Lin Segel Applied Mathematics

Downloaded 10/13/15 to 128.84.127.209. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.
php
~ MathematicsApplied~
l)in the NaturalSciencesd
to DetenninisticProblems
Downloaded 10/13/15 to 128.84.127.209. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.php
~ ~
MathematicsApplied
to DeterministicProblems
in the NaturalSciences
b ~
C. C. Lin
Massachusetts Institute of Technology
L.A. Segel
Weizmann Institute of Science
with material on elasticity by

G. H. Handelman
Rensselaer Polytechnic Institute

Sla.Jll..
Societyfor Industrialand Applied Mathematics
Philadelphia
Books in the Classics in Applied Mathematics series are monographs and textbooks declared out of
print by their original publishers, though they are of continued importance and interest to the
mathematical community. SIAM publishes this series to ensure that the information presented in these
texts is not lost to today's students and researchers.
Editor-in-Chief
Robert E. O'Malley, Jr., University of Washington
Editorial Board
John Boyd, University of Michigan
Leah Edelstein-Keshet, University of British Columbia
William G. Faris, University of Arizona
Nicholas J. Higham, University of Manchester
Peter Hoff, University of Washington
Mark Kot, University of Washington
Peter Olver, University of Minnesota
Philip Protter, Cornell University
Gerhard Wanner, L'Universite de Geneve
Classics in Applied Mathematics

C. C. Lin and L.A. Segel, Mathematics Applied to Deterministic Problems in the Natural Sciences
Johan G. R Belinfante and Bernard Kolman, A Survey of Lie Groups and Lie Algebras with
Applications and Computational Methods
James M. Ortega, Numerical Analysis: A Second Course
Anthony V. Fiacco and Garth P. McCormick, Nonlinear Programming: Sequential Unconstrained
Minimization Techniques
R H. Clarke, Optimization and Nonsmooth Analysis
George F. Carrier and Carl E. Pearson, Ordinary Differential Equations
Leo Breiman, Probability
R. Bellman and G. M. Wing, An Introduction to Invariant Imbedding
Abraham Berman and Robert J. Plemmons, Nonnegative Matrices in the Mathematical Sciences
Olvi L. Mangasarian, Nonlinear Programming
*Carl Friedrich Gauss, Theory of the Combination of Observations Least Subject to Errors: Part One,
Part Two, Supplement. Translated by G. W. Stewart
Richard Bellman, Introduction to Matrix Analysis
U. M. Ascher, R. M. M. Mattheij, and R. D. Russell, Numerical Solution of Boundary Value Problems for
Ordinary Differential Equations
K. E. Brenan, S. L. Campbell, and L. R. Petzold, Numerical Solution of Initial-Value Problems
in Differential-Algebraic Equations
Charles L. Lawson and Richard J. Hanson, Solving Least Squares Problems
J.E. Dennis, Jr. and Robert B. Schnabel, Numerical Methods for Unconstrained Optimization and
Nonlinear Equations
Richard E. Barlow and Frank Proschan, Mathematical Theory of Reliability
Cornelius Lanczos, Linear Differential Operators
Richard Bellman, Introduction to Matrix Analysis, Second Edition
Beresford N. Parlett, The Symmetric Eigenvalue Problem
Richard Haberman, Mathematical Models: Mechanical Vibrations, Population Dynamics, and Traffic Flow
Peter W. M. John, Statistical Design and Analysis of Experiments
*First time in print.

Classics in Applied Mathematics (continued)
Tamer Baar and Geert Jan Olsder, Dynamic Noncooperative Game Theory, Second Edition
Emanuel Parzen, Stochastic Processes
Petar Kokotovic, Hassan K. Khalil, and John O'Reilly, Singular Perturbation Methods in Control: Analysis
and Design
Jean Dickinson Gibbons, Ingram Olkin, and Milton Sobel, Selecting and Ordering Populations: A New
Statistical Methodology
James A. Murdock, Perturbations: Theory and Methods
Ivar Ekeland and Roger Temam, Convex Analysis and Variational Problems
Ivar Stakgold, Boundary Value Problems of Mathematical Physics, Volumes I and II
J.M. Ortega and W. C. Rheinboldt, Iterative Solution of Nonlinear Equations in Several Variables
David Kinderlehrer and Guido Stampacchia, An Introduction to Variational Inequalities and Their
Applications
F.Natterer, The Mathematics of Computerized Tomography
Avinash C. Kak and Malcolm Slaney, Principles of Computerized Tomographic Imaging
R. Wong, Asymptotic Approximations of Integrals
0. Axelsson and V. A. Barker, Finite Element Solution of Boundary Value Problems:Theory and Computation
David R. Brillinger, Time Series: Data Analysis and Theory
Joel N. Franklin, Methods of Mathematical Economics:Linear and Nonlinear Programming,Fixed-PointTheorems
Philip Hartman, Ordinary Differential Equations, Second Edition
Michael D. Intriligator, Mathematical Optimization and Economic Theory
Philippe G. Ciarlet, The Finite Element Method for Elliptic Problems
Jane K. Cullum and Ralph A. Willoughby, Lanczos Algorithms for Large Symmetric Eigenvalue
Computations, Vol. I: Theory
M. Vidyasagar,Nonlinear Systems Analysis, Second Edition
Robert Mattheij and Jaap Molenaar, Ordinary Differential Equations in Theory and Practice
Shanti S. Gupta and S. Panchapakesan, Multiple Decision Procedures: Theory and Methodology
of Selecting and Ranking Populations
Eugene L. Allgower and Kurt Georg, Introduction to Numerical Continuation Methods
Leah Edelstein-Keshet, Mathematical Models in Biology
Heinz-Otto Kreiss and Jens Lorenz, Initial-Boundary Value Problems and the Navier-Stokes Equations
J. L. Hodges, Jr. and E. L. Lehmann, Basic Concepts of Probability and Statistics, Second Edition
George F.Carrier, Max Krook, and Carl E. Pearson, Functions of a Complex Variable: Theory and Technique
Friedrich Pukelsheim, Optimal Design of Experiments
Israel Gohberg, Peter Lancaster, and Leiba Rodman, Invariant Subspaces of Matrices with Applications
Lee A. Segel with G. H. Handelman, Mathematics Applied to Continuum Mechanics
Rajendra Bhatia, Perturbation Bounds for Matrix Eigenvalues
Barry C. Arnold, N. Balakrishnan, and H. N. Nagaraja, A First Course in Order Statistics
Charles A. Desoer and M. Vidyasagar, Feedback Sy;tems: Input-Output Properties
Stephen L. Campbell and Carl D. Meyer, Generalized Inverses of Linear Transformations
Alexander Morgan, Solving Polynomial Systems Using Continuation for Engineering and Scientific Problems
I. Gohberg, P. Lancaster, and L. Rodman, Matrix Polynomials
Richard W. Cottle, Jong-Shi Pang, and Richard E. Stone, The Linear Complementarity Problem
Rabi N. Bhattacharya and Edward C. Waymire, Stochastic Processes with Applications
To our parents
Copyright 1988 by the Society for Industrial and Applied Mathematics
This SIAM edition is an unabridged republication of the work first published by the
Macmillan Publishing Co., New York, 1974.
10 9
All rights reserved. Printed in the United States of America. No part of this book may
be reproduced, stored, or transmitted in any manner without the written permission of
the publisher. For information, write to the Society for Industrial and Applied
Mathematics, 3600 Market Street, 6th Floor, Philadelphia, PA 19104-2688 USA.
Library of Congress Cataloging-in-Publication Data
ISBN 978-0-898712-29-2

SJaJ1l is a registered trademark.
Foreword
THIS volume is a classic in the pedagogy of applied mathematics. Its three

authors are internationally known leaders in their respective research com-
munities, which span mathematical approaches in astronomy, biology, and
continuum mechanics. They are also naturally gifted teachers who have been
long concerned about curricula and other issues of education and they have
been remarkably successful in showing how mathematics can be applied in
the physical and biological sciences. Students and other readers will sense the
authors' enthusiasm and conviction and will learn that relatively elementary
mathematics can often provide substantial understanding of natural phe-
nomena.
Members of SIAM, the Society for Industrial and Applied Mathematics,
are naturally pleased that the authors of this classic have chosen us to be their
new publisher. Our members naturally have a variety of different conceptions
of what is applied mathematics, and many have strong opinions regarding
what content is optimal for various students. All will agree that teaching how
to apply mathematics is important, difficult, and unfortunately somewhat
neglected. I can personally recommend Lin and Segel since I've enjoyed using
it as a text for bright undergraduates from many majors and for math students
curious about applications.
With this publication, SIAM begins a new CLASSICS series of reprinted
textbooks and researchmonographs. We do so as a service to the community
and to foster an understanding of applied mathematics. While such lofty goals
motivate us, we must also be convinced that sales of these volumes will exceed
the printing and distribution costs. Becausesuch Oassics will express ideas
reflecting SIAM's own philosophy, we will promote the series in SIAM lit-
erature and in our overall marketing program. As with all SIAM publishing
efforts, we welcome your ideas and suggestions for the Oassics series. Indeed,
Professor Marshall Slemrod of the University of Wisconsin, Madison, deserves
our thanks for recommending this fine volume to us.
Robert E. O'Malley, Jr.

Rensselaer Polytechnic Institute
vii
Preface
THIS text, an introduction to applied mathematics, is concerned with the

construction, analysis, and interpretation of mathematical models that shed
light on significant problems in the natural sciences. It is intended to provide
material of interest to students in mathematics, science, and engineering at
the upper undergraduate and graduate level. Classroom testing of preliminary
versions indicates that many such students do in fact find the material in-
teresting and worthy of study.
There is little doubt that a course such as one based on this text should
form part of the core curriculum for applied mathematicians. Moreover,
in the last few years the professional mathematical community in the United
States has emphasized the importance of some exposure to applied mathe-
matics for all mathematics students. This exposure is recommended because
of its broadening influence, and (for future university mathematicians) be-
cause of the benefits it affords in preparing for the teaching of nonspecialists.
As for scientists and engineers,there is often little differencein their theoretical
work and that of an applied mathematician, so they should find something
of value in the approaches to problems that we espouse.
There are many books that present collections of useful mathematical
techniques and illustrate the various techniques by solving classical problems
of mathematical physics. Our approach is different. Typically, we select an
important scientific problem whose solution will involve some useful mathe-
matics. After briefly discussing the required scientific background, we for-
mulate a relevant mathematical problem with some care. (The formulation
step is often difficult. Not many books actually demonstrate this, but we
try to give due weight to the challenges involved in constructing our mathe-
matical models.) A new technique may then be introduced to solve the
mathematical problem, or a technique known in simpler contexts may be
generalized. In most instances we take care to determine what the mathe-
matical results tell us about the scientific processes that motivated the prob-
lem in the first place.
We use a "case study" approach by and large. Such an approach is not
without disadvantages. No strict logical framework girds the discussion,
and the range of applicability of the methods is not precisely delineated.
Heuristic and nonrigorous reasoning is often employed, so there is room for
doubt concerning the results obtained. But realistic problems often require
techniques that cannot at the moment be rigorously justified. There is a
stimulating sense of excitement in tackling such problems. Furthermore,
mathematicians and scientists frequently use heuristic reasoning and are
ix
x Preface
frequently called upon to determine for themselves whether a method used
to solve one problem can be adapted to solve another. Some such experience
should be part of each student's education.
This work was given a lengthy title because we wished to make its limita-
tions clear at the outset. A completely balanced introduction to applied
mathematics should contain material from the social and managerial sciences,
but we have restricted ourselves to the natural sciences. In rough proportion
to applied mathematical research, the topics in this volume are drawn mainly
from the physical sciences, but there is representation from chemistry and
biology. We treat probabilistic models to a lesser extent than would be
required in a balanced presentation, and our treatment emphasizes the
relationship between probabilistic and deterministic points of view. Our work
is also limited in its almost exclusive use of analytical methods; numerical
methods are mentioned many times but are treated only briefly.
One reason for restricting the topics covered is the authors' hesitation
to tread outside their areas of expertise. Another is the fact that the work
is already lengthy, so a wider purview would necessarily be either overlong
or superficial. In any case, much further study is required of the aspiring
mathematician or scientist-we hope that our work will form a foundation and
motivation for some of that study.
STYLE AND CONTENT

In our writing we have striven in most places for careful and detailed
exposition, even at the risk of wordiness; for a rigorous proof can be built
from its skeleton, but the reasoning of the applied mathematician often can
only be mastered if it is fully described.
The nature of applied mathematics precludes an approach that is organized
in a tight linear fashion. This has certain disadvantages. But among the
compensations is a high degree of flexibility in a book such as this. In par-
ticular, there is a large measure of independence among the three parts of the
present volume, which are the following:
PART A. An overview of the interaction of mathematics and natural
science.
PART B. Some fundamental procedures-illustrated on ordinary differen-
tial equations.
PART C. Introduction to theories of continuous fields.
There is considerable fu~er independence within each part-which we have
tried to enhance, even at the expense of repetition.
Two volumes have been written. This first volume provides ample material
for a balanced and self-contained introduction to a major part of applied
mathematics. The succeeding volume, described briefly in Section 1.1, pen-
etrates further into the subject, particularly in the classical areas of fluid me-
chanics and elasticity. Seereference to Segel (1987), page 595.
Preface
xi
The chapter titles, section titles, and subheadings in the table of contents
give a good outline of this volume. It is not necessary to read the chapters
in the given order. In Part A, for example, the only sequence which must be
kept is that of the two chapters on Fourier series. It would be helpful to
begin Part B by obtaining some understanding of nondimensionalization
and scaling as treated in Chapter 6, but this is not strictly necessary. Chapter 8
can only be appreciated if the preceding two chapters have been covered,
but this chapter can be omitted if relatively simple examples of the techniques
are deemed sufficient. Chapter 9 is a prerequisite for Chapter 10, but each
of the three sections of Chapter 11 is largely independent of the others and
of earlier material.
Part C, too, offers various possibilities. For example, one can skip much
of the material if one's goal is to obtain just enough understanding of the
basic equations to permit formulation of specific problems in one-dimensional
elasticity, inviscid flow, and potential theory. Or one can just study the first
two sections of Chapter 12, to gain a glimpse of continuum mechanics.
(Note: Section 12.l contains many new ideas in a few pages.)
Some features of our approach are the following.
(a) We proceed from the particular to the general.
(b) For our major examples we attempt to choose physical problems that
are important in their own right and also permit the illustration of major
mathematical techniques. Thus the Michaelis-Menten kinetics discussed in
Chapter IO are repeatedly referred to in biochemistry, and a full treatment
of the relevant mathematical problem provides an excellent illustration of
singular perturbation theory. To give another example, we discuss the stability
of a stratified fluid in Chapter 15-both as an illustration of inviscid flow
and as a motivation for studying stability theory for a system of partial
differential equations.
(c) We make a serious effort to examine the processes of deriving the
equatfons that model certain basic scientific phenomena, rather than merely
give plausibility arguments for using such equations. As an illustration of
this spirit, we mention that the differential equation which governs mass
conservation in a continuum is derived in four different ways in Section 14.1,
and several alternative approaches are examined in the exercises of that
section. One purpose of such an effort is to engender a secure understanding of
the equation in question. Another is to help those readers who might some
day wish to derive equations that model a phenomenon which had never
before been subjected to mathematical analysis.
(d) New ideas are frequently introduced in extremely simple physical
contexts. In Part B, for example, dimensional analysis, scaling, and regular
perturbation theory are first met in the context of the problem of a point
mass shot vertically from the surface of the earth. The qualitative features of
the phenomenon are correctly guessed by most people, and the relevant
differential equation is solved exactly in elementary courses. Yet, considerable
Preface
xii
effort is required to obtain a deep understanding of the problem. This effort

is worthwhile because it generates a grasp of concepts that are useful in far
more difficult situations.
(e) We try to make explicit various concepts and approaches that are
often mastered only by inference over a period of years. Examples are pro-
vided by our discussions of the basic simplification procedure and of scaling,
in Chapter 6.
(f) In historical remarks we have focused attention on humanistic aspects
of science, by emphasizing that the great structures of scientific theories
are gradually built by the strivings of many people. To illustrate the ongoing
nature of science, we have presented certain plausible theories which are
either incorrect (Newton's isothermal speed of sound-Section 15.3) or not
fully in accord with observation (the galactic model of Section 1.2), or highly
regarded but not yet fully accepted (as in the physiological flow problem
discussed in Chapter 8).
(g) Some rather lengthy examples are worked out in detail, e.g., the
perturbation calculations in Section 7.2, in response to student objections
that they are often asked in exercises and in examinations to solve much harder
problems than they have ever seen done in the text.
(h) We have provided a number of exercises to reinforce, test, and extend
the reader's understanding. Noteworthy are multipart exercises, often based
on a relatively recent journal article, which develop a major point in a step-
by-step manner. (An example is Exercise 15.2.10.) Even if a student cannot do
one part of the exercise, he can take its result for granted and proceed. Such
exercises have been successfully used as the central part of final examinations.
PREREQUISITES
We have assumed that the potential reader has had an introductory college
course in physics and is familiar with calculus and differential equations.
Only a few exercises require knowledge of complex analysis. We make
considerable use of such topics as directional derivatives, change of variables
in multiple integrals, line and surface integrals, and the divergence theorem.
Often mathematics majors wilJ have taken an advanced calculus course that
omits some of these topics, but we have found that such students are suffi-
ciently sophisticated mathematically to be able readily to pick up by them-
selves what is required. [Potential readers who feel inadequacies in vector
calculus and physical reasoning would profit from studying Div, Grad, Curl,
and All That by H. M. Schey (N.Y., Norton, 1973).]
RELATIONSHIPS BETWEEN THIS TEXT AND VARIO US COURSES

Historically, this book grew out of the union of two courses. The first
was Foundations of Applied Mathematics introduced by G. H. Handelman
at Rensselaer around 1957. (A precursor of this course wastaught by A. Schild
and Handelman at Carnegie Institute of Technology, now Carnegie-Mellon
Preface xiii
University.) A second course, Introduction to Applied Mathematics, was

introduced by C. C. Lin at Massachusetts Institute of Technology around
1960. These courses have been taught annually, many times, by the present
authors in their respective institutions. In recent years preliminary drafts
of the present work have been used as text material. Such drafts have also
been used in applied mathematics courses taught by P. Davis at Worcester
Polytechnic Institute, D. Drew at New York University, and D. Wollkind
at Washington State University. Considerable improvements in the draft
text have resulted; the authors welcome further suggestions from users of the
printed work.
ACKNOWLEDGMENTS
The work was jointly planned. One of us (C.C.L.) wrote the initial draft
of Part A, with the exception of Section 1.3, and also Chapter 16. The other
of us (L.A.S.) drafted the remainder of the book, with the exception of
Chapter 12, written by G. H. Handelman. There was considerable consulta-
tion on revisions. L.A.S. was responsible for the final editorial work.
In writing this book, we have drawn on a background for which we are
deeply indebted to our families, colleagues, teachers, and students, and to the
writers of numerous other books. We are grateful to G. H. Handelman for
showing, from classroom experience, that an introduction to continuum
mechanics is probably best made by starting with one-dimensional problems-
and for writing out his approach as Chapter 12. Among many who have made
useful suggestions concerning this volume, we must single out Roy Caplan,
Paul Davis, Donald Drew, William Ling, Robert O'Malley, Jr., Edward
Rothstein, Terry Scribner, Hendrick Van Ness, and especially Edith Luchins.
Numbers of secretaries have performed yeoman service. The publishers have
also been most helpful, particularly our editors Everett Smethurst and Elaine
Wetterau.
The work of one of us (L.A.S) was partially supported in 1968-1969
by a Leave of Absence Grant from Rensselaer. Further support was received
during 1971..,..1972from the National Science Foundation Grant GP33679X
to Rensselaer, and from a John Simon Guggenheim Foundation Fellowship.
That year was spent as a visitor to the Department of Applied Mathematics,
Weizmann Institute, Rehovot, Israel. The hospitality and technical assis-
tance afforded there are gratefully acknowledged.
C.C.L.
L.A. s.
1994 REPRINTING
A number of misprints have been corrected and a few small changes made,
thanks to Prof. F. Kemp of the University of Tulsa. The authors welcome
further suggestio:QSfrom their readers. Write to Lee Segel, Dept. of Applied
Mathematics and Computer Science, Weizmann Institute, Rehovot, Israel.
Fax: 972-8-344-122. E-mail: [email protected].
A B
Two natural phenomena that arc c:urrcntly U?ider investigation by applied mathematicians.
As explained in Chapter I, the power of mathematic:al modeling is well illustrated by these
examples. Although stars and arncbae really are discrete collections of matter. nonetheless
much can be learned from models in which the matter is ~ded as continuously dis-
tributed in space. [a] Whirlpool galaxy and its.companion (NGC 5194/5). (Photo-
graph courtesy of the Hale Observatories.) [bl Aggregation of cellular slime :::old amebae
Polysplwndylium z;ioiaceum. (Courtesy B. Shaffer, Department of Zooiogy. Cambridge
Uciversity .)
Contents
PART A
AN OVERVIEW OF THE INTERACTION OF
MATHEMATICS AND NATURAL SCIENCE
CHAPTER 1
WHAT Is APPLIED MATHEMATICS? 3
I. I. On the nature of applied mathematics 4
The scope, purpose, and practice of applied mathematics 5 I Applied
mathematics contrasted with pure mathematics 6 I Applied mathematics
contrasted with theoretical science 7 I Applied mathematics in engineering 8 /
The plan of the present volume 8 I A preview of the fallowing volume 9 /
Concepts that unify applied mathematics 10
1.2 Introduction to the analysis of galactic structure 10

Physical laws g(l/)erninggalactic behavior 10 / Building blocks of the
universe 11 I Classification of galaxies I.I / Composition of galaxies 13 /
Dynamics of stellar systems 14 / Distribution of stars across a galactic disk 16 /
Density wave theory of galactic spirals 19
1.3 Aggregation of slime mold amebae 22

Some facts about slime mold amebae 22 I Formulation of a mathematical
model 24 I An exact solution: The uniform state 28 / Analysis of aggregation
onset as an instability 28 I Interpretation of the analysis 30
APPENDIX 1.l Some views on applied mathematics 31

On the nature of applied mathematics 3 I / On the relationships among pure
mathematics, applied mathematics, and theoretical science 32 / On the teaching
and practice of applied mathematics 33
CHAPTER 2
DETERMINISTIC SYSTEMS AND ORDINARY
DIFFERENTIAL EQUATIONS 36
2.1 Planetary orbits 36
Kepler's laws 37 I I.Aw of universal gravitation 39 / The inverse problem:
Orbits of planets and comets 39 I Planetary orbits according to the general
theory of relativity 41 I Comments on choice of methods 41 / N particles: A
deterministic system 42 I Linearity 43
xv
xvi Contents
2.2 Elements of perturbation theory, including Poincare's method for

periodic orbits 45
Perturbation theory: Elementary considerations 46 I The simple pendulum 48 I
Successive approximations to the motion of the pendulum 50 I Perturbation
series applied to the pendulum problem 51 I Poincare's perturbation theory 53 I
Generalization of the Poincare method 54
2.3 A system of ordinary differential equations 57

The initial value problem: Statement of theorems 57 I Proof of the uniqueness
theorem 6o / Proof of the existence theorem 61 / Continuous dependence
on a parameter or initial conditions 63 / Differentiability 64 / Example of
nonuniqueness 66 I Method of finite differences 66 I Further remarks on the
relation between "pure" and" applied" mathematics 68
CHAPTER 3
RANDOM PROCESSES AND PARTIAL DIFFERENTIAL
EQUATIONS 71
3.1 Random walk in one dimension; Langevin's equation 73
The one.dimensional random walk model 73 I Explicit solution 74 / Mean,
variance, and the generating function 75 / :use of a stochastic differential
equation to obtain Boltzmann's constant from observations of Brownian
motion 78
3.2 Asymptotic series, Laplace's method, gamma function, Stirling's

formula 80
An example: Asymptotic expansion via parts integration 82 I Definitions in
the theory of asymptotic expansions 83 I Laplace's "'4thod 84 f Develop1n4nt
of the asymptotic Stirling series for the gamma function 86 I Justification
of term-by-term integration 89
3.3 A differenceequation and its limit 91

A difference equation for the probability function 91 I Approximation of the
difference equation by a differential equation 92 I Solution of the differential
equation for the probability distribution function 93 I Further examination
of the limiting process 94 I Reffecting and absorbing barriers 95 I Coagulation:
An application of first passage theory 97
3.4 Further considerations pertinent to the relationship between

probability and partial differential equations 99
More on the diffusion equation and its connection with random walk 100 I
Superposition of fundamental solutions: Themethod of images 102 / First
passage time as a flux 104 I General initial valueproblem in diffusion 104 / How
twisting a beam can give information about DNA molecules 106 / Recurrence
property in Brownian motion 108
APPENDIX 3.1 0 and o symbols 112

Contents
xvii
CHAPTER 4
SUPERPOSITION, HEAT FLOW, AND FOURIER ANALYSIS II4
4.1 Conduction of heat II s

Steady state heat conduction 116 I Differential equationfor one-dimensional
heat conductionI 17 / Initial boundary valueproblem for one-dimensional
heat conduction I 18 I Past, present, andfuture I 19 I Heat conductionin
three-dimensionalspace 120 I Proof of the uniquenesstheorem 122 f The
maximum principle 123/ Solution by the method of separationof variables123/
Interpretation;dimensionlessrepresentation127 / Estimate of the time required
to diffuse a given distance 129
4.2 Fourier's theorem 131

Summation of the Fouriersine series I 31 I Proof of the lemmas I 34 I A formal
transformationI 35 I Fourierseries in the full range 135 I Summation of Fourier
series 136 I Half-rangeseries 137
4.3 On the nature of Fourier series I 37

Fourierseries for the constant/'unction 138 / Fourier seriesfor the linear
/'unction 139 I Fourierseriesfor the quadraticfunction 139 I Integration and
differentiationof Fourierseries 139 / Gibbsphenomenon 141 I Approximation
with least squared e"or 144 I Bessel's inequality andParseval's theorem 146 /
Riesz-Fischer theorem 147 I Applications of Parseval's theorem 147
CHAPTER 5
FURTHER DEVELOPMENTS IN FOURIER ANALYSIS 150
5.1 Other aspects of heat conduction 150

Variationof temperatureunderground150 / Numerical integration of the heat
equation 153 / Heat conductionin a nonuniform medium 155
S.2 Sturm-Liouville systems 159
Propertiesof eigenvaluesand eigenfunctions16o I Orthogonalityand
normalization161 I Expansion in terms of eigenfunctions 162 I Asymptotic
approximationsto eigenfunctionsand eigenvalues163 J Other methods of
calculatingeigenfunctionsand eigenvalues165
S.3 Brief introduction to Fourier transform 167

Fouriertransformformulas and the Fourier identity 167 I Solution of the heat
equationby Fourier transform 170
5.4 Generaliud harmonic analysis 171

Remarks on functions that cannot be analyzed by standard Fourier
methods 172 I Fourierseries analysis of a truncated sinusoidalfunction 173 I
Fourier integral analysisof a truncatedsinusoidalfunction 174 I
Generalizationto stationary time sequences176 I Autocorrelationfunction and
the power spectrum 178 I Verijicationof the cosine transform relation
between the power spectrum and the Qlltoco"elation 179 / Application 180
xviii Contents
PART B
SOME FUNDAMENTAL PROCEDURES
ILLUSTRATED ON ORDINARY DIFFERENTIAL
EQUATIONS
CHAPTER 6
SIMPLIFICATION, DIMENSIONAL ANALYSIS, AND SCALING 185
6.1 The basic simplification procedure 186
Illustrations of the procedure 186 J Two chastening examples 187 / Conditioning
and sensitivity 189 / Zeros of a/unction 190 I Second order differential
equations 191 / Recommendations 194
6.2 Dimensional analysis 195
Putting a differential equation into dimensionless form 195 I
Nondimensionalization of a functional relationship 198 I Use of scale models
200 I Summary 203
6.3 Scaling 209

Definition of scaling 211 I Scaling the projectile problem 211 I Order of
magnitude 213 I Scaling known functions 214 I Orthodoxy 218 I Scaling and
perturbation theory 221 I Scaling unknown functions 222
CHAPTER 7
REGULAR PERTURBATION THEORY 225
7.1 The series method applied to the simple pendulum 225
PrelimillllTies 226 I Series method 227 I Discussio11of results so far 230 /
Higher order terms 23 l
7.2 Projectile problem solved by perturbation theory 233
Series method 233 I Parametric differentiation 236 I Successive approximations
(method of iteration) 238 I General remarks on regular perturbation theory 240
CHAPTER 8
ILLUSTRATION OF TECHNIQUES ON A PHYSIOLOGICAL FLOW
PROBLEM 244
8.1 Physical formulation and dimensional analysis of a model for
"standing gradient., osmotically driven :flow 244
Some physiological facts 244 I Osmosis and the osmol 247 I Factors that affect
standing gradient flow 248 I Dimensional analysis of a functional relationship
249 I Possibility of a scale model for standing gradient /low 253
8.2 A mathematical model and its dimensional analysis 253
Conservation of fluid mass 254 I Conservation of solute mass 254 / Boundary
conditions 255 I Introduction of dimensionless variables 257 I Comparison of
physical and mathematical approaches to dimensional analysis 259
Contents
xix
8.3 Obtaining the final scaled dimensionless form of the mathematical
model 261
Scaling 262 I Estimating the size of the dimensionless parameters 264 / An
unsuccessful regular perturbation calculation 265 / Relation between
parameters 265 / Final formulation 267
8.4 Solution and interpretation 268
A first approximation to the solution 268 I Comparison with numerical
calculations 269 I Interpretation: Physical meaning of the dimensionless
parameters 272 I Final remarks 274
CHAPTER 9
INTRODUCTION TO SINGULAR PERTURBATION THEORY 277
9.1 Roots of polynomial equations 278
A simple problem 278 I A more complicated problem 281 / The use of scaling 284
9.2 Boundary value problems for ordinary differential equations 28 5
Examination of the exact solution to a model problem 285 I Finding an
approximate solution by singular perturbation methods 291 / Matching 293 /
Further examples 295
CHAPTER 10
SINGULAR PERTURBATION THEORY APPLIED TO A PROBLEM
IN BIOCHEMICAL KINETICS 302
IO.I Formulation of an initial value problem for a one enzyme-one
substrate chemical reaction 302
The law of mass action 302 I Enzyme catalysis 304 I Scaling and final
formulation 306
10.2 Approximate solution by singular perturbation methods 308
Michaelis-Menten kinetics as an outer solution 308 I Inner solution 308 I A
uniform approximation 310 / Comments on results so far 311 / Higher
approximations 311 J Further analysis for large times 315 / Further discussion
of the approximate solutions 317
CHAPTER 11
THREE TECHNIQUES APPLIED TO THE SIMPLE PENDULUM 321
11.1 Stability of normal and inverted equilibrium of the pendulum 321
Determining stability of equilibrium 322 I Discussion of results 323
11.2 A multiple scale expansion 325
Substitution of a two-scale series into the pendulum equation 327 I Solving
lowest order equations 328 / Higher approximations; removing resonant
terms 328 / Summary and discussion 330
11.3 The phase plane 334
The phase portrait of an undamped simple pendulum 335 I Separatrices 337 I
Critical points 338 / Limit cycles 339 I Behavior of trajectories near critical
points 339
xx Contents
PART C
INTRODUCTION TO THEORIES OF CON"YINUOUS
FIELDS
CHAPTER 12
LONGITUDINAL MOTION OF A BAR 349
12.1 Derivation of the governing equations 349
Geometry 349 I The material derivative and the Jacobian 352 / Conservation
of mass 355 I Force and stress 358 I Balance of linear momentum 36o / Strain
and stress-strain relations 362 J Initial and boundary conditions 366 /
Linearization 368
12.2 One-dimensionalelastic wave propagation 376

The wave equation 377 I General solution of the wave equation 377 J Physical
significance of the solution 378 I Solutions in complex form 379 I Ar,alysisof
sinusoidal waves 380 I Effects of a discontinuity in properties 381
12.3 Discontinuous solutions 388

Motion of the discontinuity surface 390 / Behavior of the discontinuity 396
12.4 Work, energy, and vibrations 401

Work and energy 401 I A vibration problem 403 I The Rayleigh quotient 404 I
Properties of eigenvalues and eigenfunctions 405 / An exact solution when
properties are constant 406 / Characterization of the lowest eigenvalue as the
of the Rayleigh quotient 406 I Estimate of the lowest eigenvalue for
mini111111n
a wedge407
CHAPTER 13
THE CONTINUOUS MEDIUM 412
13.1 The continuum model 413
Molecular averages 414 / Mass distribution functions 416 / The continuum as
an independent model 417
13.2 Kinematics of deformable media 418

Points and particles 419 / Material and spatial descriptions 420 I Streamlines
and particle paths 422 / A simple kinematic boundary condition 425
13.3 The material derivative 426

13.4 The Jacobian and its material derivative 429
APPENDIX 13.1 On the chain rule of partial differentiation 433
APPENDIX 13.2 The integral mean value theorem 435
APPENDIX 13.3 Similar regions 436
Contents
xxi
CHAPTER 14
FIELD EQUATIONS OF CONTINUUM MECHANICS 440
14.l Conservation of mass 440
Integral method: Arbitrary material region 441 / Integral method: Arbitrary
spatial region 444 I Small box metfwd 444 I Large box method 446
14.2 Balance of linear momentum 454
An integral form of Newton's second law 454 / Local stress equilibrium 456 i
Action and reaction 457 / The stress tensor 458 / Newton's second law in
differential equation form 461
14.3 Balance of angular momentum 465
Torque and angular momentum 465 I Polar fluids 466 J Symmetry of the stress
tensor 466 f The principle of local moment equilibrium 467 I Local moment
equilibrium for a cube 468
14.4 Energy and entropy 470
Ideal gases 471 I Equilibrium thermodynamics 474 / Effects of inhomogeneity
and motion 475 / Energy balance 476 I Entropy, temperature, and pressure 477 I
Internal energy and deformation rate 479 I Energy and entropy in fluids 480
14.5 On constitutive equations, covariance, and the continuum model 485
Recapitulation of field equations 485 I Introduction to constitutive
equations 486 / The principle of covariance 487 / Validity of classical continuum
mechanics 490
APPENDIX 14.l Thermodynamics of spatially homogeneous
substances 491
Experiments with a piston 49J I First law of equilibrium thermodynamics 494 /
Entropy 496 / Second law of equilibrium thernwdynanucs 500
APPENDIX 14.2 Some historical remarks 501
CHAPTER 15
INVISCID FLUID FLOW 505
15.1 Stress in motionless and inviscid :fluids 505
Molecular point of view 5o6 / Continuum point of view 5o6 / Hydrostatics 508 f
Inviscid fluids 510
15.2 Stability of a stratified fluid 515

Governing equations and their exact equilibrium solution 516 / Linearized
equations for the perturbations 518 / Characterization of the growth rate u
as an eigenvalue 521 I Qualitative general deductions 523 I Detailed results for
a particular stratification 525 I Superposition of normal modes 528 / Nonlinear
effects 530 I Worked example: A model of viscous flow instability 531
15.3 Compression waves in gases 539

