-sequence course in applied matheIIl,at.~<:_s~_Chapter 8 (on
Hamiltonian mechanics) can be examined independently of adjacent chap-
ters, but Chapter 7 is prerequisite to any other subsequent chapters. On the
other hand, those wishing primarily an introduction to optimal control could
omit both Chapters 8 and 9. The book is essentially self-contained and
includes in Chapter 0 a review of optimization in Euclidean space. It does not
employ the Lebesque integral, but in the Appendix we develop some neces-
sary results about analysis in Euclidean space and families of solutions to
systems of differential equations.
Acknowledgments
I wish once more to express my appreciation to those who first made me

aware of the elegance and power of variational methods-Daniel Frederick
at V.P.I., M.M. Schiffer at Stanford, and C. Lanczos as author. I must
also reacknowledge the contributions that William Hrusa (now at Carnegie-
Mellon University) made to the earlier work during his student days. Many
of the problems in Part I and II originated with Bill, and his assistance and
commentary during that initial production were invaluable. It has been re-
warding to hear from those including Terry Rockafellar at the University of
Washington, Frank Chorlton at the University of Aston, and Morris Kline
at New York University, whose satisfaction with the earlier work helped
motivate this extension.
I wish also to express my gratitude to Phil Loewen at the University of
British Columbia and Frank Clarke at the Universite de Montreal, as well as
to Dan Waterman, my friend and colleague at Syracuse, for their suggestions
and encouragement. Many thanks are due my other colleagues, especially
Gerry Cargo, Phil Church, Phil Griffin, Wolfgang Jurkat, Tadeusz Iwaniec,
x Preface
and Andy Vogel, who taught from this material at Syracuse and made valu-
able suggestions for its improvement. And of course, I feel deep gratitude to
and for my many students over the years, without whose evident enjoyment
and expressed appreciation the current work would not have been under-
taken. It is a pleasure to recognize those responsible for transforming this
work from manuscript to printed page: the principal typists, Louise Capra,
Esther Clark, and Steve Everson; the editors and staff at Springer-Verlag,
in particular, Jenny Wolkowicki and the late Walter Kauffmann-Buhler.
Finally, I wish to thank my wife, Patricia Brookes, for her patience and
understanding during the years of revision.
Syracuse, New York JOHN L. TROUTMAN

Contents
Preface vii
CHAPTER 0
Review of Optimization in IRd 1
Problems 7
PART ONE
BASIC THEORY 11
CHAPTER 1
Standard Optimization Problems 13
1.1. Geodesic Problems 13
(a) Geodesics in IRd 14
(b) Geodesics on a Sphere 15
(c) Other Geodesic Problems 17
1.2. Time-of-Transit Problems 17
(a) The Brachistochrone 17
(b) Steering and Control Problems 20
1.3. Isoperimetric Problems 21
1.4. Surface Area Problems 24
(a) Minimal Surface of Revolution 24
(b) Minimal Area Problem 25
(c) Plateau's Problem 26
1.5. Summary: Plan of the Text 26
Notation: Uses and Abuses 29
Problems 31
xi
xii Contents
CHAPTER 2
Linear Spaces and Gateaux Variations 36
2.1. Real Linear Spaces 36
2.2. Functions from Linear Spaces 38
2.3. Fundamentals of Optimization 39
Constraints 41
Rotating Fluid Column 42
2.4. The Gateaux Variations 45
Problems 50
CHAPTER 3
Minimization of Convex Functions 53
3.1. Convex Functions 54
3.2. Convex Integral Functions 56
Free End-Point Problems 60
3.3. [Strongly] Convex Functions 61
3.4. Applications 65
(a) Geodesics on a Cylinder 65
(b) A Brachistochrone 66
(c) A Profile of Minimum Drag 69
(d) An Economics Problem 72
(e) Minimal Area Problem 74
3.5. Minimization with Convex Constraints 76
The Hanging Cable 78
Optimal Performance 81
3.6. Summary: Minimizing Procedures 83
Problems 84
CHAPTER 4
The Lemmas of Lagrange and Du Bois-Reymond 97
Problems 101
CHAPTER 5
Local Extrema in Normed Linear Spaces 103
5.1. Norms for Linear Spaces 103
5.2. Normed Linear Spaces: Convergence and Compactness 106
5.3. Continuity 108
5.4. (Local) Extremal Points 114
5.5. Necessary Conditions: Admissible Directions 115
5.6*.Affine Approximation: The Frechet Derivative 120
Tangency 127
5.7. Extrema with Constraints: Lagrangian Multipliers 129
Problems 139
CHAPTER 6
The Euler-Lagrange Equations 145
6.1. The First Equation: Stationary Functions 147
6.2. Special Cases of the First Equation 148
Contents xiii
(a) When J = J(z) 149

(b) When J = J(x, z) 149
(c) When J = J(y, z) 150
6.3. The Second Equation 153
6.4. Variable End Point Problems: Natural Boundary Conditions 156
Jakob Bernoulli's Brachistochrone 156
Transversal Conditions* 157
6.5. Integral Constraints: Lagrangian Multipliers 160
6.6. Integrals Involving Higher Derivatives 162
Buckling of a Column under Compressive Load 164
6.7. Vector Valued Stationary Functions 169
The Isoperimetric Problem 171
Lagrangian Constraints* 173
Geodesics on a Surface 177
6.8*. Invariance of Stationarity 178
6.9. Multidimensional Integrals 181
Minimal Area Problem 184
Natural Boundary Conditions 185
Problems 186
PART TWO
ADVANCED TOPICS 195
CHAPTER 7
Piecewise C 1 Extremal Functions 197
7.1. Piecewise C 1 Functions 198
(a) Smoothing 199
(b) Norms for C1 201
7.2. Integral Functions on C1 202
7.3. Extremals in C1[a, b]: The Weierstrass-Erdmann Corner Conditions 204
A Sturm-Liouville Problem 209
7.4. Minimization Through Convexity 211
Internal Constraints 212
7.5. Piecewise C 1 Vector-Valued Extremals 215
Minimal Surface of Revolution 217
Hilbert's Differentiability Criterion* 220
7.6*. Conditions Necessary for a Local Minimum 221
(a) The Weierstrass Condition 222
(b) The Legendre Condition 224
Bolza's Problem 225
Problems 227
CHAPTER 8
Variational Principles in Mechanics 234
8.1. The Action Integral 235
8.2. Hamilton's Principle: Generalized Coordinates 236
Bernoulli's Principle of Static Equilibrium 239
xiv Contents
8.3. The Total Energy 240

Spring-Mass-Pendulum System 241
8.4. The Canonical Equations 243
8.5. Integrals of Motion in Special Cases 247
Jacobi's Principle of Least Action 248
Symmetry and Invariance 250
8.6. Parametric Equations of Motion - 250
8.7*. The Hamilton-Jacobi Equation 251
8.8. Saddle Functions and Convexity; Complementary Inequalities 254
The Cycloid Is the Brachistochrone 257
Dido's Problem 258
8.9. Continuous Media 260
(a) Taut String 260
The Nonuniform String 264
(b) Stretched Membrane 266
Static Equilibrium of (Nonplanar) Membrane 269
Problems 270
CHAPTER 9*
Sufficient Conditions for a Minimum 282
9.1. The Weierstrass Method 283
9.2. [Strict] Convexity of f(!., y, Z) 286
9.3. Fields 288
Exact Fields and the Hamilton-Jacobi Equation* 293
9.4. Hilbert's Invariant Integral 294
The Brachistochrone* 296
Variable End-Point Problems 297
9.5. Minimization with Constraints 300
The Wirtinger Inequality 304
9.6*. Central Fields 308
Smooth Minimal Surface of Revolution 312
9.7. Construction of Central Fields with Given Trajectory:
The Jacobi Condition 314
9.8. Sufficient Conditions for a Local Minimum 319
(a) Pointwise Results 320
Hamilton's Principle 320
(b) Trajectory Results 321
9.9*. Necessity of the Jacobi Condition 322
9.10. Concluding Remarks 327
Problems 329
PART THREE
OPTIMAL CONTROL 339
CHAPTER 10*
Control Problems and Sufficiency Considerations 341
10.1. Mathematical Formulation and Terminology 342
Contents xv
10.2. Sample Problems 344

(a) Some Easy Problems 345
(b) A Bolza Problem 347
(c) Optimal Time of Transit 348
(d) A Rocket Propulsion Problem 350
(e) A Resource Allocation Problem 352
(f) Excitation of an Oscillator 355
(g) Time-Optimal Solution by Steepest Descent 357
10.3. Sufficient Conditions Through Convexity 359
Linear State-Quadratic Performance Problem 361
10.4. Separate Convexity and the Minimum Principle 365
Problems 372
CHAPTER 11
Necessary Conditions for Optimality 378
11.1. Necessity of the Minimum Principle 378
(a) Effects of Control Variations 380
(b) Autonomous Fixed Interval Problems 384
Oscillator Energy Problem 389
(c) General Control Problems 391
11.2. Linear Time-Optimal Problems 397
Problem Statement 398
A Free Space Docking Problem 401
11.3. General Lagrangian Constraints 404
(a) Control Sets Described by Lagrangian Inequalities 405
(b)* Variational Problems with Lagrangian Constraints 406
(c) Extensions 410
Problems 413
Appendix
A.O. Compact Sets in [Rd 419
A.1. The Intermediate and Mean Value Theorems 421
A.2. The Fundamental Theorem of Calculus 423
A.3. Partial Integrals: Leibniz' Formula 425
A.4. An Open Mapping Theorem 427
A.5. Families of Solutions to a System of Differential Equations 429
A.6. The Rayleigh Ratio 435
A.7*. Linear Functionals and Tangent Cones in [Rd 441
Bibliography 445
Historical References 450
Answers to Selected Problems 452
Index 457
CHAPTER 0
Review of Optimization in [Rd
This chapter presents a brief summary of the standard terminology and basic
results related to characterizing the maximal and minimal values of a real
valued function f defined on a set D in Euclidean space. With the possible
exception of the remarks concerning convexity ((0.8) and (0.9», this material
is covered in texts on multidimensional calculus; the notation is explained at
the end of §1.5.
For d = 1, 2, 3, ... , let ~d denote d-dimensional real Euclidean space

where a typical point or vector X = (Xl' X 2 , ... , Xd) has the length IXI =
(L1=l IXjI2)l/2 which is positive unless X = (!) = (0,0,0, ... ,0). (We identify
~l with ~.)
On ~d, with Y = (Yl' Y2' ... , Yd)' we have the vector space operations of
componentwise addition
deC
X + Y = (Xl + Yl' X2 + Y2, ... , Xd + Yd)'
and scalar multiplication:
VaE ~.
We may also express IXI = (X· X)l/2, utilizing the scalar or dot product
which is subject to the Cauchy inequality

I(X· Y)I :::;; IXII YI· (1)
The Cauchy inequality (1) is used to prove the so-called triangle inequality
IX + YI :::;; IXI + I YI, (2a)
2 o. Review of Optimization in [Rd
an alternate form of which is

IIXI-IYII ~ IX - YI, (2b)
where
X - Y~ X + (-l)Y; (Problem 0.1).
IX - YI defines the Euclidean distance between X and Y.
When Xo E ~d, then for finite ~ > 0, the "sphere"

Sd(XO)~ {X E ~d: IX - Xol <~}
is called an (open) neighborhood of X o, and Xo is said to be an interior point
of each set D which contains this neighborhood for some ~ > O. D is open
when it consists entirely of interior points. An open set D is a domain when
each pair of its points may be connected by a (polygonal) curve which lies
entirely in D. Each open sphere is a domain, as is each open "box"
B = {X E ~d: aj < Xj < bj,j = 1,2, ... , d},
but the union of disjoint open sets is not a domain, although it remains open.
A point not in the interior of a set S, and not interior to its complement,
~d '" S, is called a boundary point of S. The set of such points, denoted as,
is called the boundary of S. For example, if S = {(X E ~d: IXI ~ I}, then
as = B = {X E ~d: IXI = I}; also aB = B. A set S ~ ~d is said to be bounded
iff it is a subset of some sphere.
We suppose that we are given a real valued function I defined on a set
D ~ ~d for which we wish to find extremal values. That is, we wish to find
points in D (called extremal points) at which I assumes maximum or mini-
mum values. With such optimization problems we should note the following
facts:
(0.0) I need not have extremal values on D.
For example, when D = ~1, then the function I(X) = Xl is unbounded in

both directions on D. Moreover, on the open interval D = (-1, 1) ~ ~1, this
same function, although bounded, takes on values as near -lor 1 as we
please but does not assume the values ± 1 on D. On the closed interval,
D = [ -1,1], this function does assume both maximum and minimum
values, but the function
I(X) =~,
Xl
I((!)) = 0,
is again unbounded.
(0.1) I may assume only one extremal value on D.

o. Review of Optimization in IRd 3
For example, on D = (-1, 1J the function f(X) = Xl assumes a maximum

value (+ 1), but not a minimum value, while on (-1,1) the function f(X) =
xi assumes a minimum value (0) but not a maximum value.
(0.2) f may assume an extremal value at more than one point.
On D = [ -1, 1J, f(X) = xi assumes a maximum value (1) at Xl = ± 1,

while on D = ~2, f(X) = xi assumes its minimum value (0) at every point
located on the X 2 axis.
The only reasonable conditions which guarantee the existence of extremal

values are contained in the following theorem whose proof is deferred.
(See Proposition 5.3.)
~ ~d is compact and f: D --+ ~ is continuous, then f as-

(0.3) Theorem. If D
sumes both maximum and minimum values on D.
In ~d, a compact set is a bounded set which is closed in that it contains

each of its boundary points. In particular, each "box" of the form
-
B ={
XdE ~ : aj ::; Xj ::; bj,]. = 1,2, ... , d }
for given real numbers aj::; bj, j = 1, 2, ... , d is compact. However, the
interval (-1, + 1) is not compact. (See §A.O.)
f: D --+ ~ is continuous at Xo E D iff for each e > 0,3 () > 0, such that when
XED and IX - Xol < (), then If(X) - f(Xo) I < e; and f is continuous on D
iff it is continuous at each point Xo E D.
The previous examples show that neither compactness nor continuity can
alone assure the existence of extremal values.
(0.4) The maximum value of f is the minimum value of - f and vice versa.
Thus it suffices to characterize the minimum points, those Xo ED for which

f(X) ~ f(X o), V XED. (3)
As we have seen, such points may be present even on a noncompact set.
(0.5) When D contains a neighborhood of X o, an extremal point of f, in which

f has continuous partial derivatives fXj = af/axj, j = 1, 2, ... , d, then for each
vector U E ~d of unit length, the (two-sided) directional derivative:
auf(Xo) ~ lim [f(Xo + eU) - f(X o)] = af (Xo + eU) I

£-+0 e ae £=0
=0.
[The bracketed quotient reverses sign as the sign of e is changed. The exis-
4 O. Review of Optimization in [Rd
tence and continuity of the partial derivatives ensures the existence of the
limit which must therefore be zero.J
Introducing the gradient vector Vfd,;! (fx" f X 2' ••• , fx), we may also ex-
press auf(Xo) = Vf(Xo )' U, and conclude that at such an interior extremal
point X o,
Vf(Xo) = (!). (4)
(0.6) The points Xo at which (4) holds, called stationary points (or critical
points) of f, need not give either a maximum or a minimum value of f.
For example, on D = [ -1, 1J, the function f(X) = x~ has Xl = 0 as its

only stationary point, but its maximum and minimum values occur at the
end points 1 and -1, respectively.
On D = ~2, the function f(X) = x~ - xi has Xo = (0,0) as its only critical

point; at X o, f has maximal behavior in one direction (X2 = 0) and minimal
behavior in another direction (Xl = 0).
In such cases, Xo is said to be a saddle point of f.
(0.7) A stationary point Xo may be (only) a local extremal point for f; i.e., one
for which f(X) ~ f(Xo) (or f(X) ~ f(Xo)) for all XED which are in some
neighborhood of Xo.
For example, the polynomial f(X) = x~ - 3x l has on D = [ - 3, 3J, sta-

tionary points at Xl = -1, 1; the first is (only) a local maximum point while
the second is (only) a local minimum point for f. (See Figure 0.1.)
(0.8) When f is a convex function on D then it assumes a minimum value at

each stationary point in D.
Figure 0.1
O. Review of Optimization in [Rd 5
For our purposes, f will be said to be convex on D when it has continuous

partial derivatives in D and satisfies the inequality
V X,XoED. (5)
A convex function need not have a stationary point, but obviously when
Xo is a stationary point of the convex function f; then by (4) Vf(Xo) = (D, so
that f(X) ~ f(Xo); i.e., (3) holds.
Observe that we can also express (5) by requiring that at each XED
f(X + V) ~ f(X) + Vf(X)· V, V VE ~d for which X + V ED. (6)
For example, in D = ~2, the function f(x l , x 2 ) = xi + x~, with gradient
Vf(X) = (2XI' 2x 2 ), satisfies for each V = (VI' v2 ) the inequality
f(X + V) = (Xl + vd 2 + (X2 + V 2 )2 = xi + x~ + 2X I VI + 2X 2 V 2 + vi + v~
=f(X) + Vf(X)· V + IVI 2
~ f(X) + Vf(X)' V
and so it is convex. At its only stationary point Xo = (D, it assumes its

minimum value. [Indeed, trivially, f(X) = xi + x~ ~ 0 = f((D).]
Moreover, with this example, there is only one minimum point. This
property is characteristic of strictly convex functions.
(0.9) When f is strictly convex on D; i.e., when (5) holds at each Xo ED with
equality iff X = X o, then f can have at most one stationary point, and hence, at
most one interior minimum point, in D.
[When Xo is a stationary point of a strictly convex function f, then f(X) >

f(X o), V XED", {Xo}. Thus f cannot assume its minimum value f(Xo) at
any other point.]
The function of the previous example is strictly convex on each set D S
~2. For other examples of convex functions and alternate characterizations
of convexity, see Problem 0.5, et seq.
if for all X in a neighborhood of X O,

(0.10) f is differentiable at Xo
f(X) = f(Xo) + Vf(Xo)' (X - Xo) + IX - Xo 13(X - X o), (7)
where 3(X - Xo) is a real valued function (defined for X '1= Xo by (7)) with zero
limit as X -+ Xo. Then the graph of f (in ~d+1) has at the point (Xo, f(Xo)) a
tangent hyperplane, namely, the graph of the affine function
T(X) ~ f(Xo) + Vf(Xo)' (X - Xo).
[For d = 1, this is just the line tangent to the curve representing the graph
of f in ~2; for d = 2, it is the plane tangent to the surface representing the
graph of f in [R3; for the general case, (7) may be used as a definition for
tangency.]
We can give the stationarity requirement Vf(Xo) = (!) the geometric inter-
pretation that the graph of f at Xo has at the point (Xo, f(Xo)) E [Rd+1 a
"horizontal" tangent hyperplane; i.e., a d-dimensional subset parallel to [Rd.
(See § 5.6.) Thus for d = 2, a marble "balanced" at (Xo, f(Xo)) should not roll
but remain "stationary." By (5), we see that a convex differentiable function
is one whose graph lies "above" its tangent hyperplanes.
NOTE: The existence of the partial derivatives offin a neighborhood of Xo
together with their continuity at Xo (as in (0.5)) are sufficient to guarantee
that (7) holds. (See A.7 and [Ed].)
(0.11) The classification of a stationary point Xo as a local maximum point,

local minimum point, or saddle point may be possible when f has higher-order
nonvanishing derivatives at Xo.
We first observe that vanishing higher-order derivatives provide no infor-

mation. On D = [Rl, for each n = 1,2,3, ... the functions
each have n derivatives which vanish at their common stationary point

Xo = (!), where f,. has a minimum value while gn has neither a (local) maxi-
mum nor a (local) minimum value.
In [Rl, it is well known that if at a stationary point, X o, f"(Xo) > 0
(f"(Xo) < 0) then f' is strictly increasing (decreasing) at Xo so that Xo is a
strict local minimum (maximum) point for f. The generalization to a higher-
dimensional space, where in addition, the possibility of Xo as saddle point
must be permitted, can best be approached through the second directional
derivative.
(0.12) Theorem. Let f: D --+ [R have continuous second-order partial derivatives

a2f i,j = 1,2, ... , d.
fx,xj = (fx,)Xj = ax.ax.;
J •
If Xo is a stationary point of fin D, and for each U = (u 1 , U2' ... , Ud) E [Rd:
+ eU) I.=0 = q(U) >

2
2 deC 0
auf(xo) = ae2f(XO 0, when lUI = 1, (8)
where
deC d
q(U) = L
i,j=l
fx,,)XO)UiUj,
then Xo is a strict local minimum point for f.
[Indeed, for each unit vector U, as e takes on real values in a one-

dimensional neighborhood of 0, j(e) ~ f(Xo + eU) takes on all the values of
Problems 7
f in a neighborhood of Xo in the (two-sided) direction of U. Now 1'(0) =

1
(8f/8e)(Xo + eU)I.=o = 0 so that has a stationary value at 0 and the require-
ment of the hypothesis is simply that 1"(0) > O. Thus f has at Xo strict local
minimal behavior in the (two-sided) direction U. This extends via a compact-
ness argument to establish that Xo is a (strict) local minimum point for f
[Ed].]
If 8U(Xo) has unlike sign for two directions, U, then f has saddle point
behavior at Xo.
(0.13) Remarks. If V = cU E !Rd, then q(V) = c 2 q(U). Hence (8) holds iff the
quadratic form q(V) > 0, V V E !Rd, V "1= (9.
When (8) holds, the (symmetric) Hessian matrix fxx(Xo) whose elements
are the second partial derivatives fx.x, J.(Xo) = fXjx.(X
, o), i, j = 1, 2, ... , d
(arranged in natural order), is said to be positive definite. Conditions which
characterize the positive definiteness of such matrices are known. (See Prob-
lem 0.10.) For the present, we observe that when (8) holds, the matrix fxx(Xo)
is invertible. [If [fxx] V = (9 for some V E !Rd, then by the laws of matrix
multiplication,
which by (8) is possible at Xo only if V = (9. This condition furnishes the

desired invertibility.] When q(V) ~ 0, V V E !Rd, the matrix fxx(Xo) is said to
be positive semidefinite.
(0.14) Unless D is open, i.e., has only interior points, then it is also necessary to
consider the extremal values of f on 8D, the boundary of D.
For example, although on !R 2 the function f(X) = x~ - xfhas its only

stationary point at Xo = (9, Xo is a saddle point, and so the maximum and
minimum values on say D = {X E R2: IXI ::; 2} can be found only along the
boundary where IXI = 2. In general for this set D, we would have to consider
the problem of optimizing f subject to the constraint g(X) ~ IXI 2 = 4.
To find the stationary points of f when so constrained to the level sets of

one (or more) functions such as g, we can employ the method of Lagrangian
multipliers. Since this method will be fully treated in a more abstract setting
(§5.7) we shall defer its further discussion. (See also [Ed].)
PROBLEMS
0.1. (a) Establish the Cauchy inequality (1). (Hint: If Y = (I), the inequality obviously
holds; assume Y #- (I), set f.l = (X· Y)/I Y1 2 , and consider IX - f.lYI2.)
(b) What can you deduce about the relationship between X and Y if equality
holds in (i)?
(c) Use the Cauchy inequality (1) to prove the triangle inequality (2a).
(d) Conclude that the reverse triangle inequality (2b) holds.
0.2. (a) Derive the inequality
IXIIYI- X· Y ~ IX - YI 2
(Hint: Show that IXII YI ~ t(IXI 2 + IY12) and add tlx - YI 2 to the right
side of this last inequality.)
(b) Use the result of part (a) to verify that
I~
IXI
_~I <
IYI -
.J2 IX -
JiXTIYi YI
for X, Y E [Rd, X #- (I), Y #- (I).
0.3. Find the maximum and minimum values (and the points at which they occur)
for
on
D = {X E [R2: IXjl ~ 2,j = 1,2}.
0.4. Find the maximum and minimum values (and the points at which they occur)
for
on
D = {X E [R2: IXI ~ 1}.
0.5. Which of the following functions are convex on D = [R2? Which are strictly
convex?
(a) f(X) = xi - x~. (f) f(X) = xi + x 2 •
(b) f(X) = Xl - X 2 • (g) f(X) = x1.
(c) f(X) = xi + x~ - 2Xl' (h) f(X) = sin(xl + X2)'
(d) f(X) = eX, + x~. (i) f(X) = xi - 2X1X2 + x~.
(e) f(X) = X1X2' (j) f(X) = aX l + bX2 + c, a, b, c E [R.
0.6. (a) Prove that the sum oftwo [strictly] convex functions on D ~ [Rd is [strictly]
convex on D.
(b) Is the sum of a convex function and a strictly convex function strictly
convex?
(c) Let f be strictly convex and c > O. Show that cf is strictly convex.
(d) Give an example to show that the product oftwo convex functions need not
be convex.
0.7*. Suppose thatf: [Rd -> [R is differentiable. Show thatfis convex on [Rd ifJfor each
X, Xo E [Rd,
f(tX + (1 - t)Xo) ~ tf(X) + (1 - t)f(Xo), V t, O<t<l.

(Hint: To prove the "if" statement, use equation (7) to express
f(tX + (1 - t)Xo) = f(Xo) + tVf(Xo)' (X - Xo) + t IX - Xo13(tX - tXo),
divide both sides of the resulting inequality by t, and let t approach zero.
For the "only if" part, set Y = tX + (1 - t)Xo, 0 < t < 1, and establish the
Problems 9
inequalities
f(Xo) ;;:::f(Y) + Vf(Y)'(Xo - Y),
(1 - t)
f(X) ;;:::f(Y) - --Vf(Y)·(Xo - Y).
t
Then, combine these last two inequalities to get the result.)
0.8*. Assume that f: jRd -+ jR has continuous second-order partial derivatives. Show
that f is convex on jRd iff for each Xo E jRd, the matrix of second partial deriva-
tives, fxx(X o), is positive semidefinite, i.e.,
d
L fx,x/XO)UiUj;;::: 0,
i,j=l
VUE jRd.
- def
(Hint: Use Taylor's theorem for f(t) = f(Xo + tU) where U = X - X o, t E jR.)
0.9*. Let D = jR2 and

2X1X2
-2--2' X,#(!J;
f(X) = { Xl X2 +
0, X=(!J.
(a) Verify that the partial derivatives af/aXl and af/aX2 both exist at X = (!J.
(b) Show that f is not continuous at X = (!J!
0.10. Let f: jR2 -+ jR have continuous second-order partial derivatives and let U =
(u l , u 2 ) be a unit vector.
(a) Verify that at Xo:
where
A = fx,x, (Xo), B = fX,X2(XO), C = f X2X2(XO)'
(b) If Xo is a stationary point of f, with both AC - B2 > 0 and A > 0,
conclude that Xo is a strict local minimum point for f (Hint: let U l = cos 0,
U 2 = sin 0.)
(c) Write the conditions of (b) in terms of subdeterminants of the 2 x 2 matrix
fxx, and conjecture a form for a corresponding set of conditions for the
general d x d matrix.
0.11. Is f(X) = IXI [strictly] convex on jRd - {OJ?
PART ONE
BASIC THEORY
Groningen,
January 1, 1697
AN ANNOUNCEMENT
"I, Johann Bernoulli, greet the most clever mathematicians in the world. Noth-
ing is more attractive to intelligent people than an honest, challenging problem
whose possible solution will bestow fame and remain as a lasting monument.
Following the example set by Pascal, Fermat, etc., I hope to earn the gratitude
of the entire scientific community by placing before the finest mathematicians
of our time a problem which will test their methods and the strength of their
intellect. If someone communicates to me the solution of the proposed problem,
I shall then publicly declare him worthy of praise."
11
CHAPTER 1
Standard Optimization Problems
"Which method is best?" is a question of perennial validity, and through the

centuries we have required for it answers of increasing sophistication. When
"best" can be assessed numerically, then this assessment may be regarded as
a real valued function of the method under consideration which is to be
optimized-either maximized or minimized. We are interested not only in
the optimum values which can be achieved, but also in the method (or
methods) which can produce these values.
When the questions arise from classical science, then it is usually one
of the fundamental quantities-length, area, volume, time, work or energy
which is to be optimized over an appropriate domain of functions describing
the particular class of method, process, or configuration so measured. Mod-
ern interests have added to this list cost, efficiency, ... , etc.
In this chapter, we shall examine several such problems-chiefly those of
classical origin which have been influential in the development of a theory to
furnish answers to the above questions. Although we can deal effectively with
only two of these in this introductory chapter, the rest will serve in formu-
lating the general class of problems to be considered in this text (§1.5).
§1.1. Geodesic Problems

Whether as a result of inherent laziness, or out of respect for efficiency, we
apparently have long wished to know which of the many paths connecting
two fixed points A and B is the shortest-i.e., has the least length. Although
in ~3 a straight line provides the shortest distance between two points (for
reasons substantiated below), in general it may not be reasonable (or possi-
13
14 1. Standard Optimization Problems
ble) to take this route because of natural obstacles, and then it is necessary to
consider the more complicated problem of finding the geodesic curves (i.e.,
those of least length) among those constrained to a given "hyper" surface. In
particular, we might wish to characterize in ~3, the geodesics on the surface
of a sphere, on a cylinder, or on a cone.
(a) Geodesics in IR.d

A curve in ~d, joining points A and B may be considered as the range of a
vector valued function yet) = (YI (t), Y2(t), ... , yit)), t E [0, 1J, with compo-
nents that are continuous on [0, 1J, such that YeO) = A and Y(l) = B. (When
d = 2 or 3, we may think of yet) as the position at "time" t, and see that the
componentwise continuity reflects our desire that the curve not have jumps.)
In particular, Yo(t) ~ (1 - t)A + tB, t E [0, 1J defines one such curve, namely,
the straight line segment determined by A and B (Figure 1.0).
With such generality, the curve may be nonrectifiable; i.e., it need not have
finite length. However, if we require more smoothness; if, for example, we
require that the component functions have continuous derivatives in (0, 1),
then we may think of Y'(t) = (Y~(t), y;(t), ... , y~(t)) as representing the veloc-
ity at time t with the associated speed IY'(t)l; hence the length of the curve
should be the distance travelled during this motion,
L(Y) ~ II IY'(t)1 dt,

(considered as a possibly improper Riemann integral). For finiteness of L, we
must require that each component of Y'(t) be integrable, and this is most
easily obtained by requiring that each component is continuously differ-
entiable on [0, 1]. Thus we are led to consider the purely mathematical
problem of minimizing this length function L(Y) over all vector valued func-
tions Y which meet the above conditions, as in Figure 1.0.
Figure 1.0
§l.l. Geodesic Problems 15
Since Yo(t) is surely admissible, with the derivative Y~(t) = B - A, we

know that if a curve of minimal length L min exists, then
Lmin:::;; II IY~(t)1 dt = IB - AI,
the Euclidean distance between A and B.

To verify the natural conjecture, that L min = IB - AI, we need only estab-
lish the inequality that for any admissible Y,
IB - AI :::;; L(Y) = II IY'(t)1 dt,
which is trivial when A = B; when A f= B, this is most conveniently done as

follows: Observe that from the fundamental theorem of calculus (A.8),
B - A = Y(l) - YeO) = II Y'(t) dt;
(i.e., Yi1) - YiO) = J6 yJ(t) dt,j = 1,2, ... , d). Thus

IB - AI2 = (B - A)·(B - A) = (B - A)·II Y'(t) dt
= II [(B-A)·Y'(t)]dt:::;; IIIB-AIIY'(t)ldt,
or
IB - AI2 :::;; IB - AI II IY'(t)1 dt,
and for A f= B, the desired inequality follows upon division by IB - AI. The
reader should verify each step in this chain. In obtaining the inequality
between the integrals, we utilized at each t E (0, 1), the Cauchy inequality in
the form
[(B - A)· Y'(t)] :::;; IB - All Y'(t)l,
and this pointwise inequality in turn implies that of the corresponding inte-
grals. [See A. 10.]
Can a nonstraight curve also be of minimal length? (See Problem 6.39.)
(b) Geodesics on a Sphere

For airlines facing fuel shortages, it is essential to know the shortest route
linking a pair of cities. Insofar as the earth can be regarded as a sphere, we
see that we require knowledge of the geodesics on the surface of a sphere
(albeit one which is large enough to be "above" the highest mountain range).
Each point Y = (YI, Y2' Y3) E 1R3 on the surface of a sphere of radius R
centered at (!) (except the north and south poles) is specified through its
Figure 1.1
spherical coordinates R, cp, and () as follows:

Y = (R cos () sin cp, R sin () sin cp, R cos cp), (1)
for unique cp E (0, n) and () E [0, 2n). Moreover, given distinct points A and B
on this surface, we suppose, as we may, that the axes are chosen so that
A is at the north pole (cp = 0), while B f= A has the spherical coordinates
(R, 0, CPl)' for CPl > 0. (See Figure 1.1.)
Then a curve joining A to B on the surface of this sphere is determined
through (1) by the pair of continuous functions ((}(t), cp(t)), t E [0, 1J, with say
cp(o) = 0; (}(1) = 0, cp(l) = CPl' (Note: To remain continuous () and cP may
have to exceed their ranges of [0, 2n) and [0, n), respectively. Also the choice
°
(}(1) = is made for convenience.) By the same considerations as in the
previous case we should require the continuous differentiability of () and cP on
[0, 1]. Then the resulting curve defined by
Y(t) = R(cos (}(t) sin cp(t), sin (}(t) sin cp(t), cos cp(t)), t E [0, 1J,
[which has at each t E (0,1) the derivative
Y'(t) = R[ - sin (}(sin cp)()' + cos (}(cos cp)cp', cos (}(sin cp)()'
+ sin (}(cos cp)cp', -(sin cp)cp'J(t)],
has the length
L(Y) = f\Y'(t)\ dt =R fal Jsin 2cp(t)(}'(t)2 + cp'(t)2 dt

~ R fal cp' (t) dt = Rcp(t) I:;
thus L(Y) ~ RCPl by our requirements on cpo (See Problem 1.7.)
§1.2. Time-of-Transit Problems 17
[Moreover, according to A.10, equality in the above occurs iff

(sin2 cp)(),2 = 0 and cp' ~ 0: or for cp E (0, n), when ()' == 0 so that () = const. =
()(1) = 0; this corresponds to the smaller great circle arc joining A to B.] Thus
we confirm the result known from antiquity that the shortest route joining
two points on a spherical surface is precisely along the (shorter) great circle
joining these points. Aircraft pilots are well advised to fly such routes, even
though they may have to travel over polar regions in doing so.
(c) Other Geodesic Problems

As the last example shows, to characterize the geodesics of a specific hyper-
surface in !Rd , it may be possible to utilize the properties of special coordi-
nates associated with that surface. For example, in !R 3 , we may use cylindrical
coordinates to search for geodesics on a cylinder, on a cone, and on a general
surface of revolution; some of the resulting problems will be examined in the
next chapter. Curiously, the original mathematical interest in finding the
geodesics on a general surface ofrevolution (as expressed by Jakob Bernoulli
in 1697) arose because of the then recent discovery that our planet is not
perfectly spherical.
The consideration of geodesics on a general surface S in !R 3 was begun by
Johann Bernoulli (1698) and his pupil Euler (c. 1728), continued by Lagrange
(1760) and treated rather decisively by Gauss (1827). When the surface can be
described by S = {Y E !R 3 : g(Y) = O} for some function g, we are required to
minimize a length integral of the form L(Y) = J61 Y'(t)1 dt as above, but
now subject to the Lagrangian constraint g(Y(t» == O. This problem forms a
branch of differential geometry, and it will be considered again in §6.7.
§1.2. Time-of-Transit Problems

If we travel at constant speed, then the geodesic routes determined in the last
section will also provide the least time of transit between given points A, B.
However, if we cannot travel with constant speed-and, in particular, if the
speed depends upon the path taken, then the problems of least distance and
least time in transit must be considered separately. In this section we shall
examine several such problems, including that of the brachistochrone (from
the Greek PprJ.xzar:or; == "shortest," xpovor; == "time") which has been very
significant in the emergence of the calculus of variations.
(a) The Brachistochrone

In falling under the action of gravity an object accelerates quite rapidly. Thus
it was natural for Galileo to wonder (c. 1637) whether a wire bent in the shape
x
A~~------~----------------.
x
Figure 1.2
of the circular arc shown in Figure 1.2 might not offer a faster time of transit
to a bead sliding down it under the action of gravity than would a straight
wire joining the same two points.
In 1696, Johann Bernoulli challenged mathematicians to find the bra-
chistochrone, that is, the planar curve which would provide the least time of
transit. His own solution was derived from an optical analogy, [see Problem
1.1]; and solutions were provided by his brother Jakob, by Newton, by Euler,
and by Leibniz. Although all of these reached the same conclusion-that the
brachistochrone is not the circular arc, but a cycloid-none of their solutions
is entirely satisfactory; however, that of Jakob Bernoulli admitted refinement
and far-reaching generalization: the variational calculus.
If we use the coordinate system shown in Figure 1.2, in which the initial
point A is the origin and the positive y axis is taken as vertically downward,
then a typical curve which might represent the brachistochrone joining A to
a lower point B = (x 1, Y 1) where x 1 and y 1 are positive, can be represented as
the graph of a continuous function y = y(x), X E [0, Xl] with y(O) = 0 and
y(x l ) = Yl. (Here we are abandoning the parametric form of representing
curves used in the previous section in favor of one less general but more
convenient. Although it is reasonable that the class of curves needed should
be so representable (Why?), the reader should consider whether something
essential is lost with this restriction.)
Assuming sufficient differentiability, this curve has length 1 and the time
required to travel along it is given through pure kinematic considerations as
T= T(y) = II ds,
o v
where v = ds/dt is the speed of travel at a distance s along the curve.
Now from calculus, for each x E [0, Xl], s = s(x) = JoJ1 + y'(~)2 d~ is
the arc length corresponding to the horizontal position x, and we may regard
v = v(x) as the associated speed. Thus with these substitutions,
_
T - T(y) -
_ IX. J1 +( y'(X)2
) dx.
o vx
§1.2. Time-of-Transit Problems 19
A~---------.----------------~
x
B
y
Figure 1.3
To express v in terms of y we resort to Newton's laws of dynamics.

Assuming that the gravitational acceleration, g, is constant during the fall,
and neglecting the effects of other forces (including that of friction), the
acceleration of a bead (of unit mass) along the wire at time t is
v = g cos 0(; also y = v cos 0(,
where 0( is the angle between the tangent line to the curve and the y axis at
this point (Figure 1.3). (We use the Newtonian notation of a dot to signify a
derivative with respect to time.)
Thus vv = gy, or upon integrating with respect to time,
v2 = 2gy + const.
But at A, v = y = 0, so that in general,
v= .,fiiY or v(x) = J2gy(x).
_ _ 1 i Xl (1 +
Thus finally,
y'(X)2)1/2
T - T(y) - M: ( ) dx, (2)
y2g 0 Yx
and we have the problem of minimizing T over all functions y = y(x) which
validate the above analysis. However, we may consider also the mathemati-
cal problem of minimizing the integral
(Xl (1 + y'(X)2)1/2 dx
Jo y(x)
over all functions y continuous on [0, Xl] for which y(o) = 0, y(xd = Yl' and
the integral is defined. This last condition requires that y have a derivative
integrable on [0, Xl] and that y be ~o with
(Xl
J 0 (y(x)t1/2 dx < +00.
In any case, there is no obvious answer although we may verify Galileo's

conjecture that a circular arc is superior to the straight line. (See Problem 1.2.)
XI X
Ar----------------T-
y
(a) (b)
Figure 1.4
Alternate forms of this problem soon suggested themselves. For example,

in 1697, Jakob Bernoulli challenged mathematicians to find the brachisto-
chrone among those curves representing travel over a fixed horizontal dis-
tance Xl' but for which the YI coordinate may vary, as in Figure 1.4(a). Later,
mathematicians investigated the problem of finding the brachistochrone
joining fixed curves as in Figure 1.4 (b).
Newton also considered the problem of finding the brachistochrone asso-
ciated with tunnels through the earth connecting fixed points. Again, the least
time is given not by the straight line, but by another type of cycloid [Sm].
It required nearly two more centuries to obtain, for these problems, the
mathematical solutions to be presented in this book. See §3.4, §6.2, §8.8, and
§9.2. At each stage of development, however, the brachistochrone remained a
testing ground for the state of the art.
(b) Steering and Control Problems

Closely related to the problem of the brachistochrone, is that from this
century concerning the best course to steer when navigating through a cur-
rent of varying speed. For example, which path joining fixed points A and B
on opposite banks of a river with varying current will provide minimum
transit time for a boat which travels with constant speed w with respect to the
water? As in Figure 1.5, we suppose the river banks parallel and utilize the
coordinate system shown, in which the Y axis represents one bank and the
line X = Xl the other. We also assume that w = 1 and that the river current r
is directed downstream and admits the prescription r = r(x), continuous on
[0, Xl]. Then the time of transit ofa boat travelling between the origin A and
the downstream point B = (Xl' YI) along a smooth path which is the graph
of a function Y = y(x) on [0, Xl] is given (after some work left to Problem 1.3)
by
(3)
§1.3. Isoperimetric Problems 21
Figure 1.5
where Q((x) = (1 - r2(x)t l /2. (In order that the crossing be possible we must
have 0 ::;; r(x) < 1.) (Why?) We are required to minimize this rather compli-
cated integral over all those functions y which are continuously differentiable
on [0, Xl], and satisfy y(O) = 0, y(x l ) = Yl. The methods of Chapter 3 will
provide access to a solution. (See Problem 3.20.)
We may also consider the more natural problems in which the river banks
are represented by curves, and permit the crossing points to vary. Finally, we
can also permit the current to vary with Y as well as x. In fact, in 1931,
Zermelo investigated the two-dimensional version of this problem which
could be equally significant to the piloting of a submarine or a light aircraft
[Cl And when we ask how to operate our craft so that it travels along an
optimal path, we enter the realm of optimal control problems, first consid-
ered by Minorsky around 1920.
(Problems 1.1-1.3, 1.8)
§1.3. Isoperimetric Problems

One of the most ancient optimization problems of which we have record is
that of finding the maximal area which can be enclosed by a curve of fixed
perimeter. According to Virgil, this problem was already of importance to
Dido of Carthage (c. 850 B.C.) and she supposedly obtained a solution on
heuristic grounds. (See Problem 1.5.) The Greek geometer Zenodoros appar-
ently knew that the circle provided a greater area than polygons having this
same perimeter, and a few centuries later Pappus (c. A.D. 390) concluded that
the circle was maximal among isoperimetric curves.
There is a simple physical analogy which supports this conjecture: Sup-
pose that a cylinder with thin inextensible impervious, but completely flexi-
ble walls is deformed so that its constant cross-sectional shape is that of the
area to be maximized. Then a section of it is placed with walls vertical on a
Figure 1.6
smooth horizontal surface and filled with water as in Figure 1.6. Under the
action of gravity, the water will seek its lowest level and if we assume none
lost by leakage at the base, then this will be accomplished by a movement of
the walls of the cylinder until the effects of hydrostatic pressure are equalized.
Since this pressure is exerted uniformly at each depth, the final configuration
must have constant curvature, i.e., it is that of a right circular cylinder.
However, the configuration which provides least depth must clearly be that
having maximal cross-sectional area. Thus the circle encloses a greater area
than any noncircular isoperimetric curve.
One mathematical formulation of this problem is as follows: We suppose
that a smooth simple closed curve oflength I is represented parametrically by
Y(t) = (x(t), y(t», t E [0, 1J, with Y(O) = Y(1) for closure (Figure 1.7).
Then according to Green's theorem [EdJ, the area of the domain D
bounded by the curve is
A(Y) = ft dx dy = fan x dy,
where aD denotes the boundary of D assumed positively oriented by the

parametrization through Y. Utilizing this parametrization we have
A(Y) = II x(t)y'(t) dt; (4)
we must maximize A(Y) over all functions Y(t) having continuously differ-
x
Figure 1.7
§1.3. Isoperimetric Problems 23
entiabl<i components on [0, 1J, which meet the closure conditions Y(O) =
Y(1), and for which the resulting curve satisfies the isoperimetric condition
L(Y) == Il IY'(t)1 dt = [ (5)

for a given [.
We shall return to the solution of this problem in §6.7.
Remark. This problem has received much attention and there have been
other less restrictive formulations. In particular, the German geometer Jakob
Steiner (c. 1840) devised several techniques to attack it [P]. One well-known
analytical solution utilizes properties of Fourier series, (see Problem 1.6.),
but in §8.8, we will present a recently discovered solution that seems more
natural.
A modern version which combines features of the isoperimetric and steer-

ing problems from the previous section is due to Chaplygin (1938). It consists
of describing the closed path which an airplane say, on reconnaisance, should
fly with constant airspeed in the presence of a wind, in order to enclose the
maximum area. When the wind speed is zero, then this problem is equivalent
to the classical isoperimetric problem (Problem 1.4).
Zenodoros considered also the higher-dimensional analogue of maximiz-
ing the volume of a solid having fixed surface area (Problem 1.9).
A different isoperimetric problem (often attributed to Euler)l consists of

finding the shape which a thin inextensible long cable or chain of weight/unit
length W, and given length L, will assume under its own weight when sup-
ported freely from end points separated a fixed horizontal distance H. As we
shall show in §3.5, this requires the minimization of the center-of-mass inte-
gral F(y) = W J~ y(s) ds over all functions y ~ 0 continuously differentiable in
[0, LJ, with y(O) = y(L) = 0, which satisfy the "isoperimetric" condition
G(y) == IL J1 - y'(S)2 ds = H. (6)
In general, the term "isoperimetric" is assigned to any optimization prob-

lem in which the class of competing functions is subject to integral or global
constraints of the form (5) or (6). The resulting isoperimetric problems admit
simple abstract formulation (§5.7) in contrast to Lagrangian problems (such
as those for geodesics on a general surface) in which intrinsic constraining
conditions must be satisfied (§6.7).
(Pro blems 1.4-1.6)
1 This problem was first proposed by Galileo (who believed that a parabolic shape would be
optimal) and it was then attacked mathematically by the Bernoullis in 1701. (See Goldstine.)
§1.4. Surface Area Problems

A higher-dimensional analogue of the geodesic problem discussed in §1.1,
might be formulated as follows:
Find the surface of minimum area that spans fixed closed curves in \R 3 •
(a) Minimal Surface of Revolution

For example, when the curves consist of a pair of "concentric" parallel circles
such as those shown in Figure 1.8, then we could ask for the surface of
revolution which "joins" them and has minimum area-or, equivalently, we
would wish to find the shape of its boundary curve. The problem in this form
was first attacked by Euler (c. 1744) who employed his recently developed
theory of the calculus of variations in its solution. .
If we utilize the coordinate system shown, then we would be led by ele-
r r
mentary calculus to minimize the surface area function
S(y) = 2n y(x) ds(x) = 2n y(x)J1 + y'(X)2 dx (7)
among all functions y which are nonnegative, continuously differentiable on

[a, b], and, for which
y(a) = a1 and y(b) = b1 •
Here, a 1 and b1 denote the given radii of the bounding circles, one of which
may degenerate to a point. (However, see Problem 1.7.)
When a1 and b1 are comparable to b - a as in Figure 1.8, it is reasonable
to expect a minimizing y to be of this form. However, when b - a greatly
exceeds a 1 and b1 , as in Figure 1.9, then it is seen that the surface area can be
Figure 1.8
§1.4. Surface Area Problems 25
Figure 1.9
made as close as we please to the area of the bounding disks-and that these
probably represent the "least" area-but the associated "boundary" curve is
not of the form admitted.
We have two alternatives: First, we can simply predict that in this case the
problem as posed has no solution and attempt to substantiate this. Or, we
can seek a reformulation of the problem in which such cornered curves
remain admissible. We shall adopt the second alternative when we return to
this problem in §7.S. For the present, we note that a framework large enough
to include this alternative must accomodate an accurate description of cor-
nered curves.
(b) Minimal Area Problem

Suppose that the competing surfaces can be represented as the graphs of
smooth functions u = u(x, y) defined on a common planar domain D, as in
Figure 1.10. Then the associated surface area is given by calculus as
S(u) = Iv J1 + u~ + u; dA,
where dA denotes the two-dimensional element of integration over D; the
boundary of D is denoted by aD, and in this text, we suppose it to be so
well-behaved that Riemann integration of continuous functions can be de-
fined over D or jj and over aD [Ed].
We would then seek the minimum for S(u) over all functions u which are
continuous on jj = D u aD, continuously differentiable inside D, and have
given continuous boundary values UlaD = y, say. We shall obtain some par-
tial results for this problem, which need not have a solution, in §3.4(e), under
the assumption that D is a Green's domain (one whose boundary smoothness
admits use of Green's theorem [F1].)
Figure 1.10
(c) Plateau's Problem

A new impetus was given to this class of problems in 1873 when the Belgian
mathematical physicist Joseph Plateau noticed that wires bent in the shape
of the bounding curves could be made to support a thin film of glycerine
which in order to minimize the surface tension should assume the shape
associated with the minimal area. These "soap film" experiments have been
continued until this day; they show that for some configurations, more than
one type of solution is possible, and in some cases the solutions can be
observed to change form as the relative geometry is altered. For example,
from our discussion of the minimal surface of revolution, we would expect
that a soap film joining a pair of circular rings that is initially cylindrical in
shape might transform into a pair of disks as the distance between the rings
is increased.
The mathematics supporting a general theory for such surfaces is outside
the scope of the present text. (See, for example, [Os].)
(Problems 1.7-1.9)
§1.5. Summary: Plan of the Text

These problems exhibit common ingredients. Each requires the optimization
of a real valued function F defined on a certain domain rlZ of functions Y,
r
usually by means of an integral of the form
F(Y) f(x, Y(x), Y'(x)) dx, (8)
for some given real valued function f. Here Y is a real [or possibly a vector]
valued function [each component of] which is continuous on [a, b] with a
continuous derivative in (a, b); and rlZ, in general, consists of those functions
§1.5. Summary: Plan of the Text 27
of this class which meet certain specified boundary conditions at a and/or b.

However, as in the case of the brachistochrone, it may be necessary to impose
further restrictions such as requiring that Y = Y ~ 0 on [a, b]. In addition,
there may be constraints of the isoperimetric form in which
derfb g{x, Y{x), Y'{x» dx =
G{Y) = a I; (9)
or of the Lagrangian form in which

g{x, Y{x), Y'{x» == 0, X E (a, b), (lO)
for given functions 9 and constants 1.
We may also find it necessary to enlarge the domain ~ to include func-
tions with corner points; i.e., discontinuous derivatives, and to permit more
freedom of the boundary points such as that required to treat the brachisto-
chrone joining fixed curves.
Finally, we should also consider the obvious extensions to integrals in-
volving higher derivatives of the functions and to functions of more than one
variable.
In the generality considered it is seen that the theory should be applicable
to any problem of finding the "best" process which meets the following
criteria:
(l) The processes can be described by a suitable class of functions.
(2) The value of each process can be measured by a weighted averaging
usually represented through an integral involving the function and one or
more of its derivatives.
Thus, for example, it should be applicable to problems of minimizing the
strain energy stored in a deformed elastic body such as a bent beam. And it
should also provide some insight as to how an economic process should
be operated over a fixed period of time in order to minimize the cost of
operation-or maximize the return on investment.
A complete solution to any of these problems requires characterizing
those functions in the domain which could produce the extremal values of
interest (either Fmax or F min ) and proving that these are indeed the extremal
values sought. In this chapter we have obtained a satisfactory solution to
only one of our problems-that of finding the geodesics on a sphere (§l.1{b».
There, a proper (and natural) selection of the domain ~ led directly to the
inequality
L (curve) ~ L (shorter great circle joining same points),
with equality iff the curve is the great circle. Moreover, it was not necessary
to "know" or guess the answer beforehand as we did when proving that the
geodesics in ~3 are the straight lines.
The general theory known as the calculus of variations has been de-
veloped over the last three centuries to handle such problems. It arose out of
efforts to duplicate, insofar as possible, the analysis of optimization of ordi-

nary functions defined on a set in ~d (a review of which is given in Chapter 0).
From the theoretical calculus of variations we learn that functions which
could provide the extremal values sought must satisfy the differential equa-
tions of Euler-Lagrange determined by the integrand f (and g, if constraints
of the form (9) or (10) are present). On the other hand, solution to these
equations need not be possible within the required class (.@), and even when
possible, need not furnish the extremal values sought.
In ~d, the minimization of convex functions is simple to analyze, and in
Chapter 3 we shall see that suitable convexity of F again provides access
to the solution of some of these problems-even when isoperimetric or
Lagrangian constraints are present.
However, this convexity is best formulated in terms of the directional
derivatives of F (its Gateaux variations), in the setting of a linear space
(Chapter 2).
After the brief technical Chapter 4, we return to this space in Chapter 5,
supply it with a norm and obtain conditions (again expressed through
Gateaux variations) which are necessarily satisfied, if a function is minimized
or maximized locally on a subset of this normed linear space. In this setting
we shall also develop the method of Lagrangian multipliers to treat optimiza-
tion with constraints of the isoperimetric type (9).
In Chapter 6, we apply the results of Chapter 5 to functions F defined
by integrals such as (8) to obtain the classical theory of Euler (1742) and
Lagrange (1755). The effects of constraints of various types are also consid-
ered. Extensions to integrals involving higher-ordered derivatives or vector
valued functions, and to multidimensional integrals will also be given. How-
ever, we reserve the generalization to piecewise C 1 (cornered) extremals for
presentation in Chapter 7, where it will also be shown that the existence of a
minimizing function demands some convexity from the integrand f (§7.6).
In Chapter 9, we demonstrate that, conversely, such partial convexity of
the integrand ensures that some solutions obtained previously do minimize
their integral functions, provided that a suitable family of these solutions is
available. These sufficiency arguments, which here incorporate earlier results
from Legendre and Jacobi, are now known as field theory and are due largely
to Weierstrass (c. 1879) and Hilbert (1900). (The so-called direct methods,
also initiated by Hilbert, but developed by Tonelli and his successors from
1915, wherein existence of a minimizing function is established through a
priori estimates, are considered to lie outside the scope of this text. See [G-F],
[Ak].)
However, one of the most satisfying uses of the Euler-Lagrange theory has
come from another direction, namely, the recognition that differential equa-
tions which describe a physical process might be regarded as the Euler-
Lagrange equations for a variational problem. This generated the variational
principles of mechanics culminating in the Hamilton (c. 1835)-Jacobi (c. 1840)
theory, which has survived the transition from classical physics to quantum
mechanics, and which properly interpreted, is still regarded as the correct
Notation: Uses and Abuses 29
theoretical foundation for the laws that govern the operation of our universe.
We shall examine this aspect of the subject in Chapter 8 (which is essentially
independent of Chapters 7 and 9).
In the concluding part (Part III), we turn to problems of optimal control
where the vector-state ¥ of a system at time t is governed by a dynamical
system of differential equations of the vector form
¥'(t) = G(t, ¥(t), U(t))
dependent upon a vector "control" function U(t). The task is to determine a
"path" ¥o and a control Uo that optimizes some performance assessment
integral of the form
F(Y, U) = r F(t, ¥(t), U(t)) dt
perhaps subject to certain endpoint or other constraints, and control restric-

tions. (For example, an engine throttle can open only so far.)
In Chapter 10 we use partial convexity of some associated integrals to
suggest a governing principle (due to Pontjragin, c. 1960) guaranteeing opti-
mality of the choices that it dictates. Then in Chapter 11, we will learn that
these Pontjragin conditions are necessary for the optimal choices, whether or
not such convexity is present. Finally we use the principle to obtain necessary
conditions when other types of Lagrangian or isoperimetric constants are
present.
To establish the above assertions, we shall use several results from ad-
vanced calculus and the theory of differential equations, proofs for most of
which will be found in the Appendix.
Notation: Uses and Abuses

In this book, we use the following abbreviations which are now standard in
mathematical literature:
E for "is an element of" or "is in" or "in";
3 for "there exists";
'if for "for each" or "for every";
iff for "if and only if";
e1 for "continuously differentiable";
e2 for "twice continuously differentiable";
en for "n times continuously differentiable";
IRd for "d-dimensional Euclidean space";
o for "end of proof";
x?a for "x approaches a from below";
x'>.a for "x approaches a from above";
def
for "defined as";
* to designate material of more than average difficulty.
In addition, we shall need notation which is not standard in order to

handle effectively composite functions such as
f(x, y(x), y'(x)) (or f(x, Y(x), Y'(x)))
and their derivatives.
First, we write f = f(x, y, z) for which the partial derivatives are as usual
denoted
fAx, y, z), /y(x, y, z), fAx, y, z), fxx(x, y, z), fxy(x, y, z), fxAx, y, z), ... , etc.
Then we evaluate at (x, y(x), y'(x)), and denote the respective results by
f[y(x)], fx[Y(x)], /y[y(x)], fz[Y(x)],fxx[Y(x)], fxy[Y(x)], fxz[Y(x)], ... , etc.
However, for a given function y = y(x), in Chapter 6 we further condense
the notation to
f(x) = f[y(x)] = f(x, y(x), y'(x)),
fAx) = fx[Y(x)] = fAx, y(x), y'(x)), .
fy(x) = /y[y(x)] = /y(x, y(x), y'(x)),
/y,(x) = fz[Y(x)] = fAx, y(x), y'(x)),
fy'y'(x) = fzz[Y(x)]; ... , etc.,
where use of the subscript (y') is consistent with regarding f(x, y, z) as
f(x, y, y'), an approach taken by some authors.
Vectors will always be indicated by capital letters. For vector valued
functions Y = Y(x), the corresponding abbreviations when f = f(x, Y, Z), are
as follows:
f(x) = f[Y(x)] = f(x, Y(x), Y'(x)),
fAx) = fx[Y(x)] = fx(x, Y(x), Y'(x)),
fy(x) = fy[Y(x)] = fy(x, Y(x), Y'(x)),
fy,(x) = fz[Y(x)] = fz(x, Y(x), Y'(x)).
where fy(x, Y, Z) is the vector withjth component /Yj(x, Y, Z); and again use
of the subscript (Y') is consistent with regarding f(x, Y, Z) as f(x, Y, Y').
Observe that with this notation, the chain rule takes the following forms
:xf[Y(x)] = fx[Y(X)] + fy[Y(x)]· Y'(x) + fz[Y(x)]· Y"(x)

or
d
dx f(x) = fx(x) + fy(x)· Y' (x) + fy,(x)· Y" (x),
assuming sufficient differentiability. (In the later chapters, we will utilize a

notation for the matrices of the second partial derivatives (and similar expres-
sions) which is consistent with that above. It will be explained as it is required.)
Problems 31
Finally, because of frequency of occurrence, we simplify (y'(X))2 to y'(X)2,

and make corresponding reductions for (y"(x)?, (y'(X))3, ... , etc., while
(Y(X))2 is giveh either of the forms y2(X) or Y(X)2, as desired.
PROBLEMS
1.1. The Brachistochrone. (The following optical analogy was used by Johann
Bernoulli, in 1696, to solve the brachistochrone problem.) In a nonuniform me-
dium, the speed of light is not constant, but varies inversely with the index of
refraction. According to Fermat's principle of geometric optics, a light ray travel-
ling between two points in such a medium follows the "quickest" path joining the
points. Bernoulli thus concluded that finding the brachistochrone is equivalent to
finding the path of a light ray in a planar medium whose index of refraction is
inversely proportional to Jy.
The optics problem can be solved by using Snell's law which states that at
each point along the path of a light ray, the sine of the angle which the path
makes with the y-axis is inversely proportional to the index of refraction (and
hence proportional to the speed). Therefore, the brachistochrone should satisfy
c sin IX = Jy, (11)
where c is a constant and IX is as shown in Figure 1.3.

(a) Assuming that the brachistochrone joining (0,0) to (Xl' yd can be repre-
sented as the graph of a smooth function y = y(x), use (11) to prove that
(b) Show that the cycloid given parametrically by

c2
x(B) = 2(B - sin B),
c2
y(B) = 2(1 - cos B),
satisfies the differential equation found in part (a) and has (0, 0) as one end
point. (It will be shown in §3.4 that c and B1 can always be chosen to make
(Xl> Y1) the other end point, and that the resulting curve is expressible in the
form y = y(x).)
(Although this does not constitute a mathematically rigorous solution to the
problem, it illustrates an important parallel between geometric optics and parti-
cle mechanics which led to the works of Hamilton.)
1.2. A Brachistochrone. (See §1.2(a).) Let Xl = Y1 = 1.
(a) Use equation (2) to compute T(y) for the straight line path y = x.
(b) Use equation (2) and the trigonometric substitution 1 - X = cos B to show
that for the circular arc y = Jl - (x - 1)2,
T(y) = (2g)-1/2 J: 1
2 (sin B)-1 /2 dB.
(c) Use a table of definite integrals to conclude that the circular arc in part (b)
provides a smaller transit time than the line segment in part (a).
(d) Use the inequality sin e : :; e, e ~ 0, to obtain a lower estimate for the transit
time of the circular arc of (b);
(e)* Find similar (but more precise) upper estimates which lead to the same
conclusion as in (c) without obtaining a numerical value for the integral in
(b).
1.3. Transit Time of a Boat. (See §1.2(b).) Use the following steps to derive equation
(3):
(a) Show that the x- and y-components of the velocity of the boat are given
respectively by cos u and r + sin u, where u is the steering angle of the boat
relative to the x-axis shown in Figure 1.5.
(b) Prove that the crossing time is given by
(Xl
T= Jo sec u dx,
(c) Show that y' = r sec u + Jsec 2 u - 1.

(d) Conclude that
T(y) = f:'[1X(x)J1 + (1Xy')2(X) - (1X2ry')(x)] dx,
where 1X(X) = (1 - r2(xW 1/2.
1.4. Chaplygin's Problem. A small airplane in level flight with constant unit airspeed
flies along a simple smooth closed loop in one hour in the presence of a wind with
constant direction and speed w :::;; 1. Suppose that its ground position at time t is
given by Y(t) = (x(t), y(t)) where the wind is in the direction ofthe positive x-axis.
(a) Argue that (x'(t) - W)2 + y'(t)2 ;: 1, t E [0, 1], while A(Y) = J6
x(t)y'(t) dt
represents the area enclosed by the flight path.
(b) Formulate the problem of finding the flight path(s) maximizing the area
enclosed. (We return to this formulation in Problem 9.19.)
(c) When w = 0, show why a solution of the problem in (b) would solve the
classical isoperimetric problem of §1.3.
(d) As formulated in (b) Chaplygin's problem is not isoperimetric. (Why not?)
Recast it as an isoperimetric problem in terms of u(t), the steering angle at
time t between the wind direction and the longitudinal axis of the plane.
(Hint: Take y(o) = Y(l) = (9 and conclude that x(t) = wt + J~ cos u('t') d-r:,
while y(t) = J~ sin u(-t') d-r:).
1.5. Queen Dido's Conjecture. According to Virgil, Dido (of Carthage), when told
that her province would consist only of as much land as could be enclosed by the
hide of a bull, tore the hide into thin strips and used them to form a long rope of
length I with ends anchored to the "straight" Mediterranean coast as shown in
Figure 1.11. The rope itself was arranged in a semicircle which she believed would
result in the maximum "enclosed" province. And thus was Carthage founded-in
mythology.
Problems 33
y
Dido's
Solution ----+ ,.,... - -~
x(s), y(s)~,
/ /
/
\\
/ /
/
/
./
Figure 1.11
For simplicity, suppose that I = n; then use the arc length s as the parameter.
(a) Show that the conjectured inequality for the vector valued function Y =
(x, y) with y(o) = y(n) = 0, is
A(Y) = - J: y(s)x'(s) ds ::;; i G) J:

= [X'(S)2 + y'(sf] ds.
Hint: Use Green's theorem.

(b) Prove that the inequality in (a) is satisfied if J~ [y'(S)2 - y2(S)] ds ;;::: 0, when
y(o) = y(n) = 0, and y is continuously differentiable on [0, n]. (We shall
establish this inequality in §9.1.)
(c) Show that equality in (a) and (b) would imply that x'(s) = - y(s).
(d) Verify that equality will hold in (a) and (b) for the trigonometric functions
x(s) = cos s, y(s) = sin s, defining the semicircle.
(e) How is Dido's problem related to that considered in §1.3? Could a solution
to one of these problems yield a solution to the other?
1.6. The Isoperimetric Inequality. (See §1.3 and §9.5.)

To derive a formulation analogous to that in the preceding problem, use the
arc length parametrization of a closed curve of length I = 2n.
(a) Show that for Y = (x, y) so parametrized, with say y(o) = Y(2n) = (D, the
isoperimetric inequality A(Y) ::;; n, would follow from Wirtinger's inequality:
Jfk [y' (sf -

0 Y(S)2] ds ;;::: 0, when y(o) = y(2n) = 0, and Jfk y(s)ds = 0,
0
where y is continuously differentiable on [0, 2n]. Hint: What is the geomet-

rical significance (for the curve) of the last integral requirement?
(b) Prove that equality throughout in (a) would require that x'(s) = - y(s), with
x(o) = x(2n) = 0.
(c) Verify that the trigonometric functions x(s) = cos s - 1, y(s) = sin s meet all
of the requirements in (a), (b), and give equality throughout. What curve do
they parametrize?
(d) Show that the integral inequality in (a) is violated for the function Yl (s) =
sin s/2. Is WYl (s) ds = O?
1.7. Degenerate Minimal Surface of Revolution. (See §1.4(a).)

When one of the end circles degenerates to a point, then we can suppose that
a typical curve to be rotated is the graph of a continuously differentiable function
Y = y(x), 0 ~ x ~ b, with 0 = y(O) ~ y(x); y(b) = b1 > O. Prove that the resulting
surface area S(y) = 2n St y(x)jl + y'(xf dx ;::: nbr. Give an interpretation of
this inequality. Hint:J1+C2 ;: : c.
1.8. A Seismic Wave Problem.

Suppose that a seismic disturbance or wave travels with a speed which
is a linear function of its depth ,., below the earth's surface (assumed flat)
along a subterranean path which minimizes its time of transit between fixed
points.
Assume that a typical (planar) path joining the points is the graph of a
continuously differentiable function y(x) = ,.,(x) + (1, a ~ x ~ b, where (1 is a posi-
tive absolute constant, and,., is the positive local depth of the path, as shown in
Figure 1.12(a); y(a) = a 1 and y(b) = b1 •
a b
0,---,-------.-------,--. o,----,----------------r-~
x y
Figure 1.12
(a) Show that the time of transit along this path is for some positive constant p,
given by T(y) = p S:(jl + y'(X)2/y(X)) dx. For further analysis of this
problem, see Problem 8.29 and §9.3.)
(b) When the x- and y-axes are interchanged, if this path can be described as the
graph of a function y = y(x) as in Figure 1.12(b), show that the time of
transit integral is for a < h given by T(y) = p S:(jl + y'(xY/x) dx. This for-
mulation is examined further in Problem 3.23.
1.9. The Zenodoros Problem.

(a) Show that the problem of finding the solid of revolution of maximal volume
enclosed by a surface of fixed total area 2nT leads to the consideration of
Problems 35
maxlmlzmg V(y) = 11: J~ y2 dx where y;;::: 0, with y(O) = y(b) = 0, subject to

the isoperimetric constraint
211:T = S(y) == 211: J: y(x) ds(x) = 211: J: y(x)J1 + Y'(X)2 dx,

where b > 0 must also be varied. (Why?)
J
(b)* Use Euler's substitution t = y ds, so that 0 :5; t :5; T, to reduce this to the
non-isoperimetric problem of maximizing
V(y) = 11: LT y2(t)X'(t) dt = 11: LT y(t)J1 - (yy'f(t) dt,
where y = y(t) is continuously differentiable on [0, T], and 0 :5; y(t), with
y(O) = y(T) = O.
The analysis is continued in Problems 6.43, 8.27 and 9.12.
CHAPTER 2
Linear Spaces and Gateaux Variations
Each problem considered previously reduces to that of optimizing (usually

minimizing) a real valued function J defined on a subset ~ of a linear space
qIj. In the present chapter we shall view problems in this context and intro-
duce the associated directional derivatives (Gateaux variations) of the func-
tions which will be required for what follows. We begin with a catalogue of
standard linear spaces presupposing some familiarity with vector space oper-
ations, with continuity, and with differentiability in !Rd.
§2.1. Real Linear Spaces

All functions considered in this text are assumed to be real valued or real
vector valued. The principal requirement of a real linear (or vector) space of
functions is that it contain the sums and (real) scalar multiples of those
functions. We remark without proof that the collection of real valued func-
tions f, g, on a (nonempty) set S forms a real linear space (or vector space)
with respect to the operations of pointwise addition:
(f + g)(x) = f(x) + g(x), \/XES
and scalar multiplication:
(cf)(x) = cf(X), \/ XES, C E!R. ([N]).
Similarly, for each d = 1,2, ... the collection of all d-dimensional real vector
valued functions on this set S forms a linear space with respect to the follow-
ing operations of componentwise addition and scalar multiplication: if F =
(fl' f2'···' h) and G = (gl'···' gd), where jj and gj are real valued functions
36
§2.1. Real Linear Spaces 37
on S for j = 1, 2, ... , d, so that F(x) = (i1 (X),f2(X), ... ,fAx)), and G(x) =
(gl (x), g2(X), ... , gAx)), V XES, then
(F + G)(x) = + G(x)
F(x)
def! 1
= ( l(X) + gl(X),f2(X) + g2(X), ... , Jd(x) + gAx))
and
(cF)(x) = (cif1 (x), C!2(X), ... , CJd(x)),
= cF(x) def .1
V XES.
It follows that each subspace of these spaces, i.e., each subset which is closed
under the defining operations of addition and scalar multiplication, is itself a
real linear space.
In particular, if continuity is definable on S, then C(S) (= CO (S)), the set of
continuous real valued functions on S, will be a real linear space since the
sum of continuous functions, or the multiple of a continuous function by a
real constant, is again a continuous function. Similarly, for each open subset
D of Euclidean space and each m = 1, 2, ... , cm(D), the set of functions on D
having continuous partial derivatives of order :5:.m, is a real linear space,
since the laws of differentiation guarantee that the sum or scalar multiple of
such functions will be another. In addition, if D is bounded with boundary
aD, and 15 = D u aD, then C m (15), the subset of Cm(D) 11 C(15) consisting of
those functions whose partial derivatives of order :5:. m each admit continuous
extension to 15, is a real linear space.
For example, when a, b E ~, then (a, b) = [a, b], is a closed and bounded
interval. A function y, which is continuous on [a, b], is in C 1 [a, b] 1 if it is
continuously differentiable in (a, b) and its derivative y' has finite limiting
values from the right at a (denoted y'(a+)) and from the left at b (denoted
y'(b-)). When no confusion can arise we shall use the simpler notations y'(a)
and y'(b), respectively, for these values, with a similar convention for higher
derivatives at a, b, when present. Observe that yo(x) = X 3/2 does define a
function in C 1 [0, 1] while Y1 (x) = X 1/2 does not.
Finally, for d = 1, 2, ... , [C(S)]d, [Cm(D)]d, and [C m(15)]d, the sets of
d-dimensional vector valued functions whose components are in C(S), Cm(D),
and Cm (15), respectively, also form real linear spaces.
We know that subsets!!) of these spaces provide natural domains for
optimization of the real valued functions in Chapter 1. However, these sub-
sets do not in general constitute linear spaces themselves. For example,
!!) = {y EC[a, b]: y(a) = 0, y(b) = 1}
is not a linear space since if y E!!) then 2y rf.!!}, (2y(b) = 2(1) = 2 "# 1.)
However,
!!)o = {y E C[a, b]: y(a) = y(b) = O}
is a linear space. (Why?)
In the sequel we shall assume the presence of a real linear space qy consist-
1 We abbreviate C((a, b)) by C(a, b), C1([a, b]) by C 1[a, b], etc.
38 2. Linear Spaces and Gateaux Variations
ing of points (or vectors), y, in which are defined the operations of (vector)
addition and (real) scalar multiplication obeying the usual commutative,
associative, and distributive laws. In particular, there is a unique vector (9
such that c(!J = oy = (9, \;f Y E qy, C E IR; we also adopt the standard abbrevia-
tions that 1y = yand -1y = - y, \;f y E qy.
§2.2. Functions from Linear Spaces

Although we may wish to optimize a real valued function J over a subset ~
of a linear space qy, it is frequently the case that the natural domain ~* of J
is larger than ~, and may be qy itself.
Example 1.
J(y) = r [sin 3 x + y2(X)] dx
is defined on all of qy = C [a, b], since each continuous function y E qy results
in a continuous integrand, sin 3 x + y2(X), whose integral is finite.
Example 2.
J(y) = r p(x)J1 + y'(X)2 dx, with p E C[a, b],
is defined for each y E qy = C l [a, b] since the assumption that y has a deriva-
tive on (a, b), which has a continuous extension to [a, b], again results in a
continuous integrand.
(Actually, J remains defined on qy, when p is (Riemann) integrable over
[a, b].)
Example 3. The brachistochrone function of §1.2(a),
_
T(y) -
1
M:
IX' J1 +r::t::\
y'(X)2
dx,
v' 2g 0 v' y(x)
is not defined on qy = C l [0, Xl] because of the presence of the term JM in
the denominator of the integrand. It is defined on the subset
~* = {y E Cl[O, Xl]: y(x) ~ 0, \;f X E (0, Xl), and f:' (y(x)fl/2 dx < +oo},
which is not a linear space. (Why not?)
Example 4. When f E C( [a, b] x 1R2), then
F(y) = r f(x, y(x), y'(x)) dx

§2.3. Fundamentals of Optimization 39
is defined on If!J == C 1 [a, b], since for each y E I1JJ, the composite function
f[y(x)] = f(x, y(x), y'(x)) E C[a, b].
However, if fE C([a, b] x D) where D is a domain in 1R2, then F is defined
only on a subset of
!!}* = {y E C1[a, b] : (y(x), y'(x)) E D, V X E [a, b]}.
Example 5. For each d = 1, 2, ... , the evaluation function L(Y) = Y(a) is

defined on !{If = (C[a, b]t It is even linear in that
L(cY + eli) = cL(Y) + CL(Y), V c, C E IR and Y, Y E CWo
Also linear are
L 1 (Y) = Y'(a; b)
and
. L(y) = r 3xy(x) dx on C[a, b].
However, most functions of interest to us in this text are (highly) nonlinear.
Example 6. If J and J are real valued functions defined on a subset !!}* of any
linear space 1f!J, then for c, CE IR
cJ, cJ + eJ, JJ, e J , sin J,
1/J, fl, tan J, need not be defined. Thus
r
are also defined on !!}*; but
J(y) = Jl + y'(X)2 dx + 2y(a)

is defined on C 1 [a, b], but
J 1 (y) = y~) is not.

Example 7. If D is a sufficiently nice bounded domain in 1R2, then as in §1.4(b),
S(u) = LJ 1 + u; + u; dA
is defined VUE C 1 (D).
(Problems 2.1-2.3)
§2.3. Fundamentals of Optimization

When J is a real valued function defined on a subset!!} of a linear space 1f!J, as
in the previous section, there may be interest in the extremal values of J
which would occur at those points Yo E !!} for which either
V Y E!!};
or
VYE~,
and in finding these points if they exist.

Since the latter points are those for which -J(y) ~ -J(yo), V y E~, it
will suffice to characterize the former-those minimum points Yo which
produce minimal values of J on ~. We say that a point Yo E ~ minimizes
J on ~ uniquely when it is the only such point in ~, or equivalently, when
J(y) ~ J(yo), V y E ~ with equality iff y = Yo.
For example, J(y) ~ Jb y2(X) dx ~ 0 = J(yo), if Yo(x) = 0, so that Yo mini-
mizes J on ~ = C[O, 1]. Moreover, it does so uniquely, because J(y) =
°
implies that p(x) = y2(X) = (see Lemma A.9 in the Appendix) so that
°
y = (!) = Yo'
In this section and in the next chapter, we will be concerned only with such
global minimum points; (the consideration of local minimum points is re-
served until Chapter 5, et seq.). We will also consider these minimal points in
problems involving isoperimetric or Lagrangian constraints.
First, let's make a restatement utilizing the linearity of!f!J.
(2.1) Lemma. Yo E ~ minimizes J on ~ [uniquely] iff

J(yo + v) - J(yo) ~ 0,
[with equality iff v = (!)].l
PROOF. For each y E~, set v = y - Yo, so that y = Yo + v, and y = Yo iff
v=@ D
Example 1. To minimize J(y) = S:y'(X)2 dx on
~ = {y E C 1 [a, b] : y(a) = 0, y(b) = 1}, (1)
we observe that obviously J(y) ~ 0, and by inspection, J(Yl) = if y~ = 0, °
but Yl = const. is not in ~.
However, if we reformulate the problem as suggested by the lemma, then
J(yo + v) - J(yo) = r
for Yo E ~, and Yo + v E 24 we should examine
[(y~(x) + V'(X))2 - y~(X)2] dx
= r V'(X)2 dx +2 r y~(x)v'(x) dx
~2 r y~(x)v'(x) dx. (Why?)
Now yo(a) = (Yo + v)(a) = yo(a) + v(a) so that v(a) = 0, and similarly, v(b) =
1 This is to be considered as two assertions; the first is made by deleting the bracketed expres-
sions throughout, while the second requires their presence.
r
0. By inspection, y~
y~(x)v'(x) dx = const. r
= const., makes
v'(x) dx = const. v(x) I: =

so that we have the inequality J(yo + v) - J(yo) ~ 0. Moreover, Yo(x) =
0, V such v,
(x - a)j(b - a) is in ~ and has y~(x) = const. Hence, by the lemma, Yo

minimizes J on !!}. It also does so uniquely, for equality demands that
r:V'(X)2 dx = 0, which by A.9 requires that V'(X)2 = 0, or that v(x) = const. =
v(a) = 0: i.e., v = (9.
Next, we make a simple observation that permits us to ignore inessential
constants.
(2.2) Proposition. Yo minimizes J on ~ [uniquely] iff for constants Co and

c "# 0, Yo minimizes c 2 J + Co on ~ [uniquely].
PROOF. If Y E ~, then
(c 2J + co)(Y) = c 2 J(y) + Co ~ c 2 J(yo) + Co = (c 2J + co)(Yo),
iff Yo minimizes J on ~ [with equality iff y = Yo]. D
Thus from Example 1, we may also say that Yo(x) = (x - a)j(b - a)
r r r
minimizes
J 1 (y) ~3 (Y'(X)2 + sin 3 x) dx = 3 y'(X)2 dx +3 sin 3 x dx
= 3J(y) + Co, on ~ of (1) uniquely.
Constraints
If we seek to minimize J on.@, when it is further restricted to a level set of one
or more similar functions G, then as was known to Lagrange and Euler, it
may suffice to minimize an augmented function without constraints.
(2.3) Proposition. If functions J and G1 , G2 , ..• , GN are defined on.@, and for
some constants AI' ... , AN, Yo minimizes j = J + Al G1 + A2 G2 + ... + ANGN
on ~ [uniquel~~' then Yo minimizes J on.~ [uniquely] when further restricted
to the set GyO = {y E ~: Gj(y) = Gj(yo),J = 1,2, ... , N}.
PROOF. For each y E~:
but when y E GyO ' then J(y) ~ J(yo), since the terms involving the Gj will have
the same values on each side of the inequality. Uniqueness is clearly preserved
if present. D
Remark. The hope is that the Aj can be found so that in addition Gj(yo) =
lj for prescribed values lj' j = 1, 2, ... , N. Reinforcement for this possibility
will be given in the discussion of the method of Lagrangian multipliers (§5.7).
Actually, Yo minimizes J automatically on a much larger set.
(2.4) Corollary. Yo of Proposition 2.3 minimizes J on ~ [uniquely] when

restricted to the set
GJo = {y E ~ : Apj(Y) ::; APj(Yo), j = 1,2, ... , N}.
PROOF. For Y E ~ the previous inequality gives us
N
J(y) - J(yo) ~ L
j=l
[APj(Yo) - Apj(Y)] ~ 0, if Y E GJo.
If J(y) = J(yo) under these conditions, then APj(Yo) = A.iGiy),j = 1,2, ... , N
(Why?), so that J(y) = J(yo). [With uniqueness, it follows that y = Yo·] D
These results illustrate an important principle: The solution to one minimi-

zation problem may also provide a solution for other problems.
The above abstract formulation is suitable for attacking problems
involving isoperimetric constraints such as the following:
Application: Rotating Fluid Column
A circular column of water of radius I is rotated about its vertical axis at

constant angular velocity, ro, inside a smooth-walled cylinder as shown in
Figure 2.1. Then the (upper) free surface assumes a shape which preserves the
volume of the fluid and minimizes the potential energy-given within an
l·
x
Figure 2.1
inessential additive constant Co by
J(y) = pn s: [gy2(X) - w 2x 2y(X)]X dx, (2)

on
!!) = {C[O, lJ : y(x) ~ O}.
Here, p is the mass density of the water, and 9 the standard gravitational
constant, while y(x) is the height of the liquid at a radial distance x from the
center. To obtain (2) we regard the column as if it were static and maintained
in the shape shown in Figure 2.1, by the interaction of the downward gravita-
tional force and the radially directed centrifugal force on each of its particles.
(We are invoking Bernoulli's principle for stable equilibrium which is dis-
cussed in §8.9.)
The volume of the (static) column is given by
G(y) = 2n s: xy(x) dx, (3)
and hence, according to Proposition 2.3, we should consider minimizing

J(y) = J(y) + AG(y), on !!). Then by Lemma 2.1, suppressing the subscript 0,
we should examine J(y + v) - J(y). Upon replacing A by pAI2 and sim-
plifying, we obtain the inequality
J(y + v) - J(y) = np s: {g[(y + vf(x) - y2(X)] + (A - W 2X2)V(X)}x dx
= np J: {gv 2(x) + [2gy(x) + (A - W 2X2)] v (x)} x dx
~ np J: [2gy(x) + (A - W 2X2)]V(X)x dx,
and this last integral vanishes V y + V E !!), when the bracketed term == 0; i.e.,
when
(4)
Moreover, equality is possible only when f~ v2(x)x dx = 0, which implies by

A.9, that v = (1), so that Yo minimizes J on !!) uniquely. It follows from
Proposition 2.3, that the free surface will be the paraboloid of revolution
defined by (4).
Now, by (3),
G(yo) = ~ (W 2X2 - A)X dx = nl 2 [w212 - ~J,
f'
9 Jo 9 4 2
and A may be selected to make G(yo) match the given fluid volume.
Observe that the minimizing shape will depend on W as we would expect
(and is of constant height when w = 0) but it is independent of the density p,
and thus would be the same for another (perfect) fluid. (The actual minimal
value of J( = J(yo», although calculable, is oflittle interest.)
The approach taken here extends also to problems involving Lagrangian

constraints on suitable function spaces. To exemplify this, we recall the nota-
tion from §1.5, that when f = f(x, y, z) E C([a, b] x 1R2) and y E C 1 [a, b],
then f[y(x)] ~ f(x, y(x), y'(x».
(2.5) Proposition. Suppose f = f(x, y, z) and 9 = g(x, y, z) are continuous on

[a, b] x 1R2 and there is a function A E C[a, b], for which Yo minimizes F(y) =
I~j[y(x)] dx on !!) s; C 1 [a, b] [uniquely] where j~ f + Ag. Then Yo mini-
mizes F(y) = I:f[y(x)] dx on !!) [uniquely] under the (inequality) constraint
A(X)g[y(x)] ~ A(X)g[yo(x)], Vx E [a, b]. (5)
r r
PROOF. If y E 2.&, then
F(y) = F(y) + A(X)g[y(x)] dx ~ F(yo) = F(yo) + A(X)g[yo(x)] dx
so that
F(y) - F(yo) ~ r A(X)(g[yo(x)] - g[y(x)]) dx
~ ° when (5) holds.

Moreover, if F(y) = F(yo) under conditions (5), then I:A(X)(g[yo(x)] -
= 0, and F(Y) = F(yo) [which with uniqueness is possible iff
9 [y(x)]) dx
y = Yo]. CI
Unfortunately, this result, although suggestive, cannot be implemented

readily, since it does not provide a method for determining a suitable A.
However, we may be able to use a known solution to another problem.
From the analysis used in the previous application, we may claim that
It
Yo (x) = x 2 - 1 minimizes F(y) = [y2(X) + (2 - 2X2)y(X)] X dx uniquely on
!!) = CEO, b] (Problem 2.12).
Setting A(X) = x in Proposition 2.5, it follows that Yo also minimizes
It
F(y) = 2xy(x) dx on !!) uniquely under the Lagrangian constraint:
g[y(x)] ~ y2(X) - 2x 2y(x) = g[yo(x)] = 1 - X4,
°
and since A(X) = x ~ on [0, b], also under the inequality constraint
g[y(x)] ~ 1 - X4.
In this section, we have shown that direct inequalities may be of use in
finding the minima for some functions. All of the functions analyzed here
were actually convex, and the convexity methods developed in the next
chapter will provide a more systematic means of obtaining such inequalities
for a large and useful class of functions. We shall also return to the basic
approach adopted here in the analysis of sufficient conditions for a minimum
in Chapter 9.
§2.4. The Gateaux Variations 45
§2.4. The Gateaux Variations

A decisive role in the optimization of a real valued function on a subset of [Rd
is played by its partial derivatives-or more generally by its directional
derivatives-if they exist (§0.5). When J is a real valued function on a subset
of a linear space all, then it is not evident how to define its partial derivatives
(unless all can be assigned a distinguished coordinate system). However, a
definition for its directional derivatives is furnished by a straightforward
generalization of that in [Rd:
(2.6) Definition. For y, v E all:

~J( ) def 1· J(y + ev) - J(y)
u y; v = 1m ,
8-+0 e
assuming that this limit exists, is called the Gdteaux variation of J at y in the
direction v.
Observe that both y and the direction v are fixed while the limit as e -+ 0
is taken in 2.6. The existence of the limit presupposes that:
(i) J(y) is defined; (6)
(ii) J(y + ev) is defined for all sufficiently small e; and then
[)J(y; v) = :e J(y + ev) 10=0' (7)
if this "ordinary" derivative with respect to the real variable e exists at e = O.

The Gateaux variation of J at y depends only on the local behavior of J
near y; however, this variation need not exist in any direction v =1= (9, or it may
exist in some directions and not in others.
As is to be expected from (7), the Gateaux variation is a linear operation
on the functions J; i.e., if at some Yo E all, functions J and J each have
Gateaux variations in the same direction v, then for constants c, CE [R,
[)(cJ + cJ)(yo; v) exists and equals c[)j(yo; v) + c[)J(yo; v). This is a direct
consequence of (7) and the linearity of the ordinary derivative. We also see
that M(y; (9) = 0, V Y at which J(y) is defined, and that when c E [R,
M(y; cv) = c[)J(y; v) provided that the variation on the right exists (Problem
2.4).
In particular, for c = -1:
[)J(y; -v) = -M(y; v) = [)( -J)(y; v),
whenever any of these variations exists.
Example 1. If J = fEel ([Rd) and Y, V E all = [Rd, then
[)f(Y; V) = lim f(Y + eV) - f(Y)

8-+0 e
is just the directional derivative of f when V is a unit vector (§0.5). Thus we

have that
!5f(Y; V) = Vf(Y)· V,
and this holds for all V E 11JJ. (Why?)
r
Example 2. IfI1JJ = C[a, b], then as in §2.2, Example 1,
J(y) = [sin 3 x + y2(X)] dx

is defined V y E 11JJ. Thus for fixed y, v E I1JJ and e =I- 0; y + ev E I1JJ (since I1JJ is a
r
linear space) so that
J(y + ev) = [sin 3 x + (y + ev)2(X)] dx is defined.
After successive cancellations we obtain the ratio
J(y + ev) -
------=-
J(y) 1 fb [( y+ev)2()x -y2()]
x dX
e e a
=2 r y(x)v(x)dx +e r v2(x) dx.
Each of the integrals in this last expression is a constant (Why?), and the limit
r
as e -+ 0 exists. Hence from Definition 2.6, we conclude that
!5J(y; v) = 2 y(x)v(x) dx, V y, V EI1JJ.
r
Alternatively, using (7), we could form:
J(y + ev) = [sin 3 x + (y + ev?(x)] dx
= r [sin 3 x + y2(X)] dt + 2e r y(x)v(x) dx + e2 r v2(x) dx,
r r
compute for fixed y, v, the derivative
:e J(y + ev) = 2 y(x)v(x) dx + 2e v2(x) dx,
and evaluate at e = 0 to get

b
!5J(y; v) = 2I y(x)v(x) dx (8)
as above.
In general, it is easier technically to use the second method for comput-

ing bJ in concrete cases simply because we are more familiar with the tech-
niques for differentiating standard real valued functions. However, the second
method as carried out here requires that (%a)J(y + av) exist for small a #- 0,
and'that it be continuous at a = O. By contrast, the first requires only the
existence of this derivative at a = O.
r
Example 3. When P E C[a, b], the function
J(y) = p(x)J1 + y'(X)2 dx

is defined V y E I1jj = C 1 [a, b]. (Exam pIe 2 of §2.2.)
r
Hence, using the second method, we form for fixed y, v E 11jj,
J(y + av) = p(x)J1 + (y + av)'(x)2 dx,

compute its derivative by differentiating under the integral sign (as in A.13)
to get
o 0 fb
o/(Y + av) = a oa [p(x)J1 + (y + W)'(X)2] dx
= fb p(x)(y + w)'(x)v'(x) dx
a J1 + (y + av)2(x)2 '
(whichis justified by the continuity ofthis last integrand on [a, b] x IR), and
evaluate at a = 0 to obtain
bJ(y; v) = fb p(x)y'(x)v'(x) dx. (9)

a J1 + y'(X)2
r r
Example 4. When fE C([a, b] x 1R2), the function
F(y) = f(x, y(x), y'(x)) dx = f[y(x)] dx,
is defined whenever y E I5J/ = C 1 [a, b]. (Example 4 of§2.2.) However to com-
r r
pute bF(y; v) by differentiating
F(y + av) = f(x, (y + w)(x), (y + av)'(x)) dx = f[(y + av)(x)] dx,

with respect to a under the integral sign, we obviously should require that the
function f(x, y, z) have partial derivatives /y, fz E C([a, b] x 1R2). Then from
the chain rule, for fixed x, y, z, v, w:
o
oaf(x, y + w, z + aw) = /y(x, y + w, z + aw)v + fz(x, y + w, z + aw)w.
With this preparation we have by A.l3 that

a
aa F(y + av) =
fb aaf[(Y
a
a
+ av)(x)] dx
= r:a f(x, y(x) + av(x), y'(x) + av'(x)) dx
= r (f,[(y + av)(x)]v(x) + fz[(Y + ev)(x)]v'(x)) dx,
(since the last integrand is continuous on [a, b] x IR), so that when a = 0, we
!5F(y; v) = r
see that F(y) = S:f(x, y(x), y'(x)) dx has the variation
[f,(x, y(x), y'(x))v(x) + fAx, y(x), y'(x))v'(x)] dx
= r(f,[y(x)]v(x) + fz[Y(x)] v' (x)) dx

We shall return to the analysis of this function in Chapters 3 and 6.
for y, v E C1[a, b]. (10)
Example 5. The function
J(y) = r sin y(x) dx + y2(b),
r
for which
J(y + av) = sin(y(x) + ev(x)) dx + (y + ev)2(b),
is defined on OJJ = C[a, b] and has at each y E OJJ and in each direction v E I1jj
the Gateaux variation
!5J(y; v) = r [cos y(x)]v(x) dx + 2y(b)v(b).
Example 6*. For OJJ = CEO, n], the function
J(y) = I" J1 - Y(X)2 dx
is not defined on OJJ, but it is defined on
!!} = {y ECEO, n]: iiyiiM ~ 1},

where
iiyiiM = max iy(x)i,
xe[O,"1
so that, for example, Yl (x) = sin x, x E [0, n] is a function in!!}, but Y2(X) = x
is not.
Moreover, for a given y E fJ4 only those directions v could be considered

for which y + ev E !!) for sufficiently small e: i.e., for which Iy(x) + ev(x) I ::s;; 1
for such e. The function y(x) == 1 is in fJ4 but its only possible direction is
v = 0; the function Yl(X) = sin x has (at most) as its possible directions, those
t
v for which v(nI2) = o. On the other hand, for y(x) = sin x and a given
v E CEO, n], then Iv(x)1 ::s;; IIvllM = M l , say, and so Iev(x) I ::s;; t when e ::s;; 1/2Ml ;
thus y + ev E!!) when e ::s;; 1/2Ml .
To proceed, we assume that for some y E!!), v E 11JJ, we know that y +
ev E !!), and consider for small e#-O the formally obtained derivative
a
ae J(y
r"
a Jo )1 - (y + ev)2(x) dx
+ ev) = ae
r"
a
= Jo ae)1-(y+ev)2(x)dx
=_ r" (y + ev)(x)v(x) dx,
Jo )1 - (y + ev)2(x)
which for e = 0 should give the value
bJ(y; v) = _ r" y(x)v(x) dx,

Jo )1 - y2(X)
provided that this process is valid. Now from A.13, it is valid if the integrand
j(x, e) = [y(x) + ev(x)]v(x)
)1 - (y + ev)2(x)
is continuous on [0, n] x [ -eo, eo] for some eo > 0 and this requires that
lIyllM < 1. However, if lIyllM < 1, then for each v E OJ/, lIy + evll M < 1 for all e
sufficiently small, by the argument above.
Thus we can at least say that when lIyllM < 1, bJ(y; v) exists V v E I1JJ and is
given above. At other y E!!) it is more difficult to consider the Gateaux
variation.
Example 7. The area function
A(Y) = f x(t)y'(t) dt
of§1.3 is defined V Y = (x, y) E I1JJ = (Cl[O, 1])2. Since for V = (u, v) E 11JJ,
A(Y + eV) = fal (x(t) + eu(t))(y'(t) + ev'(t)) dt,

it follows that
bA(Y; V) = fal [(x(t) + eu(t))v'(t) + (y'(t) + ev'(t))u(t)] dtl.=o
= fal [x(t)v'(t) + y'(t)u(t)] dt. (11)
Example 8. More generally, if f = f(x, Y, Z) C1([a, b] x then
r r
E [R2d),
F(Y) = f(x, Y(x), Y'(x)) dx = f[Y(x)] dx
is defined for all vector valued functions YE OJJ = (C1[a, b])d, (d = 2, 3, ... ).
Its Gateaux variation at V E OJJ is given by
!5F(Y; V) = r fy[Y(x)]· V(x) + fz[Y(x)]· V'(x)) dx, (12)
where fy[Y(x)] is the vector valued function with components
/YJY(x)] ~ fyj(x, Y(x), Y'(x)), j = 1,2, ... , d;

while fz[Y(x)] is that with components
fz.[Y(x)]
J
~ fz.(x,
J
Y(x), Y'(x)), j = 1,2, ... , d.
(See Problem 2.6.) This result together with the notation used to express it
should be compared with the one-dimensional case of Example 4.
Example 9. The surface area function of §2.2, Example 7,
S(u) = In J 1 + u; + u; dA
is defined VUE OJJ = C1(D), and
!5S(u; v) = f Jl
D
uxvx + uyV y dA
+ u; + u;
(13)
exists V u, v E iJjf, since the denominator can never vanish. (See Problem 2.7.)
(Problems 2.4-2.13)
PROBLEMS
2.1. Give an example of a nonconstant function in each of the following sets and
determine whether the set is a subspace of <W = C1 [a, b].
(a) C[a, b].
(b) f!) = {y E <W: y(a) = OJ.
(c) f!) = {y E <W: y'(a) = 0, y(b) = I}.
(d) C 2 [a, b].
(e) f!) = {y E <W: y(a) = y'(a)}.
bW.
(f) (C 1 [a,
(g) f!) = {y E <W: r:y(x) dx = OJ.
(h) f!) = {y E <W: y'(x) = y(x), X E (a, b)}.
Problems 51
2.2. Which of the following functions are defined on: (a) C1[a, b]? (b) C[a, b]?
L(y) = f b(sin 3 x)y(x) dx; J(y) =

fb y'(x)
dx;
• • J1 + y2(X)
F(y) = fb dx + yea); G(y) = fb log y(x) dx .
• J1 - Y'(X)2 •
2.3. For each of the following functions give a subset f!} of CfIJ (possibly CfIJ itself) on
which the function is defined and determine whether or not your f!} is a subspace
ofCflJ:
(a) F(Y) = S~IY'(x)1 dx, CfIJ = (C1[a, b]t
(b) G(y) = S~J1 + xy2(X) dx, CfIJ = C[a, bJ.
(c) H(y) = S~ log y'(x) dx, CfIJ = C 1 [a, bJ.
(d) J(u) = SDJU; - u; dA; CfIJ = C1(D), where D is a nice bounded domain of
[R2.
(e) K(y) = S~(l + Y"(X)2)y(X) dx, CfIJ = C 2[a, bJ.
2.4. Let J and J be real valued functions defined on a subset of a linear space CfIJ, and
suppose that for some y, v E CfIJ, (jJ(y; v) and bJ(y; v) exist. If c E [R, establish
existence and equality as required for the following assertions:
(a) b(cJ)(y; v) = (jJ(y; cv) = c(jJ(y; v).
(b) b(J + J)(y; v) = (jJ(y; v) + bJ(y; v).
Assuming the existence of the variations involved,
(c) is (jJ(cy; v) = c(jJ(y; v)?
(d) is (jJ(y + y; v) = (jJ(y; v) + (jJ(y; v)?
2.5. Let CfIJ = C1[a, b] and find (jJ(y; v) for y, v E CfIJ, when
(a) J(y) = y(a)3.
(b) J(y) = S~ [Y(X)3 + xy'(xf] dx.
(c) J(y) = S~J2 + x 2 - sin y'(x) dx.
(d) J(y) = S: [eXy(x) - 3Y'(X)4] dx + 2y'(af.
(e) J(y) = S: [X 2Y(X)2 + eY'(x)] dx.
(f) J(y) = sin y'(a) + cos y(b).
(g) J(y) = (S: [2yi(x) + x 2y(x)J dx)(S! [1 + Y'(X)J2 dx).
(h) J(y) = S: y(x) dx/S! [1 + Y'(X)2] dx.
2.6. In Example 8 of §2.4, verify equation (12) by formal differentiation under the
integral sign.
2.7. Let CfIJ = C 1 (15) where D is a bounded domain in the x-y plane with a nice
boundary.
(a) For J(u) = t SD(U~ + u;) dA, show that
(jJ(u; v) = Iv (uxvx + UyVy) dA, V u, v E CfIJ.
(b) For S(u) = SDJ1 + u~ + u; dA, verify that

bS(u; v) = r UxV x + UyV y dA, V u, V E CfIJ.
JD J1 + u~ + u;
2.8. Let CfIJ = (C 1[a, b])2, Y = (Y1, Y2), V = (v 1, v2), and find M(Y; V) for Y, V E CfIJ
when
(a) J(Y) = Y(a)' Y(b).
(b) J(y) = S: [Y1 (X)2 + y;(X)3] dx.
(c) J(Y) = S: [eYl(lt) - X2Y1 (x)y;(x)] dx.
(d) J(y) = S: [sin 2 Y1 (x) + XY2(X) + Y'l (x)Y2(4] dx.
2.9. Let CfIJ = C 2[a, b] and F(y) = S:f(x, y(x), y'(x), y"(x)) dx = S:f[y(x)] dx, where
f = f(x, y, z, r) in C 1([a, b] x 1R 3 ) is given. For v, y E CfIJ, prove that
(jF(y; v) = r (fy[y(x)]v(x) + f.[Y(x)]v'(x) + f,.[y(x)]v"(x)) dx.

2.10. Assume that (jJ(y; v) and (jJ(y; v) both exist for y, v E CfIJ.
(a) Verify the product formula
(j(JJ)(y; v) = M(y; v)J(y) + J(y)(jJ(y; v).
(b) Establish the quotient rule
(j (~) . v) = J(y)M(y; v) - J(y)(jJ(y; v)
- (y, - 2 '
J J(y)
provided that J(y) t= O.
(c) Supposing that h E C 1 (1R), show that
(j(h(J))(y; v) = h'(J(y))M(y; v).
2.11. If L is a linear function on CfIJ (as in Example 5 of §2.2) prove that
(jL(y; v) = L(v), V y, V E CfIJ.
2.12. (a) For p, fJ E CfIJ = C[O, b], with p > 0, find a function Yo which minimizes
F(y) = J: [p(X)y2(X) + fJ(x)y(x)] dx on f!) = CfIJ.
(b) Show that Yo is unique.

(c) What can you conclude if p < O?
2.13. Show that for the arc length function L(Y) = S5I Y'(t)1 dt of §1.1, formal differen-
tiation produces
oL(Y; V) = 1 1 Y'(t)
- - ) . V'(t) dt,
o IY'(t I
if Y E f!)* = {Y E CfIJ:, Y', t= O}. Is f!)* a subspace ofCflJ?
CHAPTER 3
Minimization of Convex Functions
By utilizing the Gateaux variations from the preceding chapter, it is straight-

forward to characterize convexity for a function J on a subset f!} of a linear
space !lJI, such that a convex function is automatically minimized by ayE f!}
at which its Gateaux variations vanish.! Moreover, in the presence of strict
convexity, there can be at most one such y. A large and useful class of
functions is shown to be convex. In particular, in §3.2, the role of [strongly]
convex integrands f in producing [strictly] convex integral functions F is
examined, and a supply of such f is made accessible through the techniques
and examples of §3.3. Moreover, the Gateaux variations of integral functions
will, in general, vanish at each solution y of an associated differential equa-
tion (of Euler-Lagrange).
The resulting theory extends to problems involving convex constraining
functions (§3.5), and it is used in several applications of interest including a
version of the brachistochrone (§3.4), and the hanging cable (or catenary)
problem of Euler. Additional applications will be found in the problem set
together with extensions of this theory to other types of integral functions.
In this chapter only those conditions sufficient for a minimum are consid-
ered, and it is shown that in the presence of strict convexity they can supply
a complete and satisfactory solution to the problems of interest. In particular,
we may be able to ignore the difficult question of a priori existence of a
minimum by simply exhibiting the (unique) function which minimizes. Actu-
ally, the direct approach developed here within the framework of convexity
extends in principle to other problems (§3.6).
1 The definitions of functional convexity employed in this book incorporate, for convenience,
some presupposed differentiability of the functions. For a convex set !lA, less restrictive formula-
tions are available, but they are more difficult to utilize. (See Problem 0.7, [Fl], and [E- T].)
53
54 3. Minimization of Convex Functions
§3.1. Convex Functions

When f E C 1 (1R3) then for Y = (x, y, z), V = (u, v, w) E 1R 3, we have
I5f(Y; V) = Vf(Y)· V; (as in §2.4, Example 1);
moreover, f is defined to be convex (§0.8) provided that V Y, V E 1R3:
f(Y + V) - f(Y) ~ Vf(Y)· V = I5f(Y; V), (1)
and strictly convex when equality holds at Y iff V = (9 (§0.9). We also observe
that minimization of a convex function f may be particularly easy to estab-
lish, in that a point Yat which Vf(Y) = (9 clearly minimizes f. (1) suggests the
following:
(3.1) Definition. A real valued function J defined on a set!!} in a linear space

OY is said to be [strictly] convex on !!} provided that when y and y + v E!!}
then l5J(y; v) is defined and J( y + v) - J(y) ~ I5J(y; v) [with equality iff
v=(9P
!!} itself may be nonconvex. (See 3.15)

Although "most" functions are not convex, a surprisingly large number of
those of interest to us are convex-even strictly convex-as the applications
will show. The following observation will prove valuable:
(3.2) Proposition. If J and J are convex functions on a set!!} then for each
c E IR, c 2 J and J + J are also convex. Moreover, the latter functions will be
strictly convex with J (for c ¥- 0).
PROOF.
(c 2 J + J)(y + v) - + J)(y) = c 2 (J(y + v) - J(y)) + (J(y + v) - J(y))

(c 2 J
~ c2 15J(y; v) + I5J(y; v) = l5(c 2 J + J)(y; v), if y, y + V E!!} (by 3.1).
This establishes the convexity of J + J (when c2 = 1) and of c2 J (when
J = (9). Moreover, when J is strictly convex and c ¥- 0, then there must be
strict inequality except for the trivial case of v = (9. D
(3.3) Proposition. If J is [strictly] convex on !!} then each Yo E !!} for which
l5J(yo; v) = 0, V Yo + V E !!} minimizes J on !!} [uniquely].
PROOF. If Y E !!}, then with v = y - Yo
J(y) - J(yo) = J(yo + v) - J(yo)
~ l5J(yo; v) = 0, (by 3.1 and hypotheses)
[with equality iff v = (9].
1 Recall the footnote on page 40.

§3.1. Convex Functions 55
Hence J(y) ~ J(yo) [with equality iff y = Yo] and this is the desired
conclusion. D
Example 1.
J(y) = r (sin 3 x + y2(X)) ix
r
for which
bJ(y; v) =2 y(x)v(x) dx, v y, V E!ljj = C[a, b] (§2.4, Example 2),
is strictly convex on !ljj since
J(y + v) - J(y) = r ((y + V)2(X) - y2(X)) dx
= 2 r y(x)v(x) dx + r v2(x) dx
~2 r y(x)v(x) dx
with equality iff f~ v2(x) dx = 0 which by A.9 is possible for the continuous
= bJ(y; v),
function v2 iff v2(x) == 0; i.e., v = (9. Thus, by Proposition 3.3, each y E !ljj
which makes
M(y; v) = 2 Lb y(x)v(x) dx = 0,
V V E IlJI, minimizes J on !ljj uniquely. Clearly, y = (9 accomplishes this and
hence it is the unique minimizing function.
On the other hand, to minimize J on
~ = {y E C[a, b]: y(a) = a1 , y(b) = bd,
we would again try to have M(y; v) = 0, but now only for those y, y + V E ~,
i.e., only for those v E ~o where
~o = {v E C[a, b]: v(a) = v(b) = O}. (Why?)
Again y = (9 would make M(y; v) = 0 but now it is not in ~ (unless al =
b1 = 0).
Example 2.
F(y) = r y'(X)2 dx
for which
bF(y; v) =2 r y'(x)v'(x) dx, V y, V E!ljj = C 1 [a, b],
is also convex on !ljj (Why?), but now the equality F(y + v) - F(y) = bF(y; v)
is possible iff S~ V'(X)2 dx = 0 and this occurs whenever v(x) = const. Thus F
is not strictly convex on O!J. However, F is strictly convex on
~ = {y E C I [a, b]: y(a) = aI' y(b) = bd,
since now y, y + V E ~ = V E ~o where

~o = {v E C I [a, b]: v(a) = v(b) = O},
and hence equality for v in ~o is possible iff v(x) = const. = v(a) = O. Thus
each y E ~ which makes c5F(y; v) = 2 S:y'(x)v'(x) dx = 0, V V E ~o minimizes
r r
F on ~ uniquely. By inspection, y' = const. will accomplish this since
(const.) v'(x) dx = (const.) v'(x) dx
= const. [v(b) - v(a)] = 0, VVE ~o.
The linear function Yo(x) = [(bi - ad/(b - a)] (x - a) + al E ~ and has

y~ = canst. Hence it minimizes F on ~ uniquely.
Example 3. A linear function L on O!J is convex, since by Problem 2.11,

L(y + v) - L(y) = L(v) = c5L(y; v), V y, V EilJI,
but it is not strictly convex. (Why?)
Example 4. The evaluation function from Example 5 of §2.4, J(y) = y2(b),

which for y, v E O!J = C[a, b] has the variation c5J(y; v) = 2y(b)v(b), is convex
since
J(y + v) - J(y) = (y + v)2(b) - y2(b) = 2y(b)v(b) + v2(b)
~ c5J(y; v);
but not strictly convex since equality occurs when v(b) = O. Clearly J(y) ~ 0,
but J assumes its minimum value of 0 whenever y(b) = O.
§3.2. Convex Integral Functions

If f = f(x, y, z) and its partial derivatives 1;" fz are defined and continuous on
r r
[a, b] x [R2, then as in Example 4 of§2.4, we know that the integral function
F(y) = f(x, y(x), y'(x)) dx = f[y(x)] dx,
r
has V y, v E C I [a, b], the variation
c5F(y; v) = (1;,[y(x)]v(x) + fz[Y(x)]v'(x)) dx, (2)

§3.2. Convex Integral Functions 57
where from §1.5 we recall the generic abbreviation

f[y(x)] ~ f(x, y(x), y'(x)), (3)
so that
fy[Y(x)] = fy(x, y(x), y'(x)) and fz[y(x)] = fz(x, y(x), y'(x)).
Hence, convexity of F requires that V y, y + V E C 1 [a, b]
F(y + v) - F(y) ~ c5F(y; v),
r r
or that
(f[y(x) + v(x)] - f[y(x)]) dx ~ (fly(x)]v(x) + fz[Y(x)]v'(x)) dx.
Now this would follow from the corresponding pointwise inequality

between the integrands in the last expression; i.e., if for each x E (a, b):
f[y(x) + v(x)] - f[y(x)] ~ f,,[y(x)]v(x) + fz[Y(x)]v'(x), (4)
or from (3), if
f(x, y + v, z + w) - f(x, y, z) ~ fy(x, y, z)v + fz(x, y, z)w,
V (x, y, z), (x, y + v, z + w) E (a, b) x 1R2, (5)
where we have incorporated the abbreviations:
y = y(x), v = v(x) and z = y'(x), w = v'(x).
Inequality (5) simply states that f = f(x, y, z) is convex when x is held
fixed. [See (6) of §0.8.] This restricted or partial convexity essential to our
development is expressed and extended in the following which uses for illus-
tration a function defined on a subset of 1R3:
(3.4) Definition. f~, y, z) is said to be [strongly] convex on S ~ 1R3 if f =

f(x, y, z) and its partial derivatives fy an:d fz are defined and continuous on
this set and there they satisfy the inequality:
f(x, y + v, z + w) - f(x, y, z) ~ fy(x, y, z)v + fz(x, y, z)w,
V (x, y, z) and (x, y + v, z + w) E S, (6)
[with equality at (x, y, z) only if v = 0.01" w = 0].
Observe that the underlined variable(s) (if any) are held fixed in the inequal-
ity while partial derivatives of f are required only for the remaining variables
(y and z). Clearly, if f itself is convex on 1R3 as in §0.8 then f~, y, z) will be
convex as above. Moreover, if as in §0.9, f is strictly convex, then f~, y, z)
will be strongly convex. However, in general, strong convexity is weaker than
strict convexity. (Why?) Also, f(y, z) is [strongly] convex on D ~ 1R2 precisely
when R!, y, z) = f(y, z) is [strongly] convex on [a, b] x D.
For instance, we will see in the next section that f(y, z) = Z2 + 4y is

strongly convex on ~2 (even though it is not strictly convex). Therefore
j(~, y, z) = Z2 + 4y is strongly convex on [a, b] x ~2 for any interval [a, b].
The significance of strong convexity is seen in the following:
(3.5) Theorem. Let D be a domain in ~2 and for given aI' bl , set

~ = {y E CI[a, b]: y(a) = aI' y(b) = bl ; (y(x), y'(x)) ED}.
r
If f~, y, z) is [strongly] convex on [a, b] x D, then
F(y) = f(x, y(x), y'(x)) dx
is [strictly] convex on ~. Hence each y E ~for which

d
dxfz[Y(x)] = fy[Y(x)]
on (a, b), minimizes F on ~ [uniquely].

PROOF. When y, y + V E ~, then inequality (6) shows that at each x E (a, b),
f[y(x) + v(x)] - f[y(x)] ~ fy [y(x)] v(x) + fz [y(x)] v'(x), (7)
[with equality only if v(x) = 0 or v'(x) = 0 so that v(x)v'(x) = 0]. Integrating
r r
(7) gives
(f[y(x) + v(x)] - f[y(x)]) dx ~ (fy[y(x)]v(x) + fz [y(x)] v'(x)) dx,

or with (2) and (3):
F(y + v) - F(y) ~ bF(y; v),
so that F is convex.
Moreover, in the presence of (7), equality between the integrals
representating these last functions is possible only when equality holds
everywhere in (7) (A.lO). [But if f~, y, z) is strongly convex, this in turn is
possible only if the product v(x)v'(x) = t(v 2(x))' == 0; then v2(x) = const. =
v2(a) = 0 when both y and y + v E ~. Thus v = (!) so that F is strictly convex
on~.]
Finally from (2), each y for which
d
dxfz[Y(x)] = fy[y(x)] on (a, b) (8)
makes
bF(y; v) = r:x (fz[Y(x)] v(x)) dx = fz[Y(X)]V(x{
= 0 when y, y + V E~. (Why?) (8')

Thus by Proposition 3.3, y minimizes F on ~ [uniquely]. o
§3.2. Convex Integral Functions 59
Neither of the convexity implications stated in this theorem is reversible.

(See Problem 3.16.)
Example 1. To minimize
F(y) = Il (y'2(X) + 4y(x)) dx

on
f!) = {y E C1(0, 1]: y(O) = 0, y(1) = 1}
we recall that f~, y, z) = Z2 + 4y is strongly convex on [0, 1J x [R2. Hence
according to Theorem 3.5, F is minimized uniquely on f!) by a solution Yo of
the equation
(0 < x < 1),
which for this f is just

d
d)2y'(x)] = 4 or y" = 2.
Upon integrating twice we obtain the general solution

= x 2 + CX + Co
y(x)
for constants c, Co to be found if possible so that y E f!). We require y(O) =
Co = 0 and (then) y(1) = 1 + c = 1, or c = O. Consequently Yo(x) = x 2 mini-
mizes F on f!) and it is the only function which does so!
(3.6) Remarks. The differential equation (8) whose solutions in f!) minimize
our convex F is known as the Euler-Lagrange equation. It is a fundamental
tool of the variational calculus and we will examine it thoroughly in Chapter
6. For the present, note that if f E C 2 ([a, bJ x D), then we may use the chain
rule (formally) on the left side of (8) and seek a minimizing y E f!) n C2 [a, bJ
which satisfies the second-order differential equation
fzx[Y(x)] + fzly(x)]y'(x) + fzz[y(x)Jy"(x) = f,,[y(x)J, (9)
(with the obvious abbreviations). Although there are standard existence theo-
rems which provide conditions for a solution y to (9) in a neighborhood of
x = a which satisfies y(a) = a 1 , these theorems do not gurantee that such
solutions can be extended to [a, bJ, or that when extendable they cari meet
the second end point condition y(b) = b1 • (In Problem 3.20, we have an
example for which even the simpler equation fz(x, y'(x)) = const. cannot be
satisfied in f!).) Thus we do not have a proof for the existence of a function
which minimizes F on f!), and indeed as we shall see, such functions need not
exist. Our condition (8) is at best sufficient, and we must consider each
application independently.
There is some simplification when y is not present explicitly, Le., when

f = f(x, z) alone (or f"= f(z)). Then fy == 0, and for an interval I, the appro-
priate requirement for the [strong] convexity of f~, z) on [a, b] x I is that

for each x E [a, b]:
f(x, z + w) - f(x, z) ~ fAx, z)w, v z, z + wEI;
[with equality at z iff w = 0]. (10)
If, in addition, fx == 0, then f = f(z) alone which is [strictly] convex on I
precisely when j~, z) = f(z) is [strongly] convex on [a, b] x I. (Why?)
With these reductions, the next results should be apparent.
(3.7) Theorem. Let I be an interval and set

q) = {y E CI[a, b]: y(a) = aI' y(b) = bl ; y'(x) E I}.
Then, if f~, z) is [strongly] convex on [a, b] x I, each y E q) which
makes fz(x, y'(x)) = const. on (a, b) minimizes F(y) = f~f(x, y'(x)) dx on q)
[uniquely].
PROOF. This follows immediately from (10) upon setting /y == 0 in the state-
ment and proof of the previous theorem. 0
(3.8) Corollary. Iff = f(z) is [strictly] convex on I and

m = [(bi - al)/(b - a)] E I,
then Yo(x) == m(x - a) + a l minimizes F(y) = f~f(Y'(x)) dx on q) [uniquely].
PROOF. If y~(x) = m E I, then Yo E q) (Why?) and fz(Y~(x)) = fAm) is constant
on (a, b). Hence Theorem 3.7 is applicable. 0
There are similar simplifications when f = f(x, y), but the associated
integrands occur less frequently in application. (See Problem 3.18.)
Free End-Point Problems

When we examine the proof of Theorem 3.5 we see that the end-point
specification was used only to conclude that the constant v2 (x) = 0 and that-
fz[y(x)]v(x)l~ = O. Hence these end-point conditions on y may be relaxed, if
suitable compensation is made in fz[Y(x)],
(3.9) Proposition. Let D be a domain in 1R2 and suppose that f~, y, z) is

[strongly] convex on [a, b] x D. Then each solution Yo E q) = {y E C I [a, b]:
(y(x), y'(x)) E D} of the differential equatiion (d/dx)fz[Y(x)] = fY[y(x)]
minimizes
F(y) = r f[y(x)] dx:
§3.3. [Strongly] Convex Functions 61
(i) on ~b = {y E~: y{a) = yo{a)}, if fz[Yo{b)] = 0 [uniquely];

(ii) 'on ~, if fz[Yo{a)] = fz[Yo{b)] = 0, [uniquely within an additive
constant].l
(As we shall see in §6.4, these "natural" boundary conditions on fz[Yo{x)]

are also necessary for the minimization.)
PROOF. Only the last assertion in (ii) requires further comment. If f~, y, z) is
strongly convex on [a, b] x D, and Yo E ~ is a solution of the given differen-
tial equation, then when Yo + v E ~ we have from (8') that
F{yo + v) - F{yo) ~ t5F{yo; v) = fz[Yo{x)]v{x)l~ = 0,
with equality only if v2{x) = const. (=> 2v 2{x)v'{x) = V{X)(V2{X))' == 0) so that
v'{x) == 0 and v{x) = const. on [a, b]. Thus
F{yo + v) = F{yo) => Yo +v= Yo + const. o
Extension of the results of this section to convex functions defined by
improper Riemann integrals is treated in Problem 3.21 *.
Example 2. Let's return to the problem in Example 1, where we found that

Yo{x) = x 2 minimizes. It happens that fz[Yo{x)] = 2Yb{x) = 4x vanishes at
x = o and so from Proposition 3.9 we see that Yo also minimizes F uniquely
on the larger set
~l = {y E C l [0, 1]: y(1) = 1}.
Similarly, we can show that Yl (x) = x 2 - 2x minimizes F on ~l = {y E

C l [0, 1]: y{O) = O} uniquely. However, our method cannot produce a
function of the form Y2{X) = x 2 + CX + Co that minimizes F on the still larger
set ~2 = C l [0, 1], since in this case, fz[Y2{x)] = 2y~{x) = 2{2x + c) cannot
be zero at both end-points. In view of Remarks (3.6) we should not be
surprised that we did not get everything we might wish. We were lucky to get
what we did so easily.
§3.3. [Strongly] Convex Functions

In order to ilPply the results of the previous section, we require a supply of
functions which are [strongly] convex. In this section techniques for rec-
ognizing such convexity will be developed.
We begin with the simpler case f = f{x, z), where as we have seen, the
defining inequality for [strong] convexity of f~, z) on [a, b] x I is (10). Now
(1O) is in turn guaranteed by a simple condition on fzz which should recall the
1 The choice of constants may be limited. See Problem 3.9.

criterion from elementary calculus for the convexity of a function defined in

an interval 1.
(3.10) Proposition. If f = f(x, z) and fzz are continuous on [a, b] x I and for
each x E [a, b], fzAx, z) > 0 (except possibly at a finite set of z values) then
f(~, z) is strongly convex on [a, b] x I.
PROOF. For fixed x E [a, b], let g(z) = f(x, z) so that g"(z) = fzz(x, z) > 0 on I
(with a possible finite set of exceptiional values). Then integrating by parts
gives for distinct z, CE I:
g(O - g(z) = f g'(t) dt = (C - z)g'(z) + f (C - t)g"(t) dt > (C - z)g'(z),
since the last integral is strictly positive by the hypothesis and A.9, indepen-
dently of whether z < Cor C< z. (Why?)
Thus with w = C- z, recalling the definition of g, we conclude that
f(x, z + w) - f(x, z) > fAx, z)w, when w =F 0, and this establishes the strong
convexity of f~, z). 0
Remark. If at some x E [a, b], fzz(x, z) == 0 on [Zl, Z2] s; I then fAx, z)
increases with z, but not strictly, so that f~, z) is only convex on [a, b] x I.
Example 1. f~, z) = sin 3 ~ + Z2 is strongly convex on ~ x ~ since
fzAx, z) = 2 > o.
Example 2. f~, z) = e'!(sin 3 ~ + Z2) is also strongly convex on ~ x ~ since
fzz(x, z) = 2e X > O.
(In fact the product of a [strongly] convex function by a positive continuous
function p = p(x) is again [strongly] convex. See Problem 3.3.)
Example 3. For r =F 0, f~, z) = Jr 2 + Z2 is strongly convex on ~ x ~, since

V ZE~:
z
fAx, z) = J
r2 + Z2 ' so that
1 Z2 r2
fzz(x, z) = Jr 2 + Z2 (r2 + r2)3/2 (r2 + Z2)3/2 > O.
Example 4. If 0 < P E C[a, b], and r =F 0, f~, z) = p~)Jr2 + Z2 is also

strongly convex on [a, b] x ~. (See Example 2 above.)
Example 5. f(z) = -J1=? is strongly convex on 1= (-1,1), since

§3.3. [Strongly] Convex Functions 63
Example 6. f0s., z) = !3 + e=z is (only) convex on ~ x ~ since fzz == O.

Indeed:
f(x, z + w) - f(x, z) = eXw = fAx, z)w,
and equality holds V w E ~.
Example 7. f(!, z) = !2 + (sin !)Z2 with fzAx, z) = 2(sin x), becomes convex
only when sin x ;?: 0; e.g., on [0, n] x ~; and is strongly convex only when
sin x> 0, e.g., on (0, n) x ~.
Example 8. Finally, f0s., z) = !2 - Z2 is never convex, since for w =f. 0,

f(x, z + w) -- f(x, z) - fz(x, z)w = -(z + W)2 + Z2 + 2zw
= _w 2 < O.
(However, - f0s., z) = - x 2 + Z2 is again strongly convex on ~ x ~.)
Of course, the conclusions just obtained are unchanged if z is replaced by

y in each occurrence.
When all variables are present in the function f(x, y, z), there are no
simplifications such as those just considered. (There is again a second deriva-
tive condition which guarantees [strict] convexity, but it is awkward to
apply. See Problem 3.5.)
The following general observations (whose proofs are left to Problems 3.2
and 3.3) will be of value:
(3.11) Fact 1. The sum of a [strongly] convex function and one (or more)
convex functions is again [strongly] convex.
Fact 2. The product of a [strongly] convex function f0s., y, z) by a
continuous function [p(x) > 0] p(x) ;?: 0 is again [strongly] convex on
the same set.
Fact 3. f(!, y, z) = a(!) + {3(!)y + y(!)z is (only) convex for any con-
tinuous functions a, {3, y.
Fact 4. Each [strongly] convex function f(!, z) (or f(!, y)) is also
[stron&ly] convex when considered as a function j(!, y, z) on an appro-
priate set.
Example 9. f(!, y, z) = -2(sin !)y + Z2 is the sum of the strongly convex

function Z2 (Fact 4) with the convex function -2(sin !)y (Fact 3), and hence
it is strongly convex on ~ x ~2 (Fact 1).
Similarly, g0s., y, z) = - 2(sin !)y + Z2 + !2 j1+?, is the sum of the
strongly convex function f(!, y, z) with the convex function !2 j1+? (Fact
2) and so it too is strongly convex on ~ x ~2.
Example 10. f(y, z) = J1 + y2 + Z2 with

y z
h(y, z) = J1 2 2 and fz(Y, z) = J1 2 2
+y +z +y +z
is more difficult to examine. For its convexity, we require that
or upon introducing the three-dimensional vectors
A = (1, y, z) and B = (1, Y + v, z + w),

that
I I > I I (yv + zw) = IAI2 + (yv + zw)
B - A + IAI IAI
1 + y(y + v) + z(z + w) A .B
IAI -j}[!'
where the dot denotes the scalar product of the vectors.
Since A· B = IAIIBI cos(A, B) ~ IAIIBI with equality iff A and Bare

codirected, it is seen that (11) does hold with equality iff(l, y, z) and (1, y + v,
z + w) are codirected, i.e., iff v = w = 0. Thus, f(y, z) = J1 + y2 + Z2 is
strongly convex on 1R2; in fact, it is strictly convex on 1R2.
Example 1l.1f0 < P E C[a, b], then
f~, y, z) = p~)J1 + y2 + Z2
is strongly convex on [a, b] x 1R2 (Fact 2 and Example 10.)
Example 12. When b #- 0, then f(y, z) = J y2 + b2Z2 has derivatives

y b2 z
h(y, z) = JY 2
+
b2
z
2 and fAy, z) = JY 2
+
b2
z
2'
which are discontinuous at the origin. However, on the restricted set 1R2 ,...,
{(O, O)} this function is again convex but not strongly convex. (See Problem
3.24.)
Example 13. When IX #- 0, f(y, z) = (z + IXy)2 is only convex in 1R2, since (5)
holds, but with equality when (w + IXV)2 = 0. (See Problem 3.16b.)
(Problems 3.1-3.19)
§3.4. Applications 65
§3.4. Applications
In this section we show that convexity is present in problems from several
diverse fields-at least after suitable formulation-and use previous results
to characterize their solutions. Applications, presented in order of increasing
difficulty-and/or sophistication, are given which characterize geodesics on
a cylinder, a version of the brachistochrone, Newton's profile of minimum
drag, an optimal plan of production, and a form of the minimal surface.
Other applications in which convexity can be used with profit will be found
in Problems 3.20 et seq.
(a) Geodesics on a Cylinder
To find the geodesics on the surface of a right circular cylinder of radius 1

unit, we employ, naturally enough, the cylindrical coordinates (lJ, z) shown in
Figure 3.1 to denote a typical point. It is obvious that the geodesic joining
points PI = (lJ I , Zl), P2 = (lJ I , Z2) is simply the vertical segment connecting
them. Thus it remains to consider the case where P2 = (lJ2 , Z2) with lJ2 "# lJI ;
a little thought shows that by relabeling if necessary, we can suppose that
o < lJ2 - lJI ::;; n, and consider those curves which admit representation as
the graph of a function Z E !!} = {z E C 1 [lJI , lJ2 ]: z(lJ) = Zj,j = 1, 2}.
The spatial coordinates of such a curve are
(x(lJ), y(lJ), z(lJ)) = (cos lJ, sin lJ, z(lJ)),
Figure 3.1
so that when Z E :!), the resulting curve has the length
With an obvious change in variables, this integrand corresponds to the

function of §3.3, Example 3, which is strongly convex. Thus by Corollary 3.8,
we conclude that
(3.12) Among curves which admit representation as the graph of a function

Z E :!), the minimum length is given uniquely for that represented by the function
which describes the circular helix joining the points.
(If the cylinder were "unrolled," this would correspond to the straight line
joining the points.) Plants take helical paths when climbing around cylindri-
cal supporting stakes toward the sun [Li].
(b) A Brachistochrone
For our next application, we return to the brachistochrone of §1.2(a). As
formulated there, the function T(y) is not of the form covered by Theorem
3.5. (Why not?) However, if we interchange the roles of x and y and consider
those curves which admit representation as the graph of a function y E :!) =
{y E Cl[O, Xl]: y(O) = 0, y(X l ) = yd (with Xl and Yl both positive) as in
Figure 3.2, then in the new coordinates, the same analysis as before gives for
each such curve the transit time
IX, 1 + y'(X)2
T(Y)=Jo 2 dx,
gx
which has the strongly convex integrand function of §3.3, Example 4, with
Figure 3.2
r = 1 and p(x) = (2gxt l /2 on (0, Xl]. Now p(x) is positive and integrable on
[0, Xl] and although it is not continuous (at 0), Theorem 3.7 remains valid.
(See Problem 3.21.)
Thus we know that among such curves, the minimum transit time would
be given uniquely by each y E ~ which makes
JxxJ1y'(x)+ y'(xf == -1c for some constant c.
Squaring both sides gives the equation
y'(X)2 X , 2 c2
1 +yx '( f = 2
c ' so 1 + Y (x) = c-x
2 2
and
y'(x) = Jx J1 + y'(X)2 =
C
J c2
x
- X
~ 0. (12)
Thus y'(0) = 0.
If we introduce the new independent variable fJ through the relation
°
x(fJ) = (c 2/2)(1 - cos fJ) = c 2 sin2(fJ/2), then fJ = when x = 0, and for fJ < n,
fJ increases with x. Also, c 2 - x(fJ) = (c 2/2)(1 + cos fJ). By the chain rule
dy
dfJ = y'(x)x'(fJ) = y'(x) (c2 sin fJ
2
)
,
and from equation (12) we get
1 - cos fJ . c2
-----:-fJ sm fJ = -(1 - cos fJ).
1 + cos 2
°
Hence y(fJ) = (c 2/2)(fJ - sin fJ) + c l , and the requirement y(O) = shows that
C l = 0.
Upon replacing the unspecified constant c by fic, we see that the mini-
mum transit time would be given parametrically by a curve of the form
X(fJ) = c2(1 - cos fJ),
{ (13)
y(fJ) = c2(fJ - sin fJ),
provided that c2 and fJ l can be found to make x(fJl ) = Xl' y(fJd = Yl' The
curve described by these equations is the cycloid with cusp at (0, 0) which
would be traced by a point on the circumference of a disk of radius c 2 as it
rolls along the y axis from "below" as shown in Figure 3.3.
For fJ > 0, the ratio y(fJ)/x(fJ) = (fJ - sin fJ)/(l - cos fJ) has the limiting
value +00 as fJ i 2n, and by L'Hopital's rule it has the limiting value of as
fJ ~ 0. Its derivative is
°
(1 - cos fJ)2 - sin fJ(fJ - sin fJ) 2(1 - cos fJ) - fJ sin fJ
(1 - cos fJ)2 (1 - cos fJ)2
/ y
(
/ "\
\
\ )
\
"'" x
Figure 3.3
which utilizing the half-angle formulae may be rewritten as
cos f) /2 ( f) f)) (f) "# n),

sin 3 f)/2 tan 2 - 2
° °
and thus is positive for < f) < 2n. (Why?) y(f))/x(f)) is positive, increases
°
strictly from to +00 as f) increases from to 2n, and hence from continuity
(through the intermediate value theorem of §A.l), assumes each positive
value precisely once. In particular, there is a unique f)l E (0, 2n) for which
y(f)d/X(f)l) = ydx l , and for this f)l' choosing c 2 = xd(l - cos f)l) will guar-
antee the existence of a (unique) cycloid joining (0, 0) to (x I' Y I)'
Unfortunately, as Figure 3.3 shows, the associated curve can be repre-
sented in the form y = y(x) only when f)l ~ n, i.e., when ydXI ~ n/2. More-
over, the associated function y E ClEO, Xl] only when ydXI < n/2, since the
tangent line to the cycloid must be horizontal at the lowest point on the arch.
Nevertheless, we do have a nontrivial result:
(3.13) When ydXI < n/2, among all curves representable as the graph of a
function Y E C l [0, Xl] which join (0,0) to (Xl' Yl), the cycloid provides
uniquely the least time of descent.
Thus we confirm Galileo's belief that the brachistochrone is not the

straight line and support the classical assertion by Newton and the Bernoullis
that it must always be a cycloid.
It is not too difficult to extend our analysis to the case ydXI = n/2
(see Problem 3.22*), and it may seem physically implausible to consider
curves which fall below their final point or those which have horizontal
sections (i.e., those which cannot be expressed in the form y = y(x)) as
candidates for the brachistochrone. However, it is true that the brachisto-
chrone is always the cycloid, but a proof for the general case will be deferred
until §8.8.
(c) A Profile of Minimum Drag

One of the first problems to be attacked by a variational approach was that
propounded by Newton in his Principia (1686) of finding the profile of [the
shoulder of] a projectile of revolution which would offer minimum resistance
(or drag) when moved in the direction of its axis at a constant speed Wo in
water. (We can suppose that Wo = 1V
We adopt the coordinates and geometry shown in Figure 3.4, and postulate
with Newton that the resisting pressure at a surface point on the shoulder is
proportional to the square of the normal component of its velocity. A shoul-
der is obtained by rotating a meridional curve oflength I defined by y = y(x)
about the y-axis as shown. At a point on the shoulder, let 1/1 denote the angle
between the normal to this curve and the positive y-axis. Then we wish to
minimize
I cos 2 I/I(o)2nx(0') cos 1/1(0') dO'.
Since cos 1/1(0') = x'(O') while 1 + y'(X)2 = sec 2 1/1(0'), we evidently wish to
minimize
on
F(y) = r
x(1 + y'(X)2t 1 dx,
fl) = {y E C 1 [a, 1]: y(a) = h, y(1) = 0, y(x) ~ O},

where we suppose that the positive constants a < 1 and h are given; (a = 0 is
excluded for reasons which will emerge). Now, if
x -2zx
f(x z ) = - - then fAx, z) = (1 + Z2)2'
, 1 + Z2'
y
1Motion
,,-
/
/
I
I
I
x
Figure 3.4
1 See the article "On Newton's Problem of Minimal Resistance," by G. Butazzo and B. Kawohl,
in The Mathematical Intelligencer, pp. 7-12, Vol. 15, No.4, 1993. Springer-Verlag, New York.
and for x> 0:

2x(3z 2 - 1) 1
fzz(x, z) = (1 + Z2)3 > 0, when Izl > .j3.
On physical grounds we expect y' ~ (Why?), l and, by Proposition 3.10,°
f~, z) is strongly convex on [a, 1J x (-00, -1/.j3]. Hence from Theorem
3.7, we know that if
Yo E;!J' = {y E;!J: y'(x) ~ -1/.j3, x E [a, 1J}
makes
, -2xy'(x) 2
fAx, y (x)) = (1 + y'(X)2)2 = const. = c' (14)
say, for a positive constant c, then it minimizes F on ;!J' uniquely. Upon

squaring and rewriting, we find that for each x E [a, 1J, u ~ 1 + y,2 should be
a solution to the quartic equation u4 = C2X2(U - 1) with u ;;;:: 1 + (1/.j3)2 =
4/3. From Figure 3.5, we see graphically that this equation has a unique root
u = u(s) ;;;:: 4/3 when s = ex ;;;:: 4 2 /3 312 = so, say, and that u(s) increases to infin-
ity with s. (u(s) can be determined explicitly by the methods of Cardano-
Ferrari, but the result is not simple. 2 Newton's own parametric approach is
\
\
\
\
Figure 3.5
1 See Problem 3.41.

2 When S ~ So = 4 2 /3 3/2 then
jA+JB-A
u(s) = J2 ' where B= S2/.j2A,
and A = (n + r)1/3 + (n - r)1/3, with
taken up in Problem 3.40.) Then, from (14), we get
y'(x) = -u 2 (cx)jcx, and (since y(1) = 0),
y(x) = I x
1 u2 (ce)
-~de
ce
=-
1
c
IC -u d
cx
(s)
2
s
s (cx ~ so), (14')
provided that c ~ soja can be found to make y(a) = h. Now u2 (s)js =

Ju(s) - 1 increases with s, and for fixed x, the first integral in (14') increases
continuously to infinity as c ? +00. Since our restricted minimum-drag
problem has at most one solution for given a, it is simpler to choose values of
c ~ Co = soja and then use (14') to determine associated values of y(x) and
h ~ y(a) by numerical integration. Each value for h exceeding that when
c = Co is achievable. A few of these minimal profiles are presented in Figure
3.6.
Each nontrivial solution thus obtained provides the profile of minimum
drag at least among those in Pfi', and this could be used in designing a
torpedo or some other missile moving in a medium under Newton's re-
sistance law. l We cannot claim that our restricted minimum drag profiles
remain optimal within a larger class, where, for example, profiles with zero
slopes are permitted. See Problems 7.27 and 9.27, and [PJ for a more
thorough discussion.
Minimum drag
0.8
profiles
0.6
0.4
0.2
0.4 0.5 0.6 0.7 0.8 0.9
Figure 3.6
1 Newton himself believed that his results might be applicable in the design of a ship's hull.
However, his resistance law is more appropriate to missiles traveling hypersonically in the thin
air of our upper atmosphere. See [FuJ.
(d) An Economics Problem
All of the classical optimization problems arose in the development of

geometry or physics and are usually concerned with optimizing one of the
fundamental quantities, length, time, or energy under various conditions. For
a change of interest, let's consider a more contemporary problem of produc-
tion planning (whose statement is taken from [Sm]).
From previous data a manufacturing company with "continuous" inven-
tory has decided that with an initial inventory ..F(O) = ..Fo and a projected
sales rate S = S(t) over a time interval [0, T], the best production rate at time
t is given by a function f!jJ = &'(t). Assuming that loss during storage occurs
at a rate which is a fixed proportion, rL, of the associated inventory ..F(t) at
time t (perhaps through spoilage), and the rest is sold at the projected rate
S(t), then we should have at time t, the simple differential relation
J'(t) = &'(t) - S(t) - rL..F(t)

(15)
(or f!jJ(t) = J'(t) + rL..F(t) + S(t)).
Now suppose that it wishes to maintain the same sales rate S(t) over a
period [0, T] but its actual initial inventory 1(0) = 10 =F ..Fo. Then from the
assumed continuity, each projected production rate function P = P(t) results
in an inventory l(t) at time t which differs from ..F(t) (at least in a neighbor-
hood of 0). With the same percentage loss, we would have as above,
P(t) = 1'(t) + rL1(t) + S(t). (15')
As a consequence, the company will experience additional operating costs
(perhaps due to handling and storage problems); these costs might be esti-
mated by a function such as
C= IT [/]2(1 - ..F)2(t) + (P - f!jJ)2(t)] dt, (16)
which takes into account the deviations in both inventory I and associated
production rate P from their "ideal" counterparts. (P is a constant which
adjusts proportions.) This is rather a crude measure of cost, but it possesses
analytical advantages.
Moreover, if we introduce the inventory deviation function y = I - ..F,
then we see from (15) and (15') that the associated production deviation is
given by
P - &' = l' - ..F' + rL(1 - ..F) = y' + rLy. (16')
Therefore the cost may be regarded as
C(y) = IT [P 2y2 + (y' + rLy)2] (t) dt (17)
which is to be minimized over :!JT = {y Eel [0, T]: y(O) = ao = 10 - ..Fo }.

Now f(1, y, z) = p2 y2 + (z + ay)2 is strongly convex on [0, T] x ~2 since

the second term is (only) convex. (Recall Example 13 in §3.3.) Here
fy = 2 [P2 Y + a(z + ay)] and fz = 2(z + ay). Therefore, by Proposition 3.9(i),
C is minimized uniquely by ayE ~T for which
fz[y(T)] = 2(y' + ay)(T) = 0 (18)
that satisfies the Euler-Lagrange equation

d
dt [2(y' + ay)(t)] = 2[(a 2 + P2)y + ay'] (t), O<t<T. (18')
From (16') we see that condition (18) simply requires that P(T) = ;?JJ(T). Is
this reasonable? Why?
If we differentiate in (18') and cancel the ay' terms from each side, then the
equation reduces to
(19)
when we substitute
y2 = a2 + p2. (19')
The general solution of the differential equation (19) is
yo(t) = c1e yt + c2e- yt , with y~(t) = y(c1e yt - c2e- yt ), (20)
and we must try to find the constants Cl and C2 so that yo(t) satisfies the
boundary conditions. We require that
Yo(O) = ao = C1 + C2,
and
o = y~(T) + aYo(T)
= c1(Y + a)e yT + c2(-Y + a)e- yT ,
or that
0= C1 (y + a)e 2YT - c2(Y - a).
From this last equation, the ratio
p ~ Y + a e2yT = C2 (21)
Y- a C1
is specified, and for this p the choices C 1 = ao/(l + p), C2 = aop/(l + p) will
satisfy both conditions. This gives the desired conclusion:
(3.14) Among all inventory functions 1 Eel [0, T] with a, p, and 1(0) = 10
prescribed, that given by
1(t) = J(t) + (1~ ~ ;0) (e yt + pe-yt), (22)
with p, Y determined by (21) and (19'), respectively, will provide uniquely the
minimum cost of operation as assessed by (16). The associated optimal produc-
tion rate is
Po(t) = &l(t) + (1 + p)-l(1o - .1"0)(1' + oc) [e yt - ey(2T-t)] (23)
Moreover, in this case the minimum cost can easily be computed. Indeed
from (17)
C(Yo) = IT [P2 Y5 + (Yb + OCYO)2] dt
and we recall that Yo satisfies (18) and (18'). We see that the integrand is just
[Yo(Yb + ocYo)]'(t), and since Yo(O) = ao, we conclude that
C(yo) = [Yo(Yb + ocYo)] 16 = -oca5 - aoyb(O),
where yb(O) = 1'(0) - .1"'(0) can be obtained by differentiating (22).
Finally recalling that ao = 10 - .1"0' we find that
Cmin = C(yO) = (10 - .1"0)

2 [y(Pp +- 11) - ]
OC , (24)
and this expression shows the effects of various choices of oc, p, T, and .1"(0) on
the minimum cost of operation. Observe that it is independent of the sign of
the initial inventory deviation.
(e) Minimal Area Problem

Our final example extends the methods of this chapter to a problem in higher
dimensions, namely, that of Plateau. In the simplified version formulated in
§1.4(b), given a bounded domain D of ~2, and a prescribed smooth boundary
function 1', we seek a function U E Cl(.O) which has these boundary values and
minimizes the surface area function
S(u) = IIv J1 + u; + u; dx dy.

Introducing ~ = {u E Cl(D) with ulaD = y} and ~o = {v E Cl(D) with
vlaD = O}, we see that this is equivalent to finding a u E ~ for which
S(u + v) - S(u) ~ 0, v V E ~o.
Now, the three-dimensional vector inequality used in establishing the
strong convexity of f(y, z) = J1 + y2 + Z2 (see Example 10 of §3.3), shows
that at each point in D:
J 1 + (ux + vx) 2
+ (uy + vy)2 - J 1 + Ux + uy ~ Juxvx + UyV y '
2 2
1 + ux2 + uy2
with equality iff Vx = Vy = O. Hence, from the assumed continuity:
S(u + v) - S(u) ~ ~S(u; v) = If
u v +u v
x x y y dx dy
D J1 + u; + u;
(as in §2.4, Example 9) with equality iff v == 0 (since Vx = Vy == 0 in the domain

D ~ v = const. = vlan == 0). Thus S is strictly convex on :!J, and again we
would seek u E:!J for which bS(u; v) vanishes, 'if v E :!Jo. Such a u would
provide the unique minimizing function for S on :!J. It would, of course,
suffice if we could find a u which is even smoother; in particular, if we could
find a u E :!J n C 2 (D) which has these properties.
For u E :!J n C 2 (D), both
U def Ux d W def Uy
= an = -----r=~:======'
J 1 + u; + u; J 1 + u; + u;'
are in C1(D) so that the integrand of c5S(u; v) may be rewritten as
Uv x + WVy = (Uv)x + (Wv)y - (Ux + w,)v. Now, if we assume that Green's
theorem holds for the domain D ([FIJ), then
fIn [(Uv)x + (WV)yJ dx dy = faD [(Uv) dy - (Wv) dxJ,
and for v E :!Jo, the line integral vanishes. Thus for v E :!Jo,
c5S(u; v) = - fIn (Ux + w,)v dx dy, (25)
and by Proposition 3.3 it is obvious that a minimum area would be given

uniquely by each u E:!J n C 2 (D) which satisfies the partial differential
equation
Ux + w, = 0 in D;
or upon substitution and simplification, which satisfies the second-order
partial differential equation
(1 + u;)uxx - 2uxuyuxy + (1 + u;)uyy = o. (26)
Equation (26) is called the minimal surface equation and it has been studied
extensively. Our uniqueness argument shows that this equation cannot have
more than one solution u in :!J, but the existence of a solution depends
upon a geometric condition on D:
(3.15) A domain D is said to be convex when it contains the line segment

joining each pair of its points.
A disk is convex while an annulus is not. If the domain D is not convex, it

is known that (26) does not always have a solution in the required set
:!J n C 2 (D), and we can draw no additional conclusions from the analysis
given here. However, it is also known that if D is convex, then (26) has a
solution in :!J n C 2 (D) for arbitrary smooth ,)" which thus describes uniquely
the minimal surface; i.e., the surface of minimal area spanning the contour
described by the graph of the boundary function ,)" among all C 1 surfaces
([OsJ). (Actually it does so among all piecewise C 1 surfaces, those described
by the graph of a piecewise C 1 function 12, which admit internal "roof-shaped"

sections. With appropriate definitions, the methods of Chapter 7 can be
extended to establish this fact.)
(Problems 3.20-3.26, 3.37)
§3.S. Minimization with Convex Constraints

Convexity may also be of advantage in establishing the minima of functions
J that are constrained to the level sets of other functions G (as in the isoperi-
metric problem). In the formulation suggested by Proposition 2.3, the next
result is apparent.
(3.16) Theorem. If D is a domain in [R2, such that for some constants Ai'
j = 1, 2, ... , N, f~, y, z) and Ajgj~' y, z) are convex on [a, b] x D [and at
least one of these functions is strongly convex on this set], let
Then each solution Yo of the differential equation

d - -
dxfz[Y(x)] = fy[y(x)] on (a, b)
minimizes
F(y) = Lb f[y(x)] dx
[uniquely] on
!!) = {y E C 1 [a, b]: y(a) = yo(a), y(b) = Yo (b); (y(x), y'(x)) E D}
r
under the constraining relations
Gj(y) ~ gly(x)] dx = Gj(yo), j = 1,2, ... , N.
PROOF. By construction (and 3.11(1)) j~, y, z) is [strongly] convex on
F(y) = r
[a, b] x D, so that by Theorem 3.5, Yo minimizes
j[y(x)] dx = F(y)
[uniquely] on!!). Now apply Proposition 2.3.

+ j~ Apj(Y)
(3.17) Remark. Theorem 3.16 offers a valid approach to minimization in the

presence of given isoperimetric constraints as we shall show by example.
§3.5. Minimization with Convex Constraints 77
However, if we introduce functions Aj = Aix) in its hypotheses, then as in 2.5
F(y) = F(y) N fba

+ jf;. AiX)gj[Y(X)J dx,
and we conclude that each solution Yo E r!J of the differential equation for
the new j minimizes F on r!J [uniquelyJ under the pointwise constraining
relations
j = 1,2, ... , N,
of Lagrangian form.
Although, in general not even one such gj[Yo(x)] may be specifiable a
priori (Why?), the vector-valued version does permit minimization with given
Lagrangian constraints. (See Problem 3.35 et seq.)
Corresponding applications involving inequality constraints are consid-
ered in Problem 3.31 and in §7.4.
F(y) = Ii (y'(X))2 dx
on
r!J = {y E c 1 [0, 1J: y(o) = 0, y(1) = O},
when restricted to the set
{y E c 1
[0,1]: G(y) ~ I1 y(x) dx = 1},
we observe thatf~, y, z) = Z2 is strongly convex, while g~, y, z) = y is (only)
convex, on IR x 1R2. Hence, we set j(x, y, z) = Z2 + AY and try to find A for
which Ag~, y, z) remains convex while the differential equation
d - -
dxfz[Y(x)] = J;,[y(x)]
has a solution Yo E r!J for which G(yo) = 1. Now since g is linear in y (and z),
Ag(X, y, z) = AY is convex for each real A. Upon substitution for j, the differen-
tial equation becomes
d A
dx (2y'(x)) = A or y"(x) = 2'
which has the general solution
The boundary conditions y(O) = 0 = C2 and y(1) = 0 = Cl + )./4 give

Yo(x) = ~). x(1 - x), which is in ~.
Theorem 3.16 assures us that Yo(x) = (-)./4)x(1 - x) minimizes F on ~
even uniquely-under the constraint G(y) = G(yo). It reinains to show that
we can choose). so that G(yo) = 1 (while ).g~, y, z) remains convex).
Thus we want
G(yo) = 1 = -
-). Jl
4 0
x(1 - x) dx = -
-),(1 1) -).
4
- - -
2 3
=-
24
or
). = -24,
and since -24g(!, y, z) = -24y remains convex, we have found the unique
solution to our problem.
(3.18) Remark. In this example we can find), to force Yo into any level set
of G we wish, since ).g~, y, z) = ).y is always convex for each value of ).. This
is not the case in general and our approach will work only for a restricted
class oflevel sets of G. (See Problem 3.29.)
The Hanging Cable
Example 2 (The catenary problem). Let's determine the shape which a long
inextensible cable (or chain) will assume under its own weight when sus-
pended freely from its end points at equal heights as shown in Figure 3.7.
We utilize the coordinate system shown, and invoke Bernoulli's principle
that the shape assumed will minimize the potential energy of the system.
(See §8.3.)
We suppose the cable to be oflength L and weight per unit length W, and
that the supports are separated a distance H < L. Then utilizing the arclength
s along the cable as the independent variable, a shape is specified by a function
y E I1Jj = C 1 [0, LJ with y(O) = y(L) = 0, which has associated with it the po-
tential energy given within an additive reference constant by the center-of-
y
H
x
(s, y(s»
-"
Figure 3.7
mass integral
F(y) = W LL y(s) ds.
However, in order to span the supports, the function y must satisfy the
constraining relation
G(y) = LL J1 - y'(S)2 ds = LL dx(s) = H,
where x(s) denotes the horizontal displacement of the point at a distance s
along/the cable, since then as elementary geometry shows, X'(S)2 + y'(S)2 = l.
Clearly ly'(s)1 ::; 1 and if 1y'(S1)1 = 1, then the cable would have a cusp at S1'
since x'(sd = 0.
Now f(§., y, z) = Wy is (only) convex on [0, LJ x 1R2 while g(§., y, z) =
-~ is by §3.3, Example 5, strongly convex on [0, LJ x IR x (-1, 1).
Thus by 3.11(1), the modified function j(§., y, z) = Wy - 2~ is
strongly convex when 2> 0. Hence by 3.16, for 2> we should seek a
solution y for the differential equation
°
d - -
d/z[Y(s)] = J;,[y(s)J on (0, L)
that is in
!!) = {y E C 1[0, L]: y(o) = y(L) = 0, ly'(s)1 < 1, V s E (0, L)}.

Upon substitution, the differential equation becomes
d( 2y'(s) )
ds J1 _ y,2(S) =W
or
2y'(s)
+c (27)
J1 - y'2(S) =
-----r==o== S
,
where we have replaced the unspecified constant 2 by W2 and introduced a
new constant c.
We know that each y E !!) which satisfies this equation for 2 > must be
the unique shape sought. Hence we can make further simplifying assumptions
°
about y if they do not preclude solution. We could, for example, suppose
y' = const., but it is seen that this could not solve (27). And we can suppose
that y is symmetric about L/2, which accords with our physical intuition
about the shape assumed by the cable. If we set 1 = L/2 it follows that
y'(l) = 0, so that from (27), c = -l; also, we need only determine y on [0, lJ,
where we would expect that y' ::; 0.
Thus from (27) we should have that
, 2 (s - l)2
Y (s) = 22 + (s _ l)2 on [0, lJ,
and so, with y(o) = 0, that
Y(S)=I' (t-l) dt=J2 2 +(t-I)21'

o J 2 2 +(t_I)2 0
or
y(s) = J2 2 + (1- S)2 - J2 2 + 12 on [0, I]. (28)
Now we can obviously suppose that 2> 0; however, we must satisfy the
constraining relation
LL Jl - y'(S)2 ds = H;
or with our symmetry assumption, we require that
I Jl - y'(S)2 ds = ~.
Upon substitution from (28), this becomes
1
1_ (1- S)2 ds = II 2 ds = H.
10 22+(1-S)2 oJ2 2 +(1-S)2 2
With the hyperbolic substitution (1 - s) = 2 sinh 0, we can evaluate the inte-
gral and fin~ that
h(0( ) <i.EH
- - -_- H
21
- s_ m
L
. h- 1 -
2
(I)
Now, h(O() d~ (sinh O()/O( is continuous and positive on (0, 00) and has by
°
L'Hopital's rule as 0( \i and 0( )" 00, the same limits as does cosh 0( viz.,
1 and 00, respectively. Thus by the intermediate value theorem (§A.l), h
assumes each value on (1,00) at least once on (0,00). Hence 30( E (0,00)
for which h(O() = L/H and for this 0(, 2 = I/sinh 0( will provide the y(s)
sought.
The resulting curve is defined parametrically on [0, IJ by
y(s) = J2 2 + (1- S)2 - J2 2 + 12,
x(s) = t Jl - y'(t)2 dt = ~- 2 sinh- 1 C~ s).

(29)
which corresponds to the well-known catenary (Problem 3.30(a)).
(3.19) Among all curves of length L joining the supports, the catenary of (29)
will have (uniquely) the minimum potential energy and should thus represent
the shape actually assumed by the cable.
Remark. This problem is usually formulated with x as the independent

variable. However, this results in an energy function which is not convex
(Problem 3.30(b)).
Optimal Performance
Example 3. (A simple optimal control problem). A rocket of mass m is to be
accelerated vertically upward from rest at the earth's surface (assumed sta-
tionary) to a height h in time T, by the thrust (mu) of its engine. If we
suppose h is so small that both m and g, the gravitational acceleration,
remain constant during flight, then we wish to control the thrust to minimize
the fuel consumption as measured by, say,
(30)
for a given flight time T.

Although T will be permitted to vary later, consider first the problem in
which T is fixed. We invoke Newton's second law of motion to infer that at
time t, the rocket at height y = y(t) should experience the net acceleration
ji =u- g, (31)
and impose the initial and terminal conditions

y(O) = y(O) = 0 and y(T) = h.
Since y(O) = 0, then y(T) = g y(t) dt, so that upon subsequently integrating
by parts we obtain
y(T) = -(T - t)y(t)/: + f: (T - t)ji(t) dt.
/
Figure 3.8
From (31) and the remaining boundary conditions, we see that
h = y(T) = f 0
T
(T - t)u(t) dt -
gT2
2'
Hence
G(u) =
deff T(T -
0 t)u(t) dt = h
gT2
+2 = k, say, (32)
and we are to minimize F on

~ = {u E CEO, T], u ~ O}
subject to the isoperimetric constraint (32).
According to Theorem 3.16, we introduce a constant A and observe that

the modified integrand
if.!., u, z) = u2 + A(T - flu
will be strongly convex for all A, since the second term is linear in u.
Moreover, iz == O. Thus, a U o E ~ which satisfies the equation !..[u(t)] =
0= 2u(t) + A(T - t) and meets the constraint (32) will suffice. (See Problem
3.18.)
We require uo(t) = -(A/2)(T - t), where A :s; 0 is to be found to have
Ok = f: (T - t)uo(t) dt = -~ f: 3
(T - t)2 dt = _ A: ,
or
6k
-2 = T 3 '
so that
() 3k(T - t)
Uo t = T3 (33)
Observe that from (30) and (32) it follows that
we may now use simple calculus to minimize this expression with respect
to T and thus obtain an optimal flight time To = (6h/g)1/2.
(3.20) Remark. We know that (33) provides the unique solution to our prob-
lem. However, observe that from (32), the corresponding maximum thrust is
3k = (3h)
uo(O) = T2 T2 + 1.5g;
§3.6. Summary: Minimizing Procedures 83
when T= To as above, uo(O) = 2g which might not be achievable. A more

realistic problem could require 0 :::;; u :::;; P( < 2g). This is a Lagrangian in-
equality which can also be treated by means of a multiplier function as
in Proposition 2.5. The resulting solution admits operating at maximal thrust
(p) until a switching time 't at which reduction can occur. The details are
reserved until Problem 7.23, and a simpler convex problem with Lagrangian
inequalities is examined in §7.4. Extensions of this approach are feasible
(Problems 3.38 and 3.39) even to problems with discontinuous controls as
will be shown in Chapter 10.
(Problems 3.27-3.32)
§3.6. Summary: Minimizing Procedures

In this chapter we have used elementary concepts of convexity to supply the
framework for making an educated guess about the solution of a minimiza-
tion problem. In the presence of strict convexity, a guessed solution is the
unique solution. However, with each specific problem, it is not essential to
establish convexity according to predetermined definitions. Inspection of the
results and applications in this chapter suggests the following procedures:
I. To minimize F on f!} !:; qy (a linear space):
First. Show that when y, y + V E f!}, then F(y + v) - F(y) ~ I(y; v) where
I(y; v) is some new expression which admits further analysis.
Second. If possible, characterize those v which permit the equality
F(y + v) - F(y) = I(y; v). (Ideally, equality at y => v = 0.)
Third. Note the restrictions on v which occur when y, y + V E!!), and
transform I(y; v) so that conditions (on y) under which it vanishes for all such
v can be discerned.
Fourth. Show that there is a y = Yo E f!} which meets these conditions.
If it exists, this Yo will be a solution (and it may be the solution) to the

problem. 0
Remarks. To obtain the basic inequality, elementary facts such as (y" + V")2
- (y")2 ~ 2y"v" may suffice. It is not essential to recognize I(y; v) as
c5F(y; v)-or indeed, even to consider this variation.
In transforming I(y; v), we may make further simplifying assumptions
about y (which do not exclude y from f!}). In particular, we may assume that
y has as many derivatives as required to integrate an expression such as
$:y"(x)v"(x) dx by parts as often as desired. "Natural" boundary conditions
for a solution y may arise in this process.
Finally, although we may not be able to guarantee a unique solution

Yo to the problem, it may be possible to use information obtained in the
second step to characterize the class of solutions; e.g., Yo + const.
II. To minimize F on f!} ~ ilJJ in the presence of further constraints

involving functions Gj,j = 1,2, ... , n:
First. Use the device of Lagrangian multipliers Aj to suggest an augmented

function F whose minimization on f!} could solve the constrained problem.
(See 2.3, 2.5, 3.17 and Problem 3.35.)
Second. Find a Yo E f!} which minimizes F on f!} (possibly under a sign
restriction on each Aj ).
Third. Determine the Aj so that Yo meets the given constraints.
Fourth. Examine the signs of the Aj (if necessary) to see that the restrictions
in the second step have been met.
If all of these conditions are satisfied then Yo is a solution to the con-

strained problem (and it may be the solution). 0
III. If a usable basic inequality cannot be obtained for F (or for F), it may
be possible to reformulate the problem-or consider a restricted version
of the problem-(perhaps expressed in other coordinates) in terms of new
functions which admit the analysis of I or II. Also, the solution of one
minimization problem usually solves some associated problems involving
constraints. 0
Success in implementing these programs requires a combination of skill,

perseverance, and luck. However, they seem to provide the only possibility of
obtaining a solution by methods which can be considered elementary, and
successful implementation is possible as the examples and problems of this
chapter demonstrate. Alternatives to these procedures require a considerably
more sophisticated theoretical framework and are at least as difficult to
utilize. (See Chapter 9 and [Ak].)
PROBLEMS
3.1. For which of the following functions f, is ffdf, y, z) convex on [a, b] x [R2?
For which will ffdf, y, z) be strongly convex on this set?
(a) f(x, y, z) = x + y - z, [a, b] = [0, 1].
(b) f(x, y, z) = x 3 + y2 + 2z 3, [a, b] = [0, 1].
(c) f(x, y, z) = ~ + x 2y2, [a, b] = [0, 1].
(d) f(x, y, z) = (x sin X)[y4 + Z2], [a, b] = [-n/2, n/2].
Problems 85
(e) f(x, y, z) = _x 4 + eXy + Z2, [a, b] = [-1,1].

(f) f(x, y, z) = yz2 + cos x, [a, b] = [0, 1].
(g) f(x, y, z) = e Y sec 3 x - z, [a, b] = [0, n/4].
(h) f(x, y, z) = _x 2y + z4, [a, b] = [-1,1].
(i) f(x, y, z) = _xy2 + Z2, [a, b] = [ -1,1].
(j) f(x, y, z) = e Xy4 - xy + 2Z2, [a, b] = [0, 1].
(k)* f(x, y, z) = (1 + sin X)y2 + (1 + cos X)Z2, [a, b] = [ - 8, 8].
(1) f(x, y, z) = x 2(cy + Z)2, c = const. [a, b] = [ -1,1].
(m)* f(x, y, z) = 2Z2 + zlzl. Hint: Consider graph.
3.2. Let f(~, y, z) and g(~, y, z) be convex on S s; [R3:

(a) Prove that f~, y, z) + g~, y, z) is also convex on S.
(b) Give simple examples of other algebraic combinations of such functions,
i.e., the difference, product, and quotient, which are not convex.
(c) IfO::; p = p(x) is continuous, then prove that when defined p(~)f~, y, z) is
convex.
(d) Can the sum in (a) be strongly convex on S when each of its terms is only
convex? Justify your conclusion.
3.3. Let f(~, y, z) be strongly convex on S = [a, b] x [R2, g(~, y, z) be convex on S,

and ex, p,}, E C[a, b].
(a) Prove that f~, y, z) + g~, y, z) is strongly convex on S.
(b) IfO 0, for z E [R, conclude that f(~, z) = g((sin ~)z)
is strongly convex on (0, n) x R Hint: Proposition 3.10. (This result
generalizes.)
3.4*. Show that if f,f" andfz are continuous on [a, b] x [R2, thenf(~, y, z) is convex
on [a, b] x [R2 iff
f(x, tYl + (1 - t)Y2, tZ 1 + (1 - t)Z2) ::; tf(x, Yl' Zl) + (1 - t)f(x, Y2, Z2)
VXE[a,b]; tE(O, 1); Yj,ZjE[R, j=1,2.
(Hint: See Problem 0.7.)
3.5*. (a) Show that iff, f,y, f,z, and!.z are continuous on [a, b] x [R2, then f~, y, z)
is convex on [a, b] x [R2 iff the matrix of second derivatives [~: i:J is
positive semidefinite on [a, b] x [R2, i.e.,
(U v) [f,y
f,z
f,zJ
fzz v
(u) ~ 0, VXE[a,b]; y,z,u,vER
(Hint: See Problem 0.8.) (Note: This condition is equivalent to requiring

that fyy ~ 0, !.z ~ 0, and L\ = f,yfzz - (f,z)2 ~ 0.
(b) Use this approach on Problems 3.1(k) and (1).
3.6-3.15. In these problems, verify that the integrand function is strongly convex (on
the appropriate set) and find the unique minimizing function for F
(a) on f!J. (b) on f!Jl' (c) on f!J2'
3.6. F(y) = Ii x- l y'(X)2 dx, c

~l = {y E l [1, 2]: y(2) = 3},
~ = {y E c l [1, 2]: y(l) = 0, y(2) = 3}. c
~2 = l [1, 2].
3.7. F(y) = g [2e X y(x) + y'(X)2] dx,
~ = {y E c l [0, 1]: y(o) = 0, y(l) = 1}. ~l = {y E c l [0, 1]: y(o) = oJ.
3.8. F(y) = n° JxJ1 + y'(4 dx, ~l = {y E C l [5, 10]: y(lO) = 6},
~ = {y E c l [5, 10]: y(5) = 4, y(lO) = 6}. ~2 = c l [5, 10].
3.9. F(y) = Ii [2Y(xf + x 2y'(X)2] dx, ~l = {C l [1,2]: y(l) = 1},
~ = {y E c l [1,2]: y(l) = 1, y(2) = 5}. ~2 = c l [1,2].
(Hint: The differential equation has two linearly independent solutions of the
form x P , P E [R.)
3.10. F(y) = Jo/ 6 [(y'(x) - cos X)2 + 4y(x)] dx,
c
~ = {y E l [0, n/6]: y(O) = 0, y(n/6) = t}.
c
~l = {y E l [0, n/6]: y(n/6) = t}.
3.11. F(y) = Ii
x- l J1 + y'(X)2 dx,
~ = {y E c l [1, 2]: y(l) = J8:,
y(2) = Js}
~l = {y E c l [1, 2]: y(l) = j8}.
3.12. F(y) = J:'l e"(x) dx,
~ = {y E c l [ -1, 2]: y( -1) = 2, y(2) = 11}.
3.13. F(y) = H/2 [y(x) + J1 + y'(X)2] dx,

~ = {y E c l [0, n
y(O) = -1, y(t) = -J3/2}.
~l = {y E Cl[O, n
y(O) = -1}.
3.14. F(y) = Jo'4
y'(X)4 sec x dx, 3
~ = {y E c l [0, n/4]: y(O) = 0, y(n/4) = 1}.
~l = {y E C l [0, n/4]: y(n/4) = 1}.
3.15. F(y) = H
[Y'(X)4 - 4y(x)] dx,
~ = {y E C l [l, 8]: y(l) = 2, y(8) = -37/4}.
~l = {y E c l [1,8]: y(l) = 2}.
3.16. (a) Demonstratethat the function F(y) = g y2(X)Y'(x) dx is convex on

~ = {y E C l [0, 1]: y(O) = 0, y(l) = 1},
although the integrand function f(~, y, z) = y2 Z is not convex on [0, 1] X
[R2.
(b)* Prove that the function f(~, y, z) = (z + 3y)2 is (only) convex on [a, b] x
[R2, but F(y) = J~(y'(x) + 3Y(X))2 dx is strictly convex on
~ = {y E Cl[a, b]: y(a) = a l , y(b) = bd.

What happens if 3 is replaced by a number rx?
3.17. (a) Show that f~, y, z) = ~z + y is convex, but not strongly convex on
[1,2] x [R2.
r
(b) Can you find more than one function which minimizes
F(y) = [xy'(x) + y(x)] dx

on
~ = {y E C l [1,2]: y(l) = 1, y(2) = 2}?
Problems 87
3.18. Suppose thatf(!, y) is [strongly] convex on [a, b] x IR and set
F(y) = f f(x, y(x)) dx.
(a) Show that each YEP} = C[a, b] which satisfies

f,(x, y(x)) = 0, \fxE[a,b]
minimizes F [uniquely] on P}.
(b) Show that each y E P}* = {u E C[a, b]: J!u(x) dx = o} which satisfies
f,(x, y(x)) = const., \f x E [a, b]
minimizes F [uniquely] on P}*.
(c) Let f(x, y) = y2 - g(x)y, where 9 E C[a, b] is a given function. Find the
unique minimizing function for F on P} and on P}*.
3.19. (a) When bl > 0, explain why the restricted surface-area-of-revolution func-
tion (see §1.4(a))
S(y) = 271: r xJl + y'(4 dx
is minimized on P} = {y E C l [1, b]: y(l) = 0, y(b) = bl } uniquely by a Yo
that makes
xy'(x) 1
--r~===;c = const. = -
Jl + y'(X)2 c
say, on (1, b).
(b)* Show that cYo(x) = cosh-l(cx) - cosh- l c provided that c> 1 can be
found to make yo(b) = bl . Graph Yo and discuss how to guarantee that c
exists.
3.20*. Minimum Transit Time of a Boat. (See §1.2(b) and Problem 1.3.) Let
where r is a given continuous function,

(1 - r(x)2)-1/2, and
° ~ r(x) < 1 on [0, Xl]' IX(X) =
P) = {y E Cl[O, Xl]: y(O) = 0, y(xd = Yl} with Xl> 0, Yl > 0.

(a) Prove that the integrand function
f~, z) = 1X(!)JI + (1X~)Z)2 - 1X~)2r(!)z
is strongly convex on [a, b] x IR.
(b) Show that each YEP) which makes (1X 3 y')(x) [1 + (lXy,)2(x)r l/2 - (1X2r)(x)
constant on (a, b) minimizes T uniquely on P}.
(c) Verify that y(x) = H(r + CIX- 2) [(1 - cr)2 - c 2rl/2(t) dt will serve pro-
vided that the constant c can be chosen properly.
(d) Show that the boundary value Yl = J~' r(x) dx is always achievable with
a proper choice of c, and find the minimizing function in this case.
(e) What happens ifr(x) = r = constant?
°
(f)* For the linear profile r(x) = (1 - 3x)/2, ~ X ~ Xl = i, show that the
admissible choices of c are restricted to -~ < c < t, and for this range of
c the integral defining y(x) in part (c) is bounded. (This demonstrates that
we cannot always choose c to meet the boundary conditions. Explain why
physically.)
3.21. (a) Verify that Theorem 3.7 remains valid for integrands of the form f(x, z) =
p(x)~, where p is continuous on (a, b], p(x) > 0, and S:p(x) dx <
00. (For example, p(x) = X- 1/2 on (0, 1].)
(b)* More generally, suppose that f(~, y, z) is [strongly] convex on (a, b) x
[R2, and Yo E C[a, b] is a C 1 solution of the differential equation
(d/dx)fz[y(x)J = J;,[y(x)J on (a, b).
r f[(yo + v)(x)J dx;?: r

Show that if [ex, f:i] ~ (a, b) and v E C 1 (a, b) then
f[yo(x)] dx + v(x).fz[yo(x)]
Thus when max Ifz[Yo(x)J1 :0::; M < +00, conclude that Yo minimizes
I:·
F(y) = S:f[y(x)J dx [uniquely] on ~* ~ OJJ = C[a, b] (1 C 1 (a, b), where
~* = {y E OJJ: y(a) = Yo (a), y(b) = y(b);
F(y) exists as an ?mproper Riemann integral}.
(This extension of Theorem 3.5 to improper integral functions F also
permits consideration of functions Yo whose derivative "blows up" at the
end points.)
(c)* Make similar extensions of Theorem 3.7 and 3.9.
(d)* Suppose f.[Yo(x)J is bounded near x = a, but only (b - x)fz[Yo(x)J is
bounded near x = b. Show that we can reach the same conclusion as in
part (b) on ~' = {y E ~*: y'(x) ...... bi as x 7" b}. Hint: Use the mean value
theorem on v near b.
3.22*. A Brachistochrone. (See §3.4(b).)
(a) Show that the time of transit along the cycloid joining the origin
to the end point (Xl' Yl) for ydXl < n/2 is given by Tmin =
j(xd2g) [Odsin(Od2)J where 01 is the parameter angle of the end point.
Hint: Use equations (12), (13).
(b) For the case Xl = Yl = 1, compute 01 and Tmin . Compare with the
answers obtained for the straight line and the quarter circle in Problem
1.2.
(c)* Use the results of Problem 3.21 to extend the analysis in §3.4(b) to the
case when ydXl = n/2.
(d)** Can you use the methods of this chapter to establish the minimality of
the cycloid when Yl/x 1 > n/2 for some class of curves?
3.23. A Seismic Wave Problem. (See Problem 1.8(b) and Figure 1.12(b).)
(a) Show that the integrand function f(~, z) = ~/~ is strongly convex
on [a, b] x [Rwhena>O.
(b) Conclude that each y E ~ = {y E C 1 [a, b]: y(a) = 0, y(b) = b1 > O} which
satisfies for some positive constant r, the equation y'/j1 + y'(X)2 = x/r on
(a, b) will minimize the time-of-travel integral T(y) = P g [j1 + y'(X)2/X] dx
on ~ uniquely.
(c) Show that the associated path is along a circular arc of radius r with center
on the y axis.
Problems 89
(d) Use the parametrization x = r sin 0 to evaluate the time-of-travel integral

T(y) = p g ds/x along this path, in terms of IX, and fl, the initial and final
values of 0, when 0 < IX < fl < n/2, as in Figure 1.12(b).
(This last result affords an experimental determination of the unknown
physical constants (J and p, by measuring the actual time of travel required for
seismic waves from explosions on the earth's surface to reach a point in a
mineshaft some distance away. Thus far, however, we must demand that the
resulting geometry be compatible with our assumption that 0 < IX < fl < n/2.)
3.24*. Geodesics on a Cone: I.
Each point on the surface S of a right circular cone of apex angle 21X has the
cartesian coordinates (x, y, z) = (r cos 0, r sin 0, ar), where a = cot IX, and r, 0
are the polar coordinates shown in Figure 3.9.
~~-----------.y
Figure 3.9
(a) If a curve on S can be represented by r = r 0 , with r E all = C l [Ol, O2],

show that its arc length is L(r) = s:~ r(0)2 + b 2r'(0)2 dO, where b 2 =
1 + a2 •
(b) Consider the convexity of f(y, z) = J
y2 + b2Z2, and explain why it would
be appropriate to consider minimizing L on
!?)* = {r E all: r(Oj) = rj , j = 1, 2; r(0)2 + r'(0)2 > OJ.
(c)* Show that a function r E !?)*, which for R = b log r satisfies the equation
d bR'(O) 1
- = on (Ol, ( 2 ),
dO J1 + R'(0)2 J1 + R'(0)2
should minimize L on !?)*.
(d) Make the substitution R'(O) = tan qJ(O) in (c), and conclude that the
resulting equation is satisfied when qJ'(O) = lib = sin IX, or when
r(O) = Cl sec(b- 1 0 + c)
for constants c, Cl'
(e) Suppose that 01 = 0, and 0 < f3~ b- 1 02 < n. Prove that an r E ~* of the
form in (d) can be found if C 1 = r 1 cos c, where
cos f3 - (rdr2)
tanc= . R
sm I'
and argue that this is possible.
(f)* Show that even though f(y, z) is not strongly convex, L is strictly convex
on ~*, so that the minimizing function found in (e) is unique. Hint: Prove
that ifr, r + v E ~*, then L(r + v) - L(r) :2: bL(r; v) with equality iff vr' =
v'r, or v = const. r.
(g) Conclude that when r 1 = r 2, the circular arc r(O) = const. is not the geode-
sic as might have been conjectured.
Geodesics on a Cone: II.
Consider the right circular cone shown in Figure 3.9. To find geodesic
curves of the form 0 = O(r) joining points (rl' 01 ) and (r2' O2), we assume with-
out loss of generality that r 2 > r 1 > 0, 01 = 0, and 0 ::; O2 ::; n.
(h) Suppose that 0 E C 1 [rl' r 2], derive the length function L(O). Is it convex?
(i) When O2 #- 0, prove that L(O) is minimized uniquely on
~ = {O E C 1 [rl' r 2]: 0(0) = 0, 0(r2) = 02},
by 0 = b sec- 1 (rlc) - b sec- 1 (rdc), provided that the constant c can be
chosen to make 0(r2 ) = O2 ,
(j) What happens when O2 = O?
3.25*. Beam Deflection. When a cantilevered beam of length L is subjected to a
distributed load (force per unit length) p(x) as shown in Figure 3.1O(a), it will
deflect to a new position which can be described by a function y E C2 [0, LJ.
According to the theory of linear elasticity, the potential energy is approxi-
mated by
U(y) = t ctJlY"(X)2 - p(x)y(x)] dx,
where Jl is a positive constant (called the flexural rigidity) determined by the

beam cross section and material, and the shape assumed by the deflected beam
will minimize U on
~ = {y E C2 [0, L]: y(O) = y'(O) = O}.
/.
y (a) (b)
Figure 3.10
Problems 91
(a) Prove that each y E fi) II C4 [0, L] which satisfies the differential equation
llyIV(X) = p(x) and the "natural" boundary conditions y"(L) = 0, y"'(L) =
is the unique minimizing function for U on fi). (The physical meaning of the
°
natural boundary conditions is that both the bending moment and shear
force are zero at the free end of the beam.) Hint: Show that U is strictly
convex on fi) and integrate the y"v" term of bU(y; v) by parts twice.
(b) Solve the differential equation from part (a) when p(x) = w = const.,
selecting the constants of integration so that the solution is in fi) and
satisfies the given natural boundary conditions. (This would be the case for
deflection of a beam under its own weight.) What is the maximum deflec-
tion and where does it occur?
(c) If at x = L, the beam is pinned and has a concentrated moment M applied
as shown in Figure 3.1O(b), the potential energy is approximated by
U*(y) = LL [tIlY"(X)2 - p(x)y(x)] dx + My'(L),
and the shape a~sumed by the deflected beam will minimize U* on

fi)* = {y E C 2[0, LJ; y(O) = y'(0) = y(L) = OJ.
Find a differential equation and a natural boundary conditon at x = L
which are sufficient to determine the unique minimizing function for U*
on fi)*.
(d) Solve the differential equation obtained in part (c) when p(x) == 0, choosing
the constants of integration so that the solution is in fi)*. What is the
maximum deflection and where does it occur?
(e) Define [strong] convexity for the function f05., y, z, r) on say [a, b] x [R3.
(f) Use your definition in (e) to conclude that the integrands of U and of U*
are strongly convex.
(g) Use your definition in (e) to characterize the [strict] convexity of the
integral function of Problem 2.9 on fi) = {y E all: y(a) = a 1, y(b) = bd.
°
Hint: If [v(x)v'(x)]v"(x) == on (a, b), then v"(x) == 0.
3.26. Dirichlet's Integral. Let
J(u) =!2 JDr (u; + u;) dA
and
where D is a Green's domain the x, y plane and y is a given smooth function

on aD.
(a) Show that each u E fi) which satisfies Laplace's equation, U xx + Uyy = 0, in
D is the unique minimizing function for J on fi).
(b) Find the minimizing function for J on fi) when D is the annulus
{(x, y): 1 < x 2 + y2 < 4}
and y is given by
O, on x 2 + y2 = 1,
y- {
In 2, on x 2 + y2 = 4.
(Hint: Look for a solution to Laplace's equation of the form u(x, y) =
u(x 2 + y2) and find an ordinary differential equation satisfied by u.)
(c)* Define [strong] convexity for a function f(}[, y, Z) on 15 x ~d+1, where D

is a Green's domain in ~d, and show that when d = 2, the integrand
function in (a) is strongly convex on 15 x ~3. (See §6.9.)
(d)* Find an analogue of Theorem 3.5 for the integral function
F(u) = Iv f(X, u(X), Vu(X» dX = Iv f[u(X)] dX

on
£1fi = {u E C I (15): ulaD = y}.
Hint: Use the divergence theorem of vector calculus.
3.27. Find the unique minimizing function for
F(y) = f xy(x) dx
on
£1fi = {y E C l [0, 1]: y(O) = y(1) = OJ,
when restricted to the set where
G(y) ~ f y'(X)2 dx = l.
3.28. Show that Yo (x) = -1 + J2 - (x -W is the unique minimizing function for
F(y) = II J1 + y'(X)2 dx
on
£1fi = {y E CI[O, 1]: y(O) = 0, y(1) =.J2 - 1},
when constrained to the set where
G(y) =deff l y(x) dx = -n - 1

-.
o 4 2
3.29. Is there a value of). which permits Theorem 3.16 to be used to minimize
F(y) = So" y'(X)2 dx

on
£1fi = {y E C l [0, n]: y(O) = y(n) = OJ,
when further constrained to the set where
I" y2(X) dx = 1?
3.30. Catenary Problem. (See §3.5, Example 2.)

(a) Verify equation (29) and eliminate the parameter s to obtain the equation
y =). cosh (x ~ h) _ J).2 + [2,
for O::s; x ::s; 2h = H. (This is a more common representation for the

catenary joining the given points.)
Problems 93
(b) Formulate the problem using x as the independent variable and conclude
that this results in an energy function U which is not convex on
fi) = {y E C 1 [0, H]: y(O) = y(H) = O}. (Hint: Use v = - Y to show that
U(y + v) - U(y) is not always greater than or equal to DU(y; v) when y,
y + V E fi).)
(c)* Use the arc length s, as a parameter to reformulate the problem of finding
the minimal surface of revolution (as in §1.4(a)) among all curves of fixed
length L joining the required points. (Take a = 0 and al = 1 ~ b1 .)
(d) Conclude that the problem in (c) is identical to that of a hanging cable for
an appropriate W, and hence there can be at most one minimizing surface.
(See §3.5.)
3.31. Determine the (unique) function y E C[O, T] which maximizes
U(y) = IT e- P' log(1 + y(t)) dt
subject to the constraint L(y) = J~ e-·'y(t) dt ~ I, where IX, p, and I are positive
constants. Hint: Problem 3.18, with 2.4, 3.17. (This may be given the interpre-
tation of finding that consumption rate function y which maximizes a measure
of utility (or satisfaction) U subject to a savings-investment constraint L(y) ~ I.
See [Sm], p. 80. y is positive when I is sufficiently large relative to IX, p, and T.)
3.32*. Dido's Problem.
Convexity may be used to provide partial substantiation of Dido's conjec-
ture from Problem 1.5, in the reformulation suggested by Figure 3.11.
Verify her conjecture to the following extent:
(a) If b > lin, prove that the function representing a circular arc (uniquely)
maximizes
A(y) = fb y(x) dx
on
fi) = {y E C 1 [ -b, b]; y(b) = y( -b) = O},
=
when further constrained to the I level set of L(y) J~bJ1 + y'(X)2 dx.
(b) If b = lin, show that the function representing the semicircle accomplishes
the same purpose for a suitably chosen fi)* (see Problem 3.21).
y
Figure 3.11
(c)* In parts (a) and (b), compute the maximal area as a function of p, the angle
subtended by the arc; show that this function increases with p on (0, n].
(d) Why does this not answer the problem completely? Can you extend the
analysis to do so?
3.33. Let I be an interval in [R and D be a domain of [R2d. For x E [R, and Y =
(Yl' ... , Yd)' Z = (Zl' ... , Zd) E [Rd a function f~, Y, Z) is said to be [strongly]
convex on I x D if f, fy = U;", ... ,J;,,), and fz = (1. " ... , f •.) are defined and
continuous on this set l and satisfy the inequality
f(x, Y + v, Z + W) - f(x, Y, Z) ~ fy(x, Y, Z)· V + fz(x, Y, Z)· W,
'v' (x, Y, Z), (x, Y + v, Z + W) E I x D
[with equality at (x, Y, Z) only if V = (!J or W = (!J].

(a) Show that if f~, Y, Z) is [strongly] convex on [a, b] x
r r
[R2d, then
F(Y) = f[Y(x)] dx = f(x, Y(x), Y'(x)) dx
is [strictly] convex on
{!) = {Y E (Cl[a, b])2d: Y(a) = A, Y(b) = B},
where A, B E [Rd are given.
(b) Iff~, Y, Z) is strongly convex on [a, b] x [R2d, then prove that each Y E {!)
which satisfies the vector differential equation (d/dx)fz[Y(x)] = fy[Y(x)]
(i.e.; (d/dx)f•. [Y(x)] = fy.[Y(x)], j = 1, 2, ... , d) on (a, b) is the unique
minimizing (unction for Fon {!).
3.34. Use the results in Problem 3.33 to formulate and prove analogous vector
valued versions of: (a) Theorem 3.7, (b) Corollary 3.8, and (c) Proposition 3.9
in §3.2.
3.35. (a) Formulate and prove a vector valued version of Theorem 3.16 in §3.5.
(b) Modify Theorem 3.16 to cover the case of a single Lagrangian constraint,
and prove your version. Hint: Proposition 2.5 with 3.17.
(c) Formulate a vector valued version of the modified problem in (b) that
covers both isoperimetric and Lagrangian constraints.
3.36. (a) Show that Yo(x) = sinh x = (eX - e- X)/2 minimizes H(Yl (X)2 + y~ (X)2) dx
uniquely for Yl E C l [0, 1] when Yl (0) = 0, Yl (1) = sinh 1.
(b) Apply 3.16 as extended above to the problem of minimizing
F(Y) = f [Y~(X)2 + Y~(X)2J dx

on
{!) = {Y = (Yl' Y2) E (Cl[O, 1])2: Y(O) = (0, 1), Y(I) = (sinh 1, cosh I)}
under the constraint: g[Y(x)] ~ YHx) - Yl(X) == on (0,1). °
(c) Compare (a) and (b) to make the "correct' choice for A.. Can you find a
means of choosing this A. without using the result of (a)?
1 That is, their components are continuous on this set.

Problems 95
3.37. Minimal Area with Free Boundary. (See §3.4(e) and §6.9.)
(a) When u E Cl(O) and v E Cl(D), verify that formally
bS(u; v) = - f D
(Ux + Wy)v dA +
(Vu' N)v du ,
+ u~ + u;
f J1
aD
where N is the outward normal to aD and du is the element of arc length
along aD. Hint: Green's theorem in its divergence form.
(b) Conclude that if Ux + Wy == 0 in D and Vu' N = 0 on a sub arc K ~ aD,
then u provides uniquely the minimal surface area among those competing
functions which agree with it on K, that part of the boundary complemen-
tary to K.
3.38*. An Optimal Control Problem with Damping.
If the motion of the rocket discussed in Example 3 of §3.5 is opposed by its
speed through the atmosphere, then the equation of motion becomes approxi-
mately
y = u - g - rx.y, for a constant at: > O.
(a) Rewrite the last equation as (d/dt)(e·ty) = e·t(u - g) and integrate it as in
the text under the same conditions to obtain the new isoperimetric condi-
tion
G(u) ~ IT (1 - e-·(T-t»)(u(t) - g) dt = at:h.
(b) Show that the function uo(t) = Ao(l - e-·(T-t») will minimize F(u) =
g u 2 (t) on C[O, T] uniquely, subject to the constraint in (a) for an appro-
priate choice of Ao.
(c) For at: = 1, find Ao and determine F(uo).
3.39. A heavy rocket-propelled sled is accelerated from rest over a straight horizon-
tal track of length 1 under velocity-proportional air resistance as in the previ-
ous problem.
(a) To optimize the fuel consumption over time T show that we might con-
sider minimizing
F(v) = IT (Ii + at:V)2 dt
on
p) = {v E Cl[O, T]: v(O) = O}
under the isoperimetric condition G(v) ~ g v(t) dt = 1, where v(t) is the
velocity at time t and at: is a positive constant.
(b) Find a minimizing velocity function vo. Is it unique? Hint: f(y, z) =
(z + at:y)2 is convex! (Why?)
(c) When at: = 1 and 1 = h, compare F(v o) with F(uo) in part (c) of the previous
problem under zero gravity conditions. Should they be the same? Explain.
3.40*. (Newton's parametrization.) In (14), let t = - y' so that
(1 + t 2)2 1
cx = = - + 2t + t 3 = cx(t) say, for c > O.
t t
(a) Show that for t > 1/A, dx/dt > 0, so that t can be used as an independent
variable when t ;::: 1/J3.
(b) Conclude that

dy 1 3
c- = - - 2t - 3t
dt t
so that ifwe replace c by 4/c, we get the parametric equations
4
y = -c ( log t - t 2 - -3t + -47) + Yl (t ~ 1/j3),
4 4
where at t = 1: x = c and Y = Yl' Plot some points on this curve when

c = Yl = 4 using a graphing calculator, if necessary, and compare with the
curves graphed in Figure 3.6.
(c) In general, we need to choose c and Yl to force our curve to pass through
given points. For simplicity, suppose that at t = 1, x = c = a < 1 and
Y = Yl = h. (This choice is further motivated in Problem 7.27). Then to
have Y = 0 when x = 1, prove that there is a unique T > 1 at which
and
Hint: Show that for T> 1, H'(T) > 0 and that H assumes each positive
value. (Recall the argument in §3.4(b).)
3.41. (Newton's minimum drag problem.) If we permit a = 0 in the problem of

§3.4(c), then we wish to minimize the drag integral
F(y) =
e1 +xdx
J 0 Y'(X)2
on
~* = {y Eel [0, 1]: y(O) = h, y(l) = 0, y'(x) :;::; OJ,
and we are now seeking the profile of minimum drag for an entire body of
revolution, not just a shoulder.
(a) Show that if we remove the last restriction from ~* and admit zig-zag
profiles Y with large slopes, then we could obtain arbitrarily small values
for F(y)!
(b) When h = 1, compare the drag values of the profiles for a cone C, a
hemisphere H, and a truncated cone T in which y(x) = h when x :;::; t.
(c) When h = 2, compare the drag values of the profiles for a cone C, a
paraboloid of revolution P, and a truncated cone T in which y(x) = h when
x :;::;t.
(d) When h = j, repeat part (c) and conjecture about the superiority of trun-
cated cones or other flattened objects. In particular, using m = h/(1 - a)
can you find the "best" truncated cone To for given h?
(e) Show that with initial flattening permitted, we would need to minimize the
Bolza-type function G(y, a) = a 2 /2 + F(y) on ~ x [0, 1J, where F and ~
are as in §3.4(c), and we require y' :;::; O. Can our previous results from
convexity be used in attacking this problem? How?
CHAPTER 4
The Lemmas of Lagrange and

du Bois-Reymond
In most of the examples in Chapter 3, we examined a real-valued function F

defined on a domain offunctions!!). We obtained for F an integral condition
in the form I(y; v) = 0, Tt v in an auxiliary domain !!)o, which is sufficient to
guarantee that each y E!!) that satisfies it must minimize F on !!).
By inspection (after reformulation if necessary) we were able to guess a
restricted class of y which could meet this condition and, in most cases, find
a particular y E !!) which would do so. This, of course, leaves open the possi-
bility that other minimizing functions might exist. In the presence of strict
convexity, we have seen that this cannot occur (Proposition 3.3). Without
strict convexity, we may have an alternative possibility. (See Problem 3.17.)
To explore this fully we should determine conditions necessary for a minimiz-
ing y and this will be carried out in the next chapter.
However, there are already related necessary conditions of both mathe-
matical interest and importance which we can consider here. For instance, in
several examples we observed that the choice h(x) = const. would make
r~ h(x)v'(x) dx = 0, whenever v(a) = v(b) = 0. Thus the constancy of h is suffi-
cient to guarantee the vanishing of all such integrals. But is it necessary? i.e.,
will the vanishing of these integrals guarantee the constancy of h? YES.
(4.1) Lemma (du Bois-Reymond). If hE C[a, b] and r:h(x)v'(x) dx = 0,

Tt v E!!)o = {v E C 1 [a, b]: v(a) = v(b) = O},
then h = const. on [a, b].
PROOF. For a constant c, the function
v(x) ~ LX (h(t) - c) dt
97
98 4. The Lemmas of Lagrange and du Bois-Reymond
is in C1[a, b] (by A.B, the fundamental theorem of calculus) with derivative

v'(x) = hex) - e on (a, b), and it satisfies the condition veal = o. It will be in
f0 0 if in addition v(b) = 0; i.e., if
r (h(t) - c) dt = 0 or e=b ~a r h(t) dt.
0::::;; r(h(x) - e)2 dx = r

Thus for this e and v, we have from the hypothesis that
(h(x) - e)v'(x) dx
= r h(x)v'(x) dx - ev(x)
Hence, from A.9 it follows that on [a, b], the continuous integrand
I: = O.
(h(x) - e)2 == 0 or hex) = e = const. as asserted. D
The next result should be compared with Theorem 3.5.
(4.2) Proposition. If g, h E C[a, b] and J~ [g(x)v(x) + h(x)v'(x)] dx = 0,

V V E f0 0 = {v E C1[a, b]: veal = v(b) = O},
then h E C1[a, b] and h' = g.
PROOF. J:
Let G(x) = get) dt for x E [a, b]. Then G E C 1 [a, b] and G' = g by
r r
A.B. Hence integrating the first term of the integral by parts gives
[g(x)v(x) + h(x)v'(x)] dx = [hex) - G(x)]v'(x) dx + G(X)V(x{,

so that
r [hex) - G(x)]v'(x) dx = 0,
and by the preceding lemma 4.1,

V V E f0o,
hex) - G(x) = const. = e, say, on [a, b].

But then
h = G + e E C1[a, b] and h' = G' = g as asserted. D
Setting h == 0 in this proposition gives the
(4.3) Corollary. If g E C[a, b] and J~g(x)v(x) dx = O.

V V E f00 = {v E C1[a, b]: veal = v(b) = O},
then g == 0 on [a, b]. D
This result admits generalization:

4. The Lemmas of Lagrange and du Bois-Reymond 99
(4.4) Lemma (Lagrange). If 9 E C[a, b] and for some m = 0, 1,2, ...
r g(x)v(x) dx = 0,
V V E!!)o = {v E Cm[a, b]: v(kl(a) = V(kl(b) = 0, k = 0, 1,2, ... , m},
then 9 == ° on [a, b]. (Here CO[a, b] == C[a, b].)

PROOF. Suppose g(c) > °
for some c E (a, b). Then from the hypothesized
continuity of g, c is contained in an interval [ex, p] £;; (a, b) in which
Ig(x) - g(c) I ~ g(c)/2 or g(x) ~ g(c)/2 > 0. On the other hand, the function
v(x) ~ {[(X - ex)(P - x)]m+l X E [ex, P],

0, x rt [ex, P],
is in cm(\R) (Why?), and nonnegative. It follows that on [a, b] the product
gv is continuous, nonnegative, and not identically zero. Thus from A.9,
° < J~ g(x)v(x) dx, contradicting the hypothesis.
° °
Similarly, the supposition that g(c) < or -g(c) > leads to a contradic-
tion and we conclude that g(c) = 0, VeE (a, b). But since 9 is continuous, it
must also vanish at the end points of the interval; i.e., 9 == on [a, b]. 0 °
Lemma 4.1 of du Bois-Reymond also generalizes but with more difficulty:
= 1,2, ...
r
(4.5) Proposition. If hE C[a, b] and for some m
h(x)v(ml(x) dx = 0, V V E !!)O,
where
!!)O = {v E Cm[a, b]: v(kl(a) = V(kl(b) = 0, k = 0, 1,2, ... , m - 1},
then on [a, b], h is a polynomial of degree < m.
PRooF*. By a translation we may assume that a = 0. The function
H(x)~ f: I'
dt 1 dt 2 ···Im-, h(t)dt, (1)
is in Cm[O, b] with derivative

H(ml(x) = h(x),
as is shown by repeated application of the fundamental theorem of calculus.
Moreover, H(O) = H'(O) = ... = H(m-ll(O) = 0, since each successive differen-
tiation eliminates one integral.
Similarly, if q is a polynomial of degree less than m, then P(x) ~ xmq(x)
vanishes at x = 0, together with pCil(x) for j < m, while p(x) ~ p(ml(x) is
another polynomial of degree less than m.
Let
v(x) = H(x) - P(x) (2)
so that
v(m)(x) = h(x) - p(x).
We must next show that with the proper choice of q we can make
°
v(k)(b) = for k = 0, 1,2, ... , m - 1, and this is possible. (See Problem 4.6*.)
Assuming that this choice has been made, the resulting v E Plo, and it follows
from repeated partial integrations that
J: p(x)v(m)(x) dx = - J: p'(x)v(m-1)(x) dx
= ... (-It J: p(m)(x)v(x) dx = 0,
since the boundary terms vanish. Thus, finally, from the hypothesis and
construction:
0::;:; f: (h(x) - p(x)f dx = J: (h(x) - p(x))v(m)(x) dx
= J: h(x)v(m)(x) dx = 0,
so that h(x) = p(x) on [0, b]. o

It is straightforward to obtain the vector valued analogue of 4.2:
(4.6) Proposition. If d = 2, 3, ... and for G, H E (C[a, bJ)d,
Lb [G(x)' V(x) + H(x)' V'(x)J dx = 0,

V V E Plo = {V E qy: V(a) = V(b) = ll!},
where qy = (C 1[a, bJ)d, then H E qy and H' = G (i.e., hi(x) = gj(x),j = 1,2, ... ,
d).
PROOF. If we restrict attention to those V = (v, 0, 0, 0, ... ,0) E qy, then the
integral condition resuces to
r [gl(X)V(X) + h1(x)v'(x)] dx = 0, V V E C 1[a, bJ with v(a) = v(b) = 0.
Hence from 4.2, it follows that

h1 E C 1[a, bJ and h~ = gl'
We can obviously apply the same technique to each component and con-
clude that hi = gj,j = 1,2, ... , m. 0
Problems 101
(4.7) Corollary. When HE (C[a, bJ)d and
r H(x)' V'(x) dx = 0,
as above, then H(x) = const. = C E [Rd.
PROOF. Set G == (i) in the proposition. o

A multidimensional version of Lemma 4.4 is presented in Problem 4.5.
PROBLEMS
4.1. Carry out the steps in the proof of Proposition 4.2 in the special case that
g(x) = sin x.
4.2. If hE C[a, b] and
f h(x)u(x) dx = 0, VUE!?Ao = {U E C[a, b]: Lb u(x) dx = O},

show that h(x) = const. on [a, bJ. Hint: Consider v(x) = J~ u(t) dt.
4.3. (a) Suppose that h E C1[a, b] and
f h(x)v'(x) dx = 0; V V E !?Ao = {v E C 1 [a, b]:

v(a) = v'(a) = v(b) = v'(b) = O}.
Use integration by parts and the proof of Lemma 4.4 to conclude that
h = const. on [a, bJ. (Do not use Lemma 4.1.)
(b) If 9 E C[a, b] and
f [g(x)v(x) + h(x)v'(x)] dx = 0,
conclude that h' = 9 on [a, b], without invoking 4.2.

4.4. Formulate and prove a vector valued analogue of the result in Problem 4.2.
4.5. (a) Prove the following multidimensional version of Corollary 4.3 for a
bounded domain D of ~d: If u E C(15) and SD UV dX = 0, V V E !?Ao = {v E
C 1 (15): VlaD = O}, then u == 0 in D. (Existence of the integrals with respect to
dX, the d-dimensional element of volume, is to be assumed.) Hint: For each
closed box B = {X E D: aj::::; xj ::::; bj ; j = 1, 2, ... , d}, consider functions
v(X) = V 1 (xdv 2 (X 2 ) ••• Vd(X d), where each Vj is a function of the type used in
the proof of Lemma 4.4 (for m = 0).
(b) Formulate and prove a corresponding extension of Lagrange's Lemma, 4.4.
4.6*. In proving Proposition 4.5 we needed to show that there is a polynomial q of
degree less than m such that P(x) = xmq(x) has at b #- 0, prescribed derivatives
P(k)(b) = Pk' k = 0, 1,2, .. , . (Indeed, we want Pk = H(k)(b) in (2).)
(a) Let qk = q(k)(b), k = 0, 1,2, ... , m - 1. By repeated differentiation show that

Po = bmqo,
Pl = bmql + mbm-1qo,
P2 = bmq2 + 2mb m- 1ql + m(m - 1)b m- 2qo,
etc.,
and that conversely these equations can be solved uniquely for the qk. Thus
the qk are known implicitly (and recursively).
(b) With the values of qk from part (a), show that
q
q(x) =
m-l
kf:O B(x - W
is a polynomial for which q(k)(b) = qk, k = 0, 1, ....
(c) Conclude that P(x) = xmq(x) will meet the required conditions.
4.7. Carry out the steps of the proof of Proposition 4.5 when m = 2.
CHAPTER 5
Local Extrema in Normed Linear Spaces
In ~d, it is possible to give conditions which are necessary in order that a

function I have a local extremal value on a subset D, expressed in terms
of the vanishing of its gradient VI (§O.5). In this chapter, we shall obtain
analogous variational conditions which are necessary to characterize local
extremal values of a function J on a subset!!) of a linear space qy supplied
with a norm which assigns a "length" to each y E qy.
In §5.1 we characterize norms and in the next two sections use them
to forge the analytical tools of convergence, compactness, and continuity,
presuming that the reader is familiar with these concepts in ~d. After some
obvious terminology concerning Oocal) extremal points (§5.4) we come to
the heart of the chapter: the observation that at such a point, the Gateaux
variation of a function must vanish in each admissible direction (§5.5). In
the next section, the norm is used to extend the development of differentia-
tion from the Gateaux variations of Chapter 2 to that of the derivative in the
sense of Frechet. In the concluding section, we introduce the method of
Lagrangian multipliers to characterize the local extremal points of one func-
tion when it is restricted (constrained) to the level sets of others.
§5.1. Norms for Linear Spaces
Analysis in ~d is described most easily through inequalities between the

lengths of its vectors. Similarly, in the real linear space qy, we shall assume
that we can assign to each y E qy a nonnegative number, denoted Ilyll, which
103
104 5. Local Extrema in Normed Linear Spaces
exhibits the following properties:
Ilyll ~ 0, v Y E!lJj with equality iff y = (1); (la)
IlcY11 = Icillyll, VeE IR, y E !lJj; (lb)
Ily + PII ::;; Ilyll + liP II, V y, P E!lJJ. (lc)
Thus II '11 is simply a real valued function on !lJj which by (la) is positive
definite, by (lb) is positive homogeneous, and by (lc) satisfies the triangle
inequality. Each function with these properties is called a norm for CW. There
may be more than one norm for a linear space, although in a specific exam-
ple, one may be more natural or more useful than another. Every norm also
satisfies the so-called reverse triangle inequality
Illyll - liP I I ::;; Ily - PII, (ld)

(Problem 5.1.)
Example 1. For CW = IRd with y = (YI' Y2, ..• , Yd), the choice Ilyll = Iyl =
(I,1=1 yJ)I/2 defines a norm, called the Euclidean norm, but the verification of
the triangle inequality (lc) is not trivial (Problem 5.2).
The choice Ilylll = I,1=llyj l also defines a norm and now all of the
properties are easily verified. In particular, for (lc):
d d
Ily + Pill = I, IYj + Pj ! ::;; I, (IYj! + IPjl),
j=l j=l
or
d d
lIy + Pill::;; I, Iyjl + I, IPjl = lIylll + IIPIlI'
j=l j=l
Still another norm is the maximum norm lIyllM = maxj=I.2 ..... d Iyjl. However,
the simpler choice, lIyll = IYII does not yield a norm for IRd if d ~ 2 because it
is not positive definite. Indeed, the nonzero vector (0, 1, 1, ... , 1) would have
a zero "norm" with this assignment.
Example 2. For !lJj = C[a, b], it is useful to think of the values y(x) as the
"components" of the "vector" y E CW. Then the choice lIyllM = max !y(x)1 =
maxXE[a.b1Iy(x)1 determines a norm, the so-called maximum norm. That it is
even defined and finite is not obvious and it requires the knowledge that:
(i) y E C[a, b] => Iyl E C[a, b];
(ii) [a, b] is compact;
(iii) a continuous function (Iyl) on a compact interval ([a, b]) assumes a
maximum value (llyIlM)'
(i) is a consequence of the (reverse) triangle inequality in IRI, while (ii) is
established in §A.O; (iii) will be established in Proposition 5.3.
§S.l. Norms for Linear Spaces lOS
Accepting its definition, it is straightforward to verify that II' 11M is

(a) positive definite: Since 0 ~ ly(x)1 ~ IlyliM, then
IlyliM = 0 = ly(x)1 = 0 or y(x) = 0, 'v'xE[a,b];
i.e., y = @;
(b) positive homogeneous:
IleyllM = max Iey(x) I = max lelly(x)1 = lei max ly(x)1 = 1c111y1iM;
and
(c) satisfies the triangle inequality:
I(y + ji)(x) I = Iy(x) + ji(x) I ~ ly(x)1 + lji(x) I
~ IlyliM + IljilIM' xE[a,b];
thus Ily + jiliM = max I(y + ji)(x) I ~ IlyliM + IljilIM'

Another choice of norm for C[a, b] is Ilylll = gly(x)1 dx. (See Problem
5.3.) Observe that Ilylll ~ (b - a) IlylIM' Are there other norms for this space?
Example 3. For C!Y = Cl[a, b], analogous norms are defined by

IlyliM = max (ly(x)1 + Iy'(x)l)
r
xE[a,bj
and
Ilylll = (ly(x)1 + ly'(x)l) dx.
To establish the triangle inequality for II' 11M, observe that for x E [a, b]:
I(y + ji)(x)1 + I(y + ji)'(x) I ~ ly(x)1 + Iji(x) I + ly'(x)1 + Iji'(x)1
= (ly(x)1 + ly'(x)l) + (lji(x)1 + Iji'(x)l)
~ IlyliM + IljilIM'
Now maximizing over x yields the inequality
Ily + jiliM = max{ly + ji)(x) I + I(y + ji),(x)l} ~ IlyliM + IljilIM,
as desired.
Observe that for x E [a, b]: ly(x)1 ~ IlylIM' Thus IlyliM = 0 = y = @; i.e.,
II' 11M is positive definite, as is 11'111' since Ilylll ~ (b - a) IlylIM' Can you devise
corresponding norms for C 2 [a, b]? For Cm[a, b]?
(5.0) Remark. Since for each m = 2, 3, ... , Cm[a, b] ~ Cl[a, b] ~ C[a, b],
then each norm for C[a, b] from the previous example will serve also as a
norm for Cl[a, b] (or for Cm[a, b]). However, these norms do not take
cognizance of the differential properties of the functions and supply control
only over their continuity.
Example 4. If 11'11 is a norm for the linear space iJJj, then Ilyll + It I is a norm
for the linear space iJJj x IR consisting of the pairs (y, t) with componentwise
addition and scalar multiplication. (See Problem 5.8.)
Example 5. iJJj = C(a, b] is a linear space which has no obvious choice for
a norm. Now, max ly(x)1 can easily be infinite as can J~ ly(x)1 dx; the function
y(x) = l/(x - a), x #- a is in iJJj and realizes both of these possibilities. In
fact, this space cannot be normed usefully.
Example 6. iJJj = (C[a, b])d, the space of d-dimensional vector functions with
components in C[a, b], is a linear space with a norm for Y = (Yl' Y2, ... , Yd)
(i.e., Y(x) = (Yl(X), Y2(x), ... , yix)), x E [a, b]), given by
IIYIIM = maxi Y(x) I,
or by
d
II YII = L maxIYj(x)l,
j=l
or by
d fb
II YII = jf:l a IYj(x) I dx,
or by
II Ylll = rCtl
The verification is left to Problem 5.6
I
Yf(X)Y 2 dx = r IY(x) I dx
Example 7. iJJj = (C l [a, b])d, the space of d-dimensional vector functions with
components in C l [a, b], is a linear space with a norm
I YIIM = max(1 Y(x) I + IY'(x)l),
or
d
II YII = L max(IYj(x)1 + lyj(x)I),
j=l
Can you think of other norms for this space?
As these examples show, the discovery of a suitable norm for a given
linear space is not immediate, is seldom trivial, and may not be possible.
Fortunately, the spaces of interest to us in this text do have standard norms,
and these have been given in the examples above.
(Problems 5.1-5.8)
§5.2. Normed Linear Spaces: Convergence

and Compactness
When a norm II' II for a real linear space iJJj has been assigned, it is straight-
forward to define an associated (topological) structure on iJJj which will permit
analysis in the normed linear space (iJJj, 11'11).
§5.2. Normed Linear Spaces: Convergence and Compactness 107
First, we define the "distance" between vectors Y and Y by Ily - YII. The
triangle inequality shows that for any three vectors x, y, Z E qy: Ilx - zil ~
Ilx - yll + Ily - zll and this has the familiar geometrical interpretation.
Next, we introduce the concept of convergence by declaring that if Yn E
I!Y, n = 1, 2, ... , then the sequence {Yn}f has the limit Yo E qy (denoted
limn-+oo Yn = Yo, or Yn -+ Yo as n -+ (0) iff llYn - Yoll -+ 0 as n -+ 00. A given
sequence need not have a limit, but if it does its limit is unique.
[Indeed, were Yo and Yo both limits of the same sequence {Yn} from I!Y, then
bythetriangleinequalityforeachn,O ~ Ilyo - Yoll ~ Ilyo - Ynll + llYn - Yoll·
The right side can be made as small as we wish by choosing n sufficiently
large; thus Ilyo - Yo II = 0; but by the positive definiteness of the norm, this
means that Yo - Yo = (!J or Yo = Yo·J
Alternatively, we can introduce the (open) spherical b-neighborhood of Yo
for each b > 0 defined by Sd(YO) = {y E qy: Ily - Yoll < b} and note that for a
sequence {Yn}: limn-+oo Yn = Yo iff the Yn are eventually in each Sd(YO); i.e., for
each b > 0, 3 Nd such that n ~ Nd = Yn E Sd(YO)·
Using these concepts, we can identify the compact subsets of qy as those

sets K with the property that each sequence {Yn} from K contains a conver-
gent subsequence {Yn.} J
with limit Yo E K; i.e., for integers n 1 < n2 < ... <
nj < nj + 1 < ... , limj-+oo YnJ = Yo E K. (Equivalently, K has the property that
each covering of K by a collection of open spheres can be accomplished by
using only a finite number of these spheres. See [F1].)
For qy = C[a, bJ, with the maximum norm (Example 2 of §5.1), the
b-sphere of a function Yo E qy is the set of functions Y whose graphs are
uniformly within b of the graph of Yo. (See Figure 5.1.) The associated conver-
gence of a sequence is accordingly referred to as uniform convergence.
Now although C[a, bJ is a linear space,
g) = {y E C[a, bJ: y(a) = 0, y(b) = I}
is not, since the sum of two of its functions cannot satisfy the boundary
8f
·r
s: -- ./ ./ __ -
'-. '- Y
--..-- - '\.
u - __ -
~----~----------------------~--.
b x
Figure 5.1
---~/ '" YE~
b x
Figure 5.2
condition at b. fl) is simply a subset of qy = C[a, b], and as such automati-

cally inherits the norm-distance properties from iJjf. Notice also that with the
maximum norm, fl) does not contain a full b-neighborhood of any of its
elements. (See Figure 5.2.) fl) is not compact, for reasons that are deferred
until the discussion of continuity in the next section.
For qy = [Rd with the standard Euclidean norm (Example 1 of §5.1), the
b-spheres are the open d-dimensional spheres which in [R3 are ordinary
spheres, in [R2, disks, and in [Rl, open intervals. With the maximum norm
IlyliM = maxj =1.2 ..... d Iyjl, the b-spheres are d-dimensional "cubes." With
either norm, the convergence of a sequence of vectors from [Rd has the
same geometrical interpretation, and the compact subsets are precisely those
which are bounded (contained inside some b-sphere) and closed (contain the
limits of each of their convergent sequences). In particular, each "box" of the
form B = {y E [Rd: aj :$; Yj :$; bj;j = 1, 2, ... , d} is compact (§A.O).
(Problems 5.9-5.12)
§5.3. Continuity
If fl) is a subset of qy, then we can consider it as the domain of various kinds
of functions. For example, a qy-valued function :F: fl) ..... qy could be defined
by requiring :F(y) = @, V Y E fl). For our purposes, the most important
functions are those which are real valued, i.e., those of the form J: fl) ..... [R, of
which we have already encountered many examples in the previous chapters.
When qy is supplied with a norm 11'11, we simply adopt the standard 6 - b
definition for the continuity of a real valued function. (See §0.3.)
(5.1) Definition. In a normed linear space (qy, 11'11), if fl) ~ I!Y, a function
J: fl) ..... [R is said to be continuous at Yo E fl) iff for each 6 > 0, 3 a b > 0 such
that
IJ(y) - J(Yo)1 < 6, V Y E fl) with Ily - Yoll < b.
§5.3. Continuity 109
We observe that the definition of continuity of J at Yo is a mathematical

statement of the intuitive requirement that the smallness of Ily - Yo II control
that of IJ(y) - J(Yo)l.
Equivalently, J is continuous at Yo E fl) iff for each sequence {Yn} from
fl), lim n _ oo Yn = Yo => limn _ oo J(Yn) = J(yo)· The proof of this fact is left to
Problem 5.13.
We also say that J is continuous on fl) iff J is continuous at each point
Yo E fl). Observe that if J: i5JI ..... IR is continuous, and fl) ~ i5J1, then the restric-
tion JI!'j: fl) ..... IR is also continuous with respect to the same norm.
Example 1. In any normed linear space (i5JI, 11·11), the norm function J(y) =
Ilyll is always continuous on i5JI and hence on any subset fl) ofi5Jl. Indeed, from
the reverse triangle inequality (ld), IJ(y) - J(Yo)1 = Illyll - IIYoll1 ~ Ily - Yoll,
and hence making Ily - Yoll small « <5) makes IJ(y) - J(Yo)1 at least as
small. (In fact, II· I is uniformly continuous on i5J1. See Lemma 5.2 below.)
Example 2. For i5JI = C[a, b] with the maximum norm II· 11M of§5.1 Example
2, the function
J(y) =f [si n 3 x + y2(X)] dx
is defined for each y E i5J1. To establish its continuity at Yo we must estimate
IJ(y) - J(Yo)1 =If [Y(X)2 - YO(X)2] dxl
~f Iy(xf - Yo (X)2 I dx
=f Iy(x) - yo(x)lly(x) + yo(x)1 dx
~ Ily - YoIIMIly + YollM f dx,

or
IJ(y) - J(Yo)1 ~ (b - a)lly + YoIIMIly - YoIIM·
Now, if Ily - YollM < 1, then IlyIIM < 1 + IIYollM (Why?)
and so
Ily + YollM ~ IlyIIM + IIYollM < 1 + 211Yo11M· Thus when Ily - YollM < 1:
IJ(y) - J(Yo)1 < (b - a)(l + 211Yo11M)lly - YollM

= AollY - YoIIM'
say, for the constant Ao = (b - a)(l + 21IYoIIM).
This last estimate shows that for each e > 0, we can make IJ(y) - J(Yo)1 < e
provided that we take Ily - Yo II < <5 ~ min(l, elAo). Hence J is continuous
at each Yo E i5J1, and so J is continuous on i5J1.
Example 3. With a given oc C[a, b], the function of §2.2, Example 2; viz.,
E
g
J(y) = oc(x)J1 + Y'(xf dx is defined V Y E qy = C 1 [a, b]. Direct examina-
tion of its continuity with respect to the maximum norm I 11M is facilitated
by the following uniform estimate for f(z) = Jl"+?:
If(z) - f(zo)1 ::;; Iz - zol, z, Zo E R
[This is an immediate consequence of the mean value theorem and the fact
that
1f'(z)1 = p::;; 1+z
1, Z E R.]
Thus for y, Yo E qy we have V x E [a, b] the uniform estimate
If(y'(x)) - f(yo(x)) I ::;; IY'(x) - Yo(x)1 ::;; Ily - YoIIM'

Hence
IJ(y) - J(Yo)1 ::;; r loc(x)llf(y'(x)) - f(Yo(x))1 dx
::;; Ily - YollM r Ioc(x) I dx = AllY - YoIIM'
say, and the (uniform) continuity of Jon qy should now be evident.
To obtain uniform estimates of the type used in this last example, we shall
make frequent appeal to the following technical
(5.2) Lemma. If K is a compact set in a normed linear space (qy, 11'11), then a
continuous function F: K -+ R is uniformly continuous on K; i.e., given E > 0,
3 () > 0 such that y, Y E K and Ily - yll < () = IF(y) - F(Y)I < E.
PROOF. We shall establish the contrapositive implication. Suppose the
lemma does not hold. Then, for some Eo > 0, and each n = 1, 2, ... , 3 points
Yn' Yn E K with llYn - Ynll < lin for which IF(Yn) - F(Yn)1 :::::: Eo' However,
since K is compact, there is a subsequence Yn. -+ Yo E K asj -+ 00, and since for
}
each j = 1, 2, ... :
1
IIYnj - Yoll ::;; IIYnj - Yoll + IIYnj - Yn)1 ::;; IIYnj - Yoll + n.'
J
it follows that Yn. }

-+ Yo asj -+ 00. But, Vj:
0< Eo ::;; IF(Yn) - F(Yn) I ::;; IF(Yn) - F(Yo)1 + IF(Yn) - F(Yo)l,

and hence F cannot be continuous at Yo since continuity would demand that
both terms on the right -+ 0 as j -+ 00. Thus we have shown that if F is not
uniformly continuous on K, then there is at least one point Yo at which it is
not continuous. D
Example 4. When f
r r
E C( [a, b] x 1R2), the function
F(y) = f[y(x)] dx = f(x, y(x), y'(x)) dx
is defined V y E qy = C l [a, b]. To establish its continuity with respect to

the maximum norm II· 11M of Example 3 in §5.1, we can use Lemma 5.2 as
follows:
f is uniformly continuous on each compact box [a, b] x [ -c, C]2 when
c > O. Thus for a fixed Yo E qy, when y E S1 (Yo) we have V x E [a, b] that both
ly(x)l, ly'(x)1 :$; ly(x)1 + 1y'(x)1 :$; IlyliM < 1 + IIYollM = co, say.
Then with c = co, it follows from the aforementioned uniform continuity
that given E:: > 0, 3 a () E (0,1) such that Ily - YollM < ()( < 1) =>
If[y(x)] - f[Yo(x)]l = If(x, y(x), y'(x)) - f(x, Yo(x), yo(x))1 < E::,
VXE[a,b].
r
This uniform estimate gives
IF(y) - F(Yo)l:$; If[y(x)] - f[Yo(x)]l dx :$; E::(b - a),

when
Ily - Yoll < () < 1,
and the continuity of F at the arbitrary point Yo E qy is established. Observe
that () depends on Co (and so on Yo) and will in general decrease as Co
increases.
Example 5. For qy = C[a, b] with the maximum norm IlyliM = maxly(x)1 of

§5.1, Example 2, the evaluation function L(y) = y(a) is defined V y E qy and L
is even linear in that for y, y E qy and c, C E IR, we have
L(cy + cy) = cL(y) + cL(y),
(since (cy + cy)(a) = cy(a) + c.Y(a)). Observe that L((!J) = O.
L is also continuous: For a linear function L it suffices to establish conti-
nuity at (!J since
L(y) - L(yo) = L(y - Yo),
so that IL(y) - L(Yo)1 is controllably small as v = y - Yo ..... (!J iff L(v) ..... 0 =
L((!J) as v ..... (!J. Here, clearly IL(v)1 = Iv(a)1 :$; Ilvil M so that v ..... (!J => L(v) ..... O.
However, if instead we use the integral norm IIyl11 = f~ ly(x)1 dx on this
same space, as in §5.1, Example 2, then this same function L(y) = y(a) remains
linear, but it is not continuous anywhere. From our observation, it suffices
to eliminate continuity at (!J, and for this it suffices to exhibit a sequence
Yn E C[a, b] for which Ilynlll ..... 0 as n ..... 00, but L(Yn) = Yn(a) ~ 1. The
functions Yn(x) shown in Figure 5.3 have this property, since geometrically
a a + lin b x
Figure 5.3
for n > (b - ar l,
IIYnlll = G) G)
(1) = 2~ --.0 as n--. 00,
while Yn(a) = 1, V n.
We are of course interested in characterizing the maximum and minimum

values of such real valued functions J and the points at which they occur.
Unfortunately, as the simple example y(x) = x, X E ( -1, 1) !:; ~, shows, a
continuous function on a subset of a linear space may have neither a maxi-
mum value nor a minimum value on this set-unless the set is compact.
(5.3) Proposition. A continuous real valued function J on a compact subset K

of a normed linear space (@', 11·11) assumes both maximum and minimum values
at points in K. In particular, these values are finite.
PROOF. To establish that J assumes a maximum value, observe that we
can at least find a sequence {yn} from K for which the corresponding se-
quence {J(Yn)} of real numbers increases to this "maximum" (which may
be infinite). However, since K is compact, we can extract from this sequence
a convergent subsequence {Yn.} J
with limit Yo in K. From the continuity of
J it follows that J(yo) = limj-+oo J(YnJ But J(Yn) must have the same limit as
the original sequence J(Yn). (Why?) Hence this common limit is J(yo), i.e., the
maximum value is assumed. The proof for the assumption of the minimum
value is similar. 0
A consequence already utilized in §5.1, Example 2, is that every real valued

continuous function on the compact interval [a,b] is bounded and it assumes
a maximum value.
However, as attractive as this solution to the problem of establishing

the existence of maxima and minima may appear, it will be of little help
to us because most of the sets of interest to us are too "large" to be compact.
One application of the proposition will be in forestalling attempts to estab-
lish compactness!
a x
Figure 5.4
For example, CEO, 1] is not compact with the maximum norm since the
continuous function J(y) = y(l) is unbounded on CEO, 1]. To see this, we
consider the functions Yn(x) = nx, x E [0, 1] for which J(Yn) = n ..... +00.
More generally, in a nontrivial normed linear space (<w, 11'11), !5Jj itself is
never compact (Problem 5.17). In particular, there is no clever assignment
of a norm to C[a, b] that makes this space compact.
Similarly, q) = {y E C l [a, b]: y(a) = 0, y(b) = 1}, which is one of the sets
of concern in Chapter 2, is not compact with respect to the maximum norm
II' 11M or the integral norm 11'111 of §5.1, Example 3. For as we have seen in
Example 1, the norm function J(y) = Ilyll is always continuous on !5Jj and
hence on any subset q) of OJ/. But JM(y) = IlyliM is unbounded on the set q)
above, as is shown by the sequence of parabolic functions sketched in Figure
5.4, which have maximum values as large as desired while remaining in q).
Clearly IlyliM ~ max ly(x)1 and so JM(y) will be unbounded on this sequence.
It is also plausible graphically that the integral norm function Jl (y) = Ilylll
(~ n ly(x)1 dx) is unbounded on this sequence, since the area under the curve
can be made as large as we wish. In particular, the apparently reasonable
problem of finding the maximum of the function J(y) = n ly(x)1 dx on q) has
no solution. (In fact, the problem of minimizing J on q) has no solution. See
Problem 5.19.)
Thus, the results of Chapter 3 in which we actually obtained minimum
values for rather complicated functions on sets such as q) become even more
remarkable. (As we saw, an underlying convexity was responsible for our
success in these cases.)
We shall proceed with the theoretically unattractive task of seeking max-

ima and minima of functions which need not have them, as the above exam-
ple shows. Note that to maximize J it suffices to minimize - J, and when the
underlying set q) is not compact it is probably of little value to establish the
continuity of a function J which we wish to minimize on !!J. (However,

continuity of related functions will be required in §5.6 and §5.7.)
Thus in general, neither continuity nor compactness alone can provide
us with useful information. As suggested by the analogous situation in ~d,
we must consider necessary conditions expressed in terms of an appropri-
ately defined concept of differentiation. We introduce first some obvious
terminology.
§5.4. (Local) Extremal Points

When qIj is a linear space and J is a real valued function on !!J !:; ifY, then a
point Yo E !!J at which J assumes its maximum value or its minimum value
is called an extremal point for J on !!J. This assignment is global in nature
and may be made without consideration of a norm. However, the presence
of a norm permits an analogous description of the local behavior of J at a
point Yo:
(5.4) Definition. In a normed linear space (ifY, 11·11) a point Yo E !!J !:; qy is said
to be a local extremal point for J on !!J if for some r > 0, Yo is an extremal
point for Jon !!Jr(YO), where !!Jr(yo) = {y E!!J: Ily - Yoll < r}; i.e., either
J(y) ::;; J(yo), V Y E !!Jr(yo) (Yo is a local maximum point for J on !!J), or
J(y) ~ J(yo), V y E !!Jr(yo) (Yo is a local minimum point for J on !!J).
Of course, each extremal point is automatically a local extremal point
whatever norm is used. However, Yo may be a local extremal point with
respect to one norm but not with respect to another. (See Problem 5.20)
Now, the Gateaux variations of Chapter 2 may also be formed without
consideration of a norm and when nonvanishing, they preclude local extremal
behavior with respect to any norm.
Suppose, for example, that at a point Yo, the function J has a positive
variation c5J(yo; v) in the direction v E qIj. From Definition 2.6, it follows
that V 6 sufficiently small, the ratio [J(yo + 6V) - J(Yo)]/6 is also positive,
so that J(yo + w) - J(yo) has the sign of 6. Hence,
V small 6 > 0, (2)
and we say that at Yo, J increases strictly in the direction v (and decreases
strictly in the opposite direction - v). When c5J(yo; v) < 0, then c5J(yo; - v) > °
(Why?) so that the preceding inequalities and assertions are reversed. In
either case, since as 6"" 0, II(yo ± w) - Yoll = 611vll "" 0, the points Yo ± 6V
in (2) are eventually in each norm neighbourhood of Yo. Thus local extremal
behavior of J at Yo is not possible in the direction v.
§5.5. Necessary Conditions: Admissible Direction 115
The function J of §2.4, Example 2, has at yo(x) = x 2 , the variation
c5J(yo; v) = 2 Lb yo(x)v(x) dx, V V E I1,Y = C[a, b],

which is clearly nonvanishing in the direction v(x) = eX. Thus, in this direc-
tion, J cannot exhibit local extremal behavior at Yo, and as a consequence, Yo
cannot be a local extremal point for J on I1,Y regardless of the norm employed
for !fIJ.
§S.S. Necessary Conditions: Admissible Direction

In minimizing a real valued function J over !!} ~ iJ,Y, where (iJ,Y, 11·11) is a
normed linear space, it is natural to consider for each y E !!} those directions
v E I1,Y in which the restricted function JI~ admits variation at y; i.e., we wish
to distinguish those directions v E !fIJ for which:
(i) y + t:v E !!}, V sufficiently small 6; and
(ii) c5J(y; v) exists.
Such directions will be termed admissible at y for !!}, or !!}-admissible at y (for
J). Observe that if v is !!}-admissible at y, then so is each scalar multiple cv for
c E !R; (!) is always admissible.
To see their significance, suppose that Yo is a (local) minimum point for

J on !!}. Then J cannot decrease strictly in any direction v admissible at Yo
for !!}, and hence from (2) (et seq.) c5J(yo; v) = 0 in such distinguished direc-
tions. The same result is obtained when Yo is a (local) maximum point on !!}.
Thus we have established the following:
(5.5) Proposition. In a normed linear space (!fIJ, 11·11), if Yo E !!} ~ !fIJ is a (local)
extremal point for a real valued function J on !!}, then
c5J(yo; v) = 0, V directions v which are !!}-admissible at Yo. 0
Our hope is that there will be "enough" admissible directions so that

the condition c5J(yo; v) = 0 can determine Yo. Observe, though, that with
this condition alone we cannot distinguish between a local maximum and
a local minimum point-or between a local minimum point and a global
minimum point. Moreover, as in !R d, we must admit the possibility of station-
ary points (such as saddle points) which satisfy this condition but may be
neither local maximum points nor local minimum points.
Clearly then, in many senses this condition is necessary but not sufficient
for, say, a (local) minimum point. However, its analysis forms a large part
of the classical calculus of variations. We shall give a geometrical interpreta-
tion for it in the next section and return to its classical treatment in the
next chapter.
(Note that Proposition 5.5 admits a local extremal point Yo E !!) which has
no nonzero !!)-admissible directions, and such points must also be considered
as candidates for local extrema.)
r
Example 1. To characterize local extrema for the function (of §2.4, Example 2)
J(y) = [sin 3 x + Y(X)2] dx

on the domain
!!) = {y E C[a, b]: y(a) = a 1 , y(b) = bd,
(where a 1 , b1 E IR are given), we know that c5J(y; v) is defined V y, v E C[a, b].
However, the only !!)-admissible directions at y E!!) are those for which
y + ev E !!) for sufficiently small e # O. Thus we require that
a1 = (y + ev)(a) = y(a) + ev(a) = a 1 + ev(a),
b1 = (y + ev)(b) = y(b) + ev(b) = b1 + ev(b),
so that v(a) = v(b) = O. Hence v is !!)-admissible at y E !!) iff
v E!!)o = {v E C[a, b]: v(a) = v(b) = O}.
Here the class !!)o is not y-dependent, and by Proposition 5.5 the condition
r
necessary for a (local) extremal point Yo E !!) is that
c5J(Yo; v) = 2 yo(x)v(x) dx = 0, V V E !!)o·
This condition is surely fulfilled when Yo == 0, but this function is not in !!)
(unless a 1 = b1 = 0). And as Lemma 4.4 of Lagrange shows there is no other
possibility; i.e., when a 1 # 0 or b1 # 0 there are "too many" admissible direc-
tions to permit any function Yo E!!) to satisfy all of the conditions c5J(yo; v) =
onecessary for a (local) extremum. Thus no such local extremum exists.
If on the other hand, we attempt to minimize J over
!!)* = {y E C[O, 1]: Inl xY(X)4/3 dx = 1},

then it is difficult to characterize the !!)*-admissible directions. (This problem
will be considered again in Example 1 of §5.7.)
Example 2*. Let's characterize local minima for the function J of §2.4,
Example 3,
J(y) = r oc(x)J1
for a given oc E C[a, b], on the domain

+ y'(X)2 dx,
!!) = {y E C 1 [a, b]: y(a) = a 1 , y(b) = bd,

with given a 1 , b1 E ~. We again know that t5J(y; v) is defined V y, v E I1,Y =

C 1 [a, b], and as in the previous example conclude that v is ~-admissible
(at y) iff
V E ~o = {v E C1[a, b]: v(a) = v(b) = O}.
Hence from Proposition 5.5, the necessary condition that y E ~ be a

(local) extremum is that
c5J(y; v) = fb ex(x)y'(x)v'(x) dx = 0,
a J1 + y'(xf
Thus from Lemma 4.1 of du Bois-Reymond, we know that this necessary
condition is satisfied only by ayE ~ for which the continuous function
ex(x)y'(x) = e = const., (3)

J1 + y'(X)2
or after obvious algebra, for which
2
,2 e
y = ex 2 - e2· (4)
We shall suppose that ex does not vanish identically on any subinterval

of [a, b]. (See Problem 5.21.) If ex vanishes at a single point, then from (3),
e = 0 and so y = const. which would require y(a) = a 1 = y(b) = b1. Thus
unless a1 = b1, solution to (3) is possible only when ex 2 (x) > e2 > 0, and
for a continuous ex this requires that either ex(x) > lei> 0 on [a, b] or
ex(x) < -lei < 0 on [a, b]. It suffices to consider the first alternative where
ex > 0 since the second is reduced to this case with the replacement of J by
-J. Now, when ex > 0 on [a, b], the integrand function is strongly convex
(§3.3, Example 4) and we know from Theorem 3.7 that each y E ~ which
satisfies (3) must supply uniquely the minimum value for J on ~. We see
that (3) is both necessary and sufficient for a minimum; that there can be
at most one such minimum point; and that there are no other (local) extremal
points for J on ~. However, as yet, we have no assurance that there is ayE ~
which satisfies (3).
Supposing that a1 < b1, and that ex(x) ~ exo > 0 on [a, b], we require from
(4) that
y'(x) = [(ex(x)/e)2 - 1]-1 /2,
or upon incorporating the requirement y(a) = a 1, that
y(x) = a 1 + f: [(ex(t)/ef - 1rl/2 dt, X E [a, b].
It remains to show that e E (0, exo) can be chosen to satisfy the other
h(e)~ r
boundary condition y(b) = b1; i.e., to make
[(ex(x)/e)2 - 1rl/2 dx = b 1 - a1 · (5)

When Q:(x) = const. = Q: o on [a, b], then h(e) = [(Q:o/e)l - 1]-1/1(b - a), is
continuous and strictly increasing on (0, Q:o) with the limiting values °
(= lime +0 h(e» and +00 (= limetao h(e». Hence in this case, there is precisely
one e for which (5) is satisfied and precisely one y E !!) which satisfies (3).
When Q: is not constant, it is more difficult to analyze h(e). (See Problem
5.32*.) However, it is important to realize that there may be no solution.
For example, when a = 1, b = 2, and Q:(x) = x on [1, 2], then we may take
Q: o = 1, and for e E (0, 1):
Q:2(X) Xl
-e 2- - 1 = -e 2 - 1 >
-
X2 - l',
hence
h(e) = f l ( 2()
1 elX - 1
Q:
)-1/2dx::;; f2 (X2 - 1 1)-1 /2 dx
: ; I2 (X2 - I t 1/2 x dx = (Xl - 1)1 / 11: = J3.