Inviscid isentropic flow of a perfect gas 540 I Waves of small amplitudes 541 /
The speed of sound 542 I Spherical waves 543 I Nonlinear war,es in one
dimension 544 I Shock waves 547
xxii Contents
15.4 Uniform flow past a circular cylinder 551

Formulation 551 I Solution by separation of variables 553 / Interpretation of
solution 555
APPENDIX 15.1 A proof of D'Alembert's paradox in the
three-dimensional case 560
APPENDIX 15.2 Polar and cylindrical coordinates 563
CHAPTER 16
POTENTIAL THEORY 566
16.1 Equations of Laplace and Poisson 566
Gravitational potential of discrete mass distributions 566 / Gravitational
potential of continuous mass distributions 567 I Theorems concerning harmonic
functions 570 I Integral representation for the solution to Poisson's equation
572 / Uniqueness 573
16.2 Green's functions 574
Green's function for the Dirichlet problem 574 I Representation of a harmonic
function using Green's function 575 I Symmetry of the Green's function 575 I
Explicit formulas for simple regions 575 / Widespread utility of source, image,
and reciprocity concepts 577 I Green's function for the Neumann problem 578 I
Green's function for the Helmholtz equation 579
16.3 Diffraction of acoustic waves by a hole 580
Formulation 580 I Selection of the appropriate Green's function 583 I
Derivation of the diffraction integral 584 / Approximate evaluation of the
diffraction integral 586
BIBLIOGRAPHY 589
SUPPLEMENT ARY BIBLIOGRAPHY 595
UPDATED MATERIAL 598
HINTS AND ANSWERS (for exercises marked with;) 600
AUTHORS CITED 604
SUBJECT INDEX 605
Conventions
EACH chapter is divided into several sections (e.g., Section 5.2 is the second
section of Chapter 5). Equations are numbered consecutively within each
section. Figures and tables are numbered consecutively within each chapter.
When an equation outside a given section is referred to, the se<..-tionnum-
ber precedes the equation number. Thus" Equation (6.3.2)" [or (6.3.2)] refers
to the second numbered equation of Section 6.3. But if this equation were
referred to within Section I of Chapter 6, then the chapter number would be
assumed and the reference would be to "Equation (3.2)" [or (3.2)]. The
fourth numbered equation in Appendix 3.1 is denoted by (A3.l.4).
A double dagger; preceding an Exercise, or a part thereof, signifies that a
hint or an answer will be found in the back of this volume.
The symbol Osignifies that the proof of a theorem has concluded.
The succeeding volume is ref erred to as "II."
A brief bibliography of books useful to beginning applied mathematicians
can be found at the end of this volume. When reference is made to one of
these books, the style "Smith (1970)" is employed.
xxiii
CHAPTER 1
What Is Applied Mathematics?
THE PURPOSE of this book is to foster an appreciation of the nature

of applied mathematics. We attempt to explain the essential processes involved
and to provide experience in applied mathematics as it is used in a variety
of physical problems.
The book is not intended as a compilation of methods-which cannot be
complete in any case. Rather, it is believed that an appreciation of the
applied mathematician's approach will provide the student with a clear
framework on which to fit his knowledge, and with a creative attitude. These
are much more valuable than an encyclopedia of methods.
Section 1.1 begins with some fairly detailed remarks on the nature of
applied mathematics. Then follows an outline of the applied mathematical
topics that are to be presented in this work. Section 1.2 consists of an intro-
duction to the study of galactic structure, and Section 1.3 analyzes the sudden
aggregation of a population of amebae. The main purpose of Sections 1.2
and 1.3 is to provide an early feeling for the type of problem now under
investigation by applied mathematicians. (The authors selected the topics
from among their own current interests.) On first perusal at least, the reader
should not concern himself unduly with detail but should concentrate on
getting a general impression.
The behavior of stars and the behavior of amebae might seem to have little
in common. It is an excellent illustration of the power of applied mathematics,
however, to realize that this is not the case. For example, we shall see that
both stars and amebae behave as if their mass were continuously distributed-
notwithstanding the fact that one can see (using a microscope) gaps between
amebae, while the nearest stars are light-years apart. But it is of little relevance
that an idealized mathematical model such as the continuum is "wrong."
The real question is: Do the errors have a significant effect upon predictions
concerning the phenomena under investigation?
For stars and amebae the idealization implied in the continuum model is
particularly striking, since it is the particulate aspect of the phenomena
that first makes itself evident. Water and air seem like continua-for most
people the existence of molecules is hearsay-but the continuum model is,
of course, also very useful for investigating phenomena involving these
media.
Another similarity between the examples of Sections 1.2 and 1.3 is that
both require an explanation of organized structure. The spiral galactic
pattern demands an investigation for reasons of aesthetics alone, but a more
basic reason is that the pattern offers an important clue to the nature of the
3
4 What Is Applied Mathematics? [Ch. 1
forces involved. The morphogenetic(form-producing) motion of the amebae

is studied because of the clues it can provide to the organized cellular move-
ment that is a basic element of developmental biology. In both areas
understanding is far from complete.
1.1 On the Nature of Applied Mathematics

Mathematics began with simple practical problems such as division of a
flock of animals among family members (number theory) and the measure-
ment of land area (geometry). Gradually, elementary ideas were organized,
and they evolved into logical structures. A monumental example of early
achievements is Euclid's geometry. The Greeks realized that the study of
mathematical theorems can be based on certain axioms. Only much later,
however, did it become clear that the axioms (such as the parallel postulate)
cannot be completely and decisively verified by experience. Indeed, it was
through a change in the parallel postulate that non-Euclidean geometry was
created; this led to many important ramifications. There is no doubt that
these pure mathematical developments are extremely important in appli-
cations.
As an increasing portion of mathematics was developed in a manner
independent of theoretical sciences, the term pure mathematics emerged. The
creative efforts of pure mathematicians are certainly impressive, but it would
be unfortunate to restrict our study of mathematics in any manner. Joint
study of mathematics and its applications can provide richer content and
greater intellectual challenge. Moreover, such study can stimulate the
development of new mathematical methods and theory. Some of these
developments will in turn find applications in the sciences.
Applied mathematics is guided by the spirit of and belief in the interdepen-
dence of mathematics and the sciences. It would be incorrect to claim that
all parts of mathematics involve such interdependence, but in the study of
applied mathematics one must give priority to such parts. This policy is
based on the assumption that the areas of mathematics that grew directly out
of the study of scientific problems have a greater likelihood of again being
applicable to other scientific problems. To give one example, later in this
chapter we see an application to a current problem in developmental biology
of stability theory for partial differential equations. This theory was developed
primarily in the classical fields of fluid mechanics and elasticity.
Historically, the development of mathematics and physics had a very close
connection. Classical examples may be found in the work of Newton (see
Chapter 2), Gauss, Euler, Cauchy, and others. More modern examples
appear throughout the study ofrelativity theory, Brownian motion, statistical
mechanics, and the associated theories of covariance, probability, and general-
ized harmonic analysis. Why is there such an intimate relationship between
mathematics and the physical sciences?
Sec. I .I] On the Nature of Applied Mathematics s
An answer often given* to this philosophical question is that

GOD IS A MATHEMATICIAN.
In other words, it is believed that there is a basic harmony and order in nature.
Consequently, the description of natural phenomena can be organized by the
logical discipline of mathematics. In social and economic contexts, by
contrast, perhaps it is the logic imposed by man in his quest for optimality
that allows so large a role for mathematics.
THE SCOPE, PURPOSE, AND PRACTICE OF APPLIED
MATHEMATICS
The scope of applied mathematics is very broad and can be well described
by borrowing the following words of Albert Einstein:
Its realm is accordinglydefined as that part of the sum total of our knowledge
which is capable of being expressedin mathematicalterms.t
These words were used by Einstein to define physics. Taken literally, this
definition would certainly include mathematical theories of biology, econom-
ics, communication engineering, etc., and it is therefore a more adequate
description of applied mathematics.
We shall now attempt to give a brief description of the objectives and the
methodology of applied mathematics, and to contrast them with those of
pure mathematics and of theoretical science.
The purpose of applied mathematics is to elucidate scientific concepts and
describe scientific phenomena through the use of mathematics, and to
stimulate the development of new mathematics through such studies. The
process of using mathematics for increasing scientific understanding can be
conveniently divided into the following three steps:
(i) The formulation of the scientific problem in mathematical terms.
(ii) The solution of the mathematical problems thus created.
(iii) The interpretation of the solution and its empirical verification m
scientific terms.
There is widespread misunderstanding that the second step is the most
important and that manipulative skill is the most valued asset of an applied
mathematician. Generally speaking, however, all three steps are equally
important. In a given class of problems, one step might stand out as more
important or more difficult than another.
A knowledge of methods and a proficiency in manipulative skill are
obviously necessary. A person with an encyclopedic knowledge of methods
" God is number" -Pythagoras; " God ever geometrizes "-Plato; " God ever arithmetizes"
-Jacobi; "The Great Architect of the Universe now begins to appear as a mathematician"-
Jeans. Quotes given (without much approval) by E. T. Bell, Men of Mathematics (London:
Penguin, 1953), p. 21.
t Out of My Later Years (New York: Philosophical Library, 1950), p. 98.
alone can provide valuable aid to other scientists using mathematics. But it is
essential to realize that this knowledge is not sufficient for an applied mathe-
matician working as an independent scientist. He must also be able to
exercise judgment in the formulation of a problem, in decisions on which
problems to attack, and in choices of what idealizations and approximations
to adopt in order to simplify the problem without losing its essentials. An
aspiring applied mathematician must try even harder to cultivate sound
judgment than manipulative skill.
Finally, it is most important to extract, from the mathematical theory, the
proper scientific conclusions and implications for empirical verification. As far
as is feasible, conclusions must be reduced to the simplest form and expressed
in the most specific terms. This step serves as a culmination of the whole
effort and as a basis for future progress. A new understanding obtained, an
insight gained, a new perspective attained-these are much more important
and satisfying than the mere derivation of certain formulas and the compila-
tion of certain useful numerical tables. The accumulation of specific quanti-
tative information must be regarded as a means to an end.
It should now be apparent that an understanding of the scientific motivation
of the problem and the ability to use heuristic reasoning, as well as mani-
pulative skill, are essential to the practice of applied mathematics. Indeed,
these elements stand out as being more important than the ability to carry
out a rigorous proof. In many cases the rigorous formulation of a mathe-
matical theory may take many years. In the meantime, the applied mathe-
matician must proceed despite the incompleteness of the logical structure.
However, he must strive to be correct and to be as careful as possible in his
reasoning.
APPLIED MATHEMATICS CONTRASTED WITH PURE

MATHEMATICS
The differences in motivation and objectives between pure and applied
mathematics-and the consequent differences in emphasis and attitude-
must be fully recognized. In pure mathematics, one is often dealing with
such abstract concepts that logic remains the only tool permitting judgment
of the correctness of a theory. In applied mathematics, empirical verification
is a necessary and powerful judge.
However, there is still a close relationship between the two disciplines.
In some cases (e.g., celestial mechanics), rigorous theorems can be proved
that are also valuable for practical purposes. On the other hand, there are
many instances in which new mathematical ideas and new mathematical
theories are stimulated by applied mathematicians or theoretical scientists.
The theory of distributions is a fairly recent example.
If a scientific problem cannot be adequately formulated in terms of existing
mathematical concepts, new concepts have to be generated, such as the
abstract "game" of von Neumann. If the mathematical problem formulated
Sec. I.I] On the Nature of Applied Mathematics
cannot be solved in terms of existing methods, or the nature of its solution

cannot be adequately understood in terms of existing theory, new methods
and new theory must be developed. (Many nonlinear problems are in this
category.) We therefore record a fourth part of applied mathematics:*
(iv) The generation of scientifically relevant new mathematics through
creation, generalization, abstraction, and axiomatic formulation.
It should be recognized that as the mathematical theory is being developed,
the first few theorems proved may not produce an impact in pure mathematics
but must be appreciated as an accomplishment that is useful for the purposes
of applied mathematics. On the other hand, much second-rate pure mathe-
matics is concealed beneath the trappings of applied mathematics (and vice
versa). As always, knowledge and taste are needed if quality is to be assured.
APPLIED MATHEMATICS CONTRASTED WITH
THEORETICAL SCIENCE
The distinction between a theoretical scientist and an applied mathematician
is often blurred, because each may work in the spirit of the other. A theoretical
physicist, when his problem cannot be solved by frontal attack, sometimes
engages in the study of related mathematical model problems-even to the
extent of following the practices of pure mathematicians-in order to build
up confidence and judgment in understanding the mathematical aspects of
the real physical problem. This type of work is often also done by applied
mathematicians for their own problems. At the same time, an applied
mathematician draws scientific conclusions from his theory in order to
compare with empirical evidence. To do this effectively, he must command
considerable scientific knowledge of the problem he is studying.
It is often the case that a theoretical scientist, from long study of his
particular subject, has a deeper knowledge of a certain discipline. An applied
mathematician, by contrast, may work in more than one discipline and
cross-fertilize each. Indeed, in these times of increasing specialization, cross
fertilization is one of the most useful and satisfying activities of an applied
mathematician.
By and large, a theoretical physicist (for example) is more attracted to the
discovery of new physical laws and principles, and therefore has more
appreciation for studies related to these, even when the attempt is only
partially successful. His work often tends to be more inductive and more
speculative in nature. The applied mathematician is more interested in the
proper mathematical descriptions of phenomena. He tends to derive the
consequences of known Jaws and principles.
There are those who regard this fourth part as the only applied mathematics and consider
the first three parts as science, not mathematics. Others find most of the work that is classifiable
under this fourth part to be of little genuine scientific relevance (so that it should be judged
primarily on its mathematical merit). The authors are not doctrinaire, but they lean more toward
the latter view. And with the exception of some material in Chapter 2, this book concentrates on
parts (i)-(iii), for it is here that there is a gap in the literature.
To give an illustration, the origin of hurricanes would not be a subject of

central interest to a modern physicist, for hurricane genesis can probably be
fully described in terms of the principles of classical mechanics and thermo-
dynamics. Yet it is a very attractive subject for the applied mathematician,
who appreciates the challenge of the nonlinear problems involved as well as
the inherent scientific interest of the phenomenon.
APPLIED MATHEMATICS IN ENGINEERING

Understanding of basic concepts is as useful to the engineer as it is to the
scientist. But for the engineer such understanding is just a means to the end
of designing structures, machines, and processes to accomplish certain tasks
with efficiencyand reliability. The requirement to develop detailed design
criteria often makes extensive l~rge scale numerical calculations indispensable.
At the very least, however, the engineer should be able to appreciate the more
qualitative model-making activities described in this work, so that he can take
advantage of relevant theoretical results. And as a lofty goal, the engineer can
aspire to combine applied mathematical and scientific insight with practicality,
as have the masters of his profession.*
THE PLAN OF THE PRESENT VOLUME
The reader will find some others' views concerning the nature, teaching,
and practice of applied mathematics in Appendix 1.1.The authors' views have
been outlined above; in a sense, the remainder of this work is an elaboration
of that outline.
This volume is divided into three parts. In contrast to the rest of the work,
which is rather detailed, Part A provides an overview of applied mathematics.
As above, it begins with an essay on the nature of applied mathematics.
This chapter continues by presenting an introduction to two examples
of current research in applied mathematics. Chapters 2 and 3 contrast
deterministic processes, as exemplified in the equations of particle dy-
namics, and probabilistic processes, as exemplified in the study of random
walk. Part A concludes with two chapters on Fourier series, a classical topic
that is not only still useful but continues to develop. The main concept
illustrated here is the principle of superposition.
In Part B an attempt is made to convey the essence of certain important
applied mathematical methods and ways of thinking. Particularly in the
discussion of the basic simplification procedure and of scaling in Chapter 6,
this involves an attempt to make explicit certain approaches that "everyone
knows" after a while, but which are rarely treated in print. The procedures
* An entertaining and informative biography of one such master is The Wind and Beyond by
Theodore von Karman (the subject of the book) and Lee Edson (Boston: Little, Brown, 1967).
Von Karman used fundamental applied mathematical investigations to estimate the behavior of
engineering systems and thereby to play a major role in airplane design (from the earliest days
through jets) and the design of turbines, pumps, tunnels, dams, and bridges.
Sec. J .J] On the Nature of Applied Mathematics 9
in question have a large intuitive element, but still there is something to be

learned by a detailed examination of their content.
Both regular and singular perturbation theory are discussed in Part B,
as is the phase plane method. Two chapters are devoted to careful discussions
of nontrivial examples that involve the various techniques and show some
processes of applied mathematics at work. One example involves an osmoti-
cally driven flow that is of interest in physiology: the other considers some
important aspects of biochemical kinetics. Throughout Part B only ordinary
differential equations are considered so that the basic ideas are illustrated in
relatively simple contexts.
Part C provides an introduction to continuum mechanics. Training in this
classical theory is important to the applied mathematician for two main
reasons: (a) Continuum mechanics remains an important subject in its own
right, and its methods (as we shall see several times in this volume) are
continually being applied in new areas; and (b) the depth that is possible in
theoretical analysis must be illustrated in a well-developed field of study.
Continuum mechanics is such a field and (unlike quantum mechanics for
example) deals with relatively familiar physical phenomena.
Part C begins with a simple but revealing study in continuum mechanics,
concerned with one-dimensional problems that arise when investigating
longitudinal motions of an elastic bar. Two chapters then present careful
derivations of the principal field equations that govern the behavior of
general continuous media (e.g., conservation of mass, balance of linear
momentum). Additional constitutive equations are added to the field equa-
tions, so" inviscid" fluids are given full mathematical description. Chapter 15
treats various examples that involve inviscid flow (stability of a stratified
fluid, sound waves, flow past an obstacle). The final chapter, a study of
potential theory, concludes with the use of Green's functions in an analysis
of the diffraction of sound waves by a hole in a screen. Partial differential
equations form the mathematical core of this part.
A PREVIEW OF THE FOLLOWING VOLUME
Many important applied mathematical ideas are illustrated only briefly
in the present volume. They are deserving of much more study. Some of this
is provided in a following volume (designated here as II}.
The following volume begins with some prerequisites to an efficient study
of three-dimensional continuum mechanics, particularly cartesian tensors.
Then follows a considerable amount of material on viscous fluid flow and on
elasticity. There is also a detailed discussion of dispersive waves, particularly
surface waves on water. This provides the most penetrating analysis of the
present work and brings the reader through much brilliant theory by classical
authors to the frontiers of present-day research. The last part of II is con-
cerned with variational methods. The more advanced theory is given a setting
that involves basic ideas of functional analysis.
IO What Is Applied Mathematics? [Ch. 1
CONCEPTS THAT UNIFY APPLIED MATHEMATICS

Our presentation is not tightly organized like a handbook but is built
(somewhat like a symphony) around certain major themes. We shall not list
all the themes and their relations; it should be part of the reader's training
and enjoyment to discover them for himself. To give just one illustration of
what we mean, however, we note that pure mathematicians classify partial
differential equations into
elliptic ( such as ::~ + ::~ = 0),

parabolic
o2 u -
. (such as ax2 au
oy = 0) '
hyperbolic ( such as ::~ - ::~ = 0).

Applied mathematicians often prefer the complementary viewpoint of regard-
ing natural phenomena as illustrating equilibrium, diffusion, or wave
propagation. We shall see how these themes-and others, such as stability,
randomness, and optimization-help to unify applied mathematics by
revealing common features of superficially disparate problems.
1.2 Introduction to the Analysis of Galactic Structure

In this section we indicate how theoretical analysis can begin to explain
aspects of galactic structure. As with all mathematical theories of natural
phenomena, we require (i) the basic physical laws that are relevant, and (ii) the
particular characteristics of the system under investigation.
Requirement (i) can be briefly disposed of, for the required laws are those
of classical physics. To satisfy requirement (ii), we shall carry the reader
on a brief quantitative survey of the universe. We then present in some detail
the formulation and solution of a problem that provides prediction of the
distribution of stars across a galactic disk. We conclude with a very brief
sketch of some current ideas about the cause of the spiral structure that is
a feature of many galaxies.
PHYSICAL LAWS GOVERNING GALACTIC BEHAVIOR
In studying galaxies, the basic physical laws that one uses are those of
classical physics: (i) Newton's laws of motion and his law of universal
gravitation, which govern all classical mechanics; (ii) the laws of electro-
dynamics according to the formulation of Maxwell; and (iii) the laws of
thermodynamics.
The reader should notice that there are very few fundamental empirically
based laws. However, we frequently have to introduce new mathematical
concepts that are essential for the study of certain physical problems. In the
Sec. 1.2] Introduction to the Analysis of Galactic Structure II
particular example that concludes this section, for instance, only physical laws
known to Newton are used. But we require mathematical concepts of much
more recent vintage, such as phase space and density in phase space.
We turn now to a survey of certain facts about our universe. It should
be remembered that the qualitative and quantitative information that we
cite is the fruit of an enormous amount of scientific effort, over centuries of
time.
BUILDING BLOCKS OF THE UNIVERSE
Our expanding universe is about 10 billion (1010 ) years of age. At the
present time, it is composed of about IO billion galaxies as the building
blocks. These galaxies, rather well-defined collections of gas and stars, are
approximately uniformly distributed in the universe (which appears to have
no border in space). There is a degree of nonuniformity: galaxies form clusters
through mutual gravitational attraction. Our own galaxy-the Milky Way
system-is a member of a cluster of 17 galaxies.
To convey a general impression of the nature of the galaxies, let us cite a
few characteristics of the Milky Way system. Its shape is that of a thin disk
with a roughly spherical nucleus at its center. (The sun is situated toward the
outer edge of the system, about 10,000 parsecs* from the galactic center.)
At a distance of about 15,000 parsecs from the center, the mass density of
the galaxy becomes quite small. The galaxy is in rotation; the linear velocity
is about 250 km/sec in the solar vicinity.
The nearest star is about 4 light years (1.3 pc) away. By contrast, the orbit
of Pluto is about 40 A.U. (2 x 10- 4 pc). Thus the solar system may be
likened to an atom of a monatomic gas. This gas is relatively rarefied, at
least in our vicinity.
CLASSIFICATION OF GALAXIES
By their appearance, galaxies are classified into four categories: (i) ellip-
ticals (including sphericals), (ii) normal spirals (disk-shaped), (iii) barred
spirals, and (iv) irregulars. Examples of these galaxies are shown in Figure 1.1.
Another example is shown in the frontispiece.
The galaxies can be classified according to an ordered scheme, such as
that shown in Figure 1.2. Galactic classification was originally based on geo-
metrical appearance alone, and detailed study shows that the geometrical
characteristics are related to other physical parameters, for example, the gas
content. This indicates that there is a dynamical basis for the differences in
galactic appearance, a difference that persists over considerable periods of
time (a few billion years).
* A parsec (pc)= 3 x 1018 cm is the distance at which the radius of the orbit of the earth
around the sun (called the astronomical unit,A.U.) subtends an angle of 1 second of arc. It is thus
about 2 x 105 A.U. A light-year is about 0.3 pc. An object traveling at the speed of 1 km/sec
covers a distance of 1 parsec in about I million years.
Sec. J.21 Introduction to the Analysis of Galactic Structure
13
REGULAR SPIRALS
EO
e
so
.'1t> ~ - <~''. .:%'
"~~)
s,
.; . .. ''
,1.
. ..?};.:
~~
: ~
,,

So Sb Sd
E3
E6 ~

580
.l)
SBo
<t,-,,J
i
SBb
.1r-
t<"
I ,
SB,
,,,,.,
.
_;
:, -..~. '..
SBd
:: ~
18,
ELLIPTICALS
BARRED SPIRALS
FIG URE 1.2. Composite of Hubble and de Vaucou/eurs galaxy classi-

fication systems. (Photograph courtesy of the Hale Observatories.)
The disk-shaped galaxies must be in sufficiently rapid rotation to maintain

their shape . Spherical galaxies presumably have no net rotation. Elliptical
galaxies with different degrees of flatness have different degrees of rotation.
Bar-shaped galaxies are in uniform rotation but not about an axis of sym-
metry. Classical examples of such rotating systems held together by gravitation
were discovered by the mathematician Jacobi (1804-1851).
COMPOSITION OF GALAXIES
Each galaxy is composed of about 10 billion stars (109 -10 11 stars, to be
more exact) . Our galaxy is among the largest. The sun is an average star,
about 5 billion years of age . Massive stats, e.g., those of 40 solar masses,
are brilliant, and they bum out their nuclear fuel very quickly. They must
therefore have been formed comparatively recently (say, within the past 10
million years). The evolution of the sun is quite slow; its nuclear energy will
not be exhausted for another IO billion years.
There is also a substantial amount of gas in the galaxies-up to more than
20 per cent in the irregular galaxies, somewhat less in the Sc spirals, down to
1 or 2 per cent in the Sa spirals, and practically none in the giant ellipticals.
Our galaxy has a gaseous content of about 3-4 per cent.
FIG u RE 1.1. Examples of galaxies of various types: (upper left) elliptical

galaxy EO pee. NGC 4486 ; (upper right) normal spiral SC NGC 5364;
(center) barred galaxy SBb(s), NGC 1300; (lower) irregular galaxy
Irr I, the Large Magellanic Cloud. (Photographs courtesy of the Hale
Observatories.)
On such a large scale, the gas behaves practically as a perfect electric

conductor, even though it is only slightly ionized and the conductivity is
small by terrestrial standards. The cosmic magnetic field produced by the
electric current in interstellar gas is a few microgauss (10- 6 gauss, the gauss
being the cgs unit of magnetic field).
Permeating the interstellar space are cosmic ray particles and electro-
magnetic radiation of various wavelengths (ranging from alpha rays, to X
rays, to optical radiation, to radio waves). These may exert an influence on the
gas and the magnetic field. We therefore use the term" interstellar medium"
to describe the whole system, which interacts only weakly with the systems
of stars. Indeed, the principal dynamical interaction between the stars and
the gas is gravitational.
DYNAMICS OF STELLAR SYSTEMS
To analyze the structure of the galaxy (e.g., to explain the spiral structure),
one must first place it in a proper mathematical context. How does one begin
to tackle such a complicated system'!
As explained earlier, the first essential step is to construct an idealized model.
This step can be taken only after we have ascertained the relevant empirical
facts and the general laws governing our system.
An important set of facts about the galaxy is given in Table I.I. We see
TABLE 1.1. Energy densities in our galaxy".
Energy density
Source (Units: 10- 12 erg cm- 3 )
Total radiation (star light) 0.7

Turbulent gas motion 0.5
Total energy of galactic rotation 1300
Cosmic rays 1
Magneticfield oo-s gauss) 4
From P. Morrison, Rev. Mod. Phys. 29, 235 (1957).
that the following constituents make successivelysmaller contributions to the

energy density: (i) the stellar system (by far the most prominent), (ii) the gas,
(iii) the magnetic field, and (iv) the cosmic ray particles and other forms of
radiation. Thus we should begin by formulating the dynamics of the most
energetic component, the stellar system. This formulation will now be
described.*
A collection of stars, separated by distances of the order of 106 -10 7 times
their radii (distances that are many times the radii of their outermost planets)
A full system of equations governing the dynamics of a galaxy can be found in a survey by
C. C. Lin, J. Appl. Math. 14, 876-921 (1966).
Sec. 1.2] Introduction to the Analysis of Galactic Structure
15
may be treated as a set of masses subject to gravitational attraction. It is

possible to write down the system of dynamical equations governing their
behavior according to Newton's laws of motion and universal gravitation.
Such systems will be discussed in detail in Chapter 2. However, this is clearly
not a suitable way to describe the galactic system, for we cannot hope to
follow the motion of each of the billions of stars. We must consider them in
some sort of collective manner, in the so-called statistical description.
For simplicity, we shall consider the idealized situation of stars of uniform
mass. (A discussion of the validity of this idealization must be left to more
specialized treatments of the topic.) Since the motion of a star is specified by
its position and velocity, we consider the number of stars within a given
range of position and velocity. For precise mathematical formulation, we
introduce the phase space of six dimensions (x, y, z; u, v, w), combining
positional coordinates and velocity coordinates. The phase space number
densityor distributionfunction'l'{x, y, z; u, v, w; t) of the stars in this phase
space at time t is defined so that 'l'(x, y, z; u, v, w; t) dx dy dz du dv dw gives
the number of stars whose three spatial coordinates are, respectively, in the
ranges (x, x + dx), (y, y + dy), (z, z + dz), and whose three velocity compon-
ents are in the ranges (u, u + du), (v, v + dv), and (w, w + dw). The distri-
bution function 'I' then gives a complete description of the stellar system at
any instant.* To describe the dynamical process, we must consider the change
of this distribution function in the course of time.
In terms of the distribution function 'I', the mass density p(x, y, z, t) of the
stars is given by
"'
p(x, y, z, t) = m,., fJf
-co
'l'(x, y, z; u, v, w; t) du dv dw, (1)
where m* is the mass of an individual star. This equation merely states that
to get the total mass density, we must sum up the mass of all the stars with
various velocities.
Each of the stars is moving in the combined gravitational field of all the
others. Except for stars coming into close encountert with a given star, the
motion of an individual star will be essentially influenced only by the gravita-
tional field due to the smeared-out distribution of matter with density
p(x, y, z, t) given by (1). This gravitational field is given by the negative
gradient of the gravitational potential V(x, y, z, t), which is related to the
density p by
rr(
r,~~~t
) __ G
-
JJJ
p(x 0 , Yo, z 0 , t) dx 0 dy 0 dz 0
, (2)
2
J(x - xo)2 + (y- Yo)2 + (z - z0 )
From material in Sections 13.1 and 14.5 one can understand that we work with the distribu-
tion function ' because interstellar distances are small on a galactic scale, but typical distances
traveled between collisions are not.
t Close encounters require specialized treatment, but their net effect is negligible in the present
problem. See S. Chandrasekhar, Principles of Stellar Dynamics (New York: Dover, 1960).
where G is the constant of universal gravitation (G = 6.7 x 10-s cgs units).

The equations of motion of each star can now be written as
dx dy dz du BV dv BV dw BV
-dt = u' -
dt
= v' -
dt
= w' -
dt
= - -ox , -dt = - -
By
, - =
dt
- -oz . (3)
Furthermore, by invoking mass conservation and the assumed absence of

collisions, one can assert that any increase of mass in a portion of phase
space must be due to flow across its boundaries. As in Exercise 14.1.12 or
Exercise 14.1.13, this can be shown to imply that
o'f'+ u o'f'+ v o'f'+ w o'f'_ (av aq, + av o'f'+ av o'f')= o. (4)

ot ox oy oz ox OU By av oz ow
Note that the gravitational field - VVis related to the distribution function 'I'
through (l) and (2). Hence the terms within the parentheses in (4) are quad-
ratic in 'f'; i.e., if 'I' is replaced by ct'I' (ct being a constant), these terms are
multiplied by a factor ct2 Thus (4) is a nonlinear equation.
Equations (3) which govern the motion of the individual stars, and Equa-
tion (4), which governs the collective behavior of all the stars, are closely
related to each other. In the language of the theory of partial differential
equations, the system of equations (3) describes the characteristic curves of(4).
From (4) one can calculate o'l'/ot at any instant provided 'f' is known,
since the gravitational potential can be calculated through the use of (l) and
(2). Heuristically speaking, we may then regard (4) as sufficient for the
description of the stellar system at all subsequent times, once 'f' is specified
at a given instant. The mathematical theory for this statement is much
harder to prove, especially when one raises such questions as stability-does
the change in the solution remain small if there is a slight change in the initial
specification of the distribution function? Such questions often cannot be
rigorously settled; yet the applied mathematician must proceed with his
work. He must, of course, have reasonable assurance that his work is
meaningful by appealing to other careful considerations beyond the desirable
but unavailable mathematical theorems.
DISTRIBUTION OF STARS ACROSS A GALACTIC DISK

As a relatively simple but realistic example of the use of the above system
of equations, let us consider the distribution of stars across a galactic disk.
This is a problem that has yet not been fully solved. There appears to be
" missing matter" in our galaxy in the solar neighborhood.* The solution
to be presented below is usually used as the essential basis to test the obser-
vations.
See, e.g., p. 15 of L. Woltjer's article in Relativity and Astrophysics, Vol. II (Providence,

R.I.: American Mathematical Society, 1967).
17
The galactic disk is 10 kpc ( = 104 pc) in radius but only 600 pc in thickness.
Thus, to a good approximation, we should be able to consider the galaxy to
vary locally only in the direction perpendicular to the disk. We shall take this
to be the direction of the z axis with z = 0 at the center of the disk. Then, in
the stationary state, q, = 'i'(z, w) and (4) becomes
aq, av o'i'
w----=0. (Sa)
oz ozow
It can be shown (Section 16.1) that the potential V(z) satisfies the Poisson
differential equation
a2 v
az2 = 4n:Gp(z). (Sb)
Our considerations now focus on a definite mathematical problem, that

posed by (5a) and (5b). (Certain additional requirements on the solution
will also be invoked later.) We shall treat these equations in two steps.
Equation (5a) can be solved by the standard method of Lagrange for
integrating partial differential equations of the first order. Let us briefly recall
the salient features of this method.
For a partial differential equation of the form
o' aq,
P(z, w, ') - + Q(z, w, ')-
0z aw= R(z, w, '), (6a)
Lagrange's methodof characteristicsis based on the characteristicequations

dz dw d'I'
-=-=- (6b)
p Q R
Let/ 1(z, w, ') = c 1 andfiz, w, ') = c2 be solutions of any two independent
ordinary differential equations implied by (6b). (Here c 1 and c 2 are constants.)
Then / 1 and / 2 can be shown to define functions ' that satisfy (6a). Thus
ft and/ 2 provide constants or integralsof the phenomenon described by (6a).
The intersections of the surfaces Ji = c 1 ,! 2 = c 2 give the characteristiccurves.*
The general solution to (6a) is given by
F(/ 1,/ 2) = 0.
In the present case, R = 0, so that one integral relation is ' = c1 [which
obviously provides a solution of (5a)]. A second characteristic equation is
dz dw
-= ' (7)
w -oV/oz
A slight extension of these remarks shows that (3) gives the characteristics of (4), as stated
above.
18 What Is Applied Mathematics? [Ch. I
which leads at once to the integral

{-w2 + V(z) = E, (8)
where E is a constant, that gives the total energy per unit mass of the star.
Thus the general solution of (Sa) is
'I' =f(E) = /H-w
2
+ V(z)]. (9)
A special case of (9) is compatible with the observed Gaussian distribution
of velocities. We take
(10)
where A 0 and f3are suitable constants. The constant f3is prescribed by the
distribution of velocities. Furthermore, by the one-dimensional version of (I),
A0 must be related to the local density by
p = m* f-oo
00 'I' dw. (11)
Indeed, we find the density distribution to be of the form

p(z) = Po e-PV<z>, (12)
where p 0 is the density where V(z) = 0, which we shall take to be at z = 0,
the middle of the galactic plane.* The quantity Po is related to A 0 and P by
Po=Aom*~ (13)
The observational determination of p 0 is based on (12), which follows

from (Sa), without the use of the Poisson equation (Sb). From (12) one can
determine V(z), and hence its second derivative, as
d2V = - .!_~In (.!!..)