Thus when b1 - al > J3,
there is no y E!!) which satisfies the necessary
condition M(y, v) = 0, V !!)-admissible directions v E !!)o; in this case, there
are no local extremal points for J on !!).
Example 3. For I1JJ = 1R1 with the standard Euclidean norm 1'1, and
!!) = {y E 11JJ: Iyl = I},
there are no !!)-admissible directions v#-(() at any y E!!) for any function
J, because if y E!!) and (() #- v E 11JJ, then y + ev ¢ !!) except at most for one
value of e, as the simple geometry of Figure 5.5(a) shows.
On the other hand, at each point in the square !!)1 = {y E 11JJ: IlyliM = I}
of Figure 5.5(b) there is always one possible nonzero !!)-admissible direction,
except at the corner points where again there are none.
(a) (b)
Figure 5.5
Example 4*. The brachistochrone function of §1.2(a), viz.,

1
T(y)=-
IX. Jl + y'(X)2 dx
fig 0 jYW
is not defined V y E I1jj = Cl [0, Xl] because of the denominator jYW. It is
defined on
{f)* = {y E 11jj: y ~ 0 and IX. Y(Xfl /2 dx < +oo},
which contains, for example, all of those functions y E I1jj for which y(x) ~
m > 0 on [0, Xl].
However, the only functions which could be considered as solutions for
the brachistochrone problem are those which satisfy the boundary condi-
tions y(O) = 0 and y(xd = Yl for some given Yl > O. Among such functions
y(x) = ax 2 for a> 0, on [0, Xl] is not in {f) = {y E {f)*: y(O) = 0, y(x l ) = yd,
since
f:' a- 1/2 x- l dx = +00.
And even for one such as y(x) = X 3/2 , which is in {f), it may be that only a
restricted class of directions v E I1jj can be {f)-admissible at y, since in order
for T(y + ev) to be defined we must have y + ev ~ 0 for small e both positive
and negative. For example, v(x) = sin(nx/x l ) (which does meet the boundary
conditions required to be {f)-admissible) is not an admissible direction for this
y since for each e > 0, x 3/2 - e sin(nx/x l ) is negative when X is sufficiently
near (but not eqmil to) O. (Why?) Consideration of admissible directions is
essential to this problem and cannot be avoided.
If y E {f), then each v E I1jj with v(O) = v(x l ) = 0 and Iv(x)1 :::;; y(x), V X E
[0, Xl] is a possible {f)-admissible direction at y (as is, of course, any scalar
multiple of such a direction). Indeed for such v and 1el :::;; t:
y(x) + ev(x) ~ y(x) - lellv(x)1 ~ y(x) - h(x) = h(x) ~ 0
and
IoXl (y + eVf l /2(x) dx :::;;.j2 IX.0 Y(Xfl /2 dx < +00.
Finally, formally differentiating T(y + ev) with respect to e under the integral
sign as in §2.4, Example 4, with
f(x, y, z) = 1
M: /¥-+Z2--,
....;2g Y
so that
-1~
fy(x, y, z) = M: 3/2 '
2....;2g y
and
z
fAx, y, z) = M: r:. ~'
....;2g....;y....;1 +z
we have
. 1 IXI [ 1 J1 + y'(X)2 y'(x), ]

c5T(y,v)= M:
v 2g 0
--2 ()3/2
Yx
v(x)+ r.:t::\J
V y(x) 1 + y'(xf
v (x) dx,
(6)
which is finite in general only if Jol y(xf3/2 dx is finite.

Thus we see that if
!?)l = {y E OJ!: y(O) = 0, y(xd = YI' y(x);;::: 0 on [0, Xl],
and f: 1
y(xf3/2 dx < +oo},
then for each y E!?)t> c5T(y; v) is defined by (6) V directions v E OJ! at y for
which v(O) = v(x l ) = 0 and Iv(x)1 :::;; y(x) (or scalar multiples of these direc-
tions). We shall return to this problem in §6.2(c), Example 4.
In summary, as these examples show, there may be "too many" nontrivial
admissible directions v to allow any Yo E !?) to fulfill the necessary condition
c5J(Yo; v) = 0; or there may be just enough to permit this condition to deter-
mine Yo, or there may be many, but not readily usable-or there may be
just one-or even none. Nevertheless, when present, they provide the most
obvious approach to attacking problems in optimization, and should always
be considered before investigating alternatives such as the method of
Lagrangian multipliers to be introduced in §5.7. Finally, as with the
brachistochrone function of Example 4*, they may be essential to the
problem.
§5.6*. Affine Approximation: The Frechet Derivative

As we have seen, the Giheaux variation in a normed linear space is analogous
to the directional derivative in ~d. In particular, without further information,
we cannot expect to use these variations to provide a good approximation
to a function which has them-except, of course. in each separate direction.
For this purpose in ~d, we required that the function satisfy the stronger
requirement of differentiability (§0.10), and we shall simply lift the definition
employed there, together with the associated terminology, to our normed
linear space (<w, 11·11).
In ~d with the Euclidean norm 1·1, a real valued function f is said to be
differentiable at Yo E ~d provided that it is defined in a sphere S(Yo) and there
f(y) = f(yo) + l(y - Yo) + Iy - YoI3(Y - Yo),
where 3(y - Yo) is a function with zero limit as y - Yo -+ (9, and I is the
§5.6*. Affine Approximation: The Frechet Derivative 121
continuous linear function defined on IRd by l(v) = Vf(yo)' v. See Problem

5.15.
Now, for (qy, 11'11) a linear function L: qy --+ IR need not be continuous
(§5.3, Example 5), and we must require this continuity. Accordingly we make
the following:
(5.6) Definition. In a normed linear space (<w, 11'11), a real valued function
J is said to be differentiable (in the sense of Frechet) at Yo E qy provided
that J is defined in a sphere S(yo) and there exists a continuous linear
function L: qy --+ IR for which
J(y) = J(yo) + L(y - Yo) + Ily - Yo113(Y - Yo), (7)
where 3(y - Yo) is a real valued function (defined when y - Yo #- (!J by this
equation) which has zero limit as y - Yo --+ (!J or as Ily - Yoll --+ o.
(5.7) Proposition. If J is (Frckhet) differentiable at Yo, then J has the Gdteaux

variation c5J(yo; v) = L(v) in each direction v E qy.
PROOF. For v E qy and e#-O sufficiently small, set y = Yo + ev in (7). Then
J(yo + ev) - J(yo) = L(ev) + IlevI13(ev).
Thus using the assumed linearity of L and the homogeneous property of
the norm (lb) we have that
J(yo + ev) - J(yo) = L(v) + ~ IlvI13(ev).

e e
Now as e --+ 0, ev = y - Yo --+ (!J and hence 3(ev) --+ 0, while lel/e = ± 1. Thus
c5J(yo; v) = lim J(yo + ev) - J(yo) = L(v)
.-+0 e
as asserted. D
It follows that the linear function L of the definition is uniquely deter-

mined. It is denoted J'(yo) and called the Frechet derivative of J at Yo'
Observe that differentiability implies continuity as we should wish:
(5.8) Proposition. In a normed linear space (<w, 11'11) if a real valued function
J is differentiable at Yo E <w, then it is continuous at Yo'
PROOF. From Definition 5.6,
IJ(y) - J(Yo)1 :$; IL(y - Yo)1 + Ily - Yo1113(Y - Yo)!.
Now as y --+ Yo, from the linearity and continuity of L,
IL(y - Yo)1 = IL(y) - L(Yo)I--+ 0;
also Ily - Yoll -+ 0, and 3(Y - Yo) -+ 0 from its definition. Thus as Y -+ Yo,
IJ(y) - J(Yo)l-+ 0,
and this establishes the continuity. o
As in ~d, the converses of these propositions need not hold. Continuous
functions are seldom differentiable. Moreover, if J admits the Gateaux varia-
tion £5J(yo; v) in each direction v E qy, the resulting function of v may be
neither linear nor continuous-and even these properties may not suffice
for differentiability. Some additional conditions are required.
Proposition 5.7 provides the key for establishing the differentiability of
a suitably defined function J at a point Yo in a normed linear space (qy, I . II).
First: Check that £5J(yo; v) exists, V v E qy.
Next: Prove that L(v) ~ £5J(yo; v) is linear and continuous in v.
Last: Show that for y f= Yo,
2(y y) deC J(y) - J(yO) - £5J(yo; y - YO) 0 as Ily - Yo I -+ O.

o - 0 = Ily - Yoll -+
This program may either be applied directly to a specific function (Exam-
ple 3 below and Problem 5.22), or it may be used to suggest alternate condi-
tions which would imply differentiability as in the following:
(5.9) Theorem. In a normed linear space (t2¥, 11'11), if a real valued function J
has at each y E Sr(YO) Gateaux variations c5J(y; v), V v E qy and
(a) £5J(yo; v) is linear and continuous in v;
(b) as y -+ Yo, Ic5J(y; u) - c5J(yo, u)l-+ 0 uniformly for u E B = {u E qy:
Ilull = I};
then J is differentiable at Yo'
PROOF*. From condition (a) we may express £5J(yo; u) = L(u) for a linear
function L: qy -+ ~. Each y E Sr(YO) ,.., {yo} may be expressed (uniquely) as
Y = Yo + tu for t = Ily - Yoll < rand u E B. (Why?)
Moreover, for each fixed u E B, f(t) ~ J(yo + tu) is differentiable on
(-r, r) since at tl E (-r, r), with B = t - tl f= 0 and Yl = Yo + t 1u, we have
Yo + tu = Yl + BU,
so that
Thusf'(td = c5J(Yl; u).

Also as t "" 0, f(t) = J(yo + tu) = J(y) -+ J(yo). (Why?) Hence we have by
the law of the mean (§A.l):

J(y) - J(yo) = J(yo + tu) - J(yo) = f(t) - f(O)
= f'(tdt, for some tl E ( - t, t),
= M(Yl; u)t as above.
Observe also that IIYt - Yoll Itll < It I = Ily - Yoll. Hence
=
J(y) - J(yo) - L(y - Yo) = J(y) - J(yo) - tL(u)
= [M(Yl; u) - M(yo; u)]t,
so that y -+ Yo = Yt -+ Yo and
= IM( . u) - M( . )1-+0
IJ(Y) - J(yo) - L(y - yo)1
Ily - Yoll Yt, Yo, u ,
if we utilize the uniformity expressed in (b) to make the last assertion. Thus
( ) ~ J(y) - J(yo) - L(y - Yo) y #- Yo,

3Y-Yo - Ily-Yoll '
has the requisite zero limiting value to satisfy Definition 5.6 for the differ-
entiability of J at Yo. D
Remark. This theorem is the most usable for our purposes. Other sufficient
conditions are known, but all involve some additional uniformity such as
that in condition (b). Without this uniformity, it is only possible to character-
ize the behavior as y = Yo + tu -+ Yo for fixed u. On the other hand, part of
condition (a) is superfluous. The linearity of c5J(yo; v) is a consequence of
condition (b). See [Y].
Conditions (a) and (b) also imply a weak continuity of c5J at Yo in the
sense of the following:
(5.10) Definition. In a normed linear space (CW, 11·11), the Gateaux variations
c5J(y; v) of a real valued function J are said to be weakly continuous at
Yo E CW provided that for each v E 1lJ/: c5J(y; v) -+ c5J(yo; v) as y -+ Yo. [See
Problem 5.34.]
Example 1. The function of §5.3, Example 2, viz.,
J(y) = Lb [sin 3 x + Y(X)2] dx,
is defined if y E £l) = C[a, b]; using the maximum norm II· 11M of§5.1, Exam-
ple 2, we know that J is continuous at each Yo E 1lJ/. Moreover, from (8) in §2.4
we know that V y, v E 1lJ/,
c5J(y; v) = 2 Lb y(x)v(x) dx
and the linearity in v is apparent (Why?) Thus to establish the continuity
in v of c5J(Yo; v) it suffices to establish continuity at v = (!) [§5.3, Example 5].

But since c5J(yo; (!) = 0,
1c5J(yo; v) - c5J(Yo; (!)I =
r
1c5J(yo;
2/
v)1
Yo(x)v(x) dx /
r
=
:::;; 2 IYo(x)llv(x)1 dx
:::;; 211Yo11Mllvll M(b - a), (Why?),

and for fixed Yo E CiJJ this last term -+ 0 as v -+ (!). Thus J as above satisfies
condition (a) of Theorem 5.9.
For condition (b) we suppose u E CiJJ with Iluli M= 1, and estimate similarly
to obtain
1c5J(y; u) - c5J(yo; u)1 = 2/
:::;; 211Y -
r [y(x) - Yo(x)] u(x) dx /
YoIIMllullM(b - a)
= 211Y - YoIIM(b - a). (8)
We observe that the last term -+ 0 as Y -+ Yo and is independent of u. Hence
the left side of (8) -+ 0 uniformly in u when IluliM = 1 as required.
It follows that J is differentiable at each Yo E CiJJ.
r
Example 2. Similarly, the function J of §5.3, Example 3, viz.,
J(y) = lX(x)J1 + Y'(X)2 dx,

for given IX E C[a, b], is defined if Y E CiJJ = C 1 [a, b], and it is continuous in
the maximum norm II· 11M [§5.1, Example 3]. We also know from Example
2 of §5.5 that V y, v E CiJJ:
c5J(y; v) = fb IX(X)Y'(x)v'(x) dx;
a J1 + y'(X)2
again the linearity in v is evident, and it will lead to continuity at v = (!)
essentially as in the previous example. Indeed:
Ic5J(yo; v) I :::;; fb IIX (x) Iy~(x) II v' (x) Idx

a Jl + Y~(X)2
:::;; fa
b
IIX(x)llv(x)1 dx, (since
IZ I
~:::;; 1, V Z E ~ ) ,
:::;; IIIXIIM r Iv(x)1 dx
:::;; A IlvlIM' say.

To establish condition (b) (of 5.9) we observe that J(z) = z/~ is

uniformly continuous on ~ since If'(z) I = (1 + z2 t 3/2 :s 1, V z E ~ and so
(by the law of the mean) IJ(z) - J(zo)1 :S Iz - zol, V z, Zo E ~. Now we make
the following estimate at Yo E ifY:
Ic5J(y; u) - c5J(yo; u)1 :S r loc(x)IIJ(y'(x» - J(yb(x»llu'(x)1 dx
:S Iluli M r loc(x)lly'(x) - yb(x)1 dx :S A IlullMy,
when Ily - YoliM < y. Thus, if Iluli M= 1,

Ic5J(y; u) - c5J(yo; u)1 < E,
when Ily - YollM < Y = EIA, and since E can be made as small as we please
condition (b) is satisfied; the differentiability at an arbitrary Yo E ifY follows
from Theorem 5.9.
Example 3. The length function of §l.1(a), viz.,
L(Y) = Ll'Y'(t)' dt forYEifY == (Cl[O, 1])d
has at each Y E £1)* = {Y E ifY: IY'(t) I ¥- O}, the Gateaux variations
c5L(Y; V) = f0
l Y'(t) ,
IY'(t)I' V (t) dt, V V E!fY,
which are linear and continuous in V by standard estimates.

L is differentiable at each Yo E £1)* in the maximum norm I YIIM' since
O:s IL(Y) - L(Yo) - c5L(Yo; Y - Yo)1
= ILl [IY'I-IY~I-I~I'(Y' - Y~)}t) dtl

= If1 [IY'IIY~I- Y"Y~](t)dtl
o I Y~(t)1
11Y' Y;'12
<f - 0 (t) dt
- 0 IY~I
:S A II Y - Yo I it L1 IY:(t) Idt
= (IIY - YoIIM)(AoIIY- YoII M), say,

so that
IL(Y) - I
L(Yo) - c5L(Yo; Y - Yo) < A II Y - Y; I -+ 0 as Y -+ Y;o.
I Y - YoliM - 0 0 M
Here, we have utilized the vector inequality

IAIIBI- A·B ~ IA - BI2 for A, B E ~d,
which is a consequence of the fact that 21AIIBI ~ IAI2 + IBI2.

Remarkably, it is still more difficult to establish the weak continuity of
<5L(Y; V) at Yo E £1)* which depends upon the previous inequality in another
form. (See Problem 5.35.)
(5.11) Proposition. When f = f(x, y, z), f" and fz E C([a, b] x ~2), then
F(y) = r f(x, y(x), y'(x» dx
is differentiable and has weakly continuous variations at each Yo E qy =

C 1 [a, b] in the maximum norm IIyllM'
PROOF. By (10) of §2.4, the Gateaux variations
<5F(y; v) = r (f,[y(x)]v(x) + fz[Y(x)]v'(x» dx
are obviously linear in v and continuous at each Yo (Problem 5.36). To show
r
that they satisfy condition (b) at Yo, note that
I<5F(y; v) - <5F(yo; v)1 ~ If,[y(x)] - fy[Yo(x)]llv(x)1 dx
by standard estimates.
+ r Ifz[Y(x)] - fz[Yo(x)]llv'(x)1 dx
Now
fy[y(x)] - f,[Yo(x)] = fy(x, y(x), y'(x» - f,(x, Yo(x), y~(x»,
and since f, is uniformly continuous on each box [a, b] x [-c, C]2 (Lemma
5.2), it follows that If,(x, y, z) - f,(x, Yo, zo)1 < e if Iyl, IYol, Izl, IZol ~ c and
Iy - Yol + Iz - zol < r = r(e).
Thus for the given Yo E rtIf, we can choose c so large that IIy - YoIIM ~
1 => ly(x)l, lyo(x)l, ly'(x)l, ly~(x)1 ~ c, V X E [a, b], and hence for a given e > 0,
conclude that :3 r > 0 such that
IIy - YoIIM < r ~ 1 => If, [y(x)] - f,[Yo(x)] I ~ e, Vx E [a, b].
Similarly, for perhaps a smaller r, we can have that
Ifz[Y(x)] - fz[Yo(x)] I
= IfAx, y(x), y'(x» - fz(x, Yo (x), y~(x»1 ~ e, Vx E [a, b].
§S.6*. Affine Approximation: The Frechet Derivative 127
Thus
I£5F(y; v) - £5F(yo; v)1 ~ B r
(Iv(x)1 + Iv'(x)l) dx
= BV, say, when Ily - YollM ~ r ~ 1,

[ ~ B(b - a) when Ilvil M ~ 1];
since B is arbitrary, it follows that
£5F(y; v) -+ £5F(yo; v) as y -+ Yo. [uniformly when IlvilM ~ 1],

and this establishes both the weak continuity and condition (b). Differenti-
ability follows from Theorem 5.9.
(When f = f(x, y) alone, the variations are weakly continuous in the
stronger maximum norm IlyliM = maxly(x)1 of Remark 5.0.) (See Problem
5.24.) 0
Tangency
By 5.6 and 5.7, a function J which is differentiable at a point Yo in a normed

linear space (<?9', 11·11), with Frechet derivative J'(yo), admits a good approxi-
mation near Yo by the affine function
(9)
which is defined V y E I[IJ. The approximation is "good" in the sense that for y
near Yo and y i= Yo,
J(y) - T(y)
Ily - Yoll = 3(y - Yo) -+ 0 as y -+ Yo·
As in Euclidean space (§0.10) the graph of T (a "hyperplane" in <?9' x IR)

may be said to be tangent to that of J at the point (Yo, J(yo» = (Yo, T(yo».
Moreover, comparison with Definition 3.1, shows that a convex differ-
entiable function is one whose graph lies "above" each of these tangent
hyperplanes in <?9' x IR.
However, there is also an intrinsic sense in which T can provide tangency

for J at Yo in <?9' itself: namely, between the respective level sets of these
functions.
Now the level set of T through Yo,
I'yo ~ {y E <?9': T(y) = T(yo)} = {y E <?9': J'(yo)(Y - Yo) = (!)},

is by definition a hyperplane, and if we introduce the corresponding level
set of J through Yo, viz.,

JyO = {y E r1JJ: J(y) = J(yo)},
we see that for y E JyO ' Y ¥- Yo,
, ( y - Yo) T(y) - J(y)

J (Yo) Ily _ Yoll = Ily _ Yoll = -3(Y - Yo),
so that if a sequence Yn E JyO ' (Yn ¥- Yo) n = 1, 2, ... , provides unit directions
.n = (Yn - Yo)/(IIYn - Yo I!) with a limit direction,., as Yn -+ Yo, it follows from
the assumed continuity of J'(yo) that
n-'oo n-'oo
By (9), T(yo + .) = J(yo) = T(yo), and so any possible limit direction. must
furnish a point Yo + • in the hyperplane T,o' Conversely, each. such that
Yo + • E T,o will make J'(yo) = O. (Why?) Accordingly we make the following
(5.12) Definition. In a normed linear space (t2¥, 11'11), if a real valued function
J is differentiable at Yo E r1JJ, then we introduce
T(y) = J(yo) + J'(yo)(Y - Yo),

and say that the level set T,o is tangent to the level set JyO at Yo' Moreover,
each nonzero direction. E r1JJ for which c5J(yo; .) = J'(yo). = 0 is called a
tangential direction to JyO at YO.1
When qy = ~3 has the standard Euclidean norm, then a function J differ-
entiable at Yo has the Gateaux variation
In this case, the linear function

L(.) = J'(Yo). = VJ(Yo)·.,
and the tangent directions. are precisely those which are orthogonal to
the gradient vector VJ(Yo). If VJ(Yo) ¥- (I), then it is perpendicular to the
plane Tyo through Yo determined by these tangent vectors, and hence VJ(Yo)
is normal to the level surface J yo through this point. Thus our definition
of tangency admits this well-known interpretation in ~3.
1 Although this definition provides suggestive terminology, it avoids the deeper question of
whether each such tangential direction , is geometrically tangent in that it is the limit of a
sequence {'n}, of the type described above. This does hold under more stringent requirements on
J. See §A.7 and Liusternik's Theorem in [1-T].
§5.7. Extrema with Constraints: Lagrangian Multipliers 129
§5.7. Extrema with Constraints: Lagrangian

Multipliers
In §5.5, we saw that in finding (local) extrema of a real valued function J
on a subset ~ in a normed linear space (<?!/, 11·11), there may be enough
~-admissible directions for J at a typical point Yo E ~ to provide a usable
characterization of possible local extremal points. However, we also saw
that even in [R2, a domain as simple as a circle has no nontrivial admissible
directions for any function J.
Observe that this domain is itself a level set of another function. For
example, the unit circle {y E [R2: Iyl = 1} is the one-level set of the function
G(y) = Iyl which is differentiable at each point of that level set. In this section
we shall develop the method of Lagrangian multipliers for characterizing the
local extrema of a function J in a normed linear space when restricted to one
or more level sets of other such functions. In this context, the level sets
involved are called constraints, and the equations defining the sets are re-
ferred to as constraining relations.
We have already encountered many examples of constraining relations
in the previous sections. For instance, a set of the form
~ = {y E C[a, b]: y(a) = aI' y(b) = bd
is the intersection of the aI-level set of the function GI(y) = y(a) with that
r r:
of the bI-level set of the function G2 (y) = y(b). The set
~ = {y E C[a, b]: y(x) dx = 1}

is the one-level set of the function G(y) = y(x) dx. In fact, many sets ~
considered previously can be described in terms of level sets of appropriately
defined functions.
To motivate the ensuing develop-ment, we consider first the problem of
characterizing a (local) extremal point Yo of a real valued function J in a
normed linear space (<?!/, 11·11) when constrained to a level set of a real valued
function G. Thus we should have that when y is sufficiently near Yo and
G(y) = G(yo), either J(y) ~ J(yo), 'if such y, or J(y) ;5; J(yo), 'if such y.
This possibility is eliminated if there exists a direction v and scalars 1',
r E [R, as small as we wish, such that upon setting y = Yo + 1'v and y =
Yo + rv, we have -
J(~) < J(yo) < J(y) while G(~) = G(yo) = G(y). (10)
However, it is more useful to consider a pair of directions v, w for which
3 pairs of scalars (1', s) and (r, ~) as small as we please such that these same
requirements (10) hold for
y = Yo + 1'v + sw,
~ = Yo + rv + ~w.
(J
Po = J(yo) p
Figure 5.6
We shall now assume that both J and G are defined in a neighborhood of

Yo, and consider for fixed directions v, w the auxiliary functions
P = f(r, s) = J(yo + rv + sw),

(1 = !§(r, s) = G(yo + rv + sW),
which are defined in some two-dimensional neighborhood of the origin in
1R2. (Why?) The pair of these functions F maps this neighborhood into a
(p, (1) set in 1R2 which contains the point
If it also contains a full neighborhood of (Po, (10)' then there are preimage
points (1', s) and (r, §) and associated y, y for which the conditions (10) are
met. This is readily seen in Figure 5.6. -
Finally, to have (1', s), (r, §) as near (0, 0) as we please we would require
that each small neighborhood of (0, 0) map onto a set which contains a full
neighborhood of (Po, (10). All of this is assured if the mapping F == (f, !§)
has an inverse defined in a neighborhood of (Po, (10) which is continuous at
(Po, (10)·1
The simplest conditions which provide this continuous local inverse are
well known, and form the content of the inverse function theorem which
we state without proof. (See [Ed].)
(5.13) Theorem. For Xo E IRd and 'r > 0, if a vector valued function F:
S«Xo)"'" IRd has continuous first partial derivatives in each component with
nonvanishing Jacobian determinant at X o, then F provides a continuously
invertible mapping between a neighborhood of Xo and a region containing a
full neighborhood of F(Xo)· D
1 A weaker open mapping result which suffices is established in §A.4.

For F = (flO f2, ... , h) we require in this theorem that the matrix with
(continuous) elements 8fJ8xj , i, j = 1, 2, ... , d, arranged in natural order,
have a nonzero determinant when evaluated at Xo' If F defines a linear
transformation of ~d into itself, then this becomes the familiar condition
for invertibility of the matrix representing the transformation.
Now, with y = Yo + rv + sw, the partial derivative
J,.(r, s) = 88 J(yo + rv + sw) = lim ![J(yo + (r + 8)V + sw) - J(y)]

r ..... 0 8
= lim J(y + av) - J(y) = c5J(y; v),

..... 0 8
by Definition 2.6 provided that this variation exists. Similarly,

,I.(r, s) = c5J(y, w); ~r(r, s) = c5G(y; v); ~s(r, s) = c5G(y; w),
provided that these variations exist.
Evaluating at (r, s) = (0,0) where y = Yo, gives us the following:
(5.14) Proposition. In a normed linear space (~ 11'11), if real valued functions

J and G are defined in a neighborhood of Yo and have there in any pair
of (fixed) directions v, w, Gdteaux variations which are continuous in this
neighborhood and satisfy the Jacobian condition
then J cannot have a local extremal point at Yo (even) when constrained to GyO '
the level set of G through Yo'
Remark. The hypotheses of Proposition 5.14 also imply that G cannot have
a local extremal point at Yo (even) when constrained to JyO ' the level set of
J through Yo'
PROOF. Since the non vanishing determinant of the hypothesis is precisely the
Jacobian determinant 8(,f, ~)/8(r, s) evaluated at r = s = 0, we can apply the
inverse function theorem (5.13) to the vector valued function F = (,f, ~)
provided that it has continuous partial derivatives in a neighborhood of
Xo = (0,0).
It suffices to establish the continuity of, say,
J,.(r, s) = c5J(yo + rv + sw; v),
for fixed v, w, in a neighborhood of (0, 0).
But if r1 , SI are such that Yl = Yo + r1 v + SI W is in St(Yo), the neighbor-
hood given by the hypothesis, then y = Yo + rv + sw is within any given r 1 of
Yl if Ir - rll, Is - sll < rd[2(llvll + Ilwll)] since Ily - Ylll ~ Ir - r1 111vll +
Is - sllllwll; (Why?) And by the continuity of c5J(y; v) at Yl we know that
given Ill> 0, 3 'Cl > 0 such that 1!5J(y; v) - c5J(Yl; v)1 < III when Ily - Ylll <
'C 1 • D
With this preparation, it is easy to give conditions necessary for a local
extremal point in the presence of a constraint. We first recall Definition 5.10:
Definition. In a normed linear space (<2Y, 11·11) the Gateaux variations c5J(y; v)
of a real valued function J are said to be weakly continuous at Yo E <2Y pro-
vided that for each v E <2Y: !5J(y; v) -+ !5J(yo; v) as y -+ Yo.
(5.15) Theorem (Lagrange). In a normed linear space (<2Y, 11·11), if real valued
functions J and G are defined in a neighborhood of Yo, a local extremal point
for J constrained to GyO ' and have there weakly continuous Gateaux variations,
then either
(a) c5G(yo; w) == 0, V W E <2Y; or
(b) there exists a constant A E ~ such that !5J(yo; v) = Ac5G(yo; v), V v E <2Y.
PROOF. If (a) does not hold, then 3 WE <2Y for which c5G(yo; w) "# o. With
this wand any v E <2Y, by Proposition 5.14 we must have that the determinant
!5J(yo; v)
Ic5G(yo; I
c5J(yo; w) = O.
v) c5G(yo; w)
Hence with Ad~ c5J(yo; w)/c5G(yo; w), it follows that c5J(yo; v) = Ac5G(yo; v),
V v E <2Y as was to be proven. D
The parameter A which appears in the conclusion of the theorem is called

a Lagrangian multiplier, and in application the theorem is usually referred
to as the Method of Lagrangian Multipliers. It is easier to apply the method
than it is to understand it! However, the following geometrical interpretation
provides some insight.
Utilizing the terminology of directional derivatives appropriate to ~d,
the Lagrange condition c5J(yo; v) = Ac5G(yo; v) says simply that the direc-
tional derivatives of J are proportional to those of G at Yo. If we suppose that
J and G are differentiable at Yo as in the last section, then in the directions 'C
tangent to the level set GyO at Yo, we know that c5G(yo; 'C) = O.
Hence !5J(yo; 'C) = 0 in these tangential directions, and this is what we
should expect for the constrained extremum point. Moreover, unless A = 0
(in which case c5J(yo; .) == 0), !5J(yo; v) is zero only in those directions of
tangency, i.e., the level set JyO of J (unconstrained to Gyo ) through Yo has
precisely the same directions of tangency and nontangency as does Gyo •
Thus in general, Lagrange's condition means that the level sets of J and
G through Yo share the same tangent hyperplane at Yo, or meet tangentially
at Yo as illustrated for ~3 in Figure 5.7.
Recalling the linearity of the Gateaux variation established in §2.4
and replacing A by - A we can also write condition (b) in the form
c5(J + AG)(yo; .) == 0, which suggests consideration of the augmented func-
tion J + AG without constraints, as in §2.3.
Figure 5.7
Example 1. We return now to the second problem of §5.5, Example 1, that of
II
minimizing
J(y) = [sin 3 x + Y(X)2] dx
over
q) = {y E I1JI = C[O, 1]: II X[Y(X)]4/3 dx = 1},
or that of minimizing J subject to the constraining relation
G(y) == II X[Y(X)]4/3 dx = 1.
We know from §2.4, Example 2, that M(y; v) exists for all y, v E I1JI and is
given by
M(y; v) = 2 II y(x)v(x) dx.
Similarly, simple computation shows that t5G(y; v) exists for y, v E I1JI and
is given by
t5G(y; v) = ~ II X[Y(X)]I/3 V (X) dx.
In the maximum norm, both M(y; v) and t5G(y; v) are weakly continuous
by Proposition 5.11. Thus by Theorem 5.15, a point Yo E I1JI which minimizes
JIGYo must satisfy either
n
(a) t5G(yo; w) = 1 xYO(X)I/3 W (X) dx == 0, V W E 11JI, [and this condition can-
not hold since by Lemma 4.4 it would imply that xYO(X)I/3 == or yo(x) == °
° on [0, 1], while Yo = (9 is not in q)]; or
(b) 3 A. E ~ such that
b(J + AG)(yo; v) = L [2Yo(x) + ;hYO(X)1/3J v(x) dx == 0, v V E i5J1.

(11)
By Lemma 4.4, condition (11) implies that 2yo(x) + 4hYo(X)1/3 == 0 on

[0, 1], and so by continuity either Yo == 0 (which we reject as before)
or yo(x) = ±( -ih)3/2. The constant -A> 0 must be chosen to make
J6xyo(x)4f3 = 1 which requires the value A = - 3. Thus the only possible
minimizing (or maximizing) functions for JIG y are yo(x) = ±(2X)3/2. (J - 3G
is not convex (Why?) and we cannot use Theorem 3.16 to further characterize
Yo·)
It is straightforward, albeit rather technical, to extend the method of

Lagrangian multipliers so that it is applicable to a problem involving any
finite number of constraints.
(5.16) Theorem. In a normed linear space (i5JI, 11'11), let real valued functions J,
G1 , G2 , ... , GN be defined in a neighborhood of Yo, a local extremal point for J
constrained to GyO ~ {y E i5J1: Gi(y) = Gi(yo), i = 1, 2, ... , N}, and have there
weakly continuous Gtiteaux variations.
Then either:
bG1(yo; v 1) bG1(yo; v2 )
bG2 (yo; v 1) bG2 (yo; v2 )
(a)
V vj Ei5JI,j = 1,2, ... ,N;

or
(b) there exist constants Ai E ~, i = 1,2, ... , N for which
N
M(yo; v) = L AibGi(YO; v),
i=1
v V E i5J1.
PROOF. If condition (a) does not hold for one set of directions v 1, v2 , ••• ,
VN E i5J1, suppose 3 one direction v E i5JI for which the (N + 1) x (N + 1)
determinant
(12)
i,j = 1,2, ... , N

(having the determinant of condition (a) in its lower right corner) is non-
vanishing. Then the inverse function theorem in IR N +1 can be used as before
to find (N + l)-tuples of scalars (r, 81' 82' ... , 8N) and (r, §l' §2' ... , §N) as near
(0, 0, ... , 0) as we wish for which the points
N
Y = Yo + rv + L: 8jVj'
j=l
N
Y = Yo
-
L: §jVj,
+ rv + j=l
satisfy the conditions
J(y) < J(yo) < J(y),
and
i = 1,2, ... , N.
We thereby exclude a local extremum for J constrained to GyO ' contradicting
the hypothesis.
Thus for the specific set of directions V1 , V2 , •.. , VN the determinant (12)
must vanish for each v E !fIJ, and if we expand it by minors of the first column,
we have upon dividing by the cofactor of c5J(yo; v) (see [N]) an equation
equivalent to condition (b), viz.,
N
15J(yo; v) - L: Aic5Gi(YO; v) = 0,
i=l
v V E!fIf
where for each i = 1, 2, ... , N, the constant
A.- = _ cofactor of c5Gi(yo; v)
• cofactor of 15J(yo; v) ,
is defined since the denominator is precisely the nonvanishing determinant
Ic5Gi(YO; Vj) I
i,j = 1,2, ... , N . 0
(3.6) Remarks. Condition (a) holds if the constraining relations are locally
linearly dependent in that there exist constants Ili' i = 1, 2, ... , N not all zero,
for which L:f=llliGi(Y) = 0, V Y near Yo. Indeed, from the linearity of the
Gateaux variation (see §2.4) it would follow that L:f=lllic5Gi(YO; v) = for °
each direction v E !fIJ. Thus for each set of directions V1 , V2' ..• , V N E !fIJ, the
rows of the determinant of condition (a) are linearly dependent and so it must
vanish.
Conversely, if condition (a) is satisfied for any set of directions V 1 , V 2 , .•• ,
VN E !fIf, then in general the rows (and columns) of the determinant are linearly
dependent. Indeed, upon expanding it by the minors of the first column as in
proof of Theorem 5.16, we would have that L:f=lllic5Gi(YO; Vj) must always
vanish for j = 1, 2, ... , N, since this represents the expansion of a determinant
having two identical columns. Thus the rows of the determinant are linearly
°
dependent (unless Ili = for i = 1, 2, ... , N).
Similarly, Lagrange's condition (b) implies that the variations M(yo; v),
c5G1(Yo; v), ... , c5GN (yo; v) are linearly dependent for each v E 11JJ. Utilizing the
geometric language of IR d, we see that when all functions J, Gi , i = 1, 2, ... , N
are differentiable at Yo as in §5.6 and 't' E ilJJ is a direction simultaneously
tangent to each level set GiyO i = 1, 2, ... , N, then it must also be tangential to
JyO ' the level set of J unconstrained by the Gi at Yo, and thus M(yo; 't') = 0 for
all such directions as we should expect.
However, when N > 1, it is possible that M(yo; v) = 0 for directions v
which are not tangential to any of the level sets GiyO at Yo' Thus we cannot
assert that Lagrange's condition implies common tangency of the level sets at
Yo, as was the case for N = 1.
Observe that upon replacement of Ai by -Ai condition (b) can also be
restated in the form c5(J + If=1 AiGi)(YO; .) == 0 which suggests consideration
of the augmented function J + If=1 AiGi without constraints, again as in
§2.3.
(5.17) Remark. If Yo E ~ £; ilJJ, and the ~-admissible directions for J at Yo

form a linear subspace l1JJo £; I1JJ (so that v, W E ilJJo => rv + sw E l1JJo V scalars
r, s E IR), then it is seen that all arguments used in proving Theorem 5.16
remain valid when the weak continuity and directions v, Vj , j = 1, 2, ... , N
are restricted to ilJJo. Hence the conclusions of the theorem hold for these
restricted directions and provide the conditions characterizing Yo as a local
extremal point of JII'A when further constrained to Gyo ' This observation will
lead to a more efficient but admittedly hybrid approach to certain problems
involving multiple constraints:
Those constraints on J which determine a domain ~ having a linear subspace
ilJJo of ~-admissible directions usable in the sense of §5.5, may be taken into
account simply by restricting the supply of directions used when applying the
method of Lagrangian multipliers to the remaining constraint(s).
Example 2*. To find the possible (local) extremal points for
F(y) = f~1 y'(X)3 dx,

on
under the constraining relation
G(y)
def
=
fO xy'(x) dx
4
= --1 ' (13)
-1 5
we may either characterize ~ by means of the two additional constraining
relations G1 (y) ~ y( -1) = 0, G2(Y)~ y(O) = i and apply Theorem 5.16 with
N = 3; or, we may utilize Remark 5.17, since clearly the EC-admissible

directions for F at any point y E EC are precisely those in
ifYo = {v E C 1 [ -1,0]: v( -1) = v(O) = O},

which forms a subspace of ifY.
We now invoke Theorem 5.16 for these restricted directions. Here
bF(y; v) = f:l 3y'(X)2V'(X) dx,

bG(y; v) = f:l xv'(x) dx,
and these variations are weakly continuous everywhere by Proposition 5.11.

Thus at a local extremal point, y, either:
E
(a) bG(y; v) = J~1 xv'(x) dx = 0, V V ifYo [which would imply by Lemma 4.1
that the continuous function h(x) = x is constant on [ -1, 0] and this is
false]; or
(b) 3 A such that b(F + AG)(y; v) = J~1 [3Y'(X)2 + h]v'(x) dx = 0, V V ifYo. E
Again by Lemma 4.1, we conclude that 3y'(X)2 + AX is constant on
[ -1, 0], or, upon replacing A by - 3A, we have for an appropriate c that
y'(X)2 = c + AX (;:?: 0 on (-1,0)). Thus on (-1,0), either y'(x) = +h -Jc
which cannot satisfy (13) (Why?); orl
y'(x) = Jc + h. (14)
Similarly, the possibility that A = 0 in (14) requires that y'(x) = and to Jc,
satisfy (13) we must take Jc= 185' but then y(x) = 18SX + const. cannot be in
EC. (Why?).
When A =F 0, integration of (14) gives for some constant c 1 :
(15)
where ex = 1- Now
2 2 2c"
y(O) = - => C 1 = - - -
3 3 3A '
while
2
y(-1)=0=>c 1 = -3/C - A)". (16)
Upon subtracting and simplifying, we obtain
A = c" - (c - A)". (17)
1 y' cannot change sign at a point Xo E (-1,0) since y'(X O)2 = c + A.xo cannot be zero.
Next, to satisfy the constraining relation (13) we require that
-14 = G(y) = fO xJc + AX dx

5 -1
= ~A[X(C + ht[ - f~l (c + h)3/2 dX]
= ~[(C - A)~ - ~(C5/2 - (c - A)5/2)]

3A 5A '
or
(18)
We need to solve the nonlinear system, (17) and (18), for c ~ 0 and A :$; c
with A -# O. By inspection, A = c = 1 constitutes one such solution, and hence
from (15), (16):
yo(x) = i(x + 1)3/2 provides a possible local extremal
point in !!} under (13).
[In fact, there is no other admissible solution to this system. l To establish
this, we use (17) to replace (c - A)~ in (18). We get
5A
A2 = 2(A - c~) + C d l + (c - A)(A - c~),
or
WA 2 = mAC~ - AC, and since A -# 0:
A = 3c~ - 2c, so that with ex = !,
c - A = 3(c - c~) = 3c(1 - Jc) ~ o. (19)
Thus 0 :$; c :$; 1; but c = 1 in (19) leads to the case A = c = 1 already
considered, while c = 0, gives A = c = 0 = y'(x) which violates the con-
straining relation (13).
Upon substitution of (19) into (17), we obtain for 0 < C < 1:
3c~ - 2c = c~ - (3c)~(1 - Jc)~ or 2c(Jc - 1) = -(3c)~(1 - Jc)~,
and with ex =l
2 = 3~ Jc(l - Jc)1 /2,
so that
4 = 27c(1 - Jc),
or finally
1 This also follows from convexity. See Problem 5.41 (e).

Problems 139
With Jc= t, this last equation becomes t 3 - t 2 + 247 = 0 which factors

into (t - if(t + t) = O. The only feasible solution is = t = i or c = ~. Jc
But then from (19), A = 3(l7) - m
= 0, a possibility already excluded.]
We conclude that Yo(x) = i(x + 1)3/2 is the only possible local extremal
point for our problem, but we still do not know if it is a local extremal point.
However, now convexity can be used to show that Yo cannot be a local
maximum point with respect to the maximum norm II·IIM of §5.1, Example
3. (See Problem 5.41 *.)
The foregoing analysis involves technical complications which, unfortu-

nately, typify the difficulties encountered when applying the method of
Lagrangian multipliers. However, had we attempted a solution using the
additional constraining functions, G1 and G2 , we would have been burdened
at the outset with two more unknowns, namely the Lagrangian multipliers
A1 , A2 associated with these functions.
PROBLEMS
5.1. Reverse Triangle Inequality. Show that if ('W, 11·11) is a normed linear space,
then
Illyll - 11.9111 :0:;; Ily - .911, Vy,.9ECfII.
5.2. Let 'W = with the Euclidean norm
r
[Rd
Ilyll = (t yJ
2
, for y = (Yl' Y2' ... , Yd)·
(a) Cauchy Inequality. Verify that
lit Yi.9il:o:;; Ilyll·II.9II, v y,.9 E 'W.

(Hint: See Problem 0.1.)
(b) Use the result of part (a) to establish the triangle inequality (Ic).
(c) Show also for this norm that
Parts (a) and (c) do not hold for all norms on 'W.
(d) Verify that IlyliM = maxi =1.2 •...• d Iyil is a norm for [Rd, which does not have
the properties in (a) and (c) above when d ~ 2.
5.3. Let'W = C[a, b].
(a) Verify that Ilylll = S:
ly(x)1 dx defines a norm for 'W.
(b) Does Ilyll = IS:y(x) dxl define a norm for 'W?
5.4. Show that Ilyll = maxxe[a.b]ly'(x)1 defines a norm for the linear space 'Wo =
{y E Cl[a, b]: S:y(x)dx = O}, but does not define a norm for'W = Cl[a, b].
5.5. (a) Verify that Ilyll = ly(a)1 + maxxe[a.b]ly'(x)1 defines a norm for 0/1 =
C 1 [a,b].
(b) Show that maxXe[a.b]ly(x)1 ~ (1 + b - a) Ilyll, V Y E 0/1. Hint:
y(x) = y(a) + r
y'(t) dt.
5.6. (a) Verify that each of the functions, II YIIM and II Y111' given in §5.1, Example 6,
gives a norm for 0/1 = (C[a, bJt
(b) Show that the remaining functions in the example also give norms for 0/1.
5.7. Suppose that both 11'111 and 11'112 are norms for the linear space 0/1.
(a) Show that Ilyll = IIyl11 + IIyl12 defines a norm on 0/1.
(b) Does Ilyll = IIyl11 . IIyl12 also define a norm for o/I?
5.8. (a) Verify the assertion of §5.1, Example 4.
(b) When (~, 11'll j ) are each normed linear spaces for j = 1, 2, find a corre-
sponding norm for the linear space 0/11 x 0/12'
5.9. With 0/1 = C[O, 1J and {Yn} = {(x/2)n},
(a) Show that Yn ...... (!J as n ...... 00, using IIyl11 = H ly(x)1 dx.
(b) Show also that yn ...... (!J as n ...... 00, using IlyliM = maxxe [o.l]ly(x)l.
5.10. Let 0/1 = C[O, 1J and {Yn} = {xn}; i.e., Yn(x) = x n, n = 1,2, ....
(a) Establish that yn ...... (!J as n ...... 00, using 11'111, but
(b) Yn -1+ (!J using II' 11M' where 11'111 and II' 11M are as in Problem 5.9.
(Note: This shows that a sequence from 0/1 may converge to Yo E 0/1 with respect
to one norm, but not with respect to another.)
5.11. Let (0/1, 11'11) be a normed linear space, and {Yn}, {Yn} be sequences from 0/1.
Show that if Yn ...... Yo and Yn ...... Yo as n ...... 00, then (Yn + Yn) ...... (Yo + Yo) as
n ...... 00.
5.12. Suppose that (0/1, 11'11) is a normed linear space, and let {Yn} be a sequence from
0/1.
(a) Show that if Yn ...... Yo as n ...... 00, then IIYnl1 ...... IIYol1 as n ...... 00.
(b) Give an example to illustrate that the converse of (a) is false.
5.13. Use Definition 5.1 to prove that in a normed linear space (0/1, 11'11), a real valued
function J is continuous at Yo E 0/1 iff for each sequence {Yn} from 0/1,
lim Yn = Yo ~ lim J(Yn) = J(yo)·
n-+co n-+oo
5.14. Let 0/1 = C[a, bJ and use Definition 5.1 to establish that J(y) = S!(sin x)y(x) dx
is continuous on 0/1 using:
(a) IlyIIM = maxxe[a.b]ly(x)l.
(b) IIyl11 = S! ly(x)1 dx.
Make a similar analysis for F(y) = S! sin(y(x)) dx. Hint: Use a mean value
inequality.
5.l5. Let (0/1, 11'11) be a normed linear space and L be a real valued linear function on
0/1 (i.e., L(cy + cy) = cL(y) + cL('y), V y, YE 0/1 and V c, C E IR). Prove that L
is continuous on 0/1 iff there exists a constant A such that IL(y)1 ~ A Ilyll,
VYEo/I.
Problems 141
5.16. Suppose that 11'111 and 11'112 are both norms for the linear space 0/1 and there is
a constant A such that IIyl11 :.,:; A Ily112, VY E 0/1.
(a) Show that if y. -+ Yo as n -+ 00 using 11'112, then also y. -+ Yo using 11'111'
(b) Prove that if a real valued function J on 0/1 is continuous with respect to
11'111, then it is also continuous with respect to I . b
5.17. Let (0/1, 11'11) be a normed linear space.
(a) Show that if K is a compact subset of 0/1, then K is bounded, i.e., there is a
constant k such that Ilyll :.,:; k, V Y E K.
(b) Conclude that if 0/1 t= {(9}, then 0/1 itself cannot be compact.
Let 0/1 = qa, bJ and K = {y E 0/1: S:y(x) dx = 1}. Is K compact if we use:
(c) IIyl11 = S: ly(x)1 dx?
(d) lIyllM = maxxE[a.b1Iy(x)l?
5.18. Let (0/1, 11'11) be a normed linear space and J, G, be real valued functions on 0/1
which are continuous at Yo E 0/1. Prove that for C E IR, the following functions
are also continuous at Yo:
(a) cJ; (b) J + G; (c) JG.
Hint for JG: ab - aobo = (a - ao)(b - bo) + (a - ao)bo + ao(b - bo).
5.19. Verify that J(y) = g ly(x)1 dx does not achieve a minimum value on
Pfi = {y E qo, IJ: y(O) = 0, y(l) = I},
although J is bounded below (i.e., J(y) ~ 0) on Pfi. Does Proposition 5.3 cover
this?
5.20*. Let 0/1 = qo, IJ and J(y) = 2y(0)3 - 3y(0)2.
(a) Prove that yo(x) == 1 is a local minimum point for J on 0/1 using lIyllM =
max xE [o.1 1 Iy(x)l. (Hint: Show that y E Sl (Yo) => J(y) ~ -1 = J(yo). Con-
sider minimizing the cubic polynomial p(t) = 2t 3 - 3t 2 on IR.)
(b) Prove that yo(x) == 1 is not a local minimum point for J on 0/1 using
lIyll1 = g ly(x)1 dx. (Hint: Consider the continuous function
-I + 2x/e, 0:.,:; x :.,:; e,
y,(x) = {
1, e < x:.,:; 1,
for each fixed e > 0 and show that lIy, - Yo 111 can be made as small as we
please by choosing e small, while J(y,) = -5 < J(yo), V e > 0.)
5.21. For Example 2 of §5.5, discuss what happens if IX vanishes identically on a
subinterval of [a, b].
5.22. Let 0/1 = qa, bJ,J(y) = S:[sin 3 x + Y(X)2J dX,andPfi = {y E 0/1: S:y(x) dx = I}.
(a) What are the Pfi-admissible directions for J?
(b) Find all possible (local) extremal points for J on Pfi. (See Problem 4.2.)
(c)* Prove directly that J is differentiable at each Yo E 0/1. (See §5.6, Example 3.)
5.23. Let 0/1 = C 1 [a, bJ, Pfi = {y E 0/1: y(a) = a1 , y(b) = bd, and J(y) = S:f(x, y'(x)) dx,
where f(x, z) and f.(x, z) are continuous on [a, bJ x R
(a) What are the Pfi-admissible directions for J?
(b) Show that if y is a (local) extremal point for J on Pfi, then J;,,(x) ~
f.(x, y'(x)) = const. on [a, b].
5.24. Let all = qa, b], ~ = {y E all: y(a) = al, y(b) = bd, and J(y) = J~f(x, y(x)) dx,
where f(x, y) and fy(x, y) are continuous on [a, bJ x IR.
(a) What are the ~-admissible directions for J on~?
(b) Show that if y E ~ is a (local) extremal point for J on ~, then J,(x) ~f
J,(x, y(x)) = 0 on [a, b].
(c)* Prove that the variations t5J(y; v) are weakly continuous in the maximum
norm. Hint: See the proof of Proposition 5.11.
(d) Conclude that if 0: E C[a, bJ, and J(y) ~ J~ o:(x)eY(x) dx, then J cannot have
a (local) extremum on ~ for any values of al, bl , unless 0: = (!J.
In Problems 5.25-5.31 find all possible (local) extremal points for J (a) on ~;
(b) on ptll'
5.25. J(y) = y'(0)2 - y'(0)3,

(a) ptl = C l [0, 1].
(b) ptll = {y E C l [0, 1]: y'(O) = y(l) = O}.
5.26. J(y) = H
cos y(x) dx,
(a) ptl = qo, 1].
(b) ptll = {y E qo, 1]: y(O) = y(l) = n}.
5.27. J(y) = Ii x- l y'(X)3 dx,

ptl= {y E C l [1, 2J: y(l) = 1, y(2) = 8}.
5.28. J(y) = H[Y(X)2 + 2xy(x)J dx,
ptl= {y E C[O, IJ: y(O) = 0, y(l) = -I}.
5.29. J(y) = Jo/ (sec2 X)y'(X)3 dx,
6
ptl= {y E C l [0, IJ: y(O) = 1, y(n/6) = t}.
5.30. J(y) = H[Y(X)3 + eXy(x)] dx,

(a) ptl = C[O, 1].
(b) ptll = {y E C[O, 1]: y(O) = y(l) = O}.
5.31. J(y) = Ii [xy'(x) - eY'(x)J dx,

ptl = {y E C l [1, 2]: y(l) = -1, y(2) = 2(ln 2 - I)}.
5.32*. In Example 2 of§5.5, let 0: E qa, b] with 0: ~ 0:0 > 0 on [a, b].
(a) Show that there exists a Do > 0 such that if 0 < bl - a l < Do, then there is
precisely one c E (0, 0: 0 ) for which (5) is satisfied (and hence precisely one
y E ptl which satisfies (3)).
(b) What happens if a l = bl ?
5.33. Suppose that (0//, 11'11) is a normed linear space for which L: all -+ IR is continu-
ous and linear (i.e., L(cy + cji) = cL(y) + CL(ji), V y, ji E all, and c, CE IR). Show
that L is Frechet differentiable at each Yo E all:
(a) by using Definition 5.6; and
(b) by using Theorem 5.9.
(c) If L=/= 0, prove that 3 Vl E all with L(v l ) = 1, and thus L(t) = 0 when t =
Y - L(y)vl> if Y E all.
(d) In Definition 5.12, take L = J'(yo) and conclude that "most" directions are
tangential.
Problems 143
5.34. If ('W, 11·11) is a normed linear space and J: 'W --+ IR has at each y E 'W Gateaux
variations which satisfy conditions (a) and (b) of Theorem 5.9, verify that
bJ(y; v) is weakly continuous at Yo. Hint: Each v E 'W may be expressed as
v = Ilvll VI' with IlvI11 = 1.
5.35*. In Example 3 of §5.6*, use the vector inequality of Problem 0.2, viz.,
I~
IAI
-~I < .j2 IA - BI ,
IBI - J[AfTBI
(!) =1= A, B, E IR",
to establish the weak continuity of bL(Yo; V) at each Yo E CfII.
5.36. Establish the linearity and continuity in v of the Gateaux variations bF(y; v)
utilized in the first part of the proof of Proposition 5.11.
In Problems 5.37-5.39, use the method of Lagrangian multipliers to determine all
possible (local) extremal points for J on ~.
5.37. J(y) = Hx 2 y(x) dx,

~ = {y E qo, 1]: HY(X)5 dx = I}.
5.38*. J(y) = HY(X)3 dx,
~ = {y E qo, 1]: Hy(x) dx = t, Hxy(x) dx = ~}.
5.39. J(y) = HY'(X)4/3 dx,
~ = {y E CI[O, 1]: y(O) = -5/4, y(l) = 5, Hxy'(x) dx = 5}.
5.40*. Suppose f = f(x, y, z) and its partial derivatives hand fz are continuous
only on (a, b] x D where D is a domain in 1R2. The improper integral F(y) =
S!f(x, y(x), y'(x)) dx may still be finite for some functions y. For given values
aI' bl , let
~* = {y E qa, b]: y(a) = aI, y(b) = bl , with y' E C(a, b] and F(y) finite.}
(a) Show that if y E ~*, then each v in
~6 = {v E CI[a, b]: v(b) = 0 and v(x) == 0 in a neighborhood of a}
is ~*-admissible at y and
bF(y; vo) = fb [f;'(x)vo(x) + h'(x)v~(x)] dx
Xo
when Vo E ~6 and vo(x) == 0 on [a, xo]. (See §1.5 for the notation.) This relax-
ation of conditions near an end point will be required for a careful analysis of
the brachistochrone where
ji+7
f(x, y, z) = M:::.
y2gy
(See Example 4* in §5.5, and Problems 6.14*, 6.15*.
(b) Formulate and prove a vector valued analogue of this result.
5.41 *. (a) For Example 2 of §5.7, prove that for A E IR, the function !(}f, z) = Z3 +
A~Z is strongly convex on [ -1, 0] x [0, 00).
(b) Conclude that when y E ~ and y'(x) ~ 0 on [ -1,0] then for an appropri-
ate A, F(Y) > F(yo), when yo(x) = t(x + 1)3/2 and y =1= Yo.
(c) Draw a sketch to show that in each II· 11M neighborhood of Yo, 3 y E ~
with F(y) > F(yo) and G(y) = J~1 xy(x) dx = G(yo).
(d) Can convexity be used to prove that Yo is a local minimum point for this
problem? Explain.
(e) Use part (b) to conclude that system (17), (18) has at most one solution A, c
for which AX + c ~ 0 on [ -1, 0]. Hint: Each solution pair (A, c) gives a
Yo E ~* = {y E ~: y' ~ O} that minimizes F = F - 3AG on ~* uniquely!
(f)* Redo the problem of this Example when - t replaces - 145 in (13).
5.42. When D is a bounded domain in IRd (for d ~ 2) with a smooth boundary,
verify formally, that Iluli M == maxxeij(lu(X)1 + IVu(X)I) defines a norm for
IJjj = C 1 (D). (D is compact.) See §6.9.
5.43. Find all possible functions that maximize J(y) = y2(1) on ~ = {y E C 1 [0, 1]:
y(O) = O} under the constraint G(y) ~ g y'(X)2 dx = 1.
CHAPTER 6
The Euler-Lagrange Equations
Jakob Bernoulli's solution of 1696 to his brother Johann's problem of the

brachistochrone (§ 1.2) marked the introduction of variational considerations.
However, it was not until the work of Euler (c. 1742) and Lagrange (1755)
that the systematic theory now known as the calculus of variations emerged.
Initially, it was restricted to finding conditions which were necessary in order
that an integral function
F(y) = Lb f(x, y(x), y'(x» dx = Lb f[y(x)] dx

should have a (local) extremum on a set
~ S; {y E Cl[a, bJ: y(a) = al; y(b) = bd.
For specified al, bl this is a fixed end point problem. However, it was already
of interest to Jakob Bernoulli to seek (local) extrema for a larger set
~b S; {y Eel [a, b]: y(a) = ad,
in describing a modified brachistochrone for which it is desired to descend

over a given horizontal distance (b - a) in minimum time, without specifying
the vertical distance to be covered (Figure 6.1 (a». This type of problem is said
to have one free end point.
There are also problems with two free end points where local extrema on
arbitrary subsets of C 1 [a, bJ are desired.
A related problem with variable end point conditions is that of character-
izing the brachistochrone joining fixed curves called transversals (Figure
6.1(b» which would require minimizing the integral with variable limits,
F(y; Xl, x 2 ) = i ,.,'

Xl
f(x, y(x), y'(x» dx = iX'
Xl
f[y(x)] dx
145
146 6. The Euler-Lagrange Equations
--Yo
--YE.@
--Y E .@b or .@,
a b
.-~-~------~~-
x
Y Y
(a) (b)
Figure 6.1
over a set
~t ~ {y E C l [Xl' X2]: 'rj(Xj , Y(Xj)) = O;j = 1, 2},
where [Xl' X 2 ] ~ IR, and the 'rj are given functions.
All of these problems admit a common variational approach: If Yo E

C l [a, b] is, say, a (local) minimizing function for one of these problems, then
with appropriate selection of al, hl and~, we may suppose that Yo E~, and,
as required, that ~ ~ ~b, or ~ ~ ~t. (This possibility is illustrated in Figure
6.1.) In each case, Yo is a (local) minimizing function for F on~, the fixed end
point problem considered initially. Consequently, from Proposition 5.5,
v V E ~o = {v E Cl[a, b]: v(a) = v(b) = O},
which are ~-admissible for F at Yo.
When f is sufficiently differentiable, there are enough such directions to
infer that on (a, b), Yo is a solution of the first and second equations of
Euler-Lagrange. These equations, whose C l solutions are by definition sta-
tionary functions for f, are the subject of the initial sections of this chapter.
The additional freedom of working in ~b or in ~t permits variation in other
"directions" specifically related to the end point freedom, and this will give
rise to the corresponding natural boundary conditions of §6.4 which the
extremal function should satisfy. Problems involving isoperimetric con-
straints are considered in §6.5 through the device of Lagrangian multipliers
and this approach is extended to cover Lagrangian constraints of a simple
form.
In the concluding sections, we examine various extensions of these meth-
ods: first, to integrals involving derivatives higher than the first (§6.6), next, to
§6.1. The First Equation: Stationary Functions 147
integrals involving C 1 vector valued functions (§6.7), and finally (in §6.9)
to integrals over higher-dimensional space. Invariance of stationarity with
respect to change in coordinates is examined in §6.8.
Many of these results were obtained first by Lagrange (1738-1813) who
began his investigations in the subject at age sixteen \ however, his successors
have added mathematical rigor to the original discoveries.
In this chapter, only those conditions necessary for a local extremum are
considered, and although the methods developed are applied to significant
problems of classical interest (including that of the brachistochrone), the final
disposition of such problems must await the discussion of sufficiency in
Chapter 9. It should be noted, however, that the initial investigators in these
fields, often regarded a function which satisfied the necessary conditions as
the extremal function sought, and the practice continues today in elementary
treatments of the subject.
Throughout this chapter, we shall supply the space C 1 [a, b] with the
maximum norm IlyliM = max(ly(x)1 + Iy'(x)l) of §5.1, Example 3, and its
vector valued counterpart (C 1 [a, b])d with the corresponding norm
I YIIM = max(1 Y(x) I + IY'(x)l).
Other norms will be introduced as needed. However, for many of our consid-
erations the particular norm in use is not significant.
§6.1. The First Equation: Stationary Functions

For simplicity, suppose initially that the function f = f(x, y, z), together with
its derivatives J;, and f.., is continuous on [a, b] x 1R2.
F(y) = r
Then for each y E!fIf = C1[a, b]:
f(x, y(x), y'(x)) dx = Lb f[y(x)] dx
r
is defined. From Example 4 of §2.4, F has in each direction v the Gateaux
variation
!5F(y; v) = [J;,(x)v(x) + fy'(x)v'(x)] dx, (1)
where for the given y E !fIf, we use the compressed notation from §1.5:
f,y(X) deC
= J;,[y(x)] an d f,y'(x) deC r
= Jz[Y(x)], (2)
(6.1) Proposition. If y E!fIf makes !5F(y; v) = 0, V v in ~o = {v E!fIf: v(a) =

v(b) = OJ, then J;" is C 1 , and
d
dxJ;,,(x) = J;,(x), x E (a, b), (3)
1 The same age at which Euler obtained his master's degree!

so that
bF(y; v) = fy'(x)v(x) 1:, VVErJ.IJ. (3')
PROOF. The first assertions are a restatement of Proposition 4.2 for the con-
tinuous functions g(x) = fy(x) and h(x) = fy.(x). But then (3) permits the
integrand of (1) to be recognized as (d/dx) [fy-{x)v(x)] and thus integrated
to produce (3'). D
(3) is the first differential equation of Euler (who obtained it heuristically

in 1736 by varying the vertices of an imagined polygonal solution curve) and
Lagrange (who obtained it in 1755 (incorrectly) by integrating the second
term of (1) by parts. Why would this not be permissible?) The correct deriva-
tion as above was given (in 1879) by P. du Bois-Reymond. As we have seen,
(3) must be satisfied along each curve which could provide a local extremal
for F on ~ (or on ~b or ~t) as in the introduction. Moreover, (3) is precisely
the equation (8) already obtained in §3.2.
(6.2) Definition. Each C 1 function y which satisfies the differential equation
(3) (i.e., for which (d/dx)fz(x, y(x), y'(x)) = fy(x, y(x), y'(x))) on some interval
will be called a stationary function for f (of x, y, y').
(An old and rather entrenched tradition calls such functions extremal
functions or simply extremals, although they may provide neither a local
maximum nor a local minimum for the problem.) Observe that we do not
require that a stationary function satisfy any particular boundary conditions,
although in each problem, we might be interested only in those which meet
given boundary conditions.
Now, as in §5.5, certain functions f with their derivatives fy and fz are
defined only for a restricted class of functions y, (e.g., y ~ 0) so that variation
of Fat y can be performed only for a reduced class of v (e.g., those for which
Iv(x)1 ::;; ly(x)I)· As the preceding discussion shows, when y is stationary and
meets the restrictions, then bF(y; v) = 0, V V E ~o for which the variation at y
is defined. However, there may also be nonstationary functions '1 which make
bF('1; v) = 0 for the reduced class of v, and these may provide the true
extremals. (See Problem 6.13.)
(Problem 6.1)
§6.2. Special Cases of the First Equation

Although every C 1 function y is stationary for f(x, y, z) = z or yz, in general,
it is difficult to find any solutions for the first equation (3). However, when
one or more of the variables of f is not present explicitly, then we can at least
§6.2. Special Cases of the First Equation 149
obtain a first integral of the differential equation. We shall analyze three such
cases in this section.
(a) When J = J(z)

Then J;, == 0 so that (3) becomes (d/dx)J;,,(x) = 0 or J;,,(x) = const. Thus
!z(y'(x» = const. and the stationary functions y have derivatives y' which lie
in the level sets of !Z,. In particular, the linear functions, for which y' = const.,
must be stationary.
Example 1. In characterizing the geodesics on a right circular cylinder of unit

radius, we were led in §3.4(a) to minimize
L(y) = f: )1
2
+ [y'(O)]2 dO
on
~ = {y E ClEO, O2]: y(O) = Yl; y(02) = Y2}.
Here
z
f = f(z) = J1+z2, fAz) = J1+z2;
hence, a necessary condition that a given y E ~ minimize L on ~ is that y be
stationary, or that
y'
) = const. or (y'f = const.,
1 + y,2
so that y' = const. In this case, the only stationary functions are the linear
functions y(O) = clO + C2 corresponding to the circular helices on the cylin-
der. (From this analysis alone, however, we cannot say that a helix provides
the minimum sought. We would need, in addition, an argument such as that
used in §3.4(a).)
(b) When J = J(x, z)

Then again fix, z) = 0 so that the stationarity condition (3) is
fAx, y'(x» = const.
Example 2. To characterize those smooth geodesics on a sphere of radius IR

which can be parametrized by functions 0 = y(cp) (see §1.1(b», we should
examine
L(y) = R io'Pl )1 + (y'(cp) sin cp)2 dcp
on
~l = {y E ClEO, CPl]: y(CP1) = O}.
Here
f = f(cp, z) = RJ1 + Z2 sin 2 cp,
so that
Rz sin 2 cp
fAcp, z) = J1 + Z 2SIn
' 2
cp
Thus the stationary functions yare those for which
Ry'(m) sin 2 m
---,=='=~"t'~====~"t'~ = const. = 0 (at cp = 0);
J1 + y'(cp)2 sin 2 cp
i.e., the stationary functions in ~1 are those for which y'(cp) = 0 which corre-
spond to the great circles. Again, the fact that such a function minimizes L
requires separate analysis, as in §l.1(b).
(c) When f = f(y, z)

Then with the abbreviation f(x) = f(y(x), y' (x» it follows from the chain rule
that when y is C 2 :
d d
dxf(x) = dxf(Y(x), y'(x» = fy(x)y'(x) + fy'(x)y"(x).
Upon substitution and cancellation we see that
d d d
dx [f(x) - y'(x)J;,,(x)] = dxf(x) - y"(x)fy'(x) - y'(x) dxfy'(x)
= -y'(x{:xJ;,,(X) - J;,(X)],
and when y is stationary, the right side vanishes by (3). Thus on each interval
of stationarity of y:
f(x) - y'(x)fy'(x) = const. (4)
Conversely, if (4) holds on an interval in which y' does not vanish, then y
is stationary. (Why?) In this case stationarity is characterized by (4) which is
a first integral of (3).1 The additional smoothness requirement that y be C2
can be removed if y is assumed to be a local extremal function. See the next
section.
Example 3. For the function f(y, z) = y2(1 - Z)2 where fz(y, z) = 2 y2(Z - 1),
the (C 2) stationary functions y = y(x) satisfy (4). Thus on each interval, for
some constant c
1 However, this integral is usually nonlinear in y' while the original Euler-Lagrange equation is
sometimes linear.
§6.2. Special Cases of the First Equation 151
so that upon simplification,

y2(1 _ y'2) = C, or y2y'2 = y2 - c.
With the substitution u = y2 (so that u' = 2yy'), we obtain the new equation
U,2 = 4(u - c) which has by inspection, the singular solution uo(x) == c. For
u > c, we get (~)' = ± 1 and hence, the general solution
y2(X) = u(x) = (x + ctl 2 + c. (5)
The constants c and Cl may be found so that y. meets given boundary
conditions. For example, the conditions y( -1) = and y(l) = 1 produce
constants c = -(if and Cl = t; but the resulting function, viz.,
°
Yo(x) = J(x + t)2 - W2 = J(x + l)(x - t)
is defined only for x ~ t, or x :$; -1, and is C2 onl)'. for x > t or x < -1.
(Why?) Moreover, the singular solution Yo (x) == .J c, cannot satisfy these
boundary conditions. For this f, there are no stationary functions in
f!) = {y E C 2 [ -1,1]: y( -1) = 0, y(l) = I}.
On the other hand, Yo is stationary for f on [1, 2] and
Yo E f!)l = {y E C l [1,2]: y(l) = 1, y(2) = 3/J2}.
We shall return to this problem in §7.3, Example 2.
Example 4*. For the brachistochrone problem as formulated in §5.5, Exam-

ple 4, we must minimize
1
T(y)=-
iX! J1 + y'(X)2 dx
on
ft9 0 JYW
f!) = {o :$; Y E C l [0, Xl]: y(o) = 0, y(x l ) = Yl; LX! (Y(X)t3 /2 dx < + 00 }.
Here
f = f(y, z) = Ji+7
Jy
(within a constant factor) and
z
fz(y, z) = r:.
~.
yyy 1 + z
The (C 2) stationary functions y = y(x) satisfy (4) so that
J1 + y'2
r:. - y , ( r:. y' )
= const.,
yy yy~l + y,2
or
1 1
r:. J = const. = -, say;
y y 1 + y,2 C
squaring gives y(1 + y'2) = c2, or
J2 Y
c - y y'=1.
(6)
With the introduction of the dependent variable () = ()(x) such that

() c2
y = c2 sin2 2 = 2(1 - cos ()), for ° :$; () < 2n,
then
y' = c2 sin ~ cos ~()'
2 2
By substituting these expressions into equation (6) we obtain
c2 (sin2 ~)()' = 1 or ; (1 - cos ())()' = 1.
Integrating gives (c 2 /2)«() - sin ()) = x - c l , for a constant c l . Replacing c 2 /2

by c2 , we get the parametric equations
X = c2 «() - sin ()) + Cl'
{ 2 O:$;():$;()l, (6')
y= C (1 - cos ()),
and we see that the only stationary functions are those which determine
cycloids. In order that y E~, we need y(O) = x(O) = 0, which implies that
C l = 0. From the corresponding analysis performed in §3.4(b), we know that
unique constants c2 and ()l < 2n can be found to make the resulting cycloid
with cusp at the origin pass through the given point (Xl' Yl). Since x'«()) =
°
c2 (1 - cos ()) > on (0,2n) it is seen that the X equation can be solved
(implicitly) for () = ()(x), with as many derivatives as desired. The composite
y(x) = y«()(x)) is C2 (at least) on (0, 2n), and except at the bottom of the
cycloidal arch, y'(x) =F 0. Thus from (4), this y(x) is a stationary function for
the problem. However, at the origin y'«()(O)) = +00, so that this function is
not in ~. Although it may represent the brachistochrone sought, we do not
yet have it within the framework of the analysis employed.
There is also a more subtle point to consider before regarding this station-
ary function as a candidate for representing the brachistochrone; it arises
from the proof of Lemma 4.1 as follows:
Our analysis that a minimizing function for T on ~ must be a stationary
°
function, utilized the fact that <5T(y; v) = for a particular v E ~o. However,
as we know from the discussion in §5.5, Example 4, at a given y E ~, the only
v E ~o which are definitely ~-admissible for variation are those for which
Iv(x)1 :$; y(x) on [0, Xl] (or their scalar multiples). Unless the particular v used
to establish stationarity meets this condition, the analysis is not conclusive,
and the true brachistochrone in this class may be provided by a nonstation-
ary function.
§6.3. The Second Equation 153
.---------------------------~~-.
x
Figure 6.2
Finally, in deciding to consider only a special class of curves, i.e., those

with graphs representable by functions of the form y = y(x) or x = x(y), we
have excluded a general curve of the form shown in Figure 6.2 which might
be the "true" mathematical brachistochrone.
We shall return to this elusive problem, which so far remains just outside
the methods being used to analyze it, in §6.4, in §8.8, and in Chapter 9.
(See, however, Problem 6.15* for a reformulation which circumvents some of
the above difficulties.)
(Problems 6.2-6.16)
§6.3. The Second Equation

When f = f(x, y, z) is C 1 and y is a C 1 solution of the first equation (3) on
[a, b], then integration gives
/y,(x) = LX fit) dt + const. (7)
When y is C 2 , then with the abbreviations

def def
f(x) = f(x, y(x), y'(x» and fAx) = fx[Y(x)],
we have
d
dxf(x) = fAx) + /y(x)y'(x) + fy'(x)y"(x), (Why?)
= fx(x) + d~ (y'(x)fy'(x», by (3).

Thus
d
dx [f(x) - y'(x)fy'(x)] = fAx),
or
f(x) - y'(x)/y,(x) = LX fAt) dt + co, for a constant co.
:i' / I
'" / I
.,'" / I
/1
/ I
- - -
y(x) = '1m / /
/ I
I
I
/ I / I
I /
~--~~---------7----~----'
IX a P b x,e
Figure 6.3
This equation resembles (7), the integral form of the first, and, moreover, it
does not exhibit explicitly the C 2 requirement on y used in its derivation.
Hence we can hope to obtain it directly. This is indeed possible (for extremal
functions) but it is surprisingly complicated to do so in view of the simplicity
of the underlying strategy: viz., to conduct the original variational operations
in terms of coordinate axes which are skewed slightly with respect to the
original x, y axes as in Figure 6.3.
and
Let
F(y) = r f(x, y(x), y'(x)) dx
~ = {y E C l [a, b]: y(a) = aI' y(b) = bl }.
(6.3) Proposition. If f E CI([a, b] x ~2) and Yo E ~ is a local extremal func-

tion for F on~, then on [a, b], Yo satisfies the second Euler-Lagrange equation
f(x) - y'(x)J;,,(x) = f: fAt) dt + Co (8)

for some constant co.
PROOF*. For the hypothesized Yo, 3 c > 0 so small that the affine transformation
x= ~ + c'1,
y='1
permits the associated extremal curve to be represented as the graph of a
function '10 E ~ = {'1 E CI[IX, p]: '1(IX) = aI' '1(P) = bl and 1 + c'1'(~) > O}.
Here, IX = a - cal and P = b - cb l .
A "nearby" curve represented by '1 E ~ should also admit the representa-
tion y E ~.l Thus if y = y(x) corresponds to '1 = '1(~), we should have that
'1(~) == y(~ + c'1(~)) so that by the chain rule, '1'(~) = y'(~ + c'1(~))(l + C'1'(m
1 Since 1 + c,,'(e) > 0, we can take y(x) = ,,(e) for the unique esuch that e+ c,,(e) = x.
§6.3. The Second Equation 155
and the derivatives are related through the equations
, l1'(e) 1
y (x) = 1 + cll'(e) or 1 + cll'(e) = 1 - cy'(x). (9)
Under the substitution x = e+ Cll(e), the integral for F(y) transforms as

follows: (Problem 6.35(a))
F(y) = r f(x, y(x), y'{x)) dx = f: Re, l1(e), l1'(e)) de = F(l1), say,
where
- 11,
f(e, C) ~
= f ( e+ Cll, 11, 1 +CcC) (1 + cC). (10)
Since Yo is by assumption locally extremal for F on ~, it follows that 110 is

locally extremal for F on §j. Hence 110 will be a solution to the first equation
for j in the integral form of (7):
lc(e, l1(e), l1'(e)) = f ],,(',11('),11'(,)) d, + cl , (11)
for some constant c 1 .

But from (10):
],,(e, 11, e
C) = fx ( + Cll, 11, 1 JCC) (1 + cC)c + !, (e + Cll, 11, 1 JCC) (1 + cC),
while
lc(e, 11, C) = cf ( e+ Cll, 11, 1 JCC) + fz (e + Cll, 11, 1 JCC)(1 + cCfl.

Using (9) to return to the original variables, (11) becomes (Problem 6.35(b))
cf(x, y(x), y'(x)) + fAx, y(x), y'(x))(1 - cy'(x))
= LX cfx(t, y(t), y'(t)) dt + f: fit, y(t), y'(t)) dt + c l ,
or with the usual abbreviations,
C~(X) - y'(x)!,,(x) - f: fx(t) dtJ = -~Y'(X) - f: !,(t) dt - cll (11')
Finally, upon substituting (7) and subsequently dividing by c, we obtain (8) as

&~ D
Observe that when f = f(y, z) alone as in case (c) discussed in the last
section, then a local extremal function y E ~, must also satisfy the equation
(d/dx)(f(x) - y'(x)!,,(x)) = 0 without additional smoothness assumptions.
Remark. When y is only stationary, this proof does not yield the second
equation unless y is C 2 • (See Problem 6.35(c), (d).) However, if fz is C 1 then y
is C2 when fzz is nonvanishing. (Theorem 7.14)
§6.4. Variable End Point Problems: Natural

Boundary Conditions
r
To find conditions necessary to minimize
F(y) = f(x, y(x), y'(x» dx

(locally) on
rl}a = {y E C 1 [a, b]: y(b) = bd,
where the value of y at a is unspecified, we know that at a local extremal
point y E rl}a, we should have
bF(y; v) = 0,
which are rl}a-admissible at y (Proposition 5.5).
When J, 1" and fz are continuous on [a, b] x ~2, then also each
v E rl}g = {v E C1[a, b]: v(b) = O} (;2rl}o = {v E C1[a, b]: v(a) = v(b) = O})
is rl}-admissible; but by Proposition 6.1, bF(y; v) = 0, V V E rl}o implies that
y is stationary on (a, b), and from (3'),
o= bF(y; v) = I,,(x)v(x) I: = - fy,(a) v(a), Vv E rl}g.
Since v(x) = b - x gives a v E rl}g for which v(a) #- 0, y must be a stationary

function that satisfies the following "natural" boundary condition at the free
end:
(or fz[y(a)] = 0). (12a)
Similarly, if y minimizes F on rl}b = {y E C 1 [a, b]: y(a) = ad (locally)
then y must be a stationary function which satisfies the natural boundary
condition
I,,(b) = 0 (or fz[y(b)] = 0). (12b)
Finally, if y minimizes F on qy = C 1 [a, b] (locally) then bF(y; v) = 0,
V V E qy ;2 rl}g ;2 rl}o. Thus y must be a stationary function which satisfies the
natural boundary conditions (12a) and, from symmetry, (12b». (All of these
conditions were utilized in Proposition 3.9.)
Application: Jakob Bernoulli's Brachistochrone

In 1696, Jakob Bernoulli publicly challenged his younger brother Johann to
find the solutions to several problems in optimization including that of the
§6.4. Variable End Point Problems: Natural Boundary Conditions 157
brachistochrone which covered a given horizontal distance, Xl (thereby initi-

ating a long, bitter, and pointless rivalry between two representatives of the
best minds of their era). Since the time-of-descent function is the same as in
§1.2 and §6.2(c); viz.,
_ 1
T(y) - M:
X
I
1 Jl + y'(xf
r:J::\ dx,
v' 2g 0 v' y(x)
we should attempt to minimize T on
.@l = {o ~ Y E ClEO, Xl]: y(O) = 0, f:' (Y(X)t3/2 dx < +oo}.

From our general analysis, we know that a minimizing function y E .@l
should be a stationary function which satisfies in addition the natural bound-
ary condition (12b)
0= fy.(xd = y'(x l ) .
Jy(x l )Jl + y'(X l )2
As in §6.2(c), Example 4, y must represent a cycloid which satisfies the
natural boundary condition y'(x l ) = 0. Since this requirement can be met
only at the lowest point on the cycloidal arch, which corresponds to () = ()l =
n, it follows from equations (6') that we require y(xd = (2In)xl' and this will
be obtained (uniquely) for the cycloid represented by (6') with c 2 = xt/n. We
still have the difficulty that y'(O) = + 00. Moreover, from this analysis alone,
we cannot conclude (nor could either of the Bernoullis)l that this cycloid
does in fact provide the brachistochrone sought. We know simply that it is
the only curve which supports the variational requirements for a minimum.
[See, however, the relevant comments in §6.2(c).] But, this cycloid meets the
condition that Xt/Yl ~ n12, and so from the analysis in §3.4(b) we can safely
say that it provides (uniquely) the least time of descent among all curves
joining the origin to (Xl' Yl) which can also be represented as the graph of
a function in C l [0, Yl]. In particular, it cannot give a (local) maximum value
for T.
Transversal Conditions*
To obtain the natural boundary conditions associated with more general end
point constraints provided by transversals such as that illustrated in Figure
6.4, it is more convenient to use Lagrangian multipliers. Here we suppose the
integral
J(y, t) = f f(x, y(x), y'(x)) dx = f f[y(x)] dx (13)
is to be minimized over
.@. = {y E c l [a, t]: y(a) = a l ; -r(t, y(t)) = O}.
1 But see Caratheodory's article in the historical references.

r(x, y) = 0
y
(a, ad
a
Figure 6.4
Assume that f and the constraining function rare CIon domains large
enough to admit all functions of interest and that Vr "# (D. If y E C 1 [a, t]
minimizes J, then varying y by functions v in ~o = {v E C 1 [a, t]: v(a) =
v(t) = O} shows as usual that y is a stationary function and thus is a solution
of (3), the Euler-Lagrange equation (d/dx)/y.(x) = fy(x) on (a, t).
For the proper natural boundary condition at the right end, we must
admit more general variations. To provide a convenient framework, we sup-
pose that the functions yare defined by extension on a fixed large interval
[a, b], and introduce the linear space
OJ! = C1[a, b] x IR,
with the norm II(y, t)ll = IlyliM + Itl. (See Problem 5.8.)
A general variation for J in this space in the "direction" (v, ~) is obtained
by differentiating J(y + /lV, t + B~) with respect to B and setting B = O. By
Leibniz' rule (A.14), we get that with the usual abbreviations:
(jJ(y, t; v, ~) = f(t)~ + f (/y(x)v(x) + fy·(x)v'(x)) dx,

and for a stationary function y, the integrand is the derivative of fy'(x)v(x)
so that
(jJ(y, t; v, ~) = f(t)~ + fy. (x) v(x) 1:. (14)
The right end point constraint may be expressed as the zero level set of the
function
G(y, t) = r(t, y(t)) = r[y(t)],
so that upon differentiating
evaluating at B = 0, we obtain
*+ B~, (y + BV)(t + B~)) with respect to Band
(jG(y, t; v,~) = rx[y(t)]~ + ry[y(t)](y'(t)~ + v(t)). (14')

§6.4. Variable End Point Problems: Natural Boundary Conditions 159
(It may be shown that variations (14) and (14') are weakly continuous.) Now,
let y be a local extremum point for J of (13). According to Theorem 5.15,
unless !5G(y, t; " .) == 0 (which would require the vanishing of both 't"x[y(t)]
and 't"y[y(t)]; Why?), then 3 A. E IR such that
!5(J + A.G)(y, t; " .) == O.
Hence restricting attention to those v E !?}o as before which vanish at a and t,
we have from (14) and (14') that
{f(t) + Jc('t"x[y(t)] + 't"y[y(t)]y'(t))}e = 0, v esufficiently small.
Similarly, if we consider variations (v, 0) for which e= v(a) = 0, then
U;,.(t) + A.'t"y[y(t)]} v(t) = 0, v v(t) sufficiently small.
Dividing these last equations by (, v(t), respectively, and eliminating A. be-
tween them shows that a local extremal point y for J on !?}t is a stationary
jUnction on (a, t) which meets the transversal condition
(15)
(15) is the desired natural boundary condition. Note that when 't"(x, y) =
b - x so that 't"y == 0, then (15) reduces to /y.(b) = 0 as obtained earlier.
Similarly, when 't"(x, y) = Y - b1 for given b1 , the terminal value t of x is
unspecified, and at (t, b 1 ) an optimal solution should meet the transversal
condition:
f(t) - y'(t)/y.(t) = O. (15')
In economics, this would be called a free-horizon problem. If the terminal
value b 1 is also unspecified, then in addition to (15'), an optimal solution must
meet the free-end condition /y.(t) = 0, at its terminal point (t, y(t)). Why?
If both end points lie on curves of this type, as in Figure 6.1 (b), then a local
extremal function will be stationary on an interval for which it satisfies (15)
at the right endpoint and the corresponding condition at the left.
Some other types of constraints amenable to the use of Lagrangian multi-
pliers will be treated in §6.7 in connection with vector valued extremals.
Example 1. If the brachistochrone joining the origin to a "lower" curve which

is the zero level set of the function 't" as in Figure 6.5, can be represented as the
graph of a function y E C 1 [0, t] for some t > 0, then we should expect that y
is stationary for the time-of-descent function f(y, z) = )1 + z2jJY (as in
§6.2(c)) and satisfies the obvious boundary condition y(O) = 0 together with
the natural boundary condition (15).
According to the analysis in §6.2(c), y must represent a cycloid joining the
origin to a point (t, y), at which
J1 +r::t:\y'(t)2 't"y[y(t)] = Jy(t))y'(t) ('t"x[y(t)] + 't"y[y(t)]y'(t)),
V y(t) y(t) 1 + y'(t)2
t(x. y) = 0
Figure 6.5
or, after obvious algebra, at which

'y[y(t)J = 'x[y(t)Jy'(t).
Thus, the point(s) (t, y) are those which permit a cycloid joining (them) to
the origin to meet the , curve orthogonally. There may be only one such
point, or there may be many. (See Problem 6.21.) Observe that this result
generalizes that obtained for Jakob Bernoulli's brachistochrone from this
section in which we found that the cycloid in question should at (t, y) have a
horizontal tangent (y'(t) = 0).
§6.5. Integral Constraints: Lagrangian Multipliers

As we observed in the preceding section, end point constraints do not affect
r r
the stationarity of the possible extremal functions for integrals such as
F(y) = f(x, y(x), y'(x)) dx, = f[y(x)J dx
but do control the boundary conditions which the extremal function should
satisfy.
However, frequently present are other constraints which operate over the
entire interval [a, b]. When each of these can also be expressed in integral
form say by requiring that a function
G(y) = Lb g(x, y(x), y'(x)) dx = Lb g[y(x)] dx,

§6.5. Integral Constraints: Lagrangian Multipliers 161
assume a prescribed value, then we can employ the method of Lagrangian

multipliers (cf. §5.7) since, in general, the linearity and weak continuity of the
variations (jF and (jG is assured by Proposition 5.11.
(6.4) Theorem. Suppose that f = f(x, y, z) and g; = g;(x, y, z), i = 1, 2, ... , N,
together with their y and z partial derivatives, are continuous on [a, b] x ~2.
Let Yo be a local extremal function for
on
F(y) =r f[y(x)] dx
q} = {y E C1[a, b]: y(a) = a 1; y(b) = bd,

when further constrained to the set
GyO = {y E C1[a, b]:
Then either:
G;(y)~ r g;[y(x)] dx = G;(yo), i = 1,2, ... , N}.
(a) the N x N determinant
I ~~;~o; v) I= 0,
I,J-1, ... ,N
whenever Vj E q}o = {v E C 1 [a, b]: v(a) = v(b) = O},j = 1,2, ... , N;
or
(b) 3 A; E ~, i = 1, 2, ... , N that make Yo stationary for the modified function
j = f + Li'=l A;g;; i.e., Yo is a solution of the equation
d - -
dxf,·(x) = f,(x) on (a, b).
PROOF. As noted, the hypotheses on f and the g; assure that the variations
(jF(y; v), (jG;(y; v) are linear in v and weakly continuous for all v in the
subspace q}o. Hence from Theorem 5.16 (and subsequent remarks), either
condition (a) holds, V Vj E q}o or 3 A; E~, i = 1, 2, ... , N, for which
(jF(yo; v) = 0, V V E q}o, where
F(y) = F(y) + ;t A;G;(y) = r j[y(x)] dx,
with j defined as in condition (b). Hence by Proposition 6.1, Yo is stationary

~! 0
(6.5) Remark. As in the general theory, the hope is that the Lagrangian
multipliers A; can be determined so that the stationary function Yo E q} will
provide prescribed values for G;(yo), i = 1, 2, ... , N.
(6.6) Remark. Ifin the theorem, q} is replaced by q}b = {y E C1 [a, b]: y(a) = ad
(as in §6.4) then q}o is replaced by the subspace q}~ = {v E C1[a, b]: v(a) = O}
and it is seen that in addition to condition (b), Yo must also satisfy the natural
boundary condition i,(b) = O. Similarly, if ~ is replaced by qy = C l [a, b],

then Yo must in addition satisfy i,(a) = i,(b) = O. Finally, if Yo is required to
meet boundary conditions such as y(a) = al, -c(t, y(t» = 0 for some function
-c, at t E [a, b], where V-c =f. (9, and
F(y, t) = f f(x, y(x), y'(x» dx,

as in §6.4, and where
Gi(y, t) = f gi(X, y(x), y'(x» dx
is prescribed at Yo for i = 1, 2, ... , N, then, in general, Yo will be stationary for

j on (a, t) and satisfy the transversal condition
j(t)-cy[y(t)] = i,(t){ -cx[y(t)] + -cy[y(t)]Y'(t)}. (16)
§6.6. Integrals Involving Higher Derivatives

It is straightforward to obtain results analogous to those in the preceding
sections of this chapter that characterize the local extremals for the function
F(y) ~ r f(x, y(x), y'(x), y"(x» dx = r f[y(x)] dx
for given fE Cl([a, b] x [R3) on domains ~ £; C 2 [a, b] defined by pre-

scribing (possibly) y(a), y'(a), y(b), or y'(b). Indeed, if y, v E C 2 [a, b], then
deC
c5F(y; v) = a/(Y
a + ev) I0=0'
or from A.13
c5F(y; v) = r (/y(x)v(x) + /y,(x)v'(x) + /y,,(x)v"(x» dx,
(see Problem 2.9); where /y,,(x) ~ f..(x, y(x), y'(x), y"(x», whenf = f(x, y, z, r),
(17)
with corresponding extensions for /y(x) and /y,(x).

For definiteness, suppose that
~ = {y E C 2 [a, b]: y(a) = al, y(b) = bl , y'(a) = aD,
where al' bl , and a~ are given real numbers, and let
~o = {v E C 2 [a, b]: v(a) = v(b) = v'(a) = O}.
Then if y E ~ is locally extremal for F on ~, from Proposition 5.5 it follows
§6.6. Integrals Involving Higher Derivatives 163
that c5F(y; v) = 0, V V E P}o. Introducing the C 1 functions
g(x) = LX f,(t) dt
and (18)
h(x) = f [f,,(t) - g(t)] dt,
r
we have upon integrating (17) by parts twice in succession that for v E P}o:
c5F(y; v) = [(f,,(x) - g(x))v'(x) + f,,,(x)v"(x)] dx + g(X)V(x{
= r [h(x) + f,,,(x)]v"(x) dx - h(X)V'(X)i:.
Here, the definition of h assures its vanishing at x = b, and hence the van-
ishing of the boundary term h(x)v'(x)l~ for v E P}o. However, since v E P}o =>
J:
v(a) = v(b) = 0, the definition of 9 is less critical and g(x) ~ f,(t) dt + const.
would also suffice. In any case, as a necessary condition that y be locally
r
extremal, we have that
c5F(y; v) = [h(x) + f,,,(x)]v"(x) dx = 0, VV E P}o, (19)
and hence by Proposition 4.5, that for some constants c and c 1:

h(x) + f,,,(x) = C1x + c, (20)
or by (18),
f,,,(x) = - f [f,,(t) - g(t)] dt + c1x + c, VX E [a, b].
Thus for x E (a, b) : f,,,(x) E C 1 and

d
dxf,,,(x) = f,,(x) - g(x) + C1 ;
similarly, the combination (d/dx)fy"(x) - f,,(x) is C 1 and
! [:xf,,,(X) - f,,(X)] = - f,(x). (21)
(6.7) Definition. (21) is the appropriate Euler-Lagrange equation for a C 1

function f (of x, y, y' and y"). Those C 2 functions y which satisfy (21) on some
interval will be termed stationary functions for f
As might be expected in this case, the additional freedom in the derivative

y'(b) gives rise to a corresponding natural boundary condition as in §6.4. To
discover it we use the Euler-Lagrange equation (21) to replace f,(x) in (17)
and then integrate by parts to get
(jF(y; v) = [0'(X) - ~h"(X)) V(X)] I:

+ Lb [(~fy"(X)) v'(x) + fy"(X)VII(X)] dx
so that
(jF(y; v) = [ 0'(X) - ~h"(X)) V(X)] I: + [h"(x)v'(x)] 1:. (21')
In our case, when v E Etla, the first term on the right in (21') vanishes and the
second reduces to h,,(b)v'(b). Hence from (19), we see that the appropriate
natural boundary condition is
h,,(b) = 0; (or f..(b, y(b), y'(b), y"(b)) = 0). (22)
Other boundary conditions are considered in Problems 6.25-6.28, together
with the derivation of the Euler- Lagrange equation for functions f involving
derivatives higher than the second. A corresponding "second" equation is
obtained in Problem 6.34.
Application: Buckling of a Column under Compressive Loading
It is a fact of experience that work is required to bend a long straight

thin elastic rod of uniform cross section and materia~ and that more work
is required to bend it further (although not necessarily in proportion). As
a result of experiments, Daniel Bernoulli (the son of Johann) concluded
(c. 1738) that the work required per unit length is proportional to the square
of the resulting mean curvature k of that length. 1 Hence the work required to
bend an entire rod of initial length I into a form whose center line is described
by a function y(x), 0 :::;; x :::;; I, as in Figure 6.6(a), is given by
~= Jl I k 2 (x) ds(x),
where the constant Jl is determined by the material and the cross-sectional

shape. 2 Also,
s(X) ~ LX )1 + y'(tf dt,
is the arc length of the center line between 0 and x, while k(x), the local
curvature at x, is from calculus, given by k(x) = IY"(X)I/(1 + y'(X)2)3/2, for
1 See [Ra], page 256, for a simple explanation.

2 Il = EI/2, where E is Young's modulus between stress and strain for the bar, and I is the
moment of inertia of the cross-section.
x p
y y
(a) (b)
Figure 6.6
each x at which y" is defined. This work may also be regarded as the potential
energy of strain stored in the rod as it is bent from an initial unstressed
configuration (supposed straight). Bernoulli conjectured that when the rod is
bent by external forces, it will assume a shape which minimizes the potential
energy. We have already utilized this principle in the analysis of the catenary
problem of §3.5, and what we have thus far would suffice to describe a
situation in which other types of strain energy (work) can be considered
negligible. (See Problem 6.29.)
However, if the bending of a column is produced through buckling under
a longitudinal compressive force of magnitude P applied to the end as in
Figure 6.6(b), then work is also done in compressing the bar. If we regard the
bent bar as an elastic spring of the "new" length
I ds(x) = I J1 + y'(X)2 dx,

then the compressive strain energy in the absence of bending may be deter-
mined from the work done by P in restoring the bar to its original length I;
viz., from
Disregarding further interaction effects, the total potential energy of the

buckled column is WB - fVc, where the negative sign accounts for the fact
that upon buckling, the strain energy of compression is released to be trans-
formed (partially) into that of bending, (For a more rigorous derivation
which permits direct interaction, see ESe], while Euler's original solution of
1744 is discussed in [Fu].)
If the bar is clamped at its lower end as in Figure 6.6(b), then y(O) =
y'(O) = 0, while if we suppose that the upper end remains essentially fixed so
that y(l) = 0, we would wish to minimize the potential energy
on
U(y) == I
0
I [
J1. (1
y"(x?
+ y'(X)2)5/2 - P(J1 + y'(X)2
]
- 1) dx (23)
~ = {y E C2[0, I]: y(O) = y(l) = y'(O) = O}.

This is the situation discussed in the text (with ~o = ~) for the function:
2
f(x, y, z, r) = J1. (1 +rZ2)5/2 - P((l + Z2)1/2 - 1). (24)
Since fy == 0, it follows from (21) that a stationary function, y, for f, makes

h"(x) continuously differentiable and satisfies the equation
d
dxfy"(x) - fy'(x) = const. = c, say, (24')
or with (24), it makes y"/(l + y'2)5/2 (and hence y") continuously differentiable,
and
y,,), y' y'
(
2J1. (1 + y'2)5/2 + 5J1.(Y? (1 + y'2f/2 + P (1 + y'2)1/2 = c.
Mter differentiating and simplifying, we find that
y'" y' {P 5y "2} c

(1 + y'2)5/2 + 2J1 + y'2 It - (1 + y'2)3 = 2J1. = c, say. (25)
The natural boundary condition (22) associated with the unspecified slope
at x = I is (from (24» given by
2J1.y"(l)
(1 + y'(l)2)5/2 =
o, or y"(I) = O. (26)
Thus it would be necessary to solve the third-order nonlinear differential

equation (25) on (0, I) where c is to be determined, if possible, to satisfy the
boundary conditions,
y(O) = y'(O) = y(l) == y"(l) = O. (27)
To carry this out is a highly nontrivial task and requires numerical approxi-
mation methods. (However, one more integration is possible. See Problem
6.34.)
To simplify the analysis, we shall make the assumption that buckling
occurs with a beam geometry for which 1y"1 is small, in particular, one for
which max 1y"1 ~ U Then for y'(O) = 0, it follows that max Iy'l ~ 1 also
(Why?) and (25) may be approximated by the following linear equation with
1 It is sufficient to assume that Iy'l ~ 1. See Problem 6.34.
constant coefficients:
y'" + w 2 y' = c, where w 2 = P/2j1., (28)
subject to the homogeneous boundary conditions (27). Integrating and using
the conditions y(l) = y"(I) = 0 gives the second-order equation
y" + w2y = c(x - I), (29)
whose general solution is known; (see, for example, [B-diPJ). It is given by
y(x) = A cos wx + B sin wx + Cx + D, (30)
for constants A, B, C, and D to be found to satisfy (29) and the remaining
boundary conditions, y(O) = y'(0) = O.
From these last two conditions, we must have
O=A+D or A=-D,
and
o= Bw +C or B = - Cjw.
Differentiating (30) twice gives
y"(x) + w 2 y(x) = w(Cx + D),
and the right side agrees with that of (29) if and only if
Cw 2 = c and Dw 2 = -cl.
Thus the solution is given by a constant multiple (c) of
y(x) = -.; cos wx -

w
~
w
sin wx + ~(x -
w
I), (31)
where it remains to select w 2 = P/2j1. to make y(l) = 0; i.e., to make

= 0, or
wi cos wi - sin wi
wi = tan wI. (32)
Now (32) has an infinite set of solutions wnl E (mr, (n + 1/2)n), n = 1,2, ... , as
is evident graphically from Figure 6.7. The least of these, wll, determines the
load at which buckling "first" occurs; viz.,
Pl = 2j1. w r~ gC;Y·
With w = Wn> (31) defines a sequence of stationary mode functions, Yn in ~
each of which satisfies the natural boundary condition (26) under the addi-
tionallinearizing assumption that
max 1y"1 ~ l.
From (31), with w = Wn (so that wnl > nn), it follows that
wi
Figure 6.7
Thus the smallness of Y" for the actual deflection curve y(x) = cYn(x) is
possible iff c itself is small, which means that the maximum deflection must
itself be small. Such linear analysis is usually termed small deflection theory,
and the approximations are often made in (23), the integral expression de-
fining U itself. (However, if the estimate max ly'12 ~ 1 is used there uncriti-
cally, the term involving P disappears. See Problem 6.30.)
Our linearization has resulted in another difficulty: Since each multiple of
YI (or Yn) is another stationary function which meets the required boundary
conditions, it is not evident in what sense the potential energy U could be
minimized by such functions. What must be realized is that once buckling
has occurred with the critical load PI in the mode described by YI' then
further bending can occur in this mode without additional load until the
nonlinear effects excluded by our analysis become prominent. In particular,
the assumptions of small deflection theory may be violated, even though they
are valid at the instant when buckling first occurs.
Another anomaly requires explanation; namely, whether buckling can
occur only at the critical loads Pn = 2jl.w;. If, for example, the column is
encased in a more rigid structure before loading, loaded by P without
buckling, and then uncased, it is in unstable equilibrium at the critical loads
Pn and buckling in the associated mode Yn can be induced, with the buckled
bar in static equilibrium. However, with the load P2 , say, the column cannot
buckle in a mode Yn for n > 2, since more energy would be required than can
be sustained by P2 • On the other hand, with this loading (or by any P > Pd
buckling in the mode YI could not retain the static equilibrium of the bar (at
least as described by small deflection theory). Thus with moderate loading P,
buckling may be prevented by supporting the bar only at the points of maxi-
mum deflection of the lower mode shapes.
§6.7. Vector Valued Stationary Functions 169
§6.7. Vector Valued Stationary Functions

If we wish to examine whether the curve shown in Figure 6.2 could represent
the (mathematical) brachistochrone, or seek geodesics on a general surface as
in §1.1{c), then we must utilize integrals depending upon vector valued func-
tions. Fortunately, it is straightforward to generalize the analysis in the
preceding sections of this chapter to obtain necessary conditions which char-
acterize a vector valued extremal function in any finite number (d) of dimen-
sions. We recall from §5.1, Example 7, that for the linear space
of elements
having derivatives
Y' = (y~, Y;, ... , y~),
a suitable norm is given by
I YII = max{1 Y{x) I + IY'{x) I}

Thus when fE C1{[a, b] x !R 2d ), to characterize a function Yo E I1JI which
provides a local extremum for the function
on
F{Y) ~ Lb f{x, Y{x), Y'{x)) dx = r f[Y{x)] dx
£Y) = {YE (C1[a, b])d: Y{a) = A; Y{b) = B},

where A, B E !Rd are prescribed, we should introduce vector directions in
£Y)o = {V E (C1[a, b])d: V{a) = V{b) = (D},
consider F{Yo + 6 V) for V E £Y)o and sufficiently small 6 E !R, and require that
«5F{Yo; V) = 0, V V E £Y)o'
As in Example 8 of §2.4,
«5F{Y; V)~ :6 F {Y + 6V)I.=o

or
«5F{Y; V) = r [fy{x)' V{x) + fr{x)' V'{x)] dx
by Leibniz' rule for differentiating under the integral (A.14), where, as ex-
(33)
plained in §1.5, fy{x) is the vector valued function with components J;,/x) ~
J;,.[Y{x)],j = 1,2, ... , d; andfr{x) is the vector valued function with compo-
J r deC r . deC
nents Jz/X) = JzJY{X)], ] = 1, 2, ... , d. (Here we regard f = f{x, Y, Z) =
f{x, Yl' Y2' ... , Yd' Zl' Z2' ... , Zd)') The dot denotes the ordinary scalar prod-
uct in !R d , and is used for convenience in notation.
Now, by assumption, both fy(x) and fY'(x) have continuous components;

hence from (33), if bF(Yo; V) = 0, V V E ~o, then Proposition 4.6 is applicable
and by generalization we obtain the following result:
(6.8) Theorem. If f = f(x, Y, Z) E C1([a, b] x D) for a domain D of 1R2d and
on
F(Y) = r
Yo E qIj = (C 1 [a, b])d is a (local) extremal function for
f[Y(x)] dx
~ = {Y E qIj: Y(a) = Yo(a); Y(b) = Yo(b); (Y(x), Y'(x)) ED},

then on (a, b), Yo satisfies the equation(s)
:xfY'(x) = fy(x); [or :x!,j(X) = fy/x),j = 1,2, ... , d 1 (34)
PROOF*. When D = 1R 2d , the result is an immediate consequence of the pre-

ceding arguments, which also apply to a general subset D if all directions
V E ~o = {V E qIj: V(a) = V(b) = (O} are ~-admissible at Yo. (Why?) But this
is true for each (open) domain D, by an appeal to the compactness of [a, b],
which we shall only outline: For each x E [a, b], the point (Yo(x), Y~(x)) is the
center of a spherical neighborhood ~D, of positive maximal radius r(x) ~ l.
Moreover, the continuity of both Yo and Y~ on [a, b] will guarantee that
r E C[a, b]; it follows from Proposition 5.3 that r assumes a minimum value at
some Xo so that r(x) ;?: r(xo) = ro > 0. Thus, when e < ro and V E ~o with
II V II ~ 1 then Yo + eVE ~ and we conclude that all such V together with
their scalar multiples are ~-admissible at Yo. 0
(6.9) Definition. Equations (34) constitute the vector valued Euler-Lagrange
equations; their C 1 solutions, Y, are called the stationary functions for f
(of x, Y, Y').
In general, Y is stationary for f on (a, b) iff bF(Y; V) = 0, V V E ~o, where

F is the associated integral function of Theorem 6.8. Indeed, such stationarity
permits the use of (34) in (33) to obtain
bF(Y; V) = fY'(x)· V(x) \: = 0, if V(a) = V(b) = (0,
while the converse assertion was established in proving Theorem 6.8.

(Problem 6.40)
There is also an analogous second equation for local extremal Y (cf. §6.3);
viz.,
f(x) - Y'(x)·fY'(x) = IX fAt) dt + c, X E (a, b), (35)
where f(x) = f[Y(x)] and fAx) = fx[Y(x)], However here, the second equa-
tion is scalar and cannot characterize stationarity of Y as in the one-dimen-
sional case (Problem 6.36).
As in §6.4, if, say, yib) is left unspecified for some value of j = 1, 2, ... , d,
there results the associated natural boundary conditionJ;,/b) = O. (See Prob-
lem 6.37.)
To consider more general boundary conditions and/or constraints of the

form G(Y) = const., we may employ the method of Lagrangian multipliers
developed in §5.7. From Theorem 5.15 and Remark 5.17, in general with a
single constraint, we should expect to characterize each local extremal func-
tion Yo of FI~ when further constrained to Gyo ' the level set of G through Yo,
by a A E IR for which
J(F + AG)(Yo; V) = 0, V VE ~o,
(since in this case ~o = OJ/o is a subspace).

In particular, if the constraining function G is itself defined by an integral
in the form
G(Y) = Lb g(x, Y(x), Y'(x)) dx,
then Yo will be stationary for the modified function f + Ag, and so should
satisfy the corresponding Euler-Lagrange equation(s):
d
dx (f + Ag)y'(X) = (f + Ag)y(X) (36)
and an analogous second equation (35). Multiple constraining functions de-

fined by integrals are amenable to a similar extension of Theorem 5.16 (and
Remark 5.17).
In applications, another symbol (usually "t") may be used to represent the

independent variable in the above formulas thereby freeing x to represent
one component of the vector valued Y. For example, in problems involving
planar curves, the use of Y = (x, y) is both more suggestive and less cumber-
some than Y = (Yl' Y2).
Application 1: The Isoperimetric Problem
For the original isoperimetric problem as formulated in §1.3, we are led to

consider the area function of Example 7 of §2.4
A(Y) ~ {l x(t)y'(t) dt
which by an easy extension of Proposition 5.11, has the weakly continuous

variations in the direction V = (u, v) given by
JA(Y; V) = t [x(t)v'(t) + y'(t)u(t)] dt.

We wish to maximize A on
!!) = {Y E O.Y = (C 1 [0, I]?: Y(O) = Y(l)},
subject to the isoperimetric condition
L(Y) = t IY'(t)1 dt = 1 (given).

Without loss of generality we can suppose that the domain D whose area
is to be maximized has the origin (9 in its boundary, and require that the
curves represented by Y originate and end there, so that Y(O) = Y(l) = (9; we
may also take !!)o = !!) = O.Yo. [Note, however, that we have not excluded the
possibility that the origin is a corner point for the curve.]
Formally (from Example 8 of §2.4),
JL(Y; V) = f 0
1 Y'(t)
IY'(t)l· V
,
(t) dt,
so that in order to apply Proposition 5.5, we must further restrict attention to

!!)* = {Y E!!): I Y'I # 0 on [0, I]}.
Curves defined by Y E !!)* are said to be smooth, and a typical curve in !!)* is
sketched in Figure 6.8.
JL(Y; V) is also weakly continuous on !!)*. (This is not immediate: see

Problem 5.35.)
~;-------~~~~~--------.
x
Figure 6.8
Now, if for some Y E ~*, c;L(Y; V) = 0, V V E ~o, then from Corollary 4.7
it would follow that the unit vector Y'/I Y'I = const. so that Y' has a constant
direction, but such functions are not in ~*. (The function Y(t) == (0,0) is in ~
but it is not in ~*.)
Hence from Theorem 5.15 and Remark 5.17, if Yo E ~* maximizes A
(locally) on ~* when restricted to the I level set of L, then :3 A E IR such that
c;(A + AL)(Yo; V) = 0, V V E ~o; Yo is stationary for the function xy' + AIY'I
and satisfies the associated Euler-Lagrange equation(s) (34):1
d(AX')'
dt IY'I = y and
d(
dt X + IAY')
Y'I = 0.
Thus y - (h'/l Y'I) = C2 and X + (Ay'/l Y'I) = c 1 , for constants C1 and c2 •

Hence c2 y' + c 1 x' = yy' + xx', or upon integrating:
x2 + y2 - 2c 1 x - 2c 2 y = const. = 0,
when evaluated at t = 0; this gives (x - C1)2 + (y - C2 )2 = ci + d and the
curve represented by Yo is seen to lie on a circle through the origin. However,
closure requires that the circle be traversed completely at least once, while
maximility in A for the given length I of traversal, could be associated only
with a single traversal. On geometric grounds, c~ + d = 12/4n 2 (Figure 6.8).
A could not not have a minimum value on ~* (or on ~). (See Problem
6.38.) Thus we support but still have not proven Pappus' conjecture that the
maximal curve is the circle. (See §8.8 and §9.5.) Observe that the circle does
not exhibit a corner point at the origin.
Lagrangian Constraints*
The method of Lagrangian multipliers may also be adapted to the case of
constraints of the form
g[Y(x)] = g(x, Y(x), Y'(x)) == 0, VXE[a,b],
where 9 E C 1 (D) for a suitable domain D of 1R2d+1. We shall consider only the
simple case of a single constraint g(Y(x)) == 0, V X E [a, b] which is required
for the discussion of Hamiltonian mechanics in §8.6. The general case will be
treated in §11.3.
(6.10) Theorem (Lagrange). For f=f(x, Y, Z) and fzo E C 1 ([a, b] x 1R 2d ),

j = 1, 2, ... , d, suppose Yo is C2 and it minimizes J
F(Y) = {b f(x, Y(x), Y'(x)) dx
1 For this problem the second equation (35) is satisfied trivially.

locally on
when subject to the constraint
g(Y(x» == 0, VXE[a,b],
where 9 = g(Y) is a C 2 function for which

Vg(Yo(x» =F (!) (d> 1).
Then:3 A E C[a, b] such that Yo is stationary for the modified function f + Ag.
Remark. If gy/Yo(x» ~ 0 for some j, then by relabelling, if necessary, we can

arrange that Y = (Yi' Y). The method of proof involves the local elimination
of this distinguished variable, and to simplify the presentation, we assume
initially that 9 has a form which readily permits this.
PROOF*. Suppose that g(Y) = Y -I/I(Y) for some C2 function 1/1. Then
g(Y) = 0 iff y = I/I(Y), and, in particular, Yo(x) == 1/1 (Yo (x» if we represent
Yo(x) = (Yo(x), Yo (x».
Introduce pj = {Y E C 1 ([a, b])d-l: Y(a) = Yo (a); Y(b) = Yo(b)} and ob-
serve that for each Y E ~ we may define y(x) = I/I(Y(x», to obtain
Y = (I/I(Y), Y) E!!) since y(a) = I/I(Y(a» = 1/1 (Yo (a» = yo(a) (and similarly
y(b) = yo(b». Moreover, by construction, g(Y(x» = y(x) - I/I(Y(x» == 0; i.e.,
these Y automatically satisfy the constraining equation. We also have
y'(x) = VI/I(Y(x»· Y'(x) expressed in terms of the gradient of 1/1.
where
- - - def
F(Y)
- -
r
For Y E pj, we may consider the unconstrained function
~ F(I/I(Y), Y) = J[Y(x)] dx,
- - -
f(x, Y, Z) = f(x; y, Y; z, Z), with y = I/I(Y) and z = VI/I(Y)· Z.
From the chain rule, it follows that in abbreviated form:
Jz = fz V1/1 + /Z,
while
Jy =/yVI/I + fy + fzH,
--def --
where H(Y, Z) = (VI/I(Y)· Z)y.
Now Yo minimizes F (locally) on pj (Why?), and hence as in 6.8, it is a
solution of the first equation in the form
d - - - -
dx/z[Y(x)] = fy[Y(x)].
Upon substitution of the preceding equations and subsequent simplifica-

tion it can be seen that Yo = (rjI(Yo), Yo) is a solution of the equation

d -
dx [!,,(x)VrjI(Y(x)) + fy.(x)]
- d -
= !,(x)VrjI(Y(x)) + fy(x) + !,,(x) dx VrjI(Y(x)). (37)
(In obtaining the last term, we have utilized the identities

0-- d _ d-
oy; (VrjI(Y(x))' Y'(x)) = j~2 rjlYiy/Y(X))yj(x) = dx rjlYi(Y(X)),
for i = 2, 3, ... , d, which hold since rjI is C 2 .)
By hypothesis, fz[Yo(x)] is C 1 , (Why?) Hence when Y = Yo, each term in

the bracketed expression of (37) is C 1 so that Yo also satisfies the equation
d
dxfy,(x) - fy(x) = - [ddx!,'(x) - !,(x)]VrjI(Y(x)).
-
Finally, since gy, == 0 and gy' == (9, while gy == - gy VrjI; (here, gy == 1), then for
each A E C[a, b],
d d
dx (f + Ag)y.(X) - (f + Ag)y(X) = dxfy,(x) - fy(x) - A(X)gy(x)
= -[:x!,,(X) - !,(x) - A(X)gy(X)]VrjI(Y(X)).

(37')
Now, since gy =1= 0 (here, gy == 1), for x E [a, b], we may define
A(X) = (d/dx)!,,(x) - !,(x)

gy(x)
I,
Y(x)=Yo(x)
(37")
and see that this A in C[a, b] forces the vanishing ofthe bracketed terms on
the right side of the last equation which in turn makes the left side vanish as
well. Upon combining these assertions we have for this A that Yo is a solution
of the equations
d
dx (f + Ag)y'(X) = (f + Ag)y(X),
and so it is stationary for the modified function f + Ag under the simplifying

assumption g(Y) = Y - rjI(Y).
In any case, the hypotheses require the nonvanishing of some gy/Yo(~))

for each ~ E [a, b]. Thus by implicit function theory ([Ed]), for each such ~
there is a locally determined C 2 function rjI for which g(Y) = 0 iff y = Yj =
rjI(Y) for all those Y near Yo(~) in ~d. Differentiating the resulting local
Figure 6.9
identity g(I/!(Y), Y) == 0 shows that gy(I/!(Y), Y) == -gy(I/!(Y), Y)VI/!(Y), or in

abbreviated form: gy = -gy VI/!.
We now consider those Y E?i near Yo which differ from Yo only in a small
interval [a, /3] containing ~, and suppose that y(x) = I/!(Y(x)) is defined in
[a, /3]. Then we can set y(x) = yo(x) outside [a, /3] to obtain a Y = (y, Y) E g)
as before. See Figure 6.9. (A = Yo(a) and B = Yo(b).)
For such Y, with] defined as above, we have
f: ][Y(x)] dx - f: ][Yo(x)] dx = F(Y) - F(Yo);
thus f~][Y(x)] dx is minimized (locally) at Yo 1[0:,111' Each such Y yields a

V == Y - Yo, vanishing at a and /3 together with its derivative, V', which is
admissible for variation (on [a, /3]). .fz[Yo(x)] is C 1 (Why?), so that on (a, /3),
Yo is stationary for 1 (See Problem 4.3.)
Each step of the preceding argument is now usable and in particular, we
can again use the A as given by equation (37") to make Yo l[o:,1I1 stationary for
f + Ag on (a, /3). But, the point ~ determining [a, /3] is arbitrary, and from
(37') we see that A(~) is given by (37") independently of which y = Yj we take
as long as gYJYo(~)] "I: O. Thus A is a well-defined function in C[a, b] and Yo
is stationary for f + Ag on (a, b). D
(6.11) Remark. It is straightforward to extend this method ofproofto N < d

constraining functions gj' j = 1, 2, ... , N of the same form, provided that a
suitable N x N Jacobian determinant of the constraining functions is non-
vanishing along the stationary trajectory. The conclusion is that there exist
N functions Aj e C[a, b] such that Yo is stationary for the modified function
When N = d, then Yo will in general be the only function satisfying the

constraining equations.
Application 2: Geodesics on a Surface

In ~3, the surface of an ellipsoid is one of many that can readily be described
as the zero-level set of a smooth function g. If we assume that there is a
geodesic curve (Yo) joining points A and B on this surface, and that along this
curve Vg #- (9, then we may use the previous theorem to characterize this
geodesic when both Yo and g are C 2 •
As in §1.1, suppose that a general curve joining A and B is the range of
Ye (C 1 [a, b])3.lts length is, of course,
L(Y) = Lb IY'(t)1 dt,

and by assumption, Yo minimizes L among such curves which lie on the
surface, i.e., for which g(Y(t» == 0, t e [a, b].
It follows that 3 A e C[a, b] for which Yo is stationary for f + Ag where
f(t, Y, Z) = IZI; thus from (34):
d
dt [(f + Ag)y.(t)] = (f + Ag}y(t),
or, since fy == gyo == (9,
Figure 6.10
or, after substitution,

d(Y' )
dt jY'[(t) = A(t)Vg(Y(t)).
Along Yo we may use the arc length s as the parameter. Then I Y~(s)1 == 1, and
for a new A, the above equation becomes
Y~'(s) = A(S)Vg(Yo(s)),
which shows that in general the principal normal to a geodesic on a surface is
in the direction of the (nonvanishing) gradient and so is normal to the surface
at each point.
Observe that we have not established the existence of geodesics for a
general surface, but we have obtained valuable insight as to the manner in
which such geodesics should lie on the surface. (See Figure 6.10.)
§6.8*. Invariance of Stationarity
For certain physical applications, and especially for Hamilton's principle

(Chapter 8), it is important to know that the property of being a stationary
function does not change with the coordinate system used to describe it. As
indicated in Problem 6.35*, this need not be the case if the function is only Cl.
However, when the function is C 2 , so that both the first and second equations
are satisfied (§6.7), then invariance can be established under the following
general transformation:
We suppose that a portion of ~d+1 containing the graph of a function
Y E (Cl[a, b])d is mapped into ~d+l under the invertible transformation
~ = cp(x, Y), x= cp(~, H),
(38)
H = 'i'(x, Y), Y = 'P(~, H),
which carries the graph of Y onto the graph of HE (C l [a, P])d as shown in
Figure 6.11. The functions cp, 'II and cp, 'l' are to be C2 • We further suppose
that, cp(~, H(~)) increases strictly from a to b with a positive derivative, and,
conversely, that cp(x, Y(x)) increases strictly from (X to p, say.
All of these conditions are realized in the case of the simple skew transfor-
mation considered in §6.3 and the proof of the next theorem is accomplished
by an appropriate extension of the formula obtained in Problem 6.35(d). The
reader should examine this formula and consider the simpler transformation
at each step of the following argument, which is straightforward but compli-
cated by the generality.
Since the point (x, Y(x)) is transformed by (38) into (~, H(~)) we must have
'P(~, H(~)) == Y(cp(~, H(~))), so that upon differentiation,
('P~ + 'PHH')(~) = Y'(x)u(~), (39)

§6.8*. Invariance of Stationarity 179
Y
-
(cp, q,)
- ---..
(cp, 'II)
H
b x P e
''It
Figure 6.11
where
u(~) = :~ cp(~, H(~)) = (cp~ + CPH· H')(~) > 0, (39')
in view of our assumptions.

Thus we have the following correspondence between triples:
, _ ( ('I'~ 'l'HH')(~)).
+ u(~)
(x, Y(x), Y (x)) - cp(~, H(m, 'I'(~, H(m, , (40)
('I'H is the Jacobian matrix having elements ot/l;/01'/j, i,j = 1,2, ... , d, arranged
in natural order with its rows indexed by i).
Now, when f E C1([a, bJ x 1R 3 ), then under the transformation x =
F(Y) =
say, if we define
r
cp(~, H(~)), [so that (formally) dx = u(~) d~J, we have
f[Y(x)] dx = f: j[H(~)] d~ = F(H),
j(~, H, Z) = f( cp(~, H), 'I'(~, H), 'I'~(~, H) +u'l'H(~' H)Z) u, (41)
where the function

u = u(~, H, Z) = cp~(~, H) + CPH(~' H)· Z, (42)
is that used to produce (39').
The desired invariance of stationarity is expressed in the
(6.12) Theorem. If Y E (C 2 [a, bJ)d and c5F(Y; V) = 0, V Yin

!?}o = {V E (C 1 [a, bJ)d: V(a) = V(b) = (O},
then under the transformation (38), H is C 2 and c5F(H; Y) = 0, V Y in
§jo = {Y E (C1[cx, fJJ)d: Y(cx) = Y(P) = (O}.
Remarks. Each V E ~o provides Y E ~o defined by

Y(~) = V(qJ(~, H(~))). (43)
Now, the graphs of V and Y need not correspond under (38), and may lie
outside the domain(s) of the transformation; however, we shall prove that for
this Y:
c5F(H; Y) = c5F(Y; 'PH V) + c5 2 F(Y; lfJR· V), (44)
where
c52 F(Y; w) d~ r UAx)w(x) + (f - Y' "ir)(x)w'(x)] dx. (45)
Here, c5 2 F(Y; w) = (f - Y'"ir )(x)w(x) I:, since this Y is C 2 and stationary

from the hypothesis so that it satisfies the second equation, (35), viz.,
(d/dx)(f - Y'"ir )(x) = fAx). 1
In (44), lfJR and 'PH must be understood as being retransformed into func-
tions of x; i.e., lfJR is actually lfJR(q;(x, Y(x)), 'i'(x, Y(x))), and 'PH is actually
'PH(q;(x, Y(x)), 'P(x, Y(x))).
Then 'PH V is a vector function in ~o as is easily verified by matrix
multiplication, and the first term of (44) vanishes by hypothesis. Similarly,
w(x) ~ (lfJR· V)(x) vanishes at a and b since V E ~o; thus by (45) et seq.
the second term vanishes as well, and we see that for this particular Y:
c5F(H; Y) = o.
Conversely, each Y E ~o arises from that V E ~o given by V(x) ~
Y(q;(x, Y(x)). (Why?) Hence we conclude that c5F(H; Y) = 0, V Y E ~o.
To avoid complications, we shall carry out the derivation of (44) only in

the case d = 1, being careful to preserve the order which would be essential
for the corresponding matrices of the higher-dimensional version. Thus
H, Z, Y, 'P, Y, Z, V reduce to 1'/, (, v, t/I, y, z, v, respectively. (However,
see Chapter 9 where analogous vector valued computations are carried out.)
PROOF*. From (43), v(~) = v(qJ(~, 1'/(~)) so that v'(~) = V'(X)u(~). Then
c5F(1'/; v) = f: (}:'[1'/(~)]v(O + ~[1'/(~)]V/(~)) d~, (46)
but from (41) and the chain rule with the usual abbreviations,
.i.,[1'/(~)] = f(x)u~ + h(x)t/I~u + h,(x)(t/I~~ + t/I~~1'//)

- fy,(X{ t/I~ +ut/I~1'/'] u~ + fx(x)qJ~u,
1 152 represents the effect in the x-direction of a variation l' defined in a different coordinate
system. We utilized this effect to derive the second equation in §6.3.
§6.9. Multidimensional Integrals 181
and
-
.r-[11(~)J = f(x)({J~ + J;,,(x)t/I~ - fy'(x) [t/I~ +ut/I~11'] ({J~,
(since from (42), u{(~, 11, 0 = ({In(~' 11)). The bracketed term in each case is seen
to be simply y'(x) by equation (39), and hence, under the transformation
x = ({J(~, 11(m, dx = u(~) d~, the integrals in (46), may be recognized as aris-
r
ing from the integrals:
{J;,(X)t/lnV(X) + fy,(X{t/lnV'(X) + (t/I~~ +ut/l~~11')V(X)]} dx
+ r {fAX)({JnV(X) + (f - fy'Y')(x) [((J~V'(X) + :~ V(X)]} dx,

where we have restored v(~) to v(x) and so v'(~) to v'(x)u(~).
To complete the proof, it is necessary only to observe that by the chain
rule (formally)
did
dx - u(~) d(
so that
and similarly,
Then we recognize each bracketed term in the last integrals as the derivative
of a product, and we obtain (jF(11; v) = (jF(y; t/I~v) + (j2F(y; ((J~v) in view of
(45).
Observe that under the transformation (38) both Y and H are assumed to
be cl, and from (39) and (40) it would follow that
y"( ) = _1_!£ y'( )= _1_!£ ('P~ + 'PHH')(~)

x u(~) d~ x u(~) d~ u(~)
is defined and continuous when H is C 2 • By the corresponding argument we
could conclude that conversely when Y is C 2 , then H is C 2 also. D
§6.9. Multidimensional Integrals

As we have seen in §1.4(c) and §3.4(e), it is frequently required to optimize
functions F defined by an integral over a domain D of [Rd where d > 1. When
the structure of D and its boundary aD are sufficiently regular, it is straight-
forward to obtain formal analogues for most of the one-dimensional results
obtained hitherto in this chapter. (Without this regularity, however, far more
sophisticated tools are required to handle the delicate questions concerning
behavior at the boundary. See [G-T].)
A typical point in IRd will be denoted by X = (Xl' X 2 , .•• , Xd) in Cartesian
coordinates, and the d-dimensional element of integration by dX. D is
assumed to be a bounded Green's domain in IRd-i.e., one for which the
boundary aD consists of (d - I)-dimensional surfaces on which integration is
possible such that Green's theorem holds in the divergence form
fD
(V, U) dX = f
aD
(U . N) du, (47)
Here C1(V) denotes the set of real valued functions u E C(V) which in D
have first partial derivatives admitting continuous extensions to 15; U =
(u 1 , u2 , ••• , ud ) is a d-tuple of such functions with the divergence
def d
V' U = L (u.)
j=l J
.
Xj'
N is the outward pointing unit normal vector on aD (which is defined except

at a set negligible with respect to the surface integration), and du denotes the
element of integration on aD (Figure 6.I2(a)). (See [Ed].)
For example, in 1R 3 , all of the above hold when D is the interior of a

rectangular box with faces parallel to the coordinate planes (Figure 6.I2(b)).
Then dX = dX 1 dX 2 dX3 and the integration over D can be expressed through
iterated integrals. aD consists of the six faces of the box which meet only at
the edges. Finally on each face, du = dX 1 dx 2 , or dX 2 dx 3, or dX l dX3 so that
the surface integrals can also be expressed through iterated integrals, and
N
N
r-----------------------~
X2
(a) (b)
Figure 6.12
Green's theorem can be verified by partial integrations. However, all of

the above also hold when D is the interior of an ellipsoid in 1R 3 , but in this
case, it is difficult to express dX, Nand du in forms convenient enough to
verify Green's theorem easily. A typical u E C 1 (15) has in D, the gradient
def
Vu = (U X1 ' U X2 ' ••• , ux ), and hence
Iluli M = max (lui + IVul)(X)
x
will supply a norm for this space. (See Problem 5.42.)
Given f E C 1 (15 x IR x IRd) and y E qeD), we shall be interested in finding

conditions necessary that u E C 1 (15) be a local extremal function for
F(u) = L f(X, u(X), Vu(X)) dX = L f[u(X)] dX

on
!')~ {u E C 1 (15): UlaD = y}.
We introduce the admissible directions (§5.5)
!')O = {v E C1 (15): vlaD == OJ,
and observe that if u is a local extremal point for F on !'), then
c5F(u; v) = ~ F(u + sv) I

uS .=0
= 0, V V E !')o·
The differentiability assumptions on f permit this variation to be calculated

by Leibniz' formula as follows:
c5F(u; v) = L [fu(X)v(X) + fvu(X)· Vv(X)] dX, (48)
where h(X) = h(X, u(X), Vu(X)) and fvu(X) is the vector-valued function
with components
fz.(X)=
J
fz.(X,
J
u(X), Vu(X)), j = 1,2, ... , d;
(where f = f(X, u, Z) = f(X, U, Zl' Z2' •• ·' Zd).
Next, we suppose that f E C 2 (15 x IR x IRd), and that u E !') n C 2 (D) (as in
§3.4(e)) so that we can integrate the second term of (48) by parts using Green's
theorem (47), for U(X) = v(X)fvu(X), as follows:
L [fvu(X)· Vv(X)] dX = Lv. (v(X)fvu(X)) dX - L [v(X)V ·fvu(X)] dX
= f aD
v(X)fvu(X)· N(X) du - f D
v(X)V "ivu(X) dX
=- L v(X)(V "ivu(X)) dX, V v E !')o, (49)
since the boundary integrand vanishes when v E !')o.

Thus, finally, if u E ~ ("\ C 2 (D), then V v E ~o
~F(u; v) = Iv [fu(X) - V "ivu(X)]v(X) dX. (50)
In particular, if u in ~ ("\ C 2 (D) is a local extremal for F on ~ then the integral

in (50) vanishes V v E ~o, and we wish to conclude from this that the brack-
eted term in the integral vanishes identically in D. This can be accomplished
by a natural extension of the proof of Lemma 4.4 of Lagrange. (See Problem
4.5.) Thus, finally, we obtain the desired analogue of the fundamental result
of Euler-Lagrange.
(6.13) Theorem. Let D be a Green's domain of ~d and suppose that functions

f E C 2 (D x ~ x ~d) and Y E qaD) are given. Then in order that Uo E C 2 (D) be
a (local) extremal for
F(u) = Iv f(X, u(X), Vu(X)) dX = Iv f[u(X)] dX

on
it is necessary that U o be stationary for f in D; i.e., that U o satisfy the equation

VXED. (51)
D
Application: Minimal Area Problem

For the minimal area problem of Plateau discussed in §1.4(c) and §3.4(e), we
take d = 2. Then f(X, u, Z) = (1 + zi + zn
1/2 has continuous derivatives of
all orders; fu = 0 and fZ j = zi1 + zi + zn-

1/2 ,j = 1,2. Denoting X by (x, y),
so that Vu = (u x , uy ), the Euler-Lagrange equation (51) for this problem is
V.( - 0
(1 + u~Vu+ U;)1/2 ) - ,
(52)
which agrees with Equation 26 of §3.4. Thus in order that Uo E C 2 (D) have a
graph with a local extremal surface area among all such functions with the
same continuous boundary values, it is necessary and sufficient that U o satisfy
the minimal surface equation ((26) of §3.4) or its equivalent. As we have noted
in §3.4(e), this equation has a solution with arbitrarily prescribed continuous
boundary values iff the domain D is convex [Os].
Multidimensional problems arise naturally when Hamilton's principle is
applied to obtain the equations governing the motions of elastic bodies. We
shall reserve further discussion until §8.9.
(Problem 6.41)
Natural Boundary Conditions
Multidimensional problems permit considerable flexibility in the specifica-

tion of the boundary conditions, and boundary point freedom from specifi-
cation gives rise to natural boundary conditions which a local extremal
function must satisfy.
(6.14) Corollary. If as in Theorem 6.13, U o E C 2 (D) is a local extremal function

for F on
~ = {u E c 1 (15): u(X) = y(X), X E K},
where K is a compact subset of aD, then
(i) in D, U O satisfies (51);
(ii) at each boundary point Xo ~ K having a neighborhood of aD in which N is
continuous: fvu(X o)' N(Xo) = O.
PRooF*. Each v E ~o = {v E C 1 (15): v(X) = 0, X E K} is ~-admissible (Why?),
and since ~o ;2 ~o, we conclude as before that U o satisfies (51).
Then, when v E ~o, it follows from (48) using Green's theorem as in (49),
that for u = uo:
0= c5F(u; v) = L [!u(X) - (V 'fvu}{X)]v(X) dX
+ faD
v(X) (fvu(X) . N (X» du
= fi v(X)fvu(X)' N(X) du,
since (i) holds and v vanishes on K. (K = aD '" K.)

Now, each Xo ~ K is a positive distance from the compact set K. [Other-
wise we could find a sequence of points in K with limit point Xo ~ K (see
§5.2).] Thu~ we can construct functions v E C 1 (15), which are positive at Xo
and vanish outside neighborhoods of Xo so small that v E ~o. By hypothesis
fvu(X) is continuous at Xo; hence if also N is continuous at X o, we can use a
version of the standard local arguments to conclude that fvu(X o)' N(Xo) = O.
(See Problem 4.5.) D
Consider the surface area function of the previous example when the
boundary values are specified only on a compact subarc K of the boundary.
Then a minimizing function u should satisfy the minimal surface equation
(52) in the domain D and have the prescribed values on K; however, on K, we
expect in general that
Vu·N
- -____---=--= = 0 or Vu' N = 0;
(1 + u;
+ U;)1/2
i.e., that the derivative of u in the direction normal to the boundary curve
should vanish. Conversely, we may use convexity as in §3.4(e), to show that
such u will in fact minimize the surface area function uniquely under these
conditions (Problem 3.37).
PROBLEMS
6.1. Give the first Euler-Lagrange equation for f when:

(a) f(x, y, z) = sin z.
(b) f(x, y, z) = X 3 Z 3 •
(c) f(x, y, z) = ~/x, (x =f. 0).
(d) f(x, y, z) = y2 - Z2.
(e) f(x, y, z) = 2xy - y2 + 3zy2.
6.2. Find the stationary functions for f which belong to !?fi if:
(a) f(x, y, z) = sin z, and
!?fi = {y E CI[O, 1]: y(O) = -5, y(l) = 2}.
(b) f(x, y, z) ~/x, and
!?fi = {y E C l [0, 3]: y(O) = 0; y(3) = 3}.
(c) f(x, y, z) = y2 - Z2, and
!?fi = {y E CI[O, n/2]: y(O) = 0, y(n/2) = 1}.
(d) f(x, y, z) = y2 - Z2, and
!?fi = {y E C l [0, n]: y(O) = y(n) = OJ.
(e) f(x, y, z) = 2xy - y2 + 3zy2, and
!?fi = {y E CI[O, 1]: y(O) = 0, y(l) = 1}.
(f) f(y, z) = ~/y, y > 0
!?fi = {y E c l [0, 2]: y(O) = y(2) = 1, y(x) > OJ.
In Problems 6.3-6.12 find the possible local extremal functions for F on !?fi.
6.3 F(y) = SO [Y(X)2 - Y'(X)2] dx, and

!?fi = {y E C l [0, n]: y(O) = y(n) = OJ.
6.4. F(y) = Ii x 3Y'(X)2 dx, and

!?fi = {y E C l [1,2]: y(l) = 5, y(2) = 2}.
6.5. F(y) = Ii y'(X)3/X2 dx, and

!?fi = {y E C I [l, 2]: y(l) = 1, y(2) = 7}.
6.6. F(y) = H/2 J1 + y'(X)2/X dx, and
!?fi = {y E CI[t, 1]: y(t) = -J3/2, y(l) = OJ.
6.7. F(y) = S~ [Y(X)3 + 3x2y'(X)] dx, and

(a) !?fi = {y E CI[O, 1]: y(O) = 0, y(l) = 1}.
(b) !?fi = {y E CI[O, 1]: y(O) = 0, y(l) = 2}.
6.8. F(y) = S~ [1 + Y(X)2]/y'(X)2 dx, and
!?fi = {y E CI[O, 1]: y(O) = 0, y(l) = sinh 1; with y'(x) =f. 0 on (0, 1)}.
6.9. F(y) = S6 [2xy(x) - Y'(X)2 + 3Y'(X)Y(X)2] dx, and
!?fi = {y E CI[O, 1]: y(O) = 0, y(l) = -1}.
Problems 187
6.10. F(y) = S6 [2xY(X)3 + e" sin y(x) + 3x2Y(X)2y'(X) + y'(x)e" cos y(x)] dx, and
(a) !!} = {y E CI[O, 1]: y(O) = 0, y(1) = 1}.
(b) !!} = {y E c l [0, 1]: y(O) = n, y(1) = j8}.
6.11. F(y) = So [4y'(x)2 + 2y(x)y'(x) - Y(X)2] dx, and

!!} = {y E Cl [0, n]: y(O) = 2; y(n) = OJ.
6.12. F(y) = S6 [Y'(X)2 - 6x 2y(x)] dx, and

!!} = {y E C l [0, 1]: y(O) = i. y(1) = ;f}.
6.13. Let F(y) = S6Jy(x) - x dx and

!!} = {y E qo, 1]: y(O) = 0, y(1) = 1, y(x) ;;:: x on [0, 1]}.
(a) Show that there are no stationary functions for f(x, y, z) = ~.
(b) Show directly that Yo (x) = x minimizes F on !!}.
(c) Verify that bF(yo; v) = 0 for all v which are !!}-admissible for the function
Yo(x) = x. (Are there any?)
(d) What does this example demonstrate?
6.14*. With the same definitions as in Problem 5.40*, duplicate the analysis in §6.1 to
prove that if bF(y; v) = 0, V V E!!}~ then J,,(x) + g(x) = Co on (a, b], where
g(x) = S~J,(t) dt when x E (a, b]. Conclude that y is stationary for f on (a, b].
(Hint: On each interval [xo, b], apply the result of Problem 4.3.)
6.15*. For the brachistochrone problem in Example 4* of §6.2, use Problem 6.14* to
show that the first equation (djdx)fy'(x) = J,(x) can be integrated as it stands
upon multiplication by
f, ,(x) = y'(x)
y JYWJ1 + Y'(X)2
Conclude that with an appropriate !!}*, the only possible minimizing function
for T on !!}* is that representing the cycloid given parametrically by equation
(6). Why is this an improvement?
6.16*. Minimal Surface of Revolution. To find a smooth planar curve represented by

y E C~[O, b], joining points (0, 1) and (b, bd with bl > 0, which when rotated
about'the x-axis will have the smallest possible surface area of revolution, we
should minimize the surface area integral
S(y) = 2n f: y(x)J1 + y'(X)2 dx

on
!!} = {y E CI[O, b]: y(O) = 1, y(b) = bl ; y(x);;:: O}
Prove that if such a curve exists, it must be of the form
y(x) = CI cosh [(X ~l C2)}

where CI and C2 are constants. (Note: Depending on the location of the points,
there may be one, two, or no curves of this type which meet the given boundary
conditions. This problem will be dealt with again in Chapters 7 and 9.)
In Problems 6.17-6.20, find all possible (local) extremal functions for F on fi).
6.17. F(y) = So/2 [Y(X)2 - y'(X)2] dx:
(a)fi) = {y E C1[0, n/2]: y(O) = O}.
(b)fi) = {y E C1[0, n/2]: y(O) = 1}.
6.18. F(y) = S6 [(y'(x) - X)2 + 2xy(x)] dx

fi) = {y E C1[0, 1]: y(O) = 1}.
6.19. F(y) = S6 cos y'(x) dx

fi) ={y E C 1 [0, 1]: y(O) = O}.
6.20. F(y) = S6 [xy(x) - y'(X)2] dx:
(a)fi) = {y E C1[0, 1]: y(O) = 1}.
(b)fi) = {y E C1[0, 1]: y(1) = 1}.
6.21. Consider the problem of finding a smooth curve of the form y(x) which will
provide the shortest distance from the origin to the parabola given by
y = x 2 - 1.
(a) What are the stationary functions for this problem?
(b) Show that there are precisely two points (t, y) and (t, ji) on the parabola
which satisfy the transversal condition (15).
(c) Find the associated curves which represent the possible extremals.
(d) Use a direct argument to show that a minimum is actually achieved for
each of the curves found in part (c).
(e) What happens if the parabola is replaced by the circle x 2 + y2 = 1?
6.22. Brachistochrone. (See §6.4, Example 1.) A brachistochrone joining the origin to
the straight line y = 1 - x is sought. Show that there is precisely one point
(t, y) on the line which satisfies the transversal condition (15) and find the
associated cycloid which might be the brachistochrone in question.
6.23. (a) Use the method of Lagrangian multipliers to find all possible (local)
extremal functions for
F(y) = f Y(X)2 dx
on
fi) = {y E C 1 [0, 1]: y(O) = y(1) = O},
when further constrained to {y E C 1[0, 1]: S6 y'(X)2 dx = 1}.
(b) Can the convexity methods of Chapter 3 be used to conclude that a mini-
mum is achieved in part (a)? Explain.
6.24. (a) Use the method of Lagrangian multipliers to find all possible (local)
extremal functions for
F(y) = f: [2(sin x)y(x) + y'(X)2] dx
on
fi) = {y E C 1 [0, n]: y(O) = y(n) = O}
when further constrained to {y E C1[0, n]: SoY(x) dx = 1}.
(b) Can the convexity methods of Chapter 3 be used to conclude that a mini-
mum is achieved in part (a)?
Problems 189
6.25. Let fE CI([a, b] x [R3). Show that the natural boundary conditions asso-
ciated with minimizing
locally on:
F(y) = r f(x, y(x), y'(x), y"(x» dx
(i) f!) = {y E C 2 [a, b]: y(a) = aI' y(b) = btl are h,,(a) = k(b) = O.
(ii) f!) = {y E C2 [a, b]: y(a) = aI' y'(a) = a'} are k(b) = 0 and
ddX h"(x) Ix=b = h,(b).
6.26. Let f E CI([a, b] x [R3). Find the natural boundary conditions associated with
minimizing
on each of the following sets:

F(y) = r f(x, y(x), y'(x), y"(x» dx
(i) f!) = {y E C2 [a, b]: y'(a) = a', y'(b) = b'}.

(ii) f!) = {y E C2 [a, b]: y(a) = aI' y'(b) = b'}.
6.27. LetfE CI([a, b] x [R4):
on
= r
(a) Show that an Euler-Lagrange equation for a function which minimizes
F(y) f(x, y(x), y'(x), y"(x), y"'(x» dx
f!) = {y E C 3 [a, b]: y(a) = aI, y'(a) = a', y"(a) = a",

y(b) = bl , y'(b) = b', y"(b) = b"},
is
where we employ the usual abbreviations, and assume sufficient

differentiability.
(b)* Without the additional differentiability, what form will this equation take?
6.28*. Let fE CI([a, b] x [Rn+1). Show that an Euler-Lagrange equation for a func-
r
tion which minimizes
F(y) = f(x, y(x), y'(x), y"(x), ... , y<n)(x» dx

locally on
f!) = {y E cn[a, b]: y(a) = aI' y'(a) = at, ... , y<n-I)(a) = dn-I),
y(b) = bl , y'(b) = b l , ... , y(n-I)(b) = b(n-I)}
is
where we employ the usual abbreviations, and assume sufficient differentiability.

Figure 6.13
6.29. A thin elastic rod of initial length I clamped at one end and pinned at the other
is deflected as shown in Figure 6.13 from its straight unstressed state. If the
center line of the rod is described by a smooth function y(x), 0 :::;; x :::;; I, then the
associated potential energy is given by
U(y) = J1
r1 Y"(X)2
Jo [1 + y'(X)2]5/2 dx,
where J1 is a constant. The physically imposed boundary conditions are y(O) =
y'(O) = 0, and y(l) = 11.
(a) Assuming that the shape of the rod minimizes the potential energy, find a
third-order differential equation satisfied by y(x).
(b) What is the natural boundary condition at x = I?
(c) Find a suitable linearized version of the differential equation from part (a)
by supposing that both ly'(x)1 and Iy"(x) I are very small on [0, I].
(d) Solve the linear equation found in part (c), choosing the constants to
satisfy the boundary conditions.
6.30. Buckling of a Column. (See §6.6.) In small deflection theory, approximations
are often made in the potential energy, rather than in the differential equation.
(a) Show that the approximation of (23) by
U(y) = I[ J1y ll(X)2 - ~ y'(X)2 ] dx
leads to the same linear differential equation (28).

(b)* Explain how the approximation in part (a was obtained. Hint: Consider
the effect of the complementary factor ( 1 + (y')2 + 1).
(c) Use (29) to determine the buckling loads when the end at x = 0 is pinned
so that y"(O) = 0 instead of y'(O) = O.
In Problems 6.31-6.33, find all possible (local) extremal functions for F on!?}.
6.31. F(y) = S~/2 [Y"(X)2 - Y(X)2] dx,
!?} = {yE C 2[0, nj2]: y(O) = 1, y'(O) = y(nj2) = 0, y'(nj2) = -1}.
6.32. F(y) = S~ [Y"(X)2 - y'(X)2] dx,

!?} = {y E C 2[0, n]: y(O) = 0, y'(O) = 2, y(n) = n, y'(n) = O}.
Problems 191
6.33. F(y) = Ii[XV'(X)2 - 12xy(x)] dx,

t,
f2 = {y E C2[1,2]: y(1) = y'(1) = 3, y(2) = 134 + In 2, y'(2) = n.
6.34. (a) Let f E C 1 ([a, b] x \R3). Show that if y E C 3 [a, b] is a solution of the Euler
-Lagrange equation (21), then y must also satisfy a second equation
f(x) - y'(x) 0'(X) - :xJ;, .. (X)] - y"(x).t;,..(x) = r: fAt) dt + c,

for some constant c, where we employ the usual abbreviations.
(b) Use this to obtain a second-order equation replacing (25) and show that
assuming Iy'l ~ 1 in this equation gives us an integral of (28). Hint: see
Problem 6.30(b).
6.35*. The Second Equation (Proposition 6.3).
(a) Verify the transformation of the integral F(y) into F('1) under the substitu-
e
tion x = + C'1m, where 1 is defined by (10).
(b) Verify the retransformation of (11) into (11'), supposing that Yo is an
extremal point for F on f2.
(c) With y E f2, '1 E 2} and c as above, take v in f20 = {v E C 1 [a, b]: v(a) =
v(b) = O}, and define v by v(e) = v(e + C'1(O), so that v'm =
v'(x)(1 + C'1'(O). Show that v E 2}o = {v E C 1[ex, p]: v(ex) = v(P) = O}.
(d)* Prove that for this v, v,
bF('1; v) = J: c["['1(m v(e) + i.,.['1(m v'(O) de

= bF(y; v) + b2F(y; cV),
where
b2 F(y; v) = deff bU,,(x)v(x) + (f -
a y:t;,.)(x)v'(x)] dx.
(e)* Use the formula in (d) to discuss the invariance of C 1 stationarity of y

under this skew transformation.
6.36. The Second Equation for Vector Valued Functions. Assume that
fEC 1 ([a,b] x \R2d)
for d > 1.
(a) Derive the second equation (35) satisfied by a function Y E (C 2 [a, b]l
which is stationary for f
(b)* Obtain the second equation when Y E I1JI = (C 1 [a, b]l supplies a local
extremal point for
on
F(Y) = r f(Y(x)] dx
f2 = {Y E 11JI: Y(a) = A; Y(b) = B}.

Hint: Use the skew transformation
{ e= x + C'1, for H = ('11, '12, '13, ... , '1d)' and a small c

H=Y
to duplicate the proof of Proposition 6.3.
6.37. Letf E Cl([a, b] x jR2d). Show that the natural boundary condition associated
with minimizing
locally on
F(Y) = r f(x, Y(x), Y'(x)) dx
~l = {Y E Cl([a, b]l: Y(a) = A, Yi(b) = bi, i = 2,3, ... , d}

(where Yl (b) remains unspecified), is h, (b) = O.
6.38. Isoperimetric Problem. (See §6.7.) Show that A cannot achieve a (positive)
minimum value on ~ or ~*. [Reason geometrically.]
6.39. Prove that in jRd, a geodesic curve which can be parametrized by Y E (Cl[O, 1])d
with Y'(t) "# l!J, is a straight line segment. Hint: See §l.1(a), and recall that the
unit vector Y'/I Y'I is tangent to the curve.
6.40. For each of the following functions f defined on a subset of jR2d+1, write the
differential equations whose solutions Y will be stationary for f. Also, give an
example of an integral function F on a set !!}, which could have such Yas local
extrema.
(a) f(x, Y, Z) = x 2 + IYI 2 + 3z l , (d = 2).
(b) f(x, Y, Z) = xlZI, (d ~ 2).
(c) f(x, Y, Z) = YllZI - (sin Zl)Y2, (d = 3).
(d) f(x, Y, Z) = IZI/~, (d = 2).
(Do not attempt to solve the differential equations.)
6.41. In (a)-(c), find a partial differential equation which is satisfied by u E ~ ("\

C 2 (D), if u is a (local) extremal function for F on ~ = {u E Cl(D): UIOD = 'l'},
where D is a Green's domain in the x - Y plane and l' is a given continuous
function on aD. (Do not attempt to solve the equation.)
(a) F(u) = JD[t(U~ + u;) - (x 2 + y2)U] dA.
(b) F(u) = JD [t(u~ + u;) + tu dA. 3]
(c) F(u) = JD [u~ - u~ + u2 ] dA.
6.42. (a) Show that a function which maximizes
F(y) = I y(s)J1 - Y'(S)2 ds

on
~* = {y E Cl[O, 1]: y(O) = y(l) = 0; y(s) ~ O},
must satisfy the equation Y/~ = r = const.

(b) Solve the equation in (a) for y' and make the substitution y(s) = r sin O(s)
to conclude that O'(s) = 1/r, or O(s) = sir + c.
(c) Show that also Y E ~*, when y(s) = r sin (sir), where r = lin. Use x(s) =
J~ J1 - y'(tf dt to eliminate the parameter s and conclude that a max-
imizing function must describe a semicircle of radius r.
Problems 193
6.43. The Zenodoros Problem (in its nonisoperimetric formulation from Problem
1.9 (b)).
(a) Show that a function Yo which maximizes
F(y) = IaT y(t) [1 - (yY')(tf)]1/2 dt

on
9)* = {y E ClEO, T]: y(O) = y(T) = 0; y(t) ~ 0, lyy'l(t) < 1}
must satisfy the equation J(1 - (yy')2)jy - const. = c > O.

(b) Conclude that when Yo(O) = 0, then yo(t) = J2t - c2t 2, so that when
Yo E 9)*, yo(t) = J2t(1 - tT 1).
6.44. Over some future time interval [0, T], a strip-mining company intends to
remove all of the iron ore from a region that contains an estimated Q tons. As
it is extracted, they will sell it for processing at a net price per ton of p(y, y') =
p - exy - fly' for positive constants P, ex, and fl, where y(t) is the total tonnage
sold by time t. (This pricing model allows the cost of mining to increase with
the extent of the mined region and speed of production).
(a) If the company wishes to control its rate of production y'(t), to maximize
its total profit as represented by
F(y) = f: p(y, y')y' dt, when y(O) = 0, and y(T) = Q,
how might it proceed? (T is unspecified, but we need y ~ 0.)

(b) If future money is discounted continuously at a constant rate r, then we
can assess the present value of profits from this mining operation by
introducing a factor of e- rt in the integrand of (a). How will this affect
optimal mine operation?
PART TWO
ADVANCED TOPICS
Paris, 1900
AN ADDRESS
"As long as a branch of science offers an abundance of problems, so long is it

alive: a lack of problems foresh.adows extinction or the cessation of independent
development. Moreover, a mathematical problem should be difficult in order to
entice us, yet not completely inaccessible, lest it mock our efforts. The mathema-
ticians of past centuries . .. knew the value of difficult problems. I remind you
only of the 'problem of the [path] of quickest descent,' proposed by Johann
Bernoulli. The calculus of variations owes its origin to this and to similar
problems. ... it often happens also that the same special problem finds applica-
tion in the most unlike branches of mathematical knowledge. So, for example,
the problem of the shortest line plays a chief and historically important part in
the foundations of geometry, in the theory of lines and surfaces, in mechanics,
and in the calculus of variations . ... And it seems to me that the numerous and
surprising analogies and that apparently preestablished harmony which the
mathematician so often perceives in the questions, methods and ideas of the
various branches of his science, have their origin in [the] ever-recurring inter-
play between thought and experience. ... "
"It is an error to believe that rigor in the proof is the enemy of simplicity.
The very effort for rigor forces us to discover simpler methods of proof. ... But
the most striking example of my statement is the calculus of variations. The
treatment of the first and second variations of definite integrals required in
part extremely complicated calculations, and the process applied by the old
mathematicians had not the needful rigor. Weierstrass showed us the way to a
new and sure foundation. By the examples of the single and double integral I
will show briefly, at the close of my lecture, how this way leads at once to a
surprising simplification. ... "
195
196 Part 2. Advanced Topics
"Mathematical science is in my opinion an indivisible whole, an organism

whose vitality is conditioned upon the connection of its parts. The organic unity
of mathematics is inherent in the nature of this science, for mathematics is the
foundation of all exact knowledge of natural phenomena. That it may com-
pletely fulfill this high destiny, may the new century bring it gifted prophets and
many zealous and enthusiastic disciples!"
DAVID HILBERT
At the Second International Congress
of Mathematicians 1
1 These are fragments ofthe celebrated lecture in which Hilbert set forth 22 additional problems
which have challenged mathematicians of all disciplines in this century. The translation of the
complete text from which they were compiled will be found in Vol. 8 of the Bulletin of the A.M.S.
(1902) pp. 437-445, 478, 479.
CHAPTER 7
Piecewise C 1 Extremal Functions
In many problems examined thus far we have required continuous differ-

entiability of the function y (or Y) defining the classes for optimization.
Already with the example of the minimal surface of revolution from §1.4(a)
we have argued that for some configurations, the minimizing curve (if it
exists) should exhibit "corners," and it is natural to wonder whether cornered
curves y and iT such as those shown in Figure 7.1 might not give improved
results for other problems. Such curves are represented readily by functions
which are piecewise continuously differentiable, or piecewise C 1 •
In the present chapter we shall extend our previous investigation to in-
clude this class of functions as possible extremals. 1 In §7.2, we show that it is
only necessary to consider such extremals when C 1 extremals cannot be
found, and in §7.3 and §7.5 prove that in general such extremals are station-
ary functions in intervals excluding their corner points, at each of which they
must meet the transitional conditions of Weierstrass-Erdmann. The tech-
niques are illustrated by application to a Sturm-Liouville problem where
consideration of piecewise C 1 extremals provides valuable information. In
§7.4, we use [strong] convexity (as in §3.2) to guarantee minimality for the
local extremals of the associated integral functions and to study problems in
which interval constraints preclude simple stationarity. (Although it is possi-
ble to introduce piecewise C k functions for k = 1, 2, ... even in a multi-
dimensional setting-and obtain corresponding extensions to other results
from Chapter 6, we shall not do so here. See [Mol)
The above material is necessarily more technical but not appreciably more
sophisticated than that which it generalizes. However, the results presented
1 An older literature refers to these as "discontinuous" extremals. Extremals which exhibit actual
discontinuities have been investigated by Krotov. See [Pel
197
198 7. Piecewise C 1 Extremal Functions
a b x
Yt
Figure 7.1
in the remainder of this chapter, namely, Hilbert's differentiability criterion

(in §7.5) and the Weierstrass-Legendre conditions necessary for a minimum
(in §7.6) are based on somewhat more difficult concepts. This concluding
material should be regarded chiefly as a prelude to Chapter 9.
§7.1. Piecewise C 1 Functions

(7.1) Definition. A function y E C[a, b] is piecewise C 1 (denoted y E C1 [a, b])
provided that there is a finite (irreducible) partition a = Co < Cl < ... <
CN +1 = b such that y may be regarded as a function in C 1 [c k , CHI] for each
k = 0, 1, 2, ... , N. When present, the interior points Ck for k = 1, 2, ... , N are
called corner points of y.
When there are no corner points, y = y E C 1 [a, b]. Moreover, if y E
C 1 [a, b], then iyi E C1 [a, b] when y changes sign only finitely often.
Observe that P' is defined and continuous on [a, b] except at a corner point
c where it has distinct limiting values y'(c ±). We shall use P'(c) to denote both
values when the distinction is not essential. Discontinuities such as these of y'
are said to be simple, and functions such as y' are said to be piecewise
continuous on [a, b], or to belong to C[a, b]. Figure 7.2 illustrates the effect
of the discontinuities of y' in producing "corners" on the graph of y. Without
these corner points, y might resemble the C 1 function, y, whose graph is
presented for comparison.
A form of the fundamental theorem of calculus remains valid for C1

functions.
(7.2) Lemma. If y E c [a, b], then y(x) = y(a) + J: y'(t) dt.

1
§7.1. Piecewise C i Functions 199
-~
--- y
'\ -
--~'
~.-T----~---------+------+---------.----~-
a b x
Figure 7.2
PROOF. When x E [c k , CHi], then

k-i
P(x) - p(a) = P(x) - P(C k ) + I
j=O
[P(Cj +1) - P(Cj )]
= fX P'(t) dt + ~f f ej
+! P'(t) dt = fX P'(t) dt. D
Ck )-1 Cj a
(Conversely, if h is piecewise continuous on [a, b], then P(x) ~ S: h(t) dt

provides apE Ci [a, b], with P' = h. See A.8 et seq.)
(7.3) Proposition. Let PE Ci [a, b]:

(i) If S:P'(X)2 dx = 0, then pi = 0 on [a, b];
(ii) If pi = 0 where defined, then P = const. on [a, b].
PROOF. (i) Observe first that (P' )2 is nonnegative and piecewise continuous so
that the integral is defined and it may be represented as the finite sum
If=o S~~+! P'(X)2 dx. This sum vanishes iff each of its terms does, and then the
continuous function P'(X)2 = 0 on (C k , CHi) (from A.9), so that its continuous
extension vanishes on [c k , CHi], for k = 0, 1,2, ... , N.
(ii) This assertion is an immediate consequence of the preceding lemma.
D
(a) Smoothing
As Figure 7.2 illustrates, each piecewise C i function is "almost" C\ it is only

necessary to round out the corners to produce the graph of a function y such
as that shown. The next construction accomplishes this analytically and
provides control over the approximation.
(7.4) Smoothing Lemma. For each y E C1 [a, b] and ~ > 0, 3 y E C 1 [a, b] such
that y = y except in a ~-neighborhood of each corner point of y where
max \y'(x)\ ~ 4 max \y'(x)\. Thus max \y(x) - y(x)\ ~ A~ for a constant A
determined by y.
PROOF. Since y has at most a finite number, N, of corner points, it suffices
to explain its modification in the given ~-neighborhood of a typical corner
point c. We suppose that ~ is so small that this neighborhood excludes a
~-neighborhood of other corner points and the end points a, b. In this neigh-
borhood we replace the discontinuous function Y' by a continuous triangular
function, such as that shown in Figure 7.3, which is determined by its "height"
hat c and the values y'(c ± ~).
For any choice(s) of h at the corner point(s), the resulting function denoted
y' is continuous by construction so that the function y defined by
y(x)~ y(a) + r: y'(t) dt is in C1[a, b]
and its derivative is the function y' just constructed. It remains to select the
value(s) of h to effect the required estimates, and this choice is most readily
understood when y has only the single corner point c. Then, clearly, when
J:
x ~ c - ~: y'(x) = y'(x) so that y(x) = y(a) + y'(t) dt = y(x); to have
y(x) = y(x), for x ~ c + ~, it is only necessary to make
f-+~~ y'(t) dt = A~ ~ f-+~~ Y'(t) dt. (Why?)
Now, by elementary geometry, the signed area A~ corresponding to the inte-

gral on the left is given by
~
A~ = M + 2 [Y'(c -~) + y'(c + ~)],
and for the given ~, A~ = A~ provided that

h = A~ _ Y'(c - ~) + Y'(c + ~) .
~ 2
y' -p'
- - - y'
a b x
Figure 7.3
§7.1. Piecewise C 1 Functions 201
Thus with M' = max IY'I, it follows that
IA~I :::;; 1~:~ IP'(t)1 dt :::;; 2~M',

so that Ihl:::;; 3M' and 1y'(x)l:::;; M' + Ihl :::;; 4M' on [c - ~, c + ~], as is
evident from the figure.
A similar choice for h = hk is possible at each successive corner point Ck'

k = 1, 2, ... , N so that by construction and obvious estimates:
Iy(x) - P(x)1 :::;; LX Iy'(t) -

N fCk+~
P'(t) I dt:::;; rIy'(t) - P'(t)1 dt
= k~l Ck-~ Iy'(t) - P'(t)1 dt :::;; 2~(5M')N = A~,
where A = 10M'N is determined by p. o

We require also (Cl[a, b])d, the d-dimensional vector valued analogue
of Cl[a, b], consisting of those functions iT E (C[a, b])d with components
Pj E Cl [a, b], j = 1, 2, ... , d. The corner points of such iT are by definition
those of anyone of its components pj . l The above lemma can be applied to
each component of a given iT and shows that iT can be approximated by a
Y E (C l [a, b])d which agrees with it except in prescribed neighborhoods of its
corner points.
In each case, these sets of piecewise C l functions form linear spaces of
which the subsets of C l functions are subspaces. Indeed, it is obvious that the
constant multiple of one of these functions is another of the same kind, and
the sum of two such functions will exhibit the necessary piecewise continuous
differentiability with respect to a suitable partition of the underlying interval
[a, b]. Moreover, the dot product iT· V of two such functions will be in
Cl[a, b]. (Why?)
(b) Norms for C1

Since Cl[a, b] s;; C[a, b], it is evident that IPI ~ max IP(x)1 is a norm for
~l def
C [a, b] and IIPII = max(IP(x)1 + lY'(x)1) can be shown to be another [Prob-
lem 7.1]. Although the first does not supply control over the differentiability
properties of the functions p, it is valuable when approximating Pby y as in
Lemma 7.4.
Another choice for a norm permitting two functions which agree except in
small neighborhoods of their corner points to be "close" in norm, is the
1 When d > 1, the curve represented parametrically by a Y with a simultaneous corner point in
all of its components need not exhibit a corner. See Problem 7.13.
202 7. Piecewise C1 Extremal Functions
integral
IIPlll ~ f (IP(x)1 + W(x)1) dx.

~ ~ def ~
Similarly, we shall supply (C 1 [a, b])d with the norms I YI = max IY(x)l,
II YII ~ max(1 Y(x) I + IY'(x)I), and II Ylll ~ J~(I Y(x) I + IY'(x)l) dx.
The maximum norms, 1·1 and 11·11, are called the strong and the weak
norms, respectively, and the functions which are locally extremal with respect
to the first [second] of these are said to be strong [weak] extremal functions. 1
Since there are many more functions in a strong norm neighborhood of a
function than in a weak norm neighborhood [Why?], it is more difficult for
a function to be a strong extremal than it is to be a weak extremal Qocally).
This classification of extremals was introduced by Kneser, a student of
Weierstrass, in 1900; it is now a fundamental part of the theory ([K]).
The above norms are not independent, and they satisfy inequalities such
as the following:
(75) (a) AlP) ::;; IIP~ 1 ::;; (b - a) IIPI~ } where A = b - a .

(b) AIYI::;; IIYlll ::;;(b-a)IIYII, 1 +b-a'
d
(c) IIYII::;; I
j=l
IIPj ll,whereY=(Pl, ... ,Pd).
(Problems 7.1-7.2)
§7.2. Integral Functions on C1

When f = f(x, Y, Z) is continuous on [a, b] x 1R 2d , and YE rfY = (C 1 [a, b])d,
then f[Y(x)] = f(x, Y(x), Y'(x)) is piecewise continuous on [a, b] and has in
general (simple) discontinuities at the corner points of f. Thus
F(Y) = f f(x, Y(x), Y'(x)) dx = f f[Y(x)] dx
is defined and finite, since a partition of [a, b] reduces this integral to a finite
sum of integrals considered previously. Now in general F is not continuous
on tfJI with respect to the strong norm I I or the II 111 norm. (F is continuous
with respect to the weak norm II II. See Problem 7.3*.) However, il!J =
(C 1 [a, b])d ~ tfJI, and the values of F on tfJI can be approximated by those of F
on il!J as shown in the next
(7.6) Proposition. If fE C([a, b] x 1R 2d ), then for each Y E tfJI and B > 0,

3 a y. E il!J, with I Y - Y,J < B, for which IF(Y) - F(Y.)I < B.
1 Some authors reverse these designations for the norms, although there is uniform agreement
about that for the extremals.
§7.2. Integral Functions on C1 203
PROOF. We use the Smoothing Lemma 7.4, to replace each component Pj

of Y by a C 1 function Yj which agrees with it except in nonoverlapping
b-neighborhoods of each corner point where
lyj(x)1 ~ 4 max IPj(x) I ~ 411.Pj l
and
IYj(x) - 'pj(x) I ~ ~b,
so that IYix )I ~ II.M + ~b. Thus V x: IYj(x)l, lyj(x)l, l.Pj(x)I, l.Pj(x)I ~ mj =
4(~b + II.Pjll). The resulting function Y = (Yl' Y2" .. , Yd) E rJ.Y and it agrees
with Yexcept in b-neighborhoods of each of its corner points Ck' k = 1, 2, ... , N,
where the previous estimates hold. Denoting maxj=1,2, ... ,dmj by m, we con-
clude from Proposition 5.3 that If I is bounded on the compact box [a, b] x
[ - m, m]2d, by m, say, and f[Y(x)] = f[Y(x)] except when Ix - ckl ~ 15 for
some corner point Ck , k = 1, 2, ... , N. Hence
IF(Y) - F(Y)I = Iktl l:k_~d (f[Y(x)] - f[Y(x)]) dxl

~ 2bN max(lf[Y(x)] I + If[Y(x)] I) ~ 4Nmb.
~,,~ "d ~ ~
Also I Y - YI ~ L.,.j=ll'pj - Yjl ~ L.,.j=l Ajb = Ab, say, by 7.5 (c) and the
above estimate. The proposition follows for Y" = Y when 15 is sufficiently
small. 0
It is straightforward to show that in the case of the geodesic problems for

IRd and for the sphere, the arguments used in §1(a), (b) remain unchanged
when the functions Yare replaced by their piecewise C 1 counterparts f. Thus
we can assert that the geodesics in these cases remain the same, namely, the
straight lines and the great circles, respectively.
There is in fact a general principle evidenced by these examples which may
eliminate the search for piecewise C 1 extremals.
(7.7) Theorem. If fE C([a, b] x 1R 2d ) and Yo is a [local] extremal point for F

on q) = {Y E rJ.Y: Y(a) = A, Y(b) = B}, then Yo is also a [local] extremal point
for F on ~ = {Y E rJ.Y: Y(a) = A, Y(b) = B} [with respect to the same 1'1 or
11'11 norm].
PROOF. The first result is a direct consequence of the preceding proposition.
Indeed, if Yo E q) supplies, say, a minimum for F on q) in the strong neighbor-
hood Sd(YO) = {Y E rJ.Y: I Y - Yol < b}, and I Y - Yol < 15, then with Y" as in
the proposition, the triangle inequality shows that
F(Y) ~ F(Y,,) - IF(Y) - F(y")1
~ F(Y,,) - e ~ F(Yo) - e,
when e is so small that Y" E SiYo) (and this can be achieved; see Problem
7.20(a)). Thus as e '" 0, we conclude that F(Y) ~ F(Yo).
The case where Yo is a weak local minimum point is left to Problem 7.20(c).
o
(7.8) Remark. The previous characterizations of local C 1 extremals given in
Chapter 5 were with respect to an unspecified norm, but as there observed,
weak local extremals Yo need not be strong local extremals; see Bolza's exam-
ple in §7.6. However, in case Yo is a global extremal for F on fl), then the norm
considerations are immaterial and we can assert that Yo is also a global
extremal for F on flJ. In particular all of the minima obtained for the convex
problems of Chapter 3 also minimize in the corresponding classes of piece-
wise C 1 functions.
(Problem 7.3)
§7.3. Extremals in C1 [a, b]:

The Weierstrass-Erdmann Corner Conditions
Theorem 7.7 does not preclude a function F from being extremized by a Yo
which is only piecewise C 1 . For example, the function (of §6.2, Example 3)
f(y, z) = y2(1 - Z)2 gives
F('p) = fl 'p 2(x)(1 - 'p'(X))2 dx,
and F has a minimum value, 0, on

fl) = {.P E e 1[ -1,1]; 'p(-1) = 0, 'p(1) = 1},
which is attained (uniquely) by the function
A ( ) _ {O, - 1 :::;; x :::;; 0,
Yo x -
x, 0:::;; x:::;; 1,
with a corner point at 0. [To establish the uniqueness, observe that if F('p) =
° for a .p E.@, then the associated nonnegative integrand must vanish on
[ -1, 1] except possibly at the corner points of'p. Thus, at each such x, either
°
'p(x) = or 'p'(x) = 1. Now 'p(1) = 1, and in the largest subinterval (a, 1] in
°
which .p(x) > we must have that 'p'(x) = 1 and 'p(x) = x. This interval must
continue until a = 0; .P(O) = 0, and the only function in 1 [ -1, 0] which e
°
vanishes at -1 and and increases at each nonzero value is 'p(x) == 0.] On
the other hand, by 7.6 we know that F(.Po) can be approximated as closely as
we wish by values F(y) for
yE fl) = {y E C 1 [ -1; 1]: y( -1) = 0, y(1) = 1}.
Thus F cannot be minimized in .@, and this explains our previous lack of
success in §6.2, Example 3.
§7.3. Extremals in C1 [a, b]: The Weierstrass-Erdmann Corner Conditions 205
When we seek conditions which are necessary in order that a function

.Po E C1[a, b] should be a local extremal, we may assume at first, that it is a
weak local extremal, i.e., that it is extremal with respect to some weak neigh-
borhood of the form Sr(.PO) = {.pEC 1 [a,b]: 11.P-.Poll <r}. In fact, each
local extremal with respect to the strong 1'1 (or the 11'111) norm is automati-
cally a weak local extremal. (Why?)
Observe that .P E Sr(.PO) iff .P = .Po + ell for fJ E C1[a, b], and a sufficiently
small e. In characterizing (weak) local extremals for the function
on
F('p) =
~
I b
~1
f(x, .p(x), 'p'(x)) dx = r f[.P(x)] dx
!?} = { .P E C [a, b]: 'p(a) = a 1 , 'p(b) = b 1 } ,

where f = f(x, y, z) and its partials 1" fz are continuous on [a, b] x jR2, we
duplicate the analysis of §6.1, and will only sketch the results.
For'p E ~ and fJ E ~o = {fJ E C1 [a, b]: fJ(a) = fJ(b) = O}, we form

F('p + efJ) = I b
f[.P(x) + efJ(x)] dx;
then, taking into account the corner points of both .p and fJ, we represent it as
a finite sum of integrals with continuous integrands, and differentiate each
:e F('p r
under the integral sign (A.13) to get upon reassembly that
+ efJ) = [I,[(.P
Hence, in the limit as e -+ 0, we obtain

+ efJ)(x)]fJ(x) + fz[(.P + efJ)(x)]fJ'(x)] dx.
c5F('p; fJ) = I b
c/,(x)fJ(x) + /,.(x)fJ'(x)) dx, (1)
where
/,(x) ~ fy(x, 'p(x), 'p'(x)) = I, ['p(x)],
and
/,.(x) ~ fAx, 'p(x), .P'(x)) = fz[.P(x)]
are again piecewise continuous on [a, b] with at most (simple) discontinuities
J:
~ the corner points of .p. Thus b(x) ~ /,(t) dt, determines a function in
c5F('p; fJ) r
C 1 [a, b], and upon integrating by parts, we have as before that
= [/,.(x) - b(x)]fJ'(x) dx.
In order that .p be a local extremal function for F on ~, it must make

c5FU; fJ) = 0, V fJ E ~o, and, in particular, for that fJ defined by
fJ(x) = IX (/,.(t) - g(t) - co) dt,

which is in .@O for an appropriately chosen constant Co. As in the proof of

Lemma 4.1 of du Bois-Reymond, we have that J~ j}'(xf dx = 0, which gives
V' = 0 (Proposition 7.3(i)). Thus
h'(x) = LX h(t) dt + co, (2)
is continuous, and in addition,

d A }
dxh'(X) exists = Jy(x), (2')
except at each corner point c of y where from the aforementioned continuity:

h'(c-) = h,(c+). (3)
(3) is the first of the Weierstrass-Erdmann conditions which must prevail at
each corner point of a piecewise C 1 local extremal function. On each interval
that excludes corner points, the local extremal function y must be C1 and
stationary for f in the sense of the previous chapter.
(7.9) Remark. Actually, the integral equation (2) could be used to character-
ize stationarity for F on .@, and many of the subsequent properties obtained
in this chapter for extremal functions yare also true for this larger class. See,
in particular, Theorem 7.12 and its proof. For example, each solution y of (2)
satisfies (3) at a corner point, and therefore (see Problem 7.25(a))
At a corner point c of a solution y of (2), the double derivative
if defined, must vanish for some value of z.
!zAc, y(c), z),
[This is a simple consequence of the law of the mean (A.3) applied to the
function g(z) = !z(c, y(c), z), which by (3) has equal values at the distinct points
z = y'(c ±) and so must have a vanishing derivative g'(z) = !zAc, y(c), z) at an
intermediate point.]
Similarly, when fEel ([a, b] x 1R2) we may duplicate from §6.3, the deri-
vation of the second Euler-Lagrange equation in integral form, to conclude
that a local extremal y for F on .@ must satisfy
!(x) - y'(x)h'(x) = LX !At) dt + const., (4)

where
!(x) = f[y(x)], and !x(x) = fx[Y(x)],
Thus (d/dx)(! - yi;,,)(x) exists = !Ax), V x E (a, b) except at each corner
point c of y at which holds the second Weierstrass-Erdmann condition:
(/ - yi;,,)(c-) = (! - yi;,,)(c+). (4')
If we use (3), and let w = y'(c - ), z = y' (c + ), we can restate (4') as follows,
f(c, y(c), w) - f(c, y(c), z) - fAc, y(c), z)(w - z) = 0; (4")
and since w =F z, we see that neither f(c, y(c),· (nor - f(c, y(c), .) can be
strictly convex. (Recall Definition 3.4). This fact can sometimes be used to
locate or even preclude corner points.
For example, the function f~, ~, z) = ~2 + t)~ is strictly convex
(in z) except when x 2 + y2 = O. Therefore an associated local extremal y
cannot have a corner point except possibly where c = Y(c) = O. The function
f(y, z) = (1 + y2)Z4 cannot have local extremals with corner points, i.e., each
local extremal is C 1 , because (1 + y2)Z4 is strictly convex in z. In this case, the
test of Remark 7.9 fails since fzAy, 0) = O. On the other hand, by the same
test, when f(y, z) = eY ~ every solution of (2) is C 1 , whether or not it
gives a local extremal. (Recall Example 3 of §3.3)
With a generalization similar to that used to obtain Theorem 6.8, we get
the following:
(7.10) Theorem. For a domain D of [R2, suppose that
f = f(x, y, z) E C 1 ([a, b] x D),
and y E tfJI = 61 [a, b] provides a (weak) local extremal value for
on
F(y) = r f[y(x)] dx
~ = {y E 0/;: Y(a) = a 1 , Y(b) = b1 ; (y(x), y'(x)) ED}.

Then except at its corner points, y is C 1 and satisfies the first and second
Euler-Lagrange equations (2') and (4). At each corner point c:
(i) h'(c-) = h'(c+); and

(ii) (/ - y'h,)(c - ) = (/ - Y'h,)(c + );
(iii) ±f(c, y(c), z) cannot be strictly convex in z.
PROOF. See 7.13 in §7.S. o
The Weierstrass-Erdmann conditions (i) and (ii) show that the disconti-
nuities of y' which are permitted at corner points of a local extremal y, are
limited to those which preserve the continuity of both h' and (/ - y'h').
Observe that by (4), when fx == 0, then the latter term is constant.
Example 1. A piecewise C 1 extremal, y, for the brachistochrone problem of

§6.2(c), where
~ z
f(x, y, z) = Jy and fAx, y, z) = Jy~
y y 1 + Z2
should be a stationary function (and hence represent a cycloid) on each
interval excluding corner points, at each of which we require the continu-
ityof
p' , 1
h' = fiJI + p,2 and ! - pt, = fiJI + p,2'
while the latter function is in fact constant since Ix == 0. Thus we require the
continuity of P', so that even on mathematical grounds this problem can have
°
only C 1 extremal functions. (An exception occurs at x = where P(O) =
and P'(O) = + 00. There, an extreme form of a corner is permitted at the cusp
°
of the cylcoid.)
Example 2. For the function F which introduces this section, where

I(x, y, z) = y2(1 - zf, and fz(x, y, z) = 2y2(Z - 1).
a piecewise C 1 extremal function p must be stationary in intervals excluding
corner points at which both
h' = -2p2(1 - P) and ! - p'j" = p2(1 - pr2)
are continuous, while the latter is constant over the interval. Since p is contin-
uous, from the first condition it follows that p' is also continuous except
perhaps at a point c where P(c) = 0, and these are the only possible corner
points. Thus unless p vanishes at some point in an interval [a, b], it is not a
local extremal; (moreover, as we have seen in §6.2, Example 3, there need not
be any local extremals in C1[a, b] which satisfy certain boundary conditions).
However, if p vanishes at a single point in [a, b], then from the second
°
condition, p2(1 - P,2) == so that at each x E [a, b]: either P(x) = 0, or
P'(x) = 1, or P'(x) = -1.
Unless p == 0, the subintervals in which p vanishes identically can only
terminate at corner points beyond which p is linear with the slope 1 or the
slope -1. Thus the most general piecewise C 1 (local) extremal function which
vanishes at some point on an interval [a, b] is one which vanishes on a single
subinterval (which may reduce to a point), beyond which it rises or falls with
unit slope. We have already encountered one function of this type on the
interval [-1, 1], viz., Po(x) = {a,x,xx ;::::::::;; 0°. Others are shown in Figure 7.4.
Observe that there cannot be more than one such function having given
boundary values at a, b, but there are none with IP(a)1 (or IP(b)l) > b - a.
Figure 7.4
-Po
y --P
- r(x, y) = 0
a c x
Figure 7.5
Natural boundary conditions corresponding to the various free end point

constraints considered in §6.4 remain the same; viz., (12a), (12b), and (IS),
respectively, where, of course, J;.) replaces 1(.) in each instance.
To see this most easily when, say, freedom is permitted only at the right
end point, as in Figure 7.5, suppose that .Po E C1 [a, b] provides a local
extremal for F, and let c be the right-most corner point of .Po. Then restricting
comparison to those competing'p which also have their last corner point at c
and satisfy 'p(c) = .Po (c), it is seen that the corresponding directions (} must
utilize the end-point freedom exactly as in §6.4. Thus the resulting natural
boundary conditions are the same.
(7.11) Remark. In order to apply the method of Lagrangian multipliers of

§S.7 to piecewise C 1 extremals with constraints, it is, in general, necessary to
use the weak norm II I in order to have the requisite weak continuity of the
s:
variations c5F('p; (}) = [!,(x)(}(x) + !,.(x)(}'(x)] dx. Again, our initial obser-
vation that each local extremal of interest is necessarily a weak local extremal
makes this permissible, and results in the now expected generalization of §6.S;
viz., the existence of Lagrangian multipliers determining a modified function
with respect to which an extremal function is stationary in intervals exclud-
ing corner points; at each corner point it satisfies the Weierstrass-Erdmann
conditions for the modified function.
Application: A Sturm-Liouville Problem
Given functions p, q, 'l' E C[a, b], with 'l' > 0 on [a, b], suppose that
.Po E C1[a, b] minimizes
F('p) ~ r ['l'(x)'p'(xf + q(X)'p(X)2] dx (S)

on the subspace
@"o = {y E 61 [a, b]: y(a) = y'(b) = O},
under the isoperimetric constraint
G(y) ~ r p(X)y2(X) dx = 1.
Then, c5G(yo; Yo) = 2 J~ p(x)y5(x) dx =F 0 and the first alternative of Theo-

rem 5.15 is excluded. Thus with Remark 5.17, 3 a constant (here denoted -A)
such that c5F(yo; (}) = 0, V V E @"o, where F = F - AG.
Introducing
j(x, y, z) = -r(X)Z2 + [q(x) - Ap(X)]y2,
it follows from Theorem 7.10 that at a possible corner point c, h[Yo(x)] =
2-r(x)y~(x) is continuous. But since -r is continuous and positive at c, this
implies that y~ itself is continuous. Thus Yo = Yo E C 1 [a, b] is stationary for
j on (a, b), and so Yo is a solution of the homogeneous linear differential
equation
d
dx (-ry')(x) = (q - Ap)(X)Y(x), X E (a, b), (5')
which satisfies homogeneous boundary conditions (such as) y(a) = y'(b) = o.

This type of problem was studied extensively by Sturm and Liouville
(c. 1835). It always has the solution y = (!). Each A for which it has a nontrivial
C 1 solution is called an eigenvalue (or characteristic value) of the problem,
and each associated solution y =F (!) is called an eigenfunction. Clearly, for
each constant c =F 0, cy is another eigenfunction for the same eigenvalue, and
if G(y) > 0, then c may be chosen (within a sign) to make G(cy) = 1. From
uniqueness of solution of (5'), y and y' cannot vanish simultaneously. See
Problem 7.21.
In particular, if our supposed minimizing function Yo =F (!), then it is an

eigenfunction for the above Sturm-Liouville problem and the Lagrangian
multiplier Ais the associated eigenvalue. Now Yo changes sign at most finitely
often on [a, b] (§A1) so that IYol E @"o. Since IYoI2(X) = y~(x), while IYol'(X)2 =
(y~)2(X), it follows that F(IYol) = F(yo) and G(lyol) = G(yo). Thus IYol is also
a minimizing function, but then, IYol E C 1 [a, b] by the argument used before,
and hence IYol is another eigenfunction for the same value A.
From this fact we can conclude that a minimizing eigenfunction Yo enjoys
a special distinction: If x E (a, b) then Yo(x) =F o. [For if Yo(x) = 0, then also
y~(x) = 0, since otherwise IYol would have a corner point at x. But this is not
possible.
Conversely, by using an ingenious computation from Picard (1896), we
can prove that an eigenfunction Y1 which is nonvanishing on (a, b) does mini-
§7.4. Minimization Through Convexity 211
mize F under the given conditions (Problem 7.19). The physical origin of the
Sturm-Liouville problem and the significance of this fact will be discussed in
§8.9.
(Problems 7.4-7.9)
§7.4. Minimization Through Convexity

When the function f~, y, z) is convex on say [a, b] x ~2 as in §3.3 then in
general, each local extremum becomes a global minimum, so that the distinc-
tion between weak and strong local extrema is superfluous. For it follows
that
is convex on
F(P) = r f[y(x)] dx
~ = {y E CI[a, b]: y(a) = a l ; y(b) = bd

for given aI' bi E ~, since from 3.4 and (1),
F(y
~ 0,
r
+ ti) - F(.P) - c5F('p; ti)
{f[Y(X)
VY E ~
+ v(x)] - f[y(x)] - (fy[y(x)]ti(x) + fz[y(X)]ti'(X»} dx
and v E ~o = {ti E CI[a, b]: ti(a) = v(b) = O}. (6)
Moreover, when fUs, y, z) is strongly convex, then F is strictly convex on ~.
[Indeed, then equality in (6) is possible only if v(x) or v'(x) = 0, V x except
at the corner points of y, V. This is seen by representing the integral in (6)
as a finite sum of integrals with continuous nonnegative integrands to each
of which may be applied the earlier argument from §3.2. Hence ti2(X) is con-
tinuous and piecewise constant on [a, b]. It follows that ti 2(x) = const. =
v2 (a) = 0, V V E ~o.]
By an analogous argument we may extend Theorem 3.5 as follows:
(7.12) Theorem. If fUs, y, z) is [strongly] convex on [a, b] x D for a domain

D f; ~2 then a
y E ~ = {y E CI [a, b]: y(a) = aI' y(b) = bl ; (.P(x).P'(x» ED},
which is stationary for f in intervals excluding corner points, I at each of which
it satisfies condition 7.10(i), minimizes F on ~ [uniquely].
PROOF. Since the hypotheses on f imply that F is [strictly] convex on~, we
need only verify that c5F(y; v) = 0, V v for which y + v E ~. The hypotheses on
!"
y assure that is in 5!J = CI [a, b] with the derivative !,. (Why?) Thus for
1 Corner points cannot occur when f0s, ~, z) is strongly convex by 7.10(iii).
r
each () E OJJ, the product /y,() is also in OJJ and by (1), and 7.2,
c5F(y; () = [/y(x)()(x) + /y,(x)()'(x)J dx
= r:x [/y,(x)()(x)] dx
= /y,(x) ()(x) = 0, I:
since ()(a) = ()(b) = 0, when y + () E ~o. o
There is a corresponding version of Proposition 3.9. (Problem 7.12).
Internal Constraints
Convexity may play an important role in problems involving internal point
constraints such as that of the following.
Example 1. To find the possible local extremal functions for:
on
F(y) = f: y'(X)2 dx
~
= {y E !lJJ = C1 [0, 2]: y(O) = y(2) = 1; y(l) = O},
which has the internal constraint y(l) = 0, we recall from §3.3, Example 1,
that J(z) = Z2 is strictly convex on lIt
Moreover, when y E ~ then y + () E ~ iff () E ~o = {() E !lJJ: ()(O) = ()(2) = 0;
0(1) = O} and then by the usual argument
F(y + () - F(P) ~ c5F(y; () = I2 2y'(x)()'(x) dx,
(with equality only if () = (0). But clearly
y'(x) = {cc 1, on [0, 1),

on (1, 2J,
2,
gives
Thus we conclude from the strict convexity of F on ~, that the piecewise

linear function Yo E ~ given by
A ( ) _ {I - x, 0::;; x ::;; 1,
Yo x -
x-I, 1::;; x::;; 2,
is the only local extremal function for F on ~, and it minimizes F on ~
uniquely, by 7.12.
§7.4. Minimization Through Convexity 213
Observe that Yo is clearly in CI but it does not satisfy the Weierstrass-

Erdmann condition of equation (3) at its corner point x = 1. This is because
the corner point in question is forced on the extremal function at the outset,
while the analysis leading to equation (3) permitted the function to seek its
own "natural" corner point(s).
The foregoing analysis extends in principle to any finite number of internal
point constraints, and to other functions f (Problems 7.17, 7.18).
For some problems involving inequality constraints of the type considered

in §2.3, the extremal function may be forced to satisfy a condition precluding
its stationarity along an entire subinterval.
We suppose that f~, y, z) is [strongly] convex as before, and that we wish
to minimize
F(y) = Lb f[y(x)] dx
on
~ = {y E CI[a, b]: y(a) = aI' y(b) = bd,
subject to the inequality constraint
g[.P(x)] ~ 0, X E (a, b),
where g~, y) is also convex.
As suggested by Proposition 2.5 and Remark 3.17, we introduce an un-
specified but nonnegative A E C[a, b], and consider instead the problem of
minimizing
F(.P) =
Specifically, we seek Yo
rE
(f[y(x)] + A(X)g[.P(x)]) dx on~.
~ which accomplishes this for a A ~ 0 with
A(X)g[yO(X)] == 0 on (a, b).
Then it would follow from Proposition 2.5 and the [strong] convexity of
k, y, z) = f~, y, z) + A~g(b y)
that Yo minimizes F [uniquely] on ~ under the given inequality constraint.
Thus we desire that in intervals excluding a corner point, Yo should satisfy

the equation of stationarity for j; viz.,
:xfz[Y(X)] - f,[y(x)] = A(X)gy[y(x)], (7)
where A ~ 0 is required to make

A(X)g[yO(X)] == o. (7')
In particular, in those intervals where g[yo(x)] ¥- 0, A must vanish and Yo is
stationary for f. However, we now permit intervals with g[yo(x)] == 0 and
Yi (b, bd
y> t(x) = x 2
a x
Figure 7.6
A -# 0, where Yo is not stationary for f Moreover, since.l. = fz' Yo has only

those corner points permitted by f (Theorem 7.10).
To fix ideas let's consider the following obstacle problem:
F(y) = r
Example 2. Minimize the strictly convex distance function
J1 + y'(X)2 dx
on?J as above under the constraint jl(x) ::;; x 2 . Then, as is evident graphically
from Figure 7.6, for some locations of the end points we must permit
a portion of the minimizing curve to lie along the parabola defined by
g(x, y) = y - x 2 = O. Clearly g~ y) is convex (as is -g~ y», but the sign of
A is still important. Here f(x, y, z) = ~ and j(x, y, z) = ~ +
A(X)(Y - x 2 ).
For this (nonstationary) part of the curve, y(x) = x 2 , and since gy == 1
while J;, == 0, we use (7), the resulting stationarity equation for j, to define
A(X) =!£ y'(x) =!£( 2x ) >0

dx J1 + y'(X)2 dx J1 + 4x 2 - .
For the remaining parts, which will be segments (Why?), we have A = 0. 1 By
taking these segments to be tangential to the parabola at their points of
contact, the resulting function Yo will in fact be C 1 • The foregoing analysis
1 The resulting A. is only piecewise continuous, but since 1. = J., the usual convexity arguments
remain valid. See Problem 7.26.
§7.5. Piecewise C 1 Vector-Valued Extremals 215
guarantees that it is the unique minimizing function for the problem. (Alterna-
tively, we may argue that since.1zz = fzz > 0, Yo cannot have corner points by
Theorem 7.10(iii).)
We have just shown how to prove that the natural conjecture for the curve
of least length joining fixed points in the presence of a parabolic barrier is
the correct one. The reader will find it instructive to analyze this geodesic
problem with barriers of other shapes where more than one subarc may be
required to lie along the barrier curve. It is more difficult to obtain necessary
conditions for problems of this type, and in particular to investigate the
behavior at the points of contact. See Problem 7.22 and the discussions in
[PeJ and [Sm].
(Problems 7.23, 7.24)
§7.5. Piecewise C 1 Vector-Valued Extremals

When f = f(x, Y, Z) C1([a, bJ x ~2d) the function
r r
E
F(Y) = f[Y(x)] dx = f(x, Y(x), Y'(x)) dx
is defined V Y E c1JJ = (C1[a, bJ)d, the linear space of piecewise C 1 vector-

valued functions Y = (Yl' Y2' ... , Yd) with the weak norm I YII of §7.1(b).
As in §6.7 we may analyze each component of a weak local extremal
function, Y, for F on
~ = {Y E c1JJ: Y(a) = A, Y(b) = B},
and conclude that at each corner point c of the component Yi we must have
the (first) Weierstrass-Erdmann condition
!,./c -) = !,j(c +),
where, of course,
j=1,2, ... ,d.
Thus at a corner point c, with the obvious notation:
/Y'(c - ) = /Y'(c + ). (8)
In each subinterval excluding the corner points of a given component Yi'
Y satisfies the jth stationarity condition
d A A ~
-d
x f, 1.. (x) = f,.(x) = !'y.[Y(x)],
1 1
so that in the subintervals excluding its corner points, Yis C 1 and stationary
and satisfies the Euler- Lagrange equation
(8')
Observe that (8) and (8') may be replaced by the equivalent vector valued
integral equation:
(9)
for some constant vector Co.
When f E C1([a, b] x /R2d) there is a corresponding version of the second

Euler-Lagrange equation (§6.7) which now takes the form
!(x) - Y'(x) -!r(x) = f: !x(t) dt + const., (10)

where
!(x) = f(x, Y(x), Y'(x)) = f[Y(x)]
and
h(x) = fAx, Y(x), Y'(x)) = fx[Y(x)]
are well-defined when x is not a corner point of f. The continuity in (10) may
be used to conclude that at a corner point c, a (local) extremal Y satisfies the
second Weierstrass-Erdmann condition:
(! - Y"!y,)(c-) = (/ - Y'-!y,)(c+). (11)
In each interval excluding corner points, the integrand in (10) is continu-
ous so that Y satisfies the second equation:
:x (! - Y" /r.)(x) = !Ax). (12)
Again with an obvious generalization there follows the
(7.13) Theorem. For a domain D of /R2d, let f = f(x, Y, Z) E C1([a, b] x D),
on
F(Y) = r
and suppose that Y provides a local extremal value for
f[Y(x)] dx
~ = {Y E (C1[a, b])d: Y(a) = A, Y(b) = B; (Y(x), Y'(x)) ED}.
Then except at its corner points, Y is C 1 and satisfies the first and second
Euler-Lagrange equations, (8') and (12). At each corner point, c, Y meets the
Weierstrass- Erdmann conditions:
§7.5. Piecewise C 1 Vector-Valued Extremals 217
(i) jy.(c-) = jy.(c+); and

(ii) (j - Y'-!Y')(c-) = (j - Y'-!Y')(c+).
(iii) ±f(c, Y(c), Z) cannot be strictly convex in Z.
Observe that when c is a corner point for only one component Yj, then as
before we may infer from (i) that if defined, each fz.z .(c, Y(c), Z) must vanish,
for some Z E ~d; (ii) reduces to requiring the conti~uity of only (j - yj 'hi)
at c; and (iii) becomes a statement of nonconvexity in Zj only. Finally, when
fx == 0, then (10) shows that (j - y,.jY') is constant.
PROOF. When Yo = Y, the radius function r(x) used in proving Theorem 6.8 is
piecewise continuous and so it again has a positive minimum value on [a, b]
guaranteeing a sufficient supply of ~-admissible directions at Yo. For (iii) see
Problem 7.25. D
It is straightforward to extend to vector-valued piecewise C 1 extremals the

method of Lagrangian multipliers (cf. Remark 7.11) and the use of convexity
in minimization problems (§7.4). (Problem 7.10.)
Application: Minimal Surface of Revolution
In 1744, Euler published a solution to the following problem: Given positive

numbers a 1 and b1 , find the planar curve joining points (a, ad and (b, b1 )
which when revolved about the x axis will have the minimum surface area.
(See §1.4(a).)
We suppose as we may that a = 0, a 1 = 1 and that bl ~ 1, so that the
configuration is as shown in Figure 7.7(a).
Then, supposing that a meridianal curve admits parametric representation
by a function Y = (x, y) E (Cl[O, 1])2, the resulting surface area is (from
elementary calculus) given by
211: Il y(t) ds(t) = 211: Il y(t) IY'(t)1 dt.
y y y
b x (r) b x (r) b x
(a) (b) (c)
Figure 7.7
However, as we have seen with similar integrals, the minimizing curve (if it
exists) may well have corners, and indeed, it is almost intuitively clear that
when b ~ b1 , the minimum area arises from the limit of curves such as shown
in Figure 7.7(b) and thus should be given by the degenerate three segment
curve of Figure 7.7(c).
Hence, ignoring the constant 2n, we are led to formulate the problem as
follows: Given positive numbers b, and b1 ~ 1, minimize
F(Y) = I1 y(t) IY'(t)1 dt

on
~ ~ {Y = (x, y) E (C 1 [0, 1])2: y(o) = (0, 1), Y(1) = (b, b1 ); with
IY'I =1= ° and y ~ 0, on [0, 1]}.
Here f(t, Y, Z) = y(zI + zn so that if YE ~ minimizes F (locally) on ~, we
1/2 ,
should expect that except at corner points it satisfies the Euler-Lagrange

equation(s) (9); viz.,
~[YX'J
dx IY'I
= 0·
' dt IY'I
~[ytJ = I f'I, (13)
and at a corner point, c, each bracketed term in these equations is continu-

ous. (However, see Problem 7.14(a).)
It follows that the first term
yx'
_______ = const. = Co say throughout [0, 1].
IY'I
°
Thus, if p or x' vanishes at any point in [0, 1], then Co = so that either p or
x' must vanish at each noncorner point.
A little experimentation shows that the only curve of this type repre-
sentable by functions in ~ is that consisting of three segments, shown in
Figure 7.7(c), for which the surface area is seen to degenerate to that of the
two end disks of Figure 7.8(a). This solution is attributed to its discoverer
y y
(b)
Figure 7.8
§7.5. Piecewise C I Vector-Valued Extremals 219
B. C. W. Goldschmidt (1831). When b is much larger than bi (~1), we have

already shown intuitively why it should supply the minimum sought.
However, when b is much smaller than 1 (:::;; btl, then the Goldschmidt
solution should not provide the minimum area, since then the area of the two
end disks will exceed that of the lateral area of the frustrum of the cone
having them as bases (Figure 7.8(b).) Thus we must consider the remaining
possibility: that Co =1= o.
When Co =1= 0, then neither y nor X' can vanish anywhere, and by (13) the
unit vector Y' / I Y'I is continuous at each corner point. But this means that the
direction of Y' is continuous, so that Y represents a curve without corners. We
shall assume that Y = Y = (x, y) E (C l [0, IJ)2. (However, see Problem 7.14).
Upon multiplying the resulting second equation (13)
(d/dt)[yy'/IY'IJ = IY'I
by its bracketed term, it may be integrated to give (yy'/I Y'lf = y2 + Cl for a
constant Cl , which together with (yx'/I Y'1)2 = c6, shows that Cl = -C6 and
( y')2 = y2 _ 1. (13')
x' c6
Since x'(t) is continuous and nonvanishing, it must be positive in order to
J6
have x(O) = 0 and x(l) = x'(t) dt = b > o. It follows that x(t) has an in-
verse in C l [0, bJ, and we can suppose the curve to be parametrized with
respect to the variable x = x(t).
For this parameter, x' = 1, and the resulting differential equation for
y = y(x) is with C = Co,
y'(X)2 = e~X)r - 1, (14)
(which may be recognized as the second Euler-Lagrange equation of the

function f(x, y, z) = y..j1+7 for the nonparametric minimal surface of rev-
olution problem. See Problem 6.16.)
Since y2(X) ~ c2, it is convenient to make the hyperbolic substitution:
y(x) = c(cosh '1(x)), so that y'(x) = c sinh '1(x)'1'(x), and the differential equa-
tion for '1 is
or
C2'1,2 = 1, for '1 =1= O.
Thus '1(x) = c-lx + Il, so that the solution of (14) is
y(x) = C cosh(c- l X + Il), (14')
for some constants c =1= 0 and Il to be determined. The boundary conditions
for y = y(x) are:
(i) y(O) = 1; and
(ii) y(b) = bl .
220 7. Piecewise C l Extremal Functions
To satisfy (i) we must have c cosh J1. = 1, so that

y(x) = cosh(x cosh J1. + J1.).
cosh J1.
As defined, y(x) is positive on [0, b]. We wish to choose J1., if possible to
satisfy (ii), i.e., to have
(b) = bl = cosh(b cosh J1. + J1.). (15)

Y cosh J1.
However, since cosh t > Itl, V t E IR, in order that (15) be solvable, it is neces-
sary to have
b > (b cosh J1. + J1.) ~ b cosh J1. -1J1.1 = b _ (_IJ1._I_), or b > b _ 1.
1 cosh J1. cosh J1. cosh J1. 1
Hence when bl ~ b - 1, there are no C I local extremal functions, and the

only possible minimizing functions for F on f!} are those which describe the
Goldschmidt curve.
It is more delicate to find conditions on b, bl which permit solutions of (15)
and hence local extremals other than the Goldschmidt curve. Sets of usable
values surely exist. For example, given b > 0, the choice J1. = 0 will satisfy (15)
when bl ~ cosh b. For these values, the curve defined by
y(x) = cosh x, 0 ~ x ~ b,
( or parametnca ( ) = bt, h b 0 ~ t ~ 1) representmg

. 11y b y {X(t) . t he catenary
y t = cos t,
shown in Figure 9.10 will provide a possible local minimal surface area.
More general conditions are known, and some permit two distinct solu-
tions of (15), and two possible minimizing catenaries such as those shown in
Figure 9.10. (See Problem 7.15.) We reserve discussion of actual minimality
of these curves until §9.6. (See [BI].)
Hilbert's Differentiability Criterion *

We have already noted in 7.10(iii) that when present, second derivatives of f
may provide useful information about the location of corner points of an
extremal function. We shall now see that at a noncorner point a condition on
the matrix fzz of the second partial derivatives fz.z.,
, J
i, j = 1, 2, ... , d, may
guarantee higher differentiability of an extremal function. In fact we have the
following result:
(7.14) Theorem (Hilbert). If fz E Cl, and YE (C I [a, b])d is a solution of the

integral equation
§7.6*. Conditions Necessary for a Local Minimum 221
then Y is C 2 in a neighborhood of each noncorner point Xo at which the matrix

fzz[Y(x o)] is invertible.
PROOF. Through implicit function theory ([Ed]), the vector valued integral
equation
fz(x, Y(x), Z) = LX jy(t) dt + Co (16)
determines Z as a unique C 1 function of x with Z(xo) = Y'(x o), in a neighbor-

hood of such Xo provided that both sides of (16) are C 1 (in (x, Z)) in a
neighborhood of (xo, Y'(x o)) and fzz[Y(x o)] is invertible.
These latter conditions are assured by the hypotheses. [Indeed
a ~ ~ ~
+ jf:1 fzyj(x,
d
axfz(x, Y(x), Z) = fzx(x, Y(x), Z) Y(x), Z)yJ(x)),
while
ax
a IXa fy(t) dt = fy(X) = fy(x,
A A ~ ~
Y(x), Y'(x)),
and both of these functions are continuous near the noncorner point Xo'
Finally, fz.z.(x,
, J
Y(x), Z) is continuous by hypothesis.]
~
Since Z(x) = Y'(x) will surely be one local solution to equation (16),
uniqueness guarantees that it is the only solution, and since Z is C 1 , it follows
that Y is C 2 in this neighborhood. 0
Remark. The continuous differentiability of fz is required only in a neighbor-

hood of (xo, Y(x o), Y'(x o)).
Observe that when d = 1, the invertibility of the matrix fzz is reduced to

the nonvanishing of the single term fzz[Y(x o)]' For example, the only sta-
tionary functions for f(x, y, z) = eY Jl+?
are necessarily C 2 since fzz > 0
and by 7.9 they cannot have corner points. On the other hand, the example
f(x, y, z) = y2(1 - Z)2 of §7.3 shows that corner points can be present where
fzAx, y, z) = 2y2 = O. A minimizing function which is only C 1 is given in
Problem 7.16.
§7.6*. Conditions Necessary for a Local Minimum

Heretofore, the conditions obtained in this chapter which are necessary in
r r
order that a function minimize
F(Y) = f[Y(x)] dx = f(x, Y(x), Y'(x) dx

locally on
?) = {Y E rfJI = (C1[a, b])d: Y(a) = A, Y(b) = B},
(in particular, (8), (9), and (10», actually characterize stationarity and do
not distinguish between maximal, minimal, or saddle point behavior-even
locally.
The early efforts (notably by Legendre in 1786) to characterize local
minimality utilized a second variation and resulted in conditions involving
the second derivatives of f However, in his lectures (c. 1879), Weierstrass
showed that a condition involving the first derivatives was even more
significant.
(a) The Weierstrass Condition
The Gateaux variations of an integral function are obtained by comparing its

values at a point Y with those at points Y + eV in a weak norm neighbor-
hood. In contrast to these (weak) variations, we wish to consider a new type
of (strong) variation by functions whose smallness does not imply that of their
derivatives-at least not at a given point, ~ E (a, b).
Specifically, given Y E r1JJ, take (~, Y(m = (0, lP) for convenience. Then for
fixed W E ~d, small h > 0 and e E (0, 1), let e = e/(l - e) and define the asso-
ciated strong variation WE (Y as follows: when
o ~ x ~ eh: W(x) = xW so W'(x) = W, }

eh ~ x ~ h: ~(x) = e(h - x)W so ~'(x) = -eW, (17)
x < 0 or x> h: W(x) =0 so W'(x) = O.
Graphs of Ware indicated in Figure 7.9. Observe that as h '" 0, I WI -+ 0,
+
but I WII 0 (unless W = lP). Therefore, if Yminimizes F on §) locally with
Graphs of W
w (e =~)
Yl
Figure 7.9
respect to the strong 1·1 norm, then for small h we have
° S; F(Y + W) - F(Y) = f: (f[Y(x) + W(x)] - f[Y(x)]) dx
=h fal (f[Y(sh) + W(sh)] - f[Y(sh)] ds,

where we have substituted x = hs (so dx = hds) to obtain the last integral.
Now, as h '" 0, we see from (17) that the first term of this integrand has
different constant limiting values according to whether s < e or s > e.
If we divide by h and take the limit we find that
° S; e[f(O,
def
(P, E
~
+ W) - f(O, (P, E] + (1 - e) [f(0, (P, E - liW) - f(O, (P, E)],
where E = Y'(O+).
If we now divide by e and take the limit as Ii '" 0, we can use the chain rule
to see that
°
S; [f(0, (P, E + W) - f(O, (P, E)] - fz(O, (P, E) . W
Finally, if we replace W by W - E and recall our simplifications we see that

t&'(0, Y(O), Y'(O+), W) ~ 0, where t&' is the excess function of Weierstrass de-
fined by
t&'(x, y, Z, W) = f(x, Y, W) - f(x, Y, Z) - fz(x, Y, Z)· (W - Z), (18)
for x E [a, b] and Y, Z, WE /Rd. Clearly, a similar construction is possible to
e
the left of = 0, and for each (e, Y(e)) along the extremal curve. Hence we
have proven the
(7.15) Theorem (Weierstrass). Let f = f(x, Y, Z) and its derivatives fy and fz
locally on
F(Y) = r
be continuous on [a, b] x /R2d. Suppose that Y minimizes
f(x, Y(x), Y'(x)) dx
~ = {Y E (C l [a, b])d: Y(a) = A, Y(b) = B}

for given A, B E /Rd, with respect to the strong 1·1 norm. Then Y satisfies the
Weierstrass condition:
t&'(x, Y(x), Y'(x), W) ~ 0, v X E [a, b], WE /Rd, (19)
(where at a corner point, either one-sided derivative of Y is permitted). 0
(7.16) Remarks. Weierstrass' condition (19) also holds when in the hypoth-
eses, /R2d is replaced by a subdomain D, but now only for those W for which
it is defined. Further restrictions may be required. See Problem 7.27.
If we recall the definition 3.4 of partial convexity as extended in Problem
3.33, we see that (19) is satisfied automatically when the function ff.b X. Z) is
convex on say [a, b] x \R2d. Moreover, if this function is strictly convex, a

locally minimizing Y cannot have corner points. We will return to such
functions in Chapter 9.
(7.17) Corollary. If f is as above and Y minimizes F on

~ = {Y E (C 1 [a, b])d: Y(a) = A, Y(b) = B}
locally with respect to the strong 1'1 norm, then Y satisfies the Weierstrass
condition:
c!(x, Y(x), Y'(x), W) ~ 0, v X E [a, b], WE \Rd. (20)
PROOF. By Theorem 7.7 we know that Y also minimizes F on ~ locally with
respect to the 1'1 norm so that the above theorem is applicable. 0
(7.18) Remark. The proofs of the local results given above fail if, in the
hypotheses the strong 1'1 norm is replaced by the weak II' II norm. [The
Weierstrass construction of Y + W is not possible within the given 11'11
neighborhood of Y.] However, if Y actually minimizes F on ~ or Yactually
minimizes F on ~, then these norm distinctions are irrelevant and the appro-
priate Weierstrass conditions are satisfied.
(b) The Legendre Condition

For fixed x, Y, Z, consider the excess function (18)
e(W) ~ f(x, Y, W) - f(x, Y, Z) - fz(x, Y, Z)' (W - Z), (21)
as a function of W E \Rd alone.
Both e(W) and its gradient ew(W) = fz(x, Y, W) - fz(x, Y, Z) vanish when
W = Z. Moreover, when defined, the second partials are given by
eW,Wj(Z) = fz,zix, Y, Z), i, j = 1,2, ... , d,
at this stationary point W = Z of e where e(Z) = O.
From (0.12), it follows that for small V = W - Z #- (D, e(W) is positive or
negative with the quadratic form
def d
q(V) = L
i,j=l
fz,z .(x, Y, Z)ViVj,
J
provided that each fz.z.

• J
is continuous (in Z). This leads to the
(7.19) Proposition (Legendre). If J, fy, and fz are continuous on [a, b] x \R2d,

and Y minimizes F on ~ locally with respect to the strong 1'1 norm, then Y
satisfies the Legendre condition:
def ~
q(x, V) = Ld
i,j=l
fz,z.[Y(x)]ViVj ~ 0,
J
(22)
at each x at which the coefficient functions fz.z• J.(x, Y(x), Z) are defined and
~
continuous (in Z) at Z = Y'(x), i,j = 1,2, ... , d.
PROOF. Were q(x, V) < 0 for some V, then the Weierstrass condition (19)
would be violated at x for some values of W, contradicting Theorem 7.15.
o
For example, when f(x, Y, Z) = X IZI 2 so that
2X' i =j, ..
fz,zj(x, Y, Z) = { 0 i =1= j, l,J = 1,2, ... , d,
and q(x, V) = 2x IV1 2 , Legendre's condition is satisfied trivially on the inter-

val [0, 2J, but cannot be satisfied on any interval containing an x < O. Thus,
there can be no (local) minimizing functions on such an interval.
Remarks. In contrast to the Weierstrass condition, (22) can also be shown to

.hold when Y is only a weak local minimum for F on ~; i.e., when Y only
minimizes F locally with respect to II' II. See [S].
The Weierstrass and Legendre conditions are not equivalent. However, if

at some x E (a, b) we have the strict Legendre condition: q(x, V) > 0 when
V =1= 0, then from the above remarks, it follows that G(x, Y(x), Y'(x), W) > 0
when 0 < IW - Y'(x)1 is sufficiently small. Observe also that the strict
Legendre condition at x implies (as in 0.13) that the matrix fzz[Y(x)J is
invertible, and hence that Y is C 2 in a neighborhood of each non corner point
x by Theorem 7.14. Conversely, (22) and this invertibility implies the strict
Legendre condition [SJ.
There is also an analogous version of Corollary 7.17.
The following example of Bolza (1902) makes clear the relative strengths
of these conditions:
Bolza's Problem
Consider
f(z) = Z2(Z + 1)2,
Here
fAz) = 4z 3 + 6z 2 + 2z and f'zz(z) = 2(6z 2 + 6z + 1). (23)
Clearly the linear function Yo(x) = mx + c is stationary for f since Yo(x) =
m = const. (§6.2(a», and each
~ = {y E CI[a, b]: y(a) = aI' y(b) = btl
contains precisely one such Yo, that with m = (b i - al}/(b - a).
Moreover, by (23) the parabolic function fzAz) = 0 precisely when z =

m± = -! ± 12- 1/2 ; also -1 < m± < O. When m:S m_ (or m ~ m+) we see
that Yo does satisfy the Legendre condition fzAy~(x)) = fzAm) ~ O. Actually,
we can use the strict convexity of f on (-00, m_] or [m+, 00) (Proposition
3.10) to assert that in these ranges, by Theorem 7.12, Yo provides the unique
minimum for
on
F(P) = Ib
f(P'(x)) dx
.§± = {p E .§: P'(x) :S m_(P'(x) ~ m+)}.

(See also Corollary 3.8.) When m < m_ (or m > m+) then Yo is a weak local
minimum point for F on .§, since all those P E.§ which belong to the
weak neighborhood where max IP'(x) - y~(x)1 = max IP'(x) - ml < m - m_
(or m+ - m) will be in .§_ (or .§+). (Why?)
However, if -1 < m < 0, then Yo cannot provide a strong local minimum
for F on.§, since obviously F(yo) = S~m2(m + 1)2 dx > 0, while each strong
norm neighborhood of Yo contains a Po E .§ for which P~(x) = 0 or -1, so
that F(Po) = O. See Figure 7.10. In this range, the Weierstrass condition is
violated for Yo. Indeed, from (19), (23) and subsequent simplification,
C(x, Yo(x), y~(x), w) = f(w) - f(y~(x)) - fz(y~(x))(w - y~(x))
= f(w) - f(m) - fz(m)(w - m)
= w 2 (w + 1)2 - m 2 (m + 1)2
- 2m(2m 2 + 3m + l)(w - m)
= (w - mf[(w + m + 1)2 + 2m(m + 1)],
and if -1 < m < 0, so that m(m + 1) < 0, the bracketed expression is nega-
tive when, say, w = -(m + 1).
Thus, in particular, when -1 < m < m_ or m+ < m < 0, Yo provides a
weak local minimum which is not a strong local minimum.
Actually, when -1 < m < 0, the bracketed expression is positive when-
ever (w + m + 1)2 is sufficiently large, and hence Yo also cannot give a strong
local maximum for F, since the Weierstrass condition for - f is violated.
y! --Yo
-- - - Po
!
a b x
Figure 7.10
Problems 227
But again, when m_ < m < m+, then - fzAm) > 0 and we can use the
convexity of - f exactly as before to conclude that Yo does provide a unique
weak local maximum for F on ~.
PROBLEMS
7.1. (a) Show that

liN = max (lP(x)1 + IP'(x)l)
xe[a,bj
and
I Pill = f (IP(x)1 + lP'(x)l) dx
define norms on Cl[a, b].

(b) Can you provide other norms for Cl[a, b]?
7.2. Verify the inequalities (7.5). Hint: Use 7.2 for a = Xo where IP(xo)1 :5; IP(x)l.
7.3*. Assume thatfE C([a, b] x [R2). Show that F: Cl([a, b]) -+ [R defined by
F(P) = f f(x, P(x), P'(x)) dx
is continuous with respect to the weak norm 11'11. (See Example 4 in §5.3.)
7.4. Show that the following functions admit only C l extremals for an associated
integral function:
(a) f(x, y, z) = J1+7".
(b) f(x, y, z) = e (l + y2 + Z2).
X
(c)* f(x, y, z) =J y2 + Z2.

(d) f(x, y, z) = 4Z2 + 2yz - y2 + x.
7.5. Find all possible corner points for extremals corresponding to
(a) f(x, y, z) = xJ1+7".
(b) f(x, y, z) = (z - 1)2(z + 1)2.
(c)* f(x, y, z) = (x 2 + y2)e z .
(d) f(x, y, z) = y2 - (cos X)Z3,
using (i) fzz =I- 0; (ii) the Weierstrass-Erdmann conditions.
7.6. (a) Find the possible (local) extremal functions for
F(P) = fl [p'(X)2 - 1]2 dx

on
~ = {P E Cl [ -1,1]: P( -1) = P(l) = OJ,
which have exactly one corner point.
(b) What is the value of F for each of these extremals?
(c) Are there any extremals for F belonging to ~ which have more than one
corner point?
7.7*. (a) Find the possible (local) extremal functions for
F(P) = !o3 [P'(X)4 - 8P'(X)2] dx

on
~1 = {P E C1 [0, 3]: P(O) = 0, P(3) = 2},
which have exactly one corner point.
(b) What is the value of F for each ofthese extremals?
(c) Are there any extremals with corner points for F on
~2 = {P E C1 [0, 3]: P(O) = 0, P(3) = 8}?
r
7.8. Discuss the possible local extremal functions for
F(P) = [2xP(x) + x 2p'(x)] dx

on
~
f!) = { pEe~1 [a, b]: p(a) = aI' P(b) = b1 } •
7.9. Consider the function
F(P) = f ([p'(X)2 + P(X)2J [P(x) - XJ2 -!P(X)3 + 2XP(x)2) dx.
(a) Find the corresponding first Euler-Lagrange equation.

(b) Show that y = x and y = rxe" are solutions of the differential equation
found in part (a), where rx is an arbitrary constant.
(c) Find all possible corner points for extremals of F.
(d) Find a possible extremal for F on
~ = {P E C1 [0, 2]: P(O) = 0, P(2) = e}.
7.10. Formulate and prove an analogue of Theorem 7.12 for piecewise C 1 vector
valued extremals. (See Problem 3.33 for the relevant definition of convexity.)
7.11. Show that if rx,
F(P) = r
p, l' E qa, bJ, then when rx < 0:
cannot have a strong local minimum on

(rx(P')2 + MP' + 1'p 2)(x) dx
~ = {P E C1[a, b]: p(a) = aI, P(b) = bd·

7.12. State and prove an analogue of Proposition 3.9 for piecewise C 1 functions,
using the methods employed in proving Theorem 7.12.
7.13*. When YE (C 1 [a, bJ)d is used to represent a curve parametrically in IRd for
d > 1, it can happen that this curve does not exhibit a corner at a point Y(c)
when c is a corner point of each component Pi' j = 1, 2, ... , d.
(a) To understand this, let d = 2, and show that the curve represented by
y. = (~., P.) E @ = (C 1 [0, 2J)2,
where for 1 ::5; rx < +00:
to 0::5; t ::5; 1,
~.(t) = P.(t) = { '
t, 1::5; t::5; 2 '
has no corners, but when rx 01- 1, ~. and P. have the corner point c = 1 at
Problems 229
which the ratio Cy~/x~) is continuous. The curve is also represented by

Y1 = Y1 E (C 1 [0,2])2.
(b)* Argue conversely that near a simultaneous corner point C = 1 of a gen-
eral YE qy where P'/X' is continuous, we should be able to use functions
similar to those in part (a) to reparametrize the curve by C 1 functions.
Hint: Argue geometrically, using (a) as an example.
(c)* Generalize (a) and (b) for d > 2.
7.14. Minimal Surface of Revolution Problem. (See §7.5.)
(a) What is the second Weierstrass-Erdmann condition for f(t, Y, Z) =
yJzi + z~? Is it useful?
(b) Make a careful analysis of the second partial derivatives of f to obtain
information about possible corner points. Can there be a Y with a corner
point in only one component? Hint: Theorem 7.13 et seq.
(c) When Co #- 0, duplicate the derivation of (13/) for a general Y satisfying
equations (13) to conclude that W/X/)2 = (p/C O)2 - 1 on (0.1).
(d)* Since the ratio in (c) is continuous, argue as in Problem 7.13, that the
curve represented by Y could also be represented by Y E (C 1 [0, 1])2 for
which (13/) is satisfied.
(e) What does this problem accomplish?
7.15. Symmetric Minimal Surface of Revolution (§7.5).
If b1 = 1, we should consider a symmetric form of equation (14/) when b < 2.
(Why?) To facilitate analysis, translate the interval [0, b] to [ -b/2, b/2].
(a) Show that in order that y defined by (14/) satisfy the symmetric bound-
ary conditions y(±b/2) = 1, we must take Jl = 0 and find f3 = 2c/b > 0
def
to make cp(f3) = f3 cosh(l/f3) = 2/b(> 1).
(b) Prove that f3 > 0 => cp"(f3) > 0, and that the (unique) minimum point
f3o( ::::::0.83) gives cp(f3o) = CPo ( :::::: 1.51) :;;; cp(f3).
(c) Conclude that for 1 < 2/b < CPo, there can be no smooth minimal surface;
while for CPo < 2/b there are two possible surfaces. What happens when
CPo = 2/b?
(d) In view of Problem 3.30(c), (d), how can there be two solutions to this
problem?
(e)* Can you make a similar analysis when b1 > I?
7.16. For the function f(y, z) = Z4 + 3y2:
(a) Show that the second equation for y E C 1 [ - 2, 2] is (y'2 _ y)(y'2 + y) =
Co = const.
(b) Take Co = 0 in (a), and find a function which satisfies this equation in
!?fi = {y E C 1 [ - 2, 2]: y( ± 2) = ± I}. Hint: Patch solutions which make
a factor in (a) vanish.
(c) Use the convexity of f, to conclude that the function Yo found in (b) must
minimize F(y) = S~d[y(x)] dx on !?fi uniquely (Chapter 3.)
(d) Show that Yo is not C 2; does this contradict Theorem 7.14? Explain.
(e)* Can you find solutions to the equation in (a) for any Co #- O?
7.17. (a) Show that a piecewise linear function
F(P) = r
Po uniquely minimizes
P/(X)2 dx
on
for given a1 , b1 , ej and Cj with a < Cl < C2 < ... < CN < b. Must Yo have
a corner point at each Cj?
(b) Whenf(,!, y, z) is strongly convex on [a, b] x [R2, and iJ is as in (a) prove
that each Yo E iJ which is stationary for f in the intervals excluding the Cj'
must minimize F(y) = S:f[y(x)] dx on iJ uniquely.
(c) Use (b) to conclude that a polygonal curve will supply the geodesic be-
tween fixed points which can be joined in order by a function in iJ.
7.18. Point Constraints. Iff E C1([a, b] x [R2):

(a) Prove that each direction () in
iJo = {{) E C1[a, b]: {)(a) = {)(b) = {)(c) = O,j = 1,2, ... , N}
is iJ-admissible for F(y) = S:f[Y(x)] dx where iJ is as in Problem 7.17(a).

(b)* Conclude that if Yo is a local extremal point for F on iJ, then Yo must
be stationary for f on each interval which excludes the Cj. Will the
Weierstrass-Erdmann conditions be satisfied at any of these Cj?
7.19. Sturm-Liouville problem. (See §7.3.)

(a) If ..l. is an eigenvalue for (5') and Yo is an associated eigenfunction, show
that F(yo) = ..l.G(yo) where F and G are defined in (5) et seq. Hint: Inte-
grate the term involving 't by parts.
(b) Conclude that if p > 0 on (a, b) while q ~ 0, then each eigenvalue ..l. > o.
(c) Suppose that Yl is an eigenfunction for (5') which is nonvanishing on (a, b).
When Y E 0//0 = 5Jlo n C 1 [a, b], consider 'tY~ as a single term, and verify
the following:
F(y) - W(y) = r ['t(y')2 + ('tY~)' ~] (x) dx
where
= r ['t(y' - y~:Y(X) + rl(x)] dx ~ r(x{,
r(x)~('ty;:2)(X)' x E(a, b).
(d) Since y E 0//0' we know that y(a) = Yl(a) = O. Use L'Hopital's rule on the
factor Y/Yl, to conclude that r(x) -+ 0 as x'" a.
(e) Since Yl E 0//0 => y~ (b) = 0, with 't(b) *- 0, think of a theoretical basis to
claim that Yl(b) *- 0 so that 't(x) also vanishes as x ,1' b. (If 't(b) = 0, but
't'(b) exists, show how to reach the same conclusion. Hint: See Problem
7.21.)
(f) Use (a), (c), (d), and (e) to conclude that when G(y) = G(Yl) = 1 say, then
F(y) ~ F(yd, V Y E fJfio.
(g) Use theorems to extend the conclusion of (f) to y E 5Jlo with G(y) = G(y 1).
(h) Give other sets of boundary conditions defining 5Jlo for which the conclu-
sions of this problem will hold.
Problems 231
7.20.* (Theorem 7.7)

(a) For that part of the proof of Theorem 7.7 considered in the text, explain
why it is possible to choose e so small that y" E S~(Yo). Hint: the triangle
inequality.
(b) Suppose I f - Yo II < (j for a fixed (j > o. Set V = f - Yo and use
Lemma 7.4 to approximate V by a V. with I V.'I ~ 41 V'I. Explain why
for sufficiently small e, II Y" - Yoll < 5(j, where Y" == Yo + V..
(c) Now, suppose that Yo minimizes F on f!) in the weak neighborhood
SsiYo). Argue that Yo minimizes F on § in the smaller weak neighbor-
hood S~(Yo) = {f E 4l!: II f - Yo I < (j}. Hint: Observe that Y" of (b) ap-
proximates f in the sense of Proposition 7.6.
(d*) Explain why this approximation does not contradict the fact that a
function f with a corner point has a weak neighborhood which excludes
all C 1 functions.
7.21. Suppose that y is a C1 solution of (5') on (a, b) with y(a) = y'(a) = O. For
x > a, set oc(x) = (q - lp)(x) and
u(x) = ({'ty')2 + y2)(X).
(a) Show that u' = 2[oc{'ty')y + (1/t)(1:y')y]
(b) Use Proposition 5.3 to verify that when x E [a, b]: u'(x) ~ M[(1:y')2 +
y2] = Mu(x), for some positive constant M.
(c) Conclude that (e-Mxu(x))' ~ 0 so that 0 ~ u(x) ~ eM(x-a)u(a) = 0; i.e., that
y=(I).
locally on
r r
7.22*. To obtain necessary conditions associated with minimizing
F(P) = f[P(x)] dx = f(x, P(x), P(x)) dx
~ = {P E C1[a, b]: p(a) = al, P(b) = bd

subject to an inequality constraint P(x) ~ 1:(x), V x E (a, b), where 1: E C 1 [a, b]
is given, introduce the "slack" variable ~ by P(x) = 1:(x) - ~2(X).
(a) Argue that it suffices to consider
F*(m =
(locally) on an appropriate ~* where

r f*[~(x)] dx
f*(x, 1'/, C) == f(x, 1:(x) - 1'/2, 1:'(x) - 2I'/C).

(b) Show that a minimizing ~o should satisfy the equation
~o(X)((~)f.[Po(X)] - /Y[Po(X)]) = O.
except at possible corner points where fz[po(x)] is required to be contin-

uous. (Here Po(x) == 1:(x) - ~5(x).)
(c) Conclude that Po will be stationary for f on subintervals where Po (x) <
1: (x), but intervals of nonstationarity where Po (x) == 1:(x) (or ~o(x) == 0)
may occur.
(d) Can you draw any conclusions about the behavior of Po at points of
transition between the intervals described in (c)? Could these be corner
points of Po?
7.23.· An Optimal Control Problem.
To solve the control problem presented in Example 3 of §3.5 under the
additional Lagrangian inequality 0 ::5; u ::5; fJ as discussed in Remark 3.20:
(a) Explain why it would be appropriate to consider the convexity of the
integrand
where Al is constant.
(b) Argue that a Uo E ~ = C[O, T] which solves the equation
(i) u(l + A) = Ao(T - t) with Ao = -Ad2, A ~ 0, will minimize
F(u) = IT u2(t) dt
uniquely on ~ subject to the isoperimetric constraint

(ii) G(u) = g (T - t)u(t) = k = h + gT2/2, and the Lagrangian inequal-
ity u2 ::5; fJ2, provided that
(iii) A(t)(uMt) - fJ2) == o.
(c) Examine (i) and (iii) carefully to conclude that a solution
u (t) _ {fJ,
o - Ao(T - t),
is possible if Ao = fJ/(T - 1:), and 1: can be found so that Uo satisfies (ii).

(Remember that we want A ~ 0 and Uo ~ 0.) How is Adefined?
(d)* For g = 2, fJ = 3, find 1: and compute F(u o) = F(u o, T) say.
(e)* Minimize F(uo, T) over the allowable range of T to obtain an optimal
flight time To.
7.24*. Repeat the analysis of the previous problem when the isoperimetric condition
of part (b) is replaced by
G(u) ~ f: (1 - e-·(T-t»)(u(t) - g) dt = rxh,
for constant rx > 0, which incorporates the aerodynamic damping discussed

in Problem 3.38. (In the final computations, set rx = 1.)
7.25. (a) Use the mean value theorem on fAc, P(c), z) to establish the assertion in
Remark 7.9.
(b) Show that 7.13(iii) holds. Hint: See the derivation of (4") and Problem
3.33.
7.26*. (a) In Example 2 of §7.4, show why the convexity arguments carry through
even though Ais only piecewise continuous.
(b) Formulate and prove a more general result of this type involving a
strongly convex f{J£, y, z), a convex g{J£, y), and a piecewise continuous
A~ o.
Problems 233
7.27**. (Newton's minimum drag problem.)

In §3.4(c), we used convexity to determine minimum drag profiles for a projec-
tile moving through a medium under Newton's resistance law within the
restricted class where y' :0;; -1/Ji (See Figure 3.4.) Physically it is clear
that we only need to consider profiles with y' :0;; 0 but Newton argued that for
a given h, a cornered profile P incorporating a segment where P' = 0 is
more efficient than one in which 0 < - P' < 1. To test his insight, suppose
Po E C1[0, 1] minimizes
F(y) =
r
1
Jo
x
1 + P'(xf dx
on
f!}* = {P E C1 [0, 1]: P(O) = h, P(1) = 0, P' :0;; O}.
We cannot claim that Po satisfies (2) because at points where P~ vanishes we

are not free to vary Po in the y-direction and remain in ~*. However, we can
vary the curve represented by Po in the x-direction, and conclude that Po
satisfies (4)/ and hence (4') at a corner point. Hereafter denote Po by y for
simplicity.
(a) Show that (4') requires continuity of (1 + 3y'2)/(1 + y'2) at a corner
point where x = c > 0, and conclude that the only such points are ones
where y' has one-sided values of m = 0 and n = - 1.
(b) Away from corner-points y is C 1 and where y' #- 0, argue that y is C2 by
replacingf by f - zf. in the proof of Theorem 7.14. Differentiate (4) and
conclude that in such intervals our y satisfies equation (14) of §3.4, so
that by the earlier analysis, - y' is strictly increasing when - y' ~ 1/J i
(c) Use the previous information to establish that y can have at most one
corner point, and that it occurs at x = a where y'(a-) = 0, y'(a+) =
-1, resulting in a single-segment profile predicted by Newton!
(d) Along the curved part of y where y' #- 0 the Legendre test (22) is applica-
ble. Conclude that necessarily - y' ~ 1/Ji
(e)* Along the same arc as in (d), the Weierstrass condition (19) holds-but
only when restricted to y' :0;; 0 and W :0;; O. (Why?) Conclude that we must
have y,2 - 1 + 2wy' ~ 0, so that y,2 ~ 1; i.e., Newton was correct! and
in Problem 9.27, we will see that his suggested profile does minimize.
1 This is best accomplished by representing the curve parametrically as in [P], but recall the
footnote following Theorem 6.12.
CHAPTER 8
Variational Principles in Mechanics
The recognition that minimizing an integral function through variational

methods (as in the last chapters) leads to the second-order differential equa-
tions of Euler-Lagrange for the minimizing function made it natural for
mathematicians of the eighteenth century to ask for an integral quantity
whose minimization would result in Newton's equations of motion. With
such a quantity, a new principle through which the universe acts would be
obtained. The belief that "something" should be minimized was in fact a
long-standing conviction of natural philosophers who felt that God had
constructed the universe to operate in the most efficient manner-but how
that efficiency was to be assessed was subject to interpretation. However,
Fermat (1657) had already invoked such a principle successfully in declaring
that light travels through a medium along the path of least time of transit.
Indeed, it was by recognizing that the brachistochrone should give the least
time of transit for light in an appropriate medium that Johann Bernoulli
"proved" that it should be a cycloid in 1697. (See Problem 1.1.) And it was
Johann Bernoulli who in 1717 suggested that static equilibrium might be
characterized through requiring that the work done by the external forces
during a small displacement from equilibrium should vanish. This "principle
of virtual work" marked a departure from other minimizing principles in that
it incorporated stationarity-even local stationarity-(tacitly) in its formula-
tion. Efforts were made by Leibniz, by Euler, and most notably, by Lagrange
to define a principle of least action (kinetic energy), but it was not until the
last century that a truly satisfactory principle emerged, namely, Hamilton's
principle of stationary action (c. 1835) which was foreshadowed by Poisson
(1809) and polished by Jacobi (1848) and his successors into an enduring land-
mark of human intellect, one, moreover, which has survived transition to both
relativity and quantum mechanics. (See [L], [Fu] and Problems 8.11 8.12.)
234
§8.1. The Action Integral 235
This expository chapter is concerned chiefly with an introduction to the

theory of Hamilton's principle as obtained from variational considerations.
The action integral for a single particle is generated in §8.1, and it is extended
to Hamilton's principle for a dynamical system in §8.2 through the introduc-
tion of generalized coordinates consistent with the constraints on the system.
In §8.3, the total energy function is obtained as a consequence of the second
Euler-Lagrange equation.
Unfortunately, applications of Hamilton's principle are usually quite com-
plicated requiring the solution of a highly nonlinear system of differential
equations of the second order. In §8.3 we illustrate by example the feasibility
of linearization of these equations, making possible their solution in re-
stricted cases. In the remaining sections we present the Hamilton-Jacobi
theory for obtaining at least partial solutions (integrals of motion) of the
nonlinear system. This is accomplished by replacing the original equations
by an equivalent first-order system-the canonical equations (§8.4), which
for some coordinates may admit immediate integration (§8.5). The search
for new coordinates for which this must occur culminates in §8.7 with the
Hamilton-Jacobi equation-a single partial differential equation of the first
order whose complete solution if available would provide such coordinates.
In §8.8 we apply Hamilton-Jacobi theory to the general problem of finding
stationary functions for a given J, and use it to uncover convexity of certain
f that enables us to resolve the brachistochrone problem among others.
Finally, in §8.9, we consider the extension of Hamilton's principle to simple
continuous media.
§8.1. The Action Integral

For the motion of a single particle of mass m which at time t has the posi-
tion vector Y(t) E [R3, we have Newton's equations of motion in the form
(d/dt)(m Y(t) = F(t). Here, F represents the force impressed on the particle to
produce the motion, and the dot denotes differentiation with respect to time
in Newton's notation. We should like these equations to be the Lagrange
equations for a function L = L(t, Y, Y); viz.,
d
dt (Ly(t)) = Ly(t), (1)
as in §6.7. Thus if Y = (Yl' Y2' Y3)' we require that

. 8L
mYi = -8. for i = 1, 2, 3,
Yi
which would occur if
L(t, Y, Y) = tm I YI 2 - U(t, Y)
for some function U.
236 8. Variational Principles in Mechanics
But we also require that with F = (fl, f2' f3):
I".
Ji = oL = _(OU) , i = 1,2, 3,
0Yi 0Yi
which means that the vector force F should be derivable from the "scalar
potential" - U.
Introducing T = !m I YI 2 called (after Leibniz) the vis viva l or kinetic

energy of the particle, we see that when F = - Uy for U = U(t, Y), then
indeed Newton's equations are precisely those which characterize the
stationary functions for L = T - U. Thus in each time interval [a, b], they
r
would be satisfied by a function Y which minimizes the action integral
A(Y) = L(t, Y(t), Y(t)) dt

on
E0 = {YE (Cl[a, b])3: Y(a), Y(b) fixed}.
However, they would also be satisfied by a function which maXImIzes
(locally) or more generally, makes stationary, A on E0, in the sense that
c5A(Y; V) = 0, 'if V E E00 = {V E (Cl[a, b])3: V(a) = V(b) = (1J}. (2)
Indeed, as we know, the Euler-Lagrange equations (1) are both necessary
and sufficient that (2) holds. (See 6.9 et seq.)
The function L = L(t, Y, Y) = T - U is called the Lagrangian for the

motion, and the first term of the action integral A(Y) = U) dt reflects r:(T -
the kinetic energy during the motion, while the second may be regarded as
arising from the potential energy injected into the system as a result of the
work done by the external forces during the motion. However interpreted,
the resulting equations of motion are those of Euler-Lagrange which must
agree with those of Newton if the principle is to be valid.
§8.2. Hamilton's Principle: Generalized Coordinates

Dynamical systems may be thought of as consisting of a large finite number
(N) of part~cles of masses mj which occupy positions lj(t) E [R3, moving with
velocities lj(t) at time t, j = 1, 2, ... , N. The associated kinetic energy of the
system is
1
T=-
LN.
m·IYI 2 (3)
2 j=l J J '
and in 1835, Hamilton postulated that if the system occupies certain posi-
tions at times a and b, then between those times, it should move along those
1 Leibniz applied this term to 2T.