. (14)
dz 2 Pdz 2 Po
One can determine f3from the distribution of stellar velocities, by using
(IO), and one can observe the number of stars per unit volume, which is
proportional to p. Evaluation of the expression in (14) is then possible.
Furthermore, this can be done for any species of stars.t Thus we have the
possibility of determining the total stellar mass by combining (14) with (Sb).
As mentioned before, existing knowledge reveals a considerable descrepancy.
The mass determined by this dynamical procedure is about I. 7 times larger
than that obtained by direct observational determination.
A potential is defined only up to an additive constant, since just its gradient is significant.
t From observations of their spectral Jines, stars are classified into a number of different
species. These species basically correspond to stars of different properties such as mass, composi-
tion, and age.
19
Equation (5b) has yet to be satisfied. Thus, to complete the analysis, one
must solve the equation
d2 V
--
dzz
= 4nGp o e-PV(z) . (15)
To do this, multiply both sides by dV/dz and integrate, with the specification
that
dV
V=-=0 at z = 0, (16)
dz
since the gravitational force also vanishes at the middle of the galactic plane.
Then (Exercise 1) there is little difficulty in showing that the distribution of
density finally takes on the form
p 2 z
-=sech -, (17)
Po Zo
where z 0 is given by
(18)
What have we gained from our theoretical analysis? Of most importance
is the discrepancy between theory and observation concerning the value of p 0 .
A naive person might think that this discrepancy proves our analysis to be of
little worth, but just the opposite is true. Rough agreement between experi-
ment and a plausible theory could be largely a coincidence, but disagreement
means that there is more to the phenomenon than meets the eye. The italicized
phrase is doubly relevant here, for there must be some invisible mass present
to account for the excess of the predicted mass compared to that observed.
The two principal conjectures are that the excess mass is due either to mole-
cular hydrogen or to dark stars. The matter has not been resolved at the
time of writing.
DENSITY WAVE THEORY OF GALACTIC SPIRALS
It is beyond the scope of this book to carry out much further analysis of
stellar systems beyond that given above. We shall, however, report here
briefly some recent developments in the theory of galactic spirals, based on
stellar dynamics and gas dynamics, as another example of the power of
mathematical analysis.
The bright spiral arms of galaxies are marked by massive young stars.
These stars are known to be rotating faster in the interior parts of a galaxy
than in the outer parts. If a spiral arm is" material," i.e., if it always consists
of the same stars, we must then have what is known as the" winding dilemma."
For trailing spiral patterns,* the spiral arms would tend to wind into a
A spiral pattern is called "trailing" if it opens up in the direction behind that of the rotation
of the stars; "leading," if in the opposite sense. A leading material spiral pattern would loosen
up under the effect of differential rotation.
tighter structure. However, the Hubble classification is based not only on the
tightness of winding, but on other physical parameters as well. Furthermore,
these physical parameters cannot change as rapidly as the geometry of a
material spiral arm. Thus we must accept the hypothesis that the spiral
pattern remains quasi-permanent.
In the 1920sthe Swedish astronomer Bertil Lindblad suggested that there
are "density waves" in the stellar system which underlie the observed spiral
structure. These are periodic temporary concentrations of stars caused by
their self-gravitation. Qualitatively, they are similar to acoustic waves, but
the physical mechanisms involved are different in the two cases. Lindblad
did not carry out any detailed mathematical analysis of the collectivebehavior
of these stars. He was therefore able to make only general conjectures about
these waves, and some of them were inevitably incorrect. Thus his ideas
remained unaccepted for a long time.
In the 1960s the density wave theory was revived by C. C. Lin and Frank
H. Shu when they succeeded in calculating a spiral structure for the Milky
Way system, in reasonable agreement with observations. The method was
also applied to other galaxies, and a number of other observable consequences
were deduced and checked against astronomical data. (See Figure 1.3.)
For a summary report of the status of the theory, see the general discourses
of B. J. Bok and C. C. Lin at the 14th General Assembly of the International
Astronomical Union.* Suffice it to note here that the theoretical studies of
these density waves offer very challenging mathematical problems. Indeed,
the problems are similar to those encountered in the study of waves in
electromagnetic plasmas. (By analogy, a stellar system may be called a
"gravitational plasma.") Thus the understanding of galactic systems requires
extensive mathematical analysis of a rather subtle kind. At the same time,
it stimulates the study of certain mathematical methods and theories that are
applicable to other important problems.
EXERCISES
1. (a) Deduce the relation (17) for the density distribution of stars across
a galactic disk. Sketch.
(b) Obtain the distribution of gravitational potential
z
V(z) = 4nGp0 z~ In cosh - . (19)
Zo
(c) Find the gravitational field at "large" distances from the galactic
plane. Discuss what is meant by "large."
* Published in Highlights of Astronomy, Vol. 11 (Dordrecht, Holland: D. Reidel Publishing

Company, 1971). For general information on spiral structure, including a description of the
density wave theory, see the article "Updating Galactic Spiral Structure" by B. J, Bok in Amer.
Scientist 60, 708-22 (1972).
Sec. 1.2] Introduction to the Analysis of Galactic Structure 21
FIGURE 1.3. The same photograph as in the frontispiece with added

curves indicating the location of the peak of synchrotron emission as
observed by D. S. Mathewson, P. C. van der Kruit, and W. N. Brouw.
Coincidence of this curve with the dust lane is strong evidence in favor of
the density wave theory of galactic spirals. Reprinted with permission
from Astron. and Astrophys. 17, 468 (1972).
1.3 Aggregation of Slime Mold Amebae

In this section we shall formulate a mathematical model to describe the
aggregation of slime mold amebae. We shall thereby gain some understanding
of a phenomenon of considerable current interest in developmental biology.
(Developmental biology includes embryology, the study of how a single cell
develops into a complicated embryo, but also encompasses the development
of simpler organisms that cannot be said to have an embryo.) We begin with
some facts about slime mold amebae that provide a basis for our mathematical
model.
SOME FACTS ABOUT SLIME MOLD AMEBAE*
One can begin a description of the life cycle of the slime mold amebae at
the spore stage, where each ameba is dormant within a protective covering.
When conditions are favorable, an ameba emerges from its spore. Of the
order of IO micrometers (10- 3 cm) across, the amebae are rather shapeless
one-celled organisms that move by extending contractile portions of them-
selves (pseudopods).
The natural habitat of the amebae is soil or dung. An important element
of the food chain on earth, they feed on bacteria by engulfing them. If food
is plentiful, the amebae continually feed and multiply by mitosis (dividing
in two). If the food supply becomes exhausted there is an interphase period of
random and somewhat feeble movement, where the amebae are more or less
evenly distributed over the area available to them. During this period, the
disappearance of the food supply triggers certain changes in the amebae.
The details of these internal changes are not known, but there is no difficulty
in observing the striking external phenomenon that results. After a few hours,
the amebae begin to aggregate into a number of collection points. Typically,t
these are more or less regularly distributed, with a spacing of a few hundred
micrometers. (See the frontispiece. Figure 1.4 gives an idea of what the be-
ginnings of aggregation look like under conditions of lower cell density.)
After aggregation has been completed, the amebae that have collected at a
given point (ranging in numbers from a few in laboratory experiments, up to
200,000) form a multicelled slug. This moves as a unit, although the formerly
free living amebae retain their cell walls within the slug. Then the slug stops
and erects a stalk, on top of which is a roundish container of spores. The
cycle is thus completed.
* "There is no philosophy which is not founded upon knowledge of the phenomena, but to get
any profit from this knowledge it is absolutely necessary to be a mathematician."-Daniel
Bernoulli. Quoted by C. Truesdell on p. 318 of Essays in the History of Mechanics (New York:
Springer, 1968).
t There are several species of cellular slime mold, with various corresponding differences in
behavior. [SeeJ. T. Bonner, The Cellular Slime Molds (Princeton, N. J.: Princeton U.P., 1967).]
Our description is appropriate for the most studied species, Dictyostelium discoideum.
Sec. 1.3J Aggregation of Slime Mold Amebae 23
t:~~!i?.
. ... ..
. =: .-.
. ..: .
FIGURE 1.4. Population of amebae just beginning to aggregate. The

scale is indicated by a 100-micrometer line segment. (Courtesy of
B. Shaffer.)
What is responsible for the organized aggregation of the amebae? This is

an important question, for "purposeful" movements occur frequently in
developmental processes. Usually, these movements take place rather
inaccessibly within a developing organism, but the cells of the slime mold
amebae will obligingly perform in a laboratory dish so that it is easy to
examine them and to experiment with them.
It has been discovered that the amebae move preferentially toward rela-
tively high concentrations of a chemical which they themselves secrete. In
some speciesthe attractant has recently been identified as cyclic 3' ,5'-adenosine
monophosphate (AMP) a chemical with many important biological functions.
(E. W. Sutherland was awarded the Nobel Prize for medicine in 1971 for his
work in elucidating some of the roles of cyclic AMP.) It is also known that a
given quantity of attractant loses its potency in a matter of minutes. This has
been traced to the activity of an enzyme that alters the nature of the AMP.
Presumably, aggregation is caused by the fact that the amebae move up a
gradient of attractant, but what determines the time of onset? What deter-
mines the spacing of aggregation centers? Can one quantify the process?
A mathematical model is needed to answer such questions. We shall proceed
to devise the simplest model that could reasonably be supposed to bear on
What Is Applied Mathematics? [Ch. 1
24
the circumstances.* If analysis of this model is encouraging, one can add

detail later. For this, the reader is referred to the paper just cited and to
subsequent papers by various authors in the same journal.
FORMULATION OF A MATHEMATICAL MODEL

Since the distance between amebae is small compared to the typical
distance between aggregation centers, we shall employ a continuum model.
Suppose that the aggregation takes place in the (x, y) plane. For simplicity
we shall assume that no quantities change with y, so that only x variation
need be considered. Exercise 3 shows that this assumption of unidimension-
ality is not at all essential, but it makes exposition easier.
Let a(x, t) denote the number of amebae per unit area at position x and
time t. Consider the amebae located in the region x 0 ::;; x ::;;x 0 + Ax, where
Ax is an arbitrary number (not necessarily small). We shall now write a
general balance law for this region; this states that the rate of change of the
net amount of a in the region equals the rate at which a flows across the
boundary, plus the net rate of creation of a within the region. In the present
case, a stands for amebae, and the net creation of amebae is equal to the
excess of births over deaths. But the balance law is "general" because it
applies to any substance whatever. For how else can a substance appear in
the region except by creation or by flow across its boundaries?
To proceed, we must define the flux density J(x 1 , t). This gives the net
rate per unit length at which a crosses the line x = x 1 Also, J is defined to be
positive if there are more amebae crossing in the direction of x increasing
than in the opposite direction. The term Q(x, t) is the net rate per unit area
at which a is being created. The desired balance law can now be written.
Considering the rectangle of Figure 1.5, we obtain
a l<o+,b
-a
r.
f a(x, t) dx = J(x 0 , t) - J(x 0 + Ax, t) + f
xo+Ax ,
Q(x, t) dx. (1)
It is convenient to use the integral mean value theorem (Appendix 13.2) to

write
a
at [a(x 1, t) ax]= J(x 0 , t) - J(x 0 + Ax, t) + Q(x 2 , t) ax;
Xo ::;; xI ::;; Xo + ax, Xo ::;; X2 ::;; Xo + Ax. (2)

We divide by axand then take the limit as ax-+ 0, to obtain the general
balance law in differential equation form,
aa aJ
ar(xo' t) = - ax (xo' t) + Q(xo, t), x 0 arbitrary. (3)
We adapt material in a paper by E. F. Keller and L.A. Segel, J. Theoret. Biol. 26, 399-415
(1970).
Sec. 1.3] Aggregation of Slime Mold Amebae 25
Fro URE 1.5. The balance law (3) considers the rate of change of mass
in a rectangle of length Ax and unit width, in the case where y variation
is absent.
In the case of the amebae, reproduction can be ignored because there is

little or none of it in the absence of food. Deaths can also be ignored, since
there are few during the time interval of interest. Thus we take
(4)
To obtain an expression for the flux J, let us first consider a situation when
attractant is absent. Then the amebae appear to move about randomly.
Owing to such "diffusionlike" movement, a concentration of amebae tends
to disperse. Thus there is a random flux J, from regions of high ameba
concentration to regions of low concentration. The magnitude of the flux at x
seems to depend on the concentration difference between x and nearby points.
We characterize this difference by aa/ax
(the simplest choice) and make the
hypothesis that
J,(x, t) = F [aa~:
t)] (5)
for some function F. Now when a = constant, J = 0-for in random motion

there will be as many amebae moving to the left as to the right. In this case
aa/ox = 0. Therefore, given (5), it is only sensible to assume that
F(O) =0. (6)
Thus the graph of F must have an appearance such as that depicted in

Figure 1.6. For sufficiently small values of s, we can approximate the graph
by a straight line. Calling the slope of the line -, we obtain F(s) = -s as
the simplest reasonable assumption concerning F; i.e.,
J.(x, t) = - oa~:t). (7)

What ls Applied Mathematics? [Ch. I
FIG URE 1.6. Possible plot of the function F defined in (5). The graph
provides negative values for s > 0 because there should be a leftward
(negative) flux if amebae density is higher at larger values of x. Similar
reasoning explains the positivity for s < 0.
Combining (3), (4), and (7), we obtain
aa a ( aa) (8)
at = OX ax .
Equation (8) is the diffusion equation, which will be studied at length
beginning in Chapter 3. This equation is used to model the concentration
variation of any kind of randomly moving set of particles, for example smoke
particles in air (see Chapter 3). The constant , which governs the vigor of
the random movement, is generally called the diffusivity.Here we call the
motility, giving precise meaning to a biological term that is often used only in
a qualitative manner.
Equation (8) was obtained under the assumption that no attractant was
present. To account for cbemotaxis, which is directional motion induced by
variations in chemical concentration, we add to Jr of (7) an additional
contribution Jc . Let p(x, t) be the density of the attractant. Arguing as
before, we pass from an initial assumption that Jc is a function of the attract-
ant gradient op/ox to the assumption that Jc is proportional to the attractant
gradient, at least for small values of the gradient. For a given gradient, if
the amebae density is twice as great, the net flux should be twice as great.
The proportionality factor should thus be a multiple of a. We are led to the
hypothesis that
(9)
Sec. 1.3] Aggregation of Slime Mold Amebae 27
The factor x measures the strength of chemotaxis. Note that in contrast to (7),
there is no negative sign in (9). This is because amebae tend to move toward
attractant concentrations (and away from ameba concentrations).
Assuming that the total flux J in (3) is the sum of the random contribution
J, and the chemotactic contribution Jc, we are led to our final equation for
the change in ameba density:
oa = ~
at ox
(aaax - xa op)
ax
. (10)
We shall take and x to be positive constants. It is not difficult to take into

account a variation with p which is probably present, but this would only
change some details of the analysis to come [Exercise 2(a)]. Note that even
with constant and X, (IO) contains a (quadratic) nonlinear term xa(op/ox),
because this term is proportional to the product of two unknown functions,
a and op/ox. [Such a quadratic nonlinearity was also found in our study of
galactic structure-compare the remarks following Equation (2.4.)] Nonlinear
terms generally make mathematical analysis more difficult, and hence more
challenging and more rewarding.
We also need an equation for the attractant density p. This will be of the
general form (3):
ap = _a,.+ Q
ai ox
(The subscript" a" refers to attractant.) The random motion of the attractant
molecules will be modeled by a proportionality of flux to gradient, as in (7):
op
J3 = -D-.
OX
The net creation term Q. has a positive contribution fa as a result of the
secretion of attractant by the amebae. Here f is the rate of secretion per unit
amebae density. What of the decay in attractant activity? We take the rate
of decay (as in radioactivity or some other spontaneous process) to be
proportional to the amount of attractant present,* via the constant k.
Thus Q. = fa - kp, and the desired equation for op/ot is
op 82 p
- = fa - kp +D- . (11)
ot ox 2
We shall take f, k, and D to be positive constants.
As we stated above, decay is actually due to the action of an enzyme. If Michaelis-Menten

kinetics (Chapter IO) are assumed, this can be modeled by making k a certain function of p.
But the essentials of the analysis are unchanged [Exercise 2(b)].
Partial differential equations (10) and (11), for the two unknown functions
a(x, t) and p(x, t), provide the basic formulation of our problem.
AN EXACT SOLUTION: THE UNIFORM STATE
It is very easy to find an exact solution of (10) and (11). This is the uniform
solution,
a=a 0 , P =Po, (12)
where a0 and Po are constants. [When (12) holds, the system is said to be in
the uniform state.] Upon substitution into (10) and (11), we find that (12) will
indeed provide a solution if
(13)
Equation (13) is physically reasonable. It says that in the uniform state the
secretion rate of attractant must be exactly balanced by the decay rate.
ANALYSIS OF AGGREGATION ONSET AS AN INSTABILITY
We identify the uniform state with the interphase period prior to aggre-
gation. We also model the onset of aggregation as the breakdown of the
uniform state due to the growth of inevitable small disturbances to ameba and
attractant density. That is, we identify the onset of aggregation with the sort
of instability of the uniform state whose investigation forms a classical part of
applied mathematics (see Sections 11.l and 15.2). The instability is presumed
to occur because of changes, during interphase, of the parameters , x.,and f
that characterize ameba behavior.
The idea behind instability theory is this. Suppose that at some initial time,
the state of the system is slightly disturbed from the uniform state. (Suppose
that there is a local clumping of amebae, for example, and an accompanying
local concentration of attractant.) Will the small disturbance tend to disappear
with the passage of time or will it become more intense? In the former case
we say that the uniform state is stable to small disturbances, in the latter
unstable. Unstable states will not be observed, for disturbances are inevitable.
In the case of instability, they will grow, so the uniform state will be replaced
by some other state.
To perform a stability analysis, we introduce the variables a' and p' by the
definitions
a(x, t) = a0 + a'(x, t), p(x, t) = Po + p'(x, t). (14)
Here a' = a - a0 , for example, measures departure from uniformity; there-
fore, it can be identified with the disturbance in ameba density.
To obtain equations for a' and p', we substitute (14) into (IO) and (11).
From the former we obtain
oa' o2a' [ o2p' oa'op']
8f = ox2 - )(. (ao + a') ox2 + oxox (15)
Sec. 1.3) Aggregation of Slime Mold Amebae
29
This equation is nonlinear, owing to the presence of the quadratic terms
a'(o 2 p'/ox 2 ) and (oa'/ox)(op'/ox). We shall assume that the disturbances
(and their derivatives) are small, in which case we call a' and p' perturbations.
The perturbations are involved quadratically in some terms and linearly in
others. Products of two small terms should be negligible in comparison with
the other terms of (15), which contain but a single perturbation function.
We thus linearizethe equation by deleting all nonlinear terms. We obtain
oa' 82 a' 82 p'
at= ax2 - xao ax2 (16)
as the linearized version of (10). As for (11), upon substituting (14) and em-
ploying (13), we obtain
op'
- = fa' - kp' + D -,
a2 p' (17)
ar ox 2
which is already linear.
We are faced with a pair of linear partial differential equations with
constant coefficients. We guess that there are solutions of the form
a' = cl sin qx e'"' p' = C2 sin qx e"', (18)
where C 1 and C2 are constants.* It is easily seen [Exercise I(a)] that there are
indeed solutions of the form (18), provided that
(O'+ q 2 )C 1 - xa0 q 2 C2 = 0, (19a)
-fC 1 + (q + k + Dq 2
)C 2 = 0. (19b)
This system of algebraic equations has the trivial solution C1 = C2 = 0.
[From (18) we see that we have merely verified that it is possible to have an
identically zero perturbation of an exact solution.] For a nontrivial solution,
the determinant of the coefficients must vanish. We thus obtain the following
quadratic equation for O":
0'
2
+ bO"+ c = O; b =k + ( + D)q 2 , c = q2 (k + Dq 2 ) - xa0 fq 2
(20)
The quadratic equation can be shown to have real roots [Exercise l(b)].
To ensure stability, both roots must be negative, so that the exponential
factor exp (qt) brings about decay of the perturbations. It is not difficult to
show (Exercise I) that a necessary and sufficient condition for stability is
c > 0, which requires that
xa0f < (k + Dq 2), q :t=
O. (21)
The reasons that lie behind such an assumption as (18) are more fully discussed in Section
15.2. We mention here that a cosine dependence in (18) will yield exactly the same results. (Com-
pare Exercise 3.) Also, more general disturbances can be obtained by the superposition of sinu-
soidal solutions, using Fourier analysis (as discussed in Chapters 4 and S).
30 What ls Applied Mathematics? [Ch. I
From (18), 2rr./qis the wavelength of the perturbation under investigation.

Since and k are positive, the right side of (21) decreases monotonically as q
decreases, toward a greatest lower bound of k. Therefore, the longer the
wavelength 2rr./q,the more "dangerous" the perturbation [for (21) is more
easily violated]. We conclude that instability is possible whenever
xaof > 1 (22)
k '
for then (21) is violated for a range of sufficiently small values of q 2 .
INTERPRETATION OF THE ANALYSIS
The view of aggregation that emerges is this. During the beginning of the
interphase period, inequality (22) does not hold and the uniform state is stable.
Triggered by the stimulus of starvation, the various parameters gradually
change. Eventually, (22) is satisfied, and aggregation commences. We identify
the size of the " aggregation territory" with the instability wavelength; our
analysis predicts that the territory size is "very large" in some sense.
The instability criterion (22) has the following interpretation. Suppose
that there is a concentration of amebae and attractant at some point. Random
" diffusion" of amebae with motility tends to disperse this concentration,
as does the attractant decay, whose strength is measured by k. Larger and k
means larger stabilizing effects. It is therefore appropriate that increases in
and k mean that the instability criterion (22) is more difficult to satisfy.
By contrast, a local concentration of attractant tends to draw amebae
toward it by chemotaxis (strength x). Also, a concentration of amebae will
provide a corresponding increase of attractant concentration because of the
higher local concentration of secretion sources. The strength of this effect is
measured by a0 f This explains the appearance of the destabilizing factors
a0 f and x in the numerator of (22). Instability ensues if destabilizing effects
outweigh stabilizing.
The explanation of aggregation provided in the previous paragraph could
perhaps have been provided without doing a mathematical analysis. But, as
is typical, only after carrying through some calculations does one seem able
to discern the physical essence of a situation. Also, the criterion (22) is a
quantitative one, capable of experimental test. No full quantitative tests have
been performed as yet, but it has been discovered that both the chemotactic
sensitivity x and the attractant secretion rate f go up by about 100-fold at the
time aggregation commences. As illustrated in our discussion of galactic
structure, future quantitative tests might reveal defects in understanding that
would never emerge from qualitative considerations.
The above analysis is based on a paper (op. cit.) that appeared in 1970
in the Journal of Theoretical Biology. Since that time, several other papers
on slime mold aggregation, by various authors, have appeared in that journal
and elsewhere.They deal with such matters as (i) a more accurate modeling of
Appendix I.I] Some Views on Applied Mathematics 31
the chemical kinetics and hence (among other things) a possible refinement of
the" very large territory size" prediction mentioned above; (ii) the streaming
type ofaggregation, as in the frontispiece; and (iii) the influence of pulsatile
attractant secretion and refractory periods. Future work will doubtless
examine the effects of nonlinearity. Two survey articles on slime mold aggrega-
tion and other collective chemotactic motions, written for a mathematical
audience, can be found in Some Mathematical Questions in Biology, Vol. III
(Providence, R.I.: American Mathematical Society, 1972).
EXERCISES
1. (a) Verify {19).
(b) Show that the roots of the quadratic equation (20) are real.
(c) Show that both roots are negative if and only if both b and c are
positive.
(d) Show that since!, D, and x are positive, if c > 0, then b > 0. Thereby
deduce (21).
2. Investigate the changes in the analysis that would occur under the follow-
ing conditions.
(aJ If and x were functions of p.
(b) If k and/were functions of p.
3. If two-dimensional variation is considered, the governing equations {10)
and (11) are replaced by
oa op
ot=V(Va-xaVp), at =fa - kp + DV 2 p.
Assume disturbances with spatial dependence sin (q 1x + q 2 y + 0), where
q1 , q2 , and(} are constants. Show that if q 2 = qf + qL then 2n/q remains
the disturbance period and the instability condition remains (22).
4. Define .6.by xa 0 fl k = I + .6.and suppose that .6. is small and positive.
[Compare (22).] Ignoring higher order terms in .6.,find an approximation
for the larger root of the quadratic. Deduce that the wavelength of the
fastest growing disturbance is approximately 2n(2D/kA) 1 ' 2 .
Appendix 1.1 Some Views on Applied Mathematics

ON THE NATURE OF APPLIED MATHEMATICS
I . . . describe applied mathematics as the bridge connecting pure mathematics
with science and technology. I have deliberately described this bridge as connecting
two areas of activity rather than leading from one to the other, because the bridge
carries two-way traffic. Its importance to science and technology is obvious, but it
is not less important to pure mathematics, which would be poorer without the
stimuli coming from the applications.
W. Prager, "Introductory Remarks" in the special issue, "Sympo-
sium on the Future of Applied Mathematics," Quart. Appl. Math. 30,
1-9 (1972).
One should not judge applied mathematicians by their contribution to mathe-

matics any more than one judges historians by their contributions to history.
Much creative effort goes into interpretation. To try to use mathematical ideas on
problems which at first sight seem too complex or obscurely stated to fit any pattern
is, for many people, an exciting challenge and it is this which I feel to be the essence
of applied mathematics.
R. W. Brockett, "The Synthesis of Dynamical Systems," ibid., pp.
41-50.
Here then are three topics on computation-algorithmic complexity, operating

system modeling, and data structuring-which could use much.heavier mathematical
participation than they now receive. Their requirements are far from the kind of
approximative analysis that has been so successful in the physical sciences. But they
will be large areas of endeavor with many people engaged in their engineering
development and in their use-larger than most of the uses of mathematics we now
have. I think we shall have to broaden our idea of what applied mathematics is to
include these subjects.
H. Cohen," Mathematical Applications, Computation, and Complex-
ity," ibid., pp. 109-121.
I feel that 'applied mathematics ' should be defined ' dynamically' and operation-
ally as the process-and the study of the process-of application of mathematics to
the other disciplines. 'Applied Mathematics' is what' applied mathematicians' do,
while 'applied mathematicians ' are people who go out and talk to those in other
disciplines who are attempting to use mathematics, and who know enough about
the discipline itself to understand the problems on a deeper level.
P. Herman, "Comments on the COSRIMS Report," Amer. Math.
Monthly, 517-21 (May 1970).
ON THE RELATIONSHIPS AMONG PURE MATHEMATICS,

APPLIED MATHEMATICS, AND THEORETICAL SCIENCE
I think that it is a relatively good approximation to truth-which is much too
complicated to allow anything but approximations-that mathematical ideas
originate in empirics .... But, once they are so conceived, the subject begins to live
a peculiar life of its own and is better compared to a creative one, governed by
almost entirely aesthetical motivations . . . . As a mathematical discipline travels far
from its empirical source [however] .... there is a grave danger that the subject will
develop along the line of least resistance, that the stream, so far from its source,
will separate into a multitude of insignificant branches, and that the discipline will
become a disorganized mass of details and complexities.
John von Neumann, "The Mathematician," reprinted in the World
of Mathematics (New York: Simon & Schuster, 1956), p. 2063.
'Applied Mathematics' is an insult, directed by those who consider themselves

' pure' mathematicians, at those whom they take for impure. Mathematics is,
always has been, and always will be pure mathematics. The adjective ' pure' is
redundant. The very essence of mathematics is abstraction, created by fancy and
tempered by rigor, but pure mathematics' as a parricide tantrum denying growth
Appendix 1.1] Some Views on Applied Mathematics 33

from human sensation, as a shibboleth to cast out the impure, is a disease invented
in the last century . . . .
C. Truesdell, "The Modern Spirit in Applied Mathematics," /CSU
Rev. World Sci. 6, 195-205 (1964).
The miracle of the appropriateness of the language of mathematics for the formu-
lation of the laws of physics is a wonderful gift which we neither understand nor
deserve. We should be grateful for it, and hope that it will remain valid in future
research and that it will extend, for better or for worse, to our pleasure even though
perhaps also to our bafflement, to wide branches of learning.
E. P. Wigner, "The Unreasonable Effectiveness of Mathematics
in the Natural Sciences," Commun. Pure Appl. Math. 13, 1-14 (1960).
ON THE TEACHING AND PRACTICE OF APPLIED

MATHEMATICS
It has always been a temptation for mathematicians to present the crystallized
product of their thoughts as a deductive general theory and to relegate the individual
mathematical phenomenon into the role of an example. The reader who submits to
the dogmatic form will be easily indoctrinated. Enlightenment, however, must come
from an understanding of motives: live mathematical development springs from
specific natural problems which can be easily understood, but whose solutions are
difficult and demand new methods of more general significance.
R. Courant, Remarks from the preface of Dirichlet's Principle
(New York: Wiley-Interscience, 1950).
There was complete unanimity on the point that presently mathemati~ majors
do not learn enough about applications . . . . There was also agreement that every
mathematician should become acquainted with at least one area of application.
Joint Report of the Committee on Graduate and Postdoctoral
Education and of the Committee on Undergraduate Education,
1970-71 Annual Report, Division of Mathematical Sciences, National
Research Council, p. 78.
The world is becoming more and more complicated. We are rapidly approaching
an unbearable and even dangerous situation . . . . Mathematics is the only subject
broad enough to prevent this. Not the mathematics actually taught but a new
mathematics inspired by all fields of applications. It must be taught as fundamental
science which provides indispensable modes of thinking and tools to cope with the
real world: the physical world, and the man-made world.
A. Engel, "The Relevance of Modem Fields of Applied Mathematics
for Mathematical Education," Educ. Studies Math. 2, 257-69 (1969).
In mathematics courses the talk is usually about problems. Not much is said
about where the problems come from or about what is done with the answers.
It seems to us that this missing part of the story is worth telling. It is particularly
important to those who intend to be mathematicians, since it explains why the
electronic computer is not going to put them on the unemployment rolls. It is also
important to those who are going into the mathematically oriented sciences, since
it explains why the electronic computer is not going to remove their privilege
(or duty, as it may be regarded) of learning a great deal of mathematics.
R. Hooke and D. Shaffer, Math and Aftermath (New York: Walker
and Co., 1965), p. 4.
I think we need some discipline as mathematicians trying to get into applied

fields. The discipline you need is to be able to turn off the mathematics as soon as
you've discovered what the really relevant physical phenomena are that go on, and
then go on for the new physical phenomena.
Remarks by H. B. Keller, "Symposium on Mathematical Pro-
fessional Societies and Their Present Role," reported on pp. 50-51 of
the 1970-71 Annual Report, Division of Mathematical Sciences,
National Research Council.
Rigor that just polishes the mathematical surface, but does not penetrate essen-
tials of the subject matter in question, is not indispensable in applications.
U. Grenander, "Computational Probability and Statistics," SIAM
Rev. 15, 134-91 (1973).
Too much of the attitude of many young mathematicians is that of always finding
out what is wrong with what is being said [by scientists) rather than what is right;
and I think one of the important things in the game of applied mathematics is to try
to understand a bit of what is right.
E. W. Montroll, "Education in Applied Mathematics" (Proceedings
of a conference sponsored by SIAM), SIAM Rev. 9, 326 (1967).
To a large extent, the community of core mathematicians has decided that it is

not its responsibility to provide instruction related to the application of mathematics;
to the same large extent, much of the instruction in methodology has become the
responsibility of the applied mathematics community. The important question is:
Will this community include in the instruction it offers illustrations of the heuristic,
inventive, reasoning so necessary to progress in science, in technology, and ultimately
in our whole society or will it retreat into a cloistered preoccupation with abstrac-
tions both in its research and its instruction? If it does retreat, then the present
generation and possibly a few future generations of students will receive little
instruction of this sort and those that enter the worlds of technology, of science,
of environmental repair and of medical advancement, will do so severely handicapped
by a grotesquely distorted education.
Our symposium is labeled 'The future of applied mathematics.' It is my view
that if, in answer to the foregoing questions, we make the latter choice, applied
mathematics has no viable future whatever-it could only become a small, rather
sterile corner of mathematics, a discipline which despite its moments of greatness,
already is populous enough to contain more sterility than it wants. But if we make
the former choice, i.e., if among all of its other activities, the applied mathematics
community provides innovative contributions to the mathematization of disciplines
which are just becoming quantitative, if it concerns itself with finding approximate
models for complicated phenomena (thereby sacrificing accuracy of detail for ease
of interpretability), if it seeks to develop techniques whereby the implications of
Appendix 1.1) Some Views on Applied Mathematics 35
given mathematical models can more easily be inferred, then applied mathematicians
will have inherited much of today's continuation of the challenges, the intellectual
achievements, the contributions to society, and the fun which, before they abdicated,
was largely the property of the mathematicians and the physicists.
G. F. Carrier, "Heuristic Reasoning in Applied Mathematics," in
the special issue, "Symposium on the Future of Applied Mathe-
matics," Quart. Appl. Math. 30, 11-15 (1972).
CHAPTER 2
DeterministicSystems and Ordinary
DifferentialEquations
J N THIS chapter we shall illustrate something of the range of applied mathe-

matics using particle dynamics as our central example. The principal purpose
of our discussion will be to analyze the process of applied mathematics.
We are not primarily interested at this point in enlarging the reader's facility,
so a thorough exposition of certain methods will be deferred.
We first examine the relation between Kepler's laws, which describe the
motion of the planets, and Newton's formulation of the theoretical law of
universal gravitation. The orbits of planets and comets are deduced by the
rather ingenious classical solution of the governing equations. When
relativistic effects are considered, however, attempting to find an exact
solution no longer is an optimal procedure. A perturbation method should
be used. The same method is called for when one wishes to calculate the rela-
tively small effect of other planets upon a given planet's orbit about the sun.
As an illustration of practical calculational procedures that have been
devised to fill a definite need, we discuss perturbation theory in a certain
amount of detail. Roughly speaking, this theory permits one to calculate
the effects of small disturbances on otherwise known situations. Remarkably
valuable information can be obtained by sophisticated variants of perturba-
tion theory-one cannot expect its content to be summarized in a few pages.
Thus our discussion is a foretaste of what is to come in the remainder of
this volume.
Particle dynamics is an example of a deterministic process wherein the
future is completely determined, given sufficient knowledge of the present.
Whether the outcome of such processes is to be calculated in detail or just
contemplated philosophically, certain purely mathematical questions arise.
We conclude this chapter by stating and proving some relevant theorems.
These involve existence, uniqueness, and continuous dependence on a
parameter of systems of ordinary differential equations.
2.1 Planetary Orbits

People of all eras have been moved by the grandeur and the beauty of
Nature when watching stars in the darkness of night. As they made careful
observations, they discovered that some of the "stars" are wanderers-the
= wanderer in Greek). Indeed, records of planetary paths
planets (1rlcxv~r17s
formed one of the earliest important collections of scientific data.
36
Sec. 2.1] Planetary Orbits 37

In the first comprehensive history of China "Shihchi" (Records of the
Grand Historian of China), written about the second century B.C., there is
one volume devoted to celestial observations. Five planets were recognized,
and it was known that the period of Jupiter is almost exactly 12 years. This
is a special case of Bode's law.*
In the West, accurate observations of planetary motions were made by
Copernicus (1473-1543), Tycho Brahe (1546-1601), and Galileo (1564-1642).
The great amount of observational data collected by Tycho Brahe formed
the basis of the famous laws of Kepler (1571-1630). These laws then led
Newton (1642-1727) to the discovery of the law of universal gravitation,
with the help of the then newly invented differential calculus. The study of
these developments provides an excellent classical case study of applied
mathematics.
KEPLER'S LAWS
In their original form, Kepler's laws may be stated as follows:
I. The planets describe ellipses about the sun as a focus.
II. The areas swept over by the radius v~tor drawn from the sun to a
planet in equal times are equal.
III. The squares of the times of describing the orbits (periods) are propor-
tional to the cubes of the major axes.t
To accomplish a mathematical formulation, we adopt a polar coordinate
system (r, (})with the sun as the origin. We treat all the planets and even the
sun itself as mass points. (Note that this is an approximation or idealization
in the formulation. It is appropriate because the radius of the sun is small
compared to the distance between the sun and a given planet.) The second
law of Kepler then states that, following the orbit (r(t), O(t)) of a planet,
r 2 0 = h, ha constant, (1)
* The radii of the planets of the Sun are fairly accurately given by the formula 0.4 + 0.3 x 2"
in astronomical units. The values of n are -oo, 0, I, 2, ... 7, 8 for Mercury, Venus, Earth, Mars,
minor planets, etc. Neptune and Pluto do not follow this law accurately. This fact probably
implies that Pluto was originally a satellite of Neptune.
t "Newton later wrote: 'Kepler knew ye Orb to be not circular but oval, and guessed it to be
Elliptical.' This is correct, but Kepler's guess was no idle guess; it came out of a hunch actively
pursued, in confrontation with all previous theories and with Tycho's new data."
This quotation is from an interesting account by Curtis Wilson, in the March 1972 Scientific
American (Vol. 227, p. 92), of how Kepler discovered the first two of his laws. Wilson's summary
of his findings includes these further remarks.
"At the root of all his [Kepler's] theorizing, however, was that initial sense of the significance
of the inverse relation between velocity and distance-an anticipatory glimmer of what would
one day be the law of the conservation of angular momentum. It was in the light of that hunch
that he was guided through 900 pages of calculation to a planetary theory better than any that
had been proposed before .... What was ... important through all the accident and luck was
Kepler's belief in the possibility of understanding, his devotion to his task that carried him through
four years of reasoning and calculation, and finally, the rightness of his initial hunch and his
ability to disentangle the confused state of things before him in the light of it."
Deterministic Systems and Ordinary Differential Equations [Ch.2
since [Exercise l(a)] h is twice the area swept over by the radius vector in
unit time. (It is appropriate to use the notation for d/dt, as introduced by
Newton.) To obtain a useful simple interpretation of this result, we introduce
the cartesian coordinates (x, y). Then (1) can be written as
tj-~=~ m
which yields, upon differentiation with respect to t,
xy-yx =0. (3)
If we define vectors r and s by
r = xi + yj, s = - yi + xj, (4)
then r s = 0, so s is perpendicular to the position vector r. But according to
(3), i' s = 0 so that i' is in the direction of r. That is, if the "equal areas"
result holds, then the acceleration of the planet is always directed toward the
sun (or away from it). Reversal of the steps in our reasoning shows that the
converse of this result is also true.
The lesson to be learned here is that the interpretation of mathematical
results may depend very muth on the form in which they are stated. The
equivalent equations (l) and (3) enable us to see quite different aspects of
the same phenomenon, with completely different interpretation and insight.
Note the importance of manipulative skill.
The first law of Kepler states that the orbit can be described by the simple
formula
p
r= , (5)
l+ecose
where p > 0 and O < e < I. This is a standard form of an equation for the
ellipse whose semimajor axis is given by p/(1 - e2 ). With a little manipulation
[Exercise 2(a)], one can show that the acceleration in the radial direction is
h2
a, = ;:- re2 = - -2 . (6)
pr
Thus the acceleration is inversely proportional to the square of the radial
distance. Note that nothing has been said about the forces between the
celestial objects. [A statement, such as (6), which describes the geometric
character of motion, or which follows from this character, is said to be
kinematical.]
The quantities h and p are characteristics of the individual planets. It is
by the use of the third law of Kepler that one can show that the ratio h2 /p
is the same for all the planets [Exercise 2(b)]. Thus, for planetary motion,
K
a,= -z,r (7)
where K is a universal constant.
Sec. 2.1] Planetary Orbits
39
LAW OF UNIVERSAL GRAVITATION
The laws of Kepler have led us to the inverse square law for the accelera-
tion of all the planets, with the same constant of proportionality. Newton,
by combining the above results with his second law of motion and assuming
the principle of superposition (i.e., that the effect of two suns is double that
of a single one), was led to formulate the present form of the law of universal
gravitation,
F. = GMm_ (8)
r r2
Here F, is the magnitude of the force acting between two point masses of
amounts M and m separated by a distance r; G is a universal constant. The
force is attractive.
The simple elegance of the mathematical manipulations is striking, especi-
ally when one compares the theory with the tedious observations. But can
this powerful law be dreamed up without these observations? The answer is
obviously no. Neither can the required calculations be made without the
use of the then newly invented discipline of calculus.
The law of universal gravitation is indeed one of the most remarkable
of the laws of nature. Why is it universal? Why is the power exactly equal to 2?
Even more remarkable is that, according to the general theory of relativity,
such an elegant principle can be replaced by another, which has even more
penetration and breadth and is at least equally elegant, if not more so. It
hardly seems frivolous to say that God is a mathematician!
On the working level, we observe that in order to arrive at our conclusions
it is necessary to keep the physical interpretation of the results in mind all the
time and to be quite skillful with the manipulations.
There is another lesson to be learned from the history of the law of universal
gravitation, giving further evidence of t.he need for adequate notation and
of proficiency with analytical manipulations. Newton tried to test his law by
applying it to the motion of the moon about the earth and comparing the
moon's acceleration with that of a body at the surface of the earth, as directly
observed. This comparison cannot be made without knowing the gravitational
potential of a continuous distribution of matter over a sphere. It is now well
known that the gravitational attraction due to such a distribution of mass is
equivalent to that due to a concentrated mass at the center. Newton was at
first not able to prove this and consequently delayed his publication of the
comparison for about 20 years.
THE INVERSE PROBLEM: ORBITS OF PLANETS AND COMETS

We start now with Newton's law (8) or the kinematical law (6), and ask
how the planets would move. We find that (i) we have to be able to solve
differential equations to provide a satisfactory answer, and (ii) not all the
40 Deterministic Systems and Ordinary Differential Equations [Ch. 2
orbits are ellipses. Some orbits are hyperbolas and others are parabolas-
and nature has provided these examples in the cometary orbits. The solution
of the relevant differential equations, as we shall now see, requires consider-
able skill in manipulation.
The formulation of the mathematical problem consists of asking for
periodic orbits for the following system of two differential equations of the
second order:
2
d r_ ,(d()) 2
=_GM (9a)
dt 2 dt r2 '
!~
r dt
(,2d())O
dt
=
'
(9b)
where the expressions on the left-hand side are the components of accelera-
tion (a,, a8) in the polar coordinate system. To follow the motion of the
planets, one has to integrate (9) for the pair of functions r(t), O(t). There
seems to be no direct way to accomplish this analytically. (Numerical integra-
tion methods can be applied, but understanding and insight would tend to
be lost.) The ingenious way out of the difficulty is to look for the orbit in the
form r = r(O). Even here, as we shall see below, a clever transformation is
needed to solve the resultant differential equation.
Equation (9b) can be integrated once to yield r 2 e= h, h a constant. (The
quantity r 2 0 is called a constant or integral of the motion.) Thus the law of
uniform area (1) is recovered. We can now use (1) to replace the time deriva-
tives in (9a) with differentiation with respect to the angular variable 0. The
resultant equation is still very complicated, but it can be simplified by intro-
ducing the variable
With this we obtain
(10)
The general solution of (10) is [Exercise 3(b)J

u = h- 2 GM[l + e cos (0 - 00 )], (11)
where e and 00 are constants of integration.

The solution (11) is periodic in (},but not all the orbits are periodic. When
e < I, (11) yields the observed elliptical orbits. But when e > 1, u may
vanish for some value of(), and the planet would then be infinitely far away;
the orbits are hyperbolas. Parabolic orbits are obtained in the marginal case
e = 1. The integration constant 00 is unimportant as it merely assigns the
value 00 to an axis of symmetry.
Sec. 2.1] Planetary Orbits
41
PLANETARY ORBITS ACCORDING TO THE GENERAL THEORY

OF RELATIVITY
When modification of Newtonian mechanics is considered in the light of
the general theory of relativity, it can be shown that equation (IO) for the
planetary orbit is changed to the form
d 2u GM 2
d</>2
+u= V (1 + eu ), (12)
where e is a small parameter (and where the influence of the other planets
still has not been taken into account). The integration of this equation is
more difficult, since it is nonlinear in u. However, it can still be done in terms
of elliptic functions. To obtain the solution in a more convenient form, the
problem should be treated by "perturbation theory," which is based on
the fact that the solution of a differential equation depends continuously
on a parameter that enters continuously in the differential equation. (We
shall soon discuss these matters further.) In the present case, when e = 0,
the solution is known. Because of the continuous dependence of u upon e,
the solution for small e can be obtained by an appropriate small modification
of the solution for e = 0.
The result obtained by the use of the perturbation theory shows that the
orbit is still very close to an ellipse, but there is an advance of the perihelion
(nearest point to sun) by the amount 2nea2 over one period. (A method for
obtaining this result is given in Section 2.2.) The observation of this advance
yields one of the principal experimental tests of the general theory of
relativity.*
COMMENTS ON CHOICE OF METHODS

The advance of the perihelion can also be established by detailed numerical
integration, or by examination of the solution of (12) in terms of elliptic
functions. However, neither method gives the results in as transparent a
form as perturbation theory, which is also extremely accurate in this case.
In other instances, analysis of a well-formulated problem with the aid of
modern computing machines can often yield more accurate numerical
results than those obtained by the use of approximate analytical methods.
It is, however, often difficult to extract useful general information from a
large amount of numerical data unless one has already gained insight into
the problem through general analytical studies. On the other hand, exploratory
* Equation (12) has been derived by K. Schwarzschild [R. Adler, M. Bazin, and M. Schiffer, In-
troduction to General Relativity (Tokyo: McGraw-Hill, Kogukusha, Ltd., 1975), p. 204.J The present
situation on the relativistic perihelion advance with regard to the comparison of experiment with theory
is as follows. For Mercury, where the effect is largest (43 seconds of arc in a century) observations are
in accord with Einstein's theory. But some controversy remains, for there is a possible flattening of
the sun that could perhaps leave room for alternatives to Einstein's theory. SeeK. Nordtvedt, "Gravitation
Theory: Empirical Status from Solar System Experiments," Science 178. 1157-1164 (1972).
numerical studies sometimes are an invaluable preliminary to successful

analytical work. The judicious balance in the use of all types of methods is
indeed most important.
N PARTICLES: A DETERMINISTIC SYSTEM

Consider now a system of N particles in gravitational interaction; e.g.,
the solar system comprising the sun and the nine major planets. Let us denote
the position vector and the velocity vector of the particle (Xby r .. and v..,
respectively ((X= 1, 2, ... , N). Then the equation governing the motion of
the (Xthparticle is
(X= I, 2, ... , N, (13)
where
(14)
i.e., the resultant force acting on the (Xthparticle is due to the contribution
F..p of all the other particles. We have, according to Newton's law,
Gm..mp
F ..p = - Ir .. -r/J l3 (r .. - rp) = -Fp .., (15)
which depends only on the positional coordinates of the particles. In other

physical situations, there may be an external force field and the forces may
also depend on the velocities of the particles (e.g., when the particles are
electrically charged and there is a prevailing magnetic field). But, in general,
there are no true forces dependent on the acceleration of the particle.
Thus we have
(X,/1,y = 1, 2, ... , N. (16)
In this most general case, we may still write the equations of motion (13)
as a system of differential equations of the first order:
dv.. dr..
m.. dt = F..(t, rp, v1), dt = v.. ; (X,p,y = 1, 2, ... , N. (17a, b)
We have 6N variables (r .., v..); (X= 1, ... , N. We expect that the subsequent
behavior of the system would be determined ifwe were given the instantaneous
positions and velocities of all the particles at any instant. Mathematically,
this means that (17) should have a well-defined solution if it is solved under
the conditions
r .. = r~0 >, v.. = v~0 > at t = t0 ; (X= 1, 2, ... , N. (18)
These conditions are called initial conditions, and the problem we have just
posed is called an initial value problem for a system of ordinary differential
equations.
Sec. 2.1] Planetary Orbits 43
That the initial value problem has a unique solution can in fact be proved,
as we shall see in Section 2.3. This proof has wide ranging consequences,
for it has led some philosophers to a deterministic view of the universe. For
what can be the place of" free will," it can be argued, if the evolution of the
cosmos can be uniquely determined by examining the configuration at any
arbitrary instant?*
But nowadays existence and uniqueness theory is also important from a
practical (rather than philosophical) point of view. The advent of computing
machines has made it desirable to ensure that a problem is precisely formu-
lated before expensive calculations are attempted. At least at the present
stage of development of computers, one cannot expect the machine to detect
difficulties in the original formulation and then to suggest suitable modifica-
tions. In practice, trial-and-error approaches are often used, for it has certainly
not been possible to prove all the required theorems. But experienced opinion
is superior to guesswork, and the ability to exercise sound judgments in the
absence of rigorous guidelines is perhaps one of the most valuable assets of
an applied mathematician from the practical standpoint.
LINEARITY
To make sure that an important concept is understood, we close this
section with a brief review of linearity. The concept will be illustrated in
connection with systems of differential equations.
If f(x) =ax+ b, a and b constants, the graph of y = f(x) is a straight
line. For this reason, f is called a linear function. When b = 0, the function
is called homogeneous; otherwise, it is inhomogeneous.
The linear homogeneous function (often just called linear) has strikingly
simple properties. If L(x) = ax, then
L(x 1 + x 2) = L(xJ + L(x 2 ), (19a, b)
or, equivalently,
(20)
Here c1 and c2 are any constants; x 1 and x 2 are arbitrary values of x.
Let x(t) be an n-element column vector (n by 1 matrix) that depends on
the scalar t. Define a vector-valued function L(x) by
dx
L(x)=- -Ax (21)
dt '
where the derivative is (as usual) taken component by component, and where
A= A(t) is an n by n matrix that multiplies x. Note that A does not depend
on x.
* Needless to say, we are only scratching the surface of the philosophical position, and we
shall not pursue the matter funher.
Deterministic Systems and Ordinary Differential Equations [Ch. 2
44
The function L also possesses the linearity properties. For if x and y are
any (differentiable)vector functions and a:is a scalar (usually a real number),
then
IL(x + y) = L(x) + L(y), L(a.x)= a.L(x). , (22a, b)
We also have the analogue of (20), which treats of what is called the linear
combinationa.x + pyof x and y (where p is another scalar):
IL(a.x+ PY)= a:L(x)+ PL(y).1 (23)
The reader has probably met some of the many consequences of the very
important linearity properties (22) or (23). It is relevant here to give an
example involving systems of ordinary differential equations. Using the
definition (21),
L(x) = 0
is a linear (homogeneous) system of first order differential equations. Property
(22a) implies that the sum of two solutions is a solution for
L(x) = 0, L(y) = 0,
implies
L(x + y) = L(x) + L(y) = 0. (24)
Property (22b) implies that the scalar multiple of a solution is a solution, for
L(x) = 0 implies that L(a:x)= ixL(x)= 0. (25)
Similarly, a linear combination of two solutions (or of many solutions) is a
solution.
EXERCISES
1. {a) Equation {I) is often derived in elementary treatments of polar
coordinates. If the derivation is unfamiliar to you, work it out for
yourself.
{b) Verify (2).
{c) Show that {I) follows from (3).
2. {a) Verify (6).
(b) Prove (7).
3. {a) Verify {IO).
(b) Derive (11).
4. In the text's theoretical discussion of Kepler's laws, the sun and the
planets were treated as mass points. Write a brief essay that provides
justification for such a treatment.
5. {a) Verify (22) and {23)when Lis defined by (21).
(b) Prove in general that (22) holds if and only if {23)holds.
{c) Prove that a linear combination of solutions to a linear system of
equations such as (21) is also a solution.
Sec. 2.2] Elements of Perturbation Theory
45
6. Eliminate () from the equations of (9) to obtain a differential equation

for radial motion. Integrate to derive Kepler's representation of radial
motion:
r = a(l - e cos E), n(t - T) = E - e sin E.
Here a is the semimajor axis of the ellipse, e is the eccentricity, n is the
frequency of the orbit, Tis the time of perihelion passage, and E (called
the eccentric anomaly) is a parameter that runs over the range (0, 21t)
for a complete orbit. (The position angle() is known as the true anomaly;
the quantity M = E - e sin E, which varies linearly in time, is known as
the mean anomaly.)
To do this, first obtain the energy equation in the form
dr) 2 2
2 2
a e - (r - a)2
(-dt
=h
r a (1 - e2 )
2 2
by noting that the radial velocity is zero at the perihelion and aphelion
(the points which are nearest and farthest from the sun, respectively).
2.2 Elementsof PerturbationTheory,IncludingPoincare's

Methodfor PeriodicOrbits
At the beginning of this chapter we considered a single planet moving
around a fixed sun. The subsequent discussion of N-particle interactions re-
minds us that our earlier considerations neglected the interaction among the
planets.* We can build on our previous work by "approximating" the force
F,. on the ixth planet by a basic force field F!0 > (due to the sun), and regard-
ing the effect of the other planets as" small perturbations." To formalize this
process, we can replace the right-hand side of (1.13) or (1.17a) by
F,.(,l) = F!0 > + .A.[F,.
- F! 0 >J
and allow the parameter .A.to vary in the interval (0, l).t The effect of the
perturbation is then obtained by comparing the solutions for .A.= 0 and
.A.= 1, respectively.
In practice, determination of the effect of small perturbations on the
motion of planets involves special difficulties.This is because the perturbation
effects accumulate over many periods of time, yet we do not expect the shape
of an individual orbit to be greatly changed. The difficulty was overcome by
Poincare, whose ideas will be discussed briefly after a presentation of the
theory in the ordinary form.
Also neglected was the solar motion, whose effect can be studied only by introducing another
set of dependent variables.
t Although not necessary, it is instructive to imagine that as.\ varies continuously from Oto I,
the solar system varies from one with but a single planet, through intermediate systems with
other planets of small mass, to the actual system of which we are a part.
PERTURBATION THEORY: ELEMENTARY CONSIDERATIONS

To clarify the ideas just discussed, let us consider the simplest example
of a single ordinary differential equation
dy
dx = f(x, y, e) (1)
to be solved under the initial conditions

Y=Yo whenx = x 0 . (2)
We wish to examine the dependence of the solution y(x, e) on the parameter e
that occurs in (1).
Suppose that the solution y(x, e) is known fore= O.* Assume that y(x, e)
can be expanded into a Taylor series in e, when the latter is sufficiently small
in magnitude:
y(x, e) = y<0 >(x,0) + ey<ll(x,O) + + e"y<">(x,
0) + . (3)
y<n>(x,
0) =~(an~)
n. Oe
. t=O
(4)
If we now substitute (3) into (1) and regard the equation as an identity in e,
we shall find that we can obtain a sequence of linear differential equations
in y<n>.
For example, disregarding all terms of order e2 or higher, we have
dy(O) dy
dx + e dx + ... = f(x,y<o> + e/1> + , e)
+ efy(x, y<0 >,O)y<1 >+ ef.(x, y<0 >,O) + ,

= f(x, y<0 >,0)
where we have used Taylor's theorem to expand f Equating equal powers
of e, we find
dy(O)
dx = f(x, y<o>,0), (5a)
dy<l>
dx = /y(x, y<o>,O)y<O+f.(x, y<o>,0), etc. (5b)
The solution for (5a) has been presumed to be known. Upon substitution
of this solution into (5b), one obtains a definite equation for the correction
y<l)_
The initial conditions to be imposed on the successive correction terms are
yC1l(x 0 , 0) = 0, y<2 >(x0 , 0) = 0, ... , y<n>(x
0 , 0) = 0, (6)
since the original condition (2) does not depend on e.
This entails no loss of generality. If the solution is known for E = e0 , define a new parameter
E =E-Eo.
Sec. 2.2) Elements of Perturbation Theory 47

Note that (Sb) can be explicitly solved in terms of quadratures under the
appropriate initial condition of (6). Indeed, calculations will show that all
the terms y<">(x,0) satisfy an equation of the form
dy<n>
-- = Ay(n) + B(n), n = I, 2, ... , (7)
dx
where A is f/x, y<0 >,0) for all n, and B<n>depends on x and all the y<t>s
preceding y<">. Thus, once y<0 > is found, the problem of obtaining the higher
approximations is relatively simple in principle, since explicit solution of
(7) is possible in terms of quadratures.
A simplifying feature of this type will be found in all other cases of the
perturbation method, including those involving partial differential equations.
Example 1. The solution of the differential equation

dy
dx = 1 +Y', (8)
subject to the initial condition

y(O) = 0, (9)
is
y = y<0> = tan x. (10)
Find a first approximation to the solution, subject to the same initial condition, of
dy
- = 1 + (1 + e)y 2'
dx
(11)
when Ie I ~ I. Compare with the exact solution.

Outline of solution. The exact solution to the perturbed problem is
y(x, e) = (1 + e)- 112
tan [(1 + e) 1 "x].
A Taylor series expansion of this solution in powers of e gives
y(x, e) = tan x + e(!x sec2 x - ! tan x) + .
Forgetting the exact solution for the moment, we assume that

y(x, e) = y<0>(x)+ ey<O(x)+ .
On substitution into (11), we find
dy(O) dy(l)
dx +e dx + ... =1 + (1 + e)[(y<D>)'+ Zey<o>y<o+ .. ],
so that
dy<O
-- = (y<Ol)1+ Zy<D>y<ll.
dx
[It is frequently the case, as here, that direct substitution provides the desired
perturbation equations more easily than substitution into a general formula such as
(Sb).] Rearranging the differentialequation for y<0 into standard form and using
the fact that y<0 >(x)= tan x, we obtain
dy(I)
- - 2y<O tan X = tan 2 X. (12)
dx
This equation has an integrating factor of
exp (-2 J tan x dx) = cos 2
x.
Indeed, (12) can be written
d
- [y 11> cos2 x] = cos2 x tan 2 x
dx
=sm. 2
x.
It is now easily seen that the solution for y 10 is preciselyas given above.
The reader should fill in the details of the above calculation (Exercise 2).
Hopefully, a good idea of how straightforward perturbation calculations
are accomplished has already emerged. But detailed "training" and further
discussion is provided in Chapter 7.
THE SIMPLE PENDULUM

Theoreticians often find it helpful to consider the operation of an idealized
device that displays interesting properties in a readily visualized form. One
such device is the spring-mass-dashpot system. As most readers will have
seen, in the simplest instance this system is governed by a second order
ordinary differential equation with constant coefficients; it therefore provides
a concrete illustration of the properties of that ubiquitous equation. Here,
for the first of several times in this book, we shall consider another example
of such an idealized device, the simple pendulum. In the present context,
consideration of the pendulum will provide an illustration (i) of the fact that
perturbation theory is often used implicitly when we formulate simplified
physical models, and (ii) of the large cumulative effect of small perturbations
and the method Poincare devised to describe this effect. The same effect
as that of (ii), and the same calculational method but in a less transparent
form, are relevant to planetary orbits.
A standard discussion in elementary physics texts is a reduction of the pen-
dulum problem to that of simple harmonic motion when the amplitude is
small, and we vaguely feel we can improve the approximation when needed.
The discussion deals with a simple pendulum as an idealized object which, by
definition, consists of a rigid, straight, weightless rod of fixed length L, with
point mass m at one end. The rod is free to rotate about a horizontal axis
at the point of suspension.
Sec. 2.2) Elements of Perturbation Theory
49
mg
FIGURE 2.1. Forces on a simple pendulum of mass mare represented by

the dark arrows. Magnitudes of the forces are indicated. The tension T
in the string is equal to mg cos 8. The unbalanced tangential force of
magnitude mg sin 8 causes acceleration of the bob.
The italicized words above represent the idealizations involved in the

formulation of the model. Estimation of the error introduced by these idealiza-
tions might be very difficult. (For example, if the rod is slightly elastic it
becomes capable of undergoing torsional, longitudinal, and bending vibra-
tions.) But we shall consider an approximation whose consequences can
be more easily analyzed. Denote by O(t) the inclination of the rod to the
vertical at time t. If s(t) denotes the distance the bob has moved from the
vertical in a counterclockwise direction, then OL = s. Equating m(d 2s/dt 2 ) to
the force component which is tangential to the path of the bob, we obtain
the following equation of motion:
miiJ = -mg sin 0. (13)
Here g is the acceleration due to gravity. The tension along the string does
not enter, because it acts normally to the path of the bob. (See Figure 2.1.)
The equation of motion may also be written in the following slightly
different form:
g)l/2
0 + wlsin O= 0 for t > 0, where co0 = (L . (14)
To obtain a definite solution, we may impose the initial conditions
f o= a, 0=n at t = o. 1 (15)
Equations (14) and (15) comprise the formulation of our mathematical

problem.
SUCCESSIVE APPROXIMATIONS TO THE MOTION OF THE
PENDULUM
In the usual treatment of the problem, the approximation sin 9 ::::() is
introduced. Then (14) becomes
0 + w~9 = 0, (16)
whose solution [subject to condition (15)] is given by
()o = a cos Wot + -n SID

.
Wot. (17)
Wo
An obvious inference from the solution is that, irrespective of the initial

conditions, the motion is periodic with the period
P= 2 = 2n(~)
0
Wo
,r
g
1 2
' (18)
This approximate solution is not adequate, however, if we are interested

in following the motion of the mass point over many periods, for the period
given by (18) is not exactly right. According to the "exact" equation (14),
the motion is indeed periodic, but the correct period is
2
p = Po . -
7t
fff/2
O
(1 - k2 sin2 l/lr112 di{,, (19)
where k 2 = sin2 (Bm/2)and (Jm is the amplitude of the oscillation [i.e., the
maximum value of IB(t)I]:
Om= max IO(t)J. (20)

r
With every swing of the pendulum, the approximate solution misestimates

the time of maximum amplitude by a further increment of P - P O [The
error in period is about0.2 per cent for an amplitude ()m = 10 = n/18 radians
(Exercise 4).] Thus the solution to the approximate equation (16) can be
expected to be nearly the same as the solution to (14) only for a number of
periods N such that N(P - P0 ) is small compared to P.
To improve on the approximation, one might write (14) in the form
0 + Q)~ () = wMO- sin 9) (21)
Sec. 2.2] Elements of Perturbation Theory 51
and then obtain a reasonably accurate value of the right-hand side of (21)
by using the first approximation 00 of (17), writing e - sine~ 80 - sin 80 .
This is justified by the argument that the right-hand side of (21), namely,
e3 es ) = w~ e3
wl(O - sin 0) = <.o~
((} - (} + -3! - -5! + - - . (22)
3! '
is of the order of 83 , while the left-hand side is of the order of Thus the e.
smaller right-hand side can be ignored altogether to obtain an initial approxi-
mation 80 (t). A better approximation should be obtained if the substitution
(} = 80 is made on the right side, etc. Thus the first attempt at an improved
approximation, 81(t), might be expected to satisfy
81 + <.o~0
1 = <.o~(0
0 - sin 80 ); 01(0) = a, 01(0) = 0. (23a, b, c)
But to make an exact evaluation of the right side of (23a) with the rough
approximation 00 gives rise to a problem that is needlessly difficult to solve.
As one might suspect, no accuracy is lost if one makes the approximation
indicated in (22) and therewith replaces (23) by
(24)
In principle, repetition of the process will provide successively more

accurate approximations. Hence the method is known as iteration (Latin
iterare, to do again), or successive approximations.
Problem (24) also arises in another approach to the improvement of the
basic approximation 80 , to be discussed now. We therefore defer considera-
tion of the solution to (24).
No TE. Chapter 7 contains a more thorough treatment of the method
of iteration.
PERTURBATION SERIES APPLIED TO THE PENDULUM
PROBLEM
In a perturbation series approach, it is convenient to introduce a new
variable,
O(t)
0(t)=-, (25)
em
where emis the maximum amplitude defined in (20). Intuitively, emseems
to provide a natural standard of comparison for the angular displacement.
The new variable 0 satisfies I0 I ~ 1, and this proves useful in estimating
the size of various terms. Considerable further justification for this type of
variable change is given in Section 6.3.
With (25), the governing equation (21) becomes
03
El+ <.o
2
O0
2
= <.o
O(0 - e-m 1 sine m 0) 2 2
= <.o8 -
Om6 +... (26)
We note that since IE>I~ I, the right-hand side of (26) is of the order* of
8!, a quantity that we assume to be very small compared to unity.
For simplicity, let us consider the case where the pendulum is released from
rest (0 = 0). It is physically clear that Bm= a, so that our problem becomes
(:) + w~E>= w~(E>- a- 1
sin a0)
oo 92n+1
= w~ L,(-1)"+ l(a2)" ' (27a)
n=l (2n+l)!
0(0) = l, 0(0) = 0. (27b, c)
To find a solution of (27), we write
) = 0(t, a) = E)(O)(t)+ aE)Cl)(t) + a 2)(l)(t) + ''' (28)
and substitute it into (27a) to obtain an infinite sequence of equations, in
the spirit of our earlier discussion of perturbation theory. The initial conditions
for the various approximations are
e<0 >(o)= 1, e<0 >(o)= o; (29a)

e<n>(O)
= 0, e<n>(O)
= 0, for n ~ I. (29b)
In this formulation, the dependence of the solution on the amplitude a is

explicitly demonstrated. Note that if all the higher powers of the amplitude
a are neglected, we have the usual simple harmonic motion.
To ensure that our method works, we might wish to satisfy ourselves
that the series (28) converges. But this turns out to be an incorrect stipulation,
for it is neither necessary nor sufficient. Even when such a series is divergent,
it can still be useful as an asymptotic series (see Chapter 3). On the other hand,
convergence does not assure its usefulness. An inordinately large number
of terms might have to be summed to obtain reasonable accuracy.
As a matter of fact, it can be shown that the series (28) is in general con-
vergent in any finite interval O ~ t ~ T, provided that a is sufficiently smalI.t
However, this does not assure us of the usefulness of the solution for ma:-"
periods. One way to see this is actually to carry out the calculations, which
show that
E)(O)(t)= COS Wot, (30)
* "Of the order of," roughly speaking, means "about the same size as," but it is useful to
give a precise meaning to "order" in this sense. See Appendix 3.1.
t This can be inferred from the analytic dependence of the right-hand side of (27a) on the
complex variable a. This general theory is beyond the scope of the present discussion. See
Coddington and Levinson (1955).
Some readers may have fully grasped the idea of straightforward perturbation theory from
the single example that has been worked out. They can attempt to verify (30), (31), and (33) by
themselves. Other readers may prefer, either now or later, to peruse Section 7.1, where the required
calculations are carried out in detail.
Sec. 2.2] Elements of Perturbation Theory 53
Thus 0(2) satisfies

())2
(:)<2 >+ ro~0(2) = cos3 w 0 t. (31)
6
Using a trigonometric identity, we can rewrite (31) as
(2) (2) ro2
0 + ())~
E) = ; (cos 3root + 3 COSWot). (32)
2
This is a nonhomogeneous linear equation with constant coefficients. The
right-hand side is of the simple type that invites the use of the method of
undetermined coefficients. The presence of the "resonant" cos (ro0 t) term
leads us to anticipate a contribution to the solution with a factor of t. And
indeed, the solution of (32) and the initial conditions (29b) turns out to be
0(2)(t) = 1 h cos ro0 t - 1 J2 cos 3ro0 t + n-Wo
t sin Wot. (33)
Noteworthy in (33) is the term proportional to cos 3ro0 t. This appearance
of more rapidly oscillating "higher harmonics" is a typical nonlinear
phenomenon.
But the term proportional to t sin ro0 t holds the center of our interest.
Alone among the contributions to the correction term a 2 0(2), this term is
not uniformly small for all time. Rather, for long times t ~ ro0 1a- 2 we must
contemplate a secular* term which, far from being a small correction, has
become of the same size as e<0 >.When this happens, the whole basis of the
approximation is invalidated. Thus the approximation e<0 > + a 2 e<2 > is an
improvement over e<0 > only for the time interval that e<0 > itself offers a
reasonable approximation (Exercise 5). We have not yet penetrated to the
core of the problem.
POINCARE'S PERTURBATION THEORY
The failure of the unmodified perturbation method occurs in all cases of
periodic motions. In connection with the study of the perturbation of plane-
tary orbits, Poincaret developed a modified perturbation theory to overcome
the difficulty. The key to his theory is the introduction of a parametric
representation that recognizes the change of period indicated by (19).
This permits the removal of any secular term that might arise in the calcula-
tions.
The Poincare representation of the solution E>(t,a) is
E) = 9(0)('r) + aE)C1l(-r)+ a 2 0{2)(-r) + , (34a)
(34b)
Saeculum means "generation" or "age " in Latin; hence the word "secular" applies to
processes of slow change.
t Poincare (1854-1912) was one of the world's leading mathematicians around the turn of
the century.
54 Deterministic Systems and Ordinary Differential Equations [ Ch. 2
where -r is a parameter. The idea is to let the solution 0<0 >(-r)take on the
simple form (34a) in the variable -r, but to distort the time scale so that the
defects in the solution are removed. We expect to eliminate all secular terms
and to get the correct period (19) to a good approximation. The calculation
is fairly complicated but can be simplified by noting that the series (34) should
really be in a 2 It also turns out that one may take all the correction terms
to be a constant multiple of -r, thus
t = -r(l + a2 h2 + ), h 2 a constant. (35)
The equation for 0(2) is then found (Exercise 6) to be
d20(2) d20(0) ,1>2
~ + ru~0(2)- 2h 2 ~ = : (cos 3ru0 -r + 3 cos ru0 -r), (36)
2
or, using (30),
d20(2) (J)2
dr + rul 0(2) = rol(t - 2h2) COS Wo 't' + 2; COS 3COo
't'.
As in (33), the cos ru0 -rterm on the right side would lead to an undesirable
secular term proportional to T sin ru0 -r. But now we can avoid this by taking
(37)
The period in -r is thus 2n/w 0 to this approximation, but is (2n/co0 )(1 + a2 h 2 )
in t. This should be compared with (19).
The reader can better appreciate the need to change the time scale by
obtaining the pendulum period using exact integration (Exercise 9). A more
interesting example involves the advance of the perihelion of Mercury's
orbit (Exercise 10). The reader may find it easier to do Exercise 10 after
having carefully studied the detailed discussion in Section 7. l of perturbation
theory for the simple pendulum.
GENERALIZATION OF THE POINCARE METHOD
The solution of specific problems is only a prelude to understanding-a
means to an end. In particular, the lessons learned from attacking a specific
problem can often be formulated in terms of general principles and developed
into general theory in order to deepen our understanding. But this cannot
always be done in a formal manner. The Poincare theory is a case in point.
The success of the parametric representation introduced by Poincare has
led to its application to other types of ordinary differential equations. for
the removal of a variety of difficulties of a similar nature. Basically, there is
a need to shift the solution surface by a mapping or distortion in the domain
of the independent variables. But it is not fruitful to try to classify all
the problems that can be solved by this approach.
We have not mentioned any theorems that rigorously justify the use of
the Poincare method. Such theorems are, in fact, available in a number of
Sec. 2.2) Elements of Perturbation Theory 55
instances. Nevertheless, many situations in which the value of the Poincare

approach can scarcely be doubted are not subsumed under present rigorous
treatments. One reason why final formalization of valuable ideas in the form
of theorems is often impracticable is that theorems of desirable generality
are very difficult to prove. A more subtle reason is that a listing of definitive
theorems seems to close a subject, whereas an applied mathematician is
always struggling to open new areas of application by suitable modification
and generalization of known results.
EXERCISES
1. Demonstrate the validity of (7).
2. Fill in the details of the solution to Example 1.
3. Prove the validity of (19). Begin by multiplying (14) by E>.
4. Demonstrate the correctness of the statement under (20) concerning
the error in period. Assume that n = 0.
5. Consider the time interval in which a 2 0<2 >,as given in (33), is a small
correction to e<0 ). Show that this interval is essentially the same as
the number of periods N mentioned under (20).
6. Carry the Poincare method through the derivation of (36). [The principal
result (37) then follows at once.]
7. Show that the more general assumption (34b) reduces to (35) when the
Poincare method is carried through.
8. A grandfather's clock is to be designed. It is to be 5 feet long and to have
a maximum swing of t foot at the tip of the pendulum. Show that the
linear approximation for the period T = 2n(L/g) 112 is not good enough
[H. 0. Pollack, Educ. Stud. Math. 2, 393-404 (1969)]. To do this,
derive and use the inequality
(ix
2
- 0 2 )ix- 1
sin IX~ 2 cos 0 - 2 cos ix~ (ix
2
- 0 2 )(1 - t sin 2 i)
and apply it to a form of the integral giving the exact period. Obtain
112
1 - cos ix
---<---< T - T0 (
-- ix )
- l
12 - T0 - sin IX '
where Tis the exact period and ocis the maximum angular displacement.
Hint: Use the fact that (sin x)/x has a local maximum at x = 0 to show
that
e I"
sin x dx ~ sin
-- ix x dx,
IX e
J"
Also show that
1- cos x ~ 2(;)sin 2
2
~.
Integrate from O to x and then from 0 to IX.