§8.2. Hamilton's Principle: Generalized Coordinates 237
= r
admissible trajectories which make "stationary" the action integral
A(Y) L(t, Y(t), Y(t)) dt,
where Y = (Y1, Y2, ... , YN) E (C 1[a, b])3N and L = T - u.

Here, U represents the potential energy given to the system (at time t)
through the work done by the external forces acting on the system, and the
function L is called the Lagrangian of the system.
The admissible trajectories are those which are consistent with the con-
straints on the system and which have the prescribed end point values at
times a, b, but stationarity for constrained systems is difficult to define.
Unfortunately, the motions of most dynamical systems must satisfy cer-
tain geometrical (or other) constraints. For example, in order that the system
move as a rigid body, the position functions lj(t), must preserve at all times
the individual distances between the particles. Even when such constraints
can be treated through Lagrangian multipliers as in §6.7, their number makes
this approach impractical.
There is the following alternative approach through the use of generalized
coordinates introduced by Lagrange (c. 1782):
We suppose that the state of the constrained system at time t admits
description by n independent kinematic variables q1' q2' ... , qn which consti-
tute the position vector Q E IRn. [For example, each position of the rigid
planar pendulum of Figure 8.1 is determined by specifying the angle e which
it makes with a fixed axis, rather than the pair of cartesian coordinates (x, y)
of its tip together with the constraining relation x 2 + y2 = 12. Of course,
there are the familiar equations x = 1cos e, y = 1sin e, which permit re-
trieval of the original cartesian description.]
Similarly, in the general case, the position vector Y should be determin-
able from Q by means of a transformation Y = G(Q), say. Thus during a
motion described by Q(t), Y = GQQ, where GQrepresents the Jacobian matrix
of the transformation. The kinetic energy of the system as given by (3) now
assumes the general quadratic form
(4)
(x, y)
x
Figure 8.1
for certain functions aij(Q) = aji(Q); i, j = 1, 2, ... , n, which are determined

from GQ and the masses mj'
[For the pendulum,
x = -l(sin ())8,
y = l(cos ())8,
and
m . m·
T = tm(x 2 + y2) = I 12 (sin2 () + cos 2 ())()2 = I I2 ()2.]
The Lagrangian of the motion may also be regarded as a new function

L = L(t, Q, Q) from 1R 2 n+1, and now Hamilton's principle takes the following
more attractive form:
Between fixed times a, b, the system should move along those trajectories
represented by generalized coordinates Q E (C 1 [a, b]t with prescribed values
at a, b, which make stationary the action integral A(Q) = f~ L(t, Q(t), Q(t)) dt.
Thus we should have

<5A(Q; V) = 0, V V E ~o = {V E (C1[a, b]t: V(a) = V(b) = (P}.
From the general theory of §6.7, on (a, b) the admissible Q should satisfy the
Euler-Lagrange equations
d
-LQ(t) = LQ(t) d(aL) aL
or - - . =-, j = 1,2, ... , n, (5)
dt dt aqj aqj
and these may be regarded as the equations of motion of the system.
Hamilton's principle, expressed in generalized coordinates, provides a

straightforward method of determining equations of motion for a general
dynamical system. Moreover, it is macroscopic in nature and does not in-
volve the specific action of the position of each individual particle (which for
a true rigid pendulum would be astronomical in number!). A difficulty in its
application is in the determination of the proper function U to represent the
potential energy of the system (which is definable only within an additive
reference constant). In some cases, this is decided by choosing that function
necessary to recover already agreed upon equations of motion in (5). This
was in effect our approach to the problem for the motion of a single particle
in the previous section. And as in that case where we required that the force
F = - Uy , there need not be a choice which renders the principle valid. In
particular, the presence of irreversible phenomena such as friction usually
precludes the use of Hamilton's principle as stated.
(8.0) Remark. In using Hamilton's principle in generalized coordinates we

assume implicitly that the number n of independent generalized coordinates
required to describe the configuration of the constrained system is fixed by
the system, and would be independent of the particular set of generalized
§8.2. Hamilton's Principle: Generalized Coordinates 239
coordinates employed. For then, two such sets of coordinates Q, Q, say,

should be relatable through an invertible transformation of the form (38) of
§6.8; we can conclude from Theorem 6.12, that those C 2 functions Q(t) which
make stationary Hamilton's action integral for the Lagrangian L(t, Q, Q),
will transform into functions Q(t), which will make stationary the corre-
spon~ing action integral when the Lagrangian is expressed in terms of t, Q,
and Q. Thus, in this sense, Hamilton's principle of stationary action may be
'said to be an invariant of the system. (When a C 1 function Q(t) actually
minimizes (or maximizes) its action integral (locally) then simpler consider-
ations show that Q(t) must do so as well for its action integral.)
Bernoulli's Principle of Static Equilibrium
Hamilton's principle represents an attempt to characterize the manner in

which the transfer between kinetic and potential energies takes place during
an actual motion. It would be satisfying to assert that actual motions mini-
mize this transfer in some sense (and in §9.8(a), we shall see that during small
time intervals, the action integral is minimized in the classical situation), but
even for a single mass particle, the actual motion along stationary trajectories
may not always minimize the action integral (Problem 9.1). (See, however, the
discussion of Jacobi's principle ofleast action in §8.5.)
If motion of the system does not occur, the system is in static equilibrium.
Then its kinetic energy T == 0, and if its potential energy, U, depends only on
position, Hamilton's principle reduces to Bernoulli's principle, which states
that this equilibrium state is one which makes stationary the potential en-
ergy, U, of the system. Here U = U(Q) may be regarded as the work done by
the external forces in bringing the system from a reference state to the posi-
tion Q. That this equilibrium state may not minimize U is evidenced by the
example of a marble balanced on top of a sphere. However, as with the case
of this marble, actual physical disturbances will transfer it from its state of
(unstable) equilibrium, to one of stable equilibrium in which it rests, say, on
a table supporting the sphere. And in this state its potential energy is mini-
mized, at least, locally, relative to small displacements. (With larger displace-
ments it could fall to the floor and further reduce its potential energy, or
alternatively, change its reference state.)
Thus we may expect that the stable equilibrium states, those capable of
being sustained indefinitely during small disturbances, should provide a local
minimum for the potential energy function U. It was appropriate to invoke
this principal of minimum potential energy in §3.5, when we were seeking the
stable equilibrium shape of a cable hanging under its own weight. Moreover,
as we shall see in §8.9, some physical systems can be in equilibrium only when
they minimize their potential energy (relative to its reference value); i.e., they
cannot exhibit unstable equilibrium states.
§8.3. The Total Energy

From Hamilton's principle we suppose that the functions Q E (C 1 [a, b])"
which describe the motions of a dynamical system are the stationary func-
tions for a suitable Lagrangian L = L(t, Q, Q). Thus when Q is C 2 (and this
assumption is usually made in dynamics) they should satisfy also the second
Euler-Lagrange equation of §6.7, in the form of equation (35):
LQ(t)· Q(t) - L(t) = - J: Llr) d"C + c;

or introducing
• deC •• •
E = E(t, Q, Q) = LQ(t, Q, Q)' Q - L(t, Q, Q),
then (6)
E(t) = - J: Lt("C, Q("C), Q("C» d"C + c.
Now, when L = T(Q, Q) - U(t, Q), and T is given by (4), then

oL oT n
-. = -. =
oqj oqj
L aij(Q)q;,
;=1
so that
n
E= L a;j(Q)q;qj - (T - U)
;,j=l
= 2T - (T - U) = T + U,
and E is the sum of the kinetic and potential energies of the system.
In any case, we call E as defined by (6) the total energy ofthe system. When
the Lagrangian, L, does not depend explicitly on time so that L t == 0, then the
second part of (6) shows that motion can occur only along those trajectories
represented by Q for which
E(t) = E(t, Q(t), Q(t» = const.,
i.e., for which the total energy is conserved. In general, we have
E(t) - E(a) = - J: Lt("C, Q("C), Q("C» d"C, (7)

or
E(t) = - Lt(t, Q(t), Q(t»,
which may be regarded as a kind of conservation law for the system. (7) is
also an integral of motion. When Q = q1 alone, then it is equivalent to the
single Euler-Lagrange equation of motion (see §6.3). However, in general,
there will be other integrals of motion.
§8.3. The Total Energy 241
Figure 8.2
Application: Spring-Mass-Pendulum System
In the system shown in Figure 8.2 which consists of a rigid pendulum of

length 1supported from a frictionless spring-mass system, the motion is quite
difficult to describe in terms of standard rectangular or polar coordinates.
However, the configuration is completely specified by the pair of generalized
coordinates (x, e) where x denotes the position of the mass m and e denotes
the deflection angle of the pendulum, both measured from the equilibrium
position illustrated.
We assume that the mass m is constrained to move horizontally by fric-
tionless guides with its motion opposed by a spring force k(x). Moreover, we
shall suppose that the mass of the pendulum arm is negligible in comparison
to that, M, at its tip. Then, the potential energy function U(x, e) is the work
required to bring the system from its reference state of equilibrium (x = e = 0)
to the configuration (x, e), i.e.,
U(x, e) = f: k(s) ds + Mgl(l - cos e), (8)
where the first term represents the work done in stretching the spring
(as given by elementary calculus) and the second term is the work done in
raising the pendulum mass M against gravity. (g is the familiar gravitational
constant.)
To obtain the kinetic energy function T it is best to express it first in terms
of the rectangular coordinates (x l' Y1) of the mass M as shown in Figure 8.2,
and then to use the geometrical relations
Xl = X + 1sin e, Xl = X + l(cos e)O,

so that (9)
Yt = 1cos e, Yl = -l(sin e)O,
to convert to the generalized coordinates. Thus from (3),

T = tm(x 2) + tM(xI + pi),
or, after substitution of (9),
T = t(m + M)x 2 + Ml(cos e)xO + tM1202. (10)
Observe that T is in the form (4). From (8) and (10) we see that the
Lagrangian for this system is given by
L = T - U = t(m + M)x 2 + Ml(cos e)xO + tMl202
-f: k(s) ds - Mgl(1 - cos e); (11)
it is seen that L does not depend on time explicitly so that L t == 0, and from
(6), the total energy E = T + U is constant for this system. We now assume
that k is continuous.
From Hamilton's principle, there are two equations of motion (5) for this
system, one for each of the generalized coordinates x, e.
That for x is
d
dt x = L x
-(L.)
or from (11) and A.8:
d .
dt [(m + M)x + Ml(cos e)e] = - k(x); (12)
while that for eis
or
d . .
dt Ml[(cos e)x + W] = -Ml sin e(g + xe). (13)
Dividing both equations by Ml and introducing p. = (m + M)/(Ml), K(X) =

k(x}/Ml, there results the coupled second-order system:
(p.x+ (cos e)O)· = - K(X),

(14)
(cos e)x + Ie = -g sin e,
for which we know one integral of motion, namely, that E = T + U is con-
stant. The particular constant depends upon the initial configuration. For
example, if the mass m is given an initial velocity Xo when the system is in
equilibrium, then
E = t(m + M)x~.
§8.4. The Canonical Equations 243
However, there are three remaining integrals of motion of the nonlinear

system (14) with the associated initial conditions, and without further simpli-
fications, these cannot be obtained explicitly. Because linear systems can in
general be solved, the further assumptions usually employed are those which
"linearize" (14). If we suppose 101 ~ 1 so that sin 0 ~ 0, cos 0 ~ 1, then (14)
becomes
e
p.x + = - K(X), x + le = -gO. (15)
If, in addition, we suppose K(X) ~ KX which will describe a so-called linear
spring with the spring constant K, then (14) is approximated by the fully
linearized system
e
p.x + = -KX, x + le = -gO. (16)
These linearizing assumptions must be retained and examined in conjunc-
tion with any solution of (16). Observe, though, that the assumptions are
different in character. The first is a definite restriction on the allowable
amplitude of the pendulum swings, while the second is satisfied in some
springs for quite large extensions x. If these same assumptions are used in
approximating the Lagrangian (11) (with 1 - cos 0 ~ 02 /2), then the linear-
ized equations (16) follow directly from Hamilton's principle. (See Prob-
lem 8.l.) The further solution and discussion of these equations is left to
Problem 8.2.
(Problems 8.1-8.3,8.12)
§8.4. The Canonical Equations

The equations of motion for a dynamical system in the Euler-Lagrange
form (5)
:t(;~) = ;~, j = 1,2, ... , n, (17)
constitute a system of n second-order differential equations which are usually

nonlinear and difficult, if not impossible, to integrate directly. Hamilton and
later mathematicians sought transformations which would simplify this task.
First, we note that in terms of the conjugate momenta Pj = oL/oqj' j =
1, 2, ... , n, the equations (17) take the form
oL . oL
Pj = -0. , Pj = -0' j = 1, 2, ... , n, (18)
qj qj
which is now a first-order system in the 2n variables Q = (ql, q2' ... , qn),
P = (Pl, P2'···' Pn)·
To transform this system into a form more amenable to analysis, Hamil-

ton made the Legendre transformation in which it is supposed that the n
equations
oL .
Pj = -0. (t, Q, Q), j = 1,2, ... , n, (19)
qj
can be solved for the qj as C 1 functions of the remaining variables t, Q, P, say,
Q= G(t, Q, Pl.
[If we regard equations (19) as constituting a transformation from ~2n+l
to ~n, then G is locally determined implicitly as a C 1 function of its variables
provided that the n x n Jacobian. determinant 10 2 L/oq;oqjl =F O. (See, for
example, [Ed].) When L = T(Q, Q) - U(t, Q) with T given by (4), then
02 L/oq;oqj = a;j(Q), i, j = 1, 2, ... , n; and the nonvanishing of the determi-
nant Iaij(Q) I is assured by the positive definiteness of the quadratic form
(19')
when expressed in the original Cartesian coordinates of equation (3); see

§O.13. In this simple but important case, (19) is just the linear system
n
Pj = L a;j(Q)q;,
;=1
j = 1,2, ... , n,
which has an explicit solution of the same form.]
With
Q= G(t, Q, P), or qj = git, Q, P), j = 1,2, ... , n,
we have
Lqp, Q, Q) = Lqj(t, Q, G(t, Q, P)) = jj(t, Q, P),
say, for j = 1, 2, ... , n, and now the system (18) becomes
qj(t) = git, Q(t), P(t)),
j = 1,2, ... , n, (20)
jJit) = jj(t, Q(t), P(t)),
which is in the so-called normal form. Moreover, when expressed in these
variables, the total energy function of (6) is given by
E(t, Q, Q) = p. Q - L(t, Q, G(t, Q, P))
= p. G(t, Q, P) - L(t, Q, G(t, Q, P)) = H(t, Q, P), (21)
say, if we use (21) to introduce the new function H called the Hamiltonian of
the system. Observe that
n
H(t, Q, P) = L Pjgj(t, Q, P) - L(t, Q, G(t, Q, P)) (21')
j=l
§8.4. The Canonical Equations 245
so that by the chain rule
oH = g.
0Pi I
+ f. (p.) _ oqj
j=l
OL)Ogj
OPi
and (22)
i= 1,2, ... ,n.
Now, since
oL . oL
Pj == ~(t, Q, Q) == ~(t, Q, G(t, Q, P)),
uqj uqj
each summation in equations (22) vanishes; also
oL oL
~ = ~(t, Q, G(t, Q, P)) = jj(t, Q, P) for j = 1,2, ... , n, as above;
uqj uqj
thus we obtain the following equations:
oH oH
-=gi and -= -J;, i=I,2, ... ,n,
0Pi Oqi
which complete the Legendre transformation. These equations do not express
laws of dynamics, but with their help, the equations of motion (20) assume
the form
oH
qj(t) = ~(t, Q(t), P(t)),
UPj
j = 1,2, ... , n. (23)
oH
Pj(t) = -~(t, Q(t), P(t)),
uqj
Thus, along a stationary trajectory where H(t, Q(t), P(t)) = H(t), say,
dH
t
oH
-d (t) = ~ +
ut
n (OH
}=1
~qj
uqj
r OH)
+ ~Pj
UPj
=
oH
~(t, Q(t), P(t)),
ut
(24)
in view of (23). This is the transformed version of the conservation law (7) and
it may be considered as a (2n + l)st equation of motion.
Equations (23) and (24) may be expressed in the condensed form
H=Ht •
or (25)
H=Ht , j = 1,2, ... , n,
and they are known as the canonical equations of motion for the system.
They are attributed usually to Hamilton (1835), although they first appear in
work of Lagrange (1809) and were also used by Cauchy (1831). They show
that instead of thinking of a dynamical system as being described by its

Lagrangian as a function of positions and velocities, it is preferable (analyti-
cally) to regard the system as determined by its Hamiltonian expressed in
terms of positions and conjugate momenta. 1
Now, formally, (21) may be restated as L = p. Q- H, and it is straight-

forward to verify that equations (25) are those of Euler-Lagrange for the
generalized action integral
A(Q, P) = r [P(t)· Q(t) - H(t, Q(t), P(t))] dt.
Moreover, functions satisfying (25) make this integral stationary on sets such
(26)
as {(Q, P) E (C 1[a, bJ)2n: Q(a) = A, Q(b) = B}; in particular, this stationarity

does not involve the specification of P at a or b. (See Problem 8.4.) Observe
that the new integrand is "naturally" of the form 2T(P, Q) - H(t, Q, P), is
independent of P, and depends explicitly (and simply) on Q.
Now, when T = t Li',j=l a;j(Q)ti;tjj and U = U(t, Q) as in §8.3, then
j = 1,2, ... , n,
so that
and from (21), the Hamiltonian is,

H(t, Q, P) = E = T + U = tP' Q+ U(t, Q)
= tp· G(t, Q, P) + U(t, Q). (27)
For example, in the spring-mass-pendulum system of the previous section
where Q = (q1' q2) = (x, e) E 1R2, the conjugate momenta are from (8) and
(10)
P1 = Lx = Tx = (m + M)i + Ml(cos e)e,
• 2' (28)
P2 = LiJ = Te = Ml(cos e)x + Ml e.
We may solve this simple linear system to get
41 = i = [M12P1 - Ml(cos e)p2J/i1,
(28')
42 = e = [ -Ml(cos e)p1 + (m + M)P2J/i1,
where i1 = M1 2 [m + M sin 2 eJ > O. It follows from substituting (28') into
1 The Hamiltonian can be defined purely geometrically when the Lagrangian L(!, Q, Z) is
strictly convex. See Problem 8.22. -
§8.5. Integrals of Motion in Special Cases 247
(27) that the Hamiltonian for this system is

H = H(Q, P) = t(p141 + P242) + U(t, Q)
= [Mf2pi - 2Ml(cos O)PIP2 + (m + M)pn/2.1
+ f: k(s) ds + Mgl(1 - cos 0). (29)
The canonical equations (25) can now be obtained. The first of these, namely
41 = x = Hp, and 42 = {) = H p2 , simply recover equations (28'). However,
P1 = -Hq, = -Hx = -k(x) and P2 = -Hq2 = -He are new equations, the
latter being rather complicated by the dependence of .1 on O. These four
first-order equations replace the two second-order equations (14) obtained
earlier, and they constitute the canonical equations of motion for the system.
(The final equation if = Ht == 0 shows that H = E = const. along the station-
ary trajectories.)
The local solution of this first-order nonlinear system with given initial
conditions may be obtained by numerical approximation. See [I-K]. This is
one of the reasons that the canonical equations are superior to the Euler-
Lagrange equations (14).
The canonical equations also offer significant advantages to those con-

cerned with constructing mathematical models for a dynamical system. It
should be observed that only usage in the last three centuries has made terms
such as "Kinetic energy of rotation" seem familiar. Hamilton's equations
suggest that the "true" dynamical variables are the momenta Pi' and for a
system, it is only necessary to know-or postulate-a suitable Hamiltonian
expressed in these variables. Experience may suggest some of the terms for a
Hamiltonian, H, and the effects of modifying an assumed H on the result-
ing equations of motion (25) can be studied easily. Moreover, in general,
Hamilton's equations cannot have more than one solution (Q(t), P(t» with
(Q(a), P(a» prescribed, so that with any given Hamiltonian, the solution
curves in (Q, P) space cannot intersect. (See A.17.) This makes them much
more stable with respect to numerical integration schemes than those of
Lagrange (14).
(Problems 8.4-8.5)
§8.S. Integrals of Motion in Special Cases

The canonical equations (25); viz.,
H=Ht , j=I,2, ... ,n,
provide immediate integrals of motion when the Hamiltonian is independent
of some of the 2n + 1 variables t, ql' q2' ... , qn' Pl' P2' ... , Pn. For example, if
H Pi == 0 for some i = 1, 2, ... , n, then qi = const. is one integral of motion,
and the motion is confined to such hyperplanes. Similarly, if Hqi == 0 for some
i = 1, 2, ... , n then Pi = const. = Ci' say, is an integral of motion. Moreover,
in this case the ith variable qi may be effectively deleted or "ignored" by
r
considering the stationary functions for the modified action integral
A(Q, P) = [P(t)· Q(t) - H(t, Q(t), P(t))] dt, (30)

where
and
each vary only in IRn - l . (See Problem 8.9.)
Jacobi's Principle of Least Action
Of special interest is the case when H t == O. Then H = H(Q, P), and from the
last of (25), H(Q(t), P(t)) = const. is an integral of motion of the system. Thus
the motion takes place along curves confined to hypersurfaces in the Q, P
space which conserve the total energy E as in Figure 8.3. In this case, t is
simply a parameter for the motion of the system and time becomes ignorable
on replacing it with t = t(,) where, is a new variable awaiting specification.
q.
Q(IX) I
Q-space
IQ(fJ)
Figure 8.3
§8.5. Integrals of Motion in Special Cases 249
Then formally, Q(t) dt becomes Q'(r) dr, and we consider instead of

(26), the problem of finding extremal functions Qo, Po for Jacobi's action
integral
.4(Q, P) = f: P(r)' Q'(r) dr (31)

on
f!) = {(Q, P) E (el[lX, pJ)2n: Q(IX) = A, Q(P) = B},
subject to the Lagrangian constraint H(Q(r), P(r)) = E. Here, IX and pare
fixed, and r is used to parametrize each curve on the hypersurface H(Q, P) =
E so that this is possible. (Earlier investigators had formulated this principle
improperly without reparametrizing, a fact first criticized by Jacobi.) It can be
shown (in principle) that such functions Qo, Po do satisfy equations (25) for a
properly chosen r (Problem 8.10).
Jacobi's principle can be given an even more geometrical flavor in the

classical case where U = U(Q) and as in (19')
2T =. t
~J=l
aij(Q)qiqj = (. t
~J=l
aij(Q)q;qi)!(t')2.
Then, as in (27), T + U = E = H = const., so that E - U(Q) =T ~ 0, while

also
2 T = p. Q= (p. Q')/t'
Thus p. Q' = 2Tt' = 2JE - U(Q)JT(t')2, and Jacobi's integral reduces to
.4(Q) = f: 2Tt' dr = f: ds(r), (32)
say, where s is the "arc length" of element whose square is

n
ds 2 = L
i,j=l
Aij(Q) dqidqj, (32')
with
i,j = 1,2, ... , n.
Because the associated quadratic form is positive definite, (32) defines a
Riemannian metric in Q space whose coefficients Aij are mass dependent. For
this case, the curves which satisfy Jacobi's principle of least action by mini-
mizing (32) are those on the constant energy surface whose projections in the
physical Q space will be the geodesics with respect to the metric. (See Figure
8.3.) Such metrics will not be spatially homogeneous in general, and Q space
may be thought of as being "curved" appropriately by the mass dependent
metric to reflect this fact. However, these motions in Q space are physical and
involve changes which can be observed.
Symmetry and Invariance

As we have seen, coordinates qi for which Hq , = 0, are ignorable, and yield
Pi = const. as associated integrals of motion. Since there are many possible
coordinate systems, it is natural to ask whether we can detect the presence of
such ignorable variables and their integrals of motion in some other coordi-
nate system. For example, when expressed in cartesian coordinates x, y, can
we already "see" that He would be zero if we transform to polar coordinates
(r, ())? Perhaps. We would surely know this if H(x, y) = k(x 2 + y2), i.e., if His
symmetric with respect to the origin.
More generally, if our entire dynamical system has an axis of symmetry,
then we should expect a corresponding ignorable coordinate and an asso-
ciated constant momentum-in this case angular momentum about the axis
of symmetry. However, to establish the presence of the axis of symmetry, it
would suffice to show that the system is descriptively invariant with respect
to the family of rotations about that axis. This led E. Noether (1918) to an
examination of in variance as a source of integrals of motion; for her principal
result, see [G-F].
§8.6. Parametric Equations of Motion

Thus far in our development, time has occupied a distinguished position.
Suppose we simply regard it as one more generalized coordinate qo, say, and
enlarge our position vector to Q = (qo, ql' ... , qn) E IRn+1.
Then, if we consider motions as given in parametric form t = qo(r), ql(r),
... , qn(1:), IX ~ 1: ~ p, Hamilton's action integral (26) is transformed into the
following:
A(Q, P) = fP
"
[t
J-I
Pj(1:)qj(1:) - H(Q(1:); PI (1:), ... , Pn(1:))q~(1:)J d1:
or
where we have introduced the new conjugate momentum

Po = - H(Q, PI' P2' ... , Pn), (34)
and enlarged the momentum vector to
P = (Po, PI' ... ' Pn) E IRn+1.
Now, suppose that this new action integral is to be made locally extremal
among all trajectories in 1R 2 n+2 which join fixed points in IRn + 1 subject to the
Lagrangian constraint of the simple form analyzed in Theorem 6.10:, viz.,
K(Q, P) == Po + H(Q, PI' P2' ... , Pn) = o. (35)
§8.7*. The Hamilton-Jacobi Equation 251
For a general K of this form we may in principle choose r so that the resulting
equations of motion become
, oK , oK j = 0, 1, 2, ... , n. (36)
qi=~' Pi = -~,
uPi uqi
(See Problem 8.10.)
Observe that the integrand in (33) does not depend explicitly on r and its
form recalls Jacobi's action integral (31) from the preceding section. How-
ever, were we to attempt an analogous geodesic interpretation of least-action
principle, we would be led to examine a form
which is not positive definite and hence could only be associated with a
nonstandard metric. Such metrics were studied by Ricci (1892) and they form
a basis for the general relativity theory of Einstein (1916). The resulting
geodesics are not purely spatial in character as were Jacobi's, but occur in the
space-time world suggested by our new coordinates.
§8.7*. The Hamilton-Jacobi Equation

There are, in practice, many sets of generalized coordinates which may be
used to describe the motion of a given dynamical system; (but see Remark 8.0
of §8.2). It is natural to ask for one, (2, P), say, in terms of which the equa-
tions of motion (36) admit simple integration. Following Jacobi, we seek that
for which the associated constraining function (35) simplifies to
(37)
since then the equations of motion (36) reduce to iio = 0, po(r) = -r + Co
with all other Pi' iii' constant.
To obtain this coordinate system from an original one (Q, P), Hamilton
had proceeded indirectly by recognizing that the trajectories which could
make stationary the integral (33) with Q(ex), Q(P) prescribed, subject to the
constraint K(Q, P) = const., will also make stationary a modified integral
f: [(Po Q')(r) - :r S(Q(r), (2(r))J dr, (38)
where S = S(Q, (2) is a single real valued C 1 function of the variables (Q, (2) E
[R2n+2; [the second term is actually a constant (after integration) so that S
does not participate in the variations]. In particular, if we could find S with
P = SQ and define P = - SQ' then, by the chain rule:
d - - - -
drS(Q(r), Q(r)) = SQ· Q' + SQ· Q' = p. Q' - p. Q',
and the new integrand of (38) is just p. Q'. Thus our search for Jacobi's
coordinates is reduced to finding an S = S(Q, Q), for which
llo = K(Q, P) = K(Q, P), when P = SQ = SQ(Q, Q).
Moreover, since motion in the new system occurs with llo == 0, it suffices
to find a complete solution S = S(Q) to the first-order partial differential
equation
Relative to the original generalized coordinates where
qo = t, and K(Q, P) = Po + H(t, ql' ... , qn' Pl' ... , Pn)

(as in (35», this equation for S = S(t, ql' ... , qn) is
aaSt + H (t, ql' ... , qn' aasql , aasq2 , aas)

qn
== 0, (39)
which is called the Hamilton-Jacobi equation.
For the spring-mass-pendulum problem of §8.4 where the Hamiltonian

function is given by (29), the Hamilton-Jacobi equation is
1 2 2 2
S, + 2.1 [Ml Sx - 2MI(cos (J)Sx S6 + (m + M) S6 ]
+ IX k(s) ds + Mgl(1 - cos (J) == O. (40)
As this example illustrates, the Hamilton-Jacobi equation (39) is not simple,

nor are its solutions readily available. However, if a complete solution to this
single equation is known, then in principle it provides a solution to the entire
system of equations of motion. By definition, a complete solution to this
equation is a function S = S(t, Q; Cl' C2' ... , cn ) + Co which satisfies the equa-
tion for each choice of the essential constants cl , c2, ... , Cn. Replacing cj by
llj' j = 1, 2, ... , n, it follows that S = S(t, Q; Q) satisfies (39) for each Q E [Rn.
Next, the n equations P = SQ = SQ(t, Q, Q) in principle determine Q as
functions of t, P, Q. Then
-def -
P = -SQ = -SQ(t, Q, Q) (41)
completes the transformation from the original (t, Q, P) coordinates to the
Q, P) coordinates of Jacobi in which the equations of motion are simply
(t,
Q(t) = C = const. E [Rn and P(t) = B = const. E [Rn, so that (41) becomes
B = -SQ(t, Q, Q)I_ = -Sdt, Q, C). (42)

Q=C
Finally, these last equations can also in principle be solved for Q as a function
§8.7*. The Hamilton-Jacobi Equation 253
of t and the vector constants B, C; say,

Q = Q(t) = G(t, B, C), (43)
and these are the equations of motion in the original space-time coordinates.
The vector constants Band C are to be chosen (if possible) to satisfy given
initial conditions on the system.
'Observe that when Ht == 0, then it suffices to find a complete solution

s(Q; C) to the reduced equation
H(Q, sQ) = E = const., (44)
since then S = - Et + s will provide a complete solution to (39).
We illustrate the method by a simple example which permits a complete
solution, namely that for a Lagrangian of a single particle given by
L(t, q, q) = q2.
Then p ~ Lq = 2q, and the resulting Hamiltonian is defined by
H(t, q, p) = pq - q2, with q = ~,

so that
p2
H = H(p) =4.
Here H t == 0, and the reduced equation (44) for s = s(q) is now
H(sq) = E or s; = 4E,
which has the complete solution s = 2c l q + co, where d = E. Then
S = - d t + 2c 1 q + Co will be a complete solution to the corresponding
Hamilton-Jacobi equation (39),
S2
St + ; = o.
For this example, n = 1 and equation (42) is simply
hl =
as (t, q; c
--a = 2c l t - 2q,
l )
Cl
which when solved for q becomes
(45)
We contend that this is a solution to the original Lagrangian equation for

L, namely, (d/dt)(2q) = 0 (Proposition 6.1). Now, (45) does indeed solve this
equation, and, conversely, each solution of this equation is of the form (45)
for appropriate choice ofthe constants hi' c1 .
The method of separation used to obtain the complete solution in this

simple case admits extension for more complicated Hamiltonians. [We]
(Problems 8.5-8.8)
§8.8. Saddle Functions and Convexity;

Complementary Inequalities
In the last sections of this chapter we have shown how the problem of finding
stationary functions for a Lagrangian L = L(t, Q, Q) may be systematically
replaced by the simpler problem of finding the stationary functions for a
Hamiltonian H = H(t, Q, P), which in turn can be accomplished by solving a
related first-order partial differential equation
St + H(t, Q, SQ) == o.
Although motivated by physical considerations and couched in the terminol-
ogy of analytical mechanics-velocity, momenta, energy, etc.-this transfor-
mation is purely mathematical, and it is equally applicable to the problem of
finding stationary functions for any f = f(x, Y, Y'). (In the next chapter we
shall see the connection between the corresponding Hamilton-Jacobi equa-
tion and the existence of a field used in establishing the minimality of a given
stationary function for f.)
For example, the integrand f(x, y, y') = y,2 corresponds to the Lagrangian
L(t, q, q) = q2 which we have examined at the conclusion of the preceding
section. As a result, we know that the stationary functions for this f must be
of the form corresponding to (45), namely,
When the integrand function ffb Y, Z) is convex, then we know from

Chapter 3 (Problem 3.33) that each stationary function for f (as "Lagran-
gian"), in fact minimizes an associated (action) integral. To see how this
convexity relates to the behavior of the "Hamiltonian" of f, assume that the
equations P = fz(x, Y, Z) determine the components of Z as C 1 functions of
x, Y, and P. Then with this Z, as in (21), the associated Hamiltonian is defined
by
H(x, Y,P)=p·Z-f(x, Y,Z) (46)
and exactly as before we conclude that at corresponding points:
Hy = - fy, Hp = Z. (46')
Consequently, at a fixed x, let (Y, Z, P) and (Yo, Zo, Po) be triples in 1R3d
related through the above transformations. Then it is straightforward to
§8.8. Saddle Functions and Convexity; Complementary Inequalities 255
verify the following identity:
f(x, Y, Z) - f(x, Yo, Zo) - fy(x, Yo, Zo)· (Y - Yo) - fz(x, Yo, Zo)· (Z - Zo)
= [H(x, Y, Po) - H(x. Y, P) - Hp(x, Y, P)· (Po - P)]
- [H(x, Y, Po) - H(x, Yo. Po) - Hy(x, Yo, Po)· (Y - Yo)], (47)
which shows that in general (see Problem 8.24)
f(~,Y, Z) is [strictly] convex precisely when both H~, X. P)

and - H (~ Y, f) are [ strictly] convex on appropriate sets.l
When a function H = H(x, Y, P) has these convexity properties on a set

S!::; 1R 2d +1, we will say that H(~, Y, P) is a [strict] saddle function on S [Sew].
For example,
is a strict saddle function on [1,2] x 1R2d. Saddle functions have their own
significance as we see in the next results.
(8.1) Theorem (Arthurs). Let H = H(x, Y, P) E C 1 (1R 2d +1 ) and suppose that

(Yo, Po) E qIj = (C 1 [a, b])2d is stationary for the corresponding action integral
on
A(Y, P) = r [P(x)· Y'(x) - H(x, Y(x), P(x))] dx (48)
~ = {(Y, P) E qIj: Y(a) = Yo(a), Y(b) = Yo (b)}.

If H(~ Y, P) is a [strict] saddle function on [a, b] x 1R 2d, then (Yo, Po) mini-
mizes A [uniquely] on
~min = {(Y, P) E~: Y' = Hp}
and maximizes A [uniquely] on

~max = {(Y, P) E~: P' = -Hy}.
PROOF. We shall establish the first implication and leave that for the
second to Problem 8.13. From its definition, we may write (with obvious
abbreviations)
A(Y, P) - A(Yo, Po) = r [H(x, Y, Po) - H(x, Y, P) - Y'· (Po - P)] dx
-r [H(x, Y, Po) - H(x, Yo, Po) - po· (Y' - Yo)] dx.
Now, Y E ~min => Y' = Hp(x, Y, P) and with the convexity of H~, X. P), it
(48')
follows that the first integral is nonnegative. Consequently, for such Y, we
1 It does not seem possible to use (47) to characterize strong convexity of f See Problem 8.28(b).
have after an integration by parts, that
A(Y, P) - A(Yo, Po) ~- Lb [H(x, Y, Po) - H(x, Yo, Po) + Po' (Y - Yo)] dx,
(48")
since the boundary term Po' (Y - Yo)l: = O. But for the stationary function,
P~ = -Hy(x, Yo, Po) (Why?), and now the convexity of -H~ Y, f) will en- ~
sure that the last integral is nonpositive. When strict convexity obtains in
both cases, then from the usual arguments of Chapter 3, we conclude that
P = Po and Y = Yo· 0
(8.2) Remarks. Theorem 8.1 does not require that H be a Hamiltonian asso-
ciated with a Lagrangian f In applications, we could take any function
P E (C 1 [a, b])d, and seek a solution
Y E!?}o = {Y E (C1[a, b])d: Y(a) = Yo(a), Y(b) = Yo(b)}
of the first-order system Y'(x) = Hp(x, Y(x), P(x», to obtain a (Y, P) E !?}min,
and hence an upper bound to A(Yo, Po). Similarly, choosing aYE !?}o, we
could use any solution P of the system P'(x) = - Hy(x, Y(x), P(x», to pro-
vide a lower bound for A(Yo, Po). For more information on this method of
complementary inequalities, see [Ar].
Now, in general, it is difficult to find (Y, P) E !?}min unless H is the Hamil-
tonian for a Lagrangian function f = f(x, Y, Z). Then each Y E !?}o provides
a P(x) ~ fy,(x) for which Y'(x) = Hp(x, Y(x), P(x» since this equation is a
consequence of the Legendre transformation; (recall (46'». This pair (Y, P) E
!?}min, and from Theorem 8.1 we would conclude that
F(Y) =derfb f[Y(x)] dx = A(Y, P) ~ A(Yo, Po) = F(Yo),

a
as we would expect from (47) and its convexity implications. However, this
same Yalso provides (in principle) a solution P1 =F P of the first-order system
P'(x) = - Hy(x, Y(x), P(x» for which (Y, Pd E !?}max. Thus A(Y, P1) is a lower
bound for F(Yo), and this could be of use in approximating an F(Yo) that
cannot be determined exactly.
It is easier to assess saddle function behavior of H on a product domain.

For example, in the one-dimensional case,
1~
H(y, p) = --v 1- p
Y
is a strict saddle function on the product domain {y > O} X {p2 < 1} since
Hyy < 0 < Hpp there; (recall Example 5 in §3.3). To find an associated
"Lagrangian" f = f(y, z), we use (46') to define
z = Hp = ~y h1 _ p2
(which has the sign of p),
and solve this equation algebraically to obtain

z
p= ------r==;;===;o
Jy 2 Z2· +
Then with this p and z, from (46) we see that
deC 1 p2 1 ~ 1 z
f(y,z) = pz-H(y,p) =- ~ +-y 1_p2 = ~ p'
yy 1- p Y yy 1 - p
so that
f(y, z) = J Y 2 + Z2 is strictly convex on {y > 0) x IR.
We can produce other functions in the same manner whose convexity is
not easy to establish by, say, the second derivative test of Problem 3.5. See
Problem 8.25. Now, let's utilize this ability and give new solutions to old
problems.
Example 1. The Cycloid Is the Brachistochrone.

To solve the brachistochrone problem of §1.2(a), we must find the
Yo E CEO, Xl] that minimizes the time-of-travel integral
1 IXI!¥+yI2
T(y)=- --dx
0 J2Y
y
on
!!}* = {y E CEO, Xl]: y(O) = 0, y(x l ) = Yl; T(y) < + oo}.
The integrand function J(1 + Z2)/y is not convex, but if we let y(x) =
y2(x)/2, so that y' = yji' the integral becomes
T(y) = - 1 IXI Jy-2 + ji'2 dx,

.;go
and we now recognize the new integrand function
f(ji, z) = Jy-2 + Z2
as being strictly convex on {ji > O} x IR!
Thus if Yo = Yo(x) describes the cycloid joining the origin to the point
(Xl' yd (see Example 4 of §6.2) then Yo(x) = J2yo(x) describes a curve that is
stationary for f on (0, xd. (This can be verified either directly or by appeal to
the in variance of stationarity in §6.8.) Finally lizl = Izl/Jy-2 + Z2 ::;; 1, and
we can use Theorem 3.5 as extended in Problem 3.21(b) to conclude that if
y E !!}*, then T(y) = T(y) ~ T(yo) = T(yo); hence Yo (so Yo) is the unique
minimizing function. We see that the cycloid is the brachistochrone because
a suitable time-of-travel integral is strictly convex. Apparently, enough speed
is attained in nearly vertical fall from (g to compensate for lower speeds in the
flatter parts of the cycloid even if we must travel upward to reach our
destination!
(x, y)(s) (i, y)(s)

y
/ /
Figure 8.4
As this example shows, sometimes a simple transformation can produce

an integral having a convex integrand function from one that does not.
Moreover, it may be easier to establish convexity of an integrand f by
showing that the associated Hamiltonian is a saddle function.
Example 2*. Dido's Problem.

One of the most important assertions of classical origin is the isoperi-
metric inequality of plane geometry (§1.3). It is closely related to the problem
of Queen Dido, concerning how to arrange a curve of length I in the upper
half-plane {y > O} with ends on the x-axis that together with this axis en-
closes the maximal area. (See Problem 3.32.)
We can suppose a typical curve to be parametrized with respect to its
arc-length s by functions x, y E C 1 [0, I] with X,2 + y'2 = 1, x(O) = 0, y(O) =
y(l) = 0 and y(s) > 0,1 S E (0, I). (See Figure 8.4.) By using Green's theorem
[Ed], it is not difficult to see that this curve encloses an area of size
A(x, y) = ILy(s)x'(s) ds I:::; Ly(s) Ix'(s)1 ds.

Moreover, the last integral is just A(x, y) when we define
x(s) = t Ix'(t)1 dt so that x'(s) = Ix'(s)1 = J1 - y'(S)2;
i.e., we can replace the given curve by a related curve that encloses an area at
least as large for which x' ~ O. Thus we only need to solve the nonisoperi-
metric problem of minimizing
F(y) = - L yJ1 - y,2 ds

on
!!} = {y E C 1 [0, I]: y(O) = y(l) = 0, y ~ O}.
1 This restriction will be relaxed later.

To learn whether f(y, z) = - y~ is convex we can use the second

derivative test of Problem 3.5 which in this case is very simple because
hY = 0. Since fyz i= 0, hyfzz - (hz)2 < 0, and we see that this f is not convex,
(nor is -f). However, f(y, z) is strictly convex (Example 5 of §3.3), and our
previous success leads us to look for a simple transformation. But which one?
The fact that yj1 - y,2 = jy2 - (yy')2, suggests trying y2(X) = 2Ji(x) so
that yy' = y' (which is just the reverse of the transformation used in Example
1!) This leads us to consider minimizing
F(y) = - I j2y - y'2 ds (49)
on
~ = {Y E C 1 [0, I]: Ji(O) = Y(l) = 0, F(y) < + oo}
and so asking whether j(y, z) = -j2y - Z2 is convex.
To find out, let's introduce the associated momentum
- z J2§p
P = fz(y, z) = j ,solve for z =
2y - Z2 Jl+Pi'
and substitute in (46) to obtain the Hamiltonian
H(y, p) = pz + j2y - Z2 = pz +: = (p + ~)~ = J2§Jl+Pi

p P 1 + p2
Now, when y > 0, H(y, p) is strictly convex as is -H(y, p), so that H is
a strict saddle function, and therefore by (47), j is strictly convex on
D={2Y<Z2}!
Dido believed that the semicircle of length I with ends on the x-axis gives
the minimizing curve, and it is easy to show that relative to arc-length, it is
parametrized by
°: :; s :::; I.
This function is stationary for f(y, z) = - y~ since
(f - y~fyo)(s) = - j/~ ,2 (s) = -1,

- Yo
and it follows that
_ ( ) _ y~(s) _ 12 . 2 (ns)
Yo s - -2- - 2n2 sm T
is stationary for j(y, z) = -j2Y - Z2 on (0, I). Here
.i:[Yo(s)] = y~ (s) = c~s(ns/I)

j2yo - y~2 sm(ns/I)
°°
is unbounded near s = and s = I. However, sfz[Yo(s)] and (I - s)fz[Yo(s)]
are bounded near s = and s = I respectively while to be in ~, y must have
Y'(O) = Y'(I) = 0 (since Y'2 < 2Y). Thus we can use Theorem 3.5 as extended
in Problem 3.21(d) to conclude that the semicircle is the unique minimizing
curve for F, and so we verify Dido's conjecture that it alone encloses the
maximal area against the x-axis.
(8.3) Remarks. 1. By Theorem 7.7, the semicircle also gives the maximal area
among polygonal or other piecewise C 1 curves (x, P) with P(s) > 0 on (0, I),
and this conclusion remains valid when we permit curves (x, P) with P(s) = 0
at some points in (0, I). Indeed at such points y'(s) = ji(s) = 0, where y = p2j2,
and when Yo = yo(s) is as before,
jm Y') - j(yo, yb) - ]'(Yo, yb)(y - Yo) - jAyo, yb)(Y' - yb)
- -,2 -
= 12 - _ -,2 _ Yo + Yo = Yo >0
v' Yo Yo /2- -,2 /2- -,2 /2- -,2 .
v' Yo - Yo v' Yo - Yo v' Yo - Yo
Thus the fundamental inequality holds even at these points.
2. It is possible to use these results to establish the isoperimetric inequal-
ity (Problem 8.26). (A different derivation of this inequality will be given in
§9.5, Example 3.) However, we can also use this approach to solve other
significant problems, including that of Zenodoros in Problem 1.9. See
Problems 8.27-8.30.
§8.9. Continuous Media

Thus far we have analyzed dynamical systems consisting of a given finite
number of "point" masses, and applied Hamilton's principle to obtain the
equations of motion. Obviously the finiteness becomes questionable when we
attempt to describe the motions of an elastic medium which is undergoing
dynamic deformation as occurs, for example, when a drum head is struck.
Fortunately, in many cases of importance, Hamilton's principle still applies
to an appropriate Lagrangian function, which is now represented by spatial
integrals. We shall illustrate how this is accomplished with two standard
elastic media, the taut string and the membrane. (The elastic bar and plate
are considered in Problems 8.20 and 8.21.) The approach remains valid for
more general continuua including elastic solids, fluids, and gases, as well as
for electromagnetic phenomena. [See C-H, We, Bi, W, O-R.] Moreover, in
the absence of motion, we obtain Bernoulli's principle which characterizes
static equilibrium for these systems.
(a) Taut String

In order to analyze the transverse planar motions of an elastic string (such as
a guitar string) which is stretched horizontally between fixed supports sepa-
§8.9. Continuous Media 261
u(t, x)
Figure 8.5
rated a distance 1 as shown in Figure 8.5, we suppose that the "vertical"

position of the string at any time t is described by u = u(t, x), 0 :$; x :$; l. Then
the resulting velocity history ut(t, x) gives rise to a kinetic energy at time t
given by
T= ~I pu~dx, (50)
where p = p(t, x) is the (local) linear mass density of the string which is
assumed continuous but may be nonuniform. (T may be considered as the
limit of the ordinary kinetic energy of a large finite number of masses
m i ~ p(t, xi)Ax i located at Xi with 0 = Xo < Xl < ... < Xn = l, and moving
with vertical velocity ut(t, Xi), when the AXi = Xi - X i - l tend toward zero.)
As usual, it is less apparent how to select the potential energy function U

to complete the Lagrangian, and we must now make additional assumptions
about the motion. First, we shall suppose that the string is so thin that the
only work required to deform it from the straight line (u = 0), to a position
described by u(t, x) is that necessary to enlongate it against a local tensile
force r = r(t, x). In particular, we consider the work in bending the string to
be negligible. (Alternatively, we could suppose that the curvature of the
deflected string is so small that ~ as determined in §6.6, is small in compari-
son with We.) In any case, we shall assume that the potential energy stored in
the string as a result of its deformation is the work required to stretch it and
hence by the same analysis as in §6.6, take
U = - We = I r(JI + u; - 1) dx, (50')
at time t. We are also assuming that there is no external loading on the string
and thereby disregard the effects of gravity. (However, see Problem 8.17(c),
(d), and ESe] for less restrictive assumptions.)
From (49) and (50) we obtain the Lagrangian
L =T- U= I npu~ - r(JI + u; - 1)] dx, (51)

and Hamilton's principle now requires that between times 0 and b, say, at
which the positions are known, the motion will make stationary the action
integral
A(u) = J: L dt = J: dt I [tput - r(J1 + u~ - 1)] dx (52)

on
!!) = {u E C 2 ([0, bJ x [0, IJ): u(t, 0) = u(t, I) = 0; u(O, x), u(b, x) prescribed}.
Upon introducing
f(x, u, Vu) = tput - r(J1 + u~ - 1), (53)
we see that this integrand function depends on X = (t, x) E 1R2 through p and
r, and on Vu = (un uX>, but does not depend explicitly on u. (See §6.9.) Thus
by Theorem 6.13, we should seek a C 2 solution u to the equation
V"fvu = 0,
or upon substitution of (53), the equation
(put)t - (ruA1 + u~rl/2)x = 0, (54)
which meets the geometrical conditions
u(t, 0) = u(t, I) = 0, 0:::;:; t:::;:; b. (55)
In addition, we suppose that the initial position is prescribed by
u(O, x) = uo(x), 0:::;:; x:::;:; I, (56)
so that from (55), uo(O) = uo(l) = o. However, instead of attempting to specify
the position at a later time b, we seek a solution valid for all later times, with
the prescribed initial velocity
O:::;:;x:::;:;l. (57)
We observe that (54) is nonlinear, and for example, will not admit a
pure time-oscillatory solution of the form u(t, x) = u(x) cos rot. Numerical
methods must be used to obtain approximate solutions to this equation.
However, as in §6.6, we can easily linearize it by supposing that the slopes
IUxl ~ l.
Then (54) reduces to
(put)t = (rux)x, (58)
which for constant p, r becomes the one-dimensional wave equation
Utt = (12U xx , with (12 = rj p. (58')
Because of its importance to the description of simple acoustical phenom-
ena, (58') was the first partial differential equation to receive serious atten-
tion. It admits the (unique) solution with given initial data first obtained by
D'Alembert (1746) in the form
u
(t, x ) =
uo(x + crt) + uo(x -
2
crt) 1
+ -2
cr
f
x + at
x-at
Vo
()
s s.
d
(59)
(Problem 8.14.)
Here we must suppose that the initial functions U o in C2 and Vo in C t are
given on all of~. (However, to satisfy the end point conditions (55), the initial
data must be extended periodically.) In particular, when the initial velocity
Vo == 0, this solution at time t may be interpreted as the average of the
translate of the initial shape U o which is travelling to the right at velocity cr,
with that of the same shape travelling to the left at velocity cr. Moreover, for
constant p and r, (58') does admit time-oscillatory solutions in the form
u(t, x) = uo(x) cos wt, (60)
provided that U o in C 2 is a solution of the linear ordinary differential equation
u~ + (wlcr?u o = 0, (61)
as is easily verified by direct substitution.
Now, for constant Il = wlcr, the general solution to (61) is well known to
be given by
uo(x) = A cos IlX + B sin IlX.
In addition, U o must satisfy the end conditions uo(O) = 0, uo(l) = O. The first
of these conditions requires that A = 0, so that uo(x) = B sin IlX. However,
the second can now only be satisfied (for U o -# (7), when sin III = 0; i.e., for
Il = Iln = nnll, n = 1, 2, ....
It follows that simple oscillatory solutions of the form (60), can occur only
for specific "natural" frequencies
w = Wn = NP(n;) = nwt , n = 1,2, ....
Observe that these natural frequencies increase with the tension, r, and
decrease as the density, p, or the length, 1, are increased, facts confirmed by
experience with, say, a guitar string. Moreover, from experimental evidence,
Mersenne, in his Harmonie Universe lie of 1636 predicted the exact square
root dependence on rip, while other properties of W t were known to the
Pythagorean school.
The mode shape associated with oscillations at frequency Wn is
Un ()
X . Iln X = SIn
= SIn . (nnx)
-1- ,
demonstrating that the sinusoidal shape accompanying higher frequencies

has more nodal or fixed points. The resulting "pure" motion is described by
(62)
These natural vibrations are to be understood in the sense that a motion

which is begun with initial deflection of pure sinusoidal shape un(x) and initial
velocity 0, should exercise small amplitude oscillations in this shape indefi-
nitely at frequency ron' as a result of precise interchange between kinetic and
potential energies. (Of course, a real motion would die out in time as a
consequence of the fact that the string is not perfectly elastic.)
Now, what happens, if the string is not released in one of these pure
shapes, but is instead plucked; that is, released in a triangular (or some other)
shape? In 1753, D. Bernoulli argued that since overtones could be detected
audibly, it was reasonable to suppose that the resulting motion might be
represented by superposition of the natural modes in the form of an infinite
series. However, Bernoulli's arguments were criticized by Euler, D'Alembert,
Lagrange, and Laplace so severely, that when Fourier reawakened interest
in such series representation as possible solutions to problems in heat con-
duction (c. 1807), his work was also regarded with suspicion. The resulting
questions of Fourier series representation exerted a profound influence on
the subsequent development of mathematical analysis, and interest continues
to this day. We state without proof, the following result (which is not the best
possible):
(8.4) Proposition. If U o E C 4 [0, I], with uo(O) = uo(l) = u~(O) = u~(l) = 0, then
the coefficients bn ~ (211) J~ uo(x) sin(mrxll) dx, satisfy for some constant M, an
estimate of the form Ibnl :$; Mln\ n = 1,2, .... Hence, the series
converges to the unique solution of (58') which satisfies

u(O, x) = uo(x), X E [a, b].
(See Problem 8.15.) o
(It is a consequence ofD'Alembert's result (59) that these initial conditions
(as extended periodically) do determine the solution uniquely. An alternate
proof of this fact is given in the next proposition.)
The Nonuniform String

We shall discuss briefly the case of a nonuniform string in which p and 't' are
permitted to vary with x but not with t. Then (58) becomes
pUtt = ('t'uxl x ,
which would also follow as the equation of motion for the modified
Lagrangian (51) in which U is replaced by fj = J~ 't'u; dx. t
(8.5) Proposition. When p and T are positive with Pt = T t = l1J, there is at most
one C Z solution u to the wave equation pUtt = (TUJx on [0, (0) x [0, I] with
given initial data (56) and (57), which satisfies the boundary conditions (55):
u(t, 0) = u(t, I) == O.
PROOF. If U 1 and U z are two solutions to this problem, then u ~ U 1 - Uz
provides a solution to the same equation with zero initial data. Introducing
the associated total energy at time t
E(t)~ T + fj = t I (pu; + TU;) dx,

it follows from A.13 that since Pt = Tt = l1J:
E'(t) = I (pUtU tt + TUxU xt ) dx
= I ut(pUtt - (TUx)x) dx + (ruJut I~,

after an obvious integration by parts using the fact that Uxt = utx .
The integral vanishes (Why?) and from (55) follows ut(t, 0) = ut(t, I) == O.
Thus E'(t) = 0, so that E(t) = E(O) = 0, since ut(O, x) is known to vanish while
u(O, x) == 0 implies the vanishing of uAO, x). But then by the standard argu-
ment we conclude that the nonnegative continuous integrand of E(t) must
vanish identically on [0, I]. From the positivity of P and T, we infer that at
time t > 0: ut(t, x) = uAt, x) == 0, so that u(t, x) = const. = u(t, 0) = O. Thus
u 1 == U z as we wished. 0
This argument admits modification. (See Problem 8.17.)
If we again consider oscillatory solutions of the form (60), we see that the
mode shape U o is now required to satisfy the equation
(TY')' = _O)zpy (63)
and the homogeneous boundary conditions
y(O) = y(l) = O.
This is a problem of the Sturm-Liouville type encountered in §7.3 for a posi-
tive eigenvalue A. = O)z. In Problem 7.19, it is shown that a solution U 1 which
is positive on (0, I) should minimize
F(y) = I TY'Z dx
on !1JJo = {y E c 1 [0, I]: y(O) = y(l) = O} under the (achievable) constraint
G(y) = I pyZ dx = 1. (64)

It follows that the corresponding eigenvalue A1 determines the fundamental

or least frequency W1 = A, and from Problem 7.19, that
A = R( )~F(U1)
1 u1 G(u 1 )
minimizes the Rayleigh ratio
R = FIG on ~ = {y E C 1 [0, 1]: y(O) = y(l) = 0; y -# (D}.
[Indeed, when y -# (D, then R(cy) = R(y), while G(cy) = c2 G(y) = 1, for some
choice of the constant c -# o. Thus on ~, R assumes the same values as does
F under the additional constraint (64).J
This reformulation of the problem initiated by Lord Rayleigh (c. 1873)
together with a corresponding characterization of the (necessarily) higher
eigenvalues An provides the basis for most results in this subject. [See §A.6,
C-H, W-S.J We observe here that each y in ~ provides an upper bound
J R(y) to the unknown fundamental frequency W 1 . Moreover, direct com-
parison can be made between the frequencies Wn and those, ron' of an appro-
priately chosen uniform string (See A.23). Finally, we may replace the fixed
end conditions by other homogeneous boundary conditions as in Problem
7.19(h». Again superposition of the pure motions is possible and there
emerges a theory of generalized Fourier series. Indeed, the Sturm-Liouville
problem considered here and its generalizations have played a vital role in
developing mathematics and its applications. See Kline, [TrJ, [WeiJ, and
[C-H].
(b) Stretched Membrane
A two-dimensional analogue of the stretched string is provided by a thin

elastic membrane or skin which is stretched tightly over a (horizontal) frame
in the shape of the boundary of a domain D of 1R2 as illustrated in Figure 8.6.
We may think of the membrane as the head of a drum for which both the
LOADING
x
Iu
DISTRIBUTED TENSILE FORCE
Figure 8.6
local density and tensile stress may vary with time, as well as position, as
would occur when a kettle drum is tightened just after it is struck.
Let X = (x, y). Then we shall assign the membrane a local areal density
p = p(t, X) so that if u = u(t, X) denotes the vertical displacement at time t,
then the associated kinetic energy is given by
T=:21 f D pUt2 dX, (65)
where dX = dx dy denotes the element of area.

We require that at the boundary iJD
UlaD = 0, Vt ;?: 0.
Again we shall neglect the strain energy of bending in comparison with that
of areal stretching against the distributed local tension. = .(t, X). However,
we shall also suppose that there is a downward distributed loading giving rise
to the pressure p = p(t, X) on the membrane, possibly, its own weight.
Then the resulting potential energy at time t is given by
U = Iv .(J1 + u; + u; - 1) dX - Iv pu dX. (66)
Here, the first term represents the work done in stretching the membrane
as in §3.4(e), while the second term is the work done in moving the distributed
load under deflection. In both cases we take as reference state that described
by u == 0.
The resulting Lagrangian is, of course,
L = T- U,
and Hamilton's principle requires that between times t = and t = b at
which the positions of the membrane are prescribed, it should execute a
°
motion which makes stationary the action integral
A(u) = J: L dt = J: Iv
dt (ipu; - .(J1 + u; + u; - 1) + pu) dX (67)
on
q) = {u E C 2 ([0, bJ x D): UlaD = l!!, u(O, X), u(b, X) prescribed},
Thus, denoting the integrand by
f = f«t, X), u, Vu),
we have from Theorem 6.13, that u must satisfy the differential equation
V"ivu = fu; upon substituting, and subsequently utilizing the linearizing
assumption that u; + u; ~ 1, it follows that u should satisfy (approximately)
the partial differential equation:
t ;?: 0, XED. (68)
(Problem 8.18).
In addition, we require that ulaD == 0, while u(O, X) = uo(X) is prescribed,

as is ut(O, X) = vo(X), in lieu of u(b, X). (68) is linear, but it is nonhomoge-
neous because of the presence of the external forcing term p = p(t, X). It can
have at most one solution for the given data when p and 1: are positive and
time independent (Problem 8.18).
When p == 0, p and 1: are constant, and Vo == 0, we may attempt an oscilla-

tory solution of the form u(t, X) = uo(X) cos wt, and see that in D, U o must
satisfy the equation
(69)
with
A = (p/1:)w 2 , and Au ~ U xx + U yy .
Of course, we require that uolaD == 0. (Problem 8.18(e».
This is known as the membrane eigenvalue problem for the domain D.

Again, only certain values of the eigenvalues A = An ~ 0, n = 1, 2, ... , will
permit a solution, and these will determine the natural frequencies, Wn =
(dn/p) 1/2, of free oscillations. The associated mode shapes uiX) determine
pure motions un(t, X) = un(X) cos wnt, corresponding to those of a drum
head which has been struck or excited in an appropriate manner. Again we
see that these natural frequencies increase with the tension 1: and decrease
with the density p of the material as experience dictates. Moreover, the
frequencies together with the mode shapes depend on the shape of the
bounding curve(s). The interesting long-open problem of whether we can
"hear" the shape of a drum; i.e., of whether we can recover the shape of the
bounding curve from knowledge of these natural frequencies, has been re-
solved quite recently. We can't! 1
The pure oscillations un(t, x) can again be combined to obtain (formal)

series solutions with undetermined coefficients which may in principle be
chosen to make the resulting solution describe a given initial displacement,
U o = uo(X), provided that U o satisfies certain regularity conditions. Theoreti-
cally, in this manner we can solve the homogeneous two-dimensional wave
equation
(70)
for constant p, 1:, with given initial displacement uo, and initial velocity vo, to
obtain the resulting unforced motions (p == 0) of the membrane.
There are associated series methods which may be employed to find forced
motions, making further use of the linearity of (68), but the actual computa-
tions, even for the homogeneous case, are quite difficult. They have been
1 See "One Cannot Hear the Shape of a Drum" by C. Gordon, D. Webb, and S. Wolpert,
Bulletin of the A.M.S., Volume 27, No.4, July, 1992 pp. 134-137.
carried out only for a few simple domains, D, such as the disk, the rectangle,
and the annulus. See [Tr], [Wei], and [C-Hl
Static Equilibrium of (Nonplanar) Membrane

°
In the absence of motion, U t == and hence from (65), T = 0. Moreover, when
p = p(X), U as defined by (66) depends only on the position u = u(X). Thus
the action integral (67) becomes
A(u) = - J: U dt = -bU(u), say,
which should now be made stationary on

q) = {u E C 2 (D): ulan = y},
where y E qoD) is prescribed.
Inspection of the integrand function of U in (66) in the form
f(X, u, Vu) = r(K)(J1 + u; + u; - 1) - p{K)u (71)
shows that it is strongly convex. (This is true for the first term by the argu-
ment of §3.4(e), while the second term is linear in u.) Hence, the only u which
can make A or U stationary on q) is that which minimizes uniquely the
potential energy U. (Problem 8.19.) Under a prescribed static loading p, the
membrane assumes a unique shape U o of stable equilibrium.
For a homogeneous membrane, r = constant, and U o will satisfy (approxi-

mately) the linearized equation (68) with U tt == 0; namely,
du = -pir,
with
ulaD == y.
This is Poisson's equation with a Dirichlet boundary condition. When the
external loading p == 0, it is Laplace's equation which is clearly satisfied by
u == 0. Hence, when y == 0, this is the unique solution reflecting the fact that
we have ignored the actual weight of the membrane.
The above analysis applies also to the nonplanar membrane in which
ulaD = y ¥= 0. Then, in the absence of loading, we are simply solving the
Dirichlet problem for D. With a smooth boundary function y, it always has a
unique solution (as does the corresponding Poisson problem [C-H]).
However, should we seek a solution to the actual nonlinear equation
governing the equilibrium position of the nonplanar membrane, then in the
absence of loading, we have in effect the minimal area problem which as we
have discussed in §3.4(e), could have at most one solution, but need not have
any solution.
PROBLEMS
8.1. (a) Show that if the approximations
M/(cos 8)xO ~ MZXO,
82
Mg/(1 - cos 8) ~ Mg1 2 ,
(for small 8) are made in the Lagrangian (11), and we suppose that
K(X) = KX, then the corresponding equations of motion are given by
(16).
(b) Explain why it is not inconsistent to use both of the approximations
cos 8 ~ 1 and cos 8 ~ 1 - 82 /2 in the same equation in part (a).
8.2. Let Q = G)'

(a) Show that equations (16) can be written in the form
Q=AQ, (72)
where the coefficient matrix A is given by
(b) Define Ql(t) = Vcos wt and Q2(t) = Vsin wt, where V#-O is a constant
vector and w is a (real) constant. Show that Ql and Q2 are solutions of (72)
iff AV = -w 2V; i.e., V is an eigenvector of A with corresponding eigen-
value _w 2 •
(c) Show that for any positive choices of m, M, I, K, the eigenvalues;' of A are
always negative. Are the eigenvalues necessarily distinct? (An eigenvalue
;. is a root of the equation det [A - 1I] = 0.)
(d) Conclude that if A has real eigenvalues - wi < - w~ < 0 with correspond-
ing eigenvectors VI and V2, then the solution of the initial value problem
Q = AQ with Q(O) = Qo and Q(O) = Qo can always be expressed in the
form
Q(t) = (c 1 cos wIt + c1 sin W1 t)V1 + (C2 cos w2t + C2 sin W2t)V2
where c 1 , c1 , C2, C2 are constants. (Hint: Eigenvectors associated with

distinct eigenvalues are always linearly independent.)
8.3. A double pendulum consists of two light inextensible rods of length I and two
bobs of mass ml and m, respectively, which are constrained to move in a
vertical plane as shown in Figure 8.7. Assume that the pivots are frictionless.
Use the generalized coordinates 81 and 8.
(a) Express T, U, and L in terms of 81 , 8, 01 , 0.
(b) Determine the differential equations of motion.
(c)* Find solutions for the linearized equations by the method of the previous
problem.
Problems 271
Figure 8.7
8.4. (a) Verify that the canonical equations (25) are indeed the first Euler-
Lagrange equations for the generalized action integral (26).
(b) Show that the variation of A at (Q, P) in the direction (V, W) is given by
M«Q, P); (V, W» = f [(Q - Hp)· W - (P + Ha)· V](t) dt + (p. V)(t) I:.
(c) Conclude that functions (Q, P) satisfying (25) make A stationary on a set
where only Q(a) and Q(b) are prescribed. Is the converse true?
z
/.
(x, y, z)
x
Figure 8.8
8.5. (Spherical Pendulum.) A bob of mass m is attached to a light inextensible rod

of length I which is free to swing in any direction about a fixed pivot as shown
in Figure 8.8.
(a) Use the relations

x = I sin cp cos e,
y = I sin cp sin e,
z = a + I cos cp,
to express the Lagrangian in terms of the generalized coordinates Q == (cp, e).
(b) Find the associated differential equations of motion and deduce that
ml 2 (sin 2 cp)iJ = const. (which corresponds to conservation of angular
momentum about the z-axis).
(c) Determine the conjugate momenta P == (Ptp' Po).
(d) Express Qin terms of t, Q, P, and find the Hamiltonian H(t, cp, e, Ptp' Po).
(e) What are the canonical equations of motion? (Compare with part (b).)
(f) What is the Hamilton-Jacobi equation (39) for this system?
8.6. Consider a dynamical system with one degree of freedom. If the Hamiltonian
does not depend explicitly on time, the Hamilton-Jacobi equation takes the
form
as + H (as)
at q, aq = O. (73)
The following argument can sometimes be used to find a complete solution of

(73).
(a) Show that S = cp(t) + s(q) is a solution of (73) iff
cp(t) = clt + Co
and
H(q, s'(q» = -C l , (74)
where Co and C l are constants. Observe that (74) is a first-order ordinary
differential equation for s.
(b) Conclude that if s = s(q; cd satisfies (74) for each choice of the constant
C l , then
s = Cl t + s(q; c l ) + Co
is a complete solution of (48).
(c) Find a complete solution of
as + e_ q
at
(aS)3
aq
= o.
8.7. Assuming that the sun is fixed at the origin, the planar motion of a single planet
of mass m about the sun may be specified by giving the polar coordinates (r, e)
of its position at each time t. The potential energy function which recovers the
inverse square law is U = - mk/r for an appropriate constant k.
(a) Show that the associated kinetic energy function for Q == (r, e) is
T = (m/2)(r2 + r2iJ2).
(b) Obtain the Hamiltonian H in terms of the conjugate momentum P ==
(Pr' Po), and write the reduced Hamilton-Jacobi equation H(Q, sQ) = E =
const.
Problems 273
(c)* Attempt a solution of this equation by separation in the form

s = s(r, 0) = R(r) + 0(0), say.
(d)* Compare the equations for the separated variables with those of
Lagrange for this problem.
8.8. The non planar motion ofthe planet of mass m about the sun (assumed fixed at
an origin), may be described in terms of spherical coordinates (r, qJ, 0) (as in
Figure 1.1), with the same energy function U = -mk/r as in the previous
problem.
(a) Prove that the kinetic energy function is T=(m/2)(r2 +r2riJ2 +r2(sin2 qJ)82).
(b) Derive the Hamiltonian
H =1(2P;
- P +-+ Pt)m k
--
2m r r2 r2 sin 2 qJ r
in terms of P == (Pro P"" P9). Hint: Use (27).
(c)* Show that the reduced Hamilton-Jacobi equation H(Q, sQ) = E = const.
may be separated successively as follows:
8s
80 = IX;
(_8s)2 + sin
8qJ
__ =p2.
2
1X2
qJ ,
for constants IX, p.

8.9. Verify that the functions which are stationary for the modified action
integrand of equation (30) satisfy the remaining equations of motion (25), for
j ¥- i.
8.10. (a) Apply the result of Theorem 6.10 to obtain equations for the stationary
functions minimizing the action integral of equation (33) subject to the
general constraint K(Q, P) = 0, in the form
for an appropriate function). = ).(-r).

J
(b) Prove that in terms of the new parameter f~ ).(-r) d7:, the equations in (a)
will take the form (36).
(c) Show that for the special choice of constraining function
K(Q, P) == Po + H(Q, PI' P2'···' P.),
equations (36) are the canonical equations of Hamilton corresponding to
taking 7: = t.
(d) Explain how this same analysis would apply to Jacobi's action integral of
equation (31) subject to the constraint H(Q, P) = const.
8.11. For a single particle of mass m moving freely under the action of a force of
potential U = U(X), we may use as generalized coordinates Q, the cartesian
coordinates X E [R3 of position.
(a) Show that the associated momenta P = mX.
(b) Conclude that the Hamiltonian is
H = H(X, P) = IP1 2 /2m + U(X).
(c) Make the Schrodinger substitution s = k log'" in the corresponding

reduced equation (44) and prove that'" satisfies the equation
for a constant E.
(d) Consider for this f,
F(u) = f U{3
f(X, u, Vu) dX,
and ignoring questions of convergence, show formally that a function '"

which makes F extremal on ~ = {u E C 2 ([R3): F(u) < oo} subject to the
constraint
G(u) == f U{3
u2 dX = 1,
should satisfy the three-dimensional equation similar to (69),
where now E is to be regarded as a Lagrangian multiplier.

(This indicates Schrodinger's method (1926) of associating a wave character to
a mass particle. The permissible eigenvalues of E which permit a solution", of
the equation meeting certain homogeneous boundary conditions define the
energy levels ofthe particle. [We].)
8.12. In the theory of special relativity in which we ignore the effects of gravitation
and postulate the constancy of c, the speed of light in a vacuum, we may
modify the Lagrangian so that the form of Hamilton's principle remains valid.
For the case of a single particle of constant (rest) mass mo, moving freely
in an electromagnetic field of vector potential U(t, X) and scalar potential
U = qqJ = qqJ(t, X), at a speed v = lXI, we may use as the Lagrangian
L = -moc 2 (1- v2 jc 2 )1/2 - U + qX'U

(where q is a constant) and take the generalized coordinates Q = X E [R3.
(a) Show that when U = (9, the particle has momentum
p = Lx = moPX, where P = (1 - v2 jc 2 t 1/2 ,
and hence even when U = 0, it has an apparent mass m = mop as a result

of its motion. When is m = mo ?; m ~ mo?; m ~ mo?
(b) Prove that the associated total energy
-def - • - mov 2
E = Lx'X -L = mc 2 + U = moc 2 + U +--r==~=.<
1+ v2 jc 2 J1 -
so that when v2 ~ c2 ,
Conclude that even when at rest with U = 0, the particle should contain
the enormous store of energy Eo = m oc2 , which is potentially available
for release.
Problems 275
(c)* Show that the corresponding equations of motion may be written
d .
-(mopX) = -q[U, + 'ilqJ - (V x !B)],
dt
where V = X, !B = curl U, and the last term is a vector product in [R3.

8.13. Complementary Inequalities (§8.8). To establish the inequality for ~m.x given
in Theorem 8.1:
(a) Consider A(YQ, Po) - A(Y, P) by simply interchanging the appropriate
subscripts in (48').
(b) Conclude that since y~ = Hp(x, YQ, Po), the first integral in (48') after inter-
change is nonnegative.
(c) Integrate the last term by parts as in the proof of 8.1 to prove that when
y E ~m... then also the second integral (with its minus sign) is nonnegative.
(d) Formulate a corresponding theorem when the convexity properties for H
and - H are reversed.
(e) Apply Theorem 8.1 to the function
H(x, Y, P) = IPI 2 -I Yl2 + cy.
8.14. D'Alembert's Solution.

(a) Verify that when p and T are constant, then u(t, x) as defined by (59) is a
solution of the wave equation (58'), satisfying (56) and (57).
(b) Take Vo :; 0, and decide how the displacement uo(x) defined initially on
[0, I] must be extended to [R if (59) is to give u(t, 0) = u(t, I) = 0, for t ~ O.
Give an example of a function U o # (!) so extended.
8.15. Fourier Series Solution. (Take I = 1, for simplicity.)

(a) Verify that the coefficients bn as defined in Proposition 8.4, do satisfy an
estimate of the form Ibnl ~ M/n4, n = 1, 2, .... Hint: Integrate by parts
four times.
(b) Use Proposition 8.4 and the hint in (a) to find a formal series solution to
(59) when uo(x) = x 3 (1 - X)3.
(c) Show that the series u(t, x) = I:'=l bnun(t, x) converges uniformly on [0, 1],
as does each of the differentiated series I:'=l bnun(t, x), I:'=l bnu;(t, x)
where the dots and primes denote differentiations with respect to t and x,
respectively, and un(t, x) is defined by (62) with I = 1.
(d) The convergence established in (c) proves that u, as defined, is C 2 • Verify
that formally at least it must satisfy (58'), with constant p, T.
(e) Observe that u,(O, x) :; 0 and u(O, x) = I:'=l bn sin nnx. Conclude that
bn = 2 g u(O, x) sin(nnx) dx, n = 1, 2, .... Hint: Integrate term-by-term
using integral tables if necessary to eliminate all but one of the terms.
(f) We would like to prove that g(x) :; u(O, x) - uo(x) = O. (Why?) Try to
invent an argument which could establish this using the fact that from (e)
J: g(x) sin nnx dx = 0, n = 1,2, ....

8.16. The vertical column of water rotating at constant angular velocity w (from §2.3)
may be analyzed via Hamilton's principle.
(a) Using the coordinates of Figure 2.1, show that at time t, the kinetic energy
(of rotation) is T = !p g 2nx(Ib (WX)2 dz) dx = npw 2 I~ x 3 y(t, x) dx, where
y = y(t, x) describes the profile of the upper free surface.
(b) Demonstrate that the potential energy (due to gravity) associated with the
deflected upper surface is U = pg I~ 2nx(Jb z dz) dx = npg I~ xy2(t, x) dx.
(c) Assuming that the positions of the rotating system are specified at times
o and b, apply Hamilton's principle to the action integral
A = A(y) = J: L dt = J: (T - U) dt,
subject to the volume constraint
G(y) = J: dt I 2nxy(t, x) dx = const.
(d) Conclude that the only profile of stationarity; viz.,

y = y(x) = (-1 + w 2x 2)f2g
for an appropriate constant 1, maximizes the action integral. Hint: Use
convexity.
(e)* Contrast the conclusion of (d) with the principle of minimum potential
energy used to obtain the same profile in §2.3.
(f) Repeat the entire analysis when w is permitted to vary with t.
8.17. Uniqueness of Solutions to the Wave Equation.
(a) Show that the conclusion of Proposition 8.5 is unaffected if (55) is replaced
by the conditions u(t, 0) = c t , u(t, 1) = C2, for constants c t , C2.
(b)* Find other end point conditions which preserve the uniqueness argument
and describe the physical situation represented by your choices.
(c) The presence of an additional (downward) distributed loading of intensity
p(t, x) adds a term - I~ pu dx to U. Conclude that the resulting equation
of motion is (pu,), - (ru,J" = p, when u; ~ 1.
(d) Argue that the conclusion of Proposition 8.5 will be unaffected by the
addition of a prescribed loading term p as in (c). Hint: What equation will
the difference u = Ut - U2 satisfy?
8.18*. The Membrane Problem.
(a) When u; + u; ~ 1, why is 0 =! I [r(u; + u;) - pu] dX a good approxi-
mation for U of (66)?
(b) Verify that with L = T - 0, the equation of motion is (68).
(c) Using the total energy function at time t given by
E(t) = T + 0,
establish a uniqueness result for solutions to (68) modelled after Proposi-
tion 8.5. Hint: Use Green's theorem to show that for the difference u of
solutions,
where N is the outward pointing unit normal to the boundary curve from D.
Problems 277
(d) Find other boundary conditions which would also guarantee uniqueness
of solution, and describe the membranes so supported.
(e) Verify that when p and or are constant while p = 0, and U o is a solution of
(69), then u(t, X) == uo(X) cos rot will be a solution of (68) satisfying the
condition ut(O, X) = vo(X) = O.
8.19. (a) Verify the strong convexity of f(}{, u, Vu) as defined in (71). (Problem
3.26).
(b) Obtain the associated (nonlinear) equation characterizing a function
which is stationary for this f (Theorem 6.13).
(c) Conclude that each solution U o E !?) of the equation found in (b), must in
fact minimize U on !?) uniquely.
(d)* Can convexity be used to give a uniqueness argument for solutions to the
time dependent wave equation in any of the forms considered in §8.9?
8.20*. Transverse Motion of an Elastic Bar. For a horizontal thick elastic bar of
constant rectangular cross-section and length 1 in which the energy of stretch-
ing may be neglected in comparison with that of bending, let u(t, x) denote the
"vertical" position of the center line at time t.
(a) Argue that for a suitable density p there should' be an associated kinetic
energy T = t S~ put dx.
(b) If the bar is subjected to a distributed downward loading of intensity
p(t, x), argue that the resulting potential energy is approximately
-
U = J(I {]:JlUxx -
0
1 2
pu) dx,
for a suitable material stiffness function Jl = Jl(t, x). Hint: See the discus-
sion of WB in §6.6.
St
(c)* For the action integral A(u) = (T - U) dt, use the definition liA(u; v) =
lim£~o (djde)A(u + ev) to prove that
liA(u; v) = J: dt f~ (putVt - JlUxxVxx + puv) dx.

(d) Suppose that u E C\ and integrate the expression in (c) by parts as re-
quired to show that
liA(u; v) = J: dt I[ -(pUt)t - (JlUxx}u + pu]v dx
+ f~ putv/: dx - J: I: + J:
(JlUxx)Vx dt I:
(JlUxx)x V dt.
(e) Conclude that stationarity of A for prescribed u at times 0 and b, and, say,
the cantilever support conditions u(t, 0) = ux(t, 0) = 0; u(t, 1) = u,,(t, 1) = 0,
requires that
(75)
(f)* When Pt = Jlt = 0, use the total energy function E(t) = T + U to give a
uniqueness result modelled after Proposition 8.5 as extended in Problem
8.17.
(g) Study the expression in (d) to obtain alternate sets of boundary conditions
which would lead to the same differential equation as in (e).
(h) When p = Jl. = const. and p = 0, determine an ordinary differential equa-
tion for U o = uo(x), if u(t, x) == uo(x) cos wt is to be a solution of (75).
(i) Can you find or guess a nontrivial solution Uo of the equation in (h) which
will permit u to meet the cantilever conditions in (e) at least for certain
values of w? Hint: See §6.6.
8.21 *. Transverse Motion of a Uniform Plate. If the membrane discussed in §8.9(b) is
replaced by a plate of uniform thickness and material, we may neglect the
energy of stretching in comparison with that of bending which is now given
(approximately) by
U= ~
2
r [(u;x + U;y) - 2(1 - 1:)(u"""uyy - UXy )2] dX,
JD
(76)
where, of course, u(t, X) denotes the vertical position of a center section at time
t, and Jl. and 1: < 1 are positive material constants.
(a) Argue that for an appropriate constant density p, the kinetic energy of
motion at time t should be approximately T = ! ID put dX.
It
(b) Set A(u) = (T - U) dt, and, neglecting external loading, reason that for
some boundary conditions, stationarity of A at u E C4 requires that u
should satisfy the equation
(77)
where
.1 2 u = .1(.1u) = .1(u""" + Uyy ).
(c) Which equation is U o = uo(X) required to satisfy in order that u(t, X) =
uo(X) cos rot be a solution of (77)?
(d) For static equilibrium of the loaded plate with pressure p = p(X), when
all functions are time independent, use convexity of the integrand of
U~U- LpUdX
to conclude that even for a nonplanar plate, only stable equilibrium is
possible, and it is uniquely characterized by a Uo which satisfies the equa-
tion:
8.22. (Geometric Hamiltonian.)

If L(t, Q, Z) is strictly convex, argue that for each fixed t, Q, P: I(Z) ~ P' Z -
L(t, Q, Z) will be maximized by that Zo for which P = L<i(t, Q, Zo). Hence
H(t, Q, P) ~ maxz [p. Z - L(t, Q, Z)] defines this Hamiltonian without
"solving" equations (19) for Q.
8.23. Find the Hamiltonian H associated with the "Lagrangian" f and decide
whether your H is a saddle function; when
(a) f(x, y, z) = z2/4y.
(b) f(x, y, z) = J~y'-=2-+-Z-=-2 (y > 0).
Problems 279
(c) f(x, y, z) = y2 + Z2.

(d) f(x, y, z) = y~.
(e) f(x, y, z) = -JY~(y > 0, Z2 < 1).
(f) f(x, y, z) = J1 + y2 + Z2.
8.24. (a) In (47) if H(2£, Y, P) is a [strict] saddle function show that f(2£, Y, Z) is
[strictly] convex. Hint: If the left side of (47) is zero, then both bracketed
expressions on the right must vanish.
(b) Establish the converse result to that in part (a).
(c)* Could these assertions hold with [strictly] replaced by [strongly]?
(d) Verify (47).
8.25. Show that H(2£, y, p) is a strict saddle function and find a related convex
"Lagrangian" f when:
(a) H(x, y, p) = J2YJi+P2, y> 0.
(b) H(x,y,p)=Jy+p2, y+p2>0.
(c) H(x, y, p) = - J1+7 J1=P2, p2 < 1.
8.26. (The isoperimetric inequality, §1.3.)

Let ({j be a closed simple (nonintersecting) curve of length 21 in 1R2, para-
metrized with respect to arc-length s by functions x, y E C 1 [0,21]. Assume that
(x, y)(O) = (!J E ~ and that P = (x, y)(l) lies on the positive x-axis as shown in
Figure 8.9. By Green's theorem, we can suppose that ({j encloses a region of
area A(x, y) = II21 y(s)x'(s) ds I.
(a) Show that as in the derivation of (49)
Al(X,Y)~lf>x'dsl:::;; IIY11x'ldS= I J y 2_(yy')2 d S
= -F(Y):::;; 12/2n,
with equality iff on [0, I], (x, y) parametrizes a semicircle. Hint: ji =
y2/2 E C1(0, I). See Remarks 8.3.
(b) Define A2 similarly and conclude that A :::;; Al + A2 :::;; 12/n with equality
iff (x, y) parametrizes a circle.
(c) Can you extend the last result to piecewise C 1 curves fll?
Figure 8.9
8.27. A Zenodoros' problem (see Problem 1.9(b))

(a) Verify that f(y, z) = - yJl - y2z2 is not convex.
(b) Make the substitution y = y2j2, so Y' = yy' and show that this Zenodoros
problem reduces to minimizing the integral
F(Y) = - IT JY(l - Y'2) dt

on
!!#* = {y E qo, T]: y(O) = Y(T) = 0, IF(y)1 < + oo}
(c) Show thatj(y, z) = -JY(l - Z2) is strictly convex on {y > O} X {Z2 < l}
by considering its Hamiltonian.
(d) Establish that yo(t) = t - (t 2jT) is the unique minimizing function. Hint:
Remarks 8.3.
(e) Return to the original variables in Problem 1.9 and conclude that the
curve which can produce the maximum value is given by a semicircle.
Hint: X~2 + y~2 = Ijy~(t) (why?), so that with part (d), x~(t) = is .fiiT
constant. Eliminate t.
8.28. (a) Use the substitution y2 = 2y to transform the (nonconvex) problem of
minimizing
F(Y) = f,0
b
(Y'2 - y2)(X) dx
-
on !!# =
{YEC1[0,b]:Y(0)=0
Y(b) = sin b
}
to the problem of minimizing
F(y) = Jor (i2

b
2y - )
2y (x) dx
on
!!# = {y E C1(0, b) n qo, b]: y(x) ~ 0, F(y) < + 00,
sin 2
y(O) = y'(O) = 0; y(b) = -2- .
b}
(b) Show thatf(y, z) = z2j2y - 2y is (only) convex on {y > O}. Hint: Look at
its Hamiltonian. (for which H{2£, y, p) is strictly convex).
(c) When b < n conclude that y(x) =-(sin 2 x)j2 minimizes F on !!# uniquely.
(d) When b = n, conclude that y(x) = (sin2 x)j2 minimizes F on !!# (but not
uniquely), and obtain the Wirtinger inequality
f: y2 dx ~ f: (Y')2 dx when y E C 1[0, n] with y(O) = y(n) = O.
Hint: We can assume that y(x) > 0, when 0 < x < n. (Why?)
(e)* Show that y(x) = (sin2 x)j2 does not minimize F on !!# when b > n, even
though it appears to satisfy the appropriate conditions.
8.29. (a) For the seismic wave problem 1.8(b), make a transformation (such as that
in Problem 8.28) for which y'(x)jy(x) = y'(x), and show that the new
J
integrand function j(y, z) = e 2y + Z2 is strictly convex on [R2.
Problems 281
(b) Conclude that a circular arc provides the (only) path of travel for such a
seismic wave.
8.30. (a) If g = g(y) is [strictly] convex and positive on an interval I, then show that
f(y, z) = J g2(y) + Z2 is [strictly] convex on I x IR.
(b) Under what conditions on g = g(x, y) will f(2£, y, z) = - J g2(2f, y) - Z2 be
[strictly] convex on a domain of definition? Give some nontrivial examples
of such g.
CHAPTER 9*
Sufficient Conditions for a Minimum
As we have noted repeatedly, the equations of Euler-Lagrange are necessary
r r
but not sufficient to characterize a minimum value for the integral function
F(Y) = f(x, Y(x), Y'(x)) dx = f[Y(x)] dx
on a set such as
~ = {Y E C 1 ([a, b])d: Y(a) = A, Y(b) = B},
since they are only conditions for the stationarity of F. However, in the
presence of [strong] convexity of f(}£, Y, Z) these conditions do characterize
[unique] minimization. [Cf. §3.2, Problem 3.33 et seq.] Not all such functions
are convex, but we have also seen in §7.6 that a minimizing function Yo
must necessarily satisfy the Weierstrass condition cS'(x, Yo(x), Y~(x), W) ~ 0,
V WE !Rd, X E [a, b], where
cS'(x, Y, Z, W) ~ f(x, Y, W) - f(x, Y, Z) - fz(x, Y, Z)' (W - Z), (1)
and this is recognized as a convexity statement for f(}£, y, Z) along a trajec-

tory in !R 2d +1 defined by Yo'
This chapter is devoted to showing that conversely, when f(}£, y, Z) is
[strictly] convex (§9.2) in the presence of an appropriate field, then each
stationary Yo in ~ does minimize F on ~ [uniquely] (§9.3, §9.4) and this will
afford a solution for the brachistochrone problem. The method extends in
principle to problems with variable end point conditions (§9.4) and to those
on which constraints are imposed (§9.5).
However, the field in question requires an entire family of stationary
functions with special properties which may, or may not, exist. A central
field will suffice (§9.6) and this provides further insight into the problem of
282
§9.1. The Weierstrass Method 283
finding the minimal surface of revolution. In §9.7 we provide conditions

which assure that a given stationary trajectory ro may be considered as that
of a central field. We encounter a new criterion, that of Jacobi, but in §9.9
we demonstrate that it, too, is almost essential for local minimization. This
embedding of ro together with the appropriate convexity of f supplies suffi-
cient conditions for the local minimization of integral functions F, both in the
weak and in the strong senses (§9.8).
To motivate the inquiry, let's look at the original method used to attack
such problems.
§9.1. The Weierstrass Method

In his lectures of 1879, Weierstrass presented the following approach to
r r
prove that a given stationary function Yo E (C 1 [a, b])d minimizes
F(Y) = f[Y(t)] dt = f(t, Y(t), Y'(t)) dt

on
~ = {Y E (C1[a, b])d: Y(a) = Yo(a); Y(b) = Yo (b)},
(where for simplicity we suppose thatf E C1([a, b] x [R2d).
Let Y in ~ be a competing function and assume that each x E (a, b]
determines a unique function '1'(-; x), stationary for f on (a, x) as in §6.7,
whose graph joins (a, Yo(a)) to (x, Y(x)) as shown in Figure 9.1.
Then, in particular, 'I'(t; b) = Yo(t), (Why?), and we consider the integral
function
(J(x) ~ - LX f['I'(t; x)] dt - f f[Y(t)] dt (a ~ x ~ b), (2)
-Y
- - 'I'
(b, Yo(b)
- - Yo
Figure 9.1
284 9*. Sufficient Conditions for a Minimum
which interpolates between
O'(a) = - f f[Y(t)] dt = -F(Y)

and
O'(b) = - f f['¥(t; b)] dt = -f f[Yo(t)] dt = -F(Yo),
so that F(Y) - F(Yo) = O'(b) - O'(a). Were O"(x) ~ 0, it would follow by the
mean value theorem that F(Y) - F(Yo) ~ O. Moreover, if also a' is continu-
ous on [a, b], then equality holds iff a' == O. (§A.1, §A.2.)
For example, when d = 1, consider the problem of minimizing the (non-

convex) function
F(y) = f: [y'(t)2 - y(t)2] dt
on
~ = {y E C1[0, b]: y(O) = y(b) = O} for b < n.
Here, the stationary functions t/J satisfy the Euler-Lagrange equation
d
dt 2t/J'(t) = -2t/J(t) or t/J"(t) + t/J(t) = 0,
with the well-known general solution

t/J(t) = Co cos t + c 1 sin t.
Since b < n, the only solution in ~ is Yo = (9, and for a given y in~, it is seen
by inspection that for each x E (0, b):
,I, ( ) def ( )
If' t;x =yx
sin
- . -t
smx
is the unique function which is stationary for f and satisfies t/J(O; x) = 0 with
t/J(x; x) = y(x). (See Figure 9.2.)
Thus, for this example, equation (2) becomes
O'(x) = - f: [t/J'(t; X)2 - t/J(t; X)2] dt - f [y'(t)2 - y(t)2] dt
= - ~2~X) IX (cos 2 t - sin 2 t) dt - fb [Y'(t)2 - y(t)2] dt,

mn x 0 x
so that after integration,
O'(x) = - y2(X) cotan x - f [Y'(t)2 - y(t)2] dt. (3)

§9.1. The Weierstrass Method 285
Graph of
y
------ -y
- - !/lb,)
--Yo
- Figure 9.2
Then by the fundamental theorem of calculus, for x E (0, b):

o"(x) = -2y(x)y'(x) cotan x + y2(X) cosec 2 x - y2(X) + y'(X)2
= (y(x) cotan x - y'(X))2 ~ 0,
with equality iff y'(x) - y(x) cotan x = O.
Now, an apparent difficulty occurs when x = 0 as a result of the multi-
plicity of stationary functions t/J which pass through (0, 0); namely, that 0'(0)
is not defined by (3). However, by L'Hopital's rule,
lim y2(X) = lim 2y(x)y'(x) = 0

x,"osinx X ,"0 cosx '
for y in ~, so that 0'(0 + ) = - F(y). Thus F(y) - F(yo) = O'(b) - 0'(0 +) ~ 0;
or F(y) ~ F(yo) = 0, with equality iff y'(x) = y(x) cotan x, V x E (0, b). But
for y in ~, this implies that y'(b) = y(b) = 0 (since b < n), and hence that
y(x) == 0 = Yo(x).
Thus we have proven that for b < n, Yo = (!J is the unique minimizing
function for F on ~. When b = n, the method is still applicable, (and provides
a prooffor Dido's conjecture of Problem 1.5), but the minimizing function Yo
is no longer unique. However, when b > n, Yo fails to minimize. (See Problem
9.1.)
Having illustrated the effectiveness of the Weierstrass approach in a simple

case, we return to the general problem and equation (2). As defined, 'P(t; x)
depends on two variables, and as above, we use the prime to denote t differ-
entiation. Supposing that 'P is C 2 , it follows that

('P')x = 'Ptx = 'Pxt = ('Px)'· (4)
Then, from Leibniz's rule {A. 14) applied to (2), we obtain
a
cr'{x) = f[Y{x)] - f['P{x; x)] - I axf['P{t;
a
x
x)] dt, (5)
and the integrand is from the chain rule, (4), and stationarity, given by
a
axf{t, 'P{t; x), 'P'{t; x)) = fy['P{t; x)] 'PAt; x) + fz['P{t; x)] ('P')At; x)
a
= at {fz['P(t; x)]· 'Px{t; x)}.
But 'P{x; x) == Y{x) by construction, so that 'PAx; x) = Y'{x) - 'P'{x; x)

(Why?); while 'Px{a; x) = 0, since 'P{a; x) = Yo{a), is constant. Hence, after
integration and substitution, (5) becomes
cr'{x) = f{x, Y{x), Y'{x)) - f{x, Y{x), 'P'{x; x))
- fz{x, Y{x), 'P'{x; x))· (Y'{x) - 'P'{x; x)),
or upon utilizing (I),
cr'{x) = c!{x, Y{x), 'P'{x; x), Y'{x)). (6)
(For d = 1, the reader should verify each step of this derivation by purely
r r
formal calculations.) Thus finally, we obtain Weierstrass' formula:
F{Y) - F{Yo) = cr'{x) dx = c!{x, Y{x), 'P'{x; x), Y'{x)) dx, (7)
°
which proves that c! ~ will imply that F{Y) ~ F{Yo), provided that an
appropriate family of stationary functions 'P{.; .) having all of the assumed
properties is available. Unfortunately, it is quite difficult to prescribe con-
ditions which ensure the existence of such families (one for each competing
Y E ~), and instead in §9.3 et seq. we shall concentrate on a less direct
approach of Hilbert, which yields Weierstrass' result even for piecewise C 1
functions Y.
(Problems 9.1-9.2, 9.5-9.6)
§9.2. [Strict] Convexity of f0i, y, Z)

The definition (I) of the Weierstrass excess function for a given f = f{x, Y, Z);
viz.,
c!{x, Y, Z, W) = f{x, Y, W) - f{x, Y, Z) - fz{x, Y, Z)· (W - Z),
and our wish to consider c! ~ 0, suggests in comparison with 3.4, the following:
§9.2. [Strict] Convexity of f0s., X. Z) 287
(9.1) Definition. f~, X; Z} is said to be [strictly] convex on a set S ~ 1R 2d +1,

when f and fz are defined and continuous in S and satisfy the inequality
C(x, Y, Z, W} ~ 0, or equivalently,
f(x, Y, W} - f(x, Y, z) ~ fz(x, Y, z) . (W - Z), (8)

when (x, Y, Z) E S and (x, Y, W) E S, [with equality at (x, Y) iff W = Z].
Usually, the set S will be of the form S = D X IRd for a domain D of IRd+1.
Since the [strict] convexity of f~, X; Z} is identical with the [strong]
convexity of f~, X; Z} as defined in Chapter 3 (Problem 3.33), the usual
linear combinations of such functions remain [strictly] convex. See Proposi-
tion 3.2 and Facts 3.11.
Example 1. f~, Y, Z} = -I YI 2 + IZI 2 is not convex, but f~, X; Z) is strictly

convex in 1R 2d +1 since fz(x, Y, Z) = 2Z so that
f(x, Y, W} - f(x, Y, Z} = IWI 2 -IZI 2 = (W + Z}·(W - Z)
= W - ZI 2 + 2Z· (W - Z)
1
~ fz(x, Y, Z) . (W - Z),
with equality at (x, Y) iff 1W - ZI 2 = 0 or W = Z.
Example 2. For d = 1, the brachistochrone function of §1.2(a}
f~, y, z} = )1 : Z2
is not convex, but f~, l' z} is strictly convex for the half space
{(x, y, z) E 1R3: y > O}
since fzAx, y, z} > o. (See Proposition 3.10 and the next example.)
Example 3. For d = 2, when Y = (x, y), the function f{J, X; Z) = J.YIZI 2 is
[strictly] convex on the half-space -
{(t, x, y, Z) E IRs: y ~ OJ, [{(t, x, y, Z) E IRs: y > On
For, by the computation of Example 1,
f(t, Y, W} - f(t, Y, Z} = JY(I W - ZI 2 + 2Z· (W - Z»
~ fz(t, Y, Z} . (W - Z),
[with equality for y > 0 iff W = Z].
Example 4. For d = 1, the function f~, y, z} = -J1 - Z2}/y is strictly con-

vex on the set S = {(x, y, z) E 1R3: y > 0, iZl < 1}. (See Exampje 5 of§3.3, and
Example 2 above.)
From §0.13, there is the following generalization of Proposition 3.10:
(9.2) Proposition. If f = f(x, Y, Z) together with its partials fz, and fziz j,
i, j = 1,2, ... , d, is continuous in a Z-convex set S s;;; 1R 2d + 1 (one which contains
the segment joining each pair of its points (x, Y, Zo), (x, Y, Zl)) and the matrix
fzz is positive semidefinite [positive definiteJ in S, then fVi, Yo Z) is [strictlyJ
convex in S.
PROOF. For (x, Y, Zo) and (x, Y, Zl) in Sand t E [0, 1J, the point
deC
Zt = (1 - t)ZO + tZ1
lies on a segment contained in S by hypothesis. Integrating by parts, we get
f(x, Y, Zl) - f(x, Y, Zo) = t :/(x, Y, Zt) dt
= (Zl - Zo)· t fz(x, Y, Zt)d(t - 1)
= (Zl - Zo)·(t - l)fz(x, Y, Zt)[:
+ fo
1 (1 - t) (. t
I.J=l
fzizj(X, Y, Zt)V;Vi ) dt,
where V = Zl - ZOo The last term is nonnegative when Zl =F Zo as a conse-

quence of the assumed semidefiniteness of fzz [and with positive definiteness
of fzz, it vanishes iff V = (9]. Hence
f(x, Y, Zd - f(x, Y, Zo) ~ fz(x, Y, Zo)· (Zl - Zo)
[with equality iff Zl = Zo]. D
A (spherical) neighborhood of a point (xo, Yo, Zo) is Z-convex. If D is a
domain of IRd+t, then S = D X IRd is Z-convex.
§9.3. Fields
The Weierstrass construction in §9.1, when possible, results in a family of
stationary trajectories (the graphs of the functions '1'(.; x)) which is consistent
in that one and only one member of the family passes through a given point
(x, Y(x)). Suppose more generally, that for a given f we have a single family
of stationary functions whose trajectories cover a domain D of IRd + 1 consis-
tently in that through each point (x, Y) ED passes one and only one trajec-
tory of the family, say that represented by '1'(.; (x, Y)) E (C1[a, bJ)d. Then the
direction of the tangent line to the trajectory at (x, Y) given by
(x, Y) ~ '¥'(x; (x, Y))
§9.3. Fields 289
determines a vector valued function in D (and one, moreover, whose values

are required for Weierstrass' formula (7) along each competing trajectory).
For a domain D of IRd +l , we shall call any C l function CI>: D -+ IRd a field in
D. When CI> = CI>(x, Y} the differential equation
Y'(x} = CI>(x, Y(x)) = CI>[Y(x}]
is called the field equation. l It is a consequence of the theory of differential
equations, that through each point (Xl' Yd ED, passes exactly one field
trajectory, the graph of the unique solution Yo of the field equation with
Yo(xd = Yl . (See A.19.)
For a given f, the field will be called stationary when each solution of the
field equation is a stationary function, so that each field trajectory is a
stationary trajectory. Conversely, when the field CI> is determined from the
tangents to a consistent family of stationary trajectories'll as above, then it is
stationary. (Why?)
While reformulating the Weierstrass theory in 1900, Hilbert observed that

the principal requirement of a usable field was that it be exact in the sense of
9.3 which follows. (His motivation for this definition is indicated in Problem
9.13, and the corresponding multidimensional generalization-also given by
Hilbert-is taken up in Problems 9.14 and 9.15.)
(9.3) Definition. A C l function CI> from a domain D of IR d+1 to IRd is said to

be an exact field for f = f(x, Y, Z) in D, if there it makes the differential
h dx + p. dY exact when
P(x, Y} == fz(x, Y, CI>(x, Y»,
and (9)
h(x, Y} == f(x, Y, CI>(x, Y)) - P(x, Y}· CI>(x, Y};
i.e., there exists a real valued function S E Cl(D} with partials Sx = hand
Sy = P, so that dS = h dx + p. dY. (S is determined within an additive
constant.)
The significance of an exact field will be demonstrated in the next section.

First, however, observe that when P and hare C l , and h dx + p. dY = dS
is exact, then equality of the mixed partials of S implies that for
P = (Pl' P2' ... , Pd)
i, j = 1, 2, ... , d. (10)
(When d = 1, the second condition is always fulfilled.)
Conversely, it is known that when the exactness conditions (10) are
satisfied in a simply connected domain D of IRd +l ; (i.e., a domain without
1 The literature in this subject does not provide a uniform definition for fields. That given here
seems most convenient for our purposes.
y y
(f) x
(a) (b)
Figure 9.3
"holes") then there exists a function S as required to make the differential

h dx + p. dY exact. [Ed]. (See also Problem 9.13.)
Example 1. The time of travel function for a seismic wave (Problem 1.8); viz.,
f(y, z) = ~/y, is C 2 for y > 0, and has as its stationary trajectories the
semicircles with centers on the x axis (Problem 9.3). Geometrically, as shown
in Figure 9.3(a), it is evident that those with fixed center (say the origin)
indexed by the radius 2, form a consistent stationary family, in the upper
half plane D = {(x, y) E ~2, Y > O}. Analytically, these are defined by
t/I(t,2) = J2 2 - t 2, It I < 2;
and that passing through the point (x, y) is obtained for 2 = Jx 2 + y2. Thus
the associated (stationary) field in D is
From (9) we form

-x/y -x
p(x, y) = fAx, y, <p(x, y)) = J 2 2
Y 1 + x /y
and
h(x, y) = f(x, y, <p(x, y)) - fz(x, y, <p(x, y))<p(x, y)
[~ Z2] 1 1
= Y - y~ z=tp(x,y) = yJl + X 2/y2 = Jx 2 + y2'
Here, it may be verified directly that p = Sy and h = Sx when S(x, y) =
log(Jx 2 + y2 + x) -log y; (or use (10)). Thus <p(x, y) = -x/y defines an
exact field for f in D. (Another consistent stationary family is given by those
semicircles with fixed (left) end point, as in Figure 9.3(b). The resulting field is
also exact, but it is defined only in a quarter plane (Problem 9.3).)
§9.3. Fields 291
In this example, stationary fields provided the exact fields. We examine the
intimate relations between these fields in the next two results.
(9.4) Proposition. When fz is C 1 , then each exact field for f is a stationary
field; i.e., each C 1 solution Yo of the field equation Y'{x) = {x, Y{x)) is a
stationary function for f
PROOF*. We have by (9) and the hypothesis that when Y~{x) = [Yo{x)], then
fz[Yo{x)] = fz{x, Yo{x), [Yo{x)]) = P{x, Yo{x)) (11)
is C 1 • Also both P and hare C 1 and again from (9)
hy = fy + <l>rfz - <l>yP - Py<l>
= fy - Py<l>,
when the partials of f are evaluated at (x, Y, {x, Y)). (In the matrices Py and
<l>y, the rows are indexed by Y; moreover, by the second set of equations (10),
the Jacobian matrix Py is symmetric and so is equal to its transpose Py.) Thus
fy[Yo{x)] = hy{x, Yo{x)) + Py{x, Yo{x)){x, Yo{x)), (12)
so that finally by substitution of (11) and (12), and the chain rule:
d d
dxfz[Yo{x)] - fy[Yo{x)] = dx P{x, Yo{x)) - fy[Yo{x)]
= PAx, Yo{x)) + Py{x, Yo (x)) Y~{x)

- hy{x, Yo{x)) - Py{x, Yo{x)){x, Yo (x)).
The terms involving Py cancel (Why?), and we obtain the equation
d
dxfz[Yo{x)] - fy[Yo{x)] = (P" - hy)(x, Yo{x)) (13)
= 0, by (10).
Thus Yo satisfies the Euler-Lagrange equation of §6.7 and so is stationary.
o
Hereafter we shall suppose fz is C 1 in all cases of interest.
The converse of Proposition 9.4 is not true in general. (See [el) However,
if ({J is a stationary field in a domain D of 1R2, then, by definition, each field
trajectory is a stationary trajectory; since the "matrix" Py is trivially symmet-
ric we may again conclude that equation (13) holds and the first exactness
condition p" - hy = 0 of (10), is met at each point (x, Yo) E D. When d = 1
and D is simply connected, it is the only requirement and we have established
the following:
(9.5) Corollary. If fz is C 1 and ({J is a stationary field for f = f(x, y, z) in a

simply connected domain D of 1R2, then ({J is an exact field for f 0
(9.6) Remark. In higher dimensions, more than simple stationarity is re-

quired to produce an exact field, and this leads to the study of Mayer fields
[S]. However, in §9.6 we shall prove that an appropriately constructed central
field determined by a consistent family of stationary trajectories emanating
from a common point is always exact.
Example 2. From the discussion for the brachistochrone in §6.2(c) we know

that the function
{1+?
f(x,y,z)=V~-y-
provides precisely one stationary curve joining the origin (0, 0) to a given
"lower" point (x, y); namely, the cycloid which is represented parametrically
by the equations .
t = A(r - sin r)
(0 ~ r ~ () < 2n), (14)
t/I = A(1 - cos r),
where A > 0 and () are determined uniquely by the boundary conditions
x = A(() - sin (),
(15)
y = A(1 - cos ().
(The previous notation has been replaced by one more amenable to our
present requirements.) The resulting family of cycloids, denoted t/I(t, Je), is
shown in Figure 9.4. The associated field is defined by the direction of t/I(t, Je)
at t = x; i.e.,
( ) '/"( ') I
({JXY='I'tA
, 't=x
=t/lt
-
tt
I
t=6
A sin r
=--:-:--c-------:-
A(1 - cos r)
I
t=6'
or ({J(x, y) = cotan ()/2, where, of course, () is determined implicitly by equa-

tions (15) as a C 1 function of (x, y). As defined, ({J is a stationary field for f in
Figure 9.4
§9.3. Fields 293
the quarter plane D = {(x, y) E ~2: x> 0, y > O} which is simply connected.
Thus ({J is exact, by Corollary 9.5. (Another exact field is discussed in Problem
9.9.)
Exact Fields and the Hamilton-Jacobi Equation*
The notation of P and h used to represent the exact differential dS = h dx +

p. dY in Definition 9.3 suggests that used for the conjugate momenta and the
Hamiltonian in Chapter 8. This is not accidental, for if we regard f(x, Y, Z)
as the Lagrangian function for a dynamical system at time "x", and, suppose
as in §8.8, that the equations P = fz(x, Y, Z) determine Z = G(x, Y, P) for an
appropriate C 1 function G, then the associated Hamiltonian is defined by
- H(x, Y, P) = f(x, Y, G(x, Y, P)) - p. G(x, Y, P).
In particular, if for some field CI>, we set P(x, Y) = fz(x, Y, CI>(x, Y)) (as in
(9)), then Z= CI>(x, Y) == G(x, Y, P(x, Y)), so that
- H(x, Y, P(x, Y)) = f(x, Y, CI>(x, Y)) - P(x, Y)' CI>(x, Y)
(16)
= h(x, Y), as defined by (9).
Moreover, when CI> makes the differential h dx + p. dY = dS, say, exact in a
domain D of ~d+l, then, of course,
P = Sy (i.e., Pi = SYi' j = 1, 2, ... , d),
and h = Sx, so that upon substituting in (16) we get that

Sx + H(x, Y, Sy) == 0 in D. (17)
This is recognized as the partial differential equation of Hamilton-Jacobi
(cf. §8.7), and we see that each exact field CI> for f determines a solution S of
this equation.
Conversely, if S is any C 1 solution of (17) in a domain D of ~d+l, then we

may define P(x, Y) = Sy, and
h(x, Y) = - H(x, Y, P(x, Y)) = Sx,
and conclude that CI>(x, Y) ~ G(x, Y, Sy(x, Y)) determines an exact field for f
In making these assertions we are assuming that the equations P =
fz(x, Y, Z) determine Z = G(x, Y, P). From implicit function theory we
know this to be true in a neighborhood of each point (x, Y, Z) E ~2d+l in
which f is C 2 and the Jacobian matrix fzz is invertible. Thus locally at least,
we may expect a direct relation between the solvability of the Hamilton-
Jacobi equation (17) and the existence of an exact field. (See also Problem
8.22.)
§9.4. Hilbert's Invariant Integral

If (x, Y) is an exact field for a given f in a domain D of IRd +1 as defined in
9.3, then for piecewise C 1 curves r in D (those which admit parametrization
by piecewise C 1 functions), the value of the line integral
I(r)~ fr (h dx + p. dY) = If dS, (18)
say, depends only on the end points of r; i.e., its value is invariant with the
particular curve in D which joins given end points. (This familiar conse-
quence of exactness is independent of particular choices for h, P [Ed].) I is
called Hilbert's invariant integral for in D.
However, with our choices (9), we see that when r is represented as the
graph of a function Y E (C1[a, b])d, then formally, on r, dY = Y'(x) dx and
I(r) = r
(18) becomes
[f(x, Y(x), (x, Y(x))) + P(x, Y(x))· (Y'(x) - (x, Y(x)))] dx

(19)
f f
a
(= f[Y(x)] dx - C(x, Y(x), (x, Y(x)), Y'(x)) dx, by (1)).
Moreover, when another such curve r o , say, is a trajectory of the field <1>,
represented by Yo E (C 1 [a, b])d, then by definition Y~(x) = (x, Yo(x)) so that
(19) reduces to
I(ro) = f f(x, Yo(x), Y~(x)) dx = f f[Yo(x)] dx = F(Yo), (20)
where F(Y) ~ J~f[Y(x)] dx.
It follows that when both rand ro have the same end points and lie in D
as shown in Figure 9.5, then by (20):
F(Y) - F(Yo) = f f[Y(x)] dx - I(ro)
= f f[Y(x)] dx - I(r), by the invariance of I;
hence, from (19), we obtain Hilbert's formula
F(Y) - F(Yo) = f C(x, Y(x), [Y(x)], Y'(x)) dx. (21)
This formula of Hilbert should be compared with that of Weierstrass J7).

Observe that it applies to piecewise C 1 functions as promised. However, as
before, we see that C ~ 0 => F(Y) ~ F(Yo).
§9.4. Hilbert's Invariant Integral 295
a b x
Figure 9.5
As we have stated, in §9.6 we will prove that every central field is exact. l
To avoid duplication we anticipate this result in formulating the next
(9.7) Theorem (Hilbert). Let f(~, J', Z) be [strictly] convex on D x [Rd where
D of [Rd+1 is the domain of <1>, an exact field (a central field) for J, which
contains the stationary field trajectory ro represented by Yo E (Cl[a, b])d.
Then Yo minimizes
F(Y) = f f[Y(x)] dx
[uniquely] on
~ = {Y Eel [a, b]d: Y(a) = Yo(a); Y(b) = Yo(b); (x, Y(x)) ED}.
PROOF. Since from convexity, $ ~ 0, it follows from (21) that F(Y) ~ F(Yo)
with equality iff
$(x, Y(x), [Y(x)], Y'(x)) = 0 (where defined).
[With strict convexity this implies that where defined, Y'(x) = (x, Y(x)) so
that Y = Y E (Cl[a, b])d represents a field trajectory. However, the field tra-
jectory through the point (a, Yo(a)) is unique, and thus Y == Yo.] D
(9.8) Remarks. Since Yo minimizes F on ~, it is necessarily stationary (§6.7);

however, not every stationary function will represent a trajectory of the given
field. Indeed, it is only by construction that a given stationary trajectory ro
can be established as a field trajectory for an exact field, and it is this fact
which makes the Weierstrass-Hilbert theory so difficult to apply. Observe
also that the size of the minimization class ~ is determined by the extent of
the field domain D, but within ~, Hilbert's theorem furnishes a strong mini-
mum in the sense of Chapter 7. In particular, it is geometrical in character.
On the other hand, in order to admit arbitrary Y' E [Rd, f(x, Y, Z) must be
defined on the set S = D X [Rd, and this would exclude some f with restricted
1 See Remark 9.6.

convexity such as -~ (for d = 1). A corresponding theorem obtains

for such f, but it is more awkward to state (Problem 9.7).
A further characteristic difficulty is encountered in the next example.
Application: The Brachistochrone*

For the brachistochrone function
f(x, y, z) = P/-
+Z2
-y-'
.
wIth fz(x, y, z) = r.:
z
~'
yyy 1 + z
we know from Example 2 of §9.3 that we have an exact field cp in the
quadrant D = {(x, y) E IFe, x> 0, y > O} for which the field trajectories are
the cycloids with cusp at the origin. Moreover, from §3.3, Example 3, it
follows that f(~, y, z) is strictly convex in D x IR. Hence, from Hilbert's theo-
rem, we can conclude that the cyloid, r o, joining the origin to a given point
(b, bd E D is represented by Yo E C 1 (0, b] which minimizes uniquely for each
a > 0, the time-of-descent integral given within a constant factor by
on
T,.(~) = r 1 + ~'(X)2
~(x) dx
(22)
~a = U E C1 [a, b]: ~(a) = Yo (a), ~(b) = b1 ; ~(x) > a}.