9. Show that a first integral of

d2(J
dt 2 + wisin O = O
is
1(d9) 2
2
2 dt = w (cos 9 - cos a)=
0 2w 20 ( sm
2-
2 a . 2
sin 9)
2,
where a is the maximum amplitude. Verify that a further transformation
(J a . .,,
sm-=sm-sm.,,
2 2
brings the solution of the problem into the form
t = --1 .
Wo
i"'
Il/2
(1 - k2sin21/1)-112di/;;
(8(0) = a)
where k 2 = sin 2 (a/2). Obtain the period of oscillation as a power series
in k 2
10. An experimental test of the general theory of relativity is the advance
of the perihelion of Mercury. According to this theory, the equation
of the orbit is (1.12):
d 2u
d<J,2+ u = a(l + eu2 ),
This differs from the classical form by the term eu2 In the Newtonian
case (e = 0), the solution is
u = u0 = a[l + e cos (q,- q,0 )],
where e is the eccentricity and </,0 determines the position of the peri-
helion. Show that the angle between two succeeding perihelions is
21ta2e to a first approximation. One method is the following.
If the solution of the relativistic equation is given in the form
+
u = a0 + a 1 cos p<J,+ a2 cos 2p<J>
and we assume
p = 1 + ep<t>+ e2p<2>+ ...
a1c= a1>+ ea11 >+ e2 aP> + ,
then
Sec. 2.3] A System of Ordinary Differential Equations
57
2.3 A System of Ordinary Differential Equations

The purpose of this section is to demonstrate the fourth part of applied
mathematics that we listed in Chapter I : how the desire to solve scientific
problems motivates the development of a mathematical theory, and the
manner in which such theories aid us in the solution of scientific problems.
We shall begin the discussion with a statement of certain theorems related
to the initial value problem for mechanics. The proofs will follow. (As far
as it is feasible, the proofs presented here wili be those actually useful for
constructing the solutions.) At the end of the section we shall briefly discuss
the general question of the value of such proofs to the applied mathematician.
THE INITIAL VALUE PROBLEM: STATEMENT OF THEOREMS

For a given (initial) point P0 (-r, ( 1 , ( 2 , .. , Cn)in the (n + !)-dimensional
space of the points P(t, z1 , z2 , , zn),the initial value problem for a system
of ordinary differential equations
dzk
dt = Ji,,(t;Z1, Z2'. .' Zn); k = I, 2, ... ,n (I)
is to find functions
Zm= Bm(t), m = l, 2, ... , n (2)
such that (1) is satisfied, and
Bm(t) = Cmat t = 'r, m = 1,2, ... ,n. (3)
We now state and subsequently prove certain theorems that are valid
for the initial value problem. In the statement of these theorems and in
subsequent discussions we shall frequently use z and { to denote collectively
the set of variables (z1 , z 2 , , zn) and ({1 , { 2 , . , Cn).
Theorem 1 (Existence). Suppose that the functions* fit, z) are continuous

in a rectangular parallelepiped defined by
R: it--rl ~a. lzk-Ckl ::s:;b; k= I,2, ... ,n. (4)
This implies that there exists an upper bound M such that
IJi,,I ::s:;M, k = 1, 2, ... , n, (5)
for P(t, z1 , z2 , , Zn) in R. Suppose further that each!i,,satisfies the following
Lipschitz condition in R:
lf(t, Z1, ,Zn) - f(t, Zi, Z2, ,Zn) I ::s:;K[J Z1 - Z1 I + + IZn - Zn I]. (6)
* Some mathematicians are always careful to distinguish the function f from f(x), the value
of this function at x. We do not find it profitable to emphasize this distinction.
(The Lipschitz condition is implied if each of the functions fk has continuous

partial derivatives.) Then there exists a solution of the initial value problem
in an interval It - -rI ~ oc,where
oc= min ( a, !). (7)
The solution functions [the functions Umof (2)] have continuous first deriva-
tives.
Theorem 2 (Uniqueness). The solution of the initial value problem is

unique.
The significance of Theorem 2 can be appreciated only when we can pro-

duce an example in which there is more than one set of functions satisfying
the same equation and initial conditions, when (of course) the conditions
in Theorem 1 are not all fulfilled. This will be done below.
Next we generalize the initial value problem to
dzk
dt = f;.(t, z, A.), (8)
Now the right-hand side of the equation depends on the parameter A..
Theorem 3 (Continuous dependence on parameters). Let the functions

fit, z, A) (k = I, 2, ... , n) satisfy the requirements prescribed in Theorem I.
Further let these functions depend continuously on the parameter A.in a
certain neighborhood IA - A.0 I < c, where i 0 and c are constants. Then the
solution functions are also continuous functions of A in some neighborhood
of 10
Theorem 3 can be used to prove that the solutions are continuous functions
of the initial values (-r, (). This continuous dependence can be regarded as
demonstrating the stability of the solution with respect to changes in initial
values, in the sense of the following theorem.
Theorem 3' (Stability). Let Zm= 9m(t; (m) denote the mth solution func-
tion (2) with the dependence on the initial condition explicitly noted. Then
given any e > 0 and fixed time T, T ~ -r, there exists a r, = f>(e,T) such that
lum(t;(m)- Um(t;(~) I < 6 for 't' ~ t ~ Twhenever 1,~- 'm I ~ f>.
Theorem 3 is an expression of our intuitive feeling that when the functions
fit, z, A) are changedslightly (through the parameter l), the solutions should
also be changed slightly. Our corresponding expectation for the initial
conditions is expressed in Theorem 3'. Let us now look more carefully at
the nature of the dependence on a parameter.
Sec. 2.3] A System of Ordinary Differential Equations 59
We shall take the initial conditions to be fixed and consider the dependence
of the solutions zm = 9m(t, .il) on .il. One might expect the nature of this
dependence to be the same as that of the functions f,lt, z, .il)in (8). Thus, if
the Ji are analytic in .il in the neighborhood of .il= .il0 (i.e., if they have a
convergent Taylor series when O s: I.il- .il0 I < r for some r ), then the solution
functions 9m should have the same property. For linear differential equations,
this is generally true. That is, the solution of the equation*
dz
dt = A(t, .il)z+ b(t, ,l) (9)
is analytic in .il,provided that A(t, .il)and b(t, .il)are analytic in .il,even though
they may merely be continuous in t. (Note that the Lipschitz condition is
automatically satisfied for linear equations when the coefficients are con-
tinuous.) In the nonlinear case, the situation is a little more complicated,
as we can see from our experience with the development of a perturbation
solution for (2.1),
dy
dx = f(x, y, e).
In present terms, the formalt calculations there depended on the fact that
the functions fi(t, z, .il)were analytic in z as well as in .il. To keep the dis-
cussion relatively simple, we shall therefore restrict our attention to the
first derivative with respect to .il.
If formal differentiation is justified, we would expect the functions
(IO)
to satisfy the differential equation obtainable formally from the differentiation

with respect to .il of the original equation (8), remembering that Ji and zm
both depend on .il. Carrying out this differentiation, we obtain
duk = L ofk Um + ofi. (11)

dt m OZm o.il
We note that (11) is a linear system in {uk} whose coefficient functions are
known functions oft, since the set {zm(t, A)}is known. If the initial values ,m
are independent of the parameter .il,then the initial conditions on the {uk}are
(12)
* This may be regarded as a single equation or a system of linear equations. In the latter case,
A is a matrix, and z and b are vectors.
t A fonnal calculation is one that is presumably valid under suitable but unspecified conditions.
Successive parametric differentiation of an equation forms the basis of an excellent way to
perform perturbation calculations. See Section 7.2.
6o Deterministic Systems and Ordinary Differential Equations [Ch. 2
In view of the above discussions, we expect to have

Theorem 4. If the n(n + 1) functions
of,. of,.
OZm' OA;
k, m = 1, 2, ... ,n;
are continuous in the variables t and {zm}and in the parameter ;.., then the
solution functions zm(t, )..) of Theorem 3 are differentiable with respect to ;..,
and the partial derivatives (10) satisfy the differential equations (11) and the
initial conditions (12).
Since generalization is almost immediate, the proofs of all the theorems

stated above will now be presented in the case of a single dependent variable.
PROOF OF THE UNIQUENESS THEOREM
The uniqueness theorem is usually the easiest to prove. In the present
case, the ideas used will also be found to be useful for the proof of the
existence theorem.
We consider the equation
dy
dx = f(x,y) (13)
and seek a solution satisfying the initial condition y = y 0 when x = x 0 Let

g(x) be one such solution. Then
g'(x) = f[x, g(x)], (14)
r
and hence
g(x) = Yo + f[t, g(t)] dt. (15)
.xo
If there were another solution G(x) satisfying the same initial conditions,
then
G(x) = Yo + r
.xo
f[t, G(t)] dt. (16)
If we now subtract (15) from (16), we obtain
G(x) - g(x) = [ [f(t, G) - f(t, g)] dt. (17)

.xo
The magnitude of the integral on the right-hand side can be appraised with
the help of the Lipschitz condition:
If(t, G) - f(t, g) I ~ Kl G(t) - g(t) 1. (18)
~Kr
Thus
IG(x)-g(x)I IG(t)-g(t)I dt. (19)
.xo
Sec. 2.3) A System of Ordinary Differential Equations 61
In the interval [x 0 , x], IG(x) - g(x) I has a maximum value that we shall
denote by IIG(x) - g(x)II. It then follows from (19) that
IG(x)-g(x)I SIIG(x)-g(x)ll Klx-x 0 1, (20)
and hence
IIG(x)- g(x)II s IIG(x)- g(x)II Klx -xol, (21)
or
[I - K(x - x 0 )] IIG(x) - g(x)II S 0.
If we now take the length Ix - x0 I sufficiently small, equal to (2K)- 1 for
example, we can make the first factor positive. Hence (21) requires that
UG(x)- g(x)II = o,
from which the desired result follows at once. O
PROOF OF THE EXISTENCE THEOREM
We have seen that the solution of the initial value problem for the differential
equation (13) satisfies the integral equation (15). Conversely, if we have a
solution of the integral equation (l 5), we have a solution for the differential
equation, with the initial co:editions implied. This can be verified by direct
calculation (Exercise I). Thus it is sufficient to prove the existence theorem
for the integral equation.
To prove that (15) has a solution, we adopt the method of successive
approximations. We start with a crude approximation y = y 0 , which at least
satisfies the initial condition.* We then calculate the sequence of functions
Y1(x) =Yo+ [ f(t, Yo) dt,

XO
Y2(x) =Yo+ [ f[t, Yi(t)] dt,

XO
(22)
Yn(x) = Yo + [ f[t, Yn-1(1)]dt,

XO
We shall prove that this sequence converges uniformly to a continuous

function in a suitably restricted interval Ix- x0 1sex, which may be
specified as in Theorem 1; i.e., ex= min (a, b/M). [Here Mis the upper bound
on I/I that was introduced in (5).] The restriction to the range b/M is needed
It is certainly not necessary to make this particular initial approximation. For example,
the same final result will obviously be obtained if y 1(x), defined in (22), is used as the initial
approximation.
only because we wish to keep the successive approximations (22) within the
bound IY - Yol < b, wherein the continuity off is assured. Otherwise,
there is no further restriction on the range of x.
To prove that the sequence of functions defined by (22) uniformly
approaches a limit in the interval Ix - x 0 I ::;;;a, let us consider the differences
between successivefunctions. For Yi(x) - Yo we have
r
IY1(x)- Yol ::,;; lf(t,Yo)I dt::,;;Mix - Xol,
Xo
(23)
where Mis the upper bound on/first mentioned in (5).

In general, we have
lyn(x) - Yn-1(x)I :s; r

xo
lf[t, Yn-1(t)] - f[t, Yn-2(t)] I dt. (24)
We need to relate the difference in the right-hand side to the difference

Yn-i(t)-Yn-i<t). This relation is supplied by the Lipschitz condition:
lf[t, Yn-1(t)] - f[t, Yn-iCt)]I :s;KIYn-1(t) - y,._i(t)I. (25)
We are thus able to establish the recurrence relation
jyn(x) - Yn-1(x)I :s;Kr IYn-1U) -Yn-i<t)i dt. (26)

xo
By combining (23) and (26), we obtain [Exercise 3(c)]

MKn-llx-xoln
lyn(x) - Yn-1(x)I :s; (27)
n.1
But the right-hand side of (27) is M/K times the nth term in the series for
exp (Kl x - x 0 I).Thus let us adopt the device ofintroducing a series expression
for the limit as n-+ ao of yn(x), namely,
co
Yo+ L [y,(x) - Yr-1(x)].
r= 1
(28)
This series is absolutely and uniformly convergent, since its terms are bounded
by the corresponding terms of another series with that property. We thus
have limn-+co yn(x) = y(x) for some function y. Further, y(x) can be shown to
satisfy the integral equation (15) by examining the limiting form of the
sequence of equations (22). Here it is only necessary to justify the inversion
of the processes of integration and taking the limit; but since the integral
is over a finite range, this follows from well-known theorems. D
The key process in the proof is the conversion of the differential equation
into an integral equation and then the application of the method of successive
approximations. The integral formulation has two advantages: (i) the initial
values are automatically incorporated; and (ii) the integration process makes
the functions smoother, and questions about the existence of a derivative

are avoided.
The proof required many technical steps. Since similar steps often occur
in other problems, it is convenient to summarize them in some form so that
they need not be continually repeated. This is one reason for developing
the theory of function spaces, in which an important concept is the norm
suitably defined (in various ways) to replace the concept of distancein ordinary
Euclidean space. Another reason is that ordinary geometric ideas can be
used to give intuitions about abstract distance properties.
In the present case, one possible norm is the quantity llg(x) - G(x)II
first used in (20). Another possible measure of "distance" is the quantity
D defined by
I
D2 = b - a f [G(x) - g(x)]2 dx.
b
a
(29)
In words, D is the root-mean-squared difference between G and g. This

norm does not occur naturally here, but it will be used in Chapters 4 and 5.
Appendix 12.1 of II provides certain details concerning function spaces
and related concepts. [In particular, Equation (59) of that Appendix gives
an abstract characterization of distance that is exemplified by both of the
norms just defined.] Material in Chapter 12 of II on variational methods
furnishes examples of the unity and clarity that can be achieved with such
concepts.
CONTINUOUS DEPENDENCE ON A PARAMETER OR INITIAL
CONDITIONS
Before we proceed to prove Theorems 3 and 3', let us note that if the solution
of the differential equation
dy
dx = f(x, y, A), (30)
under the initial condition y = Yo for x = x 0 , depends continuously on A,

then it depends continuously on (x 0 , y 0 ). For we can introduce the new
e,
variables ( ,r) = (x - x 0 , y - y 0 ) and solve the equation
dr,
ae= f (e+ Xo ' 1'/+ Yo ' A) (31)
under the initial condition r, = 0 for e= 0. The function on the right-hand

side of (31) is continuous in (x 0 , Yo)
Return now to (30). Let us treat the equation in the integral formulation
y(x, A)= Yo + r XO
f[x, y(x, A), A] dx. (32)
Consider two solutions y(x, A) and y(x, A0 ) and their difference. We have
y(x, A) - y(x, A0 ) = r
.xo
{f[x, y(x, A), A]- f[x, y(x, A0 ), A]}dx
+ r
.xo
{f[x, y(x, Ao),A] - f[x, y(x, A0 ), A0 ]} dx. (33)
The first integral on the right-hand side can be appraised with the help of
the Lipschitz condition and the second integral by simple continuity in A.
Thus, by using an argument similar to that used in the proof of the unique-
ness theorem, we have [Exercise 3(a)]
lly(x,A) - y(x, Ao)II:-,;;Klx - Xo I lly(x,A) - y(x, Ao)II+ <>Ix- Xo I,
(34)
where
If(x, y(x, A0 ), A) - f(x, y(x, A0), Ao)
I < <>. (35)
For Ix - x 0 I :-,;; 1
(2K)- one can show [Exercise 3(b)] that (34) implies that
,
jy(x, l) - y(x, A0)1::;;2<>lx- x 0 j. (36)
The quantity <> can be made as small as desired by decreasing IA - AoI

The restriction on Ix - x0 I is not crucial, since one can cover any finite
interval Ix - x 0 I < exby a finite number of intervals of the above type.
Upon passing from one interval to another, one would have to allow for a
difference in the initial condition y 0 , but this difference is of the order of<>.
Since we have only a finite number of such intervals, the cumulative difference
is still of the order of E,and does not influence the essentials of our arguments.
0
DIFFERENTIABILITY
We consider the pair of differential equations
dy
dx = f(x, y, A) (37)
and
du
dx = f.,(x, y, l)u +f;.(x, y, A), (38)
where f(x, y, A), f.,(x, y, .il.),and f;.(x, y, A) are continuous functions. Note
that (38) is obtained by formally differentiating (37) with respect to l and
writing u for oy/ol.
We consider a solution y(x, l) of (37). We also consider, for y = y(x, A), a
solution of (38) satisfying the initial condition u(x 0 , A) = 0. The existence of
these solutions is assured by the continuity of the functions involved.* We

wish to prove that
u(x, ..t)= ay(x, l). (39)
a1
The difficult part of the proof is to establish the existence of the partial
derivative ayJa..t.This will be done by direct evaluation of the limit of
liy/lll as Ill-+ 0.
Let
liy = y(x, l) - y(x, lo), !if= f[x, y(x, 10 ), l] - f[x, y(x, 10 ), 10 ].
From (33)
11y
Ill=
I"f.,(x, y,_l) liyIll dx + I"111
lil dx, (40)
XO XO
where y is a value intermediate between y(x, l 0 ) and y(x, A). We have used
the continuity of the partial derivative f,(x, y, A) so that the mean value
theorem can be applied. We also have
ZO
A.a)dx +
u(x, Ao)= [ f,[x, y(x, lo), .A.o]u(x, r %0
fi[X, y(x, lo), Ao]dx. (41)
By combining (40) and (41), we obtain for the difference

_ ( 2 ) y(x, l) - y(x, Ao)
w = u x, "1l - l _ Ao (42)
the integral equation
w= [ f,(x,y,l)wdx+ D, (43)
XO
where
D r
= xo [f,(x, y(x, A0 }, A)- f,(x, y, A0 )]u dx r
+ [f,.(x, y(x, A0 ),
zo
l 0) - !r~J
dx.
Becauseof the continuity of the functions!, and fi, the quantity D can be
made as small as we wish by reducing Ill. Since IJ,(x,y, l)I is bounded by
K, one can prove, from (43), that
lim w =0 (44)
by the same sort of reasoning as that previously used. The reader should
complete the proof (Exercise 4).
The Lipschitz condition is satisfied for (38), because the right-hand side is linear in the
dependent variable u.

EXAMPLE OF NONUNIQUENESS
One can better appreciate the usefulness of the uniqueness theorem under
the Lipschitz condition when one sees some examples of the lack of unique-
ness when the condition is not satisfied. Consider the differential equation
dy
dx =f(x,y), (45)
where
4x 3y
f(x,y) = x4 + y2 when (x, y) #- (0, 0), (46a)
f(O, 0) = 0. (46b)
It is easy to verify that f(x, y) is continuous at (x, y) = (0, 0) but does not
satisfy the Lipschitz condition. Equation (45) admits the solution
Y = c2 -Jx4 + c4 (47)
for all finite real values of c. Thus there is an infinity of solutions satisfying
the initial condition (x, y) = (0, 0). All the integral curves have zero slope
at the origin.
The lack of uniqueness can easily be understood from a heuristic point
of view, since (45) does not give a unique value for the curvature of the
integral curve at (x, y) = (0, 0). Indeed, the curvature of the solution curve
at this point is undefined. For the derivative of the slope of the solution curve
is given by the limit as (x, y)--+ (0, 0) of
dy dy
dx (x, y) - dx (O, O) f(x, y) 4x 2y
= --x- = x4 + Y2.
x-0
An existence theorem can still be proved in this case by the method of
finite differences, to be outlined in a moment. Since there is lack of uniqueness,
an initial value for d 2y/dx 2 must be prescribed (implicitly or explicitly) for
each solution.
METHOD OF FINITE DIFFERENCES
The method of finite differences is the natural approach to the solution of
differential equations from the point of view of numerical integration. It
yields at the same time a practical way of calculating the solutions, especially
with the aid of modern computing machines, and a way for proving an
existence theorem. To obtain a glimpse of this method, we may start again
with the integral formulation
Y =Yo+ rXO
f[e, y(e)] de, (48)
Let us divide the interval (x 0 , x) into n subintervals (each of length h) and

write
y = Yo + f ![,,
Xo
1 y(e)] de + + r
Xn-1
![,, y(rn dl;. (49)
This is still exact, but we shall now evaluate each of the integrals in (49)
approximately by writing
r "
+
1
f[,, y( e)] de = hf (xk, Yk). (50)
Here Yk is the value of y obtained by using the approximation (50) in (49)

up to the point xk; i.e., we have
Y1= Yo + f' xo
f(xo, Yo) de,
Y2 = Y1+ f /(x1, Y1) 2 de, (51)

x,
y(x) = Yn = Yn-1 + r
.Xn-1
f(xn-1, Yn-1) dl;.
The solution is then the broken line CL in Figure 2.2. This is taken as an
approximation to the true integral curve C. Since there is an error in y at
each approximation, the accumulation of errors would, in general, make the
approximating curve CL deviate farther and farther from the true curve C
y
Yo
FIG URE 2.2. The simplest finite difference method provides the broken
line CL as an approximation to the actual solution C of a differential
equation.
as the calculation process goes on according to (51). One might think at

first sight that this would be intolerable. Actually, since the error in y is of
the order of h 2 in each step [local formula error is O(h2 )], it can be shown that
the cumulative error is of the order of h [cumulative formula error is O(h)].
Indeed, one can prove the convergence of the approximate solution to a
unique true solution when the Lipschitz condition is satisfied. Even when the
Lipschitz condition is not satisfied, a partial sequence can always be chosen
to converge to a solution; but there is no longer any assurance of uniqueness.
An O(h) cumulative formula error can be made as small as desired by taking
h small enough. But then many calculations will be required to find the
solution in a given x interval, and the cumulative roundoff error may be
unacceptably large. To gain better accuracy, one can adopt higher order
approximations when evaluating the integrals in (49). But then, calculations
for a given interval are longer, so that they take more time and possibly
introduce more roundoff error. The reader will find extensive discussions of
such issues in books on numerical methods.
FURTHER REMARKS ON THE RELATION BETWEEN "PURE"

AND ''APPLIED'' MATHEMATICS
In this section we have gone into mathematical theory that generally would
be regarded as being of a "pure" nature. We conclude with some remarks
on the interaction between the pure and the applied aspects of mathematics.
(Also see Chapter 1.)
That most pure mathematicians would benefit from studying works in
applied mathematics seems fairly clear, for in their quest for scientific
knowledge, applied mathematicians (and other theoretical scientists) leave
unanswered many questions of a mathematical character. An applied mathe-
matician may, for example, construct a formal perturbation method which
provides a solution to a problem that agrees very closely with experiment.
The pure mathematician will often find it worthwhile to determine conditions
under which these calculations are guaranteed to be valid. Such justifications
may require proofs that certain general theorems apply in a particular case.
Or they may best be accomplished by expanding and recasting a whole seg-
ment of mathematics. An example of such a recasting is Laurent Schwarz's
relatively recent generalization of the concept of function to that of distribu-
tion, so that Dirac's virtuoso child, the delta function, could at last be
legitimatized.* Of course, much of pure mathematics is internally motivated
-but it seems unwise to neglect a rich source of meaningful problems.
What help is pure mathematics to an applied mathematician? Certainly,
it is a contribution to science to have a body of theory incontrovertibly
A good reference is M. J. Lighthill's Fourier Analysis and Generalized Functions (New

York: Cambridge U.P., 1962). The dedication of this work is a capsule history: "To Paul Dirac
who saw that it must be true, Laurent Schwartz who proved it, and George Temple who showed
how simple it could be made."

established. The magnitude of the contribution is an increasing function of
existing doubt in the theory. Indeed, if the doubt is serious, the applied
mathematician himself will try to resolve the controversy by fashioning
appropriate proofs.
At the educational level, the future applied mathematician should be
exposed to a considerable body of mathematical theory and proof (although
care must be taken not to overemphasize this aspect). He should learn typical
conditions under which the operations he may have to perform are valid.
Moreover, he should become aware of additional mathematical concepts
that may someday form a suitable framework for his theories or calculations.
For example, some applied mathematicians find use for the unifying geometric
character of the function space concepts, developed in recent decades. To
give a more classical example, the Gibbs phenomenon, which can arise in
practical calculations, cannot be appreciated without a knowledge of the
distinction between pointwise and uniform convergence. (See Section 4.3.)
And, to cite another possibility, it could be that a student will first become
aware of the calculational usefulness of an integral equation formulation, or
of a successiveapproximations approach, by virtue of having studied certain.
constructive existence theorems.
Examples of the relation between pure and applied mathematics are
provided by our discussion of theorems for differential equations. Thus it
is a beautiful theoretical result that under suitable conditions solutions
depend continuously on parameters. But this result misses a very important
scientificissue, for it merely shows that the solution is changed by an arbitrarily
small amount over a fixed time interval by a sufficiently small change in the
parameter. Left untouched is the question of the ultimate effect of a given
small change in a parameter, a question that could well be missed entirely
by one who accepts an impressive theorem as the last word. Poincare began
the study of long term and "ultimate" effects. Much more has since been
accomplished, both formally and rigorously, but the issues are by no means
resolved. [See Moser (1973).]
Another example of the relation between pure and applied mathematics
stems from the fact that standard existence and analyticity theorems for
systems of ordinary differential equations are valid only for a sufficiently
short interval of time. But if these equations describe the trajectories of
interacting particles, it seems likely that the theorems should usually hold
for arbitrarily long time intervals. For a discussion of classical results on
such matters see Chapter 16 of E. T. Whittaker's Analytical Dynamics
(New York: Cambridge U.P., 1927).
To achieve a degree of balance, we presented a few formal proofs in this
chapter. But such proofs can be found in many fine books. Hence proofs are
mainly omitted in the remainder of the present work, since we concentrate
on the relatively unexplored interaction between science and mathematics
that is the core of an applied mathematician's profession.
EXERCISES
1. Show that a solution of (15) satisfies (13) and the initial condition
y(x 0 ) =Yo. Carefully justify all your steps.
2. Prove that y(x), the sum of the series (28), satisfies the integral equation
(15).
3. (a) Verify (34).
(b) Verify (36).
(c) Verify (27).
4. Complete the proof of (39).
5. (a) Verify the statement under (46b).
(b) Verify (47).
(c) Feigning ignorance of (47), see if you can derive it.
CHAPTER 3
Random Processes and Partial

Differential Equations
IN CHAPTER 2 we considered the dynamics of individual particles, and

we showed that interaction of the particles can be modeled by a system of
ordinary differential equations. Existence of a solution to these equations
under prescribed conditions (usually specified values of the initial position
and velocity of each particle) is assured by a classical mathematical theory.
Actual determination of an adequate approximation to the solution can be
carried out, if necessary, by numerical integration. Study of the algorithms
used in numerical analysis is necessary to provide assurance that a given
approximation is accurate. Nevertheless, mathematical theory assures us
of the unique existence of the solution. The future of the particles is completely
determined by their past, so our model is deterministic.
We shall now study examples of models in which there is uncertainty in
the predicted outcome of a process. In such situations we shall refer to
random processes; the corresponding models will be termed probabilistic.
For example, we shall see that when many particles interact, it is often useful
to adopt a model in which each particle's motion is described by certain
probabilities.
Although we shall give some consideration to random processes for their
own sake, our primary goal is to investigate the connection between proba-
bilistic and deterministic models of the same phenomenon. The key fact
here is that even though a process is regarded as random, nevertheless in
many circumstances we can write down deterministic problems in partial
differential equations that yield a definite distribution function for the
probabilities associated with the process. The partial differential equation
for the distribution function is identical with the equation one would write
down in a deterministic characterization of the phenomenon.
It might appear paradoxical that a random process can be characterized
by a definite equation. But we know from experience that when individually
unpredictable events are repeated a large number of times, there usually
emerges a definite regularity. For example, when a "fair" coin is tossed
once, we have no idea whether head or tail will turn up. But after a large
number of tosses we can be almost certain that the proportion of heads will
differ only a little from t. * With this example and others like it in mind, it is
perhaps not so surprising that there is a determinable distribution of prob-
abilities which characterizes a random process.
Precise statements of this nature fall under the heading "laws of large numbers."
71
72 Random Processes and Partial Differential Equations [Ch. 3
The characterization of situations in which probabilistic models are appro-

priate is rather a subtle matter, but typically there is an element of incom-
pleteness of information. Let us once again consider the example of coin
tossing. Once a coin is tossed, its motion is determined. The process of tossing,
however, involves many parameters, such as the coin's initial orientation,
velocity, and spin; the properties of the surface on which it bounces; etc.
Usually, we have no precise knowledge of these parameters.
An even more important characteristic of situations where probabilistic
models are useful is illustrated by the fact that a slight change of the para-
meters could shift the final configuration of a coin from "head." to "tail,"
or vice versa. Another familiar example of this extreme sensitivity is found in
the kinetic theory of gases. Here each molecule can be regarded as following
exact mechanical laws of motion and exact laws of interaction (if the hard
sphere model is adopted, for instance). But a slight change in the initial con-
ditions would probably result in a tremendous change in the condition after
many collisions.
An averaging of the solutions of a number of problems with varying initial
conditions and a consequent blurring of the sensitivity of a single problem
can be regarded as underlying the successful modeling of sensitive systems
by random processes.
In Section 3.1 we formulate a model in which Brownian motion is regarded
as being caused by an assemblage of particles that move randomly on a line,
taking steps of equal length with equal probability of moving right and left.
An explicit expression is obtained for the probability w(m, N) that such a
particle has moved m steps to the right after taking a total of N steps. The
section concludes with a sketch of how theory enables one to determine
molecular properties from observations of Brownian motion.
The important concept of an asymptotic approximation is developed in
Section 3.2. This is necessary in order to understand the sense in which a
certain simplified expression approximates w(m, N) when N is large, but
the methods introduced in Section 3.2 are of far wider applicability.
In Section 3.3 we introduce a difference equation for w(m, N) and then
approximate this difference equation by a limiting differential equation.
This equation is none other than the diffusion equation commonly used in
macroscopic descriptions of the motion of many particles. We show that the
"unit source" or ''fundamental" solution of the diffusion equation provides
the same approximation to w as that mentioned in the previous paragraph.
The usefulness of the differential equation approach is illustrated by con-
sideration of random walk in the presence of reflecting or absorbing barriers.
Further explorations of the connection between probability and differ-
ential equations are made in Section 3.4. For example, it is shown how the
method of images, used in differential equations, leads to the solutions of
some probability problems. The fundamental solution, first met in a prob-
ability context, is shown to provide the key to a solution of the general
Sec. 3.1] Random Walk in One Dimension; Langevin's Equation 73
initial value problem for the diffusion equation. Section 3.4 concludes with
a discussion that links (among other things) the collision of DNA molecules,
the twisting of a beam, the probability that a randomly walking drunk will
some day return to his starting point, and the irreversible character of an
assemblage of particles in quasi-periodic motion.
3.1 Random Walk in One Dimension; Langevin's Equation