°
See Figure 9.6(a). When a = in (22), denote T,. by T.
Now, this is not quite what we wish; namely, to establish the (unique)
minimality when a = 0. The resulting problems are two-fold: first, the origin
is not in the domain of the field; and second, the cycloid (and perhaps other
curves of interest) have an infinite slope at the origin.
We can circumvent these difficulties by the following arguments:
With ro as the above cycloid, let r
be a piecewise C 1 curve joining the
origin to (b, b 1 ) contained except for its end points in D, and represented by
a b a b
x
y y
r.
(a) (b)
Figure 9.6
PE C 1 (0, b] for which T(P) < +00. Then for each a E (0, b), let Ua be the
segment (in D) joining the (possibly distinct) points (a, yo(a)) and (a, p(a))
as shown in Figure 9.6(b).
Denote by ra and t the parts of the curves ro and t, respectively, cor-
responding to x ~ a, and observe that with proper orientation of Ua , t + U a
constitutes a piecewise C 1 curve with the same end points as ra. By the
invariance of 1, Hilbert's integral (18) for cp, we obtain
1(ra) = 1(t + ua ) = 1(t) + 1(ua )
Next, with T" as defined by (22) above, we may reproduce the analysis
leading to Hilbert's theorem, utilizing in particular (19) and (20), to conclude
that
T,,(P) - T,'(Yo) = r
C(x, P(x), cp[P(x)], P'(x)) dx - l(ua ) (23)
(Problem 9.10)
Since f ~ 0, and C ~ ° in D x 1R2, we have
J: 1 + P'(X)2 fb 1 + P'(X)2
T(P) = P(x) dx ~ a
P(x) dx = T,,(P) ~ T,,(Yo) - 1(ua )·
(23')
Now as a '" 0, T,,(Yo))" T(yo), while on ua : dx = 0, and p(a, y) =
fz(a, y, cp(a, y)), so that from (18),
11(u )1 = 1ft! cp(a, y) ~I < ft! ~

a ao Jl + cp2(a, y) JY - ao JY'
(supposing that ao == yo(a) :s; p(a) = d)
:s; 21jRa) - JYo(a)l-+ 0,
(and this holds also when p(a) :s; Yo (a)).
Hence, in the limit as a '" 0, (23') gives
which establishes the minimality of the cyloid. The uniqueness requires fur-
ther analysis; [see Problem 6.15].
(Problems 9.7-9.10, 9.13-9.15)
Variable End-Point Problems

When an exact field cI> for f = f(x, Y, Z) is available in a domain D of IRd +l,
then Hilbert's approach admits modification to provide theoretical access to
YI
Figure 9.7
the problem of minimizing an integral function
F(Y, t) = f f[¥(x)] dx
on
~t = {¥ E (C 1 [a, t])d: ¥(a) = A, (t, ¥(t)) E T; (x, ¥(x)) ED},
where the right end point is confined to the transversal T defined as the zero
level set of a C 1 function r (whose gradient, Vr, is nonvanishing on T)
provided that Tis C 1 arcwise connected in D.
For then, when ro is a trajectory of the field represented by Yo E ~" and r
is the curve represented by a competing function ¥E~" under this assump-
tion, we may join their right end points by a C 1 arc UT in D n T as illustrated
in Figure 9.7.
r
With proper orientation we may consider ro and + U T as piecewise C 1
curves in D having the same end points. Since is exact we have as before
from (20)
= I(r + UT) = I(r) + I(uT ),

(where I is Hilbert's integral for <1», so that for the appropriate t:
F(Y, t) - F(Yo, b) = f C(x, ¥(x), [¥(x)], ¥'(x)) dx - I(uT ), (24)
which may be considered as Hilbert's formula in the presence of the transver-

sal defined by r.
To recover the minimizing inequality with the convexity of f0i, Yo Z) in

D x [Rd, we evidently should want that I(O"T) = o. Upon parametrizing O"T by
(x(s), ¥(s)),O :::;; s :::;; 1, we have from (18)
I(O"T) = II [h(x, ¥)X'(s) + P(x, f)· ¥'(s)] ds, (24')
and since (x', ¥') is a tangent vector to a curve in T (and hence to T itself),
which will vary with O"T' we must demand the vanishing of the integrand in
(24'); i.e., that the field functions hand P defined by (9) provide a vector (h, P)
which is normal to T at each point. But, by the well-known argument (repro-
duced at the end of §5.6), the gradient Vr = (rx, ry) is normal to T at each
point. Hence we should require that (h, P) is proportional to Vr at each point
of T, or equivalently, we require that the field meet the transversal condition
(25)
When d = 1, recalling that y~(x) = cp(x, Yo(x)), we can verify that (25) is
precisely the transversal condition (equation (15) of §6.4) which is necessary
for a local extremal of this problem. (25) may be regarded as its vector valued
generalization.
(9.9) Theorem. Let f0i, Yo Z) be [strictly] convex on D x [Rd where D 5;: [Rd+1
is the domain of <1>, an exact field (a central field) for f, which meets the
transversal condition (25) on the set T where:
(i) T = {(x, Y) E D: r(x, Y) = O},for a given C l function r with VrlT =f. (!J; and
(ii) Tis C l arcwise connected.
If ro is a field trajectory represented by
Yo E ~t = {Y E (C l [a, t])d: Y(a) = A, (t, Y(t)) E T; (x, Y(x)) ED},
then Yo minimizes F(Y, t) = J~f[Y(x)] dx on ~t [uniquely within the specifi-
cation of its interval of definition].
PROOF. Since the transversal condition (25) forces the vanishing of I(O"T) for
all C l arcs O"T 5;: T, Hilbert's formula (24) is applicable. The convexity of
f0i, Yo Z) makes <ff ~ 0, and gives the minimizing inequality
F(Y, t) ~ F(Yo, b),
(where we suppose (b, Yo(b) E T), with equality only if
<ff(x, Y(x), [Y(x)], Y'(x)) == O.
[With strict convexity this requires that Y'(x) = (x, Y(x)). Hence f = r l is
a trajectory of the field, and since both r l and ro pass through (a, A), we
know that their representing functions Y1 and Yo will agree on each common
interval [a, x] of definition. As Figure 9.7 indicates, it is possible that the
transversal T cuts a given field trajectory more than once, and then we must
accept the consequence that F(Yo, b) = F(Yl' td for b -# t 1 , even in the pres-
ence of strict convexity (Problem 9.17).] 0
The preceding arguments admit various simplifications with the form of!,
which will be taken up in Problems 9.16 and 9.17. However, it must be
admitted that field construction is difficult at best without the additional
complications accompanying the imposition of boundary conditions such as
(25). Hence our applications will be confined to those covered by Hilbert's
theorem 9.7. For a more complete discussion, see [S].
§9.S. Minimization with Constraints

We may also extend the Weierstrass-Hilbert theory (in principle) to embrace
minimization problems involving constraints of either the isoperimetric or
the Lagrangian form. We shall obtain a result covering both cases utilizing
Lagrangian multiplier function(s) (which in this chapter will be denoted by J.L
to avoid confusion). Our result is an amalgam of Hilbert's theorem 9.7 and
the approach taken in §2.3.
We assume that functions f ~ f(x, Y, Z) and gj = gj(x, Y, Z), j = 1, 2, ... ,

N, are suitably defined, and set f = f + L7=1
J.Ljgj where J.Lj = J.Lj(x, Y) is Con-
tinuous for j = 1, 2, ... , N.
(9.10) Theorem. Suppose that j~, X, Z) is [strictly] convex on D x IRd where

D of IR d +1 is the domain of <D, an exact field (a central field) for j which
contains a trajectory of the field represented by Yo E (C 1 [a, b])d. Then Yo
minimizes
[uniquely] on
F(Y) = r f[Y(x)] dx
ii = {Y E (C 1 [a, b])d: Y(a) = Yo(a), Y(b) = Ydb), (x, Y(x)) ED},
r r
under the constraining relations:
J.Lj[Y(x)]gj[Y(x)] dx = J.Lj[Yo(x)]glYo(x)] dx, j = 1,2, ... , N.

(26)
PROOF. Set F(Y) = J~j[y(x)] dx. By Hilbert's theorem, Yo minimizes
F(Y) = F(Y) + j~ r J.Lj[Y(x)]glY(x)] dx
[uniquely] on ii, and the conclusion is obvious. o

§9.5. Minimization with Constraints 301
Remarks. In view of Proposition 3.2, the [strict] convexity for j will follow
from the convexity of f~, X. Z) and Jlj(!., .Y)gj~' X. Z), j = 1, 2, ... , N;
[with the strict convexity of one of these terms]. Yo also minimizes on the
larger class of functions Y in ~ which satisfy (26) when ":5:" replaces the
equality sign as in Proposition 2.5.
In application, there are three separate cases of interest:
(9.10a) If the Jlj are constant, j = 1,2, ... , N, then we obtain [unique] minimi-
zation for F on ~ under the isoperimetric constraints
~
Gj(Y) =
deC fb glY(X)]
a
~
dx = GiYo), j = 1,2, ... , N. (27)
Since there are as many multipliers Jlj' as constraining relations (27), there is
some hope that we may find the Jlj which retain the convexity while permit-
ting the Gj(Yo ) to be specified.
(9.10b) When the Jlj = Jlj(x), j = 1,2, ... , N, then we obtain [unique] minimi-
zation on ~ under the Lagrangian form of constraining equations
j = 1,2, ... , N. (28)
Now, however, there is little reason to believe that the Jlj can be found which
permit preassigned functions gj[Yo(x)] (say gj[Yo(x)] == 0), j = 1,2, ... , N.
Indeed, each Lagrangian constraint of the form glY(x)] = 0, in general
restricts the dimensionality of the Euclidean space available for the solution
trajectories, so that unless N < d (which precludes the case d = 1) we should
not suppose that even these constraining relations can be satisfied. However,
(9.10c) When the Jlj = JliX, Y) and glYo(x)] == 0, j = 1, 2, ... , N :5: d, then
we obtain [unique] minimization for Yo under the preassigned Lagrangian
constraints
j = 1,2, ... , N, v X E [a, b]. (29)
Observe that since Yo represents a trajectory of the field CI>, (29) yields
gix, Yo(x), CI>(x, Yo(x))) == 0 and this will hold a fortiori if we require of the
field that gj(x, Y, CI>(x, Y)) == 0, in D, for j = 1, 2, ... , N. As we shall see in
Example 3, this additional field requirement may actually assist in the deter-
mination of the multiplier functions Jlj.
A problem with multiple constraints, each of one of these types, is covered

by Theorem 9.10 provided that each constraint has associated with it a Jlj of
the corresponding form permitted above.
Ftp) = Ll [P'(X)2 + P(X).P'(X)4] dx
on
~ = {P E ClEO, 1]: .9(0) = 1, y(1) = 0, Y ~ O},
when subject to the isoperimetric constraining relation
G( yA) deC
= f
1
o
xyA,4()
X dX = 1
2,
we take f(x, y, z) = Z2 + yz4, g(x, y, z) = xz4, and j(x, y, z) = Z2 + yz4 ...:

JlXZ4 for an unknown constant Jl so that
-
fzAx, y, z) = 2 + 12yz 2 + 12JlxZ 2 > 0, when X, y, Jl ~ 0.
It follows from Proposition 9.2 that for Jl ~ 0, f(}f, y, z) is strictly convex in
the quarter space D x IR where -
D= {(x, y) E 1R2: x ~ 0, y ~ O}.
The stationary functions for j are those which satisfy the Euler-Lagrange
equation
and those which are C 2 satisfy

(2 + 12yy,2 + 12Jlxy,2)y" + 3y,4 + 4Jly'3 = 0. (30)
By inspection, this equation is satisfied by the family of functions y(x) =
const., which do not belong to ~; and also by the family y(x) = mx + Xo
(for which y'(x) = m) provided that 3m4 + 4Jlm 3 = 0, which is possible for the
nontrivial choice m = -1'Jl. The stationary family oflines with slope m = -1
is consistent in the plane, and so in D, which is simply connected; and
Jl cI,;f -1m = 1 > 0. Thus by Corollary 9.5, it provides an exact field for
j = f + 19 in D. Moreover, Yo(x) = 1 - x defines a function in ~, whose
graph has slope - 1.
We conclude from Theorem 9.10a that Yo minimizes F on ~ uniquely,
when further constrained to the G(Yo)-level set of G. But G(yo) = 1.
Example 2. We may also use the field of Example 1 to conclude that Yo(x) =
1 - x minimizes F [uniquely] on ~ under the Lagrangian constraint:·
goLy(x)] == y'(X)4 = y~(X)4 == 1,
or equivalently,
gl[.Y(X)] == y'(X)4 - 1 == 0,
if we take Jl(x) = Wx in 9.10b.
We have observed frequently that a solution to one optimization problem

may solve others. A general result of this type is the following:
(9.11) Proposition. Isoperimetric problems may be reformulated as Lagrangian

problems in higher-dimensional space.
PROOF.
on
To transform the problem of minimizing
F(Y) = r f[Y(x)] dx
~ = {Y E (C 1 [a, b])d: Y(a) = A, Y(b) = B; (x, Y(x)) ED},

under the isoperimetric constraints
~ derfb ~ = 1,2, ... ,N, (31)
G/Y) = a gj[Y(x)] dx = lj' j
we proceed as follows:
To each Y E ~ we associate N new constraint variables Kj defined by
Kj(X) = t: gj[Y(t)] dt, j = 1,2, ... , N;
observe that Kj E C1 [a, b], and Kj(X) = gj[Y(x)]. Set

K(x) = (Kl (x), K2 (X), ... , KN(X)) and L = (iI' 12 , ... , IN)·
Then y* ~ CY, K) E (C 1 [a, b])d+N and the original isoperimetric problem is
equivalent to the Lagrangian problem of minimizing
on
F*(Y*) = r f[Y(x)] dx = F(Y)
~* = {y* E (C 1 [a, b])d+N: Y(a) = A, Y(b) = B; (x, Y) E D;
K(a) = (i), K(b) = L},

under the Lagrangian constraints
~ der ~
gj [Y*(x)] = gj[Y(x)] - Kj(X) == 0 on (a, b), j = 1, 2, ... ,N. (32)
For if Yo* = (Yo, Ko) minimizes F* on ~* [uniquely] under these
Lagrangian constraints, then Yo E ~, while for each· Y E ~ we have Y* E ~*
and
(33)
= K/b) - Kj(a) = lj.
Thus Yo minimizes F on ~ [uniquely] under the isoperimetric constraints
(31).
Conversely, if Yo in ~ solves the latter problem, then its counterpart

Yo* E ~* satisfies the Lagrangian constraints, (32), while for each y* =
(Y, K) E ~* which satisfies (32), Y E ~ and it satisfies the isoperimetric con-
straints (31) as above. Thus Yo* solves the Lagrangian problem for F*
[uniquely]. 0
304 9*. Sufficient Conditions for a·Minimum
Theorem 9.10(c) permits the construction of a more usable field for the
isoperimetric problem as reformulated than does Theorem 9.10(a). However,
it is more difficult to construct higher-dimensional exact fields, and more-
over, as we have observed in §3.5, it is especially difficult to use the freedom
in selecting J1. effectively. Despite these complications, this approach will yield
a solution to the classical isoperimetric problem.
The Wirtinger Inequality
Example 3*. In Problem 1.6 it was shown that the classical isoperimetric
inequality would follow from the Wirtinger inequality
F(y) = 1 0
2"
[Y'(X)2 - y(xf] dx ~ 0
on
!!) = {y E c 1 [0, 2n]: y(O) = y(2n) = O},
subject to the isoperimetric constraint
G(y) == 1 0
2"
y(x) dx = o. (34)
Since y = (9 is admissible, this is a minimization problem.
To transform this to a Lagrangian problem as in 9.11, we introduce

for y E ~ = {y E C1 [0, 2n]: P(O) = y(2n) = O}, the constraint variable, (here)
denoted IX
P1(X)== 0 y(t)dt,
and consider instead the problem of minimizing
on
t 2
" [P'(X)2 - P(X)2] dx
~* = {Y = (y, P1) E (C 1 [0, 2n]f: Y(O) = Y(2n) = (9},

subject to the Lagrangian constraint
g[y(x)] == y(x) - P~(x) = O. (35)
(Asterisks have been suppressed where possible.)
To do this, we shall find a central field 1 for the modified function

j(x, Y, Z) = Z2 - y2 + 2J1.(x, Y)(y - zd, (36)
1 See Remark 9.6.

where J.l is as yet unspecified, and the factor of 2 is introduced for conve-
nience. The stationary functions for j satisfy the usual equations, (§6.7) which
in this case are
!2y' = -2y + 2J.l + 2J.l y (Y - y~),
and
In addition, we require that the constraining equation, Y - y~ == 0, be

satisfied along each stationary trajectory, so that the terms in parentheses
vanish. From the second of the resulting equations we conclude that J.l is
constant along each stationary trajectory. Then, the first becomes y" + y = J.l,
which has the well-known general solution
y(x) = y cos x + A sin x + J.l, (37)
for trajectory constants J.l, A, y.
We shall construct a stationary field whose trajectories emanate from the

origin. To have y(O) = 0 in (37), we must take y = - J.l, so that with A == (J.l, A):
t/1(x; A) ~ y(x) = J.l(1 - cos x) + A sin x,
t/11(X; A)~ Yl(X) = f: y(t) dt = J.l(x - sin x) + A(1 - cos x).

(38)
Equations (38) determine the stationary family, 'P(x; A), indexed by the
parameter A. By construction, '1'(0; A) = (!). Moreover, for each x E (0, 2n):
the Jacobian matrix,
A)~[t/1!l c~s x
'PA(x,
so that its determinant,

t/1 1" t/1 1.
t/1;.](X; A) = [1 -
X - SIn X
.
sin x ],
1 - cos x
L\(x)~ det 'PA(x; A) = (1 - cos X)2 - sin x(x - sin x)

(39)
2(1 - cos x) - x sin x > 0, (n ~ x < 2n)
{
= 2 sin x(tan x/2 - x/2) > 0, (0 < x < n).
Hence, L\(x) > 0 on (0, 2n), and it follows that through each point (x, Y) E D =
(0, 2n) x ~2 passes a unique trajectory of the family-namely, that with
parameter A given as the unique solution of the linear system (38). An easy
inversion gives A = A(x, Y), where
J.l(x, Y) = [(1 - cos x)y - (sin X)YIJ/L\(X),
(40)
A(X, Y) = [ -(x - sin x)y + (1 - cos X)YIJ/L\(X).
The stationary field <D = (({J, ({Jl) so determined is given from (38) and
(40) by
<p(x, Y) = t/J'(x; A(x, Y)) = J1(x, Y) sin x + Jc(x, Y) cos x
(41a)
= [(sin x - x cos x)y + (cos x - 1)Y1J/A(x),
while in view of (35),
<P1(X, Y) = t/J~(x; A(x, Y)) = y. (41b)
In the next section we shall verify that this is a central field for j in D, and
from (36) we observe that j0s:., X. Z) is convex (but not strictly convex) in D,
since this is true of the term Z2 while the remaining term is linear in z 1. Hence
Hilbert's theorem (9.7) is applicable to j in D x [R2.
It follows that if rab is a trajectory of the field represented by a

Yo E (C 1 [a, bJ)2, with 0 < a < b < 2n, and fab S D has the same end points
as rab and is represented by Y E (C 1 [a, bJ)2, then
f j[y(x)] dx ~ f j[Yo(x)] dx = f f[yo(x)] dx (Why?).
As usual, this is not the desired inequality which must be obtained by a

limiting argument made rather delicate by the fact that A(O) = A(2n) = O.
We first note that Yo = (9 E ~*, and it has a graph ro on the x axis which
(except for its end points) is a trajectory of the field in D, since Yo = ,¥(x; (9),
as given by (38).
Next, ify E ~ satisfies (34), define Y1(X) = Joy(t) dt, so that Y = (y, yd E ~*,
and the constraining relation y~ = y hold. We also note for future use that
since y(O) = 0, then
and hence
ly(x)1 = If: p'(t)dtl ~ Mx,
IY1(X)1 ~M f: t dt ~ Mx 2.
Here, M = maxIP'I is fixed by y.
Let f be the graph of Y, and for 0 < a < b < 2n, let tb be that part of f
for x E [a, b]. Finally, let Ua and Ub be the segments joining the end points of
tb to (a, (9) and (b, (9), respectively, as shown in Figure 9.8.
With proper orientation of these segments we may use Hilbert's integral
[ for the field of j (as in the analysis of the brachistochrone in §9.4) to
conclude that
f j[y(x)] dx ~ f j[Yo(x)] dx + [(ua ) + [(ub ),
or
21t X
Figure 9.8
The desired inequality will follow if we prove that I(O',J --+ 0 when x = a '" 0,
and when x = b )" 2n. We shall concentrate on the first and more difficult of
these assertions. Since iz(x, Y, Z) = 2(z, - Jl), by (36), then from (9), the field
function P(x, Y) = iz(x, Y, CI>(x, Y)) = 2(<p, - Jl)(x, Y), where Jl and <p are
given by (40) and (41a).
Since dx = 0 on 0'", we have by (18):
1(0',,) = Lx p. dY = 2 Lx (<p dy - Jl dyd

= ± {[(sin x - x cos X)y2 + 2(cos x - I)YYl + (sin x)yn/L\(x)},
when Y = Y(x). (42)
This last follows because 2(<p dy - Jl dYl) is (for fixed x) an exact differential
of the function in braces, as may be verified by differentiation and comparison.
Using the estimates for Y(x) obtained previously, we have for small x, that
11(0',,)1 :$; M2 x 2 [(sin x - x cos x) + 2(1 - cos x)x + x 2 sin xJ/L\(x).
By using, say, L'Hopital's rule, it can be shown that lim" ... o L\(x)/x4 is non-
zero, while lim"",o ([term in bracketsJ/x 2 ) = O. It follows that 11(0',,)1--+ 0 as
x'" o.
To establish the corresponding results as x )" 2n, we note that since
Y(2n) = (!) we have
1 Y(x) 1 :$; IJf,,2" Y'(t) dt I :$; M(2n - x), while lim 2L\(X) is nonzero.
")'2,, n-x
Thus in particular, we have established the Wirtinger inequality: F(y) ~ 0,

V Y E [!} for which f~" y(x) dx = O. Moreover, if equality holds and F(yo) = 0,
then by the methods of §6.7, we know that Yo is stationary for j with J.l =
const., and hence must be of the form y of (38) with $5" Yo(X) dx = 21tJ.l = O.
Thus Yo(x) = A sin x. Conversely, any function of this form gives F(yo) =
A$5" (cos 2 x - sin 2 x) dx = O. (A direct ad hoc proof of the Wirtinger in-
equality is given in [H-L-PJ, while a field for the isoperimetric problem
itself is presented in Problem 9.20.)
§9.6*. Central Fields

With the brachistochrone problem as analyzed in §9.3, we saw that we could
find a unique stationary trajectory (cycloid) joining the point (X = (0, 0) to a
given point in D = {(x, y): x > 0, y > O} by making a suitable choice of the
parameter A = A(X, y), say. The functions ",(.; A) representing these trajec-
tories determine in D, the stationary field
cp(x, y) = ""(x; A(X, y)), (43)
which, by Corollary 9.5, is exact. As we have observed, in higher dimensions,
more than stationarity is required for exactness of a field, and we shall now
obtain this for a field determined by a suitable generalization of the family",.
(9.12) Definition. For a given f = f(x, Y, Z), a family of stationary functions

qt(.; A) indexed by a parameter A E !Rd, is a central family with center
(X = (a, A) E !Rd +1, if qt(a; A) = A, V A, and, in addition, both qt = qt(x; A)
and qtx are C l .
(9.12') Definition. For a given f = f(x, Y, Z), a central field in a domain D

is a stationary field whose trajectories are defined by a central family (qt; A)
for which qtA is invertible "in D" (i.e., at each (x 1, Ad for which (x 1, Yl ) E D
when Yl = qt(Xl; Ad)·
A = A(x, Y) is then called a parameter for the field.
Having A in !Rd makes possible the following result:
(9.13) Lemma. Let qt(.; A) be a central family for f. If the matrix qtA with
elements 8"'d8Aj , i, j = 1,2, ... , d is invertible at (Xl' A l ), then in a neighbor-
hood Dl of (Xl' Yl == qt(Xl' A l )), for values of A near A l , A = A(x, Y) is
defined and Ct, and
(x, Y) = qt'(x; A(x; Y)) ~ qtt(t; A(x, Y))lt=x (44)
is a central field in D l .
PROOF. By implicit function theory, the equation qt(x; A) = Y determines

A = A(x, Y) as a (unique) C l function in a neighborhood Dl of (Xl' Yd such
§9.6*. Central Fields 309
'I'(x, A(x, Y)) == Y. (45)

It follows that through each (x, Y) E D1 , passes one and only one trajectory
of the possibly smaller family, namely, that with parameter A = A(x, Y).
Thus CI> is defined in Dl by (44) and it is C 1 by the chain rule and our
differentiability requirements on '1'. It is stationary by construction, and 'I'A
is invertible in a possibly smaller neighborhood Dl by continuity. D
This lemma shows that however indexed, a central family will determine a
central field in a neighborhood of each point associated with an invertible
matrix 'I'A. Outside this neighborhood, the associated trajectories may inter-
sect in a quite complicated manner, and in particular, they must intersect at
the center 0(. Moreover, other trajectories of the original family might inter-
sect these even in this neighborhood. The situation is illustrated in Figure 9.9.
Example 1. In the last section, we saw that with A = (p., A) E 1R2, the functions
of equations (38); viz.,
lT/(X.
T ,
A) = {P.(l - cos x) + A sin x, X E [0, 2nJ,
p.(x - sin x) + A(l - cos x),
are stationary for a certain j, and clearly '1'(0; A) = (!) E 1R2, while both 'I' and
'l'x are C 1 everywhere. It follows that '1'(.; A) is a central family for this j with
center 0( = (!).
Moreover, by the computation of (39), we know that for x E (0, 2n) the
Jacobian
and hence 'I'A is invertible.
a x
Figure 9.9
We already know that each point (x, Y) E D = (0, 2n) x 1R2, determines a
unique A, so that the field <II == (cp, CPl) is well defined by (41). It follows that
<II is a central field for j in the domain D which is that required for the
application.
Lemma 9.13 shows that the field parameter function A is in C1(D), and
this can now be used to show that <II is exact. More generally, we have the
following.
(9.14) Theorem. A central field is exact.
PROOF. For a given f = f(x, Y, Z), let <II be a central field for f in a domain D
of IRd + 1 with center 01: = (a, A), where we may suppose a < x, V (x, Y) E D.
By definition, <II is determined by a central family '1'(.; A) (with center 01:)
where by Lemma 9.13, the field parameter A E C1(D). Thus to each point
(x, Y) E D is associated the stationary function '1'(.; A(x, Y)) E (C1[a, X])d;
then
x
S(x, Y) =
def
fa f['I'(t; A(x, Y)] dt (46)
is defined in D, and depends on (x, Y) only through A(x, Y) and the upper
limit of the integral. We shall prove that <II is exact by demonstrating that in
D, S has partials of the correct form (9). It is here that we shall use the
differentiability of'll and A guaranteed by 9.12 and 9.13.
Recall that f[Y(t)] == f(t, Y(t), Y'(t)) and introduce the abbreviation
t= (t, A(x, Y)). Then by A.13 and the chain rule, using the prime to denote
"t" differentiation, we get
where 'I'A and Ay are the matrices of partial derivatives with, respectively,
columns and rows indexed by A, and ('I")A is the matrix of t derivatives
corresponding to 'I'A- Here Sy, fy, and fz are row matrices.
Now, by assumption ('I")A = ('I't)A = ('I'A)t = ('I'A)', and 'I'(t) =
'I'(t; A(x, y)) is stationary on (a, xl Hence, fy['I'(t)] = (d/dt)fz['I'(t)], and
so (47) becomes
But
'I'A(ii) = (!), since 'I'(a; A) = A, V A;
and if Id denotes the d x d identity matrix,
'I'A(i)Ay(x, Y) = Id , since 'I'(i) = 'I'(x; A(x, Y)) = Y.
Finally, since
(x, Y) ~ qt'(x; A(x, Y)) = qt'(x);
Sy(x, Y) = fz[qt(x)] = fz(x, Y, (x, Y)) (48)
= P(x, Y), as in (9).
Similarly, by A. 14, from (46):
S,,(x, Y) = f[qt(x)] + t' {fy[qt(t)]qtA(t) + fz[qt(t)] (qt')A(t)}A,,(x, Y) dt
= f[qt(x)] + fz[qt(t)]qtA(t)C: Ax(x, Y).

Here Ax is a column matrix. Again qtA(ii) = (9, but now differentiating
qt(x; A(x, Y)) == Y with respect to x shows that
qtA(x)A,,(x, Y) = -qt'(x; A(x, Y)) = -(x, Y).
Hence
Sx(x, Y) = f(x, Y, (x, Y)) - fz(x, Y, (x, Y))· (x, Y)
(49)
= h(x, Y), as in (9).
Equations (48) and (49) prove that is exact by Definition 9.3. D
The reader should verify each step of this derivation in the simple case
d = 1, at least by formal operations.
Theorem 9.14 justifies the incorporation of the term "central field" in the
statements of Theorems 9.7, 9.9, and 9.10, which should now be reexamined.
Observe that both fields considered in Example 1 of §9.3 are exact, but only
one, that obtained from the semicircles with fixed (left) end point !X is central
(Problem 9.3); moreover, it is the one with smaller domain. (Attempts to
enlarge this domain by including say the semicircles with the same fixed right
end point !X result in a pair of disjoint quarter planes.)
In order to apply Hilbert's theorem (9.7) for central fields, we must be able
to find or construct them. For a given f, and center !X = (a, A) E IRd +l, the
collection of "all" stationary functions qt E (C 1 [a, b])d with qt(a) = A (and
possibly variable b) constitutes a potential central family. It may always be
parametrized by A = qt'(a), or, some other choice may be more natural.
However, whatever choice is made, the resulting family, qt = qt(x; A), with its
derivatives, qtx = qt,,(x; A), must be C 1 in all variables. In explicit cases this
technical requirement is usually met automatically.
Supposing that qt(.; A) is a central family, then each domain D covered

consistently by its trajectories constitutes the domain of a stationary field <1>.
However, in order to claim that it is a central field, we must establish that
A E C1(D), or by Lemma 9.13 that the matrix qtA is invertible "in D."
Alternatively, we may simply examine the matrix qtA for (x, Y) domains of
invertibility. Each point (Xl' Yl ) of invertibility provides through Lemma
9.13, a local central field, but each domain of invertibility provides only a
possible central field. Coverage of the domain is assured, but consistency
must still be established (and need not be present).
An additional complication to applications is the fact that the center Q(
must be located on the extension of a trajectory of interest but be outside the
domain of a usable central field containing this trajectory. In our analysis of
the brachistochrone in §9.4 we have presented one method of confronting this
problem. However, the usual approach is to extend the given trajectory
slightly, make its new end point a, say, the center, and show that a resulting
central family can still produce a central field in a domain which contains the
original trajectory. We shall adopt this approach in the following application
and in the embedding construction of the next section.
Smooth Minimal Surface of Revolution

Example 2. We know from our discussion in §7.5 that when the positive
configuration parameters b, bl permit, there may be a smooth curve ro
parametrized by Yo E C l [0, b] which minimizes the area of the surface
obtained by revolving it around the x axis. Yo is stationary for j on [0, b],
where j~, y, z) = y~ is strictly convex on [0, b] x (0, (0) x IR from
Example 3 of §3.3. -
By equation (14') of §7.5, each positive function y, stationary for this j,
must be of the form
y(x) = c cosh(cx + j1.), (50)
for some choice of the constants c > 0, c == c-t, and j1., and hence must graph
as a catenary. We seek a central family t/J(.; J..) of such functions with center
Q( = (a, a l ), say, with a < 0, whose graphs are consistent on [0, b]. On geo-
metrical grounds, it is evident that this is possible only in domains D of 1R2,

such as that shown in Figure 9.10. D is unbounded above, but it is surely
bounded below by that catenary through Q( which "first" touches the x axis.
A central family through Q( in the form (50) is given by
t/J(x; J..) = c cosh(c(x - a) + J..), for al = c cosh J.., (51)
since t/J as defined, is C2. Observe that t/J'(a; J..) = sinh J...
Using the first of the addition laws
cosh(t + 't) = cosh t cosh 't + sinh t sinh 't,
(52)
sinh(t + 't) = sinh t cosh 't + cosh t sinh 't,
equation (51) may be rewritten as follows:
t/J(x; J..) = al (cosh u + sinh u tanh J..), (53)
where u = al(x - a) cosh J.., and al == (ad- l .
t/I(x; 0)
a b x
Figure 9.10
For fixed x > a, u, and hence, "'(x; A), increases strictly to +00 as A
traverses the positive real axis. It follows that for each such x, "'(x; A) assumes
each value ~ "'(x; 0) = a l cosh al (x - a), precisely once for some choice of
A ~ O. Thus for A ~ 0, ",(.; A) provides a central family of catenaries which
remain consistent in the domain
Da = {(x, y) E 1R2: x > a, y > a l cosh al(x - a)}.

To show that ",(.; A) determines a central field in Da , it only remains to
establish that the scalar "';. is invertible in Da , i.e., that it is non vanishing. But,
from (53) and the chain rule, we have
"';.(x; A) = a l {u tanh A(sinh u + cosh u tanh A) + sinh u(cosh Af2};

after some manipulation involving (52), we get
"';.(x; A) = al(cosh Af2[u sinh A sinh(u + A) + sinh u], (54)

which for A ~ 0, is positive with u; so, \:Ix > a.
Now fix Al ~ 0 and set Yl = ",(.; Ad. Using Hilbert's theorem (9.7), we
conclude with the
(9.15) Proposition. The catenary r l represented by Yl provides (uniquely) the

minimum surface of revolution among those admissible piecewise smooth curves
joining its end points in Da· D
When A1 > 0, Y1 provides a strong local minimum for the surface area
function, in the sense of Chapter 7. (Why?)
It is considerably more delicate to introduce trajectories for A < 0 in order
to enlarge Da since they may intersect each other, as well as those already
present. However, although, as in Problem 7.15, some configurations may
permit another catenary i\ to join these same end points, it is clear from the
above that it cannot provide a lesser surface area if it is contained in Da , or in
any legitimate enlargement of Da. For a complete discussion, see [Bl]. (A
similar problem is analyzed in Example 1 of §9.9.)
Note that we still have not obtained a result for the given catenary ro
unless Yo = l/I(.; Ao} for some Ao. To ensure this, we must enlarge the interval
of definition of Yo, and choose a1 = yo(a} for some a < O. However, to con-
struct the family with this center as above requires that Yb(a} = sinh Ao ;::: O.
Geometrically, it is seen that this is assured if and only if yb(O} > O. In the
next section we shall construct such families for a general f.
§9.7. Construction of Central Fields with Given

Trajectory: The Jacobi Condition
As we have seen in this chapter, in order to obtain a complete solution for a
minimization problem involving a convex integrand f~, X; Z} with given
end point conditions we should first obtain a stationary trajectory ro which
meets these conditions and then show that ro is a trajectory of a central field
 for f (or of some other exact field for f). In the previous examples this was
accomplished usually by appeal to geometric arguments which enabled us to
"see" the central family producing the field. However, it is clearly of impor-
tance to characterize if possible, those f and ro which afford such fields, at
least in a neighborhood of roo (The minimal surface example of §9.6 shows
that even on geometrical grounds this may be the best achievable.) In the
present section we shall obtain sufficient conditions to guarantee this local
embedding of ro under further assumptions on f.
First, we utilize results from the theory of systems of ordinary differential

equations to construct a central family '1'(.; A} of stationary functions con-
taining a representative Yo for roo This, in principle, is always possible.
Second, we examine the matrix 'I'A and show that its invertibility along ro
guarantees the field construction. This invertibility is not always present;
however, as we shall prove in §9.9, it is almost essential if Yo is to provide even
a weak local minimum.
We operate with a fixed Yo E (C 1[a, b]}d which is stationary for a given

§9.7. Construction of Central Fields with Given Trajectory 315
f = f(x, Y, Z) assumed to be C 3 at least in a neighborhood of

rt'o = {(x, Yo(x), Yo(x», x E [a, b]} !:; 1R 2d +1 ;
We further suppose that fzz, the matrix of second partial Zj derivatives, is
invertible along rt'o, and so by continuity, in this neighborhood. Then by
Theorem 7.14, Yo is C 2 , and so it satisfies the Euler-Lagrange equations of
§6.7 for f in the condensed form
Y'(x) = Z(x),
(55)
:xfz(x) = fxz(x) + fzy(x)Z(x) + fzz(x)Z'(x) = fy(x),
where fzy is the matrix with elements fz., yJ., i, j = 1, 2, ... , d having its rows
indexed by i, while the remaining matrices are columns.
Using the hypothesized invertibility of fzz, we conclude that (Yo, Zo = Yo)
satisfies the system
Y'(x) = Z(x),
(56)
Z'(x) = G(x, Y(x), Z(x»,
where
G(x, Y, Z) = f ZZl [fy - fxz - fzyZ], (57)
with the partial derivatives of f evaluated at (x, Y, Z).
Now, (56) is a first-order system for the vector function T == (Y, Z) E

(C 1 [a, b])2d. Moreover having f in C 3 ensures that Gis C 1 in the neighbor-
hood of invertibility of fzz. Upon setting A = Yo(a) and Ao = Yo(a) = Zo(a),
we can use the theory of such systems as presented in §A.5, to assert the
following:
(9.16) Proposition. Under the above conditions, Yo is contained in a central

family \{I(.; A) for 1, defined on [a, b]. Moreover, V A: \{I(a; A) = A = Yo(a)
and \{I'(a; A) = A.
PROOF. The theory provides a b > 0 so small that for each A E IRd with
IA - Aol < b, there exists a unique solution (Y, Z) of the system (56) on [a, b]
which meets the initial conditions Y(a) = A, Z(a) = A; set this Y = \{I ( .; A).
The differentiability of \{I and \{Ix required by Definition 9.12, also follow
from the general theory, since \{IA'; A) = Y' = Z. 0
Of course, the trajectories of this family need not cover a domain con-
sistently. They obviously intersect at IX = (a, A), and we are indexing the
family by their slopes at IX. The situation is illustrated in Figure 9.11 where it
is seen that the trajectories appear to remain distinct, at least for small x-a.
In view of Lemma 9.13, the determinant of the matrix \{IA plays a crucial
role, and it suffices to consider it when A = Ao. Upon differentiating (56) with
respect to A and using the chain rule, we obtain for the matrices, \{IA and ZA'
Y
Graphs of
(a,A)-=:::::::-~
--....,,--::::::=----1 } ,¥(.; A)
----==----:::::~S:...:::;~ Yo = ,¥(.; Ao)
a b x
Figure 9.11
evaluated at Ao, the system

'P~(x) = ZA(X),
(58)
Z~(x) = Gy(x)'PA(X) + GZ(X)ZA(X),
where the partials of G are evaluated along ~o (Why?) and so are known a
priori. In addition, we have that
'PA(a) = (D, since 'P(a; A) = A, V A;

and (59)
ZA(a) = Id , since Z(a; A) = 'P'(a; A) = /<., V A.
Observe that (59) may be written as the linear system
,
T (x) = Gy(x)
[(D Id ]
Gz(x) T(x) = G(x)T(x), (60)
say, where T == ('PA, ZA) E ~N, with N = 2d 2 . Id is the identity matrix of

order d.
(9.17) Proposition. If for I> > 0 the coefficient functions in the matrix G are
continuous on an interval I. = [a - 1>, b + 1>], then on this interval, each solu-
tion T of the system (60) admits an estimate
I T(x) 1 ::;; eMlx-all T(a)l, for a constant M = M(I» > O. (61)
PROOF. Let N = 2d 2 • Each coefficient function 9u in the matrix G is bounded
on I. by MjN 2 , say. (Proposition 5.3). The dot product of (60) by
T(x) = (tl (x), t 2(x), ... , tN(x)) is
~ ~ IT(xW =
2 dx
T(x)· T'(x) = f
i.j=l
gij(x)ti(x)tj(x)
::;;MIT(xW,
since
1tix)1::;; I T(x) I, j = 1,2, ... , N.
§9.7. Construction of Central Fields with Given Trajectory 317
Thus for x > a:
so that
or
IT(x) I ~ eM(x-a)1 T(a)l.
A similar argument gives the desired inequality (61), when x < a. 0
Recall that our coefficient functions J, and hence, G, together with its
partials, Gy and Gz , are defined in a neighborhood of '1&'0' Thus we may
suppose that Yo, and, by Proposition 9.16, that '1'(.; A), together with 'I'A
and ZA, are defined on [a - /:, b + /:] for some /: > O. Then in view of (59),
Proposition 9.17 proves that each solution to (58) on this interval of length
I = b - a + 2/: has components bounded by eMl , for M = M(/:). (Why?) Fur-
thermore, each component bij(x) of the matrix
def
BW=~W~W+~W~W ~~
is also bounded by some M = M(/:). Integrating (58) and incorporating the
initial conditions (59) gives
'I'A(X) = 'I'A(a) + LX ZA(t) dt,
or
'I'A(X) = (x - a)Id + LX dt f B(-r) d-r, (63)
where B is given by (62).
(9.18) Lemma. L\(x, a) = det['I'A(x)] of. 0, for 0 < Ix - al < 2/Md, where
M = M(/:).
PROOF. If L\(x, a) ~ 0 for some x of. a, then 'I'A(x) is not invertible and there
exists a U = (U 1 , U 2 , .•• , U d ) E ~d with II UII = maxj=l, ... ,d Iujl = 1, for which
'I'A(X)U = (!). From (63), we have
(x - a)U = -Lx dt f B(-r)U d-r,

so that for x > a:
I It
x
Ix - al = Ix - a111U11 ~ a dx a IIB(-r)UII d-r ~ Md
(x - a)2
2 ;
i.e., (x - a) ~ 2/Md, and again the prooffor x < a is similar. D
(9.19) Remark. Since the constants M = M(/:) used in the preceding argu-
ments apply to each solution of (58) on the interval [a - /:, b + /:], it follows
that the solution Yo can be extended to a larger interval [a, b], where a < a,
such that a corresponding central family with center a. = (a, Yo(a)) will have
~(a, a) =f. O. Hence, by Lemma 9.18, that portion of ro near a. = (a, Yo(a)) is a
trajectory of a central field for f
To discuss a total embedding of r o, we shall require one additional fact

about ~.
(9.20) Proposition. ~(x, a) is (jointly) continuous on [a - 1>, b + I>Y

PROOF. For each fixed a, ~(x, a) is, by definition, the determinant of the
solution qJA of the linear system (58) with the boundary conditions'llA(a) = (!),
'P~(a) = I d • Clearly, each component of qJA and, hence, its determinant,
(which is just a sum of products of these components) is continuous in x.
Moreover, by analogy with (63):
qJA(X) - qJA(X O) = (x - xo)Id + IX dt

Xo
f:a
B(r) dT,
so that
IlqJA(X) - qJA(xo)11 ~ Ix - xol + Mix - xol(b - a + 21»
~ Mllx - xol, for Ml = Ml(I»·
(In estimating the inner integral, we observe that la - tl ~ b - a + 21>.)
To obtain the continuity in afor fixed x = Xo, it is simpler to return to the
general linear system. (See Problem 9.21.) However, once this is established,
the joint continuity of each component of'llA (and hence of ~) follows from
the triangle inequality:
II'PA(x, a) - 'PA(X O , ao)11
~ II'PA(x, a) - 'PA(X O, a)11 + II'PA(x O, a) - 'PA(X O, ao)ll
~ Mllx - xol + II'PA(x O ' a) - 'PA(X O, ao)ll,
since the right side may be made as small as we wish, when
Ix - xol + la - aol is sufficiently small. o
(9.21) Proposition. If ~(x, a) =f. 0 for x E (a, b], then ro is a trajectory of a
central field for f
PROOF. It follows from the hypothesis that ~(x, a) =f. 0 on [a, b] for some
a < a. Suppose to the contrary, that for a sequence an /' a, as n --+ 00, we
had ~(xn' an) = 0 for some Xn E [a, b], n = 1, 2, .... Then by Lemma 9.18,
IX n - ani ~ 21Md, so that IX n - al ~ 11Md, ifn is large enough, i.e., Xn E 10 ~
[a + (Mdr l , b]. But this interval 10 is compact and as established in §A.O,
there is a subsequence x nk --+ Xo E 10 , as k --+ 00. Finally, by the joint continu-
ity of ~, we have ~(xo, a) = limk--+oo ~(xnk' anJ = 0, which contradicts the
hypothesis, since Xo > a.
§9.8. Sufficient Conditions for a Local Minimum 319
Then by Lemma 9.13, the trajectories of a central family 'P('; A) with A

near Ao and center at IX = (Ii, Yo(Ii)), cover consistently a neighborhood of
each point on the trajectory ro of Yo = 'P('; Ao). 'PAis invertible in each of
these neighborhoods and their union constitutes a domain D of a central
field; obviously ro E D. 0
(9.22) Definition. When L\(x, a) #- 0, V X E (a, b], then ro is said to satisfy the
Jacobi condition for f
By Lemma 9.18, it is always satisfied if b - a is sufficiently small. Unfortu-

nately, this condition is not always satisfied on larger intervals even when I is
simple.
For example, I(y, z) = Z2 - y2 has stationary functions I/I(x; A) = A sin x
on [0, 3n/2], satisfying 1/1(0; A) = 0 and 1/1'(0; A) = A. But L\(x, 0) = I/I;.(x; Ao) =
sin x and L\(n, 0) = O.
On the other hand, L\(x, 0) #- 0 on (0, b] for each b < n. Hence, for each Ao,
the associated trajectory ro is a trajectory of a central field, but the extent of
the field will depend on the particular Ao.
In §9.9, we will see that the nonvanishing of L\(x, a) on (a, b) is in fact
necessary, if Yo is to provide a weak local minimum.
§9.8. Sufficient Conditions for a Local Minimum

We now have the ingredients required to provide sufficient conditions for
local minimization. As in §9.7, we assume that Yo in (C 1 [a, b])d is stationary
for a given I = I(x, Y, Z) which is C 3 in a neighborhood of
~o = {(x, Yo(x), Y~(x)) E [R2d+1: x E [a, b]}.
We know from Proposition 7.19 and subsequent remarks, that if Yo provides

a local minimum, then necessarily
VVE [Rd, X E [a, b].
Conversely, if the matrix Izz of these second partials is positive definite

(i.e., equality holds iff V = (9) at a single point on ~o, then, by continuity, it
remains so in a spherical neighborhood of this point. Thus as in §0.13, it is
invertible, and by Proposition 9.2, IUs., Y, Z) is strictly convex in this neigh-
borhood. Although this will suffice to establish that in the neighborhood Yo
is a weak local minimum, i.e., a local minimum with respect to the weak norm
II YII = max(1 Y(x) I + IY'(x)I), we must permit Y' with arbitrary values in [Rd
in order that Yo could minimize with respect to the strong norm, 1YI =
maxIY(x)l.
It is convenient to treat first the results which can be obtained from
conditions specified at a point. The analogous results for the full trajectory
require the Jacobi condition.
(a) Pointwise Results

(9.23) Theorem. If Yo is stationary for a C 3 function f = f(x, Y, Z) and fzz is
positive definite at y = (a, Yo(a), Y~(a», then for sufficiently small b - a, Yo
on
F(Y) r
provides a unique weak local minimum for
= f(x, Y(x), Y'(x» dx
~ = {Y E (C 1[a, b ])d: Y(a) = Yo(a), Y(b) = Yo (b)}.

If, in addition, f(}£, X; Z) is [strictly] convex on D x /Rd, where D is a neighbor-
hood of at = (a, Yo(a», then Yo provides a [unique] strong local minimum.
PROOF. As noted above, f(}£, X; Z) is strictly convex in a neighborhood of y
(which we may suppose to be Z-convex). We may also assume that fzz is
invertible in this neighborhood and, hence, along ~o, for b - a sufficiently
small, so that by 9.18 and 9.21, at = (a, Yo(a» is contained in the domain Do
of a useable central field. Hilbert's comparison, F(Y) > F(Yo), is possible for
all Y in ~, (Y "# Yo), with graphs f in Do, for which (x, Y(x), Y'(x» is in the
given neighborhood of y, V x E [a, b]. (At the corner points, we require this
for both one-sided derivatives.) But this defines a weak neighborhood of Yo
in~.
If, in addition, f(}£, J', Z) is [strictly] convex in Do x /Rd, then we may
remove these restrictions on Y', and apply Hilbert's theorem to all Y in ~
which are in a strong neighborhood of Yo' 0
Observe that the positive definiteness of fzz at a point ensures that on a

small interval determined by the point, Yo is the unique weak local minimiz-
ing function. This result was known to Legendre (1786) who believed errone-
ously that it would extend to the full trajectory. By contrast, the similar claim
for Yo as a strong local minimum is not guaranteed unless f(}£, X; Z) is strictly
convex on some D x /Rd, and this cannot be inferred from a pointwise condi-
tion. However, Proposition 9.2 should be recalled. (Problem 9.23.)
Application: Hamilton's Principle

When the Lagrangian expressed in generalized coordinates is given by
. 1
L(t, Q, Q) =2 ;.#1 aij(t, Q)tj;tjj -
n
U(t, Q) where aij = aj;, i, j = 1,2, ... , n.
then [~zz] = [LQQ] = [a;j] is always positive definite. (See §8.4.) Thus
L{J., Q, Q) is strictly convex on D x /Rn for some domain D of /Rn+1, by 9.2.
Hence, supposing that L is C 3 , it follows that each point on a stationary
trajectory is contained in the domain of a useable central field. By Theorem
§9.8. Sufficient Conditions for a Local Minimum 321
9.23, each stationary trajectory provides a strong local minimum for the
action integral
A(Q) == r L(t, Q(t), Q(t)) dt.
when b - a is sufficiently small, among nearby trajectories with the same end
points at a, b.
Thus Hamilton's principle of stationary action becomes in this case a
principle of (strong) local minimal action, even though the total action might
be maximized by this same stationary function.
(b) Trajectory Results
An extension of Theorem 9.23 which provides corresponding information

about the entire trajectory r o, for a given stationary function Yo E (C 1 [a, b])d,
is more difficult to obtain, and must be nonexistent in some cases. Even those
stationary functions which minimize in neighborhoods of each point may not
do so on an interval, and so we cannot hope to make their trajectories those
of central fields. We must invoke the Jacobi condition of the last section
(which depends only on information along ~o). Remarkably, this is the only
additional requirement.
(9.24) Theorem. If Yo E (C 1 [a, b ])d is stationary for a C 3 JUnction f = f(x, Y, z)

and
(i) fzz is positive definite along ~o;
(ii) A(x, a) of. 0, x E (a, b];
on
F(Y) r
then Yo provides the unique weak local minimum for
= f(x, Y(x), Y'(x)) dx
~ = {Y E (C1[a, b])d: Y(a) = Yo(a), Y(b) = Yo (b)}.

If, in addition, f~, X, Z) is [strictly] convex in D x ~d, where D is a domain
containing ro (the graph of Yo), then Yo provides a [unique] strong local
minimum.
PROOF. (i) implies by Proposition 9.2 that f~, X, Z) is strictly convex in a
spherical neighborhood of each point on ~o. The union, 0/10, of these spheres
will be a Z-convex domain containing ~o in which f~, X, Z) is strictly
convex by Proposition 9.2. Moreover, the projection of 0/1 0 on the (x, Y)
space is a domain containing ro as illustrated in Figure 9.12.
Then, (i) and (ii) together assure, by Proposition 9.21, that ro is a tra-
jectory of a central field (in perhaps a smaller domain Do). Again we may
z <¥to
b
x
Figure 9.12
apply Hilbert's comparison F(Y) > F(Yo) to all Y E ~ '" {Yo} which are in a
weak neighborhood of Yo. If, in addition, f~, y, Z) is [strictly] convex in
Do x !R d , then Hilbert's theorem (9.7) applies to all Y E ~ which are in a
strong neighborhood of Yo. That such a neighborhood exists requires an
appeal to compactness. D
We defer application of this result until the end of the next section.
§9.9*. Necessity of the Jacobi Condition

In the previous section, the Jacobi condition, A(x, a) "# 0 on (a, b], was used
to embed a given stationary trajectory ro in the domain of a central field on
which Hilbert's theorem could act to show that Yo provides a local minimum.
If we give up the nonvanishing at x = b, the reduced condition is necessary
for Yo to be a weak local minimum (or a weak local maximum).
on
F(Y) r
(9.25) Theorem (Jacobi). Let Yo be a weak local extremum for
= f[Y(x)] dx
~ = {Y E (C 1 [a, b])d: Y(a) = A, Y(b) = B}

and suppose that f is C 3 in a neighborhood of ~o, while fzz is invertible along
~o. Then A(x, a) "# 0 on (a, b).
PROOF. (We may assume that Yo is a weak local minimum.) Suppose to

the contrary, that A(a*, a) = 0, for some point a* E (a, b). Then the matrix
'PA(a*; Ao) is not invertible, and it follows that 'PA(a*; Ao) U = mfor some
(column) vector U E !R d of unit length.
§9.9*. Necessity of the Jacobi Condition 323
Set
V(X) = 'PA(x; Ao)U, so that V(a) = V(a*) = (!),
and observe that since 'Px is C 1 , then
Now, it is convenient to return to the original Euler-Lagrange equations

for the stationary family 'P(x; A) as in §6.7; viz.;
d
dxfz['P(x; A)] = fr['P(x; A)]. (64)
Upon differentiating (64) with respect to A and subsequently evaluating at

A o, we obtain
:x [fZY(x)'PA(X) + fzz(x)'P~(x)] = fyy(x)'P A(X) + fYz(x)'P~(x), (64')
where the double partials of f are matrices with columns indexed by the
second subscript, and rows by the first, evaluated on ~o. Post-multiplying this
equation by the constant column vector U and incorporating the properties
of Vas above, we obtain
:X [fZY(x) V(x) + fzz(x) V' (x)] = fyy(x) V(x) + fyz(x) V' (x). (65)
It may be verified that (65) is the first Euler-Lagrange equation for the
quadratic function
der - - -
q(x, V, W) = Vfyy(x) V + 2 Vfyz(x) W + Wfzz(x) W, (66)
where V, Ware (column) vectors in [Rd and V, Ware the corresponding row
vectors. (See Problem 9.22.) However, for our purposes, it suffices to observe
that since Vfyz(x) W = Wfzy(x) v,
r' q(x, V(x), V'(x)) dx = r' {V(X) [fyy(x)V(x) + fyz(x)V'(x)]

+ V'(X) [fZY(x)V(x) + fzz(x)V'(x)]} dx = 0,
(66')
if we utilize (65) together with the fact that V(a) = V(a*) = (!).
Next, for V E ~o = {V E (C 1 [a, b])d: V(a) = V(b) = (!)}, let

h(6) = F(Yo + 6V), for small 6.
Then by Taylor's formula with remainder, for fixed 6 > 0,
~ 62
F(Yo + 6V) - F(Yo) = h(6) - h(O) = h'(0)6 + h"(B)2"'
where 0 < 6 < 6. But h'(O) = c5F(Yo; V) = 0, since as hypothesized, Yo mini-
mizes F locally on ~ (weakly) and hence by Theorem 7.7, on

~ = {Y E (C 1[a, b])d: Y(a) = A, Y(b) = B}.
S:
Now, h"(O) = Q(V) ~ q(x, V(x), V'(x)) dx (Problem 9.22). Were h"(O) <
°
0, then by continuity of h", we would have h"(E) < for sufficiently small
B > 0, implying that F(Yo + BV) < F(Yo) contradicting the assumed mini-
mality. It follows that h"(O) ~ 0, or equivalently, that Q(V) ~ 0, V V E ~o.
However,
~ def{V(X), x::;; a*, . . ~ .
Vo(x) =
(!), x ~a ,
* IS m ~o, smce V(a*) = (!), (67)
and Q(Vo) = S:·

q(x, V(x), V' (x)) dx = 0, from (66'). Hence Vo minimizes Q on
~o and so, at the (possible) corner point a*, it must satisfy the Weierstrass-
Erdmann condition (7.13), which from (65) and (67) is
lim [fZY(x) V(x) + fzz(x) V' (x)] = lim [fZY(x)(!) + fzz(x)(!)] = (!),
x?a* x\"a*
or, since V(a*) = (!), we conclude that fzz(a*)V'(a*) = (!).
Butfzz(a*) is invertible by hypothesis so that V'(a*) = V(a*) = (!). Since V

is the unique solution of the homogeneous second-order linear equation
(65) on [a, a*], satisfying these conditions, it follows that when x E [a, a*],
'¥A(X)U = V(x) = (!). Thus, for x near a, '¥A(X) is not invertible and so
A(x, a) = 0, contradicting Lemma 9.18. (See A.17.) 0
(9.26) Remark. Observe that the Weierstrass-Erdmann argument cannot be

invoked if a* = b. The case considered in §9.1 of f(y, z) = Z2 - y2 where
°
A(x,O) = sin x while F(y) = So [y'(X)2 - Y(X)2] dx ~ shows that the van-
ishing of A(x, a) at b( = n) must be permitted.
(9.27) Remark. (65) is called Jacobi's equation. Its linearity (with fzz inver-
tible on fili'o) assures that for each j = 1,2, ... , d, it has a unique solution lj on
[a, b], with lj(a) = (!) and lj'(a) = Ej , where Ej is the unit vector of ~d in the
jth coordinate direction. The matrix having the lj(x) as columns is precisely
'¥A(X; Ao) and so A(x, a) = det[V1 (x)iV2(x)i'" iv,,(x)]. The Jacobi condition
may be restated in terms of the linear independence of these lj.
Example 1. f(y, z) = [y(1 + Z2)]1/2 is strictly convex on (0, (0) x R (Exam-

ple 4, §3.3.) The stationary functions for f on say [0, 1] are C 2 by 7.14 and
satisfy the second equation (of §6.3) f(x) - y'(x)fy.(x) = const., or after simpli-
fication, (y/l + Y'(X)2)1/2 = 1/Jc, say, so that
1 + y'(X)2 = cy(x). (68)
With the geometric substitution y'(x) = tan O(x), we obtain

cy(x) = sec 2 O(x),
§9.9*. Necessity of the Jacobi Condition 325
which, when differentiated becomes

cy'(x) = 2 sec 2 O(x) tan O(x)O'(x),
or (when y'(x) = tan O(x) i= 0),
d c
dx tan O(x) = sec 2 O(x)O'(x) ="2 = CI , say.
Thus y'(x) = tan O(x) = cl(x - co), for a constant Co, or finally, using (68)
once again, we obtain
(69)
We shall consider those functions of this form capable of providing local
(minimal) values for
F(y) = L f[.p(x)] dx
on
~ = {y Eel [0,1]: y(O) = 1; y(1) = t}.
To satisfy the condition y(O) = 1, we must take cic~ = 2c I - 1 in (69), and
after some manipulation, we obtain for C I = (A 2 + 1)/2, the central family of
quadratic functions:
(A 2 + 1)x 2
l/J(x; A) = 4 + AX + 1, (70)
in which the parameter Ahas been chosen to have l/J'(O; A) = A. Observe that
A~ 2
A(x,O) = l/J;.(x; A) == T + x, = 0 when X = -I. (71)
If one of the associated parabolas passes through the point (x, y), then
from (70), we have after competing the square, 4y = [(AX + 2)2 + x 2 ]. Thus
no curve of the family enters the region in which 0 < 4y < x 2 , and exactly one
curve of the family passes through each point on the boundary of this region
where 4y = x 2 • However there will be two curves of the family passing
through each point of the complementary region in which 4y > x 2 • In partic-
ular, for the point (1, t), there are the curves corresponding to the parameters
A = -1 and A = - 3 shown in Figure 9.13. From (71), l/J;.(x; -1) i= 0 on
(0, 1]. Hence the Jacobi condition is satisfied, and we are assured by Theorem
9.24 that YI(X) = l/J(x; -1) = x 2 /2 - x + 1, provides a strong local minimum
value for F on ~. But also from (71), l/J;.(x; -3) = 0 when x = l Thus by
Theorem 9.25, the curve Y3(X) = 5x 2 /2 - 3x + 1 cannot provide even a weak
local minimum (or maximum) value.
If we consider the corresponding problem in the parametric version of

minimizing
G(Y) = 1 0
5/2
y(t)I/21 Y'(t)1 dt,
Figure 9.13
on
~ = {Y = (x, Y) Eel [0,1])2: Y(O) = (0, 1); y(i) = (1, t), Y ~ O},
then, as in the application of §7.5, we must admit the Goldschmidt curve to
consisting of the horizontal segment along the x axis and the two vertical
segments at its ends. By direct calculation in which we utilize the arc length
as the parameter t, we obtain
G(Yo) = II yI/2 dy + 12 0 dx + I /2 yI/2 dy = ~(1 + 2- 3/2 ),

and this is less than
F(YI) = II [Y1(x)(1 + y~(X)2)]1/2 dx = J2C: f YI(X) dx
=.j2 II (x 2/2 - x + 1) dx = (i).j2.
(In making this last computation we have used (69), with 2c I = 1 + A,2, where
A, = -1.) We can approximate the Goldschmidt function Yo by smooth func-
tions, Y for which the integral G(Y) = F(y) has values as near G(Yo) as we
wish (Proposition 7.6), and we conclude that YI gives only a strong local
minimum value for F. While it is true that for this problem, the upper of
two possible (stationary) parabolas through the same pair of points always
§9.10. Concluding Remarks 327
provides a local minimum value, it may sometimes provide a minimum less

than that for the Goldschmidt curve (Problem 9.26). A complete discussion is
given in [P].
The situation illustrated in this problem occurs with some more general
strongly convex functions of the form p(x, y)~ with p > 0, including
the minimal surface of revolution function where p(x, y) = y. In particular,
when there are two catenaries through the same points (as in Problem 7.15),
then the upper satisfies the Jacobi condition (9.22) and will give a strong local
minimal surface area which mayor may I?-ot be less than that provided by the
Goldschmidt solution, while the lower will not supply even a weak local
extremal value. These arguments are facilitated by a geometric interpretation
of the Jacobi conditions. See [P], [BI].
§9.10. Concluding Remarks

In Chapter 3, we demonstrated that convexity of f0i, Y, Z) provides an
elementary sufficiency proof for minimization even with variable end point
conditions and constraints. In this chapter, we have obtained analogous
results when f0i, X, Z) is convex, provided that the existence of a suitable
field can be established.
Although this field theory approach appears to be the only one capable
of supplying the results sought within the scope of the text, it involves sub-
tleties which as we have seen, make each application to a specific problem
extremely difficult. It should be used only as a last resort.
In this book, an effort has been made to present those results which can be
established without use of the Lebesgue integral. However, we have not
examined the special class of homogeneous integrand functions which yield
integrals independent of the particular parametrization of the underlying
trajectories. The study of the associated curve dependent integrals is pre-
sented in such works as [S], [G-F] and [Ak], and the relation between a
resulting field theory and the Hamilton-Jacobi equations for the trajectories
is investigated thoroughly in [C] and in [Ru]. In the latter work will also be
found a corresponding development for multidimensional integral functions.
Convexity plays a significant role in the examination of certain multidimen-
sional problems in which the underlying domain of integration is permitted
to vary. See [G-F] and [P-S].
The principal benefit derived from the introduction of the Lebesque inte-
gral is the possibility of establishing existence of a minimizing function in a
larger class of functions than those admitted in this book. These methods
were first carried through successfully by Tonelli (c. 1922) and his results
will be found in [Ak]. They depend on the fact that integrands, f, convex
as in this chapter (§9.2), will produce integral functions, F, which are semi-
continuous in a weak sense on sets of a (Sobolev) space of functions which are

compact in the same weak sense. Then, an extended version of Proposition 5.3
may be invoked to guarantee the existence of a minimum, at least where the
integrand admits suitable estimates. Moreover, it may be possible to show
that a minimizing function is a C 1 solution of the classical Euler-Lagrange
equations of Chapter 6. In any case, a basis is obtained for systematic
approximation to the minimizing value which provides theoretical justifica-
tion for the numerical procedures of Ritz and Galerkin and their extensions.
([Mi], [Ak], [C-H], [G-F] and [We].)
Most normed linear spaces of functions used in this book are Banach
spaces, and from the amalgamation of the algebraic theory of linear struc-
tures with the standard techniques of analysis has emerged the rich and
powerful theory of functional analysis for such spaces. When its methods
are invoked, then, for example, it is possible to establish the existence of
Lagrangian multiplier "operators" for problems with Lagrangian con-
straints. This theory (of Liusternik) is presented in [E-T] and [I - T], where
also will be found corresponding results for optimal control problems, and
much, much, more. [V] should also be consulted. ,
Most of the general results alluded to above, remain valid when the ele-
mentary convexity of the integrand functions used in this book is replaced
by a more geometric convexity of the form indicated in Problem 0.7. ([B-P].)
Moreover, there is a theoretical construction which permits an "arbitrary"
integrand f to be replaced by a new integrand f** which is convex in this
extended sense. To obtain this, it is necessary to examine a related problem
for a conjugate integrand f* which acts on the dual of an appropriate real
Banach space. This theory, which is presented in [E-T], [1-T], and [R],
uses some of the deeper results in functional analysis. It also extends into the
realm of nonsmooth analysis where derivative sets are assigned pointwise to
functions that are not differentiable at these points in any reasonable sense.
See [Ro] and [CIa] for illuminating presentations.
Finally, we should mention a modern form of the homogeneous integrand
problem, that of minimization of quadratic functionals by means of Hilbert
space methods. (Hilbert spaces are Banach spaces in which the norm is
obtained from a scalar or inner product exactly as in Euclidean space.
Because of the additional geometric structure afforded by the inner product,
and the corresponding Cauchy-Schwarz inequality, these spaces exhibit
most features which might be expected from a Euclidean space of infinite
dimension.) This approach is presented in the works of Mikhlin, [Mi], and its
extension will be found in [V], and in [K-S]. Its methods are directly
applicable to multidimensional problems, and the resulting Euler-Lagrange
equations are, in general, elliptic partial differential equations. [Mo] gives
an exhaustive treatment of multidimensional problems from a modern
perspective.
Many other applications have been made of variational methods and a
representative sample of suggestive titles will be found in the bibliography. In
Problems 329
particular, [O-R] contains an extensive list of works in which variational

methods have been applied to problems in mechanics and related areas, and
[Ka-S] provides a similar list for problems in economics and management.
The foregoing remarks are merely suggestive and cannot do justice either
to the comprehensive scope of this field, or to its contributors. Despite classi-
cal origins, the principles of the variational calculus remain a vital force in
both mathematical and philosophical thought. Although the problems are
now clearly delineated, methods for their solution are far from exhausted,
and there remains a formidable gap between the known theoretical methods
and their satisfactory application in specific instances. New techniques will
be developed, and, it is to be hoped, that they will admit expression in
even simpler forms than those now employed. The subject both needs and
deserves an elegance of expression which is comparable to the idealism
embodied in its concepts.
PROBLEMS
9.1. (a) Show directly that for b > n, the function
F(y) = S: [y'(t)2 - y(t)2] dt

is not minimized on
~ = {y E C l [0, b]: y(O) = y(b) = O}

by Yo = (I). Hint: It is easy to construct a function y E ~ such that F(y) < O.
(b) For a single particle of mass m = 2, whose rectilinear motion is opposed by
a linear spring with spring constant k = 2, verify that F is Hamilton's
action integral from §8.1. What does this problem show about Hamilton's
principle?
(c) Make a limiting analysis as x ? b = n as in §9.1, to show that when b = n,
F(y) ~ 0 still holds on ~, with equality iff y(x) = ). sin x. (This proves
Dido's conjecture of Problem 1.5.)
9.2. (a) Use the Weierstrass method of §9.1 to prove that y = cx minimizes
F(y) = f J1 + y'(xf dx
on
~ = {y E Cl[O, 1]: y(O) = 0, y(l) = c}.
(b) Use Hilbert's theorem (9.7) to reach the same conclusion.
(c) What problem are we solving?
9.3. (a) Show that the stationary trajectories for
f(y, z) = ~/y, y>O,
are semicircles with centers on the x axis.
(b) Prove directly that those having the origin as a common (left) end point
determine an exact field qJ for f in a quarter plane.
(c) Use a theorem to reach the same conclusion as in (b).
9.4. (a) Examine Facts 3.11 and make appropriate vector valued extensions for
[strictly] convex functions f(!., X; Z).
(b) Prove your results as in Problems 3.2 and 3.3.
9.5. By the method of §9.1, consider the function
F(y) = J: [y'(X)2 + 2y(x)y'(x) - 16Y(X)2] dx

on
~ = {y E Cl[O, b]: y(O) = y(b) = O}.
(a) Show that Yo = (!) provides a strong minimum for F on ~, if 0 n/4.
9.6. Show that Yo(x) = sin 2x - 1 provides a strong maximum for
F(y) = f/4 [4Y(X)2 - y'(X)2 + 8y(x)] dx

on
~ = {y E C l [0, n/4]: y(O) = -1, y(n/4) = O}.
9.7. Suppose that f(!', X; Z) is [strictly] convex in a set S = D x Do where D is the

domain of an exact field (central field) for f, and Do is a convex domain in [Rd,
as in 9.2.
(a) If Yo E (Cl[a, b])d represents a field trajectory ro in D, show that Yo mini-
mizes F(Y) = S:f[Y(x)] dx [uniquely] on a certain set ~o.
(b) Prove that Yo in (a) is a [unique] weak local minimum point for F on ~o.
Hint: Use Hilbert's formula (21).
9.8. (a) If f(!., Z) is [strictly] convex on D x [Rd, use Hilbert's theorem to prove
that each stationary function Yo for f minimizes F on an appropriate ~
[uniquely]. Hint: There is a simple exact field for such f.
(b) Obtain the same result as in (a) without using the methods of this chapter.
9.9. (a) Construct an exact field qJ for the brachistochrone problem which utilizes
a family of geometrically similar "concentric" cycloids.
(b)* Define qJ(x, y) so obtained. (See §9.3.)
9.10. Verify the analysis leading to (23).
9.11. (a) Use Problem 6.42(b) to show that a central family for f(s, y, z) =
- y~ is given by I/I(s; A) = A-I sin(As) for A > 0, where the center is
at the origin.
(b) Graph three members of this family noting that 1/1'(0; A) = 1.
(c) Argue graphically that when 0 < y < s, there is a unique A with
o < AS < n for which AY = sin (AS), i.e., y = I/I(s; A). Hint: make the substi-
tution u = AS.
(d) Explain the geometrical significance of the restriction 0 < y < s.
(e) Examine I/Il and conclude that this central family determines a central
field for f in D = {(s, y) E [R2: 0 < y < s}.
Problems 331
(f) Show that f(§., y, z) is strictly convex on D x [R, and apply Hilbert's theo-
rem to make some comparisons about curves in D.
(g)* Sketch a limiting argument in terms of the integral
F,(y) = -1 h y(s)j1 - y'(sf ds,
for some Il > 0, which could lead to a proof that Yo minimizes - F on ~*

of Problem 6.42. Hint: If Yo = yo(s) defines the semicircle and P = P(s)
another curve both joining the point (/,0) to the origin and lying other-
wise in D, consider the small segments u, on which s = Il and s = 1 - Il.
What would you like to prove concerning these segments?
(h)* Show that Hilbert's integral 1 for the field
cp(s, y) = I//(s; 1)1 ;'y=sin).s

= j1 - 12y21
ly=sin).s
on each "vertical" segment u, considered in (g) reduces to
1(u.) = f p(s, y) dy = f jr(y)2 - y2 dy
for small positive a and b, where r(y) = 1- l l ly =sin l., or 1- l l ly =sin l(l-.), are
the related radii of the arcs defining the central family.
(i) Argue geometrically that as Il '" 0, 11(u.)1 --> 0, while F.(y) --> - A(y). Con-
clude that Dido's conjecture in the form A(P) :::; A(yo) with equality iff
P= Yo, is true. (See Problem 1.5.)
(j) Compare this solution with the approach taken in Problem 6.42. In what
sense are both incomplete solutions of Dido's Problem?
9.12. The Zenodoros Problem. (See Problems 1.9, 6.43, and 8.27.)
(a) Use Problem 8.27(d) to show that a central family for f(x, y, z) =
-(y(1 - Z2»1/2 is given by the parabolic functions t/J(x; 1) = x - 1x2, for
1> o.
(b) Graph three members of this family noting that t/J'(O; 1) = 1.
(c) Show by direct computation that through each point (x, y) E [R2 with
o < y < x, passes precisely one trajectory of this family generating the
field cp(x, y) = 2(y/x) - 1.
(d) Establish that this is a central field for f in D = {(x, y) E [R2: 0 < y < x}.
(e) Argue that f~, y, z) is strictly convex in D x (-1, 1), and apply Hilbert's
comparison to related curves in D.
(f) For T> 0, let ro be the graph of the function Yo(x) = x - x 2/T =
t/J(x; T- 1) on [0, TJ, and let f be the graph of a function P E ~ =
{P E C1 [0, T]: P(O) = P(T) = 0, with (x, P(x), P'(x» E D x (-1, 1).} Finally,
for a> 0, let Ua be the segment joining these curves at x = a. Show that
for proper orientation of Ua' where 1 is Hilbert's integral for cp.

(g) Prove that for a suitable function p,
II(u.)1 = IL. Pdyl ~ f~ M
= 2MIJa - p(a) - Ja - yo(a)I-+O as a"" O.

(h) Conclude that the function Yo minimizes F(P) = g f[P(x)] dx on ~.
(i)* Discuss the restriction, 0 < y < x, in terms of the original variables of
Problem 1.9, and show that this is equivalent to the requirement, np2 (x) <
2n So y(e)J1 + y'(e)2 de, which is always true. Hint: See Problem 1.7.
9.13*. In a domain D of [Rd+1, suppose that we are given arbitrary C 1 functions h =
h(x, Y) and P = P(x, Y), for which the line integral
I(r) = L (h dx + P'dY)
has a value that depends only on the end points of C 1 curves r s; D.
(a) Consider only those curves r which can be parametrized by Y E qy =
(C1[a, b])d; argue that for fixed A, B E [Rd,
b
F*(Y) =
def
f• [h(x, Y(x» + P(x, Y(x»· Y'(x)] dx
is constant on
!?# = {Y E 11,1/: Y(a) = A, Y(b) = B, (x, Y(x» ED},
and hence l5F*(Y; V) = 0, 'v'V in a set !?#o.
(b) Show that the resulting first equation for a typical Y E !?#, is
P" - hy + (Py - Py) . Y' = 0,
where the terms are evaluated at (x, Y(x», and Py is the Jacobian matrix
with rows indexed by Y, while Py is its transpose.
(c) Since (a, A), (b, B) could be any points in D with a < b, conclude that h, P
must satisfy the exactness conditions (10). Hint: Reason geometrically.
(d) Next, let III be a stationary field in D for a given f = f(x, Y, Z) (assumed C 1 ),
and suppose that Y E!?# is a solution of the associated field equation (§9.3).
Show that F(Y) = S:f[Y(x)] dx = F*(Y) given in (a), if h is defined as in
(9), where P is as yet unspecified.
(e) Use conditions (10) [obtained in (c)] to prove that when Y is as in (d) and
V(x) ~ P(x, Y(x» - fz[Y(x)], then V' (x) = - lIl y (x, Y(x» V(x) with trivial
solutions V == (I). Now define P. Hint: See Proof of 9.4.
9.14*. To obtain a multidimensional version of Problem 9.13, proceed as follows:
(d> 1)
(a) For arbitrary C 1 functions h = h(X, u), P = P(X, u), defined in a domain
D of [Rd+1 require that the integral
F*(u) ~ t (h(X, u(X» + P(X, u(X»· Vu(X» dX
be constant on sets
!?# = {u E C1(R): ul oR = i' fixed},

Problems 333
where R is a domain in [Rd, and conclude as in Problem 9.13, that this

requires V . P = hu in R.
(b)* Conversely, when R is simply connected, apply Green's theorem in the
form of equation (47) of §6.9, to show that when V' P == hu in R, then we
might expect that 1(S) <I,;! Is (- h, P)· (n, N) dO' is invariant with all those
hypersurfaces S S; aG, where G is a Green's domain in R, and the "edges"
oj S are Jixed. Here (n, N) is the outward normal to S from G. (Reason
geometrically, concentrating on the case d = 2.)
(c) When S is the graph of a function u E C1(R), show that 1(S) = F*(u). Hint:
argue that dO' = )1 + IVul 2 dX while (n, N) is proportional to (-1, Vu).
Again, concentrate on the case d = 2.
9.15. To obtain a multidimensional form of Theorem 9.7 which is valid for C 1

integrand functions J = J(X, u, Z) as in §6.9, reproduce Hilbert's reasoning as
follows: Let = (X, u) be a C2 function in a simply connected domain D of
[Rd+!, for which V' P == hu(as in Problem 9.14), when P(X, u) ~ Jz(X, u, (X, u)),
def
and h(X, u) = J(X, u, (X, u)) - P(X, u)· (X, u).
(a) Show that each solution Uo E C 2 (R) of the field equation, Vu(X) =
(X, u(X)), is stationary for J (Theorem 6.13).
(b) Suppose that this Uo is in ~ of Problem 9. 14(a). Show that F*(u o ) =
F(u o ) = SRJ[UO(X)] dX.
(c) Use the constancy of F* on ~ to prove that when also u E~:
F(u) - F(u o) = t 8(X, u(X), (X, u(X)), Vu(X)) dX,

where
def
8(X, u, Z, W) = J(X, u, W) - J(X, U, Z) - Jz(X, U, Z). (W - Z).
(d) Define an appropriate convexity for J(X, u, Z), and formulate a multi-
dimensional version of Hilbert's theorem.
(e) Solutions u of the above field equation are now difficult to produce. When
d = 2, give a condition on the field which is necessary for it to have any
C 2 solutions.
(f)* How might an exact field for J be obtained from a family of stationary
functions l/I = l/I(T; (X, u))?
9.16. (a) When J = J(x, z) E C 1 ([R2), and -r = -r(x, y) is cl, with V-r =f. f!J, show that
the family of lines which cut T, the zero level set of -r, orthogonally will
determine an exact field for J which meets the transversal condition (25) of
§9.4 in any domain D which they cover consistently.
(b) Find such a family for the parabolic function -r(x, y) = y - x 2 , and deter-
mine domains which they cover consistently.
(c) Make the construction of (b) for the function -r(x, y) = x 2 + y2 - 1.
(d) For J(z) = Z2, with -r as in (b), what related minimization problems can you
solve?
(e) Repeat part (d) when J(z) = J1+"7.
9.17. Suppose that is an exact field for J in a domain D of [Rd+!, where J(~, r. Z)
is [strictly] convex in D x [Rd.
(a) If satisfies the transversal condition (25) for a C 1 arcwise connected
transversal T S; D, which cuts one field trajectory r represented by Y E
(Cl[a, b])d twice, verify directly that
fl J[Y(x)] dx = f' J[Y(x)] dx,
where (tj , Y(t) E T, j = 1,2, and tl < t 2 • Hint: Use the invariance of Ion
the corresponding subtrajectories Ij of r.
(b) If for some fixed b, the set of points (b, Y) E D is convex, then it defines a
transversal T. Show that the transversal condition (25) reduces to requiring
that JAb, Y, lIl(b, Y» = (!J, V (b, Y) E T. How could this transversal condi-
tion have been anticipated? (See §6.4.) If III meets this condition, which
minimization problems can you solve?
(c) What condition should III satisfy on T = {(x, B): Xl :::;; X :::;; X2} in order
that Theorem 9.9 guarantee minimization for Yo, if Yo(t) = B, but
Xl<t<X2?
(d)* Prove that the cycloid provides a minimum for Jakob Bernoulli's brachis-
tochrone problem from §6.4. Hint: Use a field from Problem 9.9(a), in
conjunction with the limiting analysis for the ordinary brachistochrone
problem given in §9.4.
9.18. Verify that the functions defined by (38) satisfy (for constant f-L) the equations
of stationary for j of (36), under the constraint (35).
9.19. Chaplygin's Problem.
(a) Using the formulation with Lagrangian constraint from Problem 1.4, show
why it might be appropriate to consider the flight path of maximal area as
being stationary for the modified function
j(t, Y, Y') = -xy' + ).(t, Y)[(x' - W)2 + y'2 - 1],
where Y = (x, y), and Y' = (x', y') is regarded as a variable. ). is an un-
known function which we may suppose Cl.
(b) Demonstrate that when). > 0 on a domain D of [R3, then j(!, X, Y') is
strictly convex on D x [R2.
(c) Under the additional constraint (x' - W)2 + y'2 == 1, prove that the sta-
tionary functions for j satisfy with f-L = 2), the equations
f-L(x' - w) = - y + C2;
for trajectory constants Cl, c2 •
(d) Show that f-L2 = (x - clf + (y - C2)2, and differentiate along the trajectory
to get
f-L' = wy' so that f-L = wy + r,
for a trajectory constant r.
(e) Conclude that under the constraint in (c), each stationary trajectory must
project onto an arc of a y ellipse whose ratio of minor to major axes is the
wind-constant ~.
(In particular, an associated closed flight path through the origin would be
along the ellipse of the type in (e) fixed by the flight time and the initial flight
path angle. It is considerably more difficult to reach this conclusion with the
isoperimetric formulation of the problem. See [Sm] for details.)
(f) Explain what would be necessary to prove that flight along such an ellipti-
cal path would in fact maximize the enclosed area.
Problems 335
9.20. Isoperimetric Problem.

(a) Take the Lagrangian formulation suggested by Problem 9.19, with w = O.
Show that with the "steering angle" parametrization x'(t) = cos cx(t),
y'(t) = sin cx(t), then J1.CX'(t) = 1, where now J1. is a trajectory constant which
we may assume positive.
(b) Conclude that if Y(O) = (!), the equations of motion are given in complex
form by
v(x(t) + iy(t» = vz(t) = - ieiY(eivt - 1),
where v = 1/J1. and i' is a real trajectory constant; i = p.

(c)* Prove that each point (t, x, y) E [R3 with t > 0, (x, y) #- (!), may be joined to
the origin by a unique trajectory of this type.
(d)* Prove that the resulting central family provides a central field for this
problem. Hint: See the computation in §9.5, Example 3.
9.21. Let L\(x, a) = det 'I'A(X, a) as in Proposition 9.20.
(a) For a and ao E I, = [a - e, b + e], verify that with V(x) ~ 'I'A(X, ao) -
'I'A(X, a), then T(x) ~ (V(x), V'(x» is a solution of (60) with T(a) =
('I'A (a, ao), 'II;" (a, ao) - Id).
(b) Argue that as a -+ ao, IT(a)l-+ o.
(c) Verify that for x E I" 3 M > 0 for which IT(x) I :::;; eMI T(a)I. Hint: Proposi-
tion 9.17.
(d) Conclude that as a-+ ao, each component ofthe matrix 'I'A(X, a) approaches
the corresponding component of the matrix'llA (x, ao).
(e) Why does (d) imply that L\(x, a) -+ L\(x, ao), as a -+ ao?
9.22. The Jacobi Equation.
(a) Verify that equation (65) is the Euler-Lagrange equation for the function q
of (66), first when d = 1, and then for d > 1.
(b) With h(e) = F(Yo + BY), show by formal differentiation of h'(e) that
h"(O) = Q(V) = f q[V(x)] dx,
again first when d = 1, then for d > 1. Explain why Q(V) might be denoted
b2 F(Yo; V).
9.23. Legendre's Approach.

(a) Show that when d = 1, with the substitutions p(x) = hY[Yo(x)], s(x) =
hz[Yo(x)], r(x) = fzz[Yo(x)], equation (65) reduces to
(rv')' + (s' - p)v = o.
(b) Suppose that r > 0 and the Jacobi equation in (a) has a positive solution
VI on [a, b]. Verify that the function u = U 1 ~ -r(v~/vI) - s satisfies the
Riccati equation u' = - p + (s + u)2/r.
(c) With r, U1 as in (b), prove that for q of (66),
Q(v) = f q[v(x)] dx = f rev' + (s + u l )v/r]2(x) dx,

when v E C1[a, b], and v(a) = v(b) = o.
(d) Conclude that for such v, Q(v) > 0 unless v == O. Hint: If v'(x) = - a(x)v(x)
then v(x) = const. e-A(x), where A(x) = Sa(x) dx.
(e) Argue, (with Legendre), that ifr(x) = fzz[Yo(x)J > 0 on [a, bJ, and a solu-
tion Ul of the Riccati equation in (b) is available, then the stationary
function Yo will minimize F(y) = S:f[y(x)] dx on a typical set !?), in each
direction v.
(f) Why would the conclusion in (e) not establish Yo as a weak local minimum
point for F on !?)? What might strengthen Legendre's approach into a
proof for this minimality?
(g)* Attempt a formal extension of Legendre's argument to the case where
d = 2.
9.24. (a) For the Sturm-Liouville function /(x, y, z) == 't(X)Z2 + [q(x) - lp(x)]y2,
with q, p, 't E C[a, bJ (as in §7.3), verify that the solutions of the Jacobi
equation (65) will be stationary for j
(b)* When't > 0 on [a, bJ, relate Picard's argument of Problem 7.19 in terms
of a nonvanishing stationary (eigen)function Yl, to that of Problem 9.23(b)
in terms of a nonvanishing solution V 1 of Jacobi's equation.
(c) Explain how Picard's inequality for F of Problem 7.19(f) follows from
that of Legendre for Q in Problem 9.23(d).
9.25. Suppose that f = f(x, Y, Z) is C 2, that fzz[Yo(x)] is positive definite V x E [a, bJ

(as in 9.2), and that r o, the graph of Yo, is a trajectory of some exact field for f
Conclude that "most" of ro is the trajectory of a central field for f Hint:
Combine 9.7, 9.25, and 9.21.
9.26. For Example 1 in §9.9,

(a) Verify that when 1 = - 3, the function from (70), Y3 = 1/1 ( .; -3) Eel [0, 1J,
gives an integral F(Y3) > F(yd = mJ2.
The two parabolas defined by (70) with 1 = - 2 and 1 = - 6 intersect at
fJ = (to -ft,).
(b) Show that the corresponding integral for the Goldschmidt curve joining
(0, 1) to fJ is greater than F(Y2), (on [0, tJ), where Y2 = 1/1(.; - 2).
(c) Let Y6 = 1/1(.; -6) on [0, n Is Jacobi's condition satisfied for Y6?
9.27*. (Newton's drag profile minimizes.) In Problem 7.27 we saw that the only
candidate for minimizing F(P) on !?)* is the cornered curve Yo for which
Yo(x) = h, x ~ a, and y~(a + ) = -1. Moreover, for a ~ x ::;; 1, Yo is described
by (14') of §3.4(c), and so has a stationary extension above the line y = h
(whose graph is indicated in Figure 3.6), as well as one below the x-axis. It is
given parametrically in Problem 3.40(b) where c = a and Yl = h.
(a) When Yl = h, the family of curves for c> 0 in Problem 3.39(b) cover
consistently a simply-connected domain D :2 {O < x ::;; 1,0::;; Y ::;; h}. Ex-
plain, and graph several curves from this family.
(b) Explain why the slopes ({J of these curves define a field in D that is station-
ary, hence exact. Also explain why ({J(x, h) = -1 and why ({J ::;; 0 in D.
Conclude that when Y = h, the field functions in (9) (here denoted by p and
h) are, respectively, x/2 and x.
(c) Let P E ~*, and suppose that like Yo, P(x) = h, for 0 < x ~ c say. Let a be
r,
the segment from (c, h) to (a, h) parametrized in this direction), the graph
of Pfor x ;::: c, and r o, the graph of Yo for x ;::: a. Then show that l(r) =
l(ro) + l(a), where I is Hilbert's invariant integral for ({J. Use previous
Problems 337
information to conclude that
1(u) = r
x dx = s: x dx - f: x dx
so that as in (21), F(P) ~ F(yo) as desired. Hint: Recall that e ~ 0 for the
appropriate pi, cp.
(d) If PE g)* is not constant on some initial x-interval, we can modify the
previous construction by introducing the vertical segment Ue from (c, y(c))
to (c, h) for small c < a and letting t and re be the graphs of y and Yo
respectively for x ~ c. Show that 1(re) = 1We) + 1(ue) and that
fh 2clcp(c, y)1 lui 1
11(uJI :;:; J0 (1 + cp2(C, y))2 dy :;:; ch (since 1 + u2 :;:; 2: when u E R).
Let c ...,. 0 and conclude that as before F(P) ~ F(yo).
PART THREE
OPTIMAL CONTROL
AN OBSERVATION
"Since the fabric of the universe is most perfect, and is the work of a most
wise Creator, nothing whatsoever takes place in the universe in which some
form of maximum and minimum does not appear."
LEOHARD EULER, 1744
339
CHAPTER 10*
Control Problems and Sufficiency

Considerations
The discipline now identified as optimal control emerged during the decade
1940-1950, from the efforts by engineers to design electromechanical appara-
tus which was efficiently self-correcting, relative to some targeted objective.
Such efficiency is clearly desirable in, say, the tracking of an aircraft near a
busy airport or in the consumption of its fuel, and other economically desir-
able objectives suggest themselves. The underlying mathematical problems
were attacked systematically in the next decade by Bellman [Be], by
Hestenes [He], and by a Russian group under Pontjragin [Po]. Their results
were quickly adapted to characterize optimal processes in other fields
(including economics itself) and the feasibility of optimal control is now a
standard consideration in contemporary design strategy.
In examining the associated idealized problems, it is natural to employ
the techniques of the variational calculus to obtain models for what can
occur. We have already attacked two such problems successfully by such
methods-the production problem of §3.4(d), and the fuel consumption prob-
lem in §3.5. Indeed, most (deterministic) problems in optimal control admit
formulation as one of steering a system so as to minimize a performance
integral over an interval (in time), in the presence of Lagrangian constraints,
with certain additional target conditions and control restrictions (§10.1). (We
shall not consider problems involving multidimensional integrals.) Control
constraints are usually a reflection of physical limitations, and although
their presence imposes severe complications on the theoretical derivation of
necessary conditions (§11.1), it seems far less inimical to sufficiency consider-
ations. (We avoid attempts at presenting general existence theory which for
optimal control problems is truly formidable; see [Ce] and [CIa]). In this
chapter we concentrate on developing effective sufficiency methods which
can usually be attempted, and which, when successful, wi11lead to a solu-
tion-in many cases-the unique solution to the problem.
343
342 10*. Control Problems and Sufficiency Considerations
We demonstrate (in §1O.2) that representative problems can be solved by

reformulation utilizing elementary convexity as explored in this book (espe-
cially in Chapters 3 and 7). In §10.3 we develop an underlying sufficiency
theory for a class of such problems (on a fixed interval) expressed in terms of
the Pontrjagin function h which arises naturally in this context. The theory is
applied to provide a complete solution to an important linear state-quadratic
performance problem without control constraints. When subsequently ad-
mitting such constraints in §1O.4, we encounter the fundamental control
minimization principle which characterizes h along the optimal trajectory; we
show that it too may provide sufficiency criteria, even for a nonconvex
performance integrand.
The material in this chapter and the next is essentially independent of that
in Chapters 8 and 9, but it demands the setting from Chapter 7.
§10.1. Mathematical Formulation and Terminology

For our purposes, a control problem arises whenever the state of a system at
time t, as described by a vector Y E IIl d, evolves according to a prescribed law
given usually in the form of a first-order vector differential equation,
Y'(t) = G(t, Y(t), O(t)), (1)
a
under assignment of a (vector valued) control function with O(t) E IIlk. Gis
given and assumed C 1 and Y is C1 (as in Chapter 7), but a
may be only
piecewise continuous (C) on each time interval. Controls which have the
same values except at common points of discontinuity will be considered
identical.
This becomes a problem in optimal control when we wish to find those
controls 0 0 which produce states Yo that optimize a performance criterion
r
assessed by an integral of the form
F(Y, 0) = f(t, Y(t), O(t)) dt, (2)
where f is a given real valued function (usually supposed C 1 ), and a and b

might be permitted to vary to allow the state vectors Y to meet various
initial/terminal conditions, Y(a), Y(b). In addition, the controls or state vec-
tors might be subjected to other restrictions, such as 10(t)1 ::;; 1, or U(t) E 0/1,
or'p > 0, etc. In general, we will assume that the initial state is prescribed, but
allow various terminal conditions, and replace [a, b] by [0, T], where the
target time T may be permitted to vary.
The problem is said to be autonomous when both the state function G
and the performance function f (as well as any other interval constraints) do
not depend explicitly on t. It is linear when both f and G are linear in Yand
U, and it is time-optimal when the performance function f == 1 on [0, T].
§1O.1. Mathematical Formulation and Terminology 343
Observe that system dynamics involving higher-order state derivatives

can in principle be replaced by an equivalent first-order law of evolution of
the form (1) but with a higher state dimension. However, the special character
of the admissible controls 0 make it inadvisable in general to regard the pair
cy. 0) as a larger "state" vector, especially since Y' is present explicitly, while
0' is probably nonexistent.
Nor is it appropriate in general to integrate the state equation incor-
porating state boundary conditions and thereby convert a Lagrangian con-
straint into an isoperimetric one, since the solution to the resulting problem,
even when available, might not satisfy the Lagrangian constraints.
The reduction of problems in a given discipline to the form considered
above is a matter which usually requires significant insight. However, what
we wish to stress in this chapter is that an important key to obtaining
solutions to each problem is suitable reformulation. In particular, as we
will show by representative examples, reformulation emphasizing aspects of
convexity to determine sufficient conditions may again provide access to
complete solutions.
By contrast, because of the restrictions on the controls, it is far more

difficult to characterize necessary conditions for these same problems, since
the optimal controls usually lie (at least partially) on the boundary of the
control region, rlIi. Those controls which lie entirely on the boundary ex-
tremes are called bang-bang controls after the model of a light switch which
is effective only when fully "on" or "off." Some rather deep theorems provide
conditions under which an optimal control must be bang-bang-or; at least
can be replaced by a bang-bang control which is also optimal. We shall
obtain a simple result of this type in Corollary 11.14. For the more general
case, see [M-S], [Ne].
In order to simplify the presentation, the dimension of the function spaces
as well as carats indjcating piecewise continuity, differentiability, etc., will
be suppressed when possible. Thus Y E (C 1 [0, T])d will be replaced by Y E
C1 (0, T], and differentiation with respect to t will be denoted by a dot.
Finally, differential equations such as (1) will be understood to hold only
when t is a point of continuity of U.
To better appreciate the significance of convexity in this exposition, the
following observations may be of benefit:
On a fixed interval [0, T], the typical problem in optimal control is seen
to be that of minimizing
F(Y, U) = IT f(t, Y(t), U(t)) dt
on a set ~ incorporating initial and terminal conditions (usually on Yalone),

subject to Lagrangian constraints of two forms:
(a) the state laws symbolized by [Y, U] = 0;
(b) the state-control restrictions symbolized by (Y, U) ~ 0;
where for each (Y, U) E!!J, each of the expressions [ ] and < ) is an
(integrable) real valued function of t on [0, T].
Now, if for some (continuous) Lagrangian multiplier functions p and Ji,
(Yo, Uo) E !!J minimizes
F(Y, U) = F(Y, U) + LT (p(t)[Y, U] + Ji(t)<Y, U») dt
on !!J [uniquely], it follows that (Yo, Uo) will minimize F(Y, U) on !!J
[uniquely] under (a) and (b) provided that for t E [0, T]:
(a') [Yo, Uo] = 0; and
(3)
(b') Ji> 0 with Ji<Yo, Uo ) = 0;
(except possibly at a finite set of values of t).
Indeed, then <Y, U) :$; 0 => Ji< Y, U) :$; 0 = Ji< Yo, Uo ), so that
F(Y, U) :?: F(Y, U) :?: F(Yo, Uo) = F(Yo, Uo).
For the vector valued versions of (a) and (b) we simply add additional
terms to j, each with its own multiplier function as in §2.3, resulting in a new
integrand of the form
j=j+p.[] +M·< ),
which can then be subjected to a similar analysis.
Of course, if the new integrand j is convex in the sense of this book, then
minimization of F can, in general, be obtained from a (Yo, Uo) E !!J which
satisfies the Euler-Lagrange equations for j together with the corresponding
Weierstrass-Erdmann conditions of §7.5 at any corner points.
Also, sufficient strong convexity will guarantee uniqueness of the minimi-
zation, and this requires only convexity of each term in j plus strong convex-
ity of one of these terms. Moreover, as in §7.4, these arguments remain valid
even when the terms of j[Y(t), U(t)] are only piecewise continuous, as can be
seen by the usual partitioning of the integrals.
When the target time T is not fixed, it may be possible to solve the
problem as if it were, and then optimize over T as was done with the perfor-
mance problem in §3.5. If this is not possible-and it cannot be for time
optimal problems in which it is this T itself being minimized-then there
may be a transformation which replaces the problem by a (convex) one
over a fixed interval in some other independent variable. If this fails, then
there are certain other sufficiency theorems including some of the field theory
type of Chapter 9, but they are usually more difficult to implement.
§10.2. Sample Problems

It is difficult to identify a canonical set of problems in optimal control
because of the relatively recent origin of this field. Moreover, only a few of
those which have been formulated yield satisfactory explicit solutions, and
§1O.2. Sample Problems 345
those presented here may be more representative of their solvability than

of the entire field. Nevertheless, their solutions do exhibit some of the charac-
teristics to be expected, principally in the discontinuities (with or without
bang-bang behavior) of the optimal controls, and in the changes in character
of the optimal solutions with the relative size or "geometry" of the given
conditions. The latter behavior has already been encountered in investigating
the minimal surface of revolution (§1.4(a) and §7.5) but its presence clearly
complicates both the analysis and the presentation of results.
Whenever possible, we solve these problems by methods that involve
preliminary skirmishing, with direct attack once convexity seems at hand, as
opposed to the usual approach of forcing the problem into the procrustean
format of general control theory. In some of these problems, the optimal
control is continuous, but in the Bolza problem of part (b), it is neither
continuous nor bang-bang.
Problems in parts (a), (b), and (c) are presented without considering possi-
ble applications, but the remaining problems do have physical origins. The
first extends the rocket analysis of §3.5 to a corresponding problem for which
the optimal control is bang-bang. The next (whose statement is taken from
[Ber]) shows that both continuous and bang-bang controls can be optimal
for different time intervals. The last, that of an oscillator is selected because
various natural questions concerning its optimal behavior(s) require different
methods for satisfactory answers. The time-optimal problem for both it and
the simpler nonoscillatory "docking" case (discussed in §11.2) are to be found
in most expositions of optimal control.
The final problem utilizes a preliminary minimum principle instead of
convexity to obtain the optimal time of transfer for certain simple state
equations including those for a spinning asymetric body in space.
(a) Some Easy Problems

(
Suppose that we wish to minimize

F(y, u) = Il y2(t) dt
on
f!) = {(y, u) E ClEO, T] x CEO, 1], with y(O) = 0, y(1) = 1},
subject to y(t) = u(t), with u(t) ~ 1.
We note that these last two conditions can be replaced by the single
Lagrangian inequality,
y(t) - 1 ~ 0 (4)
(thereby eliminating u from the problem); thus we examine the modified

integrand
(considering y as a variable), which for any continuous j1., is strongly convex

on [0, 1J x 1R2. (Why?) It follows that each Yo in

r!}* = {y E Cl [0, 1J: y(o) = 0, y(1) = 1}
which satisfies the Euler-Lagrange equation
d ~ ~
d/ = h
Y or fJ, = 2y on (0,1)
with
I1(Y - 1) = 0,
H
will minimize F*(y) = y2(t) dt on r!}* under (4).
By inspection, yo(t) = t satisfies all conditions if we take l1(t) = 1 + t 2
(since fJ,(t) = 2Yo(t) = 2t). Thus, this Yo, with Uo = Yo = 1, provides the unique
minimum for this problem.
It also provides the unique maximum since in fact for this problem, if
y E r!}* and y :::;; 1, then
1 = y(1) = Ll y(t) dt:::;; Ll 1 dt = 1,
where equality is possible only if y(t) == 1 or y(t) = t = yo(t). Therefore Yo is

the only competing function, and naturally it excels in all respects. This
solution is independent of the optimization question being posed, and in
H
particular, it is independent ofthe integral y2(t) dt; this fact will have more
significance in our discussion of Theorem 11.6.
Note also that the above simple device of setting Y = U, transforms most
problems in variational calculus considered in the previous chapters into
those which have the appearance of problems in optimal control. This
supplies a convenient source of counterexamples.
For example, it is easily verified by the techniques of Chapter 3, that
the convex function
is minimized uniquely on
r!} = {y E Cl[O, 1]: y(O) = 1, y(1) = e}
by
Hence
uo(t) = yo(t) = et
provides the unique optimal control for minimizing
F(y, u) = Ll [y2(t) + u2(t)J dt on r!} x C[O, 1J
under the state law y = u, and, if desired, the control restriction

/u(t)/ :::;; 4.
§10.2. Sample Problems 347
(This is a simple example of the linear state-quadratic performance problem

to be investigated in §10.3.)
Moreover, some related problems can be solved by brute force once the
state equation is used properly. For example, the minimization of
F(y, u) = Ll (u 2 - y3) dt when u E CEO, 1], -1 :$; u(t) :$; 0,

on
fl) = C1 [0, 1]: y(O) = y(1) = 1, y(t) ~ Op
{y E
under the state equation y = yu, appears to be difficult primarily because of

the - y3 term. However, if we integrate this term by parts, we see that when
y E fl):
F(y, u) = Ll (u 2 + 3(t - 1)y2y) dt - (t - 1)y3(t{
= Ll (u 2 + 3(t - 1)y3 U) dt - 1.
Now, when y ~ 0 ~ u, the new integrand is ~ 0 with equality when u(t) =

uo(t) = 0, so that yo(t) = 0 or yo(t) = yo(1) = 1. (This problem resembles a
more significant one of Zeidan that yields to the same attack. See [K-P] and
related works. However, it also resembles the Wirtinger problem of §9.5
which does not give itself up so easily!)
(b) A Bolza Problem
The methods employing convexity may extend also to a problem such as the
following of Bolza type in which the performance integral is augmented by a
function of the endpoint values.
Minimize F(y) = y2(2) + L2 y2(t) dt
on fl) = {y E C 1 [0, 2]: y(O) = 1}

under the state law y(t) = u(t) with the control restriction lu(t)1 :$; 1.
Here it is simplest to replace the last two requirements by y(t)2 - 1 :$; 0,
and consider the modified function
F(y) = F(y) + L2 ll(t)(y(t)2 - 1) dt,
1 The requirement that y(t) ~ 0 is superfluous for this problem. From the state equation, we see
that y ~ - y when u ~ -1, on any interval [0, t] in which y ~ O. But then y(t) ~ y(O)e-', so this
interval cannot terminate.
which is convex on !!) when Jl ~ O. (Why?) We want y = Yo E !!) such that
t5F(Y; v) = 2y(2)v(2) + I2 [2y(t)v(t) + 2Jl(t)Y(t)v(t)J dt

= 0, 'if v E!!)o = {v E C1 [0, 2J, v(O) = O},
with Jl(t)(Yo(tf - 1) == O. (Why?)
Integrating by parts, incorporating the fact that v(O) = 0, we obtain
tt5F(y; v) = [y(2) + Jl(2)Y(2)Jv(2) + I2 [y(t) - (Jly)"]v(t) dt
Thus we want JlY to be continuous with (JlY)" = y except at corner points, and
JlC? - 1) == 0, with y(O) = 1 and y(2) + Jl(2)Y(2) = o.
Suppose Jl(O) > 0, so tht either y(O) = + 1 or y(O) = -1. In the former case
JlY increases (why?) so that Jl can never vanish and y(t) = + 1; therefore,
y(t) = t + 1, but this violates the terminal condition. In the latter case,
y(t) = -1, and both y(t) = 1 - t and Jl(t) = (1 - t)2/2 + c decrease until Jl
vanishes and permits a corner point at t = t 1 , say. For example, if c = 0,
then tl = 1, y(t 1 ) = 0, and we could take Jl(t) = y(t) = 0, for t > 1, since this
satisfies all requirements. Thus, one possible solution is to take
yo(t) = 1 - t and Jl(t) = (1 - t)2/2, t ~ 1,
=0 = 0, t ~ 1,
but are these the only choices? Observe that
F(yo + v) - F(yo) - t5F(Yo; v) = v 2(2) + I2 [v 2(t) + Jl(t)V(t)2J dt
=0 iff v == O.
Thus F(yo + v) = F(yo) iff v == 0, so that the solution Yo is unique, and
_ . _ {-1 on (0, 1),
Uo - Yo - 0 on (1, 2),
is the unique optimal control.
Here the optimal control uo, although discontinuous, is not of the bang-
bang type, since U o = 0 is not on the boundary of the control region dfI =
[ -1, 1]. (However, it would be of this type on the smaller control region
dfI = [0, 1].)
(c) Optimal Time of Transit

The state X(t) = (x(t), Xl (t)) of a system at time t is governed by the linear
equation
X(t) = (1 + u(t), U 1 (t)),
under the control U(t) = (u(t), U 1 (t)) with the constraint IU(t)1 = 1.
In order to find the minimum time T required to transfer this system from
its initial state at the origin (X(O) = (9) to a prescribed state (X(T) = B =
(b, bd), we can transform this problem to one on a fixed interval as follows:
Under admissible motions we have
1= u2 + uf = (x - 1)2 + xf,
or 2x = x2 + xf ~ 0, which implies that since x(O) = 0, x(T) = b,
b= LT x(t) dt ~ 0, (5)
°
and only such b can be permitted. Moreover, b = => x == on (0, T), (why?),
so that Xl == 0, as well. Therefore bl = xl(T) = Xl(O) = 0, and the problem is
°
trivial.
Thus for b > 0, we shall assume that x > 0, and consider x as the new
independent variable, while Y = t(x) and Yl (x) = Xl (t(x» will be the new state
variables, governed by the new state equation
Y? = 2y' - 1 (6)
on the fixed interval [0, b]. (Why?) Since
{
xy' = 1,
y~ = x1 Y',
we shall permit discontinuities in y~ provisionally. Then we have the simple
convex problem of minimizing
Jor y'(x) dx = 2"1 Jor (y? + 1) dx = F(Yl)

b b
T= (7)
on
~l
q) = { Yl E C [0, b]: Yl(O) = 0, Yl(b) = bd,
where the Lagrangian constraint (6) can now be considered after Yl has been
determined. By the methods of §3.3, it is easy to show that the unique
minimum is
Jor b2 + b2
b
To = F(Yo) = y~(x) dx = Yo(b) = l2b 2 = p, say.
It follows that Pis the desired minimum time To required to reach state B
under the mild restriction that x > 0; it is obtained uniquely with the linear
trajectory Xo(t) = (tiP, bltlbP)·
Remarks. In solving this problem, we were fortunate in that it is convex

when expressed in terms of new variables on a fixed interval. Although
it is usually possible to reduce a time optimal problem (or other problem
involving a varying time interval) to one for a fixed interval (see the proof
of Theorem 11.10), the resulting integrand is seldom convex. Some special
forms of time-optimal problems will be discussed in §11.2.
(d) A Rocket Propulsion Problem
In §3.5 we examined a problem of minimum fuel consumption for a vertically

rising rocket, neglecting dynamic effects of the fuel mass loss in so doing. A
seemingly better model for this same rocket obeys the dynamic law
my = -prh - mg, (8)
where p is a given positive constant, and m = m(t) is the mass of the rocket
and fuel at time t. Then -rh, the rate of fuel consumption, is nonnegative,
controllable, and limited by the design of the engine; its effect in providing
thrust is clearly visible in (8).
Since m(t) ~ mR , the mass of the rocket alone, we can divide (8) by m and
obtain
y= u - g, (9)
where u = - prh/m ~ 0, is a thrust control with, say, u(t) ::;; p; g is the gravita-
tional acceleration which we still assume constant. Note that (9) is identical
to the dynamic law used in the earlier analysis, but our new model admits
more realistic applications.
For example, if the initial mass of the fueled rocket is M o, what is the
maximum altitude which can be reached by this rocket with the consumption
of a fixed mass M of fuel during the first stage of ascent, and how should the
fuel be burned to achieve it?
For each program m(t) of fuel consumption over the fixed time interval
[0, T], with m(T) = Mo - M, there is a control u E CEO, T], with
IT u(t) dt = p 10g(Mo/(Mo - M)) (Why?)

(lOa)
= k > 0, say,
Under this isoperimetric constraint, we wish to find that control u which, by
(8), (9), and (lOa) maximizes
y(T) = IT (T - t)y(t) dt
= IT (T - t)(u(t) - g) dt = IT (T - t)u(t) dt - gT2/2, (lOb)

on
r!}T = [u E CEO, T]: 0 ::;; u(t) ::;; p}.
To use convexity, we consider instead the problem of finding that Uo E r!}T
which under (9) and (lOa) minimizes
F(u) = IT (t - T)u(t) dt.

If we ignore the control constraints, 0 ::;; u(t) ::;; p, and examine for con-
stant A, the modified integrand
j(!, u) = (! - T)u + AU
(which is convex), for that U o which makes 0 = .!..[uo(t)] = t - T + A, we see
that this is not feasible. (Why?) Hence the inequality constraints on u appear
to be an essential feature of this problem.
To take them into account most effectively observe that the pair of
inequalities 0::;; u ::;; p is equivalent to the single quadratic inequality
u(u - P) ::;; 0, since the possibility of u < 0 < u - P is untenable.
Then according to the approach taken in §10.1, we should consider the
modified integrand
j(!, u) = (! - T + A)U + /1(t)(u 2 - pu)
which is strongly convex on [0, T] x ~ when /1 E CEO, T] is positive.
If we can find a A, /1, and U o E C[0, T] such that
o = .!..[uo(t)] = (t - T + A) + /1(t)(2u o(t) - P), (11 a)
while
/1(t)u o(t) [uo(t) - P] = 0 with /1 > 0, (l1b)
and (lOa) is satisfied, we have a unique solution. (Why?) Observe that by
(lla), /1(t) = 0 only when t = ,~f T - A for A < T, so that for any other t,
(l1b) requires uo(t) = 0 or p. But then /1 > 0 in (l1a) requires that
Uo
(t)-_{P, t E (0, ,),
0, t E (" T).
Finally, it remains to select A, if possible, to make
k= IT uo(t) dt = t p dt = p, = P(T - A);
since this simple equation has the unique solution

A = T - k/ P< T,
we are assured that U o is optimal for the problem. Moreover, although
/1(') = 0, /1 is otherwise positive on (0, T) so that by subdivision of the inte-
grals, uo, is unique. Note that it is discontinuous and switches in value from
p to 0 at , = kiP.
The corresponding maximum altitude at time T is from (lOb) given by
Ymax(T) = IT (T - t)(uo(t) - g) dt
=P t (T - t) dt - gT2/2
= P(T't' - ,2/2) - gT2/2.

Finally, if we maximize this parabolic expression in T, we obtain the absolute

maximum amplitude of
Ymax = (P - 2g)k 2 /2pg
attained at time T = To = k/g.
Observe that although U o = Pis constant during the burn interval (0, 't"),
rho, the optimal rate of fuel consumption, is not. Instead, from (9), follows
(d/dt) log mo = - Pip, so that on (0, 't") = (0, P/k),
mo(t) = Moe- flt / p
or
rho(t) = -(p/p)Moe- fit / p •
Note that here, the optimal control Uo must be bang-bang for the physi-
cal control region o/L = [0, P]. If formulated as a control problem this yields
an F(y, u) which is (only) convex and in fact u can be arbitrary on ['t", T],
resulting in a singular control problem [B-J].
(e) A Resource Allocation Problem

At time t, a fraction u(t) of a quantity being produced at a rate y(t) is allocated
for investment to improve productive capacity, the rest being sold for profit.
It is desired to choose an investment fraction U o and a production rate Yo to
maximize the total profit over a fixed time interval [0, T].
At time t, goods for profit are produced at the rate (1 - u(t»y(t) (why?)
and thus we wish to minimize
F(y, u) = - IT (1 - u(t»y(t) dt
on
~ = {(y, u) E Cl [0, T] x CEO, T]: y(O) = a l ~ 0, y(t) ~ O},
subject to the control restriction
The choice
°: ;
u(t) ::;; 1, and an investment constraint.
'"
y(t) = u(t)y(t),
allows the production rate to increase directly with the amount available for
investment.
This is clearly of the form which identifies it as a problem in optimal
control. However, the product u(t)y(t) which appears in the integrand makes
convexity arguments difficult. Fortunately, in this case, we can use the con-
straint to replace the product.
Moreover, since y ~ 0, we can replace the control restriction
by the Lagrangian inequalities °: ;
y(t) ::;; y(t), and we obtain the simpler
°: ;
u ::;; 1
problem of minimizing
F(y) = IT [y(t) - y(t)] dt
on
q) = {y E Cl[O, T]: y(O) = al, Y ~ O}
under the pair of Lagrangian inequalities
y(t) - y(t) ~ 0, - y(t) ~ 0
expressed in the most useful form.
According to our general analysis in §7.4, we should introduce Lagrangian
multiplier functions, A(t), p,(t), and try instead to minimize the modified
f:
integral
F(Y) ~ F(y) + [A(t)(y(t) - y(t)) - p,(t)y(t)] dt
= LT [(1 + A(t) - p,(t))y(t) - (1 + A(t))y(t)] dt.

Now, the associated integrand
j(!, y, z) = (1 + AW - p,W)z - (1 + AW)y
is convex, and if we can find Yo E q), A ~ 0, p, ~ 0, for which Yo minimizes F
on q) with
A(t)(YO(t) - yo(t)) == - p,(t)Yo(t) == 0,
then we will have a solution to the problem. For minimization, we require
only the Euler-Lagrange equation
(12)
with continuity of 1 + A - p, at the corner points, together with the natural

boundary condition i.,[yo(T)] = (1 + A - p,)(T) = O.
At T: p,(T) - 1 = A(T), and if A ~ 0,
then: p,(T) ~ 1; hence near T, we must have yo(t) == 0 to obtain
p,(t)Yo(t) == 0;
but then yo(t) = Co near T, so that A(t)(Yo(t) - yo(t)) == 0, with Co i= 0, implies
that A(t) == 0 near T, and p,(T) = 1. Thus, near T, the Euler-Lagrange equa-
tion (12) reduces to - {L = -1, so that p,(t) = 1 - (T - t), in an interval
which terminates at a switching time 't' < T. It is possible that 't' ~ O. Other-
wise, for 't' < T, we expect A i= 0, so that Yo = Yo, and thus yo(t) = coet, where
we can take Co = yo(O) = al, if all ofthe other conditions can be met.
Since Yo i= 0, it follows that p, == 0, and (12) again simplifies to
(1 + A)' = - (1 + A),
with the exponential solution
1 + A(t) = ce- t ,
where the corner condition gives ce- t = T - 't' (why?), or c = (T - 't')e t •
Requiring continuity of p, at 't' (and hence of A) gives 0 = 1 - (T - 't') provided
354 10*. Contrpl Problems and Sufficiency Considerations.
that the real switching time. = T - 1, when T ~ 1. Finally, when T:::;; 1,

we take Jl(t) = 1 - (T - t), and A(t) = 0, which gives an optimal solution,
yo(t) = a 1 •
When T> 1, there is a single switching time. = T - 1, and

a1 et, t:::;;.,
Yo (t) = {
ale" .:::;;t:::;;T,
provides an optimal solution since the auxiliary functions
O, t:::;;.,
{
Jl(t) = 1 - (T - t), .:::;; t :::;; T,
A(t) = {-1 +
0,
(T - .)e t - t
, t:::;;.,
• :::;; t :::;; T,
are nonnegative and continuous as required. Observe that in this latter case,
the optimal control function is bang-bang;
_ Yo(t) _
()-
uot ---
t<" {1,
yo(t) 0,. < t :::;; T,
is discontinuous with values entirely on the boundary extremes of the control
region o/J = [0, 1]. It dictates that initially all of the output should be used
to improve plant production capability, after which all material produced
should be sold for profit. In a real situation, the work force might object to
this "ideal" solution when T is large!
It remains to consider uniqueness of the optimal solution. What we have
actually established is that if y E (51 [0, T], with y(O) = Yo(O) = a 1 , then with
v = y - Yo:
rT ~
F(y) - F(Yo) = J
0 [(1 + A- Jl)v - (1 + A) v] (t) dt
= [(1 + A - Jl)v] (t) [ = o. (Why?)
Hence, F(y) = F(yo) = F(yo), and under the required inequalities 0 :::;; Y :::;; y,
since A and Jl are continuous and nonnegative, it follows that
F(y) = F(yo) iff A(Y - y)(t) == - Jly(t) == O.
However, when T:::;; 1, Jl =I- 0, so that Y == 0 or y(t) = y(O) = yo(t); and simi-
larly for T > 1, we conclude that
{ ~y=O,
= y, 0:::;; t :::;; "
.:::;;t:::;;T,
which again means that y(t) == yo(t). (Why?)
Thus we have obtained the unique solution Yo, to the given problem and
hence the unique optimal control Uo. A related application which arises in
the fishing industry is explored in Problem 10.29.
(f) Excitation of an Oscillator

If a particle at rest at the origin is forced into rectilinear motion whose
position y at time t is governed by the equation
}i+y=u, (13)
where the forcing term u E C is bounded, say, lu(t)1 ::;; 1, then the resulting
motion is in general oscillatory in character.
Indeed by variation of parameters [C-L], the motion with
y(O) = y(O) = 0
is given uniquely by
y(t) = I sin(t - 1:)U(1:) d1:, (14)
as can be verified by using Leibniz' formula (A. 14) to differentiate the integral
at each point of continuity of u.
In particular, if u == + 1, then the substitution 81 = t - 1: gives
y(t) = I sin 81 d8 1 = 1 - cos t,
and the oscillatory behavior is evident.