In Brownian motion, small particles move about irregularly in a liquid or
gas. Perhaps the most common manifestation of this phenomenon is the
dancing of dust particles in a beam of sunlight.
The relatively large dust particles move because of myriads of impacts
by the molecules that comprise the surrounding medium. In making a
mathematical model of this situation, we should without doubt regard the
impacts as occurring randomly. There is no possibility of computing the tra-
jectory of each molecule, and no interest in doing so, if one wishes to under-
stand the broad outlines of the phenomenon.
THE ONE-DIMENSIONAL RANDOM WALK MODEL
Let us try to construct the simplest possible model that bears on the
situation. Thus, although the motion is three-dimensional, let us consider a
projection of the particle, which moves randomly along a line. A reasonable
idealization of the motion consists of a sequence of steps, with randomly
determined lengths and durations. We shall simplify the picture still further
by considering steps of a fixed length Ax that occur at a fixed time interval
!it. We are thus led to the classical one-dimensional model of Brownian
motion, in which a particle moves according to the following rules:
(i) During the passage of a certain fixed interval of time !it, a particle
takes one step of a certain fixed length !ix along the x axis.
(ii) It is equally probable that the step is to the right or to the left.
We shall study the characteristics of a particle that executes the unbiased*
random walk described by (i) and (ii). To do this we need only the rudiments
of probability theory since a very simple type of probability problem is involved.
In a given situation a certain set of possible events is selected for attention.
These events are often labeled "favorable." Because all events are regarded
as "equally likely," the probability of a selected "favorable" event is given
by the ratio
number of favorable events
total number of events
We wish to determine the probability w(m, N) that a particle will arrive at
a point m steps to the right of its starting point after taking N steps. (Points
The walk is biased unless steps in an permitted directions are equally probable.
to the left are associated with a negative integer m.) Every possible N-step
walk is equally likely, and the "favorable" event is arriving exactly m steps
to the right in such a walk.
In this section we shall derive an explicit formula for w(m, N). We shall
then introduce the device of a generating function to compute the average
and root-mean-squared values ofm.
The concept of" random walk" is colorful and it is particularly interesting
when regarded as describing the fate of a drunk staggering down a street,
with equal probabilities of stepping forward or backward. As one might
expect, random walks can be found concealed in a wide variety of situations.
To give just one further example, consider a gambling game in which a
"fair" coin is tossed and player A gives a dollar to player B, or takes a dollar
from him, depending on whether the coin is heads or tails, respectively.
If a particle is stepped to the right (left) when A wins (loses) a dollar, then
w(m, N) provides the probability that A has won m dollars after N tosses.
The reader will find it worthwhile to keep this additional example in mind
as he reads this chapter.*
The most important physical implication of the study of Brownian motion
resides in the fact that such motion provides a visible macroscopic mani-
festation of individual molecular forces. (By contrast, only the bulk effects
of molecular motion are manifested by observations of such variables as
temperature and pressure.) The most direct theoretical approach to these
matters involves a differential equation with a random forcing term, the
Langevin equation. We sketch the essence of the relevant theory at the end
of this section.
EXPLICIT SOLUTION
We now work out an explicit formula for w(m, N), the probability that a
particle will arrive at a point m steps to the right of its starting point after
taking N steps. Suppose that the particle takes a total of p steps to the right,
p;?:: 0, and hence a total of N - p steps to the left. It would then reach a
point m steps to the right ( - N :::;;;; N), provided that
m :::;;;;
m=p-(N-p); i.e.,p=t(N+m). (l)
Note that m is a signed integer, but p is a nonnegative integer. Also m =
2p - N is even (odd) when N is even (odd). Thus m ranges from - N to N
in steps of 2. For example, if a total of N = 3 steps are taken, the possible
values of displacement are m = - 3, - l, 1, 3.
The number of paths with p steps to the right and N - p to the left is
given by
N! =CN (2)
p!(N-p)!- p
For other examples and a good introduction to the literature, see M. Barber and 8. Ninham,
Random and Restricted Walks (New York: Gordon and Breach, 1970).
Sec. 3.1] Random Walk in One Dimension; Langevin's Equation 75
Recall that c;
is called the binomial coefficient, because of its role in the
binomial expansion
N
(x + Yl = I c;xN-PyP. (3)
p=O
The total number of different N-step paths is 2N. Dividing by the number
of favorable events, we obtain the result
CN
w(m, N) = 2!, where p = ,t(N + m). (4)
The number desired for (2) is the same as the number of full box-empty box
patterns formed by p balls in N boxes, one ball per box. (A ball in the kth box is
analogous to the kth step being to the right.) We start by considering p distinguish-
able balls. The first ball, pi, can be placed in any of N boxes; the second, P2, in
any of N - 1 boxes; etc. Thus p distinguishable balls can be placed in N boxes in
the following number of ways:
N!
N(N - 1) [N - (p - I)] = . (5)
(N-p)!
Interchanging distinguishable balls permutes the order in which they appear but
does not change the pattern of full and empty boxes. There are p ! permutations of
p balls. Summing up:
number of
number of ways p .
. h bl b number of full permutations of
d tsti~gm~ a d ~ ~IIs = box-empty box distinguishable
cban e P ace m patterns balls within a
oxes
pattern
or
N!
----CNp'
(N-p)!- .,
yielding (2).
Let us check that the sum of all probabilities is unity. Using (3), we find
that indeed
N N N
I"
m= -N
w(m, N) =
p=O
L c:ut = L c:ut-pHY
= (t + tl
p=O
= 1. (6)
o TE . The double prime on the sum is used to indicate that the summa-
N
tion index takes every other value.
MEAN, VARIANCE, AND THE GENERATING FUNCTION

Suppose that f is a function whose domain is the possible outcomes m
of the random walk. Then the expected value off is defined by
N
<f>N = I" f(m)w(m,
m=-N
N). (7)
Random Processes and Partial Differential Equations [Ch. 3
The term "expected value" or average value is used, because in a large

number A of N-step walks, one would expect the particle to arrive m steps
to the right of its starting point Aw(m, N) times. The average value of f(m)
would be expected to be
N
A- 1
:["
m=-N
f(m)[Aw(m, N)] = <f>N (8)
It is customary to omit the subscript N in (7). Also, customary usage

employs (f(m)) instead of(/). For example, if f(m) = mk, (!) is replaced
by (mk). To give another example, the average value of p(m) = t(N + m),
the number of steps to the right, can be written in the following equivalent
ways:
N N
(p) =
m=-N
I" p(m)w(m, N) = L pc:ut.
p=O
Here we have used formula (4) for w(m, N). Passage from the first sum to
the second has been made by changing variables from m top.
The most important cases of (!) are the average or mean displacement
(m) and the mean square displacement or variance (m 2 ). The expected value
of f(m) = mk is called the kth moment.
In order to evaluate the various moments, we introduce the generating
function
N
G(u) = L c:ctluP.
p=O
(9)
This function is formed by attaching a power of a newly introduced variable

u to the probability w(m, N) and summing over all permissible ranges of the
parameter p (which is preferable to m because p goes in steps of unity).*
Alternatively, (9) can be viewed as a Fourier series of form
(10)
p
with u = eit and aP = 0 for p > N. The generating function method is

therefore related to the Fourier transform method to be considered later.
It is not difficult to follow manipulations using the generating function, but it is amazing
that anyone would have thought of this device. Amazement is lessened upon learning that the
first person to make use of the generating function was the genius Euler. In Vol. 1, Chapter VI,
of his splendid book Induction and Analogy in Mathematics (New York: Oxford U.P., 1954),
G. Polya discusses the motivation for using a generating function. He states that "a generating
function is a device somewhat similar to a bag. Instead of carrying many little objects detachedly,
which could be embarrassing, we put them all in a bag, and then we have only one object to
carry, the bag. Quite similarly, instead of handling each term of the sequence a0 , ai, a 2 , ,
a., ... individually, we put them all in a power series :I:. ax", and then we have only one mathe-
matical object to handle, the power series." And to deal with power series, we can use the extra-
ordinarily powerful theory of analytic functions. For a guide to the extensive literature on genera-
ting functions see E. B. McBride, Obtaining Generating Functions (New York: Springer, 1971)
and the review of this monograph by J. Wimp in SIAM Rev. 14, 663-67 (1972).
Sec. 3.1] Random Walk in One Dimension; Langevin's Equation
77
To demonstrate the use of the generating function. we note that
N N
G'(l) = L pC;<tl=
p=l m=-N
[" pw(m,n) = (p). (11)
But since p = ,t(N + m),

N N
(p) = ,tN [" w(m, N) + ,t [" mw(m, N) = ,tN + t(m). (12)
m=-N m=-N
If G(u) can be written in a simple form, we shall therefore have no difficulty

in evaluating G'(l) = (p) and hence the desired average value (m). But
another use of the binomial expansion (3) gives
G(u) = (I + ulHl, (13)
which is a closed form for G, in contrast to the series form (9). Equation (13)
allows us to conclude at once that G'(l) = N/2. From (11) and (12), then, we
see that
N
(p) =2 (m) = 0. (14a, b)
It is to be expected that the average displacement (m) is zero, for motions

to the left and to the right are equally probable.
Example. Find <m2 ) 1 2

' and interpret the result.
Solution. From (9)
N
G"O) =
p=O
oc:m"= <P<.P-
2 P<.P- n>= - 2
But, from (13),
G"(u) = N(N-1)(1 + u)N- 2 (21)" ,G"(l) = N(N-

4
1)

Using the result <P) = N/2 of (14a),


= tN 2
- tN,
However,
m = 2p - N so that (m 2) = 4(p 2 ) - 4N(p) + N 2
That is,
(m 2 )=N 2
+ N-2N 2
+N2 = N.
The root-mean-squared displacement from the origin, i.e., the standard deviation, is
(IS)
which is much less than the number of steps N. This is a consequence of the can-
cellation of efforts due to the random leftward and rightward motion of the
particle.
USE OF A STOCHASTIC DIFFERENTIAL EQUATION TO OBTAIN

BOLTZMANN'S CONSTANT FROM OBSERVATIONS OF
BROWNIAN MOTION
We now turn to a very brief exposition of how observations of Brownian
particles can lead to a determination of molecular properties. This possibility
was provided by Einstein, and the experiments were carried out by Perrin.
An outline of their approach follows.
Assume that the macroscopic resistance on the particle is proportional
to the velocity-as is predicted by classical hydrodynamics (see Section 3.6
of II). Introduce the ensemble average, <), a mean calculated from many
repetitions of the same experiment. Observe the particle for a period t to
obtain either (i) the net displacement projected along a line (denoted by x),
or (ii) the net displacement itself (denoted by r ). Einstein showed that these
observations of diffusion should obey the statistical law
<x2 ) = t<r2 ) = 2Dt, (16)
where the diffusion coefficient D is given by
k0
D=-. (17)
f
Here 0 is the absolute temperature, and f is the coefficient of resistance such
that the force acting on a particle moving at velocity v is -fv. For spherical
particles of radius a moving in a medium with viscosity coefficient , f is
determined by Stokes' law [Equation (3.6.10) of II]:
f= 61ta. (18)
In (17), k stands for Boltzmann's constant. (It is a fundamental physical
result that the average kinetic energy of a gas molecule is equal to fk0.)
Perrin was able to verify (16) and thus to determine the Boltzmann con-
stant k over wide ranges of 0, , and a. Since k equals the previously known
universal gas constant R, divided by Avogadro's number N 0 , the determina-
tion of k led at once to a value for N O
Result (16) is essentially found in Equations (3.8) and (3.28) by examining
random walk in an appropriate limit, but a fuller description requires another
approach. The modern theory of the Brownian motion of a free particle
generally starts with Langevin's equation,
dv
m dt = -fv + F(t), (19)
where v denotes the velocity of the particle and m its mass. The influence of
the surrounding medium is split up into two parts, the macroscopic dynamical
friction -fv and a random force F(t). It is assumed that F(t) is independent
of the position and the velocity of the moving particle, and that it varies
extremely rapidly compared to the variation ofv.
Sec. 3.1) Random Walk in One Dimension; Langevin's Equation 79
It would take us too far afield to reproduce the required theory of the Lange-
vin equation in this volume. The reader must be referred to references such
as pp. 22-27 of Chandrasekhar's article in Wax (1954). But in the next two
paragraphs we shall provide a very.sketchy treatment in an attempt to convey
the essence of the relationship one finds between macroscopic and micro-
scopic processes.
If we multiply ( 19) scalarly with x, the displacement of the particle from
the mean position of a group of particles, and consider the ensemble average
over the group, we obtain
2
m (~ (x v) - (v )) = -f<x v) - (x F). (20)
We now assume that (x F) = 0 and that ((dx/dt) (dx/dt)) becomes

stationary. We can then solve (20) for the quantity (x v). This quantity
is proportional to the diffusion coefficient; for, from (16),
(xv) = 1(t<:;
x)))
= 3D. (21)
The solution for (x v) shows that eventually (x v) alsobecomes stationary

and takes on the value
2
m(v )
(xv) =--. (22)
f
Unfortunately, we have no way of evaluating the right-hand side of (22)
in this simple derivation. We may argue, however, that the random motion
of the macroscopic particle will attain an energy which is equal to that of
the molecular motion. Thus we postulate that
2
tm(v ) = fk0. (23)
Combining (21), (22), and (23), we obtain (17).
EXERCISES
1. (a) Show that w(m, N) is an even function of m and interpret this result.
(b) Verify (13).
2. Given w(m, N) as in (4):
(a) Find (p 3 ) and (m 3 ).
(b) Find (p 4 ) and (m 4 ).
3. Consider a random walk where the probability p of taking a step to the
right is not necessarily the same as the probability ,t = I - p of taking a
step to the left.
(a) Show that now
(b) Introduce a generating function and use it to find the mean and
variance of the displacement m.
{c) Repeat Exercise 2 in this case.
4. (a) The Bessel function of the first kind, of order p, is defined as follows:
co ( - l)k(xfz)2k+11
(24)
Jp(x) = k~O k! (k + p)! .
Show (formally) that this series provides a solution to Bessel's

equation,
2 d2y dy 2 2
x - 2 +x- + (x - p )y = 0. (25)
dx dx
(b) If pis an integer n, show that
J_n(x) = (- ItJnCx). (26)
(c) Demonstrate that
~r/2e-xf2r = I, I, (- ll(x/2)i+k ,J-k = f rnJnCx), (27)

j=O k=O j! k! n=-co
which provides the generating function for the Bessel functions of

integer index n.
(d) In (27), replace r by exp (iO) and deduce the following integral repre-
sentation:
Jn(x)=2._
21t
f'
-,r
ei(xsin9-n9)d(); n=O, 1, 2, .... (28)
(e) By a change of variable, obtain the alternative formula
(29)
(We shall need the formula for n = 0 in Section 16.3.)
3.2 Asymptotic Series, Laplace's Method, Gamma

Function, Stirling's Formula
We have seen that the probability w(m, N) that a Brownian particle will
arrive m steps to the right of its starting point after taking a total of N steps
is given by
N!
w(m N)=---~
' p!(N-p)!2N'
where p = t(N + m). (1)
Sec. 3.2] Asymptotic Series, Laplace's Method, Gamma Function 81
We wish to approximate this result when N, p, and N - pare large; we are

thus led to the problem of approximating the factorial of a large number.
This is a general problem that occurs in many other situations. Years ago,
Stirling showed* that the natural logarithm of a factorial can be represented
by the series
ln n! - -1 In (2n) + ( n + -1) Inn - n + - 1 - -- 1 3 + . (2a)

2 2 12n 360n
The dominant terms give Stirling's formula or Stirling's approximation,
(2b)
By applying (2) to the factorials in (1) [Exercise l(a)], we obtain the approx-
imation
2
w(m, N) "' ( nN
)112
exp
( m2)
N . (3)
2
Table 3.1 shows that this approximation is an excellent one, even for such
seemingly moderate values as m = 4, N = 10. Nevertheless, the series of
which the first terms are given in (2) is known to be divergent for all values
of n. In what rigorous sense, then, is (2) a valid mathematical expression?
TABLE 3.1. w(m, N) for N = 10.
m Exact formula (1) Asymptotic formula (3)
0 0.24609 0.252
2 0.20508 0.207
4 0.11715 0.113
6 0.04374 0.042
8 0.00977 0.010
10 0.00098 0.002
To understand the nature of the Stirling approximation (2), and to handle

many other problems, we must consider the concept of an asymptotic
series. We begin our discussion with a relatively simple example, in which
an approximating series is generated by parts integration. As there is an
James Stirling (1692-1770) was an English mathematician of the period after Newton.
Stirling actually derived a series that has smaller coefficients than those in (2), but the coefficients
are more difficult to obtain. "The ingenuity needed to obtain the series with the mathematical
resources then available, not even the general definition of z! being known, is astonishing. The
usual form is most accurately described as de Moivre's form of Stirling's series, sin<:eall the
principles used in finding it were given by Stirling" in a letter to de Moivre (Jeffreys and Jeffreys,
1962, p. 467).
explicit expression for the remainder, it is not difficult to see the asymptotic
nature of the approximation. Following a formal definition of asymptotic
series, we turn to a heuristic description of Laplace's method for generating
asymptotic series for certain important integrals. The method is illustrated
on the gamma function integral, an extension of the factorial function.
A demonstration of the method's validity is outlined. As a result of this, the
Stirling formula is proved.
AN EXAMPLE: ASYMPTOTIC EXPANSION VIA PARTS

INTEGRATION
Suppose we want to evaluate the integral
f(x) = f""1-iex-1 dt (4)

"
for large positive values of x. The improper integral converges [Exercise
l(b)]. Using integration by parts, we obtain the following expression:
1
J
OC)
f(x) = - - t- 2 e"-t dt. (5)

x "
The magnitude of the integral in (5) is smaller than that in (4) by at least a
factor of 1/x, since t ~ x in the interval of integration. Consequently, we can
deduce from (5) that
I
- =f(x) + J"" C 2 e"-' dt ,;::,:,f(x).
x "
x- 1 by successive partial inte-
We can improve the approximation f(x) ,;::,:,
grations. One finds [Exercise 2(a)] that
f(x) = Sn(x) + Rn(x), (6a)
where
S (x) = ~ - ~ + 2! - ... + (-l)n-l(n -1)!

n x x2 x3 xn ' (6b)
Rn(x) = f(x) - Six)= (- ltn! f""t-<n+l)e"- 1

dt. (6c)
"
The sequence Sn(x) diverges as n-+ oo, for all values of x. Yet we feel that
Sn(x) must be a good approximation in some sense when xis large. We shall
therefore examine the remainder Rn(x) as x-+ oofor afixed value of n.
We know that
It"" r<n+l)ex-, dtl :s;; t"" ,c(n+1)1 ,e"-,1 dt.


Since t ~ x, I t-<n+t) I ~ x-<n+t>, so that we can write
I Rn(x) I < n! x-Cn+l) f' <r-,

.>:
dt = n! x-(n+t>. (7a)
Thus, if x--+ co, RnCx)--+0 for any fixed value of n. In fact, as long as x ~ 2n,
one can show [Exercise 2(b)] that
IRnCx)I < r<n+l)n- 2, (7b)
which is small even for moderate values of n. When n = 3, for example,
provided that x ~ 6.
Equation (7a) shows that the remainder in our series is less than the first
term neglected. This is not always the case, however. The generalizable
property of (7a) is that for fixed n, the ratio of the remainder to the last term
retained approaches zero as x --+ co. Let us formalize this statement in a defin-
ition.
DEFINITIONS IN THE THEORY OF ASYMPTOTIC EXPANSIONS
Consider
A1 A2 An An+t An+l
Ao+7+ x2++ xn+ xn+1+=Sn(x)+ xn+1+. (8)
The expression Sn(x) is said to provide an asymptotic expansion of f(x)

(as x--+ co) if
lim ~[f(x) - Sn(x)] = 0 (n fixed). (9a)
In other notation, requirement (9a) can be written
f(x) - Sn(x) = o(x-n) as x--+ co (n fixed). (9b)
In case (9) holds we write

n
f(x)- L A,x- 1
, X-+ CO. (10)
i=O
The symbol " ,.., " is read "is asymptotic to." If (10) holds for every n, we
write
00
f(x),.., L A,x- 1
, X-+ CO, (11)
i=O
and we speak of the asymptotic power series development off(as x-+ co).
In practice, the series in (11) is usually divergent for any fixed large
x. On the other hand, if this series does converge to f for x sufficiently
large, i.e., if
lim lf(x) -
n~a:,
_IA;x-il
i=O
= 0 for x large and fixed, (12)
then (11) is asymptotic as x-+ oo (Exercise 3).

As we have seen, in
I,,""t- 1~-t dt- nt(-1)"- 1

(n - l)!x-n
we have an example of a series that diverges for all x but is asymptotic as

x-+ oo. The distinction between a convergent series and an asymptotic
series must be kept clearly in mind. Toprove that a seriesexpansionis asymptotic
as x-+ oo, we examine the remainderf(x) - S"(x), after a fixed number n of
terms, when x becomes very large. By contrast, a convergenceproof requires
examinationof the remainderat a fixed value of the independentvariablewhen
n becomes very large.
The definition of asymptotic series can be extended to include x-+ - oo.
Moreover, one often wishes to consider Izl -+ oo, where z is complex.
(In the latter case, there is usually a restriction on the range of values of arg z.)
Asymptotic expansions as z approaches an arbitrary point z0 may also arise,
but it is often convenient to reduce to the case where the variable approaches
oo by introducing 1/(z - z0 ) as a new variable.
To give one more definition,
N
f(x) "'g(x) L A"x-",
n=O
X-+ 00, (13)
means
--
f(x)
g(x)
X-+ 00. (14)
Asymptotic expansions can often be obtained by judicious use of parts

integration. (See Exercise 7.) We shall now describe a different approach,
which is of wider applicability and which provides a relatively straightforward
derivation of Stirling's series.
LAPLACE'S METHOD
We present a method, due to Laplace, for obtaining the asymptotic
expansion of certain integrals containing a large parameter. Suitably general-
ized, this method is one of the most frequently used in applied mathematics.
In its standard form, Laplace's method is used for integrals which can be
placed in the form
=f
p
F(l) g(t)e-J.f<t>
dt, ).> 0. (15)
"
Sec. 3.2] Asymptotic Series, lAplace's Method, Gamma Function 85
An approximation is sought in the case in which the positive constant A

is large. The idea used in the whole class of methods of which Laplace's
method is a paradigm is this: transform the integral if necessary until, for
large l, the dominant contribution to the result comes from a small portion
of the path of integration. Suppose that portion is in the neighborhood of
a point t0 Then simplify the integrand in that neighborhood by using a
Taylor expansion about t 0
In carrying out explicit calculations on (15), we shall assume that /(t)
has an absolute minimum at an interior point t 0 of the interval [oc,Pl, with
/'(t 0 ) = 0, f"(t 0 ) > 0. Approximating /(t) by the first two nonvanishing
terms in its Taylor series, we see that exp [-l/(t)] behaves like Q(t) exp
[ -l/(t 0 )], where
2
Q() _ [-}.f"(t 0 )(t - t0) ]
t=exp .
2
Figure 3.1 depicts graphs of Q(t) when
J,r(to)
- - = 1, 10, and 100.
2
For large l, Q(t) is seen to be appreciable only within a very small dis-
tance of t 0 Within a narrow band about t 0 it should be permissible to
1.0
0.5
0 0.5 1.0
t -1 0
2
FIGURE 3.1. Illustration of the fact that exp[-A(t-to) ] decays
rapidly from its maximum value when A is positive and large.
approximate g(t) of (15) by g(t 0 ). We are thus led to the conjecture that
F(..l) = f./l g(t)e- ..J<t>

dt
Cl
,..,g(t 0 ) exp [-1f(t 0 )J ffl
Cl
Q(t) dt.
But since only a narrow interval about t 0 makes a substantial contribution

to the integral, we can obtain a further simplification at little cost in accuracy
by replacing the limits of integration by - oo and oo. With the change in
variable
u = (t - to)
[f";o)] 112
our result becomes [Exercise S(a)]
F(..l),.., uUo)exp [-..lf(to)J [f"~to)r'

f~oo
2
exp (-..lu
2
) du. (16a)
Evaluating the integral by (34) below, we finally obtain*
F(..l),.., g(to) exp [-..lf(t 0 )] J

L~); 1
112
as ..l-+ oo. (16b)
Rather than provide a proof for (16b), we pass at once to the development
of Stirling's series. We shall perform manipulations that are similar in spirit
to those just completed.
DEVELOPMENT OF THE ASYMPTOTIC STIRLING SERIES FOR
THE GAMMA FUNCTION
Our first step, by no means an easy one, is to find an integral representation
for n !. This is accomplished by introducingt the gamma function r(s), where
r(s) = f
0
00
x1- 1 e-x dx. (17)
It takes only a straightforward application of parts integration [Exercise 2(c)]

to verify that when sis a positive integer n, then
r(n) = (n - l)!. (18)
But the integral in ( 17) is defined for more than positive integral values of s.
An important special case is
(19)
Formula (16b) can be generalized so that it applies to complex integrals by choosing an
integration path of" steepest descent" near "saddle points" that are complex analogues of the
stationary points off. In the present case, the path of integration is fixed, but it happens to be
the path of steepest descent.
t Given our earlier remarks on the origin of the generating function, the reader will not be
surprised to learn that it was Euler who obtained the integral representation (17).

a formula with a trick proof [Exercise 2(d)]. Indeed, (17) remains meaningful
even ifs assumes complex values, as long as Re(s) > 0.
We are interested in the behavior of r(s) for large positives. To obtain an
integrand containing an exponential withs as a factor, we write (17) as
r(s) = s:xs- le-x dx = ft)e<s-l)lnxe-x dx,

which shows that
r(s) = f'
0
e-F dx, where F(x, s) = -In x + s- 1 In x + s- 1x.
Considered as a function of x, F(x, s) is stationary at the point x 0 , where
8F =O i.e., x 0 =s- I. (20)

ax '
In contrast to the fixed location of t 0 in our treatment of (15), the point
x 0 "moves" as s-+ oo. To fix its location, we make x = s - 1 correspond
to the fixed point t = I by the change of variable
x
t=-- (21)
s-1
With this (17) becomes
r(s) = (s - l)'J, (22)
where
J= J: exp [-(s - I)f(t)] dt, f(t) =t - In t. (23)
The minimum off(t) occurs for

f'(t 0) = 0 or to= I. (24)
At the minimum,fhas the value/(1) = 1. The integrand of (23), therefore,

has a maximum at t = I and drops off extremely rapidly as t goes away from
this point.
To bring out the dominant features of the integrand, we write
.).= s - I, f(t) = f(t 0) + [f(t) - f(t 0 )], (25)
with which our integral becomes
J(l) = e-Af(to) (' e-A[f(r)-f(to)ldt. (26)
(We have assumed thats - 1 =;. is positive, but since we are interested in
larges, this is not an additional restriction.) To emphasize the key fact that
88 Random Processes and Partial Differential Equations [Cir. 3
[/( t) - /(t 0 )] is a nonnegative function with a single minimum, we introduce

a new variable w that satisfies
w2 = f(t) - f(t 0 ). (27)
We specify w completely by
w= -Jl{i)-T<tJ, (28)
w= J f(t) - f(to),
Using the fact that w increases continuously and monotonicaily from - oo
to oo as t increases from Oto oo [Exercise 4(a)], we rewrite (26) as
J(l) = e-J.f(ro) (' e-;..w> !!!_dw. (29)

_
00
dw
The next task is to expand dt/dw as a power series in w. Now

w2 =f(t) - /(t 0 ) = t - 1 - Int, (30)
so
wi = t(t - 1)2 - !(t - 1)3 + .... (31)
Consequently [Exercise 4(b)], we may write
t - I = ../2w(l + a w + a 1 2 w2 + ). (32)
The coefficients a 1 , a2 , may be determined by substituting (32) into (31)
and comparing terms [Exercise 4(c)].
From (32) we can easily calculate a power series for dt/dw. Assuming that
term-by-term integration is permitted, we find that the evaluation of (29)
reduces to computation of integrals of the form
Im(A)= f"" e-;..w>~ dw; m =0, I, 2, .... (33)

-oo
An elementary substitution shows that I 0 is a multiple of r(t). Therefore,

using the expression for r(t) in ( 19),we find
Io(l) =J;fi. (34)
Similar substitutions reduce determination of Imto computation of r integrals.

Alternatively, one can formally differentiate (34) with respect to l (Exercise 5).
The result is
I,,.(>.)= 0, m odd;
_ 1 3 . 5 (m - 1)
I,,.(>.)- 2m/2)..m/2 >., '
(!)112m even. (35)
Sec. 3.2) Asymptotic Series, Laplace's Method, Gamma Function 89
Putting all our results together (Exercise 5), we obtain
J= (2;)e-}.[l + O(l-
112
1
)], (36)
so that
In r(s) = (s - !) In (s- 1) - (s - 1) +!In 21t+

2 2
o(-
s-1
1
-). (37)
In particular, whens= n + I, nan integer, we recover the leading terms of

Stirling's series (2a).
JUSTIFICATION OF TERM-BY-TERM INTEGRATION
Series like (32), in general, have a finite radius of convergence, and thus
our termwise integration is not justifiable if one insists on working with
convergent series. Yet, for large A.,the only important contribution to the
integral is near w = 0, and here the series is expected to converge. We have
a strong intuition that all is well. Indeed, the result (35) can be justified (and
even extended, as in Exercise 6) in the asymptotic sense. To do this, instead
of using the series (32) we make a slight change of viewpoint and write t - 1
as a finite sum plus a remainder:
t- 1 =J'iw[l + ... + 2
02n w n + 02n+lw2n+l + Bl2n+iw)w 2
n+
2
1. (38)
Coefficients a; with odd subscripts provide no contribution to the integral
in (29), since the corresponding integrand is an odd function. Thus the last
nonvanisbing contribution from the polynomial in (38) is proportional to
112
/ 2 n(J.)and this is O(rn- ), by (35). The remainder makes a contribution
proportional to
I~"'Bl2n+2<w>w2n+2e-)."'2 dw = ;,-n-3/2 I~
'9l2n+2 ~ ,2n+2e-,2 dt. (39)
For (39) to be O(;.-n- 3 12), as desired, all we need is that the integral on the
right side converges. This will certainly be the case if ~ 2 .. +it/.Ji) is bounded
for large t, or even if it becomes infinite as long as this is not in any way
comparable with the growth of exp (t 2 ).
The literature contains a number of theorems that give sufficient conditions
for the validity of various manipulations which yield asymptotic expansions.
As usual, however, we stress here the spirit of the approach, since in any case
published theorems must often be extended before they apply to a case met in
practice. The extension may be very difficult. If so, many applied mathemati-
cians may not make the effort to carry it out.
This is quite typical of many manipulations in applied mathematics.
The formal steps can be justified, but perhaps the usual rule is to accept the
plausible, unless there is reasonable doubt to the contrary, i.e., unless a
serious issue can be raised. If you wish, you may imagine that the following
qualifying statement is often implied in work in applied mathematics:

"The result obtained is plausible and tentative; but we have as yet no reason
to doubt its validity." In many cases of extreme complexity, the mathematical
justification, even of a general process, may take years to accomplish. It is
obvious that one cannot wait for it before going ahead. The attitudes and
practices that we adopt are consequently fundamentally different from those
adopted by pure mathematicians.
EXERCISES
1. (a) Verify (3).
(b) Show that the improper integral in (4) is convergent.
2. (a) Verify (6b) and (6c).
(b) Verify (7b).
(c) Verify (18).
(d) Show that
I= f'
-o:::,
e-:xl dx = f.., e-yz
-co
dy = Jn
by considering / 2 and introducing polar coordinates in the (x, y)
plane. Thereby verify (19).
3. Show that if (8) converges for x sufficientlylarge, then (8) is an asymptotic
series as x -+ oo.
4. (a) Verify (29) and (31).
(b) Provide a formal proof that justifies the inversion necessary to
obtain (32).
(c) Determine the coefficientsa 1 and a 2 in (32).
5. (a) Verify (34) and (36).
(b) Derive the formulas in (35) by reducing the integrals to gamma
functions.
(c) Derive (35), for even m, by formal differentiation of (34).
(d) Justify the formal procedure of part (c).
6. Extend the text's derivation of (37), and obtain more terms of Stirling's
series (2a).
7. Use integration by parts to obtain asymptotic expansions as x-+ oo of the
following integrals:
..,
(a) The complementary error function 2n- 112
f
:x
e- 12
dt.
(b) The Fresnel integrals
C(x) = ( cos G 2
1 ) dt, S(x) = ( sin G 1 )
2
dt.
8. (a) Verify (16a) and (16b).