However, if at some time (J', we switch from u = + 1 to u = -1, then the
subsequent oscillations will be about y = -1, but the amplitude will vary.
For example, (J' = n produces for t ~ n, y(t) = -1 - 3 cos t; but then, we
can switch again to get a still larger amplitude, etc.
Thus we might ask for the largest y(T) which can be obtained at a given
time T under the above restrictions, and which control(s) u will be optimal in
producing it. But from (14), we see that
y(T)::;; LT Isin(T - 1:)1 d1: = yo(T),
and Yo(T) is obtained uniquely under the optimal control uo(t) =

sgn sin(T - t).
For example, when T = 2n, Yo(2n) = 4, and it is obtained solely from
the control
-1 on (0, n),
Uo =
{ +1
on (n, 2n).
However, if we consider instead the equally reasonable problem of maxi-
mizing at time T, the energy measured by E ~ (y2 + y2 )/2, then although
under (13),
E(T) = IT E(t) dt = IT (y + ji)Y dt = IT yu dt, (15)
we cannot use the above approach as successfully since Iyul is not as easy to
analyze. From (14) follows
y(t) = J: cos(t - 't')u('t') d't', (16)
and since E = yu (why?), we can obtain after some manipulation left to

Problem 10.6 the useful formula
2E(T) = (IT u(t) cos t dtY + (IT u(t) sin t dt y. (17)
Moreover, for t ~ T ~ n/2, both sin t and cos t are nonnegative, and we
see that when lui ~ 1,
E(T) ~ t[ (IT cos t dtY + (IT sin t dtY] = 1 - cos T = Eo(T)
and thus maximal energy Eo(T) is achievable with either of the controls
Uo == + 1 or U o == -1. When T > n/2, this maximal energy problem is signifi-
cantly more difficult than its predecessors. (A related time-optimal problem is
solved completely in [Y].) We shall return to it in §11.1 after we have de-
veloped a more general approach for attacking it.
For the present, we shall convert it into the standard form by introducing
the new variable
y = Yl'
so that (18)
Yl = U - y.
Thus, with Y = (y, yd, we wish to minimize
F(Y, u) = - IT Y1u dt
on
!!) = {(Y, u) E c [0, T]
1 x CEO, T], with Y(O) = (f) and lu(t)1 ~ I}
under the new linear state equation (18), which takes the matrix form
y=[_~ ~]Y+ul~l·
We see that this problem is autonomous, and that the terminal state Y(T)
is unspecified.
(g) Time-Optimal Solutions by Steepest Descent

Suppose k = d and the state equation has the simple form
Y(t) = I/I(t, Y(t)) + U, (19)
with the control restriction I UI ~ 1. Then it may be possible to solve the
time-optimal problem for which Y(O) = A is given and Y(T) = (!) is required,
by arguing that we would like to force IY(t) I to descend from its value
IAI ¥- 0 to 0 as quickly as possible.
Since except at corner points, for Y(t) ¥- (!),
d 1d 2 •
IYldtlYI =2dt(IYI )= Y·¥,
we obtain from (19) , that with V = Y/I YI,

d
dt IYI = V· I/I(t, Y) + V· U
~ V·I/I(t, Y) - 1 iflUI ~ 1, (20)
with equality precisely when, U = - Y/I YI, if we employ the Cauchy inequal-
ity (§0.1) to infer that V· U ~ -IVII UI ~ -1, when IUI ~ 1 for the unit
vector V = Y/I YI.
Therefore, if we could solve the state equation (19) with U = - Y/I YI, and
meet the appropriate boundary conditions at some To, then the solution Yo(t)
would be the (unique) optimal trajectory, To would be the minimal time of
transfer, and Uo(t) = - Yo (t)/I Yo(t)l, t < To, would supply the (unique) opti-
mal control.
However, if we could just solve
d Y
dt IYI = wi"
I/I(t, Y) - 1, (21)
with IY(O) I = IAI, IY(To)1 = 0, then by the same argument, To will at least
be the (unique) minimum time of transfer. Now we can surely do this if, say,
Y·I/I(t, Y) = 0 (22)
since then (21) is just (d/dt) IYI = -1 with the trivial solution
IY(t)1 = IY(O) I - t = IAI - t so that To = IAI. (23)
Moreover, the optimal control Uo(t) is that which opposes the (unit) state
vector Y(t)/I Y(t) I at each subsequent instant. (This provides an illustration of
control synthesis: a state dependent prescription for the optimal control at
each instant.)
Condition (22) is realized in the problem where a spinning fully asymme-
trical body (say, a satellite) in space is to be brought to (spin) rest by the
application of a (vector) torque U.
Figure 10.1
Indeed, if the body has distinct moments of inertia Ij about three body
fixed principal axes through its center of mass as illustrated in Figure 10.1,
then the associated angular momentum vector Y = (Yl' Y2' Y3) is governed
by Euler's equations. (See [eel) The first requires gyroscopic coupling due to
the assumed asymmetry 12 "# 13 ; it is
and the rest are obtained by cyclic permutation of the indices (1, 2, 3).
Since the inertial coefficients are constant, it is easy to see that (22) requires
only
and the sum of the first two terms does cancel the last.
It follows from (23) that if all control torques lUI::;; 1 are available, then
the minimal time of transfer from a spin state A is To = IAI, and it is given by
that control Uo(t) = - Y(t)/I Y(t)1 which opposes the spin state direction at
each subsequent instant. For further discussion of this problem when one
axis of symmetry (12 = 13 ) is allowed, see [A- Fl
In addition to (22), other conditions may lead to a solution by this method
in related problems where it is desired to drive a system from a given state A
to the origin. Some of these are explored in Problems 10.7-10.9, while time-
optimal problems for different control regions are investigated in §11.2. Ob-
serve, however, that the fundamental device utilized in the above analysis,
was that of choosing at each instant t an admissible control Uo(t) which
minimizes a certain expression related to the given functions. The develop-
ment of a suitable generalization (the Pontjragin principle) is taken up in
§10A.
§1O.3. Sufficient Conditions Through Convexity 359
§10.3. Sufficient Conditions Through Convexity

In the preceding section we have seen several examples of control problems
which (perhaps after suitable transformation) yield to the convexity methods
of this text. We shall now subject the general control problem on a fixed
interval as formulated in §10.1 to the same methods, by imposing on it
whatever additional convexity is required to assure theoretical success.
We begin with the simple problem in which there are no control restric-
tions (CIlI = IRk), and derive conditions sufficient to provide a complete solu-
tion to the problem in which the state equation is linear while the perfor-
mance integrand is quadratic in the state-control variables.
As the basic analysis in §2.3, revealed, for the fixed interval [0, T], if we
can minimize
F(Y, U) = LT j[Y(t), U(t)] dt
on some set ~ specifying initial/target conditions on Y, where for some
Lagrangian multiplier function, PEel [0, T],
j(t; Y, U; Z) ~ f(t, Y, U) + P(t)· (G(t, Y, U) - Z), (24)
then in the absence of further state/control constraints, we minimize
on~,
F(Y, U)
under the state equation
= f: f[Y(t), U(t)] dt
Y(t) = G[Y(t), U(t)], t E (0, T).

Observe that (24) can be rewritten
j(t; Y, U; Z) = h(t, Y, U) - P(t)· Z,
where (25)
h(t, Y, U) ~ f(t, Y, U) + P(t)· G(t, Y, U).
Moreover, j(t; Y, U; Z) is [strongly] convex, precisely when h(t, Y, U) is
[strictly] convex, and if the state function G is linear in Y and U, then
only the [strict] convexity of f(t, Y, U) is required.
Assuming sufficient differentiability, from (25) we get
jy = -P and lu = m,
and the Weierstrass-Erdmann conditions on j (§7.5) are fulfilled by the
requirement that P be continuous. The corresponding Euler-Lagrange equa-
tions (§6.7) are, in condensed form,
d - -
d/y =fy, or -p = hy(t, Y, U), (26a)
d - -
d/u = fu, or m= hu(t, Y, U), (26b)
and they must be considered together with the state equation

Y = G(t, Y, U) (26c)
as constituting the differential system governing the problem (except at cor-
ner points of Y and/or points of discontinuity in U). Equation (26a) is called
the adjoint equation, and P is called the adjoint function. P can be given
interpretations related to particular disciplines. For example, see [K-S].
With this preparation, we can get our first major result.
(10.1) Theorem (Sufficiency). Suppose that D is an open set in ~d, 0/1 £; ~k,
and h(f, Y, U) is [strictly] convex on [0, T] x D x 0/1. Then each solution Yo,
P E C1 [0, T], Uo E C[0, T], of the system (26a, b, c) minimizes
F(Y, U) = IT f[Y(t), U(t)] dt
[uniquely] under the Lagrangian inequality
on:
P(t)· (G[Y(t), U(t)] - yet)) ~ ° (27)
(i) !!) = {(Y, U) E c [0, T]

1x CEO, T]: (Y(t)· U(t)) E D x 0/1}, if P(O) =
P(T) = (r);
(ii) !!)T = {(Y, U) E !!): YeO) = Yo(O)}, if P(T) = (r);
(iii) !!)o = {(Y, U) E !!): YeO) = Yo(O), YeT) = Yo(T)};
or on
(iv) !!)", = {(Y, U) E !!): YeO) = Yo(O), CI>(Y(T)) = (r}}, if P(T) = ACI>y(Yo(T)),
for some A E ~l that makes A· CI> convex. In the last case CI> is an I-vector valued
function on an open set containing Yo(T) such that CI>(Yo(T)) = (r), and the
columns of Cl>y are indexed by Y.
PROOF. For the stated conditions, if (Y, U) and (Yo, Uo) are in !!), then
F(Y, U) - F(Yo, Uo)
~ F(Y, U) - F(Yo, Uo)
= IT [h[Y(t), U(t)] - h[Yo(t), Uo(t)] - P(t)· (Y(t) - Yo(t))] dt
~ IT [hy[Yo(t), Uo(t)]· (Y(t) - Yo(t))
+ hu[Yo(t), Uo(t)]· (U(t) - Uo(t)) - P(t)· (Y(t) - Yo(t))] dt

(by convexity)
= - J(T dtd (P(t)· (Y(t) -

0 Yo(t)) dt = - P(t)· (Y(t) - Yo(t))
IT0
(by (26a, b))
= ° under (i), (ii), and (iii);
[with equality = equality in integrands = Y(t) = Yo (t), U(t) = Uo(t), t E

[0, T], if we recall the convention concerning points of discontinuity in
U and Uo.]
For part (iv), A acts as Lagrangian multipliers for the additional
constraints cI>(Y(T)) = O. As in §3.4(d), we consider minimization of the
augmented function of Bolza type
F*(Y, U) ~ F(Y, U) + A· cI>(Y(T)),
for which we have by the previous inequality and the hypothesized convexity
of cI> . A, the comparison
F*(Y, U) - F*(Yo, Uo)
~- P(t)· (Y(t) - Yo(t){ + AcI>y(Yo(T))· (Y(T) - Yo(T))
= [ - P(T) + AcI>y(Yo(T))] (Y(T) - Yo(T)) = 0
[with equality only if Y(t) = Yo(t), U(t) = Uo(t), t E [0, T]]. o

Remarks. In condition (iii), both the initial and final states are prescribed;
in both (ii) and (iv), the initial state is prescribed, but in (ii), there is essentially
no restriction on the final state, while in (iv) (which has significance only
when d > 1) the final state is restricted to lie in the zero level set of cI>. In each
of the latter cases, and in (i), we require that the Lagrangian multiplier
function P satisfy boundary conditions which compensate for the freedom
permitted in Y. If d > 1, we only need to compensate in each component of
Y, and we could establish similar results when there is a componentwise
mixture of initial/target conditions of the above types. It is also easy to
compensate for a target inequality such as y(T) :5: Yo(T) by requiring that
p(T) ~ O.
In the usual case, where only the actual solutions Y of the state equation
are of interest, we obtain minimization under the stated conditions among
those Y which satisfy Y(t) = G[Y(T), U(t)]. However, our formulation of
Theorem 10.1, permits consideration of certain state differential inequalities
provided that corresponding multipliers P(t) can be found with components
Pj of the correct signs. With further modifications of the integrand j by
addition of terms such as p.(t)I/J(t, Y, U), it is straightforward to formulate
corresponding theorems which permit such state-control inequalities as
I/J(t, Y(t), U(t)) :5: O. (See Problem 10.18.)
Linear State-Quadratic Performance Problem

A special but important nonautonomous control problem is that governed
by a linear state equation of the form
Y(t) = A(t)Y(t) + lB(t)U(t) (28)
for appropriate matrix functions A and IB, together with a "quadratic" per-
formance function of the type associated with energy assessments:
For symmetric positive definite matrices 0 and § with elements in C[O, T],
f(!, Y, U) will be strictly convex on [0, T] x ~d+k. (Why?)
Then
hU, Y, U) = f(!, Y, U)+ PW· (AW Y + IBW U)
will be strictly convex on [0, T] x ~d+k for any P in Ct, and the equations
(26a) and (26b) are, respectively,
- P(t) = Y(t)§(t) + P(t)A(t), (28a)
(!J = U(t)O(t) + P(t)lB(t), (28b)
where P, Y, and U are row vectors to be determined to satisfy these equations
together with (28) (the state equation) and with certain initial/target condi-
tions on Y, P.
Since 0 is invertible (§0.13), we can rewrite (28b) as
U = - PIBO-t, (29)
and substitute into (28) to obtain the first-order system which we abbreviate
Y= A Y - IB(Qtl jBp,
(30)
using bars to denote transposes.

Each solution (Yo, Po) of the linear system (30) that meets the boundary
conditions of Theorem 10.1(i), (ii), or (iii), supplies through (29) a solution of
the system (28) and (28a, b). As we know from the strict convexity of h(!, Y, U),
this will be the unique solution to our optimal control problem.
From the general analysis of linear systems as given in §A.5 (in particular,
(A.19)) we can assert that system (30) has a unique solution on any interval
[0, T] for any given initial conditions (Y(O), P(O)). However, we cannot pre-
scribe both Y and P at either endpoint, and this awkward feature of optimal
control systems complicates the analysis. Fortunately, here, we can give
simple arguments that guarantee existence of solutions meeting the split
boundary conditions of Theorem 10.1, and so, the unique solution of
the control problem. In case (i), Yo(t) = P(t) = (!J and Uo(t) = (!J minimizes
uniquely; (but see Problem 10.19).
Upon introducing the enlarged vector V = (Y, P) E ~2d, the system (30)
written in condensed form V = II\\V for an appropriate matrix 11\\, has for
each j = 1, 2, ... , 2d a unique solution J.j on [0, T] with J.j(0) = E j , the unit
vector in the jth coordinate direction in ~2d. Moreover, each solution V of
the condensed system may be expressed as a (unique) linear combination
V = If!i Cj J.j of these base solutions, with coefficients C = (c 1 , C2 ,.··, C2d) =
V(O). Note that then

V(t) = V t V(O) where V t = [V1 (t) : V2(t): ... : V2d (t)].
In particular
V(T) = 1Y(T) 1= V 1Y(O) 1 = [V1 V 2JI Y(O) 1

P(T) T P(O) V3 V 4 P(O) ,
say, where V T has been partitioned into four d x d matrices.
Now in case (ii), where Y(O) and P(T) = (9 are prescribed, we need to solve
the smaller system
(31)
which is possible provided that V 4 is invertible. However, for the admissible
choice Y(O) = (9, each solution P(O) of the reduced system (9 = V 4 P(O)
provides the unique solution of the associated control problem which, by
inspection, is clearly Yo(t) = P(t) = (9. Thus, P(O) = (9, and V 4 is invertible.
Similarly, in case (iii), where Y(O) and Y(T) are prescribed, we must be
able to choose P(O) to ensure that Y(T) = V 1 Y(O) + V 2P(O), which will be
possible if V 2 is invertible. Again, the admissible choice Y(O) = Y(T) = (9
gives P(O) = (9 as the only solution of V 2P(O) = (9, and thus the invertibility
of V2 is assured by the same appeal to uniqueness of solution of the control
problem.
Case (iv) is the most difficult to examine, and to remain within the frame-
work of linear analysis, we assume that CI>(Y) = twlI Y + L for a given vector
L E [RI and a given constant matrix twlI of maximal rank I (:::; d). Then Cl>y(Y) =
twlI, so that we must find a solution (Yo, P) of the system (30) which meets the
prescribed state conditions Yo(O), with Yo(T) E .91, the affine space in [Rd for
which twlI Y = - L, together with the adjoint condition, that P(T) = AtwlI for
some A E [RI. Thus we must find a A for which
Y(T) = V 1 Y(O) + V 2P(O) = Y Ed,
(32)
P(T) = V3 Y(O) + V 4 P(O) = AtwlI.
From the previous analysis, both V 2 and V 4 are invertible, so that if a
solution exists, we would have
P(O) = V21(y - V 1Y(O)) = Vi1(AtwlI - V3 Y(O)), (33a)
or
Y = V 1Y(O) + V 2Vi 1(twlIA - V3 Y(O)). (33b)
But then, twlI Y = - L gives the following equation for A:
twlIV2Vi1MA = -L + M[V2Vi1V3 - V 1]Y(0), (33c)
which may be solved, provided that the matrix twlI o = twlIV2Vi 1 twlI is
invertible. For the admissible case Y(O) = (9, L = (9, we note that Y(T) =
(9 E d, and thus as before, Yo(t) == P(t) == (9 will give the unique solution to
the optimal control problem, while each solution A of twlIoA = 0 gives, by
(33b), a Y, which will be an admissible Y(T) for the system (32), and hence
provide a solution to the control problem. We conclude that P(T) = (!J = MIA,
or since MI is of maximal rank I :s; d, that A = (!J; i.e., Mlo is invertible.
Since Mlo is invertible, we can always solve (33c) for A, and find Y Ed
from (33b), resulting in a compatible P(O) from (33a) which satisfies the
system (32) with Y(T) = Y. The resulting Yo(t), Uo(t) provide the unique
solution to the control problem.
For other linear systems and split-boundary conditions, such arguments
may not be possible-and even when they prevail there is in general no
closed method for obtaining the desired solutions to such two-point bound-
ary value problems. For a survey of the required numerical "shooting"
methods, see [Ke]. We note however, that in principle the fundamental
matrix WT can be approximated quite accurately for explicit linear systems
which are not too large, and thus used to obtain P(O) as an explicit solution
to a set of simultaneous linear equations, from which follows
Yo(t) 1 = [W] 1 y(o) I·

1 P(t) t P(G)
The unique optimal control Uo(t) is then given by (29) and it is always
continuous.
In cases (ii), (iii), and (iv), in which Y(O) is prescribed, we can obtain the
same results under slightly weakened hypotheses. Let's summarize our
findings
(10.2) Theorem. Let A, IB, 0, and § be given matrix functions with elements in
CEO, T], of sizes d x d, d x k, k x k, and d x d, respectively, where at each t,
O(t) is symmetric positive definite while §(t) is symmetric positive semi-definite.
Then there exists a unique optimal control Uo E C 1 [0, T] and trajectory Yo E
C 1 [0, T] that minimizes
F(Y, U) = IT [V(t)O(t) U(t) + Y(t)§(t) Y(t)] dt
under the state equation Y(t) = A(t) Y(t) + lB(t) U(t), on:
(ii) ~T = {(Y, U) E C1 [0, T] x CEO, T], with Y(O) prescribed};
(iii) ~o = {(Y, U) E ~T with Y(T) prescribed};
(iv) ~M = {(Y, U) E ~T with MlY(T) = -L}.
where L E IRI is given and MI is a given matrix of maximal rank I :s; d.
PROOF. Only the weakened hypothesis on § requires comment. The
integrand
f(f, Y, U) = V O(t) U + Y§(t) Y
remains convex, and is in fact semi-strongly convex in that
f(f, Y, U) = f(f, Yo, Uo) => U = Uo (why?).
§10A. Separate Convexity and the Maximum Principle 365
Thus by arguments for uniqueness of the solution, we now can conclude

only that equality of performance implies that V(t) = Vo(t). But the compet-
ing state functions Y(t) must satisfy the linear equation
Y(t) = A(t) Y(t) + lEB(t) Vo(t) with Y(O) = Yo (0).
Therefore Y(t) == Yo(t) on [0, T], so that full uniqueness still obtains, and we
can claim invertibility for the various (sub)matrices required in the foregoing
analysis. 0
(10.3) Remarks. The methods of linear analysis extend to provide corre-

sponding results for this problem when the state equation takes the form
Y(t) = A(t) Y(t) + lEB(t) V(t) + E(t) for given E E C[O, T].
(See [C-L].) As we have noted, the approach indicated in each case is numer-
ically feasible, and in simple examples with small constant matrices, can in
fact be carried out explicitly. See Problems 10.14, and 10.15 This is one
example of an optimal control problem in which control synthesis is achiev-
able, in that if the system is in state Y at time '"C, then there is an associated
optimal control Vo(t, '"C) say which can be used to bring the state to its
assigned target area at time T.
Such control synthesis although clearly desirable cannot always be ex-
pected, especially if there are not enough conditions present to guarantee
uniqueness of the optimal control.
Thus, we have theoretical access to a complete solution for certain linear

state-quadratic performance problems under the tacit assumption that there
are no restrictions on the admissible controls (other than having piecewise
continuity). Otherwise, Vo as defined through (29) might violate thes~ as-
sumptions. It may be possible to accomodate control restrictions such as
IV(t)1 :::;; 1 by adding a Lagrangian function such as ,u(t)(1 VI 2 - 1) to the
integrand (25), and proceeding as in the previous examples. Here, we would
°
try to find ,u ~ so that ,u(t)(l Vo(tW - 1) == 0, and we have already experi-
enced the complications such conditions impose on the ensuing analysis. In
the next section, we shall give a new approach to handling control con-
straints of an even more general form.
§10.4. Separate Convexity and the

Minimum Principle
As we have seen, an optimal control problem (with convexity present) is
characterized naturally by the auxiliary function
h(t, Y, V) = f(t, Y, V) + P(t)· G(t, Y, V)
expressed in terms of the performance function f, the state transformation

function G, and the Lagrangian multiplier (vector) function P awaiting further
specification. In this section, we uncover an alternative to the convexity
of h with respect to U, in the form of the minimum principle of Pontjragin,
which dominates the remainder of the book. We use this principle to solve a
problem with a nonconvex performance integrand.
In Theorem 10.1, we required convexity of h(f, Y, U) on an appropriate
set, which entails separate convexity in Yand U, and it is this which we now
wish to examine.
(10.4) Proposition. Suppose that for some Uo E CEO, T] and P E Cl [0, T],
h(f, Y, UoW) is convex on [0, T] x D x 0/1, where D is open in ~d and 0/1 !:; ~k,
then each Yo E Cl [0, T], which satisfies the adjoint equation
P(t) = - hy(t, Y(t), Uo(t», (34)
makes ho(t) ~ h(t, Yo(t), Uo(t» continuous on [0, T]; and constant in the auton-
omous case if additionally,
on (0, T).
PROOF. With j(t, Y, U, Z) = h(t, Y, U) - P(t)· Z, it follows that for Y E
ClEO, T],
F(Y, Uo ) - F(Yo, Uo ) ~ LT {hy[Yo(t), Uo(t)] - P(t)· (Y(t) - Yo(t)} dt
= P(t)· (Y(t) - Yo(t» I: = 0,
if Y(O) = Yo(O) and Y(T) = Yo(T). Hence Yo satisfies the second Euler-
Lagrange equation for j, which in the integral form given in §7.5, is that
ho(t) = j[Yo(t), Uo(t)] -iz[Yo(t), Uo(t)] . Yo(t)
= I.i:[Yo(r), Uo('t")] d't" + co,

for some constant co. Thus ho is continuous, and it is constant in the autono-
mous case where
Yo = G(Yo, Uo) and h(t, Y, U) = f(Y, U) + P(t)· G(Y, U),
since then
.i:[Yo(t), Uo(t)] = P(t)· G(Yo(t), Uo(t» - P(t)· Yo(t) = 0,
(except at corner points). D
In particular, under the hypotheses of Theorem 10.1, we may expect the

optimal ho to be continuous, possibly constant, and this can assist in the
detailed analysis of particular candidates for optimality, since discontinuit-
§10.4. Separate Convexity and the Minimum Principle 367
o a
Figure 10.2
ies in Uo are limited both in location and values to those for which
h(t, Yo(t), Uo(t - )) = h(t, Yo(t), Uo(t + )).
Example 1 (Optimal Highway Design). A construction company wants to

build a highway in terrain that has local altitude of oc = oc(t), over a horizontal
distance of T meters as indicated in Figure 10.2. Assume that land prepara-
tion costs (excavation and fill) for a highway of altitude y = y(t) at t E [0, TJ
is proportional to (y(t) - OC(t))2 per meter, and that safety concerns dictate
that the grade LPI not exceed a = 1/10.
Then the company would try to find y E @ = C1 [0, TJ that minimizes
F(y, u) ~ IT (y - OC)2(t) dt, with y(t) = u(t) E 1111 = [ - a, a]. (35)
Here, f(t, y, u) = (y - oc(tW, so the problem is not autonomous and for a

suitable P E @,
h(f, y, u) = (y - OC(f))2 + P(f)u
is convex. Therefore the choices of an optimal grade strategy U o = Yo are
limited to those for which
ho = (Yo - OC)2 + pUo is continuous;
in particular, pU o must be continuous which means that U o can have a discon-
tinuity only at points where p vanishes. Of course, we suppose that p satisfies
the adjoint equation (26a)
- p = hy = 2(yo - oc), (36)
and we can use Theorem 10.1 (i) if we take
p(O) = p(T) = 0, (36')
since Yo is unspecified at either endpoint. However, we would also need to
invoke (26b), 0 = hu = p, which means that yo(t) == oc(t), thereby eliminating
our ability to limit the grade-size. We will return to this problem after we
examine an alternative to (26b).
In the context of the previous proposition, suppose that for each non-
corner point t E [0, T]
(37)
Then under convexity of h(t, Yo(t), U) on 1111, it follows by §0.8, that this
function of U is minimized at Uo(t), or that
min h(t, Yo(t), U) = ho(t) = h(t, Yo(t), Uo(t)); (38)
Uel!/!
moreover, under continuity of ho, this must hold for all t, since then the
equivalent inequality
VUE dIt.
persists. (Why?)
This last inequality shows that the control Uo is instantaneously optimal in

minimizing h(t, Yo (t), U) at each t, and (38) affords our first glimpse of this
principle credited to Pontrjagin (c. 1955), 1 (which is formulated solely as a
minimum principle in this book.) Observe that if it holds for all U near Uo(t)
(in particular, if dIt is open) and h(t, Yo(t), U) is differentiable with respect to
U, then by ordinary calculus we recover (37). However, it can hold even when
these conditions are not met, as for example, when f(t, Y, U) = -I U I. More-
over, if dIt is convex and h is continuously differentiable in the components of
U, then by the mean value theorem,
U E dIt. (39)
This result of Loewen shows that Theorem 10.1 holds as stated when the
minimal inequality replaces equation (26b). (See Problem 10.28.)
Now, let's see how this helps the highway design project of Example 1.
First, we note that 1111 = [ -a, a] is convex, and that for any Yo E ri,
h(t, yo(t), u) = (Yo - CLf(t) + p(t)u (40)
is minimized on 1111 by u = uo(t) = ± a if p(t) ~ O. (A minimizing U o is undeter-
mined at points where p vanishes.) Therefore according to this analysis, the
optimal highway consists of segments of maximal grade where Yo = ± a,
joining sections that follow the given terrain (over which p = 0). Theorem
10.1 as modified by (39) shows that this descriptive choice is optimal, and in
fact it is unique because h(r, y, ~) is strictly convex. (See also, the next propo-
sition.) For the terrain of Figure 10.2, Yo is as indicated, where the interval
endpoints, and (1 can be approximated numerically. (See Problem 10.20.)
Observe that through the minimum principle we have obtained a useful
qualitative description of the optimal highway without actually solving the
problem.
1 In fact, the first version of the principle appeared in 1950 in a then little-known work of
Hestenes (see [He]).
In using (39) to modify Theorem 10.1, we still required joint convexity of

h(t, Y, U). The minimum principle permits alternatives. For instance, if
h*(t, Y) = min h(t, Y, U), (41)
Ue"lt
is defined, then we have the following result first considered by Arrow and
Kurtz. (For some generalizations, see [S-S].)
(10.5) Proposition. In Theorem 10.1, suppose that h*(t, Y) is [strictly] convex

on D. Then the conclusions remain valid [with uniqueness in Yonly] when
instead of (26b) and the convexity of h, the minimum principle (38) holds; i.e.,
h*(t, Yo(t» = ho(t) = h(t, Yo(t), Uo(t».
PRooF*. When (Y, U) and (Yo, Uo) are in f!} and we suppress the argument t
in the integrands, we have as before that
F(Y, U) - F(Yo, Uo ) ~ F(Y, U) - F(Yo, Uo)
= IT [h(t, Y, U) - h(t, Yo, Uo)] dt - IT p. (Y - Yo) dt
~ IT [h*(t, Y) - h*(t, Yo) dt - IT p. (Y - Yo) dt,
(by (41) and (38»
~ IT h~(t, Yo)-(Y - Yo) dt- IT p·(Y - Yo) dt,
(by the hypothesized convexity of h*(t, Y».

It is not obvious, but we show below that
(= - P from (26a». (42)
Therefore, the right side of the last inequality becomes -p·(Y - Yo)(t)I&,
which vanishes in cases (i), (ii), or (iii) of Theorem 10.1, and establishes our
result [since equality implies that Y = Yo under strict convexity of h*(t, Y)].
For case (iv), we can introduce the Bolza function F* and modify the argu-
ment as before to reach the same conclusions.
It remains to establish (42), and in doing so, we can hold t E [0, T] fixed.
Then, for all Y sufficiently near Yo = Yo (t),
Y.~(1- e) yo + eY, 0< e :$; 1,
is also near Yo so that with Uo = Uo(t):
h(t, y., Uo) - h(t, Yo, Uo ) ~ h*(t, Y.) - h*(t, Yo)
~ h~(t, Yo)· (Y. - Yo) = eh~(t, Yo)· (Y - Yo)·
If we now divide bye> 0 and let e '" 0, we find by the chain-rule that
h~(t, yo, Uo )· (Y - Yo) ~ h~(t, Yo)· (Y - Yo)
and this can hold for all Y near Yo, iff the derivative vectors on each side are
identical. D
(10.6) Remarks. Since the form of h* depends on the unknown adjoint func-
tion P, it is not easy to predict that h* will exhibit the desired convexity.
However, it clearly does so when
h(t, Y, U) = h'(t, Y) + h"(t, U) (43)
and h'(f, Y) is [strictlyJ convex on D.
F(y, u) = Il [y2 - Y - (1 - U)2J (t) dt
on
= {y E C1 [0, IJ; u E CEO, 1]: y(O) = O}
fi)
under the state equation y = u, where u(t) E o/J = [0, 2J, we note that the
problem is autonomous. Moreover, if p E C1 [0, IJ, then
h(t, y, u)~ y2 - Y - (1 - U)2 + p(t)u
= h'(y) + h"(t, u), (44)
where h'(y) = y2 - Y is strictly convex, but
h"(f, u) = -(1 - U)2 + p(Ou
is not convex.
In fact h"(f, u) is a concave parabolic function in u, so that for 0 :$; u :$; 2,
it is minimized at u = Uo where Uo = 0 or Uo = 2. Thus, by Proposition 10.4,
if Yo E C1 [0, IJ, and Yo = u, then
ho = y~ - Yo - (1 - UO)2 + pu o = Co, a constant,
when p satisfies the adjoint equation
-p = hy = 2yo - 1 (with p(l) = 0). (45)
Since (1 - u o? = 1 when Uo = 0 or 2, we see that y~ - Yo + pu o = Co + 1 is
also constant, and conclude that switching from U o = 0 to U o = 2 can occur
only at points where p vanishes. Finally, we know that if it exists, the
minimizing Yo with yo(O) = 0, is unique. Hence we can first try the simplest
possibilities.
Case 1. uo(t) == O. Then Yo = Uo = 0 so that yo(t) = yo(O) = 0, and by (45),
P= 1 with p(l) = 0, so that p(t) = t - 1 :$; O. However, from (44), we see that
for t < 1,
h(t, Yo, u) = -(1 - U)2 + (t - l)u
is minimized only when u = 2 and has the minimum value of -1 +
(t - 1)2 #- ho(t), since ho(t) is constant. This solution does not satisfy the
minimum principle (38). It may be the optimal solution but we cannot use
our proposition to verify its optimality (or that of the other simple case
uo(t) = 2. (See Problem 10.22.)
Case 2. We must look for a solution with at least one switching point a,
say, where p vanishes. Since p(1) = 0, it is simplest to try p(t) = 0 for t ~ a.
Then 2Yo(t) - 1 = - P = 0 or yo(t) = t, so that uo(t) = Yo(t) = 0 for t ~ a.
But then for t ::;; a we should try uo(t) = 2 so that yo(t) = 2t, and continuity
of Yo at a requires that a = t. Finally, for t::;; t: P(t) = - 2Yo(t) + 1 =
-4t + 1 so that p(t) = -2t 2 + t + c, and p(1/4) = 0 => C = -i so that
p(t) = - 2(t - t)2 ::;; O. Now for t::;; t, h(t, 2t, u) is minimized when u =
U o = 2 (since p ::;; 0) and it has the minimum value ho(t) = -l Similarly for
t ~ t: h(t, t, u) = (t)2 - t - (1 - U)2 is minimized when either u = 2 or u =
U o = 0 and it has the minimum value ho(t) = -l Therefore for all t E [0, 1J
this Yo satisfies the minimum principle. Consequently, the unique solution to
our problem is given by
yo(t) = 2t; and uo(t) = 2, t< t,
_ 1..
- 2, =0, t > t,
and the optimal control is bang-bang.
One final observation may be of value. As we have seen, in analysis of

optimal control problems, the Lagrangian multiplier P acquires equal status
with the state vector Y. In some sense we are optimizing with respect to
Yand P through the control U, and it is sometimes useful to reflect this fact
by writing the auxiliary function as
H(t, Y, U, P) = f(t, Y, U) + p. G(t, Y, U).
Then note that formally, equations (26a) and (26c) become

P= -Hr' (46)
which resemble those canonical equations of §8.4 for a kind of "Hamiltonian"
H. Although the presence of U prevents this from providing a complete
analogy, it is possible to incorporate some features of Hamiltonian mechan-
ics in analyzing problems of optimal control. Indeed, it is this partial analogy
which seems to be responsible for the original motivation of the Pontrjagin
(maximal) principle; (particularly illuminating discussions are given in [YJ
and in [LoJ). Further sufficiency theorems incorporating convexity-type
inequalities can be formulated in terms of hypothesized solutions to an
associated Hamilton-Jacobi equation, but these too, cannot be readily
implemented in the solution of actual problems. However, there are related
sufficiency theorems utilizing suitable field theory-like extensions of the
Weierstrass method of §9.1 which can be carried through. (See [Le].) A
sufficiency result free of convexity assumptions is given in Problem 10.25.
PROBLEMS
10.1. Minimize F(y, u) = g u 2(t) dt under y = -2y + u with y(O) = 1, y(l) = O. Is

your solution unique?
10.2. Minimize F(y, u) = g (2 - 5t)u(t) dt, with y(O) = 0, y(l) = e 2, under:
(a) y(t) = 2y(t) + 5e 2'u(t), lu(t)1 ::;; 1. Is your solution unique?
(b*) y(t) = 2y(t) + 4e 2'u(t), lu(t)1 ::;; 1. Hint: Let v(t) = e- 2'y(t).
10.3. To maximize F(y, u) = gy2(t) dt under y = u with y(O) = 0, lu(t)1 ::;; 1, show
that ly(t)1 ::;; Itl, and hence F(y, u) ::;; ? When does equality hold?
10.4. In minimizing F(y, u) = g(u(t) + 1) dt, under y = -u with y(O) = y(l) = 0,
show that each control u E C[O, 1] with g u(t) dt = 0 is optimal.
10.5*. An old cylindrical concrete tank of height 30 feet contains water of depth y
used for drinking; it leaks at a rate proportional to y. The initial depth is 12
feet and over the next 100 days we can supply the tank with water from a
controllable but limited source. Suppose that the resulting state equation is
approximated by y = -O.ly + u, ft/day, where u(t) E [0, 1] represents the net
inflow limited by goo u(t) dt = 60, say. How should we control u to maximize
the mean depth 100- 1 goo y(t) dt during the time interval [0, 100]? Hints: Let
Y1 = eO/'y, and write the control inequality as u(u - 1) ::;; O.
10.6. Obtain the energy formula (17) of §1O.2(f) by writing the square of a single
integral as a double integral.
10.7. (a) (Solution by steepest descent.) If (22) holds, then (20) becomes IYI" =
V· U. Use the Cauchy inequality (§0.1) to get I YI" ;::: -I UI (with equal-
ity... ). By integration, conclude that nlUI dt;::: IAI under given hypo-
theses (with equality... ). Formulate an optimal control problem for the
spinning body in which the fuel consumed (by say torque producing jets)
in bringing the body to spin rest at time T, as measured by the above
integral, is minimized. Which controls Uo will be optimal; i.e., how will
they be directed?
(b) Combine the inequality in part (a) with the Schwarz inequality for inte-
grals [Rud] to obtain
IAI2 ::;; Tn IUI 2 dt (with equality only if IU(t) I = const.).
Conclude that the energy integral on the right is minimized for fixed
T;::: IAI by the controls
u. (t) = _~ Yo(t) .
o T IYo(t) I
10.8. Under the same conditions as in Problem 10.7, show that for k > 0, the con-
trols Uo(t) = -mYo(t)/lYo(t)1 with m = min(k, 1), minimize Ink + IUI 2 (t)] dt.
Determine the associated minimum value(s) and optimal time(s) of transfer.
Argue that this control prescription is unique.
10.9. For the spinning body in §1O.2(g), a measure of resistance to motion can be
admitted if a Y is subtracted from the right side of the state equation(s), where
a is a positive scalar constant. Carry out the time optimal analysis for the
resulting problem. Hint: Equation (20) reduces to IYlo = -al YI + V· U.
Now let v(t) = ea'i Y(t)l.
Problems 373
10.10. To maximize y(2), if Y = u with lui:::; 1, where

y(O) = y(O) = 0; y(2) = 1:
(a) Explain why it would suffice to find (Yo, uo) meeting the above conditions
which minimizes
F(y, u) = J: [ -u + p(y - u) + J-t(u 2 - 1)] (t) dt
with p, J-t continuous and J-t ;::: 0 on [0, 2], such that
J-t(t)(u~(t) - 1) == O. (47)
(b) Examine the convexity of the associated integrand j(!, u, ji).
(c) Consider ~ = F(y, u) - F(yo, uo), and integrate the p(y - Yo) term by
parts to show that if y also meets the above conditions, then
~ ;::: p(2)(y - Yo)(2) + J: [p(y - Yo)] (t) dt

when
+ p) + 2J-tuo == o.
-(1 (48)
(d) Conclude that ~ ;::: 0, if also p(t) = c(2 - t) for any constant c.
10.11. (e) Observe that Uo = ± 1 satisfies (47) and with (d), (48), provided that
2J-t = (1 + p)u o = (1 + c(2 - t))u o ;::: O.
Argue that Uo = sgn(1 + c(2 - t)) suffices.
(f) We must determine c so that
Yo(2) = 1= J: yo(t) dt = J: (2 - t)uo(t) dt,
or after substitution from (e):
1= L2 't" sgn(1 + c't") d't".

Show that this requires c = -llJi
(g) Find the optimal yo(2), and explain why Yo is unique. Hint: Why is the
associated optimal control U o unique?
(h) Note that if Yo(2) = 2, then any c ;::: -t suffices in (f). Does this contradict
uniqueness?
10.12. To minimize F(y, u) = J~ [(y - e- t )2 + u2] dt under
y(t) = g(t, u) ~ (1 + sin t)u(t) with y(O) = o.
(a) Reformulate this as a related problem for
F(y, u) = F(y, u) + f p(t) [g(t, u) - y(t)] dt;
(b) Explain how convexity might be used to characterize a minimum for F

when p(n) = o.
(c) Obtain the Euler-Lagrange equations for the integrand of (b);
(d)* Try to solve these equations.
(e) Should the solution be unique? Explain.
10.13. Minimize F(Y, u) = g u (t) dt where

2
Y(O) = (y, yd(O) = (1) and Y = (Y1' u),

using multipliers P = (p, pd with P(2) = (1). Is optimal control U o unique?
bang-bang?
10.14. Minimize F(Y, u) = g u (t) dt, where
2
Y(O) = (1),
and y + 5Y1 = 15 at time t = 2. Hint: Theorem 10.2 with M = [1,5].

10.15*. (a) To minimize
F(u) = f l00
o x
dx
+ (25 + x/4)u(x)
on C[O, 100], with
G(u) = f l00
o x
u(x) dx
+ (25 + x/4)u(x)
=1
,
where 0 ::;; u(x) ::;; 1, show that it suffices to find U o which minimizes
F(u) = F(u) + AG(u) + J0 rlOO

J.L(x)(u(u - l))(x) dx
for an appropriate constant A, and continuous J.L ~ 0, such that

J.LUo(uo - 1) == O.
(b) For the appropriate j, set 1. = 0 and note that J.L(x) = 0 only when
x = , = 25/A1' where Al = A - 4- 1. Otherwise, we must take uo(x) = 1
1 =,
for x < " uo(x) = 0, x > " in order to have J.L ~ O. Show that G(u o) =
= 20(e 5/4 - 1), so that A = 251' + 1/4 is known.
(c)* For convexity, show that f.. ~ 0 when u ~ O. (This is immediate when
x < ,; for x > " it is essential to use the fact that J.L(x) = (XA1 - 25)/x 2 .)
(d) Observe that J.L obtained in (b) is continuous, and conclude (through
(3.10)) that U o is indeed the unique optimal "control" for this problem
(which is given a physical origin in Problem 11.26).
10.16. Under the conditions of Problem 10.13, find P(2) to minimize
F*(Y, u) = IY(2) - (5, 2W + F(Y, u).
10.17. (a) Transform the problem of minimizing
Sa2" (y2 + u2)(t) dt under ji + y= u
with say y(O) = y(O) = 0, y(2n) = 1, into one for which Theorem 10.1
applies. Does Theorem 10.2 apply? Explain.
(b) Find an optimal trajectory for this problem. Is it unique? Hint: Eliminate
u and transform to a Bolza problem with (ji + y)(2n) = O.
10.18. (a) Explain how the hypotheses of Theorem 10.1 should be modified to admit
a single state-control inequality such as
",(t, Y(t), U(t)) ::;; 0, where", is Cl.
Problems 375
(b) How should the proof be modified?

(c) How will the conclusions change?
(d) Use your result to reexamine Problem 10.10.
(e) Make a similar analysis of Proposition 10.5.
10.19. Consider Theorem 10.2 in case (i) of Theorem 10.1. Does a solution (yo, Uo)
exist? Is it unique?
10.20. For Example 1, in §1O.4:
(a) Explain why (J should be chosen to make So (Yo - IX) dt = 0, and how you
might locate (J by using the graph of IX in Figure 10.2.
(b) How should 't" be chosen?
10.21. (a) In minimizing
F(y, u) = IT lui (t) dt under y = u, lu(t)1 ;5; 1,
with given y(O) and y(T) = ll!, show that we should choose
uo(t) = {O Ip(t)1 < 1,

-sgn p(t), Ip(t)1 > 1,
for an appropriate adjoint function p. Use this fact to solve the problem
when y(O) = 1, and T = 1.
(b) What happens if T = 2? if T = t?
10.22. (a) In Example 2 of §1O.4, verify that uo(t) = 2 does not satisfy the minimum
principle for h.
(b) Solve the corresponding problem when y is added to the integrand of F.
10.23. Minimize H(u 2 - 2u)(t) dt, under the isoperimetric inequality Htlul(t) dt;5; i.
Hint: Introduce y with y(t) = tlul(t), y(O) = O.
10.24. Minimize say - (2t - 1)]2 dt, subject to y(O) = y(l) = 0, with y ~ 0, by
introducing y = u. Hint: Problem 10.18.
10.25. An elementary sufficiency criterion can be obtained as follows:
(a) Set h = h - p. Y and as in Theorem 10.1, show that
F(Y, U) - F(Yo, Uo) = IT [h(t, Y(t), U(t)) - h(t, YO(t), Uo(t))] dt
- P(t)· (Y(t) - Yo(t)) 1:.

(b) Conclude that if h(t, Y, U) ~ h(t, Yo(t), Uo(t)) when (t, Y, U) E [0, T] x
D x <¥t, then (Yo, Uo) minimizes F(Y, U) on fi)"" if P(T)· (Y - Yo(T)) ;5; 0
when (Y) = ll!, under the given Lagrangian inequality.
(c) Here, there are no convexity assumptions. How would you modify the
hypotheses to obtain a unique Yo? optimal control?
10.26. (a) Use Theorem 10.2 to minimize
I2 IU(t)1 dt under Y= U E [R2 where U = (u, ud,
with Y(O) = ll! and (e Y - Yl)(2) = O. (Take !PI = 1.)

(b) This problem has a simple graphical interpretation. Use the result of
Problem 10.25 to contrast what happens when instead (e Y - yd(2) = 2.
10.27. (Optimal oil production.) An offshore oil field estimated to contain K barrels
is to be pumped out by w identical rigs over a large number of days, T. Each
oil rig costs R dollars to install and has a maximum production rate of m
barrels per day of oil that will sell at a fixed price of D dollars per barrel. At
discount rate r, the company should control both wand v = vet), the fraction
of its total productive capacity in use at time t, to maximize
D IT e-r'v(t)mw dt - Rw where y = v(mw), yeO) = 0,
with 0 ~ vet) ~ 1 and yeT) ~ K. (See [S-S] and Problem 6.44.)

(a) Show that this leads to a nonconvex (and nonstandard) minimization
problem in optimal control.
(b) Explain why it is enough to introduce W = mw, B = R/mD, and minimize
F(y, W) = BW - IT e-r'y(t) dt, where yeO) = 0, yeT) ~ K,

and 0 ~ yet) ~ w.
(c)* Verify that (Yo, Wo) with 0 ~ Yo ~ Wo, Yo(O) = 0, Yo(T) = K, minimizes
provided that we can find /-l = /-let) ~ 0, v = vet) ~ 0, and p > 0, for which
B= IT /-let) dt, /-l(Yo - Wo) == vYo == 0 and (/-l - v)(t) = e- r• - p.
Hint: See §10.2(e).

(d) Show that one optimal strategy is to take Yo == Wo, t ~ t; Yo = 0, t > t,
where Wo = K/t; find t implicitly when Br < 1. (Assume T> t, and
ignore the fact that Wo = Wo/m should be an integer.) Is this the only
optimal strategy?
10.28. To establish (39): if Uo = Uo(t) and U are in d/i, then for 0 < e ~ 1 note that
U. ~ (1 - e)Uo + eU is also in d/i. Show that for some l < e,
o ~ h(t, Yo, U.) - h(t, Yo, Uo) = ehu(t, Yo, Ui )· (U - Uo),
where Yo = Yo(t). Now divide by e and let e -+ O. Where is the convexity of d/i
used?
10.29*. (A fish harvesting problem.) A large population of fish is to be taken
commercially over the next 4 years. During this period, its size y (in millions)
at time t can be described approximately by the equation
y = y _ y2 _ uy, o ~ t ~ 4,
where u = u(t) represents the fraction of the existing population to be har-
vested at time t, and the quadratic term accounts for possible natural loss due
to overcrowding, fighting, etc. (If there is no fishing, this population could
stabilize at one million.) The fishing industry wants to control u to maximize
its yield H(uy) dt, where yeO) = a is assumed known and positive, and 0 ~
u(t) ~ 1.
Problems 377
(a) Show that this leads to a nonconvex minimization problem in optimal

control, and identify the source(s) of nonconvexity.
(b) Following the model in §1O.2(e) show that it suffices to minimize
F(y) = f (y2 - Y + y)(t) dt
under the Lagrangian inequalities

y2 _ Y + y S; 0,
Conclude that it suffices to find Yo > 0 with Yo(O) = a which minimizes
F(y) = f [(1 + J1. - V)(y2 + y) - (1 + J1.)y] dt,

for appropriate continuous multiplier functions J1. ~ 0, v ~ 0, if
J1.(y5 - Yo + Yo) == v(Y5 + Yo) == 0 (~J1.vYo == O~J1.V == 0).

(c) Note that the integrand 1(1, y, y) is strongly convex when p ~
(1 + J1. - v) > o. Explain why it then suffices to find Yo > 0, P > 0,
satisfying p = 2py - (1 + J1.), with Yo(O) given, p(4) = 0, and J1., v as in (b).
(d)* For y(O) = i, conclude that
(1 + 3e- t )-1, 0 s; t s; log 3,
{
yo(t) ~ 1/2, log 3 s; t s; 2,
l/t, 2 s; t S; 4,
provides the unique minimum sought (so that U o ~ 1 - Yo - (Yo/Yo) is
the unique optimal control). Hint: Calculate Y5 - Yo + Yo and Y5 + Yo
in each subinterval, and conclude that J1. = 0 except on [0, log 3], while
v = 0 except on [2, 4]. Then use the reduced equation for p to find these
functions on the remaining intervals. Check that they are nonnegative,
while p > 0, with p(4) = o.
(e)** Obtain the solution in (d) directly by solving the equations for
(nonnegative) p, J1., v, Yo from the previous parts. Outline: Note that
J1.V == 0 with p(4) = O~near 4, v> 0, J1. = 0, so that Yo = -Y5 which
can be integrated. Then solve p equation for (1 - v) and determine t2 at
which V(t2) = O. Argue that in next interval J1. = v = 0 so that Yo = t.
Finally, permit J1. > 0, and conclude that in such intervals (where v = 0),
Yo = - Y5 + Yo, which is an integrable equation (for x = l/yo) with a
unique solution having Yo (0) = i. Then solve p equation for (1 + J1.) and
find tl < t2 at which J1.(t2) = 0, while yo(t 1) = t. Assemble results.
(f)* Analyze the simpler case when y(O) = t.
CHAPTER 11
Necessary Conditions for Optimality
What conditions are necessary for optimal performance in our problems? In

Chapter 10 we saw that if a control problem can be formulated on a fixed
interval and its defining functions are suitably convex, then the methods of
variational calculus can be adapted to suggest sufficient conditions for an
optimal control. In particular, the minimum principle of §10.3 and §10.4 can
guarantee optimality of a solution to the problem. In §11.1 we will discover
that this principle is necessary for optimality whether or not convexity is
present, even when the underlying interval is not fixed (Theorem 11.10). Then
in §11.2, we examine the simple but important class of linear time-optimal
problems for which the time interval itself is being minimized and the adjoint
equation (a necessary condition) can be used to suggest sufficient conditions
for optimality. Finally, in §11.3, we extend our control-theory approach to
more general problems involving Lagrangian inequality constraints, and in
Theorem 11.20 we obtain a Lagrangian multiplier rule of the Kuhn-Tucker
type.
The material in this chapter (and in §A.7) is significantly more difficult
than that of previous chapters, even though it represents only a small selec-
tion of the results to be found in [Po], [Ber] , [He], and [M-S], among
others.
§11.1. Necessity of the Minimum Principle

In this section we wish to show that the sufficient conditions obtained in
§10.4 in the presence of convexity, are in general necessary for an optimal
solution whether or not the integrand exhibits convexity. Recall that this
378
§11.1. Necessity of the Minim urn Principle 379
was also the case for the fixed interval problem in the variational calculus
(Chapters 3 and 6).
Specifically, we shall prove the necessity of the minimum principle, which
asserts that optimality of a control Uo guarantees its pointwise minimization
of some related function h, in that
h(t, Yo(t), U) ~ h(t, Yo(t), Uo(t», V UEi1IJ, (1)
where i1IJ is the control set, and Yo is an associated optimal state function.
Moreover, we expect that for performance and state functions f and
G, h might take the form
h(t, Y, U) = f(t, Y, U) + P(t)· G(t, Y, U) (2)
for some function P to be determined, but which probably satisfies the
adjoint equation
(3)
along the optimal trajectory, with related target conditions.
Now, under classical assumptions that i1IJ is open and h is differentiable in
U, (1) implies that h u(', Yo, Uo) == (f). Then the minimum principle is just
the Weierstrass necessary condition (7.15), for the modified integrand
j(t, Y, U, Y) = h(t, Y, U) - P(t)· Y, (4)
if we think temporarily of U as U. Indeed, the relevant combination for
W = (U, V) E lRk+d, is the excess function given by
j(-, Yo, U, V) - j(-, Yo, Uo, Yo) -icu,rk, Yo, Uo, Yo)·(w - (Uo, Yo»
= h(', Yo, U) - h(', Yo, Uo) - P'(V - Yo) + p·(V - Yo)
= h(', Yo, U) - h(', Yo, Uo) (assuming that hu (', Yo, Uo) = (f)1
and this is nonnegative when (1) holds.

Finally, since from (4),
j(t, Y, U, Y) = f(t, Y, U) + P(t)(G(t, Y, U) - Y), (5)
we see that the minimum principle involves the existence of Lagrangian
multiplier functions P, for which the modified integrand j satisfies the
Weierstrass condition. This too, is something we might expect, but we have
not established it previously with Lagrangian constraints of the form Y =
G(t, Y, U); moreover, we know that i1IJ might not be open and that hu need
not exist.
For these reasons, the necessary conditions for optimality of a control
must be established independently, and we should not anticipate success by
elementary methods.
We begin by considering the effects of certain permissible control modifi-
cations on an autonomous state equation and obtain a precise description of
1 If '¥t is convex, then (39) of Chapter 10 can be used to eliminate this assumption.
380 11. Necessary Conditions for Optimality
the resultant states at later times. We use this in part (b) to establish the
minimum principle for autonomous problems on a fixed interval (albeit with
a slightly different h), and apply our results to the oscillator energy problem
from §10.2(f). Finally in part (c) we transform the general problem into the
case covered in part (b) and give illustrative applications which indicate the
techniques employed when using these conditions, as well as some of the
difficulties encountered in doing so.
We will sometimes refer to a state Y E C1 [0, T] as a state trajectory.
We may also assign a vector-valued function or a matrix-valued function any
smoothness property common to its elements.
(a) Effects of Control Variations

Suppose that some control Uo E CEO, T] produces a solution Yo E C1 [0, T]
of an (autonomous) state equation
Y = G(Y, Uo) on (0, T), (6)
where both G and Gy are continuous on D x o/J for a domain D of ~d and a
bounded control set o/J £; ~k. T is finite.
We wish to characterize analytically, the state target values Y(T) which
arise through solutions of (6) with Y(O) = Yo(O), when the control Uo is
replaced by constant values V E o/J over small subintervals [r/, r] of continu-
ity, as illustrated in Figure 11.1.
Intuitively, we expect that small intervals of replacement should produce
small target effects, but we require a far more precise mathematical descrip-
tion. The results are presented in Propositions 11.2, 11.3, and 11.4, and if
desired, they can be accepted without proof on first reading.
We consider first the effects of a single interval of replacement [r/, r],
where r' = r - ec for some given c > 0, and a positive e so small that Uo is
continuous on [r/, r]. The result, denoted U., is again an admissible control
and it produces a unique associated state trajectory y. that satisfies the
- - - U,
~k ----- Uo
O~----------~--~--------------~r-~
I .'. ----'!..L t
Figure 11.1
§11.1. Necessity of the Minimum Principle 381
equation Y = G(Y, U.) on (0, T), with y(o) = Yo (0). Indeed, when t ~ ,', we
just take y'(t) = Yo (t). On Eo', ,], the small interval of replacement, we can
solve the equation Y(t) = G(Y(t), V) with Y.(,') = Yo(,'), by the methods of
§A.5, and obtain
Y.(,) = Yo(,) + f (G(y'(t), V) - G(Yo(t), Uo(t)) dt. (7)
For small s, Y.(,) will be sufficiently near Yo(,) to permit the embedding
methods of Theorem A.18, to be applied to obtain the solution y. on the
successive subintervals of continuity of Uo over [" T]. In particular, the
resulting target value y'(T) is known, and it can be compared with Yo(T).
(11.1) Lemma. There are G-dependent positive constants M and <50 such that
on Eo', ,],
a(t) = I y'(t) - Yo(t) I ~ Ma, when Cl: ~ <50 ,
PROOF. Replacing, with t in (7), we get
y'(t) - Yo(t) = f [G(Yo(s), V) - G(Yo(s), Uo(s))] ds
+ f [G(y'(s), V) - G(Yo(s), V)] ds. (8)
The first integrand on the right is continuous, hence bounded by Mo (say) on

the interval [0, T], while the second may be estimated by some
1'1 y'(s) - Yo(s)l, exactly as in (A.16).
It follows that for t E Eo', ,J:
a(t) ~ MoCl: + l' f a(s) ds = l/!(t), say, (9)

so that
~(t) = 1'a(t) ~ 1'l/!(t) or (e-ytl/!(t))" ~ 0.
Hence
(10)
if say cs ~ 111' = <50 , and we replace eMo with M. D
Henceforward, we suppose a ~ <50 ,

If we examine (8) when t = " we see that
Y.(,) = Yo(,) + sc~G(" V) + S3(S), (11)
where for smalls, we have approximated the first integrand by
~G(" V) ~ G(Yo('), V) - G(Yo(')' Uo(,)), (12)
and used the preceding lemma to estimate the second integrand uniformly by
some ye. The final expression in (11) collects the error terms associated with
°
these approximations into some 3(e) that approaches with e; (ec = 7:' - 7: is
just the length of [7:', 7:J).
Next, we note that on the remaining interval [7:, TJ both y. and Yo satisfy
the same state equation
Y = G(Y, Uo)
but with initial value Y(7:) = A, for different A.
By the arguments used in Theorem A.20 we know that as a function of A,
the resulting solution Y(t, A), say, is differentiable with respect to the compo-
nents of A, and when evaluated at Ao = Yo(7:), W< ~ lA(·, Ao) is the unique
solution to the linear system
W= Gy(Y, U)W on (7:, T) (13)
with W«7:) = ~d (the d x d identity matrix), when the columns of Gy are
indexed by Y. It follows that since A = y'(7:) is given by (11), then on [7:, TJ
d Y.(.) I
-d = lA(·, AO)-d
dAI = cW«·)L\G(7:, V);
e .=0 e .=0
i.e.,
(14)
for some function 3(t, e) that approaches 0, as e -+ 0. When t = T, let W; =
W;(T) and 3(e) = 3(T, e), so that
y'(T) = Yo(T) + ecW<L\G(7:, V) + e3(e). (14')
At this point, we digress to obtain a simple result.
(11.2) Proposition. If PEel [0, TJ is a solution of the adjoint equation

p= - Gy(Yo, Uo)P on [0, TJ, (15)
then P(7:) = P(T)W<.
PROOF. Since under transposition, P = -PGy(Yo, Uo), we have by (13)
that (PWS = PW< + PW< = 0, while Wlr) = "d. Thus P(7:) = P(7:)W«7:) =
P(T)W«T) = P(T)W<. D
Now we return to our principal development. From (14), it follows that if

we next replace Uo on a "later" disjoint interval of continuity [7:'1' 7:1J by
some constant value V1 E 1111, then at 7:'1 = 7:1 - eC 1 the new y. agrees with the
old so that we can repeat the construction to obtain for t > 7: 1:
y'(t) = Yo(t) + e(cW«t)L\G(7:, V) + c 1 W<,(t)L\G(7:1 , Vd) + e3(t, e).
In particular, this would apply when t is the target time T, and we can
evidently repeat the process at a finite set of 7: i • This observation establishes
the following:
(11.3) Proposition. If Uo is replaced by constant values V; E I1It on disjoint

intervals of continuity [r; - IX;, -r;], i == 1, 2, ... , n, then the new control U.
produces a state trajectory y. with target value
n
y'(T) = Yo(T) +6L c;"fi + 63(6), (16)
;=1
where 6 > 0, c; ~ 0, and
i = 1,2, ... , n, (16')
with terms defined in (12) and (13), while 3(6) -+ 0, with 6. D
The next argument is rather deep analytically, although it is not difficult

to appreciate its geometrical plausibility. We suppose as we may that
Yo(T) = (I), and observe that for fixed "fi E ~d, i = 1,2, ... , d, as the c; ~ are
varied, the set
°
K~{x=.f
.=1
c;"fi:c ~O,i= 1,2, ... ,d}
1
forms an infinite pyramid in ~d with apex at (I), and "edges" in the directions
of the "fi as indicated in Figure 11.2. This pyramid need not have an interior,
but when it does, we can establish the following:
(11.4) Proposition. If K contains a point Xo in its interior, then for some

admissible control U., the corresponding y'(T) = pXo, if p > is sufficiently
small.
°
PROOF*. To have an interior, the edge vectors "fi of K must be linearly
independent, and we note that for some small p, pXo is also in the interior of
YoCT) = {} YoCT) = f7
Figure 11.2
the compact subpyramid
Ko = {x = I: it Ci"li: 0 ~ ~ 1:I: 0, 0 ~ Ci ~ 1, it Ci = I},
also shown in Figure 11.2.

Of course, we know that for 1:0 sufficiently small, we can get all of the
target values from (16) (with n = d); they take the form
d
Y,,(T) = I: L Ci"li + 1:3(1:),
i=1
but the correction term cannot be ignored. Instead, we consider the function
fwhich assigns to each X E K o, the target value from (16) with the same 1:, ci ,
so that f(X) = X + 1:3(1:). Since for 1:0 sufficiently small, and 0 ~ Ci ~ 1, we
can take a single function 3, it follows that this f: Ko --+ IRd is continuous.
Moreover, for X E 8K o, the boundary of K o,
If(X) - XI = 1:13(1:)1 ~ 1:0 < IX - pXol,
if 1:0 is sufficiently small and p> 0 is even smaller. Then examining Figure
11.2 again, we can "see" that f does not move the boundary of Ko far enough
to remove pXo from f(Ko); i.e., some X E K o, has the target value f(X) = pXo
as required.
(To complete this argument rigorously we must appeal to a form of
Brouwer's Fixed Point Theorem as given, say, in [L-M]; this circumstance
gives analytical depth to the proposition, and to subsequent results depen-
dent upon it.) 0
(11.5) Remark. This proposition also holds if K is replaced by the target

cone, $" also indicated in Figure 11.2, where
$" = {X f =
;=1
ci"li: C; ~ 0, i ~ n = 1,2, ... }, (17)
and the "Ii are any of the vectors defined by «12), (13), and (16)) for some rio
V;. For if Xo is in the interior of $", then there must be d linearly independent
"Ii forming a subpyramid K containing Xo in its interior. With a slight shift
if necessary, we can suppose these V; to be associated with distinct r i , so that
the construction in the proposition is valid for K, and so for $".
(b) Autonomous Fixed Interval Problems
To understand the relevance of the next result, note that the autonomous
optimal control problem from §1O.1 admits (Mayer) formulation as that of
finding a Y = (y, Y) E C1 [0, T] with minimal y(T) among those which satisfy
an augmented state equation

Y= G(Y, U) on (0, T), (18)
with prescribed Y(O), and, say, Y(T), if the target is fixed.
Indeed, if G = (f, G) so that y = f(Y, U) is the "first" state equation in
(18), and we set y(O) = 0, then y(T) = $6 f(Y(t)), U(t)) dt = F(Y, U). Hence
minimizing y(T) under (18), is equivalent to minimizing F(Y, U) under the
usual state equation Y = G(Y, U) with prescribed Y(O) and, say, Y(T). How-
ever, as we shall now show, this Mayer problem admits attack by the geomet-
rical methods just developed. The tilde over G will be suppressed but d will
be replaced by d + 1 to facilitate application to the control problem as
originally formulated.
(11.6) Theorem (Pontrjagin). Let D be a domain of IRd +1 and I1Ii be a bounded

set in IRk. Suppose that on a fixed interval [0, T], (Yo, Uo) minimizes y(T)
among those with Y = (y, Y) in
fl) = {(Y, U) E C1 [~, T]x _CEO, T]: (Y(t), U(t)) E D x 11Ii}
III Y(O) = Yo(O), <D(Y(T)) = @ ,

Y = G(Y, U) on (0, T).
(Here G and Gy are continuous while either <D == @ or <D is a C 1 I-vector valued
function having <Dy(Yo(T)) of maximal rank I :::;; d.) Then, there exists a solution
P = (p, P) of the adjoint equation
p= - Gy(Yo, Uo)P on (0, T), with P(T) = A<DY(Yo(T)), (19)
for some A E IRI, such that the minimal inequality
P(-r)· G(Yo(-r), V) ~ P(-r)· G(Yo(-r), Uo(-r)), V VEI1Ii, (19')
holds at each point -r of continuity of Uo.
Remark. In fact, p. G(Yo, Uo) is constant so that (19') holds at all t. (See
Remarks 11.11.)
PROOF. Assume that Yo(T) = i9, and consider first the simpler cases where the
target is either fixed (<D(Y) == Y) or free (<D(Y) == @). In these cases, the target
cone % of 11.5, cannot contain in its interior the "downward" pointing
vector Xo = (-1,0, ... ,0). [For then by Proposition 11.4 (as extended), there
would be a modified control U., with associated state target value y'(T) =
( - p, 0, 0, ... ,0) for some p > O. But then Y(T) = Yo(T) = @, while
y(T) = - P < 0 = Yo(T) contradicting minimality.] Thus, there is a unit vec-
tor N orthogonal to a d-dimensional subspace ff for which
N·X~O, v X E %, with N = (n, N) and n ~ o. (20)
/RxY
%
"-\\ Edge
\ view
\
\
yz
Figure 11.3
(When == 0, we can take N = (1, 0, ... ,0)) ([N]). In case d = 2, (20) guaran-
tees that each vector in the cone is within an angle of nl2 of the fixed
upward-slanted vector N, as illustrated in Figure 11.3.
(11.7) For the more general <1>, the requirement that fMI = <l>y(Yo(T)) (= <l>y({D))
have maximal rank I ::;; d means that locally (through implicit function theory
[Ed]), Yo the {D-Ievel set of through {D can be represented as a (smooth)
r (= d - I)-dimensional submanifold in /Rd. Moreover, !/' has an associated
r-dimensional tangent space !T which is orthogonal to the subspace spanned
by the I linearly independent (column) vectors of fMI. A similar argument to
that given above shows that :f{" cannot contain in its interior any downward
directed tangent vector Xo E /R X fT.l We conclude that there is again a
vector N = (n, N) with N orthogonal to /R x !T for which (20) holds. But
then N = Ay(Yo(T)) for some A E /R I.
In Figure 11.3, we illustrate the latter situation for d = 2, where the target
surface is defined by the single equation CP(Yl, Y2) = O. Maximal rank at (D
means that Vcp({D) l' {D and as we know say, by the discussion at the end of
§5.6, this vector is orthogonal to the r (= I)-dimensional subspace !T (a line)
tangent to Yo the O-level set at {D. Since the interior of :f{" cannot contain
vectors in the "lower" half of the plane /R x !T, then (20) holds for some
N = (n, N), where N is orthogonal to !T and thus is given by some AVcp({D).
1 For if so, then ff also contains a small downward directed target vector Y,(T) which lies in
[J;1! x 9", and this is clearly inadmissible. (Why?)
It is now very easy to complete the derivation of the minimal inequality

(19'). First, we observe that the linear homogeneous adjoint equation (19)
always has a unique solution over [0, T] on successive subintervals of conti-
nuity of Uo with prescribed target value P(T) = N as obtained above. But,
then by Proposition 11.2, we know that at each point of continuity 1: > 0 of
Uo, and for any V E 11/1,
P(1:)~G(1:, V) = P(T)Wt~G(1:, V)
= N·"Y (by (16'»
~ 0, since l ' E :%,
so that the result follows from definition (12). D
Now let's see what these results tell us about our original autono-
II}ous fixed interval problem. Note that in its Mayer formulation, when
Y= G(Y, U) only, then with P = (p, P), we must have from (19) that
P·=-G·P=O
y ,
so that p(t) = const. = p(T) = n, the first component of N, which by our

arguments cannot be slanted downard. Thus p(t) = ..10 ~ 0, and this means
that either ..10 = 0, which although atypical can occur;1 or by division and
redefinition of P, and A as required, we can have ..10 = 1. Moreover, we
always have (..1 0, P(T)) = N =1= @E /Rd+1, so that for the free target case
(<1> == lD), where P(T) = lD, we must take ..10 = 1. Let's summarize our findings.
(11.8) Theorem. Suppose that we have an autonomous control problem on the

fixed interval [0, T] for performance and state functions f and G, respectively,
with a target that is either fixed or free. If (Yo, Uo) is a minimizing solution
then there is a Pontjragin function
h(t, Y, U) ~ Aof(Y, U) + P(t)· G(Y, U) (21a)
that obeys the minimum principle
v, U EI1/1, (21b)
and makes ho constant.
P is a C1 solution of the adjoint equation
P= -h y (·, Yo, Uo ) on [0, T], (21c)
for which (..1 0, P(T)) =1= lD . ..10 is 0 or 1, with ..10 = 1 and P(T) = lD in the free
target case.
PROOF. The constancy of ho is a consequence of more general results to be
established in part (c). (See Remarks 11.11.) Therefore (21b) which holds at
1 See the first example in §10.2(a). Such problems are referred to as being abnormal because their
optimal criteria do not depend on the performance measure being optimized.
each point t of continuity of Uo by (19') also holds at a point of discontinuity,

in that it is satisfied by either limiting value of h(t, Yo(t), U). D
(11.9) Remarks. 1. If fixed (or more general) target values are defined by
requiring that (Y(T)) = (!) for some C l I-vector valued function for which
the matrix <l>y(Yo(T)) has maximal rank I ::;; d, then by (19),
P(T) = Ay(Yo(T)) for some A E ~I. (21d)
Maximal rank is essential for this conclusion. See Example 1 below.
2. The Lagrange multiplier theorem (5.16) admits similar formulation for
Ai = Pi' with
Then alternative (a) permits Ao = 0, while if (a) is not satisfied, we must take
Ao 1= 0, and, if desired, Ao = 1.
3. Observe that under classical smoothness conditions, we have just es-
tablished a Lagrange multiplier rule for Lagrangian constraints of the form
Y = G(Y, U). It guarantees the Weierstrass condition, but for the modified
integrand j = Ao! + p. (G - Y) (where now U is to be thought of as U). The
more general situation is discussed in §11.3.
Example 1. As an illustration, recall the linear-quadratic problem at the end

of §10.2(a) where we found that yo(t) = uo(t) = e t minimizes
F(y, u) = !a [y2(t) + u2(t)] dt

l
uniquely on
!'!) = {(y, u) E ClEO, 1]: y(O) = 1, y(1) = e}
under y = u, with lui::;; 4,
Thus, for some P E Cl [0, 1] and Ao (= ° or 1),
h(t, y, u) = Ao(y2 + u2) + p(t)u
obeys the minimum principle for uo(t) = e t ; hence,
v lui::;; 4:
Since letl < 4 for t E [0, 1], it follows that
so that
hu(t, yo(t), e t ) = ° or 2A ouo(t) + p(t) = 0,
p(t) = - 2A ouo(t) = - 2A oe t and p(1) = - 2A o.
Since (Ao, p(1)) 1= (0, 0), we must take Ao = 1, even though this is a fixed
target problem. If we use cp(y) = y - e to fix the target, we can take A = -2
and satisfy (21d). However, if we use cp(y) = (y - e)2 to fix the target, then
CPy(Yo(1)) = 2(Yo(1) - e) = 0, and we cannot find a suitable A because the
maximal rank condition is not satisfied.
Note that
ho(t) = (e t ? + (e t )2 - 2e t et = 0.
Example 2 (Oscillator Energy Problem). At the end of §10.2(f), we considered

how to maximize the energy of an oscillator that can be reached from a state
of rest in a specified time T. If we set
11/1 = [-1,1], Y = (y, yd,
f(Y, u) = -Y1U,
and
G(Y, u) = (Yl, U - Y),
then Theorem 11.8 is applicable to each
(Yo, uo) E [!) = {(Y, u) E Cl [0, T] x CEO, T]: Y(O) = 0, u(t) E 11/1}
that minimizes
F(Y, u) = IT - Yl (t)u(t) dt on [!),
under the state equation Y = G(Y, u).

For this free-target problem, we must take AO = 1 and, for each
P = (p, Pl) E Cl [0, T],
we have
h(t, Y, u) = - Yl U + P(t)Yl + Pl (t)(u - y)
= P(t)Yl + Pl(t)y - U(Yl - Pl(t»
= l(t, Y) - U(Yl - Pl (t», say. (22)
In our earlier analysis, we found that
Yo(t) = Ito {sin«t

cost-r
- r»} uo(r) dr, (23)
if we recall equations (14) and (16) in §1O.2(f).
Similarly, we know that P satisfies the adjoint equation(s)

P= -hy[Yo, uo] = Pl'
(24)
Pl = - hy,[Yo, uo] = u o - p,
which are seen to be identical to those satisfied by Yo. Since P(T) = (!), we
conclude that
P(t) = -iT t
{sin«t - r»}uo(r) dr.
cos t - r
(25)
To determine possible optimal controls, we invoke the minimum princi-

ple. When we examine the dependence on u of
h(t, Yo(t), u) = l(t, Yo(t» - U(Yl (t) - Pl (t»,
we see that in order that uo(t) minimize this expression when lui ~ 1, we
should choose
uo(t) = sgn(YI (t) - PI (t». (Why?)
Thus, from (23) and (25), we find that
uo(t) = sgn (LT cos(t - r)uo(r) dr)

= sgn(cos(t - a», (26)
where cos a = c/Jc 2 + S2, sin a = s/Jc 2 + S2, if
~} = LT {:~::}uo(r) dr
are not both zero. If we exclude the non optimal case U o == 0, then we must
find a if possible that makes these equations for U o compatible. We see that
the optimal control is of the bang-bang type; in fact, it is constant, + 1 or
-1, on successive intervals of length n, exactly as our preliminary analysis in
§1O.2(f) suggested.
For example, when T = 2n, we can ask whether for a = n/2
U
o= {+
-1
1 on (0, n),
on (n, 2n),
could be optimal?
F or this we need
L2" cos(t - r)uo(r) dr = L" cos(t - r) dr - J,,2" cos(t - r) dr
= 2 sin t + 2 sin t = 4 sin t,

to have the sign of uo(t) on intervals of constancy. It does, and hence this
control could be optimal when T = 2n, or when T = n. (Why?)
On the other hand, this same U o cannot be optimal when T = 3n/2 since
J(3"1
0
2
cos(t - r)uo(r) dr = 2 sin t - "
f 3"12
cos(t - r) dr
= 3 sin t + cos t
which equals -1 when t = n, and so by continuity, is negative in a neighbor-
hood of t = n, while uo(t) = + 1, for t E (0, n).
From Theorem 11.8, we also know that ho(t) = h(t, Yo(t), Uo(t» is con-
stant. In particular, ho(O) = ho(T), and since Yo(O) = P(T) = (!), from (22), we
see that
or from (23) and (25), that
uo(O +) IT (cos r)uo(r) dr = uo(T-) IT cos(T - r)uo(r) dr. (27)
This equation provides a simple test which possible optimal controls must
meet. For example, when T = n/2, the control
{ +1 on (0, n/6),
u= -Ion (n/6, n/2),
cannot be optimal since fO/2 cos ru(r) dr = 0, while the corresponding inte-
gral on the right side of (27) is non vanishing. On the other hand, for T = n,
this condition becomes
uo(O+) t' cos ruo(r) dr = -uo(n-) t' cos ruo(r) dr,
so that either the integral vanishes (which it does when U o is constant) or

uo(O + ) = - uo(n -), and this is true of all other candidates permitted by
uo(t) = sgn cos(t - IX). Thus all of these uo(t) are still candidates for opti-
mality (and for the latter, ho(t) # 0).
However, for T = 5n/4, we can have uo(O + ) = ± uo(5n/4 -), (why?), so

that the corresponding integrals in (27) must be checked for both types of
possible U o. Supposing that initially U o = + 1, then after tedious computa-
tions outlined in Problem 11.2, we find that only one candidate of each type
survives the test, and upon comparing the associated energies, that only one
of these, namely
+1
{ -1 on (0, u), h
U o= (
on u,5n ,
/4) where tan u = -(1 + y'2), (27')
could be optimal.
Unfortunately, without either a sufficiency theorem, or an existence theo-
rem, we still cannot guarantee the optimality of this sole survivor.
(c) General Control Problems
In Theorem 11.8, the interval [0, T] is fixed, which precludes direct applica-
tion to say, time optimal problems. However, it is straightforward to use
the theorem (and subsequent remarks) to obtain an extension to the non-
autonomous case where T also varies, and thereby strengthen the previous
results. It is less straightforward to present the new results.
(11.10) Theorem (Necessity of the Minimum Principle). Hypotheses. Let D be

a bounded domain of [Rd and d/t be a bounded set in [Rk. Suppose that
F(Y, U, T) ~ IT f(t, Y(t), U(t)) dt
is minimized by (Yo, Uo, To) on

(Y, U) E C1 [0,
T] x CEO, T], for T> O;}
!'!)", ~ { with Y(O) = Yo(O), (T, Y(T)) = (9 E [R';
and (Y(t), U(t)) E D x d/t
Y(t) = G(t, Y(t), U(t)) on (0, T). (28a)
Assume that f, i(t,y) and G, G(t,y) are continuous on [0, b] x D x d/t for some
b> To, while either == (9, or is a C 1 function in a neighborhood of Bo ~
(To, Yo(To)) E [Rd+l such that (t,y)(Bo) has maximal rank I ::::;; d + 1.
Conclusions. Then there exist (A,o, A) =F (9 in
P E C1 [0, To], with P(To) = Ay(Bo) such that
[Rl+l with Ao = ° or 1, and a
P = -hy[Yo, Uo] on (0, To), (28b)

where
h(t, Y, U) ~ Aof(t, Y, U) + P(t)· G(t, Y, U),
obeys the minimum principle that VUE d/t
h(t, Yo(t), U) ~ ho(t) ~ h(t, Yo(t), Uo(t), t E [0, To]. (29)
Moreover, ho E C1 [0, To], with ho(To) = - A, <l>t(Bo), while
- ho = Ao.t;[Yo, Uo] + p. Gt[Yo, Uo], (30)

and
(30')
PROOF. Observe that when = (cp, iii) with cp(t, Y) = t - To, and the prob-
lem is specialized to that of the autonomous fixed interval case (T = To)
considered previously, then the initial conclusions are just a statement of the
results already obtained at each point of continuity of Uo. Hence we can
assume that this part of the theorem is valid under these restrictions, with
proper attention to dimensionality.
We will obtain the new results by reduction to this case on [0, 1], through
the parametric substitutions
t = t(r); X(r) = Y(t(r», V(r) = U(t(r)).
Here t is regarded as a new state variable governed by the law t'(r) = w,
where w is to be considered as a new nonnegative control in CEO, 1]. Through
the chain rule, (28a) is replaced by
X'(r) = G(t(r), X(r), V(r))w(r) (31)
and it is seen that if to(r) = rTo (so that Wo = To), with corresponding sub-
stitutions for X o, Yo, then (to, Xo; wo, Yo) minimizes
F(t, X; w, V)~ I1 f(t(r), X(r), V(r))w(r) dr (= F(Y, U, T), if T = t(l))

on
[f)= {(t, X; w, V) E C1 [0, 1] x C[O, 1], with t(O) = 0, X(O) = Yo(O) E IR d, }
II> (t(l), X(l)) = (9 E IRI, and (w, V)(r) E [0, b] x o/.t '
under the state laws t'(r) = w, and (31). Observe that the transformed prob-
lem is autonomous as desired, since its defining functions do not depend
explicitly on r.
°
Hence by Theorem 11.8 and Remarks 11.9, with appropriate substitutions,
3 A.o = or 1, A E IRI, and functions (qo, Q) E C1 [0, 1] with
(A. o; qo(1), Q(l)) "# (9, (32a)
such that
qo(1) = A . <l>t(To, Xo(l)); Q(l) = Ax(To, Xo(1))
(since t o(1) = To); and
q~(.) = -hk, to: X o, Vo)To, Q'{-) = -h y (·, to, X o, Vo)To, (32b)
where
h(r, t, X, V)w ~ [A.of(t, X, V) + Q(r)· G(t, X, V) + qo(r) ·l]w,
obeys the minimum principle on [0, 1] for the control pair (w, V), relative to
the trajectory to(r) = rTo and Xo(r) = Yo(rTo), at each point r of continuity
of Yo.
In particular, if we take V = Vo(r) = U(rTo), and consider both w(r) =

T ~ To and w(r) = T::;; To in the resulting inequality, it follows that at each
point of continuity, h(r, rTo, Xo(r), Vo(r)) = 0, (why?), or, with
t = rTo and P(t) ~ Q(t/To), (32c)
that ho(t) = - qo(t/To) is continuous on [0, To], where ho is as defined in (29).
Similarly, if we take w(r) = To and consider the minimum inequality for V
compared with Vo(r), then we see that after cancellation, and resubstitutions,
h as defined in (29) obeys the desired minimum principle relative to this same
h o. (Problem 11.3.)
Finally, from (32b) and (32c) follow the equations (28b) and (30); the fact
that (A.o, ho(To), P(To)) "# (9 is an immediate consequence of (32a) and this in
turn would be violated were (A.o, A) = (9. (Why?) 0
(11.11) Remarks. 1. Equation (30) may be written

ho(t) =- Ht(t, Yo(t), Uo(t), P(t)), (33)
if h is replaced by H(t, Y, U, P) as in the final part of §10A. When the problem
is autonomous so that it == 0, Gt == (9, then (30) implies that ho is constant on
[0, To]. If, in addition, <l>t == m, then we can conclude that ho = ho(To) = 0, so
that (Ao, P(To )) =f. m. However, for the fixed interval case which requires say
 = (<p, <ill with <p(t, Y) = t - To, then <l>t = (1, <ilt) =f. m, and we know that
for such problems ho need not vanish (Example 2). Observe that all conclu-
sions about the behavior of ho were obtained under the assumption that the
various integrands are defined on some time interval (0, b) ;2 (0, To]. For
certain nonautonomous fixed interval problems, this need not be valid. If,
for example, f(t, Y, U) = JTo2 - t 2j(Y, U), then more general methods are
required to obtain analogous results. See [He].
2. This theorem required boundedness of the control set 1111. When 1111
can in addition be described by Lagrangian inequalities of the form 'P(U) :::;;
mE IRm, then under sufficient smoothness assumptions, we can replace the
minimum principle as stated by iiu(t, Yo(t), Uo(t)) = m, where ii = h + M· 'II
for suitable multiplier functions M on [0, To]. Recall that a similar suffi-
ciency formulation was made in §10.1 and subsequently used to attack sev-
eral of the problems in §10.2. This extension, as a necessary condition for
optimality, will be established in §11.3.
3. In this theorem, the optimal target time To could be any positive num-
ber, and this might not be realistic in certain applications. Problems in which
To is required to lie in some given interval are explored in [S-S].
Example 3. Suppose that for Y = (y, YI) and given B = (b, bl ) E 1R2, we wish
to minimize
F(Y, u, T) = IT (1 + yi(t)) dt
on
q} = {(Y, u) E CI[O, T] x C[O, T) with T> O:}
Y(O) = m, Y(T) = B, lu(t)1 :::;; 1 '
under the state law

y = Cy, )\) = (YI, u) on (0, T). (34a)
(A possible physical setting for this problem is presented in the discussion

of the docking problem in the next section.)
Here, we have an autonomous fixed-target problem with a free terminal
time and we can take (Y) = Y - B. If we suppose that (Yo, uo, To) supplies
an optimal solution, then we may apply Theorem 11.10; accordingly, we
introduce Ao (= 0 or 1) with P = (p, pd E Cl [0, To], and form
h(t, Y, u) = Ao(1 + yi) + P(t)YI + PI (t)u.
Then on (0, To), P satisfies the adjoint equations, abbreviated as follows:
- P = hy = 0 => p(t) = c, const.,
(34b)
-PI = hy, = 2Ao Y! + p = 2A oYI + c.
Also, after simplification (p(t) = c)

h(t, Y, u) = -to(1 + yi} + CYI + PI (t)u, (35)
satisfies the minimum principle which here reduces to the requirement that
PI (t)u ~ PI (t)uo(t), If lui :5; 1,
and this demands that
uo(t) = - sgn PI (t). (35')
Moreover, ho(t) = h(t, Yo(t), uo(t» is C1, with ho = 0 by autonomy; and
ho(T) = 0, since c))t = (1), so that ho(t) == o. (See Remarks 11.11.)
Now, if -to = 0, then, since YI(O) = 0,

0= ho(O) = PI (O)uo(O + ),
and by (34b), PI = - c or PI (t) = C 1 - ct which can change sign at most once,
at r = cdc. Thus by (35'), uo(t) is constant, either + 1 or -1 in each interval
excluding r. But then, PI (0) = 0 => C I = 0 => r = O. Hence, in this case, no
switching is permitted and we must take uo(t) == + 1 or uo(t) == -1. However,
this implies through the easily integrated state equations (34a), that To = Ibll,
while
either b = bf!2 and bl > 0,
(36)
or b = - bf!2 and bl < O.
For other targets B, we must consider -to = 1. Then (34b) gives -PI =
2YI + c (=> PI is continuous) which combined with (34a) yields -PI = 2YI =
2uo, or finally by (35'),
(37)
On each interval in which PI is of constant sign l1, this equation has only
parabolic solutions of the form PI (t) = l1t 2 + at + a l . Moreover, the parab-
olas permitted cannot vanish twice unless PI vanishes in the intervening
interval. (Why?) Since PI is continuous, it follows that the only possible PI are
those shown in Figure 11.4 where a possibly degenerate zero interval [r, r']
is flanked by tangential semiparabolas opening in either direction.
/
/
" Uo =-1 ./ Uo = 0
...... .;'
't 't'
Uo = +1
Figure 11.4
Uo = + 1 or U o = -1 in each complementary interval. The period [T, T']

°
By (35'), such P1 correspond to optimal controls U o = on [T, T'], with
corresponds to one in which by (34a), Yt (t) is constant while y(t) varies

linearly. In the complementary interval (0, T),
y(t) = at 2 /2, l"
lor a = + 1 or y = ayd2.
2
(38)
Y1(t) = at, -
°
At a switching time T where U o changes from + 1 to or -1, we must have
° °
P1(T) = and P1(T) ~ (Why?) Thus from (34b) and (35):
ho(T) = °= 1 + Y~(T) + CYt(T),
(39)
-PI(T) = 2YI(T) + C ~ 0.
Moreover, having P1 negative on (0, T) corresponds to taking a = + 1 in (38)
so that Y1(T) = T > 0. It follows that
° ~ 1 + Y~(T) - 2Y~(T) => Y~(T) ~ 1. (39')
°
Thus such changes can occur only when < Y1 (T) ~ 1, and a similar analysis
shows that in the opposite case, where U o changes from -1 to or + 1, °
° < - Y1(T) ~ 1. It follows that switching occurs according to this prescrip-
tion or not at all. Moreover, Yt = ± 1 define the only lines of constancy for
an optimal trajectory in the Y-plane, and y(t) = ± t + Co only along such
lines.
We can use convexity to argue that for the given time To, there can be at
most one optimal control. For if we introduce Il as in §1O.1, we see that the
modified integrand
i<f, Y, u, Y) = 1 + y~ + P(f)(YI - y) + P1 W(u - yd + ll(f)(u 2 - 1)
is for Il ~ 0, convex on [0, To] x ~3, and strictly convex in Y1.
The Euler-Lagrange equations for j in Y and Y1 are just (34b), while that
for u is
°
.1u = = P1 + 2lluo,
which is automatically satisfied on intervals where U o = 0, and we can take
= 0 there. Otherwise, from (35'),
Il
21l = -pdu o = Ipil > 0,

and thus, ll(t)(uJ(t) - 1) == 0.
It follows that if (Yo, uo) E ~To' then it minimizes go
(1 + y~(t)) dt on ~To
among those satisfying (34a), and uniquely in Y1. But YI = U o so that U o
(hence Yo) must be uniquely determined.
Assembling these facts, leads us to form the state diagram of Figure 11.5,
where the primary parabolas corresponding to U o == + 1 and U o == -1, re-
spectively, are shown in bold lines.
The switching lines Y1 = ± 1 are dashed, and the remaining parabolas
are those which arrive at other targets with Uo = + 1 or -1. From the
§11.2. Linear Time-Optimal Problems 397
Uo = +1
- - - --:~---I
Uo = -1 /" B
-'
Figure 11.5
uniqueness of the optimal control-trajectory, and a careful analysis utilizing

the symmetry of those parabolas of different families passing through the
same point, it follows that the typical optimal trajectories to various target
regions are as indicated by the arrows.
For example, to reach Bl by a path utilizing the line Y1 = + 1, would
result in an integral for f&O yi(t) which is at least as large as that along
Yl = y2/2 and the time go
1 dt would clearly be larger. Thus the dotted
alternative is excluded.
For further discussion of this and related examples, see [M-S]. The
simpler case in which g
yi(t) dt only is minimized is taken up in Problem
11.1(h), where it is shown that the necessary conditions preclude optimality
for all targets B except those on the primary parabolas. The corresponding
time-optimal problem is investigated in the next section.
§11.2. Linear Time-Optimal Problems

In this section, we confront some problems in which the target time T is being
minimized and which are not replaced by equivalent problems on some
fixed interval. To avoid further complications we will look only at time-
optimal problems governed by linear state equations and consider in detail
only the autonomous case, although certain results have nonautonomous
(and nonlinear) extensions.
Even within this narrow range, the problems require special additional
conditions of a geometrical nature in order to guarantee the optimality or
uniqueness of a control which meets the necessary conditions of Theorem
11.10. The controls which emerge are in general bang-bang and piecewise
constant; with further restrictions, it is possible to limit the number of
switches in value which an optimal control can exhibit. These results are
applied to obtain a complete solution to a docking problem, and other
applications are discussed.
Problem Statement
We wish to find the minimal time T required to bring a system from a

given state in IRd at time t = 0 to the origin under the linear autonomous
state equation
Y{t) = AY(t) + IBU(t) on (0, T), (40)
where A and IB are constant matrices and U(t) E 1111 ~ IRk.
The state equations in all of our previous examples in Part III can be put
in the form (40), so t~at it still permits wide application.
Suppose that (Yo, Uo, To) gives a minimizing solution to this problem. To
apply Theorem 11.10 we should take f = 1, AO = 0 or 1, and form
h(t, Y, U) = Aol + P(t)·(AY + IBU).
Then the adjoint equation (28b) takes the homogeneous form -1' = 1'A,
which has the general solution
(40')
for an appropriate square matrix function e- tA and row matrix v. l From (29)
we see that for each U E 1111:
h(t, Yo(t), U) = AO + Ve-tA(A Yo(t) + IBU)
~ ho(t) = AO + Ve-tA(AYo(t) + IBUo(t»,
or
(41)
Also ho = 0 (why?), so that by (30'),(AO, V) =f. (!).
For this linear autonomous case, there are additional conditions under
which the minimum principle in the form (41) is effectively sufficient.
(11.12) Theorem. Suppose that the origin is an interior point of the control set
1111 in IRk and that the matrix tMI = (IB: AlB: A21B: ... : Ad-lIB) has rank d. If for
some vector V E IRd '" (!), Ve-tAIBU ~ Ve-tAIBUo(t), t E [0, To], VUE 1111, then
1 Formally
where the indicated series of matrices converges uniformly in t on each compact interval. Also
and etA are inverse matrices [C-L].
e- tA
(Yo, Uo, To) minimizes g dt under (40) on

q}o = {(Y, U, T) E Cl[O, T] x C[O, T] x (0, (0);
Y(O) = Yo (0); Y(T) = (9; U(t) E 1JIt}.
PROOF. With the same matrix e- tA used above, the state equation may be
rewritten (d/dt)(e-tAY(t» = e-tAIBU(t) so that upon integration over [0, T],
with Y(T) = (9 (and e- OA = ~d)' we obtain
(9 = Y(O) + !aT e-tAIBU(t) dt, (42a)
and similarly, when Y(O) = Yo (0):

(To (T
- Yo(O) = Jo e-tAIBUo(t) dt = Jo e-tAIBU(t) dt. (42b)
Now suppose that this admissible control U, produces a state trajectory Y

with Y(O) = Yo(O) and Y(T) = 0 in the time T < To. To reach a contradiction
we utilize our hypotheses as follows:
Since U = (9 E 1JIt, we see from (41) that
V t; (43)
hence from (42b), we have upon premultiplication by the constant (row)

vector V, that
-V· Yo (0) = {!aT + f:O}(Ve-tAlBuo(t»dt
~ !aT Ve-tAIBU(t) dt = - V· Yo(O). (Why?)
Then equality holds throughout so that in particular,
0= f TO
T Ve-tAIBUo(t) dt,
and by (43),
Ve-tAIBUo(t) == 0 on [T, Tol
Using (41) again, we see that U E IJIt => Ve-tAIBU ~ 0, and where also - U E 1JIt,
we get Ve-tAIBU == 0 on [T, Tol Since (9 is an interior point oflJlt, this must
hold for all U near (9 so that the row vector Ve-tAIB == (9 on [T, Tol By
translation, with tl = (T + To)/2,
c))(t) ~ Ve- t1 Ae- t AlB,
= VI e-tAIB == (9, near t = O.

In particular, C))(O) = c))'(0) = ... = c))(d-l)(O) = 0, and since
[~e-tA
dt
= _Ae- tA ]
t=O
= -A,
upon successive differentiation, we see that Vl = Ve- t, A is orthogonal to the

columns of IB, AlB, A 2 1B, ... , Ad-lIB, and hence to the columns of MI, hypoth-
esized to be of rank d. It follows that Vl = (!) => V = Vl e tl A = (!), and this is
the desired contradiction. 0
(11.13) Remarks. The condition on MI will be discussed after Corollary 11.14.

The same techniques are effective for certain nonlinear state equations of the
simple form Y{t) = A Y(t) + G(V(t», when G((!) = (!). Details and applica-
tions are given in Problems 11.4 and 11.5.
Example 1. In 1R2, consider the state equation Y = V where A = 0 and

IB = ~, so that MI =" [IB : 0: : OJ = D : 0: : OJ has rank 2. If we take
1111 = {V E 1R2: Iud::; 1, i = 1, 2} then for each nonzero V = (Vl' v2 ), and
V E 1111:
Ve-tAIBV = V· V = V1U l + V2 U2 ~ -IVll-lv21.
For instance, when V = (1,0), each control of the form Vo(t) = (-1, u2 (t»,
with its associated state Yo(t) = (To - t, - ITo u2(r) dt) on [0, ToJ, is optimal
among these producing state trajectories Y, with Y(To) = (!), and Y(O) = Yo (0).
In particular, Vo(t) = (-1,0) is optimal among those on [0, 2J for which

Y(O) = (2, 0), but it is not unique in this respect. (Problem 11.6.) Under
slightly more restrictive conditions, we can obtain a much more attractive
result.
(11.14) Corollary. Suppose that 1111 of Theorem 11.12 is a closed box (or a
compact convex polyhedron) in IRk with vertices ~. If the vectors IBE, AIBE, ... ,
Ad-1IBE are linearly independent for each edge vector E joining adjacent
vertices, then the optimal control Vo(t) is unique, and piecewise vertex-valued
on 1111.
PROOF*. For each t E [0, ToJ, Ve-tAIBV is linear in V and continuous on 1111.
Hence by Proposition 5.3 it is minimized at some Vo(t) E 1111, and for the
hypothesized shapes of 1111, the minimum value must occur at some vertex ~.
(See Problem 11.7.) If this same minimum value occurs at another vertex, it
must do so at an adjacent vertex Vi so that by subtraction, Ve-tAIBE = 0 for
the edge vector E = Vi - ~.
Suppose this nonuniqueness occurs for more than a finite set of values of
t = tn' Then since the number of edges is finite we can assume that all tn
are associated with the same edge E, i.e., that q>(tn) ~f Ve- tnA IBE = 0, 'V n =
1, 2, .... From this it follows that the function q>(t) = Ve-tAIBE, which is
real analytic in t, vanishes identically [Rud]. But if q>(t) = Ve-tAIBE == 0,
§11.2. Linear Time Optimal Problems 401
then by differentiating successively and evaluating at t = 0, we find as

before that VIBE = VAIBE = ... = VAd-1IBE = 0. Thus, if V = (Vb· .. , Vd),
then Lt=l viAi-1lBE = 0, and by the hypothesized linear independence, we
must have V = (!), a contradiction. Therefore, for all but a finite set of
t E [0, To]; Vo(t) = ~, for a unique vertex ~.
If tj E J, a maximal open subinterval of (0, To) that excludes these
exceptional values and V(tj) = ~, then for Vi -# ~, we see from (41) that
Ve-tjAIBVi > Ve-tjAIB~; by continuity, this strict inequality holds for all t
near tj, so that Vo(t) = ~ for all such t. (Why?) It follows that in the open
interval J, Vo(t) is locally constant, hence constant (= ~), since possible
points of discontinuity have been precluded. This provides the desired
description of an optimizing control Vo.
If Vet) is another optimizing control on [0, T], then obviously T = To, so
that by (42b) we have
JorTo Ve-tAIB(V(t) - Vo(t)) dt = 0;
by (41), the integrand is nonnegative. From the standard argument we

conclude that Ve-tAIBV(t) = Ve-tAIBVo(t), but as we have just established,
this is possible only for that Vo described above, i.e., Vet) = Vo(t). 0
Remarks. This result shows that for such problems, an optimal control
should be of the bang-bang type since it is mostly vertex-valued. If we
replace d/t by {V: IVI ~ 1} then such distinguished vertices are not present,
and we have already considered related problems in §10.2(g).
The matrix M of Theorem 11.12 is of rank d if the edge-vector condition

of this corollary is satisfied by any E -# (!), but not conversely. (See the previ-
ous example.) Similar results hold in the nonautonomous linear-state case
where the given matrices A and IB are continuous functions of t, but the
required condition generalizing that for E is less geometrical and more
awkward to verify. (See [BerJ) Through an even less accessible condition,
Diliberto [TO] uses the methods of §11.1 to obtain a bang-bang principle
for control sets d/t which are nonconvex polytopes.
Example 2 (A Free-Space Docking Problem). An ion propulsion vehicle in

free space equipped with forward and reverse thrust engines has given posi-
tion and velocity at time 0, and it is to be brought to rest in minimum time
at a target position (taken to be the origin.)
If we assume rectilinear motion only and neglect gravitational and fuel
mass-loss effects, then its position yet) at time t is governed by Newton's
second law: yet) = u(t), where mu(t) is the net thrust produced by the engines,
and m is the mass of the vehicle (assumed constant). By normalization we can
suppose lu(t)1 ~ 1.
If we set Y(t) = (y(t), Yl (t)) where Yl (t) is the velocity at time t, then the
state vector Y = (y, yd E ~2 obeys the linear law
Y(t) = (Yl (t), u(t)) = A Y(t) + lBu(t),
where
A= [~ ~J and IB = I~I so that AlB = I~I,

and if ott = [ -1,1], we can apply the preceding results. Here E = ±2 E ~l
provide the only edge vectors, and the vectors {IBE, AIBE} = ±2{(0, 1), (1, O)}
are linearly independent in ~2. Thus by Corollary 11.14 we expect an optimal
control uo(t) on [0, To] to be bang-bang, in fact to have only "vertex" values
of + 1 and -1 except at a finite set of points. If present, it is unique.
Since A2 = {} (why?), then e- tA = n - tA, only, = [~ ~tll By Theo-

rem 11.12, any Uo E CEO, To] is time-optimal if we can find some V =
(v, vd -# (!), such that V lui ~ 1:
Ve-tl\lBu = V·(-t, 1)u;?: V·(-t, 1)u o(t)
or
(44)
The linear (switching) function a(t) = V 1 - tv can vanish at most once on any
interval, and otherwise, its sign determines uo(t) = + 1 or -1. (Why?)
All that remains is to show that by suitable choice of a switching time
1', we can indeed reach the origin from the given initial state A E ~2, using
only those controls of the form
a, t < 1',
uo(t) = {
-a, t > 1',
where for each A, a = ± 1. The time To required to do so is then minimal,
and the associated control is unique.
Under the simple dynamics permitted here it is clear by time reversal that
this is equivalent to asking whether we can travel from the origin to a
given state A by such controls.
Obviously, we can do so for all states in ~2 where uo(t) == + 1, i.e., for all
states A which lie along the upper parabolic arc with parametric equations
{ Y(t) = t /2,
2 2
( ) (so Y = yd2, Yl > 0),
Yl t = t,
or along the lower parabolic arc Y = - yil2, Yl < 0, obtained when uo(t) =
-1, both shown in Figure 11.6. The other parabolic arcs shown in this figure
are those obtained by reversing the thrust direction at various points along
1 See the footnote on page 398.

Figure 11.6
the principal pair. On geometrical grounds, it is evident that each point

A -# (!) can be joined to the origin by one (and only one) such combined
curve. Moreover using analytic geometry, we can obtain the (unique) optimal
time To associated with each A. The details are given in Problem 11.8. The
unique optimal control will be the negative of that used to reach A.
As this example shows, it is of considerable value to know how many

switching times an optimal control can have. In general, this is a difficult
problem. However, suppose that o/J is a closed box of the form
o/J = {U E ~k: -mj ~ uj ~ mj , with mj > O,j = 1,2, ... , k},
whose vertices are just the points where each uj = ± mj , j = 1, 2, ... , k.
Then the edge condition of Corollary 11.14 is equivalent to requiring that
for each column Bj of IB, the vectors B.e. ISI.Bj , ... , ISI. d - 1 Bj are linearly indepen-
dent (Problem 11.9). Moreover, with P(t) = Ve- tA as in (40'),
k k
P(t)IBU = L (P(t)· B)uj = L O"j(t)Uj,
j=l j=l
(45)
where O"j(t) = P(t)· Bj is the switching function for the jth component of Uo(t),
so-called because as (45) shows, if O"j(t) -# 0 then
O"j(t)Uj ~ -mj sgn O"it) = O"j(t)(Uo)it),
which means that the jth component of Uo is
-mj , when O"it) > 0,
{
+ mj , when O"j(t) < o.
When the eigenvalues of lSI. are all real, there can be at most d - 1 zeros of
each O"j (which is continuous), and so at most d - 1 switches in each compo-
nent of the optimal control Uo. Some details are given in Problem 11.10, and
generalizations are presented in [Le] and in [Y]. In particular, ott as above

can be replaced by any closed rectangular box. (But unless it has (I) as an
interior point, we cannot use Theorem 11.12, to guarantee optimality, and
without some other sufficiency argument-or an independent existence
proof-we cannot assert that we have found an optimal control.) Even for
the simple case considered here, existence theory is difficult. (See [M-S].)
Moreover unless A. has eigenvalues with only nonpositive real parts, there
need not be a solution. (Problem 11.14.)
The eigenvalues of A. are imaginary in the case of the corresponding

time-optimal problem for the oscillator of §10.2(f) governed by the state
equation
(y, )\) = (Yl, u - y),
Indeed, the matrix
A.=[_~ ~J
has eigenvalues A where
IA. - AnI = I =~ ~ AI = 0,
or A2 +1= 0, so that A = ± i.
Here again
~=I~I so that M=[~:A.~]=[~ ~J

has rank 2, and thus each minimizing control for h on [0: To] will be the
unique optimal control. However, the state-plane analysis is somewhat more
complicated. (See [Po].) For this problem, sign changes in uo(t) occur over
successive intervals of length n. Thus, the number of switches is not even
bounded, but will vary with the location of the initial state, and this is
characteristic of problems with nonreal eigenvalues. (Problem 11.10.)
§11.3. General Lagrangian Constraints

In addition to endpoint constraints, a problem in the variational calculus (or
in optimal control), may also be subjected to a Lagrangian constraint of the
form G(t, Y(t), Y(t)) = (I), or more generally, to a Lagrangian inequality such
as
G(t, Y(t), Y(t)) ~ (I) (46)
interpreted componentwise; (in control problems, Y is replaced by U).
When the Y terms do not appear explicitly, then such problems can often
be attacked through implicit function theory, as was indicated in the proof of
§11.3. General Lagrangian Constraints 405
Theorem 6.10. (See Problem 11.15.) Otherwise, the problem is considerably

more difficult.
In this section we will use the minimum principle of Theorem 11.10 in
conjunction with some more facts about cones in ~d (collected in §A.7)
and implicit function theory to obtain existence of multiplier functions asso-
ciated with minimizing in the presence of Lagrangian (in)equalities such as
(46). From our introductory discussions in §2.3, §7.4, and §10.1, we know that
such multipliers can be effective in producing sufficient conditions on fixed
interval problems. We now want to consider their necessity in a more general
setting.
We begin in part (a) by using Theorem 11.10 to extend the Pontjragin
principle to optimal control problems in which the control set o/J can be
described by C l inequalities. Then in part (b), we use these results to produce
multipliers for the general problem in the variational calculus governed by
Lagrangian (in)equalities of the type (46). Finally, in part (c), we indicate
further extensions that could apply to analogous problems in optimal con-
trol. This material represents a small selection of the results that can be found
in [He].
In addition to the results from §A. 7, we will need the following:
(11.15) Lemma. If G is a real m x n matrix with m linearly independent rows

and n columns!, then GG is invertible.
PROOF. Suppose that for some X E ~m, we have XGG = (!J. Then XGGX =
(X G)· (X G) = 0; i.e., X G = (!J. But since G is of maximal rank m, this is
possible iff X = (!J, and the invertibility of GG follows as in (0.13). 0
(a) Control Sets Described by Lagrangian Inequalities
We suppose that all terms are as defined in Theorem 11.10. Then from
Theorem A.28, we can easily get the following:
(11.16) Corollary. In Theorem 11.10, suppose that the control set o/J can be
described by Lagrangian inequalities 'P(V) ~ (!J E ~m, where 'P is C l and
'Pu(Vo(t)) has maximal row rank m while fu and Gu are continuous. Then there
exists a (unique) M = (Ill' 1l2' ... , Ilm) E CEO, ToJ, with M ~ (!J, for which
hu( ., Yo, V o) = (!J on [0, ToJ, where
h(t, Y, V) = h(t, Y, V) + M(t)· 'P(V). (47)
Moreover
i = 1,2, ... , m.
PROOF. The minimum principle (29) implies that for each t, h(t, Yo(t), V) is
minimized at Vo(t) under the constraint 'P(V) ~ (!J. By Theorem A.28, a
I In this case we will say that G has maximal row rank m (::s; n).
unique M ~ (f) and h exist as required to satisfy

huh Yo, V o) = huh Yo, V o) + M(' )'I'u(Vo) = (f)
(where the rows of'l'u are indexed by V), with J1.i(t)t/!i(VO(t» == 0, i = 1,2, ... ,
m. But since 'I'u(Vo('» is hypothesized to have maximal row rank then Qo('),
its product with its transpose, is invertible (by Lemma 11.15) and the inverse
is necessarily continuous with Vo; hence, M(') = - h u(', Yo, Vo)iJiu(Vo)Q 01 (.).
However, the matrices on the right are continuous except at the discontinui-
ties of V o, so that M E C[O, ToJ 0
Note that when a component t/!i is inactive at Vo(t) in that t/!i(VO(t» < 0,
then J1.i(t) = O. In §10.1 and in several of the applications in §1O.2, we have
seen how this extension affects the search for optimal solutions.
(b)* Variational Problems with Lagrangian Constraints

In this part, we use the optimal control results just obtained, to establish
the existence of appropriate multipliers for problems in the variational calcu-
lus with general Lagrangian constraints as mentioned in §6.7. To do so we
require some rather technical extensions of standard implicit function theory.
(These can be avoided in a simpler case taken up in Problem 11.16.)
In the next theorem and its corollary, all sets are in Euclidean spaces,
and all mappings are vector valued.
(11.17) Theorem. Suppose that :F = :F(T, V) is C 1 and that :Fy(T, V) is

invertible on a nonempty set So on which :F(T, V) = (f). If So ;2 {(T, (f): T E
50} for some set 50, then there exist Il > 0 and a (locally) unique C 1 function
F defined on a neighborhood ff of 50 such that :F(T, F(T» == (f), with
IF(T)I < Il, V T E 50.
PROOF. The local existence of F is a consequence of ordinary implicit function
theory [Ed], which guarantees that near each To E 50, the (f)-level set of :F
near (To, (f) can be represented by the graph of a unique C 1 function, F(', To),
say. When a compact part of 50 is covered by a finite number of such
neighborhoods, the relevant locally defined functions must agree in any com-
mon region of definition (why?) so that F is well defined in a neighborhood
of this part, and it is locally unique. By considering extensions to larger
compact sets, it follows that an F is defined and (locally) unique as
required. 0
Note: An analogous argument was used in proving Theorem 9.24, and

Figure 9.12 may help to visualize the construction.
(11.18) Corollary. If t§ = t§(X, V) is C 2 and t§u(X, V) has maximal row rank

on So, the nonempty (f)-level set of t§, then there exists a C 1 function G defined
on a neighborhood of So such that

~(X, G(X, U)) == (D; and when (X, U) E So: G(X, U) = U.
PROOF. Set T = (X, U), suppose the columns of ~u indexed by U, and
consider ff(T, V)~~(X, U + ~u(T)V), which is e 1 (since ~ is e 2 ). Also
ffy(T, V) = ~u(T)<ju(T) is invertible on So (by Lemma 11.15) and ff(T, (D) =
~(X, U) = (D on So, so that we can apply the theorem to obtain an F defined
in a neighborhood of So. Then
G(X, U) ~ U + ~u(X, U)F(X, U)
has the desired properties, since·
~(X, G(X, U)) = ~(X, U + ~u(X, U)F(X, U)) = ff(T, F(T)) == (D;
(X, U) E So => ~(X, U) = ff(T, (D) = (D,
but then ff(T, F(T)) = (D => F(T) = (D, by local uniqueness of F; i.e.,
G(X, U) = U as desired. D
By using the previous results, we can transform a problem in the varia-

tional calculus with Lagrangian constraints into a standard problem in con-
trol theory. (It may be helpful to first read Problem 11.16.) To facilitate later
comparison, we will employ notation which differs slightly from that used in
previous chapters to formulate the problem.
(11.19) Suppose that Yo E C1 [0, To] minimizes
ff(Y, T) ~ IT I(t, Y(t), Y(t)) dt
on
q} _ { c
Y E 1 [0, T]: T> 0, with y(o) = Yo(O), }
- (Y(t), Y(t)) E D, and (T, Y(T)) = (D E ~l '
under the Lagrangian constraint
~(t, Y(t), Y(t)) == (D on (0, T).
Here D is a domain in ~2d while I = At, Y, Z), Iy and Iz are assumed
continuous, and = (t, Y) is e 1 while (t,y)(To, Yo(To)) has maximal row
rank 1 ::;; d + 1. Similarly, we suppose that ~ = ~(t, Y, U) is e 2 , and that
~u(t, Y, U) has maximal row rank m ::;; d whenever ~(t, Y, U) = (D, in a neigh-
borhood So of/{fo, the trajectory of Yo in ~2d+l. According to Corollary 11.17,
with X = (t, Y), there exists a e 1 function G(t, Y, U) defined in a neighbor-
hood of So, such that
~(t, Y, G(t, Y, U)) == (D E ~m, while G(t, Y, U) = U on So. (48)
Next, let
f(t, Y, U) ~ At, Y, G(t, Y, U)). (49)
We claim that with Uo = Yo, the triple (Yo, Uo, To) minimizes
F(Y, U, T) = IT f(t, Y(t), U(t)) dt,

on
§j = {(Y, U) E C1 [0, T] x CEO, T]; T> 0, with Y(O) = Yo(O);}
U(t) E OU(t), <D(T, Y(T)) = @ E ~l,
under the state law

Y= G(t, Y, U), (50)
where
OU(t) is a bounded neighborhood of Yo(t) such that
(Yo(t), U(t)) E D, when U(t) E OU(t).
Indeed, if (Y, U, T) E §j, then defining Y = U, gives (Y, T) E q} and if Y =
G(t, Y, U) then
F(Y, U, T) = IT I(t, Y, Y) dt = ff(Y, T) ~ ff(Yo, To),

since by (48),
f§(t, Y, Y) = f§(t, Y, G(t, Y, U)) == @.
However, before we can simply invoke Theorem 11.10, we must address

the complications which arise from the time-varying control set OU(t). Note
first that since Yo is C, we can find a fixed finite set of subintervals of [0, To]
on each of which OU(t) = OUj does not vary with t. Then if we reexamine the
proof of those results about propagation of vectors V; (Propositions 11.3-
11.5) we see that the arguments are not really affected by the finite set of OUj
along [0, T], since all we require is that the actual V; be chosen within the
relevant OUj • Moreover, the arguments about target values utilized specific V;,
not their local bounds. Thus we can conclude that Theorem 11.6, and (hence)
Theorem 11.10 remain vaid.
In particular, from the latter, we know that there exist P E C1 [0, To]
°
and ..1.0 (= or 1) for which
h(t, Y, U) ~ Aof(t, Y, U) + P(t)· G(t, Y, U) (51)
satisfies the minimum principle for admissible U; i.e., for t E [0, To]:
h(t, Yo(t), U) ~ ho(t) ~ h(t, Yo(t), Uo(t)),
whenever U E OU(t). Thus if we set

A(t, Y, U) = Ao/(t, Y, U) + P(t)· U,
so that (51')
A(t, Y, G(t, Y, U)) = h(t, Y, U),
we can say that A(t, Yo(t), U) ~ A(t, Yo(t), Yo(t)), VUE OU(t) such that
f§(t, Yo(t), U) = @, since by (48), this implies that G(t, Yo(t), U) = U.
Moreover, by hypothesis and construction, at each point of continuity t,

Au(t, Yo(t), U) will be continuous in U near Yo (t). Thus from Theorem A.28
and our assumption that tSu(t, Yo(t)·) has maximal row rank at Yo(t), (and so
in a U neighborhood), it follows that there exist unique multipliers M(t) such
that
Au(t, Yo(t), Yo(t» == (!J (52)
where
h(t, Y, U) ~ A(t, Y, U) + M(t)· tS(t, Y, U). (53)
Thus from (51) and (51'),
Aolu + P + MtSu = 0, (54)
when evaluated at (t, Yo(t), Yo (t».
However, again, since tSu has maximal row rank, we conclude from
Lemma 11.15 that 0, its product with its transpose, has an inverse which is
continuous with Yo; i.e.,
(55)
is continuous with Yo.
Also from (54),
P = - iu[Yo(-)]
if we introduce
i= Aol + M·tS, = A - p. U. (56)
Next, from (51') and (53) we have
h(t, Y, U) = A(t, Y, G(t, Y, U»,
since
tS(t, Y, G(t, Y, U» == (!J,
and from our original optimal control characterization (Theorem 11.10) it

follows that
-P(t) = [h y = Ay + AuGy] I(t, Yo, Yo)
= Ay(t, Yo, Yo) = iy[Yo(t)] (by (52».
Combining this with (56) and replacing U by Y, we see that except at corner
points,
i
i.e., Yo is stationary for in such intervals.
At corner points, h[Yo(·)] is continuous, (with P), while from the conti-
nuity of ho(t) = i[Yo(t)] - iy[Yo(t)]· Yo(t) follows the second Weierstrass-
Erdmann condition of §7.5.
From equation (30) for ho in Theorem 11.10, we can show that Yo also
I
satisfies the second Euler-Lagrange equation for Similarly, the minimum
i
principle gives the Weierstrass necessary condition on of Theorem 7.15.
(Problem 11.20.)
It remains only to obtain the appropriate transversality conditions on 1

By Theorem 11.10, there is a A E ~I such that (A.o, A) oF (!) for which
h[Yo(To)] = -P(To) = -Ay(To, Yo(To)),
while (57)
i[Yo(To)] - iy[Yo(To)]· Yo(To) = ho(To) = -A· <l>t(To, Yo(To))·
We have established the following major result:
(11.20) Theorem. Suppose that (Yo, To) is minimizing for the problem formu-
lated in (11.19). Then there exist A. o (= 0 or 1) and Lagrangian multi-
pliers ME C[O, To] such that Yo is stationary for i = A.ol + M· '§ except
possibly at its corner points. At these corner points, both iy[Yo(·)] and
(i - (iy)· Y)[Yo(-)] are continuous. Moreover, for i.
Yo satisfies the
Weierstrass condition (7.15), together with the second Euler-Lagrange equa-
tion, and it meets the transversal conditions (57), for an appropriate A, with
(A. o, A) oF (!). 0
(c) Extensions
(11.21) Various extensions of Theorem 11.20 are possible. For example,
suppose that instead of the fixed left end condition Y(O) = Yo(O), the
competing Yare required to satisfy a condition of the form <D(a, Y(a)) == (!) for
some C l vector valued function <D whose Jacobian matrix at (ao, Yo(a o)) has
maximal rank:$; d. The 1 defined as above on [ao, To] will satisfy transver-
sality conditions analogous to (57) at (ao, Yo (ao)) for an appropriate vector A.
It is also possible to consider minimization of a Bolza type functional in
which .'F is replaced by
.'F + qJo(T, Y(T)), where qJo is Cl.
Then the theorem holds as stated if A . is replaced by A.o qJo + A· in the
transversality conditions (57). (Details are indicated in Problem 11.22*.)
Corresponding problems with additional isoperimetric constraints may
be handled by replacing these with Lagrangian constraints in a higher-
dimensional vector space as in Proposition 9.11. The resulting multipliers
associated with such constraints will be constant as expected. A simple
example in discussed in Problem 11.23.
(11.22) Lagrangian inequalities of the form '§(t, Y(t), Y(t)) :$; (!), can be
treated with the device of slack variables introduced by Valentine [CCV],
and used in Problem 7.22*.
To accomplish this, we let Vi = wl, i = 1,2, ... , m, and for the new function
,§*(t, Y, Z, W) = ,§(t, Y, Z) +V (58)
we consider the problem of minimizing .'F as before on
!!}* = {(Y, W; T): (Y, T) E!!}, WE C[O, T]}
under the Lagrangian constraint

~*(t, Y(t), Y(t), W(t» == @. (59)
Observe that 'Vgr = ('Vg i, 2wiE i), where Ei is the unit vector in the ith direc-
tion in IRm, so that the new gradients are still linearly independent.
We consider W as though it were W in a problem in which W is not
present explicitly, but could be obtained from W by integration, with say,
W(O) = @.
If (Yo, To) minimizes ff on ~, under ~[Y(t)] ::;; @, then defining Vo(t) =
-~[Yo(t)], gives a Wo for which (Yo, Wo, To) minimizes ff on ~* under (59).
For any (Y, W, T) E ~* which satisfies (58), equation (59) gives a (Y, T) E ~
satisfying ~[Y(t)] ::;; @, since V ~ @.
Thus Theorem 11.20 is applicable, and we have multipliers ME C[O, T]
i
for which all conditions satisfied by are now satisfied by
I*~ Aol + M·~* = i + M· V, (60)
i
but with respect to both Yand W (as W), where = Aol + M . ~, as before,
is independent of W.
In particular, the first Weierstrass-Erdmann condition relative to W is
that I: be constant when evaluated along the minimizing trajectory. How-
ever, since Vi = wl, this means that 2J1.iWi = const., when so evaluated; more-
over, either J1.i == 0, or g;[Yo(t)] vanishes at some point along Yo, for otherwise
this constraint would remain inactive throughout. Thus, upon proper evalua-
tion we must have
J1.iWi == 0 or J1.iwl == 0, so that
J1.i(t)g;[YO(t)] == 0 on [0, Tol (61)
But, then the second equation for 1*, the Weierstrass condition for 1*, and
the second transversality condition (57) for 1* reduce to those for ialone; it
follows that for this new problem, Theorem 11.20, holds in its entirety and in
addition we have (61). Moreover, by a careful repetition of the transforma-
tion to the optimal control problem, we can conclude that M ~ @. (See
Problem 11.24.)
(11.23) Finally, there is an optimal control analogue for each of the above
extensions which can be obtained by applying the same techniques. Let's
consider only one such problem, that for which (Yo, Uo, To) minimizes
F(Y, U, T) = IT f(t, Y(t), U(t» dt

on
~ = {(Y, U, T): (Y, U) E c1 [0, T] x C[O, T]; T> O;}
Il> Y(O) = Yo (0), cD(T, Y(T» = @,
subject to the Lagrangian inequality
~(t, Y, U) ::;; @, (62a)
together with the state law

Y = G(t, Y, U), O:5;t::;; T. (62b)
Observe that (62a) permits description of state-dependent control sets
1111. Here, it is supposed that C§ and G are C 2 (although slightly weaker
hypotheses suffice), where C§u(t, Yo(t), Uo(t)) is of maximal rank m ::;; k while f,
fy, and fu are continuous. The additional assumptions are those made in
Theorem 11.10.
We can reduce this problem to the one covered in (11.22) if we consider U
as if (from which an "honest" U could be obtained by integration with say
U(O) = (D), and the resulting (Y, U) as an enlarged state vector E ~d+k. Then
equations (62) can be combined into the single Lagrangian inequality
C§*(t Y. Y. U) ~ { C§(t, Y, U) .} < (D

, , , G(t, Y, U) - Y -
and since
C§"':
(Y,U)
[.!- :
= -Id I :!~J
Gu
has maximal rank m* = m + d ::;; d + k along the optimal trajectory, the

previous results are applicable. Accordingly, there exist multipliers M* =
(M, P), say (which are continuous except at the discontinuities of U), and a Ao
( = 0 or 1) for which the modified integrand
i~ Aof + P'(G - Y) + M'C§
satisfies the usual necessary conditions along the optimal trajectory, together
with the appropriate transversality conditions at To. In addition, from (61), it
follows that
i = 1,2, ... , m. (63)
In particular, we expect that iu is constant and hence (D (since <l>u == 0)
along the optimal trajectory, while both - P = iy and i - iy' Y - iu' U
are continuous. Observe that in view of (63), the latter reduces to ho =
Aof[Yo, UoJ + P(·)· G[Yo, Uo]. Next, P(To) = A(T,y)(To, Yo (To)) for an ap-
propriate A, with (Ao, A) -:f (D. In addition, we expect that except at corner
points (Yo, Uo) is stationary for i; i.e.,
. d - - -
-P = d/y = fy = h y, say,
if
(64)
where
h ~ Aof + p. G, as usual;
moreover,
hu = iu = (D, by the above argument.
Problems 413
Finally, we have from the Weierstrass necessary condition for j that with
V = (w, U) say, near (Yo(t), Uo(t»,
h(t, Yo(t), V) - h(t, Yo(t), Yo(t), Uo(t»
= j(t, Yo(t), V) - j(t, Yo(t), Yo(t), Uo(t» + P(t)· (W - Yo(t»
;:::: 0 (since P = - jy and ju = (!J along the optimal trajectory.)
As two special cases of the latter inequality, we obtain (when U = Uo(t»
that
h(t, Yo(t), W, Uo(t» ;:::: ho(t), v W near Yo(t)
and (when W = Yo(t» that
h(t, Yo(t), Yo(t), U) ;:::: ho(t), v U near Uo(t).
We have now closed the circle of ideas begun in §10.1, by proving that the
conditions shown there to be sufficient on a fixed interval with convexity
are indeed necessary for more general problems. In [He] other "smooth"
problems are considered wherein the various given functions are supposed
C 2 as required. Related problems with C state variables are examined in
[S-S], and numerical methods that implement some of these results are
indicated in [G-L].
Associated nonsmooth problems are explored in [Ce], [CIa], [Lo], [1-T],
[R], [Wa], [Z] among others. Such problems demand far more sophisticated
tools, and constraints of the type considered here impose substantial diffi-
culties for the development of a governing theory. Nevertheless, after proper
interpretation, the results are essentially the same; viz., constrained optima
guarantee existence of associated multipliers for which a modified perfor-
mance function is optimized as though such constraints were not present. In
this theory, the role of the multipliers receives functional analytic clarifica-
tion, but the solution of any given problem, even when smooth, remains as
difficult as before. Each problem must be analyzed by methods peculiar to
itself. What is most needed is simple effective sufficiency criteria, especially for
nonconvex problems on variable intervals, or a better understanding of the
transformations which might be used to reduce such problems to those for
which such criteria are already known. Peterson and Zalkind (see [Le])
supply a valuable comparative survey of earlier work, while recent results of
Zeidan (see [CIa] and [Lo]) provide hope for further success.
PROBLEMS
11.1. After suitable reformulation if required, apply Theorem 11.8 to Problem:

(a) 10.1; (b) 10.3; (c) in §1O.2(e); (d) 10.29; and (e) 10.10 (see Remark 11.9).
After suitable reformulation, if necessary, apply Theorem 11.10 to Prob-
lem (f), 1O.2(a); and (g) 10.14. Try to identify necessary conditions with those
previously found to be sufficient. (h) In Example 3 of §11.1, show what
happens when we replace F by F(Y, u, T) = g yi(t) dt. Hint: Equation (39').
11.2. (Oscillator energy problem when T = 5n/4.)