(b) Provide a formal proof that () 6b) is valid if suitable restrictions are
made on/ and g.
Sec. 3.3] A Difference Equation and Its Limit 91
3.3 A Difference Equation and Its Limit

One of the essential elements that underlies the success of theoretical
analyses is the practice of approaching the same problem in a variety of ways.
The strength of the various methods will then supplement each other and
enable one to reach a deeper understanding. So far we have attempted to
follow the motion of a particle as it started out from the origin. We have
obtained exact and approximate expressions for w(m, N), the probability
that a particle will arrive m steps to the right after taking a total of N steps.
If we now turn our attention to the particle's immediate behavior after it
has reached the point i steps to the right of the origin, we shall be enlightened
in an entirely different manner. We shall find that random walk can be de-
scribed by a certain difference equation. This equation can be approximated
by a differential equation, which turns out to be the diffusion equation.
Various solutions of the diffusion equation will be shown to provide solutions
for probability problems of considerable interest.
A DIFFERENCE EQUATION FOR THE PROBABILITY FUNCTION
We assert that the following difference equation,
w(x, t + Lit) = tw(x - Lix, t) + tw(x + Lix, t), (l)
governs the change of
w(x, t) ~r)
=w(:X, (2)
in finite steps of time Lit. In (I) and (2) we have defined

x = m(Lix), t = N(Lit). (3)
Thus w(x, t) is the probability that the particle which starts at the origin at
time t = 0 is located at point x at time t. Our difference equation (1) merely
states that to reach the point x at the time t + Lit, at time t the particle must
either reach the point x -Lix [probability w(x - Lix, t)] and then take a step
to the right [probability t], or it must reach the point x + Lix and then take
a step to the left.
Equation (1) governs the evolution of w.To complete the formulation of
the problem in these terms, we must add the condition that the particle
begins at the origin:
w(O,O) = I ; w(x, O)= o, x:f::0. (4)
We expect to be able to recover the exact solution (1.4) for w by solving the
difference equation (I) subject to the initial conditions (4). Readers may not
be able to provide the solution from first principles, but they can easily
verify that (1.4) satisfies (1) and (4) (Exercise 1).
92
APPROXIMATION OF THE DIFFERENCE EQUATION BY A
DIFFERENTIAL EQUATION
In our previous discussion we obtained the approximate solution (2.3) by
applying Stirling's formula to the exact solution (1.4). A very useful result
emerged, but surely there must be a more direct way to obtain the approxi-
mation. It should require less work to obtain a less detailed result, and indeed
the present formulation provides a direct and natural approach to the
approximation (2.3).
We are interested in the solution only after the particle has taken a large
number of steps. Thus we should consider the limit N-+ oo. From (3),
however, if we leave !;.t fixed in this limit, then t-+ oo. We wish to examine
the solution at an arbitrary time t, so we are led to approximate (1) in the
limit
N-+ oo, t fixed, so /;.t -+0. (5)
Recall that m tu is the coordinate of the particle after N steps, assuming

that the particle starts from the origin. If the particle takes a very large total
number N of steps, we expect that there is an excellent possibility that the
(positive or negative) integer m may be very large in magnitude. We certainly
do not wish to exclude such a possibility; therefore, we also consider the
limit m-+ oo. Because we wish to be able to describe the situation at any
fixed point x, from (3) we are led to approximate (1) when
m -+ oo, x fixed, so tu-+O. (6)
The limits (5) and (6) require us to approximate (1) when tu and flt are
small. To accomplish this, we naturally think of expanding the functions
in (1) by Taylor's formula. [It gives us pause that wis so far defined only at a
set of discrete points in the (x, t) plane, but we can circumvent this problem
by imagining a polynomial that smoothly passes through all these values
of w.] We thus approximate (1) by (Exercise 2).
w,(x,t) at+ 0(!;.t) 2 = t)(tu) 2 + O(tu) 3

fwx.:Cx, (7)
We note that if we divide (7) by at and assume that

. (tu)2
hm A = D, D 0, (8)
4x-o 2 at
4r-o
we obtain the plausible-looking result wr= Dwxx This is encouraging,

but we must proceed a little more cautiously.
First, we observe that it is no longer appropriate to keep the quantity w
as the dependent variable. This is because the probability of finding the par-
ticle at a particular point is expected to approach zero as the number of
Sec. 3.3] A Difference Equation and Its Limit
93
points becomes infinite. Since only every other point on the x axis can be
occupied at any instant, we introduce the quantity
w
u=-- (9)
2Ax
In terms of u, the probability at time t of finding a particle on or between

a = i ~ and b = k ~ can be expressed by
k
I" u(m .6x, t) 2 Ax.
m=i
(10)
[Remember that the sum takes in every other value of m from i to k. We

take both i and k to be even (odd) if N is even (odd).]
We must also take note of the fact that assumption (8) seems at first to
have little intuitive appeal. But if this assumption is not made, then our
limiting process does not seem to yield a sensible result. Let us therefore
accept (8) for the moment. If further examination of our limiting results
confirms our impression that we are making progress, we can return to the
justification of (8). On the other hand, if our formalism proves unedifying,
there is no point in convincing ourselves of its legitimacy.
Using (9) and (8), then, the leading approximation to (7) for small ~
and tit is
(I 1)
Taking the limit of (10) as /u-+ 0, we find that u(x, t) is related to the prob-
ability U(a, b; t) of finding the particle in the interval (a, b) at time t by
U(a, b; t) = f u(x, t) dx.

b
a
(12)
More informally, the probability density function u is such that u(x, t) dx is

the probability that at time t the particle is located between x and x + dx. *
SOLUTION OF THE DIFFERENTIAL EQUATION FOR THE

PROBABILITY DISTRIBUTION FUNCTION
What if a particle starts at the origin and takes a large number of steps
at random? We would expect that the distribution of probabilities would
satisfy (11). Certainly,
f' -oo
u(x, t) dx = 1, (13)
Note the analogy with the mass density function p(x, t), which can be regarded as providing
the mass of material located at time t between x and x + dx.
94
for the particle must be somewhere along the x axis. Moreover, since the
particle started at the origin we must require that*
lim u(x, t) = 0, X#: 0. (14)
r.j.O
If we could solve the differential equation (11) subject to the conditions

(13) and (14), we would expect to achieve our goal of proceeding directly
to an approximation for large N, without first obtaining the exact solution.
Those reasonably familiar with partial differential equations can find the
desired solution by looking for a function u of self-similar form. But to see
whether our program is sensible, we can work backward. In present terms,
the approximate solution (2.3) is [Exercise 3(a)J
U ~ Uo(X, t), u0 (x, t) .= (4,rDt)- 1 2

' ( x2).
exp - - (15)
4Dt
And it is not difficult to verify that u0 satisfies the differential equation (11),
the normalization condition (13), and the initial condition (14) [Exercise
3(b)]. It can also be shown that the solution is unique.
We have uncovered a method for directly obtaining an approximate
solution to random walk problems: set up a difference equation, perform
appropriate limiting operations, and then solve the resulting problem in
partial differential equations.
Since our approach seems successful, we shall want to examine it more
carefully and also to try it out on some new problems. This will be done
in the remainder of this chapter. For the moment, let us look a little more
closely at the solution u0 We note that the area under the curve u = u0 (x, t)
is always unity, but that u0 becomes more sharply peaked as t ! 0. The maxi-
mum value of Uo (at x = 0) increases as t! 0 like i- 112 , but the "width"
decreases like t 112 t Thus, as t ! 0, u0 approaches a limiting "delta function"
that has unit area even though it is zero at every point except the origin.
As t increases, the width of u0 increases from its initial zero value like 1112 ,
a fact already indicated by Equation (1.14a).
FURTHER EXAMINATION OF THE LIMITING PROCESS
We have seen that the approximation (15) can be obtained in two ways,
by applying Stirling's formula to the exact solution and by solving the limiting
differential equation (11) under appropriate conditions. But for accurate
approximation via Stirling's formula, we had to require not only that the
total number of steps N was large but also that p and N - p were large.
Recall from Section 3.1 that the latter two numbers are the total number of
steps taken to the right and left, respectively. Thus if p or N - p is small,
The downward arrow in (14) means that t approaches zero through positive values.
t The width may be defined by setting x 2 = 4Dt, for instance, to obtain the position where the
exponential factor in (15) is 1/e of its maximum.
95
most of the steps are either taken to the left or to the right, and the displace-
ment m is comparable either to - Nor to N. As a consequence, we may inter-
pret the above-mentioned requirement on the use of Stirling's formula
as excluding those rare cases where the displacement is comparable to the
total distance traveled. This requirement was not mentioned in our derivation
of the limiting differential equation. Why not?
To see the answer, let us write out in full the limiting process we used.
This was
tu-+ 0, At-+ 0, m-+ oo~ N-+ oo, m(tu)-+ x, N(At)-+ t,
(tu)2 (16)
2(M)-+ D.
With this
tu 2D
----+OO (17)
At tu '
so that the speed with which the tiny steps are taken is infinite in the limit.
Also, and more importantly,
~ = m Ax flt = :: At -+ O. (18)
N N At Ax t Ax
Thus, our limiting process does, in fact, implicitly use the assumption that the
displacement from the origin, m tu = x, is small compared to the total
distance traveled, N Ax.
The previous discussion provides some insight into the reason why our
limit must be such that (Ax)2/At remains finite. From (18) the requirement
that m be small compared to N means that Ax/At must approach infinity.
This makes our requirement of finite (Ax) 2 / At a little less mysterious. But
a better reason for this requirement is that the ratio of the net displacement m
to the standard deviation N 1' 2 is x[M/t(~x) 2 ]1'2 [Exercise 3(c)]. Equation
(3.3) shows that most net displacements have magnitudes less than a small
multiple of N 1' 2, which provides good motivation for considering a limit
wherein m/N 1' 2 is fixed.
REFLECTING AND ABSORBING BARRIERS
To give an indication of the power afforded by the limiting differential
equation, let us consider a random walk problem in which a particle starts
from the origin, as before, but in the presence of a perfectly reflecting barrier
at x = L + t tu, L > 0. The presence of this barrier means that a particle
that reaches x = L at time t and then moves to the right will again be found
at x = L at time t + At. A particle also can attain the position x = L by
moving to the right from x = L - tu. Consequently,
w(L, t + At) = tw(L - tu, t) + tw(L, t). (19)
Random Processes and Partial Differential Equations {Ch. 3
Upon introducing the probability distribution function u = w/(2 Ax) via (9)
and letting Ax-+ 0, this condition becomes
u,,(L, t) = 0. (20)
All other conditions remain the same, except that (13) must be replaced by
f L
-ao
u(x, t) dx = I, (13a)
since all particles remain to the left of x = L.

Finding a solution u = uR(x, t) to the differential equation (11) that
satisfies (13a), (14), and (20) is-as we shall see in Section 3.4--not too much
more difficult than finding a solution to (11) when the last condition is
omitted. The result is that the distribution function in the presence of a barrier
at x = L is given by
uR(x, t) = Uo(X,t) + Uo(X- 2L, t). (21)
It is possible to find an exact solution to the random walk problem with a
reflecting barrier and then to obtain (21) by using Stirling's approximation.
But, given some familiarity with differential equations at least, a smaller
amount of ingenuity is perhaps necessary to obtain (21) from an appropriate
solution to the differential equation (11). For both the exact and the approxi-
mate problems, a reflection idea is generally used to obtain a solution.
Let us now consider the situation when there is a perfect absorberjust
to the right of x = L. By this we mean that the particle is removed if it moves
to the right from x = L. Thus the only way that a particle can arrive at x = L
is first to arrive at x = L - Ax and then move to the right. In this situation,
then, the governing difference equation holds for x =s;;L - Ax, and there is
the further condition that
w(L, t + At) = tw(L - Ax, t).
Upon introducing u = w/(2 Ax) and neglecting higher order terms in Ax, we
obtain the boundary condition
u(L, t) =0. (22)
The solution of (11), (14), and (22) is (for - oo < x =s;;L)

uA(x, t) = Uo(X,t) - Uo(X- 2L, t). (23)
Equation (23) gives the approximate probability that a particle, having
started from the origin at time zero, is located between x and x + dx at
time t. Also of interest is the probability F(L, t) dt that a particle is absorbed
at x = L in the time interval (t, t + dt), given that it started at the origin
when t = 0. Another interpretation of F is that it represents the probability
97
that a particle makes its first passage to L in the time interval (t, t + dt).
As we shall demonstrate in Section 3.4, the required formula is remarkably
simple. It is
F(L,t) = -D (-OU,4.). (24)

OX x=L
Thus the first passage time is proportional to the gradient in the probability
distribution function for random walk in the presence of an absorbing
barrier.
COAGULATION: AN APPLICATION OF FIRST PASSAGE

THEORY
We shall indicate one of the numerous applications of first passage time

calculations by considering the coagulation exhibited by colloidal particles
when an electrolyte is added to the solvent. [Our discussion will be very
sketchy. For more details see pp. 60ff. of Chandrasekhar's article in Wax
(1954).] The theory, due to Smulchowski, begins by assuming that the elec-
trolyte causes a sphere of influence of radius R to appear around each particle,
and that particles stick when they approach within a distance R. Suppose
that the sphere of influence of a single particle is centered on the origin and
that a sea of other particles, of concentration v, moves randomly around it.
Regarding the sphere as an absorbing barrier and the random sea as associated
with a diffusivity D 1 , the rate at which particles arrive at p = R (p the radial
polar coordinate) is
p
2
4nD 1 ( -,:1- ow) , (25)
up p=R
where we have used the formula for the gradient in spherical coordinates.
In (25), w is the solution of
Wr = D1V2 w, p>R.
w = v at t = 0, p>R; (26)
w=O atp = R, t >0.
H the particle, assumed stationary in the first instance, is itself undergoing

Brownian motion with diffusivity D 2 , it can be shown that D 1 in (25) and (26)
should be replaced by D 1 + D 2 Knowing the coalescence rate, one can then
write down a set of kinetic equations that describe the rate of formation of
k-fold particles by coalescence of i- andj-fold (i +j = k), and the disappear-
ance of k-fold particles due to their coalescence with other particles. Some
simplifying assumptions allow solution of the equations, and the results are
in reasonable accord with experiment.
EXERCISES
1. Verify that (l) and (4) are satisfied by the function w(m, N) defined in
(1.4).
2. Verify (7).
3. (a) Verify that (2.3) implies (15).
(b) Show that the function u0 defined in (15) satisfies (II), (13), and
(14).
(c)Show that (l.15) can be interpreted as showing that the mean
squared displacement in time tis t(ll.x')2/(ll.t).Use (3) to verify that
mN-112= x[il.t/t(ax)2p12.
(d) If, by chance, you are already familiar with similarity methods [see
Section 3.4 of II], use them to derive (15).
4. Using ( 15), evaluate the first four moments for diffusion from a point
source. Compare them with those obtained in (l.14), (1.15), and
Exercise 1.2.
5. (a) Consider a particle executing an unbiased random walk on a cubic
lattice, between points a distance ll.x apart. Let At be the duration
of a step. 'Bygeneralizing the text's discussion of the one-dimension-
al problem, show that
(27)
where
(28)
and u is suitably defined to represent a probability density.

(b) Comment on the difference between (28) and the corresponding
one-dimensional formula (8).
(c) What do you think that the two-dimensional version of (28) and
(8) is? Test your conjecture.
6. Consider one-dimensional random walk in which the particle moves a
step to the left, a step to the right, or stays where it is, with equal prob-
ability. Show that the limiting probability distribution again satisfies
the diffusion equation. Why is the change in definition of the diffusivity
D plausible?
7. Consider the random walk of a particle along the x axis. Suppose that
the particle moves a distance 11to the left or right with probability
q or I - q, respectively. One step takes T time units.
(a) Write a difference equation for w(m, N), the probability that the
particle is a distance mil from the origin at time NT (m and N
are integers).
t(b) Find the limiting differential equation as A-+ 0, T-+ 0, A2 /2T = D,
mil. -+ x, NT -+ t. Show that q must approach t in a certain way
if there is to be a limit of the type sought.
Sec. 3.4) Further Considerations Pertinent to Differential Equations 99
8. Consider random walk along the x axis in a field of constant force.

In time -r a particle moves a distance 11to the right or left with prob-
ability t - /111or t + Pli, where /3 is a constant. Write a difference
equation for the probability that the particle is a distance mA from
the origin at time N-r, m and N integers. Find the limiting differential
equation as 11-+0, -r-+ 0, 112 /2-r = D, m/1-+ x, N-r-+ t.
9. An elastically bound particle moves to the right or left, if it is at Ka,
-R ~ K ~ R (R a fixed integer), with probability f(J - KR- 1) or
t(l + KR- 1). The duration of each step is -r. Convince yourself that
"elastically bound" is a sensible description. If 11-+0, -r-+ 0,
R-+ oo, 112 /2-r = D, (R-r)- 1 -+y, etc., derive the limiting equation
OU_ o(xu)
ot -y~+D
(o
2
ox 2
u)'
(29)
for an appropriate probability density u.

10. It is not easy to derive honestly, but it is straightforward to verify, that
()J
u x, t =
.--
Y (l
2nD - e
-211)-112 [ -yx
2
exp 2D[l - exp ( - 2yt)]

] (30)
satisfies (29) and the auxiliary conditions u;;;:::O; J~..,u(x, t)dx = 1;

lim 1_ 0 u(x, t) = 0, x #: 0.
(a) Verify this.
(b) Compare this result with that of the free particle. In particular,
show that (30) approaches the free particle result as y -+ 0.
(c) Comment on the behavior of the two results as t-+ oo for x fixed.
11. (a) Suppose that a particle moves in an unbiased random fashion
in the plane, taking a step of length A every -r seconds. Justify
the equation
tJ,(x,y, t + -r) =
2
1,r f2"tj,(x + a cos 0, y + A sin 0, t) dO. (31)
O
In doing so, define tj,. Start by explaining why a particle should

step in a direction between O= 00 and (J= 00 + d(J with prob-
ability d0/2,r.
(b) Introduce a probability density function. By approximating (31),
show that this function satisfies the diffusion equation.
3.4 FurtherConsiderationsPertinentto the Relationship

BetweenProbabilityand PartialDifferentialEquations
It is quite remarkable that the process of Brownian motion is governed
in the limit by a partial differential equation of a fairly simple form. It is
even more remarkable that this equation is common to many other physical
processes such as heat conduction or diffusion of one gas through another,
for these are quite different physical processes, at least superficially. The fact
that many natural processes can be described by similar or even identical

equations is essential to the very existence of applied mathematics as a scien-
tific discipline. For, if there were no such natural unifying tendency, the
mathematical study of various phenomena would be hopelessly fragmented.
In the present section we continue to study the relations between random
walk and partial differential equations. We begin by reviewing the derivation
of the diffusion equation from a macroscopic or continuum viewpoint.
We stress that different microscopic models can give rise to this same macro-
scopic equation. The blurring of inessential microscopic detail is one reason
why the theory of a number of rather different phenomena is subsumed in a
study of a few differential equations.
We shall examine various solution techniques for the diffusion equation.
The interplay between probability and differential equation theory is illus-
trated in the process. For example, a probability problem gives rise to the
source solution, which is the key to solving the general initial value problem
for the diffusion equation. On the other hand, certain special solutions for
the diffusion equation provide answers to probability questions.
We close with an examination ofa remarkable set ofrelationships between
various probability problems on the one hand, and Laplace's equation and
the inhomogeneous Laplace equation (Poisson's equation) on the other.
MORE ON THE DIFFUSION EQUATION AND ITS CONNECTION

WITH RANDOM WALK
We shall examine the limiting partial differential equation

2
au= D(a u) (1)
at ax 2
more carefully. We have already encountered this equation in Chapter 1.
Let us review the main points of its derivation from a macroscopic perspective.
Let u(x, t) be the amount per unit length of some quantity. Let J(x, t)
denote the flux of this quantity, the net rate at which the quantity is passing
from the left of x to the right of x, at time t. * If the quantity is conserved, the
rate at which it accumulates in an interval must equal the rate at which it
passes into the interval through its boundaries. Thus, for arbitrary L\x,
d J"+4%
- u(x, t) dx = J(x, t) - J(x + L\x, t).
dt "
In the limit as L\x - 0, this yields
au aJ
at= - ax (2)
It is often useful to regard u as the amount per unit volume of a quantity whose distribution
only depends on a single coordinate x. Then J is a flux density, the flux per unit area placed
perpendiculary to the x coordinate line.
Sec. 3.4] Further Considerations Pertinent to Differential Equations IOI
We suppose that there is a tendency for relatively high concentrations

of u to disperse, i.e., there is a flow from high concentrations to low concen-
trations. The simplest reasonable hypothesis one can think of to describe a
process in which flow is "caused " by differences in u is to assume that the
magnitude of J is proportional to au/ax. Denoting the proportionality
constant by - D, we therefore write
au
J= -D-. (3)
ax
With D positive, (3) correctly shows a flow of u to the left if u increases with
x (au/ax>0) and a flow to the right if u decreases with x.
Combining (2) and (3), we obtain (l) (assuming that D is a constant).
This diffusion equation is the one we obtained as our limiting description of
random motion. As mentioned in the first paragraph of this section, (1) is
known to provide an excellent description of many physical phenomena.
For example, if u denotes the concentration of solute particles immersed in
solvent [in which case (3) is called Fick's Law], then (I) describes the diffusion
of solute. If u denotes temperature [in which case (3) is called the Newton-
Fourier law of cooling], then ( 1) describes the diffusion of heat. And in Chapter
1 we have mentioned the application of (1) to the random motion of motile
microorganisms.
When dust particles move in air they are subjected, as a result of collisions
with surrounding air molecules, to a vast number of tiny nudges. Paint part-
icles in water receive similar nudges from water molecules. When we speak
of the diffusion of dust particles or paint, we have in mind a spreading over
times that are very long compared with the typical time interval between
nudges and a spreading over distances that are very long compared with the
typical amount that a particle moves between nudges. Thus, in retrospect,
we can agree with the reasonableness of our finding that the diffusion equation
results from a random walk in which step size !ix and step interval fl.t are
regarded as small.
Heat diffusion is harder to perceive clearly. Molecular collisions should
certainly cause some sort of random change in the local kinetic energy, but
it is not obvious that heat "walks" in the same way that dust particles do.
It is helpful in this connection to consider Exercise 3.6, where the reader is
asked to show that the diffusion equation ( 1) results from a random walk
wherein with equal probability the particle moves to the left, moves to the
right, or stays where it is. This provides one illustration of a general feature:
the same limiting or macroscopicequationis obtainedfor a number of different
microscopicmodels. There is a blurring of microscopicdetail. Thus a theoreti-
cal unity is obtained by ignoring differences (say between molecular details
of heat and solute diffusion), which are in any case irrelevant for many pur-
poses.
102
We wish to emphasize that a random walk model of Brownian motion is

appropriate because of the vast disparity between the two scales involved.
To fix ideas, imagine that one is examining a randomly moving dust particle
through a certain magnifying glass M 1 . Only sufficiently large motions can
be resolved by M 1 ; so, too frequent observation of the particle is pointless.
With this in mind, suppose that one finds that it is sensible to plot the course
of the particle by determining its position every t 1 seconds. A more powerful
magnifying glass Mi might be utilized. Now one can profitably measure the
position of the particle every t 2 seconds, t 2 < t 1 A yet more powerful glass
M 3 reveals yet more detail and makes it appropriate to determine particle
position every t 3 seconds, t 3 < t 2
All these " macroscopic " time scales t; are very large compared with the
typical time -r between molecule-particle collisions. Thus one can pick an
intermediate time At such that -r ~At~ t;; i = I, 2, .... Based on /1t, one can
construct a random walk model of the dust particle motion in which a "step"
is defined as the typical displacement after a time interval of length At. Two
salient features of the random walk model are that successive steps are
regarded as independent and analysis proceeds via the diffusion limit. Succes-
sive steps can safely be treated as independent, since there is effectively no
correlation between events separated in time by At seconds. (By contrast, the
effects of successive nudges are not necessarily independent.) The diffusion
limit is appropriate, because many steps are required to produce a typical
displacement.
To reiterate, diffusion limits (or, equivalently, Stirling's approximation)
can be employed, because one can define a " step time" At that is small com-
pared with the macroscopic scales ti And the hypothesis of independent
steps can be used because At is large compared with the typical time -r between
individual molecular nudges.
A final point should be made. Ever increasing detail* will become manifest
as one selects successive magnifying glasses M;. The particle path will be
revealed as increasingly long and distorted, and hence the speed as in-
creasingly great. This is the reason why one can accept the infinite particle
velocity that is a formal concomitant of the Brownian model in the diffusion
limit.
SUPERPOSITION OF FUNDAMENTAL SOLUTIONS: THE

METHOD OF IMAGES
Equation (1), interpreted as describing the flow of heat, will be discussed
at some length in C~apters 4 and 5. Here we wish to illustrate the interplay
between probability and differential equations by considering the solution
u0 (x, t) given in (3.15) (and repeated here for reference),
If the "magnifying glass" is powerful enough to resolve the effect of a single nudge, then no
further detail is revealed by taking an even more powerful glass. But at this stage there is no longer
a separation of macroscopic and molecular scales, and the entire basis for the standard random
walk model has disappeared.
Sec. 3.4] Further Considerations Pertinent to Differential Equations
103
u0 (x, t) = (41tDt)- 1 2
'
-x2)
(4Dt
exp -- . (4)
Remember that this solution satisfied the "delta function" initial condition
(3.14) and also that the integral of u0 over the entire x axis was unity. We
can thus interpret u0 (x, t) as a unit source solution, which gives the amount of
mass that diffuses out at time t to a location between x and x + dx, assuming
that a unit amount of mass was initially concentrated at the origin. This
interpretation matches the random walk interpretation if we imagine the unit
mass as composed of a large number M of tiny particles. Then the probability
that a single particle is between x and x + dx gives the percentage of the M
particles that have this location.
The function u0 is also called the elementary or fundamental solution
to ( l ), since more general solutions can be built up from suitable combina-
tions of solutions like u0 To see this, we first observe that the diffusion
equation (1) is invariant when x and tare changed into x - eand t - . This
translation invariance is easily verified by direct substitution (Exercise 1).
Also, it is physically obvious, since the diffusion process is the same no
matter where we fix our origin of space measurement and no matter when
we start our clock.
Because of the translation invariance, u0 (x - e,
t - -r) is also a solution.
This new solution represents the effect at point x and time t (t > -r) of a source
introduced at point eat time ...
Next, we take account of the fact that the diffusion equation (I) is linear
so that the principle of superposition holds. That is, if u1(x, t) and ui(x, t) are
two solutions of(l), then
(5)
is also a solution. So is any finite linear combination of solutions. As an

example, for (1) we may consider the solution
(6)
which is a superposition of the effects of two sources, one located at e 1 and

e
one located at 2 The strengths of the two sources, i.e., the amount of mass
concentrated at e1 and e2, have the values C1and C2, respectively.
At the end of the previous section we sought a solution to (1) with a
vanishing x derivative at x = L. We can now solve this problem by super-
position, using the method of images. For if one unit source is placed at
x = O and another image source of equal strength is placed at x = 2L,
then the resulting mass distribution will always be symmetric about x = L. *
A function symmetric about x = L must have a zero derivative there, and
Note that the symmetry implies that exactly half the total mass will always remain to the
left of x =L. Thus the total "probability mass" in x :5,L is unity, as is required by (3.13a).
104
this is the required boundary condition. We thus obtain the required solution
u0 (x, t) + u0 (x - 2L, t) of Equation (3.21).
We leave to the reader the demonstration that the solution (3.23), where u
is required to vanish at x = L, can also be readily found by the method of
images. But we wish to point out how one can obtain the result (3.24) for the
first passage time F(L, t).
FIRST PASSAGE TIME AS A FLUX
The function F(L, t) was defined so that F(L, t) dt is the probability that a
particle which starts from the origin at time zero is absorbed at x = L in
the time interval (t, t + dt). An alternative interpretation of F can be given
in terms of a situation involving a unit amount of mass that is initially
concentrated at the origin, which is diffusing in the presence of a mass sink
at x = L. In this situation F provides the rate at which mass is leaving the
system at x ==L, by being absorbed in the sink there. In other words, F is
the flux of mass at x = L, in the direction of increasing x:
F(L, t) = J(L, t). (7)
But we know that the flux is proportional to the gradient of the appropriate
distribution, which in this case is the absorbing-barrier function u..4.(x,t)
of Equation (3.23). We at once obtain the desired result of Equation (3.24),
F(L,t)=-D- OU..4.)
( OX x=L .
More ingenuity would have been required to obtain this result from purely
probabilistic considerations [see Chandrasekhar's article in Wax (1954).]
GENERAL INITIAL VALUE PROBLEM IN DIFFUSION
To see further use of the source solution, let us consider the general one-
dimensional problem of solute diffusion. Imagine an infinite region in which
the solute is initially distributed in some manner:
u(x, t) = g(x) at t = 0, -oo <x< oo. (8)
How does the solute distribute itself? What is desired is a solution to the
diffusion equation (1) subject to the initial condition (8). We can obtain
quite a general distribution of mass by considering a large number of sources
ofstrengthJ(,,)a, located at points ,;=ia,, i=O, 1, ... , M. Now
u0 (x -ei, t) gives the mass that diffuses x - t units, in time t, from a unit
initial concentration. The total mass at x from all the sources is thus given by
M
L
i=-M
Uo(X - 'i' a,.
t)J(,;) (9)
Sec. 3.4) Further Considerations Pertinent to Differential Equations
105
In the limit as M-+ oo, ae-+ 0, this would become*

I>Uo(X-
-ox,
e, t)f(e) de. (10)
Roughly speaking, in (10), /(e) de

is the strength of the source located at
x = {. For the initial condition (8), then, we should take g = f We therefore
conjecture that the solution to the initial value problem is
u(x, t) = f OX)
-ox,
uo(x - e,t)g<e)ae. (11)
Formula (9) can also be obtained from a probabilistic point of view.

We know that u0 (x - e;,
t) provides the probability that a particle is found
x- e;units to the right of where it starts (at time t), given that the particle
has taken many random steps. If/(0 ae
particles started at then of these ei,
Uo(X- e;, t)/(e;) ae are expected at point X. Thus (9) sums the individual
contributions to the particles at x from the various initial concentrations at
e; ae.
= i We are again Jed to the conjecture that (11) provides the solution
we seek.
To verify that the function defined in (I I) does in fact solve our problem,
we should first substitute it into the equation (1). There is no difficulty if
g(e) vanishes outside a finite interval, for then differentiation under the
sign of integration is justified. Even for the infinite interval, the process of
formal differentiation is not difficult to justify. One merely has to show that
the infinite integrals, obtained after the formal differentiation process, are
convergent.
A more urgent problem is to verify that (11) satisfies the initial condition
lim
t+O
f >
-ox,
Uo(X- e,t)g(e) de= g(x). (12)
To accomplish this, let us introduce the variable '1 by
2 (x - e)2
(13)
'1 = 4Dt '
which is suggested by the exponential function in u0 (x - e,t). [Compare (4).)

Our solution (11) then becomes
1
J- f
OX)
+ 11.Jwt) a,,.
2
u(x, t) = e-" g(x (14)
1t -ox,
The strength of each individual source is made to approach zero in the limit. Otherwise,
an infinite number of sources would give rise to an infinite amount of mass. As it is, we have
adjusted the relation between source strength and number of sources so that the total amount of
mass remains finite as the number of sources become infinite.
If we now take the formal limit t -+ 0, we easily obtain the desired answer,
using (2.34).
In practical applied mathematics, the discussion of the previous paragraph
is often deemed sufficient verification that the initial condition is satisfied,
since justification of the formal limit seems straightforward. Presumably, all
one has to do is to divide the interval of integration in (14) into the three
parts ( - oo, -A), ( -A, A), (A, + oo), and take care with the double limiting
process t-+ 0, A-+ oo. (The reader may wish to carry out the proof.)
We have seen how the superposition principle enables us to obtain a variety
of useful solutions to the diffusion equation. Superposition will be treated
much more fully in the next two chapters. Here we continue our examination
of the use of differential equation approximations to probability problems.
HOW TWISTING A BEAM CAN GIVE INFORMATION ABOUT
DNA MOLECULES
A paper that is unusually rich in interconnections among various sciences
and various branches of mathematics is " Diffusion Out of a Triangle"
by W. Smith and G. S. Watson [J. Appl. Prob. 4, 479-88 (1967)]. We now
present a brief exposition of the content of this paper.
Sometimes separate strands of long helical DNA molecules are present in
solution. There are two types of such molecules, with complementary chemical
structure. Thermal agitation brings these molecules together from time to
time. It is presumed that only if complementary portions slide into apposition
will the two separate molecules successfully "zip up" into a double helix.
Otherwise, the molecules will eventually slide apart. We wish to calculate the
frequency of collisions (called "successful") in which the zip-up process
occurs.
The problem is idealized into a consideration of two rigid rods of unit
length. Let (x, y) denote a randomly chosen intersection point, 0 ~ x ~ 1,
0 ~ y ~ 1. Suppose that thermal motion of the solvent models leads to a
random walk of the point (x, y). Which will happen first, a coming into
register (x = y) or a sliding apart [(x, y) passing out of the unit square]?
By symmetry, an alternative form of the question is as follows? If a point
is dropped at random on an isosceles right triangle and then executes an
unbiased random walk, what is the probability that it will leave the triangle
along the hypotenuse (successful collision) rather than along one of the legs
(unsuccessful collision)?
More generally, let us consider the following problem. A plane domain D
has a boundary r that is composed of two parts r 1 and r 2 Let a particle
execute an unbiased random walk starting from a point (x, y) in D. Determine
P(x, y), the probability that a particle starting at (x, y) crosses the boundary
first through I' 1 *
* In terms of concepts that were introduced above, this problem can be regarded as requiring
the probability that the particle is ultimately absorbed along r 1 rather than along r 2
107
In setting out to solve this problem, imagine a square grid to be super-

imposed on D, with a grid spacing of 6. Suppose that the particle executes a
random walk on this grid (Figure 3.2). If the particle starting from (x, y)
eventually crosses the boundary along r 1 , then it must have moved to one
of the four points adjacent to (x, y) and then crossed the boundary. Thus
P(x, y) = t[P(x - 6, y) + P(x + 6, y) + P(x, y - 6) + P(x, y + 6)]. (15)
-
'--
,v r-.... ""'-.
~I
1
\ ~
D
r2 '- ... v
_...
- .:l
FIG URE 3.2. Diagram for a random walk problem. It is desired to

x
determine the probability that a particle which starts in the interior will
first cross the boundary on the heavy portion r 1
As has become our custom, we assume that the step length Li is small and
retain only the lowest order terms in an approximation of the governing
difference equation. After a short calculation (Exercise 7) we thereby obtain
from ( 15) the Laplace equation
a2p a2p
-+--0
ox2 oy2 -
(16)
The nearer the particle starts to r 1 (r 2 ), the more (less) probable it is that it will
first cross the boundary at r 1 This leads to the boundary conditions
Jim P(x, y) = I, lim P(x, y) = 0. (17a, b)

(x,yJ-+T1 (x,y)-+T2
Suppose that the particles are initially distributed in D according to the

probability distribution function f(x, y). That is (roughly speaking), let
f(x, y) dx dy be the probability that the particle is initially located between
x and x + dx and between y and y + dy. Then the probability P that a
particle will leave D along r 1 is given by the formula
P= fJP(x, y)f(x, y) dx dy. (18)

D
1o8 Random Processes and Partial Differential Equations [Ch. 3
We have reduced the DNA collision question to the solution of (16)

subject to the boundary conditions of (17), and then a substitution of the
result P(x, y) into the integral of (18). A transformation of the problem will
render its solution more tractable-as it is, the discontinuous boundary
conditions of (17) make the solution awkward.
The basis for the transformation is the symmetric Green'sformula
ff(QV P - PV Q) du= fr(Q ~: - P !;)ds,

2 2
(19)
D
in which o/on denotes exterior normal derivative. In (19) we take P to be

the function defined above, while Q is required to satisfy
Q =0 on r. (20)
With this, a short calculation [Exercise 7(b)] shows that
P={ P(:~) ds. (21)
If Q can be determined, our problem is solved by (21). But the problem

for Q is not only relatively simple compared to the problem for P, it is a
problem that appearsin many physical contexts. For example, it arises in the
flow of viscous flui down a pipe with cross section D (Exercise 3.2.14 of II)
and (when/ is constant) in the torsion of a prismatical bar(Section5.2ofll).
Using published solutions of (20) for constant f, Smith and Watson cal-
culated P = 0.41062 for the case when the particle is initially dropped at
random on an isosceles right triangle (the biologically interesting situation).
The details need not concern us. What we wish to emphasize is how classical
ideas of potential theory, and classical solutions in the theory of elasticity,
illumine ,considerations in the seemingly distant matter of a probability
calculation in biochemistry.
RECURRENCE PROPERTY IN BROWNIAN MOTION

We can now solve some elementary differential equations to answer an
interesting question in probability theory. The question is, if a particle starts
executing a random walk from a given point x, what is the probability that
it will eventually touch another given point y? The answer depends on the
dimension of the space in which the walk is executed. We shall treat the one-
dimensional case, but our discussion can be readily generalized to higher
dimensions.
Suppose that the particle starts at a point x in the set D described by
D: e < Ix:J< R.
By the one-dimensional version of the argument used to obtain ( 16) and
(17), the probability that a particle which starts at x will hit the boundary
109
IxI = e before the boundary IxI = R is the solution P(x) of the following
problem:
d 2P
dx 2 = 0, x in D; (22a)
P = I when IxI = e, P = 0 when IxI = R. (22b, c)
We seek a linear function that satisfies the boundary conditions. The result is
P(x)=R- lxl. (23)

R-e
As R --+ oo (for x and e fixed), P --+ 1. This means that the particle is virtually
certain to touch the interval IxI = e before it "escapes." Since the starting
point is arbitrary, we can conclude that a particle is virtually certain to
pass arbitrarily close to any given point on the line. And, by repeating the
argument, we conclude that the particle is virtually certain to do this arbi-
trarily often. We say that one-dimensional random walk has the recurrence
property.With the tools at our disposal, we can show (Exercise 9) that the
recurrence property also holds in two dimensions, but not in three.
The more "directed" character of diffusion in lower dimensions is a phe-
nomenon of importance in applications. To give an example, G. Adam and
M. Delbruck* have provided quantitative results along the lines we have dis-
cussed in an effort "to propose and develop the idea that organisms handle
some of the problems of timing and efficiency, in which small numbers of
molecules and their diffusion are involved, by reducing the dimensionality
in which diffusion takes place from three-dimensional space to two-dimen-
sional surface diffusion."
The recurrence property seems to threaten theoreticians, however, as well
as help them. It appears that random walk possesses a quasi-cyclic character,
and this seems entirely out of keeping with its use as a model for such ir-
reversible processes as diffusion. In fact, Poincare proved that there is a
stronger property of the recurrence type. A collection of interacting particles,
making up a dynamical system, will return arbitrarily closely to its initial
state with probability 1. (A few initial states must be excluded-the excluded
states forming a set of Lebesgue measure zero.) How, asked Zermelo, can
one reconcile the recurrence property of dynamical systems with the irreversi-
bility of continuum theories?
The resolution of the "Zermelo paradox" is based on the fact that the
mean time to recurrence is very long in situations that we regard as irreversible.
For example, it is true that if you lower a spoonful of cream into a cup of
coffee that the cream particles should eventually return to form a spoon shape,
In A. Rich and N. Davidson (eds.), Structural Chemistry and Molecular Biology (San
Francisco: Freeman, 1968).
but the expected time to do this is very large compared to the age of the uni-
verse. Thus there is no real contradiction with a continuum model for diffusion,
which predicts that the cream will spread more or less evenly throughout
the coffee in a matter of minutes. [For a more detailed exposition, see Kac's
article in Wax (1954).]
The discussion of the previous paragraph returns us to our original derivation
of the connection between random walk and diffusion. Both the approaches
from random walk to a diffusionlike result-whether by Stirling's formula
or by passage from a difference equation to a limiting differential equation-
both approaches gave approximations that were inaccurate for those rare
particles which almost always went to the right or almost always went to
the left. Misestimating the effects of very rare events will not have a perceptible
effect on over-all results, and this is why the diffusion model is so successful.
But as responsible thinkers, we must not lose sight of the limitations of our
models.
EXERCISES
1. Provide a formal verification of the fact that (1) remains the same if the
variables (x, t) are replaced by (x - ~' y - t), where ~ and t are
constants.
2. (a) Verify that the function uR of (3.21) has the properties claimed for
it in the text.
(b) Use the method of images to derive Equation (3.23). Verify directly
that u.. has the desired properties.
3. Provide a formal proof that (11) provides a solution to (l), providing
g is suitably restricted.
4. Complete the demonstration that taking the limit of ( 14) as t __.0 indeed
provides a proof that the initial condition ( 12) is satisfied by the function
u defined in (11). Assume that g(x) is a continuous function of x that
is bounded for all values of x (including \xi -+ oo).
5. Show that if g(x) has a discontinuity at a point x 0 , the limiting value
of the integral in ( 14) is
-![g(x 0 - 0) + g(x 0 + O)],
the average of the values obtained when approaching x 0 from the left
and right.
6. Using ( 11), write down the density u(x, t) of a solute if initially
u= 1 for O :s;x :s;1,
u=O forx<O and x> 1.
Express the answer in terms of the error function,
e= 2n- 1' 2 Lexp< -s 2 ) ds.