(a) Suppose cr E [n/4, T). To examine posible optimality of
UtI = {+ 1 on (0, cr),

-Ion (cr, T),
compute
LT cos 't"U,,(t) dt and LT sin 't"U,,(t) dt
as functions of cr, and using u,,(O + ) = - u,,(T -), determine cr from
equation (27).
(b) For the same choices of UtI' use (17) of Chapter 10 to calculate
2E,,(T) = a + b sin(cr - p) and note that it is maximized by the cr
found in part (a).
(c) Make a similar analysis for
-I on (cr, cr + n),
- { +1
i1,,-
otherwise,
when cr E (0, n/4), and note that max E,,(T) ~ max E,,(T) from part (b).
What can you conclude?
11.3. Verify the assertions made in the proof of Theorem 11.10 concerning the
properties of ho and h as obtained by applying the minimum principle to hw.
11.4. In Theorem 11.12 replace the linear state equation by Y(t) = AI Y(t) +
G(U(t», where G is ~1, with G( - U) = - G(U), V U near (r) E 0/1.
(a) Observe that G«(r) = (r), and verify that the inequalities in the proof re-
main valid when G(U) replaces IEBU in each occurrence. Conclude that on
[T, To]: Ve-tAG(U) == (r), V small U.
(b) Show that if IEB ~ Gu«(r) is used to define M, then the conclusions remain
valid.
(c) State your result as a theorem.
11.5. (a) Observe that the hypotheses on M in both Theorem 11.12 and the previ-
ous problem preclude having IEB = O. Could AI = O? Explain.
(b) The new result in the previous problem applies to the state equation
y = -2y + sin u, with say lui ~ 1. (Why?) Solve the associated time-
optimal problem with y(O) = 1.
(c) Give a vector-valued nonlinear state equation to which the new theorem
applies. Give one with a vector-valued control.
11.6. Find alternate time-optimal controls for the problem in Example 1 of §11.2.
Verify that the edge condition in Corollary 11.14 is not satisfied for this problem.
11.7. (a) When 9 is a compact convex polyhedron in IRd with vertices ~,j = 1,2,
... , n say, then for each fixed nonzero V E Rd, show that the function
I(U) = V· U is linear and continuous (on 9). Conclude that it assumes its
minimum value m at some point Uo on 9.
(b) Suppose that Uo = (r), and note that U = - p V ¢ 9 for any p > 0, so that
Uo = (r) must be a boundary point of 9. Also I(U) = I(Uo) = m, whenever
(U - Uo) 1. V, which means that 1 also assumes its minimum value at any
such U. Since all such U must lie on the boundary of 9 (why?), conclude
that we can suppose Uo is a vertex of 9.
Problems 415
(c) If this same minimum occurs at another vertex, why must it do so at an

adjacent vertex U1 (one for which the segment joining Uo and U1 is also
on the boundary of each face containing these vertices)?
11.8. In Example 2 of §11.2:
(a) Verify that u == + 1 provides a unique parabolic trajectory of the state
equation through each point A = (a, ad in the upper region shown in
Figure 11.6, while u == -1 accomplishes the same for the lower region.
(b) In the upper region, show that the parabola through A meets the
primary trajectory for u == -1 at the point A = (a, ad where a1 =
- J ai/2 - a, and find the corresponding point when A is in the lower
region.
(c) Show that the time of travel IX = T(A) from (f) to A along the combined
trajectories is given by T(A) = J2ai - 4a + al; or J2ai + 4a - al,
according to whether A is in the upper or lower region.
(d) Using such trajectories, time optimality can be established independently
of the results of §11.2 by showing that when t E [0, IX], for each trajectory
Y(t) of the state equation joining (f) to A, we have (d/dt)T(Y(t)) ~ 1,
where T(Y(t)) is the associated time of travel from (f) to Y(t) given in
part (c).
(e)* Obtain the last inequality by proving that
(U - 1)( 2;1 + 1), in the upper region,

~ T(Y(t)) - 1~ { J2Yl - 4y
dt (u + 1)( 1),
2Yl _ in the lower region.
+
J2yi 4y
Hint: (d/dt)T(Y(t)) = 'T,,(Y(t))Yl(t) + 'T",(Y(t))u(t).
11.9. (a) When 0/1 is a closed coordinate box centered at the origin with sides of
length 2mi' i = 1, 2, ... , k, explain why the edge vectors E joining adjacent
vertices are essentially the coordinate vectors in each direction.
(b) What does the edge condition in Corollary 11.14 become for such 0/1?
11.10*. (Analysis of switching function u of equation (45).)
(a) Suppose that the eigenvalues A.i of A are real and distinct, and B is a
column of lB. Show that for vectors Ci ,
u(t) = Ve-liI. B = V· Ct cie-).,') = it Pie-).,I, say.
Hint: e-liI. B is a solution of X = - AX, where A is similar to a diago-

nal matrix. (See [C-L].)
(b) Prove by induction that this continuous switching function u associated
with B is either == 0, or it cannot vanish at more than d - 1 points
t E [0, 00). Why does this limit its number of sign changes?
(c) If the eigenvalues are real but not necessarily distinct, then from [C-L],
e
u(t) = L qi(t)e-).,I
i=l
for appropriate e < d and polynomials qi associated with the distinct A.i,
i = 1,2, ... , e. Argue that the conclusion in part (b) applies to this case as
well.
(d) If nonreal eigenvalues occur, then they must do so in conjugate pairs

A = fl. ± iv for real fl., v, giving terms in u(t) such as e- Il ' cos(vt + IX). Why
would this eliminate bounds on the number of sign changes?
11.11. A system is governed by the state law .Y'= u, with lui :S 1. Analyze the time-
optimal problem of bringing this system from state A = (y(O), y(O), ji(O)) to
(!) = (0,0,0). Characterize possible optimal controls uo, including bang-bang
behavior, number of switches, etc. Is an optimal control unique (if it exists)?
11.12. Perform an analysis similar to that in the previous problem for the system
governed by the state equation Y = (y, yd = (Y1 + u, - y + U1), where I1It =
{(u, u 1 ): lui :S 1, Iud :S I}.
11.13. Perform an analysis similar to that in Problem 11.11 for the system governed
by the state equation Y = (y, Yl) = (Y1 + u, u), where lui :S 1.
11.14. (Time-optimal controls need not exist.)
(a) Show that for lui :S 1, the equation y = Y + u has solutions from y(O)
to y(T) = 0 only if ly(O)1 < 1. Hint: Multiply by e-' and integrate to ob-
tain the estimate ly(O)1 :S 1 - e-'. Note that here the eigenvalue of A. is
positive.
(b) For y = - y + u, show that the corresponding argument is not conclu-
sive.
11.15. (Lagrangian constraints G(Y) = (!) or G(x, Y) == (!).)
(a) Extend the argument used in proving Theorem 6.10 to the case where g
is replaced by G = (gl' g2' ... , gm) (m < d), where Gis C2 and at each
x E [a, b], the gradients Vg;(x), i = 1,2, ... , m, are linearly independent.
Observe that now the continuous vector function, A = (A1' ... , Am) is
selected to make
d
dx}"Y·(x) - }"y(x) = A(x)Gy(Yo(x))
for some locally defined y E IRm such that y = 1fI(Y) "solves" G(y, Y) = O.
(b) Extend the argument in part (a) to cover the case when G = G(x, Y) for a
C 2 function G with Gy[Yo(x)] of maximal row rank m < d at each x.
(c)* Explain how you could use slack variables to consider Lagrangian
inequalities such as G(Y(x)) :S (!) or G(x, Y(x)) :S (!). Hint: See (11.22) in
§11.3.
11.16*. (Simple differential constraints: ~(t, Y(t), Y(t)) = (!).)
(a) Suppose that for some choice of variables Y = (X, V) E IRd - k X IRk, the
constraint function ~ takes the form ~(t, Y, Y) = G(t, Y, V) - X. Then
show that the problem of minimizing
9'(Y, T) = SaT I(t, Y(t), Y(t)) dt under ~[Y(t)] = (!),

with say Y(O), Y(T) prescribed, can be transformed into a control prob-
lem of minimizing
F(Y, U, T) = SaT f(t, Y(t), U(t)) dt,

Problems 417
where f(t, Y, U) ~ I(t, Y, G(t, Y, U), U) under the new state law Y =
(G(·, Y, U), U) with the control set 0/1 = IRk.
(b) Apply the minimum principle (Theorem 11.10) to a possible minimizing
(Yo, Uo , To), to obtain an associated Lagrangian multiplier function
(P, Q), say, and a Ao = 0 or 1 for which (since 0/1 is open): Aofu + GuP +
Q = lTJ, when evaluated at (Yo, Uo).
(c) Establish that (P, Q) satisfies the equations
(P, Q)" = -(h x , h y ) for h = Aof + p. G + Q. U.
(d) Introduce M = P + Ao!x[Yo] and conclude that the above equations
combine to form the Euler-Lagrange equations for
i~ Aol + M·(G - X) = Aol + M·t;;.
(e) Discuss the related transversal conditions, and formulate your result as a
theorem.
11.17. (a) Obtain conditions which are necessary for Y = (y, yd Eel [0,1] to mini-
mize F(Y) = H'p
2 dt under.Pi - .P = 0 with Y(O) = lTJ, y(l) = 1, Yl(1) = 2.
(b) Show that these conditions will also be sufficient provided that an appro-
priate multiplier, J.L ~ 0, can be found.
(c) Explain how to find an optimal solution from these conditions.
11.18. (a) Apply an analysis similar to that used in the previous problem to
minimize F(Y) = HI YI 2 dt under .P~ + .P = 0 with Y(O) = lTJ and
y(l) = 1.
(b) What is the minimum value?
11.19. (a) Obtain necessary conditions for the time-optimal problem of a system
whose state Y = (y, Yl) at time t is governed by the nonlinear state equa-
tion Y = (u, u 2 ) where lui:::;; 1 with Y(O), Yl (T) prescribed.
(b) Can you find a sufficiency argument?
11.20. In the arguments leading to Theorem 11.20 verify that as defined does i
satisfy the second Euler-Lagrange equation and the Weierstrass necessary
condition.
11.21. (Bolza functionals, I)
(a) Suppose that for some C 1 function cp, (Yo, Uo, To) minimizes
F*(Y, U, T) = cp(Y(T)) + IT f(t, Y(t), U(t)) dt,
where U E 0/1, on
!!}* = {(Y, U) E C1 [0, T], T > 0, Y(O) = A,}

<J>(T, Y(T)) = lTJ
subject to Y = G(t, Y, U) on (0, T).
Introduce a new state variable Y and control u = .p, form Y = (Y, y),
a a
= (U, u), and show that (Yo, o , To) minimizes
FCY. a, T) = IT [f(t, Y(t), U(t)) + u(t)] dt

on a related ?), under the above augmented state equations, provided that
Yo(O) = 0, y(To) = q>(Y(To)), U E R
(b) Apply the minimum principle from Theorem 11.10 to
h= Ao(f + u) + p·G + pu,
noting that then
h.(·, Yo, Uo) = Ao + p = o.
(c) Examine the associated transversality conditions for the augmented
<ij(y) = (<J)(Y), q>(Y) - y), and verify that the conclusions of Theorem
11.10 are valid as stated if A<J) is replaced by Aoq> + A<J).
(d) Argue that the same is true when <J) = <J)(t, Y) and q> = q>(t, Y). Hint:
Introduce t as a new state variable as in the proof of Theorem 11.10.
11.22*. (Bolza Functionals, II)
Assume that Theorem 11.10 remains valid for Bolza functionals of the type
considered in Problem 11.21, when A<J) is replaced by Aoq> + A<J). Reexamine
the arguments leading to Theorem 11.20, and conclude that it, too, is valid for
$'*(Y, T) = q>0(T, Y(T)) + IT I(t, Y(t), Y(t)) dt,
under the same replacement.

11.23. Transform the problem of minimizing
$'(Y, T) = IT I(t, Y(t), Y(t)) dt under ~(t, Y(t), Y(t)) = f!J

on some set ~ subject to the single isoperimetric constraint
g g(t, Y(t), Y(t)) dt =1, by introducing a new variable y satisfying
y= g[Y(·)], with y(O) = O.
11.24*. Repeat the arguments leading to Theorem 11.20 when ~ is replaced by
~*(t, Y, U, W) defined in (58). W is to be considered as a new control in the
state equation Y = G*(t, Y, U, W) where G* is an appropriate function
defined implicitly by ~* = f!J. Conclude that for a suitably defined A*,
A*(t, ¥O(t), U) ~ A*(t, Yo(t), Yo(t)), V U such that ~(t, ¥O(t), U) :s; f!J, and hence
that the associated multiplier function M ~ f!J.
11.25. Obtain necessary conditions for the time-optimal problem of a system whose
state Y = (y, yd at time t is governed by Y = (Yl' u), with Y(O), Y(T),
prescribed, when u 2 :s; Y1.
11.26. (a) A cup fIlled with tea at 100 DC is to be cooled to a temperature 0 DC by
adding a fixed amount of milk. An equation for the temperature y in the
cup permitting heat loss due to overflow can be approximated by
y = - y - 25u - tuy, with y(O) = 100, y(T) = O. Obtain necessary
conditions for a time-optimal solution under an external liquid flow u
controlled by O:s; u(t) :s; 1, with g u(t) dt = 1, and show that U o = 0,
o :s; t :s; 0.69.
(b) Note that y < 0, permits x = y to be used as the independent variable in
an equivalent fixed interval problem. Formulate the new problem and
compare with Problem 10.15. What can you conclude concerning
uniqueness and sufficiency of the control U o obtained in part (a)?
Appendix
In this part, we present complete proofs of several standard results from

advanced calculus and basic analysis which are utilized in the text. Frequent
appeal is made to the proven facts that a continuous function on a compact
set is uniformly continuous (5.2), is bounded, and assumes maximum and
minimum values (5.3). However, these results can be invoked as needed, since
compactness of closed and bounded sets in IRd is established at the outset.
Section A.5 is devoted to establishing the existence and uniqueness of
solutions to systems of differential equations as required for Chapters 9 and
11. The next section provides a brief discussion of the Rayleigh ratio as
connected with the higher eigenvalues of the Sturm-Liouville problem from
§7.3. The appendix concludes with an examination of tangent cones in IRd
that is needed (only) in §11.3.
§A.O. Compact Sets in ~d

To establish the compactness of the bounded interval [a, bJ s IR (as defined
in §5.2), it suffices to prove that the unit interval I = [0, IJ is compact.
Indeed, a scale change and a translation permits us to regard [a, bJ as the
unit interval in another coordinate system, and convergence is clearly pre-
served under these elementary operations.
(A.O) Lemma. 1= [0, IJ is compact.

PROOF. Each x E I has the binary expansion x = Lk=l bk 2-k, where bk = 0
or 1, k = 1, 2, ... , and in case of ambiguity, we choose that expansion which
terminates; e.g., we take 2-(m-l) instead ofLk=m2-k, m = 1,2, ....
419
420 Appendix
Now, for each n = 1, 2, ... , suppose that Xn E f has the binary expansion
Xn = Lk=l bnk 2- k •
We will select a subsequence Xl' X2, ... of these Xn which converges to a
limit point X E f.
If there is an infinite set of the Xn with first binary coefficient bIn = 0, then
choose Xl to be that Xn from this set with least index n l . Otherwise, there is
an infinite set of the Xn with bIn = 1, and Xl should be taken to be that from
among them with least index. In either case let 51 be the first binary coeffi-
cient of Xl'
Next, let X2 be that Xn of least index n2 > n l for which bIn = 51 and
b2n = 0, if there is an infinite set of such Xn; otherwise, for which bIn = 51
and b2n = 1. Let 52 be the second binary coefficient of x2.
This selection process can be continued indefinitely, since at each stage we
consider only those remaining Xn which form an infinite set. There results a
subsequence xm = Xn ,and an associated sequence of binary coefficients 5m ,
m = 1, 2, ... , d etermmmg
?' .
t h e number x
- def"
= L.,k=l
<Xl b-k 2- k Ef.
But by construction, X and xm will have the same initial m binary coeffi-
cients 51' 52' ... , 5m, and the remaining binary coefficients cannot exceed 1.
Hence, by an easy estimate,
<Xl
Ix-xml::;; L 2·2- k =4·2- m , which-+O asm-+CX). D
k=m
Remarks. The selection process used in proving the lemma may be visualized
by successively bisecting (sub)intervals of f, and retaining at each stage, the
leftmost interval which contains an infinite set of the x n. The 5k index these
selections, and, more importantly, define the limit point X (as the sum of a
series of positive terms). We are assuming here (and elsewhere) that the set of
real number is complete in that it contains a sum for each convergent series
of its elements. For further discussion of this somewhat subtle point, see
[Ed].
Compactness is the basis for many surprising results in analysis. For

example, suppose f is differentiable at each point of a compact interval, [a, b]
and f(xn) = 0 for distinct points Xn E [a, b], n = 1,2, .... Then a subsequence
{xnJ has a limit point Xo E [a, b] at which both
· f( X ) = 0 an d
f( Xo ) = 11m f'( Xo ) -- l'1m f(x nk ) - f(xo) -- 0.
nk
k-+<Xl k-+<Xl X nk - Xo
Compactness is essential for this conclusion: f(x) = sin x and f'(x) = cos x
never vanish simultaneously even though f(nn) = 0 for n = 1,2, ....
If we consider successively each component of a sequence Xn E B =

::;; Xj ::;; bj , j = 1, 2, ... , d}, n = 1, 2, ... , then it is straightforward
{X E /R d : aj
to establish the compactness of the closed box B. [The proven compactness
§A.1. The Intermediate and Mean Value Theorems 421
of [aI' bl guarantees a convergent sequence of the first components of the

]
Xn with a limit point Xl E [aI' bl ]. Then from this subsequence of the second
components, we may extract a convergent subsequence with the limit point
X2 E [a2' b2]. Continuing successively, we obtain a dth sub-sub-subsequence
{Xn}~=l' say, for which the dth components converge to a limit point
Xd E [ad' bd], while each of the other components converges to its previous
limit. (Why?) It follows that limn Xn = X = (x I, X2, ... , Xd) E B by applica-
tion of the triangle inequality.]
(A.1) Theorem. Each closed and bounded set K £; ~d is compact.

PROOF. Since K is bounded, K £; B for some (large) closed box B. If Xn E
K £; = 1, 2, ... , the compactness of B guarantees a convergent sub-
B, n
sequence {XnJ, say, with limit point Xo E B. But K is closed and by defini-
tion contains the limits of each of its convergent sequences. Hence Xo E K.
o
§A.l. The Intermediate and Mean Value Theorems

In the text, we are often required to show that a given function f assumes a
prescribed value in an interval I. In this section we establish two standard
results of this type.
First, an observation: When f is continuous at Xo E I and f(x o) > 0, then
f(x) > 0, V x in a neighborhood of Xo.
[Indeed, by the reverse triangle inequality we have
f(x) ;::: f(x o) - If(x) - f(xo)1 > f(~o), when If(x) - f(xo)1 < f(~o),
and this holds V x in a neighborhood of xo.]
A similar analysis shows that when f(x o) > c, then f > c in a neighbor-
hood of x o, and the corresponding result holds when f(x o) < c.
(A.2) Intermediate Value Theorem. Iff E era, b] then f assumes each value c
between f(a) and f(b).
PROOF. For definiteness, suppose that f(a) > c > f(b). Then by continuity (as
above) f> c on (a, x) for some x sufficiently near a. We wish to find the
largest such interval. To do so rigorously, we let P = {x E (a, b): f> c on
(a, x)} and consider the (open) set 1= UXEP(a, x), which with Xl' and
X 2 > Xl also contains the interval [Xl' x 2 ]. (Why?) Hence I is an open
interval of the form (a, x o) for some Xo E (a, b] on which f > c. Were f(xo) < c
we could use continuity as above to conclude that f < c, on some (x, x o), and
this is a contradiction. Similarly, were f(x o ) > c, we could conclude that
f> c on a larger interval (a, x) with x > x o, but then x E P, and x E (a, x o],
which is another contradiction. Thus f(xo) = c as desired. D
422 Appendix
f may of course assume the value c at many distinct points, unless f is

strictly increasing on [a, bJ (or strictly decreasing on [a, bJ). In particular,
each sign change of f on the interval must be accompanied by an interior
point at which f vanishes.
If f is also differentiable on the compact interval [a, bJ, then we can gain
further insight about its behavior. For example, unless f changes sign only
finitely often on the interval, we can use our earlier compactness argument to
produce a point at which both f and f' vanish.
Similarly, we have the following:
(A.3) Mean Value Theorem. Iff is differentiable on an interval I, then when

a < b and [a, bJ ~ I : m ~ [f(b) - f(a)]/(b - a) = f'(x*) for some x* E (a, b).
PROOF. On the compact interval [a, bJ, by (5.3), the continuous function
g(x) = f(x) - m(x - a) assumes maximum and minimum values. Since
g(a) = f(a) = g(b), at least one of these must occur at an interior point x*.
°
Then, by the usual argument, g'(x*) = = f'(x*) - m, as desired. D
(A.4) Corollary. Iff' == ° on an interval I, then f is constant on I.

PROOF. For fixed a E I, we may apply the previous result to each x = bE I, if
a < x. Then for an associated x* we must have f(x) - f(a) = (x - a)f'(x*) =
0, or f(x) = f(a), for any a < x. D
°
(A.5) Corollary. Iff is differentiable on an interval I and f'(x) "# in I, then
when [a, bJ ~ I, f assumes each value between f(a) and f(b) precisely once on
[a, b].
PROOF. From A.3, we could not have f(a) = f(b), for any distinct points a,
b E I. But f is continuous so that A.2 applies. D
It is straightforward to obtain a multidimensional version of the mean

value theorem for a C l function on a convex set (one which contains the
segment joining each pair of its points.)
(A.6) Proposition. If S is a convex open set in IR d, and f E Cl(S), then when X o,

Xl E s, f(Xd - f(Xo) = Vf(X*)'(X l - X o), for some point X* on the line
segment joining Xo and Xl'
PROOF. A typical point on the line segment is given by Xu = uX l + (1 - u)Xo,
(0 ~ u ~1), and each such point is in S. Thus
def
g(u) = f(XJ = f(uX l + (1 - u)Xo)
satisfies the conditions of the mean value theorem on 1= [0, 1J, so that
f(X l ) - f(Xo) = g(1) - g(O) = g'(u*)(1 - 0) for some u* E (0, 1). But since f
is C\ it follows that g'(u*) = Vf(Xu.)· (Xl - Xo) and setting Xu. = X* estab-
lishes the proposition. D
§A.2. The Fundamental Theorem of Calculus 423
(A7) Corollary. Iff E Cl(S(XO))' where S(Xo) is an open spherical neighbor-

hood of X o, then f is differentiable at Xo. (See 0.10.)
PROOF. S(Xo) is convex. (Why?) With Xl = X we have from A6, that when
X#- X o, and U = (X - Xo)/IX - Xol, then
f(X) = f(Xo) + Vf(Xo)-(X - Xo) + IX - Xol(Vf(X*) - Vf(Xo))· U,
for a unit vector U, and an X* "between" X and Xo. From the Cauchy
inequality it follows that the correction term 3(X - Xo) ~ (Vf(X*) -
Vf(Xo))· U may be estimated by
and 3(X - Xo) ~ (9 as X ~ X o, since IX* - Xol ::; IX - Xol. Thusfis differ-
entiable at Xo. D
§A.2. The Fundamental Theorem of Calculus

In this appendix (as elsewhere in the text) we assume as known that a real
valued function f which is piecewise continuous on a compact interval [a, b]
generates a unique real number denoted J~f(x) dx called its (Riemann) inte-
gral which may be regarded as the signed area "under the curve" of its graph.
When these conditions are met oil each compact subinterval [a, b] c (IX, /3),
then it may also be possible to define the improper Riemann integral
f: f(x) dx = lim (f f(x) dX)' as a'" IX and b ? /3.

As defined, the integral is a linear function of its integrand in the sense of
§2.2. The following results from advanced calculus are consistent with an
areal interpretation of the integral, and we shall assume as known those basic
additivity and mono tonicity properties and estimates for the integral which
can be inferred from this interpretation [FI].
(A8) Theorem (Fundamental Theorem of Calculus). If f is a (Riemann) inte-

grable function on [a, b], and F(x) ~ J:f(t) dt, then F'(x o) = f(x o) at each
point Xo E (a, b) where f is continuous. Iff E c l [a, b], then
f f'(x) dx = f(b) - f(a).
(fis also integrable over [a, x] ~ [a, b] so that F is defined.)

PROOF. For each x E [a, b] with x#- X o, we may express the difference
quotient
F(x) - F(x o) = f(x o) + _1_ IX [f(t) - f(x o)] dt. (1)

x - Xo x - Xo Xo
424 Appendix
Since f is continuous at Xo by hypothesis, for each e > 0, 3 (j > 0 such that

It - xol ::;; Ix - xol < (j => If(t) - f(xo)1 < e.
Thus, for 0 < x - Xo < (j, the term with the integral can be estimated by
Ix ~ xo,1 Ix: [f(t) - f(x o}] dtl ::;; Ix ~ xol Ix: If(t) - f(xo)1 dt
< e IX dt = e·
-Ix - xol Xo '
reversing the limits on the integral shows that the same estimate is valid when
0< Xo - x < (j. Hence as x -+ Xo, the integral term in (1) approaches zero
and we conclude that
F'( Xo ) -- 1·1m F(x) - F(xo) -- f( Xo.)
X-+Xo X - Xo
Finally, when fECI [a, b], then by the result just established, F(x) ~
J~f'(t)dt is defined on [a, b] with derivative F'(x) = f'(x), x E (a, b). Hence
by A.4, f(b) - f(a) = F(b) - F(a) = J~f'(x) dx. 0
Remark. When Xo is a or b, the initial derivation-restricted to those x for

which F(x) is defined-shows that f(x o) is the corresponding one-sided deriv-
ative of Fat Xo. The first part also holds when the real valued function F(x)
is defined by an improper Riemann integral f(t) dt. J:
(A.9) Lemma. If 0::;; p E C(a, b) and J~p(x) dx = 0, then p(x) = 0, V X E (a, b).
(The existence of the possibly improper integral is presupposed by the

hypotheses.)
PROOF. For x E (a, b),O ::;; P(x) ~f J~ p(t) dt ::;; J~ p(t) dt = 0, since p ~ O. Hence
P(x) == 0, so that by Theorem A.8, p(x) = P'(x) = O. 0
(A.10) Proposition. If f and g are in C(a, b) and f::;; g, then J~f(x) dx ::;;
J~ g(x) dx provided that both integrals exist and are finite; with equality iff
f(x) = g(x), V X E (a, b).
PROOF. 0::;; p ~ g - f is also in the linear space C(a, b). Thus as in Lemma
A.9.
0::;; r (g(x) - f(x)) dx =
with equality of the integrals on the right

r r g(x) dx - f(x) dx,
iff g(x) - f(x) = 0, V X E (a, b), or

f == g on (a, b). 0
Remark. The conclusions of A.9 and A.1O hold for piecewise continuous
functions p, and f, g, respectively, except at the points of discontinuity.
§A.3. Partial Integrals: Leibniz' Formula 425
§A.3. Partial Integrals: Leibniz' Formula

When a function f(x, y) can be integrated with respect to say the first
variable (x), then the resulting integral defines a function of the second
variable (y) whose properties we wish to ascertain.
(A.ll) Lemma. Iff is continuous on the compact rectangle [a, b] x [u, v] £;

1R2, then
g(y) =
deC
f f(x, y) dx
a
b
(2)
determines a continuous function on [u, v].

PROOF. For each y E [u, v], f(·, y) is continuous on [a, b] and hence the
integral in (2) is defined. Since f is uniformly continuous (Lemma 5.2) it
follows that for each e > 0, 3 b > 0 such that If(x, y) - f(x, ji)1 < e when
Ig(y) - g(ji)1 = r
y, ji E [u, v] and Iy - jil < b. Thus for such y, ji:
I [f(x, y) - f(x, ji)] dx I

:S r If(x, y) - f(x, ji)1 dx
:s e Lb dx = e(b - a),
and this implies that g is (uniformly) continuous on [u, v]. D
Since g of (2) is continuous, we can form the iterated integral
f dy Lb f(x, y) dx = f g(y) dy
which has the same value:

r
and by symmetry, the other iterated integral
dx LV f(x, y) dy
(A.12) Theorem. If f is continuous on the compact rectangle [a, b] x

[u, v] £; 1R2, then
Lb dx f f(x, y) dy = LV dy Lb f(x, y) dx. (3)
PROOF. By the above remarks, both iterated integrals exist and it remains
only to establish their equality.
426 Appendix
We shall regard u as fixed and permit v to vary, say v E [u, vo]. Then from
Theorem A.8,
:v rf dy f(x, y) dx =
(since by Lemma A.ll, the inner integral function is continuous).

f f(x, v) dx = g(v),
The theorem will follow if we can show that the left side of (3), F(v), say,
has the same derivative, since both sides vanish at v = u. (Here we use the
well-known fact that functions with the same derivative on an interval differ
by a constant (See A.4).)
For small t of. 0, with v and v + tin [u, vo], we may express the difference
quotient as follows:
F(v
----- =
fb dx-1 fV+I [f(x, y) - f(x, v)] dy
+ t) - F(v) + fb f(x, v) dx. (4)
tat v a
Now, given 8 > 0,3 !50 such that If(x, y) - f(x, v)1 < 8 when Iy - vi < !50 and
y, v E [u, Vo], V X E [a, b] by the aforementioned uniform continuity of f on
[a, b] x [u, vo] (Lemma 5.2). Thus when 0 < t < !50 , the iterated integral in
(4) can be estimated by
f bdx-1 fV+I If(x, y) -

f(x, v)1 dy :$; dx-
fb 8 fV+I dy = 8(b - a),
a t v a t v
and the same estimate holds for 0 < It I < !50 . Hence, recalling the definition
of F, we see that in the limit:
d
-d
fb dx fV f(x, y) dy = lim F(v + t) - F(v)
= F'(v) =
fb f(x, v) dx
v a u 1.... 0 t a
as desired. D
Interchanging the order of integration as permitted by this theorem may
be interpreted as integrating "under" the integral sign. It may also be possible
to differentiate under the integral sign.
(A.13) Theorem. Iff = f(x, y) and its partial derivative h are continuous on
[a, b] x [u, v] then
def fb
g(y) = a f(x, y) dx (5)
is C 1 in the interval (u, v) with the derivative
g'(y) = : fb f(x, y) dx = fb fy(x, y) dx. (6)

y a a
PROOF. h(y) ~ S:h(X, y) dx is continuous on [u, v] (by Lemma A.ll) and
r rf r
hence by the theorem just proven,
h(y) dy = dy fy(x, y) dx = f dx fy(x, y) dy.

§AA. An Open Mapping Theorem 427
However, the last inner integration may be carried out by A.8, and the last
iterated integral becomes
r [f(x, v) - f(x, u)] dx,
so that from (5) follows
r h(y) dy = g(v) - g(u).
For fixed u, by Theorem A.S, we may differentiate (7) with respect to v to

(7)
get g'(v) = h(v) and when v = y, this is the desired result. 0
Finally, from Theorem A.13 and the chain rule we obtain Leibniz' formula:
(A.I4) Theorem (Leibniz). If f = f(x, y) and fy are continuous on [a, b] x

[u, v] and h E Cl(u, v) with a :$; h(y) :$; b, V Y E (u, v), then
d fh<Yl fh<Yl
d f(x, y) dx = fy(x, y) dx + h'(y)f(h(y), y). (S)
y a a
PROOF. Set g(y, z) = J~f(x, y) dx so that gy(y, z) = J:fy(x, y) dx by A.13 while

gAy, z) = f(z, y) by A.S. Hence from the chain rule,
:yg(y, h(y)) = g},(y, h(y)) + gAy, h(y))h'(y),

which, after obvious substitutions, is (S). o
§AA. An Open Mapping Theorem

As We observed in §5.7, the full strength of the inverse function theorem
(5.13) is not required to establish the existence of Lagrangian multipliers.
The following suffices:
(A.I5) Theorem. For a neighborhood S = S(Xo) £; [Rd, let F = (fl' ... , h) E

(Cl(S))d have an invertible Jacobian matrix F'(X o) (with elements (8j;/8x)(Xo),
°
i, j = 1, 2, ... , d, and rows indexed by i). Then 3 m > such that for each small
t> 0, YI E Smt(F(Xo)) => Yl = F(X I ), for some Xl E St(Xo),
Remarks. The inverse function theorem guarantees, in addition, that the

point Xl is unique, and that the resulting inverse function is itself C l . We will
utilize the continuity of the partials of the Jj only at XO'
PROOF. L\(X) ~ det[F'(X)] consists of sums and products of the elements of
F'(X) and so will inherit their hypothesized continuity. Moreover, .1(Xo) ¥- 0,
since F'(Xo) is invertible, and thus, by continuity as in §A.I, .1(X) ¥- in an °
428 Appendix
open neighborhood
S,(Xo) = {X E [Rd: IX - Xol < r},
so that F'(X) is also invertible in this neighborhood. The continuity of the
partial derivatives (at Xo) in conjunction with the mean value theorem
ensures that each component Jj is differentiable at Xo as in A.7, and we may
infer that F is differentiable at Xo in that
for an appropriate vector valued function 3 having limit (!) as X - Xo -+ (!).

(F(X) and X - Xo are column vectors). When X -# X o, we may rewrite (9) as
F(X) - F(Xo) = F'(X )U + 3(X - X) (10)

IX - Xol 0 0,
for the unit vector U = (X - Xo)/(IX - Xol).
Now, F'(Xo)U -# (!), since F'(Xo) is invertible (§0.13), and hence the real
valued function p(U) ~ IF'(Xo)UI is positive and continuous on the compact
surface of the unit sphere, {U s; [Rd: lUI = I} (§A.O). Thus, by Proposition 5.3,
p assumes the necessarily positive minimum value, 4m, say, so that from (10),
and the reverse triangle inequality
=
IF(~ ~~~o)1 ~ IF'(Xo)UI -13(X - Xo)1 > 4m - m = 3m(> 2m), (11)
on a sphere St(Xo), with t < r, so small that 13(X - Xo)1 < m.
We claim that with m, t, as above and Yo = F(Xo), each point YI E Smt(YO )

is obtained as YI = F(X I ) for some point Xl E St(Xo). To establish this asser-
tion (which completes the proof of the theorem), we appeal once again to
Proposition 5.3 as follows:
On the compact sphere {X: IX - Xol ::;; t}, the continuous function
,u(X) ~f IF(X) - YII assumes its minimum value at a point, Xl' say. More-
over, if IXI - Xol = t, we would have from the reverse triangle inequality:
or, by (11),
,u(XI ) > 2mlXI - Xol- mt = 2mt - mt = mt,
which contradicts our knowledge that the minimum value,
,u(Xd::;; ,u(Xo) = !F(Xo) - YII = IYo - YII < mt.
Thus the minimum value occurs at Xl with IXI - Xol < t, and Xl must also
give the minimum value of
§A.5. Families of Solutions to a System of Differential Equations 429
Hence \7fl2(X1) = 2(F(Xl) - Y1))F'(Xd = (9 (where now F(Xd - Y1 is a

row vector). But F'(X1) is invertible, since IXl - Xol < r; therefore F(Xl)-
Y1 = (9, or Yl = F(X 1 ) as asserted.
Finally, observe that while m is fixed, t ( < r) may be taken to be as small
as we wish. 0
Remark. A further application of the mean value theorem can be used to

prove that the point Xl is uniquely determined in St(Xo) for a given t > 0,
and this establishes the existence of a local inverse function, F-l, in Smt(lO)'
The elements of the inverse matrix to F'(X) constitute the Jacobian matrix for
F-t, and their continuity will follow from Cramer's rule. For this analysis,
which would establish the inverse function theorem, and a derivation of its
corollary, the implicit function theorem, see [Ed].
§A.5. Families of Solutions to a System

of Differential Equations
In §9.7 we obtained central families as solutions of a first-order system of
differential equations of the form
T'(x) = G(x, T(x)), X E [a, b],
with T(a) = L prescribed. (12)
Here, for a fixed N = 1,2, ... , G = G(x, T) is a given N-vector valued func-
tion which is continuous on a domain D of [RN+l.
We suppose that To E ff' = (C1[a, b]t is a known solution to (12), with

To(a) = L o, and that for some 150 > 0, D contains the compact strip Ko =
{(x, T): IT - To (x) I :::; 150 , x E [a, b]}, as shown in Figure A. I.
For each L with IL - Lol < 150 , we seek a C 1 solution T E ff' to (12) with
T(a) = L whose graph is contained in Ko. We further desire that these solu-
a b x
Figure A.I
430 Appendix
tions, denoted T(-; L), have a Jacobian matrix TL with continuous elements,
and to obtain this we shall require that G have a Jacobian matrix GT with
continuous elements (and columns indexed by T).
(A.16) Lemma. If the elements of GT are continuous on K o, then for (x, To)
and (x, Td in Ko:
G(x, Td - G(x, To) = (Il GT(x, T,J dU) (Tl - To), (13)
where T" def

= uT1 + (1 - u)To, u E [0, 1]; so that
IG(x, T1 ) - G(x, To)l::; yIT1 - Tol, (14)
for a constant y > 0.
PROOF. For each u E [0, 1], (x, T,,) E Ko (Why?), and by the chain rule
(djdu)G(x, T,,) = GT(x, T,,)(Tl - To) for the constant vector Tl - To. For a
typical component, g, of G, we have
g(x, T1 ) - g(x, To) = Il :u g(x, T,,) du = (Il gT(X, T,,) dU). (Tl - To),
and this gives (13).

Moreover, the continuous function IgT(X, T,,)I is bounded on the compact
strip Ko by 5.3, and hence from standard estimates:
Ig(x, T1 ) - g(x, To)1 ::; (Il IgT(X, T,,)I dU} Tl - Tol
::; Ygi Tl - Tol, say.

(14) follows for a sufficiently large y. o
(A.17) Proposition. With G as above, equation (12) has at most one solution
T E !!I with prescribed T(a) = Lo.
PROOF. Let To be the known solution, and take K o, 150 , y as above. Observe
that upon integration of (12) with To(a) = L o, we obtain
To(x) = Lo + LX G(s, To(s)) ds, x E [a, b]. (15)
Each solution T E !!I of (12) with T(a) = Lo also satisfies (15), and from
continuity (x, T(x)) E K o, for a maximal interval [a, b], with b ::; b. Conse-
- ~
quently, on [a, b] we may use (14) to estimate V(x) = T(x) - To(x), as follows:
v(x) ~ IV(x) I = ILX (G(s, T(s)) - G(s, To(s))) dsl
::; y LX IV(s) I ds = y LX v(s) ds = a(x), say.

When A.8 is applied, we get

a'(x) = yv(x) :::;; ya(x) or (e-YXa(x))':::;; 0,
which, when integrated, gives
IV(x) I = v(x) :::;; a(x) :::;; eY(X-aJa(a) = 0. (Why?)
Thus T(x) = To(x) on [a, b], and in particular: T(b) = To(b). Were b < b, we
would conclude again by continuity that (x, T(x)) E Ko for a larger interval
than [a, b] contradicting its supposed maximality. Hence b = band T = To.
o
This proposition also shows that each solution T('; L) considered above
will be fixed uniquely by its initial value at x = a. We will now prove that
such solutions exist.
(A.18) Theorem. Let To E:Y be a solution of (12) with To(a) = L o, and G, K o,

«50 , y be as above. For each L with IL - Lol = «5 < «5oe- Y(b-a J, there exists a
unique solution T E :Y of (12), with T(a) = L, and for this solution
I T - To I = max IT(x) - To (x) I < «5e y(b-a J.
xe[a,bl
PROOF. We shall obtain T as a (continuous) solution to the integral equation
T(x) = L + LX G(s, T(s)) ds. (16)
It then follows that T(a) = L, and from the fundamental theorem of calculus
that T E:T, with T'(x) = G(x, T(x)), x E [a, b]. The uniqueness is then a
consequence of A.17.
Let r = b - a.
The solution to (16) is obtained by successive approximation from (15) as
follows
To(x) = Lo + LX G(s, To(s)) ds;
(17)
1',.+1(X) = L + LX G(s, 1',.(s)) ds, n = 0,1,2....
Observe that Tl (x) = To(x) + L - L o, so that

IITl - Toll = IL - Lol = «5 < «50 , i.e., (x, Tl(X)) E Ko. (18)
Next, by (17), (14), and (18):
IT2 (x) - Tl (x) I = ILx [G(s, Tl(S)) - G(s, To(s))] dsl

432 Appendix
so that by the triangle inequality (§5.1, Example 6):
IIT2 - Toll ~ IITI - Toll + IIT2 - TIll ~ «5(1 + yr) ~ «5e yr < «50 •
Thus (x, T2 (x)) E K o, and we may continue inductively to conclude that
for n = 0,1,2...
[y(x - a)]n
1T,,+I(x) - T,,(x) I ~ , «5, (19)
n.
while
I T,,+1 - Toll ~ «5e yr < «50 , so that (x, T,,+1 (x)) E Ko·
From (19), we see that when n ~ 1,
n-l
T,,(x) = To(x) + L
k=O
[1i+l (x) - li(x)],
is the nth partial sum of an absolutely (and uniformly) convergent series;

hence T(x) = limn-+oo T,,(x) exits, and I T" - Til -+ as n -+ 00. Also °
IT(x) - To (x) I = lim IT,,(x) - To (x) I ~ «5e yr ;
n-+oo
i.e.,
IT - Toll ~ «5e y(b-a) < «50 , so that (x, T(x)) E Ko. (20)
Finally, T, as the uniform limit of continuous functions T", is itself continu-

ous (see [Ed]), and using (17) and (14), we have from the triangle inequality
IT(x) - L - LX G(s, T(s)) ds I

~ IT(x) - T,,+1 (x) I + ILX [G(s, T,,(s)) - G(s, T(s))] ds I
~ IT(x) - T,,+1 (x) I + y(b - a)llT" - Til. (Why?)
In the limit as n -+ 00; the right side vanishes and we conclude that
T(x) = L + S: G(s, T(s)) ds as required. 0
(A.19) Remark. A slight modification of the construction used in this theo-

rem yields a proof of local existence for a solution T to (12) with T(a) = L,
even when the solution To(x) is not available. Indeed, if we set To(x) == L, and
define T,,(x) recursively by (17) then it is seen that
IT1(x) - To (x) I ~ LX IG(s, L)I ds < «5 0, when Ix - al ~ r,

for a sufficiently small r. The remaining estimates will be as before, and the
functions T,,(x) will converge to a continuous solution T(x) of (16), but only
in the neighborhood {Ix - al < r}. However, Proposition A.17 still applies
and we can conclude that the solution to (12), with T(a) = L, is uniquely
determined in this neighborhood. (It is possible to give growth conditions on
G, which permit the extension of this solution to the entire interval [a, b]. See
[C-LJ and [Ak].)
The continuous dependence of the solution T = T('; L) (established by

Theorem A.18) on its initial data, L, is apparent. The estimate (20); viz.,
liT - Toll ::;;; bey(b-a) = ey(b-a)IL - Lol,
shows this when L is near L o, and by the uniqueness previously established,

we may obtain a similar estimate for L sufficiently near Lo, when
ILo - Lol < b.
To establish the differentiable dependence on L requires more effort, and

we shall do so for a typical component I, of L. If E is the associated unit
vector in Ill N , and we set T(x, h) = T(x; Lo + hE), for small h in Ill, then we
must guarantee the existence of the directional derivative.
V(x) ~ lim V(x, h), (21)
h-+O
where
h) = T(x, h) - To(x)
V (x, h (h oF 0). (22)
It is not difficult to see by formal operations that if it exists (and is

continuous) then V should satisfy the linear integral equation
V(x) = E + LX GT(t, To(t)) V(t) dt. (23)
By means similar to those employed in proving A18, we may show that

for a given unit vector E, equation (23) has a unique continuous solution V.
(The estimate corresponding to (14) is trivial, and the solution discussed in
A19 is valid for the entire interval.)
(A20) Theorem. The solutions T = T('; L) of (12) have a Jacobian matrix

TL = TL('; L) with jointly continuous components, ot;/8lj , i, j = 1,2, ... , N.
PROOF. Following the above discussion, we shall first establish (21), assuming
the existence of V as a solution to (23). Thus with T(x) = T(x, h) = T(x; L),
for L = Lo + hE, we have from (22) and (16),
V(x, h) - V(x) = LX {[ G(s, T(s)) - G(s, To(s))] h- 1 - GT(s, To(s)) V(s)} ds.
Introducing
T,.(s) = uT(s) + (1 - u) To(s) for u E [0, 1J, (24)
434 Appendix
and using an analogous form of (13), we get after some manipulation:

V(x, h) - V(x)
= IX [(It GT(s, T..(s)) dU) V(s, h) - GT(s, To(s)) V(s) ] ds
= IX (f [GT(s, T..(s))] dU) (V(s, h) - V(s)) ds
+ f: (f [GT(s, T..(s)) - GT(s, To(s))] dU) V(s) ds, (25)
since GT(s, To(s)) is constant with respect to the U integration.
From A.1S, we know that when Ihl = IL - Lol = ~ < ~oe-y(b-a), then
I T - Toll <
~Ko = IhlKo, where KO = ey(b-a), and by (24),
liT.. - Toll = luIIIT(·, h) - Toll < IhlKo for U E [0, 1]. (26)
From the uniform continuity of the components of GT on the compact set
Ko (5.2), then for a typical row gT of GT and each e > 0, we can conclude by
(26) that IgT(s, T..(s)) - gT(S, To(s)) I ::; e, when Ihl is sufficiently small.
Thus we may estimate (25) in the manner used to obtain (14), and see that
for sufficiently small h, and an appropriately large y,
vh(x) ~ IV(x, h) - V(x) I ::; y (IX vh(s) ds + ell VII) = a(x), say.
Then, as in the proof of A.17, a'(x) = yvh(x) ::; ya(x), or (e-yxa(x))' ::; 0, so that
vh(x) ::; a(x) ::; ey(X-a)a(a) ::; ey(b-a)eyll VII. Since e may be made as small as we
°
wish, we see that as h -+ 0, Vh(X) -+ and in fact I V(', h) - VII -+ 0.
Recalling the definitions (22) and (21) of V(x, h) and V(x) respectively, we
have proved that for an appropriate E, the partial derivative 1/ = 1/(x; Lo)
exists and provides a continuous solution to equation (23)
1/(x; Lo) = E + IX GT(s, To(s)) 1/(s; Lo) ds.

Similarly, for each L near L o, with the same E,
1/(x; L) = E + IX GT(s, T(s; L)) 1/(s; L) ds,

and by a repetition of the arguments used above in estimating (25), it can be
°
shown that for each e > 0, there is a ~ > so small that IL - Lol < ~ makes
I T(-; L) - To I ::; eo(::; ~o), and consequently, 111/('; L) - 1/('; Lo)ll ::; e.
§A.6. The Rayleigh Ratio 435
Therefore, for this b, e:

11i(x; L) - 1i(xo; Lo)1 ::; 11i(x; L) - 1i(x; Lo)1 + 11i(x; Lo) - 1i(xo; Lo)1
::; e + 11i(x; Lo) - 1i(xo; Lo)1 ::; 2e,
say, when also Ix - xol is small. Thus, in particular, the components of TL are
(jointly) continuous as required. D
§A.6. The Rayleigh Ratio

In the discussion of the nonuniform string in §8.9, it was shown that the
natural frequencies of motion are directly related to the eigenvalues of a
Sturm-Liouville problem of the type presented in §7.3. Moreover, both in
Problem 7.19 and in §8.9(a), we have indicated the role played by the asso-
ciated Rayleigh ratio in characterizing the fundamental frequency, or least
eigenvalue, A. In this section, we will examine briefly, how this same ratio is
related to the higher eigenvalues. Although we consider only the problem
studied in §7.3, the methods employed admit far reaching generalizations
which form the basis for most modern results in this subject. (See [W -S].)
We first establish that the eigenfunction associated with a given eigenvalue
is essentially unique, and that the eigenfunctions for distinct eigenvalues
exhibit a natural orthogonality. We then use this orthogonality to define
(inductively) the higher eigenvalues as successive minima of the Rayleigh
ratio, and obtain an equivalent characterization by means of the Courant-
Weyl maximin principle. We then use comparison to establish the existence
of arbitrarily large eigenvalues and this provides useful summation formulas
for coefficients related to eigenfunction expansions. Finally, we apply these
formulae to give a proof for a Wirtinger inequality.
For given functions p, q, r E C[a, b], with p, r, positive, and q ~ 0, we

recall that the problem of minimizing the Rayleigh ratio
S~[ry'2 + qy2](x)dx
R(y) = S: [py2] (x) dx (27)
on
~ = {y E CfYo: y =f. l!J}, where CfYo = {y E CI[a, b]: y(a) = y(b) = O},
could be solved by a positive solution YI E CfYo of the Euler-Lagrange
equation
(ry')' = (q - Ap)y on (a, b), (28)
when the eigenvalue A = Al = R(yd (Problem 7.19). In fact, with the normal-
ization G(yd = 1, where
G(y)
def
= f
a
b
[py2] (x) dx, (29)
436 Appendix
Y1 is within a sign, the unique eigenfunction in iflJo for this eigenvalue A1.
[Otherwise, when A = A1, the equation (28) would have two linearly indepen-
dent solutions, both in iflJo. It would then follow that all solutions to this
second-order linear equation on [a, b] would be in iflJo, contradicting the fact
that a solution y can be constructed by the method of §A.5 with y(a) =f. O.
(See [C-L].)]
This argument is equally applicable to any eigenvalue A, so that the asso-
ciated eigenfunction y is uniquely determined within a sign by the normal-
ization G(y) = 1. In what follows we suppose the eigenfunctions to be so
normalized. (We also suppress the argument x in the integrands.)
(A.21) Lemma. The eigenfunctions y, y associated with distinct eigenvalues

S:
A, 1, respectively, are p-orthogonal in that pyy dx = O.
r r
PROOF.
(A -1) pyy dx = [(APY)Y - (1py)y] dx
r = [(rf),y - (ry')' y] dx =((rY')y -

where we have used (28) and the analogous equation for 1, y, to produce the
(ry')Y) I: =0,
central term, and subsequent integration by parts to obtain the boundary
terms which for y, y E iflJo, must vanish; other terms which arise in this process
cancel. Since A =f. 1, the orthogonality follows. D
Thus, in seeking another eigenvalue, it would be reasonable to consider

minimizing R on q) subject to the constraint
G1(y) == r PYY1 dx = 0, (30)
or equivalently, to consider minimizing
F(y) == r (r(y')2 + qy2) dx

on the subspace iflJo subject to the constraints G(y) = 1, G1 (y) = O. By Theo-
(31)
rem 6.4b with Remark 5.17, it follows that if Y2 accomplishes this minimiza-
tion then 3 Lagrangian multipliers J1. and - A2 such that Y2 is a C 1 solution
of the equation
(32)
(The constraints eliminate the alternative conclusion 6.4a, since
<5G(y;y)
1<5G <5G(Y;Ydl=12G(Y) 2G (Y)I=12 01=2)
1
1(y; y) <5G1(y; yd G1(y) G(yd 0 1 .
Here J1. = 0. [Indeed, with two integrations by parts repeatedly utilizing the
constraint G1 (y) = 0, we have from (29) that
J1. = J1.G(Yd = J1. Lb PYf dx = Lb [(ry'rYl - qYY1] dx
= - Lb [ry'y~ + qYY1] dx
= r [(TY~r - qYl]Y dx = - Lb A1PY1Y dx = 0,
since Yl is a solution of (28) for A = A1 . (The boundary terms vanish as usual
with Y and Yl')]
It follows that Y2 is an eigenfunction to the eigenvalue A2, and as in
Problem 7.19(a), A2 = R(Y2) ~ R(yd = A1 . (Why?) Moreover, A2 > A1 , for
otherwise Y2 would be an eigenfunction to Al which is essentially different
from Yl (since G1 (Y2) = 0). Continuing in this manner, we can define A3 =
min{R(y): Y E q), Giy) = 0, j = 1, 2} where
GiY)~ Lb PYYj dx,

and conclude that if a minimizing function Y3 exists, then it is necessarily an
eigenfunction to the eigenvalue A3 (> A2 > Ad. Finally, assuming that for
j < n, eigenfunctions Yj have been found to eigenvalues Aj (and normalized to
have G(Yj) = 1), we set
An = min {R(y): Y E q); Gj(y) = 0, j < n}. (33)
In order to continue this construction we must at each stage produce a
minimizing (eigen)function, or at least have a proof that one exists. Instead,
investigators have sought methods which could be used to approximate these
eigenfunctions and eigenvalues. For example, numerical methods can be
employed to reduce (28) to an ordinary eigenvalue problem for a large N x N
matrix. A corresponding approximation for the orthogonality conditions
Gj(y) = 0, shows that the solution is to be sought in a subspace of Ill N , and
matrix methods are available to attack these problems. There are also other
methods available to establish the minimality. (See [G-F].)
However, if we assume that the relevant eigenfunctions do exist, then the
associated eigenvalues can be approximated both from above by (33) and
from below, as in the following "maximin" result.
(A.22) Proposition (Courant). For j = 1, 2, ... , n - 1, let CPj E era, b]. Then
J1.n ~ min {R(Y): Y E q); <l>j(Y) ~ Lb YCPj dx = 0, j < n} :$; An'
PROOF. We may select constants Ci not all zero so that the function Y = CiYi L7
meets the orthogonality conditions <l>j(Y) = L7=1 ci<l>iYi) = 0, j = 1, 2, ... ,
438 Appendix
n - 1, since this underdetermined system of homogeneous linear equations

always has a nontrivial solution. Then, for this Y E OJ/o, we know that
Jin ~ R(y), since
G(y) = f bpy2 dx = . ~ n
CiCj
fb PYiYj dx = .L ct =f. 0,
n
(34)
a 1,J=1 a 1=1
in view of Lemma A.21 and the assumed normalization G(Yj) = Gj(Yj) = 1,

j = 1,2, ... , n. Similarly from (31) and (28),
or
n
F(y) ~ An L ct, since Al < A2 < ... < An·
i=l
Thus Jin ~ R(y) = F(y)/G(y) ~ An by (34). D

Remarks. Writing Jin = Ji(({Jl, ({J2'···' ({In-d, so that by (33):
An = Ji(PY1, PY2, ... , PYn-d,
we see that An is the maximum (over all sets of continuous functions ({Jl'
({J2' ... ' ({In-d of the minimum for the Rayleigh ratio offunctions orthogonal to
these ({Jj' for j < n. This explains the appellation "maximin" applied to the last
result. (There are also corresponding minimax characterizations of the higher
eigenvalues. See [W-S].) Moreover, as in §7.3, we may replace the ({Jj by
piecewise C l functions <pj-in particular, by piecewise linear functions-
which are easier to produce and to control computationally. Thus, this prop-
osition may be implemented and leads to the method of finite elements; it is
also related to the approximation methods of Ritz and Galerkin. See [C-H]
and [S-F].
Now, when P = Po and 't" = 't"o are (positive) constants, while q = 0, then
(28) has for eigenvalues 1n > 0, the elementary eigenfunctions
Yn(x) = an cos WnX + bn sin WnX,
where Wn = me/I,
1= b - a;
and the coefficients an and bn are selected to have Yn E q). (See §8.9(a).)
(A.23) Corollary. When Po = max p, and 't"o = min 't", then 1n ~ An' n = 1,
2, ... , so that An -+ +00, as n -+ 00.
PROOF. For this Po, 1"0 and q = 0, let R be the associated Rayleigh ratio which
is defined since Po > O. Observe that when Y E !?):
R(y) = (1"0 f y,2 dx )/(po f y2 dX) ~ ~~~ = R(y),

so that clearly 11 = minyE.@R(y) ~ minyE.@R(y) = A1 . To obtain the corre-
sponding result for the higher eigenvalues, we set
qJj = POYj, j < n,

so that
1n = min{R(y): y E!?); <D/y) = O,j < n} ~ Jin ~ An'
by Proposition A.22. D
(A.24) Theorem. If An --+ +00 as n --+ 00, and qJ E C[a, b], set
Cn = Gn(qJ) = f PqJYn dx, n = 1,2, ....

Then:
(a) L:'=1 c~ ~ G(qJ) = J~pqJ2 dx (Bessel's inequality).
(b) If qJ E Wo, then L:'=1 c~ = J~ pqJ2 dx (Parseval's formula).
PROOF. For N ~ 1, set qJN = qJ - L~=1 CnYn, and observe that for j ~ N:
G/qJN) = fba PqJNYj dx = c j - f cnGj(Yn) = c

n=1
j - cj = 0, (35)
by the assumed orthonormality of the Yn' Hence
o ~ G(qJN) = fb PqJN (qJ -

a n-1
t CnYn) dx = fb PqJNqJ dx,
a
or
(36)
Thus L~=1 C~ ~ G(qJ) and (a) follows as N --+ 00. (Observe that we have used
only the orthonormality of the Yn.)
If, in addition, qJ E Wo, then we may suppose that G(qJN) =f. 0, V N. [Other-
wise we would have L~=1 c~ = G(qJ), while qJN = 0 (Why?), so that cj =
Gj(qJ) = L~=1 cnGj(Yn) = 0 ifj > N, and (b) would follow.] Then qJN E Wo and
by (33) and (35), AN ~ R(qJN) = F(qJN)/G(qJN), or by (31),
o ~ G(qJN) ~ AN 1 fb (1"qJ;J + qqJ~) dx =

a
AN 1(F(qJ) - t C~An) ~
n-1
AN 1F(qJ),
where we have made repeated use of the devices of integration by parts in

conjunction with the vanishing of the boundary terms, the differential equa-
440 Appendix
tion (28), and the orthogonality of the eigenfunctions. Hence by (36),

N
0::; G(qJ) - L c; = G(qJN) ::; AN 1 F(qJ) -+ 0 as N -+ 00,
n=l
and (b) follows. 0
In particular, we note that when a = 0 and b = 1 then the eigenfunctions

for the reduced problem of P = Po = , = '0 = 1, and q = 0 are the normal-
ized functions Un (X) = .j2 sin nnx obtained in §8.9(a). It follows that to each
qJ E 0//0 = {y E C 1 [0, 1J: y(O) = y(1) = O}, is associated a set of coefficients
Cn = .j2 H qJ(x) sin nnx dx such that HqJ2(X) dx = L::"=l c;. This lends cre-
dence to the (proveable) conjecture that qJ(x) = L::"=l Cn sin nnx. Moreover, in
the general setting we might expect to prove that functions qJ E 0//0' admit the
eigenfunction representation qJ = L::"=l CnYn. Although with proper interpre-
tation of convergence this, too is the case, we shall not pursue these series
considerations, and instead refer the interested reader to [C-H].
However, the preceding theorem does afford a proof for a Wirtinger in-
equality which is analogous to that considered in §9.5. As above, we note that
the trigonometric functions Yn(x) = (1/Jn) sin(nx/2), constitute the normal-
ized eigenfunctions on [0, 2nJ when q = 0 and p == , == 1, for the eigenvalues
An = n2/4, n = 1, 2, ....
(A.25) Corollary (A Wirtinger Inequality). If

YE 0//0 = {y E C 1 [0, 2nJ: y(O) = y(2n) = O}
and
f 2"
o y(x) sin(x/2) dx = 0,
then
F(y) = f 0
2"
[(y'? - y2J dx ~ 0,
with equality iff y(x) = C2 sin x.

PROOF. Since y E 0//0' we have for it the Parseval formula (A.24b):
f 2" y2 dx = L c;, 00
with Cn = r.:. f2" y(x) sin-
1 nx
2 dx, (37)
o n=l V n 0
n = 1, 2, ... , where by hypothesis, C 1 = O. Similarly, a simple computation

shows that the complementary functions Yn(x) = (l/Jn) cos(nx/2) are also
orthonormal on [0,2n]. Hence, for qJ = y', we have the Bessel inequality
(A.24a):
f o
2"
y,2 dx ~
00
L c;,
n=l
(38)
§A.7. Linear Functionals and Tangent Cones in IRd 441
where
en = - 1 f2" y'(x) cos-dx
nx
.In 0 2
n 1 f2" nx n
"2.Jn 0 y(x) sin 2 dx = "2 cn , n = 1,2, ... ,
since y(O) = y(2n) = O. Thus combining (37) and (38) and recalling that C1 =
0, we have
f
2" (y,2 - y2) dx ~
o
L
00
n=2
(e: - c:) = L
00
n=2
[(n)2
-2 - 1] c: ~ 0,
with equality iff cn = 0 when n > 2. But it then follows from (36) that
q>2(X) = y(x) - C2 sin x satisfies G(q>2) = H"(q>2)2 dx = G(y) - (C2)2 = 0, (by
(37)) so that q>2 == 0; i.e., y(x) = C2 sin x. 0
Remark. There is a similar proof for the standard Wirtinger inequality used
in Problem 1.6, involving the general theory of Fourier series. (See [H-L-P].)
As we observed there, some restriction on the functions in iflJo is necessary
for its validity. Fortunately, the integral condition H" y(x) sin(x/2) dx = 0
can always be achieved in the solution of the classical isoperimetric problem,
without violating the additional requirements that y(O) = y(2n) = O. [We
suppose that the origin is placed on a typical bounding curve of length
2n, and "load" the curve non uniformly with the mass function sin(s/2), where
s is the arc length from the origin. Then, if the x axis is taken through the
centroid of the loaded curve, it follows that H" y(s) sin(s/2) ds = 0.]
§A.7. Linear Functionals and Tangent Cones in ~d

In this section, we combine results from linear algebra with our earlier efforts
to describe tangency (§5.6) to obtain multipliers associated with minimization
in !Rd with inequality constraints.
Definition. If A E !Rd, then I(X) = A· X, X E !Rd, determines the linear func-

tional I on !Rd. When Ai E !Rd, i = 1, 2, ... , m, then L(X) = (11 (X), 12(X), ... ,
Im(X)) provides the vector valued counterpart, where li(X) = Ai· X, X E !Rd.
We say that the Ii are linearly independent when the Ai are linearly inde-
pendent, i = 1, 2, ... , m ~ d. We write A ~ (9, when each component of A
is nonnegative, and give a corresponding componentwise interpretation to
other vector-valued inequalities. Let M = (11-1,11-2, .•• , I1-m) E !Rm •
(A.26) Proposition. Suppose the Ii as above are linearly independent, i = 1, 2,

... , m ~ d, and I(X) = A . X.
442 Appendix
(a) If L(X) = (9 => I(X) = 0, V X E ~d, then I + M· L == 0, for a unique

ME~m.
(b) If L(X) ~ (9 => I(X) ~ 0, V X E ~d, then I + M· L == 0, for a unique

M ~ (9 (M E ~m).
PROOF. (a) If A is not in the m-dimensional subspace g spanned by the A;,

i = 1, 2, ... , m, then there is a vector X orthogonal to !/, but not to A.
However, then L(X) = (9 while I(X) i= 0, contradicting the hypothesis. It
follows that A = L7'=l (- /l;)A; for some /l;, or that
m m
I(X) = A·X = L (-/l;)(A;·X) = ;=1
;=1
L (-/l;)/;(X) = -M·L(X).
Clearly, having I(X) = - M. L(X) => (M - M)· L(X) == (9 which violates

linear independence unless M = M.
(b) The inequality hypothesis implies that of part (a) since L( - X) = (9 <;:;>-
°
L(X) = (9 => I(X) ~ and -1(X) = l( - X) ~ 0. Thus L(X) = (9 => I(X) = (9,
and the representation I(X) = -M· L(X) follows for a unique M E ~m. Were,
say /ll < 0, we could find an X orthogonal to the (m - I)-dimensional
subspace spanned by A;, i ~ 2, for which Al . X < 0. But then L(X) ~ (9,
while I(X) = - /ll (Al . X) < 0, again contradicting the hypotheses. Thus
/ll ~ 0, and by extension, M ~ (9 as desired. 0
We now wish to obtain "one-sided" characterizations for some of the

concepts of tangency introduced in §5.6. By Theorem A.l, the surface of the
unit sphere in ~d is compact. Suppose that G = (g;, ... , gm) has (Frechet)
differentiable components at Xo E ~d and G(Xo) = (9. Assume that Xo is not
an isolated point of the sublevel set S = {X E ~d: G(X) ~ (9}. There are
sequences Xn E S, n = 1, 2, ... , (Xn i= X o), with Xn -+ Xo as n -+ 00, for which
the corresponding unit vectors (Xn - Xo)/IXn - Xol have convergent
subsequences with unit vectors T as limits. Then
%0 = {X = cT: c ~ °and T is any unit vector obtained as above}
defines the tangent cone to S at Xo.
It follows as in §5.6, that T E %0 => G'(Xo), T ~ (9, where G' is the
Jacobian matrix of G having as its rows the gradients Vg;(Xo) = A;, say,
i = 1,2, ... , m. Indeed, for Xn E S; X i= Xo:
0> g;(Xn) - g;(Xo) = V .(X ). (Xn - Xo) + (X - X )

- IXn - Xol g, 0 IXn _ Xol 3 nO,
so that as n -+ 00: Vg;(Xo)· T ~ 0, i = 1, 2, ... , m. We can now obtain a

converse assertion of the type alluded to in the footnote at the end of §5.7.
(A.27) Proposition. If G is C l near Xo and G'(Xo) is of maximal row rank

m ~ d, then T E %0 iff G'(Xo)T ~ (9.
§A.7. Linear Functionals and Tangent Cones in [Rd 443
PROOF. We know that if T E %0, then G'(Xo) T ~ {9. Conversely, suppose that
T'~ G'(Xo)T ~ (9, for some T with ITI = 1, and let li(X) = Vgi(XO)·X,
i = 1,2, ... , m. As X ranges over ~d, L(X) = (11 (X), ... , Im(X)) ranges over ~m,
since the Ii are linearly independent by hypothesis. Thus there exist Xj E ~d,
j = 1, 2, ... , m for which the vectors L(X) are linearly independent in ~m, so
that the m x m matrix G with elements gij = Vgi(X0)· Xj' i, j = 1, 2, ... , d, is
invertible.
Next, observe that (0, (9) E ~m+1 is in the (9-level set of
F(t, V) ~ G ( Xo + j~1 VjXj + tT) - tT' (39)
(since G(Xo) = (9), and the Jacobian matrix Fv(O, (9) = G, (since
(8jJ8v j )(0, (9) = Vgi(XO)· Xj). Hence by implicit function theory [Ed], we can
represent a neighborhood of this level set as the graph of a unique C 1 func-
tion V(t) defined for It I < e, say; i.e., F(t, V(t)) == (9 for It I < e, and V(O) = {9.
Differentiating and evaluating at t = 0 we obtain
(9 = G'(Xo) C~ vj(O)Xj + T) - T' = j~ vj(O)G'(Xo)Xj = GV'(O)

(since T' = G'(Xo)T). But G, the transpose of G, is nonsingular by construc-
tion so that V'(O) = (9.
Finally, for 0 ~ t < e:
m
X(t) ~ Xo +L vit)Xj + tT E S,
j=1
(since G(X(t)) == tT' ~ (9) and as t '" 0,
X(t) - Xo - m m ,
--- = T +
t
L (vit)lt)Xj -+ T + j=1
j=1
L Vj(O)Xj = T.
But this means that (X(t) - Xo)!lX(t) - Xo I -+ Til TI = T E %0· 0
These results enable us to relax the hypotheses on the Lagrange multiplier
rule in ~d. (See Theorem 5.16.)
(A.28) Theorem. Suppose Gis C 1 and G'(Xo) is of maximal row rank m ~ d,

while f is real valued and differentiable at Xo E ~d. If f(X) ~ f(X o), when
G(X) ~ G(Xo) E ~m, then there exist unique Lagrangian multipliers, M =
(11-1' ... , I1-m) ~ (9, such that vj(Xo) = (9, where j = f + M· G, and l1-igi(XO) =
0,Vi=1,2, ... ,m.
PROOF. Assume that G(Xo) = (9, let Land S be as in the proof of the
preceding proposition, and set I(X) = Vf(Xo)· X. Then from the proposition,
L(T) ~ {9 => T E %0, and there is a sequence Xn E S, Xn -# X o, for which
(Xn - Xo)/IXn - Xo I -+ T as n -+ 00, if ITI = 1. By hypothesis, f(Xn) ~ f(X o),
so that as n -+ 00, 0 ~ (f(Xn) - f(Xo))/IX n - Xol-+ Vf(Xo)· T; i.e., L(T) ~
(9 => I(T) ~ 0, V unit vectors T, and hence by homogeneity, for all T E ~d.
444 Appendix
Thus by Proposition A.26(b) there exist unique M E [Rm for which M ~ (9 and
m
Vf(Xo) = - L fJ.iVgi(XO),
i=l
or
Vj(Xo) = (9, if j = f + M· G.
Now, were some giXo) < 0, then by continuity we should have gj <
near Xo , and such constraints would be inactive in the minimization. It
°
follows that we can restrict the gi to those for which gi(XO ) = 0, and apply
that part already established to these gi' But then, by uniqueness, we can
°
simply set Jlj = when gj(Xo) < 0, and the conclusion follows. 0
Bibliography
[AK] N.!. Akhiezer

The Calculus of Variations. Blaisdell Publishing Company: Boston, 1962.
[A-F] M. Athans and P. Falb
Optimal Control. McGraw-Hill Book Company: New York, 1966.
[A-S] P. Aris and w.e. Strieder
Variational Methods Applied to Problems of Diffusion and Reaction. Springer
Tracts in Natural Philosophy, Vol. 24, Springer-Verlag: New York, 1973.
[Ar] A.M. Arthurs
Complementary Variational Principles, 2nd ed. Oxford University Press:
Oxford, 1980.
[Bar] S. Barnett
Introduction to Mathematical Control Theory. Clarendon Press: Oxford,
1975.
[B-P] V. Barbu and T. Precupanu
Convexity and Optimization in Banach Spaces. Sijthoff and Nordhoff:
Bucuresti, 1978.
[Ba] R. Pallu de la Barriere
Optimal Control Theory. W.B. Saunders Company: Philadelphia, 1967.
[B-J] D.l Bell and D.H. Jacobson
Singular Optimal Control Problems. Academic Press: New York, 1975.
[Be] R.E. Bellman
Introduction to the Mathematical Theory of Control Process. Academic
Press: New York, 1967.
[Ber] L.D. Berkovitz
An Introduction to Optimal Control Theory, Springer-Verlag: New York,
1974.
[Bi] M.A. Biot·
Variational Principles in Heat Transfer. Clarendon Press: Oxford, 1972.
[BI] G.A. Bliss
Calculus of Variations. Open Court Publishing Company: Chicago, 1925.
Lectures on the Calculus of Variations. University of Chicago Press:
Chicago, 1946.
445
446 Bibliography
[Bo] O. Bolza
Lectures on the Calculus of Variations, 3rd ed. Chelsea Publishing Com-
pany: New York, 1973.
Vorlesungen uber Variationsrechnung. B.G. Teubner: Leipzig and Berlin,
1909.
[B-dP] W. Boyce and R. DiPrima
Elementary Differential Equations and Boundary Value Problems, 2nd ed.
John Wiley and Sons: New York, 1969.
[CCV] Contributions to the Calculus of Variations, 4 vols. University of Chicago
Press: Chicago, 1930-1941.
[C] e. Caratheodory
Calculus of Variations and Partial Differential Equations of the First Order:
Vols. I, II. Holden-Day: San Francisco, 1965, 1967.
[Ca] L.B. Carll
A Treatise on the Calculus of Variations. John Wiley and Sons: New York,
1890.
[Ce] L. Cesari
Optimization: Theory and Applications. Springer-Verlag: New York, 1983.
[Cia] F. Clarke
Optimization and Nonsmooth Analysis. John Wiley and Sons: New York,
1983.
[CI] J.e. Clegg
Calculus of Variations. Oliver and Boyd: Edinburgh, 1968.
[C-L] E. Coddington and N. Levinson
Theory of Ordinary Differential Equations. McGraw-Hill Book Company:
New York, 1955.
[Co] R. Courant
Calculus of Variations and Supplementary Notes and Exercises. New York
University Lecture Notes, 1957.
[C-H] R. Courant and D. Hilbert
Methods of Mathematical Physics, Vol. I. Interscience Publishers: New
York, 1953.
[Cr] B.D. Craven
Mathematical Programming and Control Theory. Chapman and Hall:
London, 1978.
[Ed] e.H. Edwards, Jr.
Advanced Calculus of Several Variables. Academic Press: New York,
1973.
[EI] L.E. El'sgol'c
Calculus of Variations. Addison-Wesley Publishing Company: Reading,
MA,1962.
[E-T] I. Ekeland and R. Temam
Convex Analysis and Variational Problems, North-Holland: Amsterdam,
1977.
[Ew] G.M. Ewing
Calculus of Variations with Applications. W.W. Norton: New York, 1969.
[FI] W.D. Fleming
Functions of Several Variables, 2nd ed. Springer-Verlag: New York, 1977.
[F-R] W.D. Fleming and R. Rishel
Deterministic and Stochastic Optimal Control. Springer-Verlag: New York,
1975.
[F] M.J. Forray
Variational Calculus in Science and Engineering. McGraw-Hill Book Com-
Bibliography 447
[Fo] A.R. Forsyth

Calculus of Variations. Dover Publications: New York, 1960.
[Fox] C.Fox
An Introduction to the Calculus of Variations. Oxford University Press:
Fair Lawn, NJ, 1950.
[Fu] P.Funk
Variationsrechnung und ihre Anwendung in Physik und Technik. Springer-
Verlag, OHG.: Berlin, 1962.
[G-F] LM. Gelfand and S.V. Fomin
Calculus of Variations. Prentice-Hall: Englewood Cliffs, NJ, 1963.
[G-L] J. Gregory and C. Lin
Constrained Optimization in the Calculus of Variations and Optimal Control
Theory. van Nostrand Reinhold: New York, 1992.
[G-T] D. Gilbarg and N. Trudinger
Elliptic Partial Differential Equations of Second Order. Springer-Verlag:
New York, 1977.
[G] S.H. Gould
Variational Methods for Eigenvalue Problems, University of Toronto Press:
Toronto, 1957.
[H] J.S. Hadamard
Le~ons sur Ie Calcul de Variations, A. Herman et Fils: Paris, 1910.
[H-K] G. Hadley and M. Kemp
Variational Methods in Economics. North-Holland/Elsevier: Amsterdam,
1971.
[H-L-P] G.H. Hardy, J.E. Littlewood, and G. Polya
Inequalities. Cambridge University Press: Cambridge, 1934.
[He] M.R. Hestenes
Calculus of Variations and Optimal Control Theory. John Wiley and Sons:
New York, 1966.
[I-K] E. Isaacson and H.B. Keller
Analysis of Numerical Methods. John Wiley and Sons: New York, 1966.
[I - T] A.D. Ioffe and V.M. Tihomirov
Theory of Extremal Problems. North-Holland: Amsterdam, 1979.
[Ka-S] M. Kamien and N. Schwartz
Dynamic Optimization: The Calculus of Variations of Optimal Control in
Economics and Management. North-Holland: New York, 1981.
[Ke] H.B. Keller
Numerical Solution of Two Point Boundary Value Problems. SIAM: Phila-
delphia, PA, 1976.
[K-P] P. Kosmol and M. Pavon
Lagrange approach to the optimal control of diffusions. Acta Appl. Math.
32 (1993), 101-122.
[K-S] D. Kinderlehrer and G. Stampacchia
An Introduction to Variational Inequalities and Their Application. Aca-
demic Press: New York, 1980.
[K] A. Kneser
Lehrbuch der Variationsrechnung. Friedr. Vierweg & Sohn: Braunschweig,
1900,1925.
[L] C. Lanczos
The Variational Principles of Mechanics. University of Toronto Press:
Toronto, 1949.
[L-M] E.B. Lee and L. Markus
Foundations of Optimal Control Theory. John Wiley and Sons: New York,
1967.
448 Bibliography
[Le] G. Leitmann
The Calculus of Variations and Optimal Control. Plenum Press: New York,
1981.
[Li] L.A. Liusternik
Shortest Paths: Variational Problems. Macmillan: London, 1964.
[Lo] P. Loewen
Optimal Control via N onsmooth Analysis. American Mathematical Society,
Providence, RI, 1991.
[M] K. Maurin
Calculus of Variations and Classical Field Theory, J. Lecture Notes Series,
# 34. Aaarhus: Danemark, 1972.
[M-S] J. Macki and A. Strauss
Introduction to Optimal Control Theory. Springer-Verlag: New York, 1982.
[Mi] S.G. Mikhlin
Variational Methods in Mathematical Physics. Macmillan: New York, 1964.
The Problem of the Minimum of a Quadratic Functional. Holden-Day: San
Francisco, 1965.
The Numerical Performance of Variational Methods. Walters-Noordhoff:
Groningen, 1971.
[Mo] c.B. Morrey, Jr.
Multiple Integral Problems in the Calculus of Variations and Related
Topics. University of California Press: Berkeley, 1943.
[Mor] M. Morse
The Calculus of Variations in the Large. American Mathematical Society:
Providence, RI, 1934.
[Mu] F.D. Murnaghan
The Calculus of Variations. Spartan Books: Washington, DC, 1962.
[N] J. N ering
Linear Algebra and Matrix Theory, 2nd ed., John Wiley and Sons: New
York, 1970.
[Ne] L. Neustadt
Optimization: A Theory of Necessary Conditions. Princeton University
Press: Princeton, NJ, 1976.
[No] A.R.M. Noton
Variational Methods in Control Engineering. Pergamon Press: New York,
1965.
[O-R] J.T. aden and J.R. Reddy
Variational Methods in Theoretical Mechanics. Springer-Verlag: Berlin,
1976.
[as] R. Osserman
A Survey of Minimal Surfaces. Van Nostrand Reinhold: New York, 1969.
[P] L.A. Pars
An Introduction to the Calculus of Variations. Heinemann: London, 1962.
[Pe] J.P. Petrov
Variational Methods in Optimal Control Theory. Academic Press: New
York, 1968.
[Po] L.S. Pontjragin with V.G. Boltjanskii, RS. Gamkre1idze, and E.F.
Mischenko
The Mathematical Theory of Optimal Processes. Pergamon-Macmillan:
New York, 1964.
[P-S] G. Polya and M. Schiffer
Convexity of functionals by transplantation. J. d'Anal. Math., 1953/54,
247-345.
Bibliography 449
[R] K. Rektorys
Variational Methods in Mathematics, Science and Engineering. Reidel
Publishing Company: Boston, 1975.
[Ro] R.T. Rockafellar
Convex Analysis. Princeton University Press: Princeton, NJ, 1970.
[Rud] W. Rudin
Principles of Mathematical Analysis, 2nd ed., McGraw-Hill Book Com-
[Ru] H. Rund
The Hamilton-Jacobi Theory in the Calculus of Variations. D. Van
Nostrand Company: Princeton, 1966.
[Rus] J.S. Rustagi
Variational Methods in Statistics. Academic Press: New York, 1979.
[S1 H. Sagan
Introduction to the Calculus of Variations. McGraw-Hill Book Company:
New York, 1969.
[S-S] A. Seierstad and K. Sydsaeter
Optimal Control with Economic Applications. North-Holland: New York,
1987.
[Se] L.A. Segel
Mathematics Applied to Continuum Mechanics. Macmillan: New York,
1977.
[Sew] M.J. Sewell
Maximum and Minimum Principles. Cambridge University 'Press, New
York,1987.
[Sm] D.R. Smith
Variational Methods in Optimization. Prentice-Hall: Englewood Cliffs, NJ,
1974.
[Smi] P. Smith
Convexity Methods in Variational Calculus. Research Studies Press Ltd.:
Letchworth, Hertfordshire, England, 1985.
CSt] W. Stacey
Variational Methods in Nuclear Reactor Physics. Academic Press: New
York,1974.
[S-F] G. Strang and G.J. Fix
An Analysis of the Finite Element Method. Prentice-Hall: Englewood
Cliffs, NJ, 1973.
[T] L. Tonelli
Fondamenti di Calcolo delle Variazione, I. II, N. Zanichelli: Bologna, 1921,
1923.
[Ti] V.M. Tikhomirov
Fundamental Principles of the Theory of Extremal Problems, John Wiley &
Sons: New York, 1986.
[TO] G. Leitmann, Ed.
Topics in Optimization. Academic Press: New York, 1967.
[Tr] J.L. Troutman
Boundary Value Problems of Applied Mathematics. PWS Publishing Com-
pany, Boston, 1994.
[V] M.M. Vainberg
Variational Methods for the Study of Nonlinear Operations. Holden-Day:
San Francisco, 1964.
Variational Method and Method of Monotone Operators in the Theory of
Nonlinear Equations. John Wiley and Sons: New York, 1973.
450 Bibliography
[W] K. Washizu
Variational Methods in Elasticity and Plasticity. Pergamon Press: New
York, 1967.
[Wa] J. Warga
Optimal Control of Differential and Functional Equations. Academic Press:
New York, 1972.
[W-S] A. Weinstein and W. Stenger
Methods of Intermediate Problems for Eigenvalues. Academic Press: New
York, 1972.
[Wei] H. Weinberger
A First Course in Partial Differential Equations. Blaisdell Publishers: New
York, 1965.
[We] R. Weinstock
Calculus of Variations with Applications to Physics and Engineering.
McGraw-Hill Book Company: New York, 1952.
[Y] L.c. Young
Lectures on the Calculus of Variations and Optimal Control Theory. W.B.
Saunders: Philadelphia, PA, 1969.
[Y - M] W. Yourgrau and S. Mandelstam
Variational Principles in Dynamics and Quantum Theory. Dover Publica-
tions: New York, 1979.
[Z] E. Zeidler
Nonlinear Functional Analysis and its Applications, III: Variational Methods
and Optimization. Springer-Verlag: New York, 1985.
Historical References
In addition to the appropriate comments and surveys to be found in the previous

works, the following references are of particular interest:
C. Caratheodory
The beginning of research in the calculus of variations, (and) Basel und der Beginn der
Variationsrechnung. Ges. Math. Schriften, II. C.H. Becksche Verlagsbuchhandlung:
Miinchen, 1955.
H. Goldstine
A History of the Calculus of Variations from the 17th through the 19th Centuries.
Springer-Verlag, New York, 1980.
M. Kline
Mathematical Thought from Ancient to Modern Times. Oxford University Press:
New York, 1972.
D.J. Struik
A Source Book in Mathematics 1200-1800. Harvard University Press: Cambridge,
MA: 1969.
G.F. Temple
100 Years of Mathematics. Springer-Verlag: New York, 1981.
I. Todhunter
A History of Progress of the Calculus of Variations in the Nineteenth Century.
Cambridge University Press: Cambridge, 1862.
R. Woodhouse
A Treatise of I soperimetrical Problems and the Calculus of Variations. Cambridge
University Press: Cambridge, 1810.
Jean d'Alembert (1717-1783) Carl Jacobi (1804-1851)
Daniel Bernoulli (1700-1782) Adolf Kneser (1862-1930)

Jakob Bernoulli (1654-1705)
Johann Bernoulli (1667-1748)
Gilbert A. Bliss (1876-1951) Joseph Louis Lagrange (1736-1813)
Oskar Bolza (1857-1942) Pierre-Simon Laplace (1749-1827)
Adrien-M. Legendre (1752-1833)
Gottfried Wilhelm Leibniz
Augustin L. Cauchy (1789-1857) (1646-1716)
Richard Courant (1888-1972) Joseph Liouville (1809-1882)
P.G.L. Dirichlet (1805-1859) Isaac Newton (1642-1727)

Paul du Bois-Reymond (1831-1889) Emmy Noether (1882-1935)
Albert Einstein (1879-1955) Lord Rayleigh

Leonhard Euler (1707-1783) (John William Strutt)
(1842-1919)
Bernhard Riemann (1826-1866)
Pierre de Fermat (1601-1665)
Maurice Frechet (1878-1973)
Erwin Schrodinger (1887-1961)
Charles Sturm (1803-1855)
Galileo Galilei (1564-1642)
Carl F. Gauss (1777-1855)
George Green (1793-1841) Karl Weierstrass (1815-1897)
William R. Hamilton (1805-1865) Ernst Zermelo (1871-1953)

David Hilbert (1862-1943)
451
Answers to Selected Problems
Chapter 0
0.3. Maximum occurs at (- 2, 2); f( - 2, 2) = 238 • Minimum occurs at (-1, - 2);

f(-1, -2) = -l
0.5. (a), (e) not convex.
(b) convex, but not strictly so.
(c), (d) strictly convex.
Chapter 2
2.1. Only (a), (c), (f), are not subspaces.

2.5. (a) .5J(y; v) = 3y(a)2v(a).
(c) .5J(y; v) = -t S:(2 + x 2 - sin y'(x)) cos y'(x)v'(x) dx.
(e) .5J(y; v) = S: [2x 2y(x)v(x) + eY'(x)v'(x)] dx.
(g) .5J(y; v) = 2(f: [2y'(x) + x 2 y(x)] dx)(S: [1 + y'(x)]v(x) dx)
+ (S: [2v'(x) + x 2v(x)] dx)(S: [1 + y'(X)]2 dx).
Chapter 3
3.6. Yo (x) = x 2 - 1; Yl(X)=O; Y2(X) = const.

3.7. y(x) = -1 + (2 - e)x + e\ Yl(X) = eX - ex - 1; incompatible.
3.9. y(x) = t(19x - 12x- 2); Yl (x) = t(x + 4x- 2); Y2(X) = O.
3.11. y(x) = )9 - x 2.
3.13. y(x) = -~.
452
3.15. y(x) = i(l1 - 3x 4/3 ).
3.19. y(x) = x 3 - 3x.
x 3 X2)
3.25. (b) WL(X4
2JL 1~ - -f
+ 21 ,where Xl = x/L;
M .
(d) 4JL x 2(1 - xd wIth max. at X = 2L/3.
3.27. y(x) = {j (x 3 - x).
3.32. (c) A(fJ) = [2(/3 - sin /3)/2/32.
3.38. (c) F(u o) = 2[h + g(T - 1 + e- T)]2.

2T _ 3 + 4e T _ e 2T
3.39. (c) They are the same.

3.41. (d) Fe = 10 > Fp = t loge93 ) > FT = 153'
Chapter 5
5.25. (b) Each y E ~1'
5.27. y(x) = 7e(x 3/2 - 1) + 1 with e = (2 3/2 - 1)-1.

5.29. y(x) = 1 + sin x.
5.31. y(x) = x(ln x-I).
5.37. y(x) = m Jx.

115
5.39. y(x) = i(5x 4 - 1).
Chapter 6
6.1. (a) cos y'(x) = c.

y'(x)
(c) = c.
xJl + y'(X)2
(e) x - y(x) = O.
6.3. y(x) = C1 sin x.
6.5. y(x) = 2X2 - 1.
6.7. (a) y(x) = x.
(b) None.
6.9. y(x) = -i(x 3 + 5x).
6.11. y(x) = 2 cos(i).

454 Answers to Selected Problems
6.17. (a) y(x) = C1 sin x.

(b) None.
6.19. y(x) = nnx, n = 0, ± 1, ±2, ....
6.21. (c) Y = ±xl,J2.
6.23. (a) y(x) = ±,J2 sin (nnx), n = 1, 2, 3, ....
nn
6.29. (d) y(x) = iOY (31- x).
6.31. y(x) = cos x.
6.33. y(x) = tx 2 + In x + t(4x- 1 + 7x - 8).

,,2
6.34. (b) (1 ::'2)5/2 + P(J1+?) = cy' + Co for constants Co, c.
6.41. (a) U xx + U yy + x 2 + y2 = O.
(c) 3u x uxx - 6u;u yy - u = O.
Chapter 7
7.5. (a) c = 0 with (i) or (ii).

(d) Any point c with (i); only when cos c = 0, with (ii).
7.6. (a) jl(x) = ± {I - x, 0:::; x :::; 1,

x + 1, -1:::; x < O.
(c) Yes, many.
7.7. (a) Only two possible solutions, each with jl'(x) = ±2.
(c) No.
(Note: For 7.6 and 7.7, express the Weierstrass-Erdmann conditions in terms ofm
and n, the distinct limiting values of jl' at a potential corner point, and determine these
values.)
X, x:::; c, 1
7.9. (c) jl(x) = { x where c = eC - •
e - 1, x ~ c,
7.23. (d) 1: = T if T2 :::; 2h and 1: = T - JT 2 - 2h otherwise; (e) To = 3J2h15.
Chapter 8
8.3. T =tml202 + tm112(02 + OD + m1F001 cos(O - Od,

u= mgl(1 - cos 0) + m1gl(2 - cos 0 - cos Od.
(c) Linearized equations are
.. m1··
0+ -01 = -gOII
m
and
8.5. T = tmI2(q,2 + sin 2 cp{)2),

U = mgl(1 + cos cp).
1- 2
(f) St + m2 [S; + sin- 2 cpS;] = -mg(1 + cos cp).
Chapter 10
sinh 2(1 - t). . . .

10.1. Yo(t) = . h gives umque mimmum.
sm 2
10.3. Equality holds iff yo(t) == tor yo(t) == -to
10.5. Take uo(t) = 1, t < 60; and uo(t) = 0, t > 60.
10.12. (d) The equation p- (1 + sin t)2p = -2e- t is not elementary.
10.14. uo(t) = 15(1 + 4e t- 2)/(8e- 4 - 11).
10.16. P(2) = 2(Yo(2) - (5, 2)) = ( - 22/7, 16/21); uo(t) = -l1t/7 + 58/21.
10.19. Uo(t) = 0 is the unique optimal control, but Yo(t) = 0 might not be the only
minimizing trajectory.
10.21. (a) yo(t) = 1 - t.
(b) If T = 2, solution is not unique;
if T = t, solution does not exist.
10.23. Minimum is -t which occurs if uo(t) = 1 - t.
10.26. (a) Minimum occurs where Yl(2) = _y2(2) = e-Yi (2) = a, say. The minimum
value is a/J1+7ii, but neither optimal control nor optimal state is unique.
Chapter 11
11.1. (a) yo(t) = [sinh 2(1 - t)]/sinh 2.

(e) Ao = 1, uo(t) = sgn[1 - rl/2(2 - t)], Yo(2) = 2 - 2J3.
(f) uo(t) = sgn(5 - 2t), so
yo(t) = - 5te 2t, t,

t ~
= (5t - 4)e 2t, t~l

(h) Optimal trajectories occur only along primary parabolas.
11.5. (b) To = t 10g(1 + 2/sin 1).
11.11. uo(t) = + 1 or -1 with at most two switching points, will give the unique
optimal control from any reachable A.
11.12. Optimal controls are unique and vertex-valued, but cannot bound number of
switching points.
Index
Action integral 235 Jakob Bernoulli's 156,20

in generalized coordinates 238 problem 17,66,151,257,296
in parametric form 250 with transversal 159
Action principle
least 248
Canonical equations of motion 245
stationary 238
Catenary 80, 313
Adjoint
Catenary (hanging cable) problem 78
equation 360
Cauchy inequality 1
function 360
Central family 308
Admissible directions 115, 136
construction 314
Admissible trajectories 237
Central field 308
Affine approximation 120
construction 314
Affine function
in [Rd 5 versus exact field 310
Chaplygin's problem 23, 32, 334
in linear space 127
Closed set 3, 108
Augmented function 41
Column buckling problem 164
Autonomous control problems 342,
Compact set
384
in Euclidean space 3,419
in normed space 107
Bang-bang controls 343,400 Conjugate momenta 243, 250
Beam deflection problems 90,190 Consistent trajectories 288
Bolza functional 347,410,417 Constraint
Bolza's problem 225 convex 76
Boundary inequality 42, 44, 213
of a subset of [Rd 2 internal 212
point 2 isoperimetric 23,42, 301
Bounded Lagrangian 27,44, 173, 301
function 112,419 Continuity of function on a set
set 2, 108, 421 in [Rd 3
Brachistochrone 17,31,66 in normed space 108
function 38, 119 uniform 110
457
458 Index
Continuous media 260 Eigenfunction 210, 436

elastic bar 277 Eigenvalue
stretched membrane 266 of matrix 270
taut string 260 of membrane 268
uniform plate 278 of Sturm-Liouville problem 210,
Control 265,436
optimal 339 End-point problems
steering 341 fixed 145
synthesis 365 free 60, 145
variations 380 variable 156, 297
Convergence 107 Energy
uniform 107 kinetic 236
Convex constraint 76 potential 42, 165,237
Convex function total 240
in [Rd 5 Equilibrium, static
in linear space 54 Bernoulli's principle 43, 239
strictly 5, 54, 287 stable, unstable 239, 269
strongly 57,94 Euclidean
Convex set distance 2
in [Rd 75,285,422 length 1
Corner point 198,201 norm 104
conditions at 207,216 space 1
Critical point 4 vector 1
Curve Euler-Lagrange equation
catenary 80, 313 first equation 59, 148
cornered 197,228 second equation 153
cycloid 67, 152, 159,257 Exact differential 289
geodesic 14, 177 Exact fields 289
smooth 172 and Hamilton-Jacobi equation 293
Extremal
global 204
D'Alembert's solution to wave local 4,114,203
equation 263 point in [Rd 2
Derivative, Frechet 120 point in normed space 114
Derivative in [Rd value (in [Rd) 2
continuous 37
directional 3
from right (left) 37, 424 Fermat's principle 31
partial 3 Field
second-order 6 central 292, 308
Dido's problem 32, 93, 258 construction 314
Docking, problem 401 definition 289
Differentiable function equation 289
in [Rd 5,423 exact 289
in sense of Frechet 121 functions h, P 289
Differential equations 429 parameter 308
Dirichlet's integral 91 stationary 289
Domain trajectory 289
in [Rd 2 Function
Green's 25, 182 affine 5, 127
simply connected 289 augmented 41, 132
du Bois-Reymond Lemma 97 continuous 3, 108
general 99 differentiable 5, 121
Index 459
evaluation 39 triangle 1, 104

integral 56, 202 Wirtinger 33, 304, 440
linear 39, 56, 111 Integral
piecewise C 1 198 action 235
piecewise continuous 198 constraints 160
Pontjragin 387 function 56
space 36 of motion 240, 247
stationary 148, 163, 169, 184 multidimensional 101, 181
[strictly] convex 54, 287 Interior point 2
[strongly] convex 57,61 Internal constraints 212
vector-valued 36, 169,215 Invariance of stationarity 178
Inventory problem 72
Inverse function theorem 130
Gateaux variation 45 Isoperimetric
weakly continuous 123 constraint 23, 76, 301
Generalized coordinates 237 curve 21
Geodesic curves 13 problems 21, 301
on cone 324 versus Lagrangian problems 23,302
on cylinder 65 Isoperimetric Problem 21, 33, 171, 304
in Euclidean space 14
on sphere 15
on surface 17, 177 Jacobian
Goldschmidt curve 219, 326 determinant 130
Gradient vector 4 matrix 427, 433
Green's Jacobi's
domain 25, 182 condition 319, 322
theorem 22 equation 324
principle of least action 248
Hamilton-Jacobi equation 251

Kinetic energy
and exact fields 293
of continuum 261
Hamiltonian 244,371
of particle 236
Hamilton's principle
of system 236
in generalized coordinates 238
as local minimum 321
Hessian matrix 7
Lagrange Lemma 99
Hilbert's
Lagrangian
address 195
and Hamiltonian 247,254
differentiability criterion 220
of particle 236
formula 294, 298
of system 237
invariant integral 294
Lagrangian
theorem 295
constraint 27, 44, 173, 404
(multidimensional) 333
inequalities 44,213,405
Highway design problem 367
multipliers 42, 129, 160,410,443
Legendre
approach to sufficiency 335
Implicit function theorem 429 condition 224
Inequality transformation 244
Cauchy 1 Leibniz' formula 427
complementary 256, 275 Level set offunction
constraints 42, 44, 77 in [Rd 7
isoperimetric 33 in normed space 128
460 Index
Linear control problems 342 in normed space 107

Linear function 39, 56, 111 Newton's
Linear functionals 441 equations of motion 19
Linear space offunctions 36 profile of minimum drag 69, 96, 233,
Linear state problems 361 336
Linear time-optimal problems 397 Normed linear spaces 106
Linear (vector) space 37 compact sets 107
continuity 108
convergence 107
Matrix
Norms 51
Hessian 7
for C 1 105, 106
invertible 7
for C1 201
positive (semi)definite 7
strong versus weak 202
symmetric 362
Maximal row rank 405, 442
Maximum Open
local 6,114 mapping theorem 427
point 6,114 neighborhood 2, 107
value 2 set in [Rd 2
versus minimum 3 Optimal control 339
Mean value theorem 422 simple problems 81,232, 345
Minimal area problem 25,74, 184 with damping 95, 232
with free boundary 95, 185 Oscillator problems 355, 389
Minimal surface
equation 75
ofrevolution 24, 187,217,312 Performance problem 81
degenerate 34 Plateau's problem 26
symmetric 229 Pontjragin principle 365,378,392
Minimum drag problem 69, 95, 233, Potential energy
336 Bernoulli's principle of 239
Minimum point for bent column 164
global 40 for buckled column 165
in [Rd 5
for deflected beam 90
in linear space 40 for hanging cable 78
local 6,114 for stretched membrane 267
strict local 6 for taut string 261
unique 5,40
Production planning problems 72, 193,
Minimum principle 376
necessity of 378, 392 Propulsion problems 81,350,401
sufficiency of 365
Minimum value
in [Rd 2
Quadratic form 7, 224,244
Multipliers, Lagrangian Quadratic-performance problems 347,
existence 132, 134,410 361,388
method of 42, 132
use of 76, 133
Rayleigh ratio 266, 435
Resource allocation problems 352, 376
Natural Rotating fluid column 42, 276
boundary conditions 61, 156
frequency 263,268
modes 263, 268 Saddle
Neighborhood (open) function 255
in [Rd 2 point 6
Index 461
Seismic wave problem 34,280 Triangle inequality 1, 104

Smoothing 199 reverse 104
Slack variable 231,410
Stationarity, invariance of 178
Stationary Uniform
action 238 continuity 110
field 289 convergence 107
function 148, 163 estimate 110
point in [Rd 4 Unique
trajectory 245, 288 minimization 40
Steepest descent 357,372 solution 265,431
Steering problems 20
Strong
convexity 57, 61, 94 Variable end point problems 156,391
local extremal 202 Variation, Gateaux
norm 202 at local extremal point 115, 116
variation 222 definition 45
Sturm-Liouville problem 209,435 weak continuity of 123
Switching Wave equation
function 403,415 D'Alembert's solution 263
points 370 one-dimensional 262
two-dimensional 268
uniqueness for 265
Tangent Weak
cone 442 local extremal 202, 322
direction 128 norm 202
graph 127 variation 222
hyperplane 5, 127 Weakly continuous variation 123
level set 128 Weierstrass
to trajectory 288 condition (necessary) 222, 379
Target excess function 223, 286
cone 384 formula 286
fixed 387 method (for sufficiency) 283
free 387 Weierstrass-Erdmann conditions 204,
time 342 216
Time-optimal problems 342, 397 first 206, 216
Trajectory second 206,216
admissible 237 Wirtinger inequality 1, 104,440
field 289
stationary 245, 288
Transversal Zenodoros' problem 34,193,280,
arcwise connected 298 331
conditions 157,299 Zermelo's steering problem 21, 32
(conttnuedfrom page ii)
LidllPilz: Applied Abstract Algebra. Samuel: Projective Geometry.

Macki-Strauss: Introduction to Optimal Readings in Mathematics.
Control Theory. ScharlaulOpolka: From Fermat to
Malitz: Introduction to Mathematical Minkowski.
Logic. Sigler: Algebra.
MarsdenIW einstein: Calculus I, II, III. Silvermanffate: Rational Points on
Second edition. Elliptic Curves.
Martin: The Foundations of Geometry Simmonds: A Brief on Tensor Analysis.
and the Non-Euclidean Plane. Second edition.
Martin: Transformation Geometry: An Singerffhorpe: Lecture Notes on
Introduction to Symmetry. Elementary Topology and Geometry.
MiIlmanIParker: Geometry: A Metric Smith: Linear Algebra. Second edition.
Approach with Models. Second Smith: Primer of Modern Analysis.
edition. Second edition.
Moschovakis: Notes on Set Theory. StantonIWhite: Constructive
Owen: A First Course in the Combinatorics.
Mathematical Foundations of Stillwell: Elements of Algebra:
Thermodynamics. Geometry, Numbers, Equations.
Palka: An Introduction to Complex Stillwell: Mathematics and Its History.
Function Theory. Strayer: Linear Programming and Its
Pedrick: A First Course in Analysis. Applications.
Peressini/SullivanlUhl: The Mathematics Thorpe: Elementary Topics in
of Nonlinear Programming. Differential Geometry.
PrenowitzlJantosciak: Join Geometries. Troutman: Variational Calculus and
priestley: Calculus: An Historical Optimal Control. Second edition.
Approach. Valenza: Linear Algebra: An
ProtterlMorrey: A First Course in Real Introduction to Abstract Mathematics.
Analysis. Second edition. WhyburnIDuda: Dynamic Topology.
I\-rotterlMorrey: Intermediate Calculus. Wilson: Much Ado About Calculus.
Second edition.
Ross: Elementary Analysis: The Theory
of Calculus.

Troutmal CalculusOpControl

Uploaded by

Copyright:

Available Formats

Troutmal CalculusOpControl

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Troutmal CalculusOpControl

Uploaded by

Copyright:

Available Formats

Undergraduate Texts in Mathematics

Anglin: Mathematics: A Concise History Elaydi: Introduction to Difference

Variational Calculus and

Mathematics Subject Classifications (1991): 49-01

Library of Congress Cataloging-in-Publication Data

Printed on acid-free paper.

© 1996, 1983 Springer-Verlag New York, Inc.

ISBN 0-387-94511-3 Springer-Verlag New York Berlin Heidelberg

Although the calculus of variations has ancient origins in questions of Ar-

In Part I, we provide the standard results of variational calculus in the

I wish once more to express my appreciation to those who first made me

Syracuse, New York JOHN L. TROUTMAN

(a) When J = J(z) 149

8.3. The Total Energy 240

10.2. Sample Problems 344

Answers to Selected Problems 452

Review of Optimization in [Rd

For d = 1, 2, 3, ... , let ~d denote d-dimensional real Euclidean space

which is subject to the Cauchy inequality

an alternate form of which is

When Xo E ~d, then for finite ~ > 0, the "sphere"

(0.0) I need not have extremal values on D.

For example, when D = ~1, then the function I(X) = Xl is unbounded in

(0.1) I may assume only one extremal value on D.

For example, on D = (-1, 1J the function f(X) = Xl assumes a maximum

On D = [ -1, 1J, f(X) = xi assumes a maximum value (1) at Xl = ± 1,

The only reasonable conditions which guarantee the existence of extremal

~ ~d is compact and f: D --+ ~ is continuous, then f as-

In ~d, a compact set is a bounded set which is closed in that it contains

Thus it suffices to characterize the minimum points, those Xo ED for which

(0.5) When D contains a neighborhood of X o, an extremal point of f, in which

auf(Xo) ~ lim [f(Xo + eU) - f(X o)] = af (Xo + eU) I

For example, on D = [ -1, 1J, the function f(X) = x~ has Xl = 0 as its

On D = ~2, the function f(X) = x~ - xi has Xo = (0,0) as its only critical

For example, the polynomial f(X) = x~ - 3x l has on D = [ - 3, 3J, sta-

(0.8) When f is a convex function on D then it assumes a minimum value at

For our purposes, f will be said to be convex on D when it has continuous

and so it is convex. At its only stationary point Xo = (D, it assumes its

[When Xo is a stationary point of a strictly convex function f, then f(X) >

if for all X in a neighborhood of X O,

(0.11) The classification of a stationary point Xo as a local maximum point,

We first observe that vanishing higher-order derivatives provide no infor-

each have n derivatives which vanish at their common stationary point

(0.12) Theorem. Let f: D --+ [R have continuous second-order partial derivatives

+ eU) I.=0 = q(U) >

then Xo is a strict local minimum point for f.

[Indeed, for each unit vector U, as e takes on real values in a one-

f in a neighborhood of Xo in the (two-sided) direction of U. Now 1'(0) =

which by (8) is possible at Xo only if V = (9. This condition furnishes the

For example, although on !R 2 the function f(X) = x~ - xfhas its only

To find the stationary points of f when so constrained to the level sets of

for X, Y E [Rd, X #- (I), Y #- (I).

f(tX + (1 - t)Xo) ~ tf(X) + (1 - t)f(Xo), V t, O<t<l.

f(tX + (1 - t)Xo) = f(Xo) + tVf(Xo)' (X - Xo) + t IX - Xo13(tX - tXo),

0.9*. Let D = jR2 and

Standard Optimization Problems

"Which method is best?" is a question of perennial validity, and through the

§1.1. Geodesic Problems

(a) Geodesics in IR.d