~
erf
Sec. 3.4] Further Considerations Pertinent to Differential Equations III
Then verify directly that the equations and boundary conditions are
satisfied.
7. (a) Expand (15) in powers of A and obtain (16). [Usually, in numerical
analysis, the difference equation (15) is used as an approximation
to (16). Here we work backward and approximate (15) by the
differential equation (16).J
(b) Verify (21).
8. Derive (19) from the divergence theorem HD
V v du= tr n v ds.
Begin by considering the two special cases v = QVP and v = PVQ.
9. t(a) The two-dimensional analogue of (22) requires finding a function
P of the polar coordinate r that satisfies
d 2P _ 1 dP
d,2 +r dr =0, e<r<R,
P(e) = I, P(R) = 0.
Find this function and use it to show that the recurrence property
holds in two dimensions.
t(b) Extend the above calculations to the three-dimensional case by
finding a suitable function P of the spherical coordinate p. Show
that e/p is the probability that a particle that starts a distance p
from the origin will ever approach to within e of the origin,
e < p. Deduce that the recurrence property does not hold in three
dimensions.
10. (This exercise presents an alternative approach to the diffusion limit
ofrandom walk.) Suppose that a large number of particles are executing
a one-dimensional random walk, with step length llx and a time
interval At per step. Let
N(x, t) = number of particles at point x, time t;
v(x, t) = net number of particles that move from x to x + llx in
the time interval (t, t + At).
Furthermore, define p and J by
- N(x, t) J(x, t) = v(x, t).
p (x, t ) - llx ' at
(a) What is the interpretation of p and J?
(b) Show that to first approximation ap/at = -aJ/ox.
(c) Justify the following assumption in unbiased random walk:
v(x, t) = tN(x, t) - tN(x + llx, t).
(d) By taking an appropriate limit, derive a relation between J and p,

and thereby show that p satisfies the diffusion equation.
11. Repeat Exercise IO for the case of biased random walk, where the
particle moves to the right with probability p and to the left with
probability q, q '# p, p + q = 1. (Compare Exercise 3.7.)
12. The motion of certain amebae is known to be affected by the concen-
tration c of certain chemical attractants. Let us idealize the ameba as an
organism that takes a step A to the right or to the left with an average
frequency f = f(c). Suppose that the organism has a chemical receptor
at each end and that the distance between receptors is ocA.Finally, let
a(x) be the density of amebae centered at x. (There is also a depend-
ence on time t, but we do not indicate this in our notation.)
(a) What is implied by the assumption that the flux J is given by
J =[
x-.i.
f[c(s + }aA)]a(s) ds - f x+.i.
"
f[c(s -1-aa)]a(s) ds.
(b) From the equation above, show that to lowest order in a,

da de
l= --+xa-,
dx dx
where
2
(c) =f(c)A , x(c) = (oc- l)'(c).
(c) Try to explain why x changes sign with oc- 1.
(d) Discuss the case a= 0. [Chapter 1gives some idea of the significance
of the "chemotaxis" phenomenon modeled here. For further de-
velopment along the lines just outlined, see E. F. Keller and
L.A. Segel, J. Theoret. Biol. 30, 225 (1971).]
13. How might the random walk concept be used in a quantatitive dis-
cussion of polymer molecules? (These molecules are composed of a
long chain of identical submolecules.)
Appendix3.1 0 ando Symbols

Definition1. cp(x)= O[lj,(x)]as x-+ x 0 if there exists a positive constant M
such that Icp(x)I $ Ml lj,(x) I whenever x is sufficiently close to x 0
REM ARK s. "</>= O(i/t)" is read "<pis big-oh of t/J."Most often x0 is
either zero or infinity. Reference to x 0 is frequently omitted if the x 0 is
clear from the context. Sometimes the phrase '' </> is of the order of t/1" is
used instead of</> = 0(1/1)if and only if </>It/I
= 0(1/1).If l/f(x0 ) '# 0, </> is bounded
in a sufficientlysmall neighborhood of x 0
Examples. (i) ef,(.x)

= 0(1) as x-+ xo means that <{,is bounded in the neighbor-
hood of Xo . (ii) x 2 sin x = O(x 3) as x-+ 0. (iii) x 2 tanh x = O(x 2) as x-+ oo.
(iv) The O(x') term in the Maclaurin expansion for 1/(1- 2.x) is (2x)' (x-+ O is
understood).
Appendix 3.1] 0 and o Symbols 113
Definition 1 is extended in various ways. Thus, for vectors, cj,(x)=

0[\),(x)] as X-+Xo requires that I cl>(x)I ~ Ml \),(x)I for Ix - Xol small.
Moreover, </>(x)= O[tf,(x)] for x in the region R means that there exists a
positive constant M such that I</>(x) I ~ Ml tf,(x)I for all x in R. Also,f(x, e) =
O[g(x, e)] as x-+ x 0 means that there exists a positive quantity M, which can
depend one but not on x, such that lf(x, e)I ~ M(e)lg(x, e)I whenever x
is sufficiently close to x 0 If an M can be found that is independent of a,
then we say that/= O(g) as x-+ x 0 , uniformly in e.
Definition2. q,(x) = o[tf,(x)] as x-+ x 0 if for every positive constant M

we have jq,(x)I ~ Mltf,(x)I whenever xis sufficiently close to x 0 In words,
" q,is small-oh of 1/J."
Particularly when x 0 = oo, the above definition is often altered so that there
is no requirement on q,and tf, when x = x 0 Then q,= o(tf,) if and only
if lim,.,
...,.,0 4>/1/1
= 0.
Examples. (i) x 2 In x = o(x) as x-+ 0. (ii) e-" = o(l) as x-+ co.
EXERCISES
1. Interpret and prove the following (Erdelyi, 1956, p. 8).
(a) O[O(cJ>)]= O(cJ>).
(b) O(cJ>)O('I')= O(cJ>'I').
(c) O(cJ>)+ O(cJ>)= O(cJ>)+ o(cJ>)= O(cJ>).
2. Make up and prove some more lemmas like those in Exercise 1.
CHAPTER 4
Superposition,Heat Flow, and
FourierAnalysis
THE MAJORITY of scientific problems that can be treated by standard

mathematical methods are linear. For linear homogeneous problems, the
principleof superpositionholds; i.e., if special solutions to the problem are
known, then a linear combination of such solutions yields new solutions.
Furthermore, the solution in the inhomogeneous case can be obtained by a
linear combination of solutions to the homogeneous problem and a special
solution of the inhomogeneous problem.
Perhaps the most important landmark in the development of the basic
theory for solving linear problems is the publication in 1822 of Fourier's
Theorie Analytique de la Chaleur (Mathematical Theory of Heat). Not only
does this book provide an exemplary formal treatment of a general class of
" boundary value problems," but it also serves to initiate the theory of a
mathematical approach possessing great generality. In its original and general-
ized forms, Fourier's approach is indispensable to modern science and engi-
neering. It is one of the most important landmarks in the development of
applied mathematics.
The early history of Fourier analysis is typical in its demonstration of
differences in emphasis between pure and applied mathematicians. E. T. Bell
sums up the situation this way.*
It was while at Grenoble that Fourier composed the immortal Theorie Analytique
de la Chaleur (Mathematical Theory of Heat), a landmark in mathematical physics.
His first memoir on the conduction of heat was submitted in 1807. This was so
promising that the Academy encouraged Fourier to continue by setting a contri-
bution to the mathematical theory of heat as its problem for the Grand Prize in 1812.
Fourier won the prize, but not without some criticism which he resented deeply but
which was well taken.
Laplace, Lagrange, and Legendre were the referees. While admitting the novelty
and importance of Fourier's work they pointed out that the mathematical treatment
was faulty, leaving much to be desired in the way of rigor. Lagrange himself had
discovered special cases of Fourier's main theorem but had been deterred from
proceeding to the general result by the difficultieswhich he now pointed out. These
subtle difficultieswere of such a nature that their removal at the time would probably
have been impossible. More than a century was to elapse before they were satis-
factorily met.
* In E.T. Bell, Men of Mathematics (N.Y.: Simon and Schuster, 1961),pp.197-98. Copyright
1937 by E.T. Bell. Reprinted by permission of Simon and Schuster, Inc.
114
Sec. 4.1] Conduction of Heat 115
In passing it is interesting to observe that this dispute typifies a radical distinction

between pure mathematicians and mathematical physicists. The only weapon at the
disposal of pure mathematicians is sharp and rigid proof, and unless an alleged
theorem can withstand the severest criticism of which its epoch is capable, pure
mathematicians have but little use for it.
The applied mathematician and the mathematical physicist, on the other hand,
are seldom so optimistic as to imagine that the infinite complexity of the physical
universe can be described fully by any mathematical theory simple enough to be
understood by human beings. Nor do they greatly regret that Airy's beautiful
(or absurd) picture of the universe as a sort of interminable, self-solvingsystem of
differential equations has turned out to be an illusion born of mathematical bigotry
and Newtonian determinism; they have something more real to appeal to at their
own back door-the physical universe itself. They can experiment and check the
deductions of their purposely imperfect mathematics against the verdict of exper-
ience-which, by the very nature of mathematics, is impossible for a pure mathe-
matician. If the mathematical predictions are contradicted by experiment they do
not, as a mathematician might, turn their backs on the physical evidence, but
throw their mathematical tools away and look for a better kit.
This indifference of scientists to mathematics for its own sake is as enraging to
one type of pure mathematician as the omission of a doubtful iota subscript is to
another type of pedant. The result is that but few pure mathematicians have ever
made a significant contribution to science-apart, of course, from inventing many
of the tools which scientists find useful (perhaps indispensable). And the curious
part of it all is that the very purists who object to the boldly imaginative attack of
the scientists are the loudest in their insistence that mathematics, contrary to a
widely diffused belief, is not all an affair of grubbing, meticulous accuracy, but is as
creatively imaginative, and sometimes as loose, as great poetry or music can be on
occasion. Sometimes the physicists beat the mathematicians at their own game in
this respect: ignoring the glaring lack of rigor in Fourier's classic on the analytical
theory of heat, Lord Kelvin called it "a great mathematical poem."
As Bell points out, Fourier's work was at first criticized for its lack of rigor.
Yet its importance as a new approach and a powerful method was undeniable
even from its very beginning, and it sparked developments in pure mathematics.
The solution of specific problems, and the subsequent stimulation to more
general developments, cannot be more clearly exemplified than in the study
of Fourier analysis, a subject that should be familiar to every student of
applied mathematics.
4.1 Conductionof Heat

Let us begin by considering a simple problem of heat conduction, involving
a metal rod, insulated on the sides, with one end in a bath at temperature
01 and the other end in a bath at temperature 02 We shall formulate a
mathematical representation for this problem and for its three-dimensional
generalization. Next we discuss some of the mathematical issues that arise.
JI6 Superposition, Heat Flow, and Fourier Analysis [Ch. 4
In particular, we prove a uniqueness theorem. We solve the one-dimensional

problem by the method of separation of variables, and we show how the use
of dimensionless variables facilitates interpretation of the solution.
STEADY STATE HEAT CONDUCTION

In the steady state, the temperature 8 in the rod will be found to have a
uniform gradient, and the rate of flow of heat will be proportional to this
gradient. Thus, suppose that the end point of the rod located at x = 0 is at
temperature 81 and the end located at x =Lis at temperature 82 (see Figure
4.1). Then the temperature at a distance x from one end is
(1)
(_ -----i-1-!--..w(~MeaA
I I
I I
I I
I I
I I
I I
FIGURE 4.1. One-dimensional heat conduction along a rod.
(The subscript s is added to denote the steady state.) The temperature

gradient is
(2)
The heat flux density J (the rate of heat flow per unit area per unit time, in the
direction of increasing x) is found to be proportional to the temperature
gradient. We write
d8s
J= -k-, (3)
dx
where k is the coefficient of heat conductivity. The negative sign is introduced
so that k will be positive if (as expected) heat flows in the direction of de-
creasing temperature. Thus, with k > 0, J is positive in (3) when d8/dx is
negative.*
An enormous amount of work has gone into measuring various "phenomenological co-
efficients" like k. For a summary of a nearly one thousand page compilation of results on pure
substances alone, see "Thermal Conductivities of the Elements" by R. W. Powell and Y. S.
Touloukian, Science 181, 999-1008 (1973).
Sec. 4.1] Conduction of Heat
II7
DIFFERENTIAL EQUATION FOR ONE-DIMENSIONAL HEAT
CONDUCTION
Now consider the adjustment of the temperature distribution before the
steady state is reached. The temperature at any point x at any instant tis now
denoted by 6(x, t). Can we find an equation governing the function O(x,t)?
We assume that (3) still holds for each instant, so that the instantaneous
heat flux density is given by*
J= -k-.
06 (4)
ox
Consider a section of a cylindrical rod between x and x + Ax (Figure 4.1).
The heat flow across the surface at x into this section is at the rate
A(-koO)
ax
, ,c
(5)
where A is the (constant) cross-sectional area. The subscript x indicates that

the expression within the parentheses is to be evaluated at x.
The heat flow across the surface at x + Ax out of the section has the same
form as (5) except that the expression is evaluated at x + .::ix.Thus the net gain
of heat is at the rate q, where
4= A [(kOX
ao) -(kOX
ao)].
x+Ll.x ,c
(6)
If no sources or sinks of heat are present, then the net gain of heat per unit
mass is proportional to the rise of the temperature e.Thus
. a
q = ot[cOp(AAx)], (7)
where p is the density of the material and c is its specificheat. (A unit increase
of temperature increases the amount of heat in a unit mass of material by
c units.) If we equate (6) and (7), divide by Ax, and take the limit as .::ix--+0,
we obtain
pc ao= !_ (k06) (8)

a1 ax ox
It has been implied that the rod is composed of a homogeneous material and
therefore p, c, and k are constants; but since (8) is a local equation, it holds
Equation (4) is an example of a constitutive equation. Such equations describe the behavior
of a particular material, in contrast with universal physical laws, such as conservation of mass,
which describe the behavior of all materials. To see that (4) is not universally valid, note that if
we were dealing with a liquid, we would have to add to the heat flux density J a contribution
due to convection. It is not an easy matter to establish constitutive equations. We shall discuss
many such equations in the course of this work.
118 Superposition, Heat Flow, and Fourier Analysis [Ch. 4
even for an inhomogeneous cylindrical rod (when p, c, and k are functions

of x).
In the case of a homogeneous material we may write (8) as
ao 02 0
Ot = K ox2 ( 9)
Here K = k/pc is usually called the coefficientof thermal diffusivityin order to

recognize the similarity of (9) with the equation of diffusion of matter. Typical
values of Kare 1.14 cm 2 /sec for copper and 0.011 cm 2 /sec for granite.
INITIAL BOUNDARY VALUE PROBLEM FOR ONE-DIMENSIONAL

HEAT CONDUCTION
Having found the equation governing O(x, t), we now turn to the complete
formulation of the problem.
We are dealing with the solution of (9) in a domain D in the (x, t) plane
(see Figure 4.2) defined by
D:O<x<L, 0<t < 00.
Curve C
Fm URE 4.2. Mathematical formulation of the problem of heat conduc-

tion: the domain, its boundary, and the boundary conditions.
The boundary of this unbounded domain is a (heavily drawn) curve C that is

concave (upward). From physical considerations, it is clear that the initial
temperature of the rod must be known in order to define the problem. Let
this be
8(x, 0) = g(x). (IO)
The ends of the rod are kept at fixed temperatures for all t > 0 so that
0(0, t) = 01 , O(L, t) = 02 , fort> 0. (11)
The rod may have been exposed to the baths beginning only at t = O; therefore
g(O) and g(L) need not be equal to 61 and 62 , respectively.
Sec. 4.1] Conduction of Heat II9
Our mathematical problem is now formulated. To emphasize this, let us

repeat the equations [(9), (IO), and (11)]; find O(x, t) for O < x < L, 0 < t < co,
where
ae a e 2
for O < x < L, 0 < t < co.
ot= K OX2
O(x, 0) = g(x) (g given) for O < x < L.
0(0, t) = 01 , O(L, t) = 02 for t > 0.
In deference to the physical interpretation of the problem, the conditions

(11) are often referred to as the boundaryconditions(being imposed at the
boundary of the rod), and condition (IO) is referred to as the initial condition
(being imposed at the initial instant). Mathematically, they are both conditions
prescribed on the concave boundary C of the domain D. A close look at theory
(not presented here) is required to clarify this confusion in terminology.
PAST, PRESENT, AND FUTURE

Suppose that after the evolution of heat distribution in the bar has been
allowed to go on for some time, the temperature at the left end is changed to
01 at the instant t 1 , t 1 > 0. Physical experience indicates that the temperature
distribution in the rod will be influenced in the future (t > t 1) but not in the
past (t < t 1). Thus the equation must have a "sense of time," physically
speaking. Indeed, it is physically clear that the solution in the domain
D1 = {(x, t) IO< x < L, 0 < t < t 1}
is entirely determined by the conditions specified on the heavily marked curve
C 1 of Figure 4.3. Changes in the specification of 8(0, t) and 8(L, t) for t > t 1
can have no influence.
e =e;
e =11;
0 8-=g (x) L
---x
FIG URE 4.3. Difference between past and future illustrated by a
change of boundary conditions. Fort< ti the solution is only affected by
changes along Ci, the heavily marked portion of the boundarr.
Superposition, Heat Flow, and Fourier Analysis [Ch. 4
120
The solution in the domain D 2 is defined by a similar specification of

conditions on a concave curve. Thus we may make the conjecture that the
solution of the equation (8) in a domain D bounded by a concave curve C and
a line t ==constant (see Figure 4.4) is determined by specifying the value of
(} on C. This is indeed a correct result: The unique existence of the solution
can be proved for a sufficiently smooth curve C and continuous prescribed
values.
i
;.
I
'
FIG URE 4.4. General mathematical formulation of the problem of heat

conduction: the domain and its boundary.
HEAT CONDUCTION IN THREE-DIMENSIONAL SPACE

It is not difficult to generalize the formulation of the heat conduction
problem to cover the case of three spatial dimensions. As we shall show
in a moment, the final equation governing the change of the temperature
O(x, y, z, t) at the point (x, y, z) and time t is
oO= ,c(o
2 2
0 + 0 0 + 0 8')
2
or (12)
ot ox oy oz2
2 2
in the case of a homogeneous material. [Equation (12) generalizes (9).] In

the case of an inhomogeneous medium, we have the following generalization
of (8):
pc 08 ==~ (koe)+ ~ (k08) + I (kae).

ot ox ox oy oy oz oz (13a)
In vector notation, (13a) is

08
pc ot= V (kV8). (13b)
If the temperature over a surface enclosing a given volume is prescribed for

all time and the temperature inside is given at the initial instant, we would
Sec. 4.1] Conductionof Heat
121
expect the solution to be completely specified. Again, this is true, but the
proof of the existence theorem is not easy.*
The derivation of (13) can be accomplished by considering a rectangular
box with two of its corners at (x,y, z)and (x + ax,y + Ay, z + Az) and using
substantially the same reasoning as before.t [See Exercise 2(b).] The deriva-
tion can also be accomplished by applying the divergence theorem to the
fallowing equation for the rate of heat increase in a region R bounded by an
arbitrary smooth closed surface oR:
frfffcp8d1: = ~ k :: da = ~ kn VBda. (14)

R 8R /JR
In (14) n is the unit outward normal. This equation uses the assumption that
the heat flux through an element of surface in the direction of n is -k(o8/on)
where o/ondenotes the normal derivative [compare (4)]. From (14),
fff[cp :: -V (kV8)] d1:= 0. (15)

R
Since R is arbitrary,the integrandmust be zero if it is continuous.Otherwise,

near a point P O , where the integrand does not vanish, we must be able to
find a small region R 0 in which the integrand has the same sign as its value at
P0 The integral (15) cannot vanish when R = R0 , leading to an absurdity.
The complete formulation of a problem depends not only on the derivation
of the equation but also on the specification of initial and boundary condi-
tions. One common boundary condition is
(a) The temperature is specified for all time at each point on the bound-
ary. (It may vary from point to point.)
Another possible boundary condition is
(b) The boundary is insulated, so that o9/on = 0 at every point on the
boundary.
A third and more general condition is
(c) The heat loss from the boundary is proportional to the amount by
which the surf ace temperature is higher than the ambient tempera-
ture.
Particularly noteworthy existence proofs, for the wave and Laplace equations as well as
the heat equation, arc supplied in a remarkable paper by R. Courant, K. Friedrichs, and H. Lewy
["On the Partial Diff"crential Equations of Mathematical Physics," translated by P. Fox in
IBM Journal II, 215-34 (1967); the original paper appeared in the German journal Math. Ann.
100, 32- 74 (1928)]. Existence emerges automatically from proofs that the solutions to appropriate
finite dift"crenceschemes approach limiting values, as the mesh size decreases, which satisfy given
partial differential equations and initial boundary conditions. Connections with Green's functions
and random walk arc also explored.
t In keeping with the "broad-brush" approach of the first three chapters, derivations arc
merely outlined here. For a more careful treatment of the required ideas, sec Chapter 14,
particularly Section 14.l.
This is the Dubois Reymoncl lemma, a result that we shall use many times.
There are many other possibilities such as the prescription of the time vari-
ation of the temperature over the bounding surface or a portion thereof. It is
difficult to cite all the possible cases, and it is much harder to prove all the
necessary existence theorems, as required by the approach of the pure math-
ematician. For the applied mathematician it is sufficient to adopt as a working
hypothesisthat one condition analogousto (a), (b), or (c) may be specified on the
boundary. The validity of this hypothesis will be accepted unless a reasonable
doubt can be demonstrated. Indeed, this hypothesis will be accepted for
an inhomogeneous medium as well, and even for cases where the specific heat
(say) is a function of temperature.
In general, it would be satisfying to have the unique existence of the
solutions proved under prescribed conditions; but this is not always possible.
Often, the applied mathematician obtains the conditions to be prescribed from
consideration of the scientific problem. He sometimes checks the appropriate-
ness of his equations by proving that they have at most one solution (unique-
ness). (If more conditions are prescribed than are necessary for uniqueness,
either these extra conditions are redundant or they cannot be fulfilled.)
Although the existence of solutions cannot be taken for granted, proofs are
surprisingly difficult. Applied mathematicians often do not consider the gain
in confidence that these proofs furnish to be worth the effort.
Since uniqueness is easy to demonstrate, we now present the proof. Then
we state and comment upon another general result, the maximum principle.
PROOF OF THE UNIQUENESS THEOREM

Consider the problem of heat conduction in a region R with prescribed
temperature on the bounding surface S. Let 81 (x, y, z, t) be a solution
satisfying the initial condition 8 1 = g(x, y, z) at t = 0. To prove uniqueness,
let us assume that there is another solution 8i(x, y, z, t) satisfying the same
condition. Then their difference 8 = 82 - 81 would also satisfy the heat
equation (13), but with the following boundary condition and initial
condition:
8 = 0 on S, (16a)
8 = 0 at t = 0. (16b)
Uniqueness is proved if(} can be shown to be identically equal to zero. To
show this, we first multiply (13) by 8 and integrate over the region R. We find
[Exercise 5(a)] that
-i
2
fffpc --a(e2) a8
at d-r = -e + k(} -anda ' (17)
R i1R
where
8 = ffI k(V8) d-r.
2
R
123
The surface integral vanishes because of (16a). Since pc is independent oft,

we may integrate (17) with respect to t in the interval (0, t 1) and obtain
(t fff 2
pc() d-r:)
t,
- 0= - f' 0
e dt (18)
R
if we recall the initial condition (16b). The left-hand side of (18) is non-
negative and its right-hand side is nonpositive, if t 1 > 0. Thus both sides must
be equal to zero. Since t 1 is arbitrary we must have
e = 0, at all times. (19)

Thus V() = 0 at all times, and hence ()(x, y, z, t) can only be a function of time.
The vanishing of the left-hand side of (18) then shows that [()(t1)] 2 = 0, and
hence () = 0. 0
THE MAXIMUM PRINCIPLE

An important result in the theory of partial differential equations can be
conjectured from the intuitive feeling that heat is never intensified as a result
of diffusion. That is, we expect temperature in the interior of a domain never
to exceed the largest of the temperatures that occur on the boundary or are
assigned initially. With some ingenuity, one can indeed prove the following
result. If ()(x, t) is defined and continuous for O :S x :SL, 0 :S t :S T, and if ()
satisfies the heat equation (9) in the interior of this rectangular region, then ()
assumes its maximum and minimum when t = 0 or when x = 0 or when x = L.
(We state our results in one dimension for simplicity, but they are valid for
higher dimensions, too.) For a proof, see texts on differential equations, e.g.,
Tychonov and Samarski (I 964, pp. 165-67).
With the aid of the maximum principle, proofs of uniqueness and con-
tinuous dependence on initial and boundary conditions are easily con-
structed (Exercise 11). Bear in mind, however, that the maximum principle
has been stated only for solutions that are continuous in the closed rectangle,
so one is permitted to make deductions from this principle only for the class
of solutions that are so restricted. This restriction does not apply to the
uniqueness proof that was given above, which employed the energy method.
As has been illustrated to some extent, a major task in the theory of partial
differential equations is to prove that the solutions to various problems exist,
are unique, and depend continuously on initial and boundary data. Such
problems are said to be well posed.
SOLUTION BY THE METHOD OF SEPARATION OF VARIABLES

There is no obvious way to approach the basic heat conduction problem
posed above, namely, solution of the heat conduction equation (9), the
initial condition (10), and the boundary conditions (11). The ingenious method
used by Fourier is based on the idea of superposing an infinite number of
appropriate simple solutions to the linear equation (9). The key to his method
is the determination of an infinite number of such solutions by the technique
of separation of variables.
For reasons that we shall give below, in problems of the type we are
investigating, the Fourier method applies directly only to the part of the
solution that decays to zero, leaving the steady state solution e.
Let us denote
this decaying transientby v(x, t); i.e.,
v(x, t) = O(x, t) - Os(x). (20)
e.
Since is a steady solution of the heat equation (9), the transient satisfies the
same equation:
ov o2 v
-=1'- (21)
ot ox2
The initial condition is now
for O <x <L, v(x, 0) = f(x); f(x) = g(x) - e.(x). (22)
The boundary conditions (11) become
v(O,t) = v(L, t) = 0 for all t > 0. (23)
Note that, unlike (II), these boundary conditions are homogeneous, in the
sense that they are satisfied by the zero function.
We follow a step-by-step procedure to solve the problem posed by (21), (22),
and (23). (Exactly the same procedure is used to solve a related problem in
Section 15.4. Taken together, the two illustrations should provide an adequate
exposition of the method.)
STEP A. Assume a product solution:
v(x, t) = X(x)T(t). (24)
STEP B. Substitute into the governing differential equation. From (24) and
(21),
XT' = ,cX"T (
X' = dX T' = <!\. (25)
dx' di)
STEP c. Separate variables. We write (25) in the form
X"(x) T'(t)
(26)
X(x) = K:T(t).
The variables are now separated, so that the left side is a function only of the
independent variable x and the right side is a function only of the independent
variable t. It appears that each side of (26) can be varied more or less at will,
by altering the independent variables x and t. If this were truly the case, a
125
contradiction would arise. Thus both sides of (26) must equal a separation
constant k, so that we obtain the separation equations
X"=kX, T'=kKT. (27a, b)
STEP D. Determine permissible values of the separation constant from a

problem consisting of a separated equation and suitable homogeneous boundary
conditions. The homogeneous boundary conditions are provided by (23).
They give
X(O)T(t) = 0, X(L)T(t) = 0 for all t > 0,
or (excluding T =0)
X(O) = X(L) = 0. (28)
These conditions must be applied to the solution of (27a). As is usually the

case, it is helpful to give separate consideration to the possibilities k > 0, k = 0,
k<O.
If k > 0, we write k = 2 , > 0. The general solution of the resulting
equation X" =X 2 is
X = C1 cosh x + C2 sinh x.
The boundary conditions (28) require that C1 = C2 = 0, so that X = 0.
If k = 0, we must solve X" = 0 subject to (28). In this case, too, there is only
the trivial solution X = 0.
If k < 0, we write k = - rx2 , rx> 0. The general solution of the resulting
equation X" = - cx2 X is
X = C1 cos rxx+ C2 sin ixx. (29)

The requirement X(O)= 0 is only satisfied if C1 = 0. The requirement X(L) =
O gives
C2 sin cxL= 0. (30)
One possible deduction from (30) is C2 = 0. But since C1 = 0, this would

again provide a trivial solution. Nontrivial solutions result only if the separa-
tion constant k = - a.2 takes on certain special values. Such solutions are
permitted, in the light of (30) and since ex> 0, only if
mn
rx=- m = I, 2, 3, .... (31)
L'
With (31) we indeed obtain nontrivial solutions (29) of the form
. mnx
(
Xix)= Bmsm L' Bmarbitrary constants. (32)
STEP E. Solve the remaining separated equation when the separation

constant takes on the values determined in (d). With (31), since k = - 1X2 , (27b)
becomes
m2n2
T'= -7 KT. (33)
Every solution of this equation is a constant multiple of

m2n2 )
exp ( -vKt. (34)
STEP F. Superpose all possible product solutions and attempt to satisfy

the remaining initial or boundary conditions. "Superpose" means (roughly
speaking)" add together." Thus we assume for v the infinite series
00 2 2
v(x, t) = I rnnx exp ( - -m n-
Bm sin -- Kt
)
. (35)
m=t L L2
Since the heat equation (21) is linear, a.finite series of the form (35) would
certainly provide a solution (since each term is a solution). Under "suitable
conditions" infinite series have the same properties as finite sums; we proceed
under the assumption that such conditions will hold here, at least in some
interesting instances.
The boundary conditions (23) are homogeneous so that the series (35)
should satisfy (23), since each term does separately.*
The initial condition (22) is not satisfied, but we have all the constants Bm
at our disposal. If we attempt to satisfy this condition, we have
00
mnx
f(x) = I
m=1
Bmsin-,
L
0 <x <L, (36)
as a necessary condition. From this we can proceed formally to find the value
taken by a typical coefficient, say Bn, if we multiply by sin (nnx/L) and
integrate from O to L. It can be shown (Exercise 8) that Bn is given by
2 fL . nnx
Bn = L f(x)smLdx. (37)
0
It is now evident why we chose to apply the Fourier method to the transient solution v; for
v satisfies the homogeneous boundary conditions (23), but the full solution (Jsatisfies the inhomo-
geneous boundary conditions (11). The latter conditions do not have the linearity property of
being satisfied by the sum of functions that individually satisfy them. Note that the homogeneity
of the X equation (27a) and the accompanying boundary conditions (28) meant that the trivial
solution was always a possibility. Our efforts to obtain a nontrivial solution led to the deter-
mination of a discrete infinite set of permissible separation constants and thus to the infinite
series (35)--another illustration of the key role played by homogeneity. Beginners often forget
that before standard application of the Fourier separation of variables procedure (a)-(f) can be
made, it must be arranged that (with two independent variables) one of the separated equations
has homogeneous boundary conditions. (The general rule is that " all but one" of the separated
equations must have homogeneous boundary conditions.)
127
This follows from
. mnx . nnx) = { L/2 ifm = n, (38a)

( sm y,smy
mnx nnx)
( cosy,cosy O if m # n, (38b)
L mn:x . n1tx
I
-L
cos --
L
sm -
L
dx = 0, (38c)
where we use the notation*
(J, g) = f f(x)g(x)
L
0
dx. (39)
With this notation, (37) takes the following form (where mis used instead ofn)
Bm = 2L - i (J(x),
sin m;x)
(40)
We have thus found a tentative solution to our problem as
O(x, t) = Os(x)+ ~ Bm(sin 7x)e-<m,./L) 2

xr, (41)
where the Bm are determined from the initial condition (22) by means of
Equation (40). Is (41) really a solution? To give a positive answer we must be
sure that it is possible to express the function f(x) in the form (36). This
would certainly be true if f(x) were the sum of a finite number of sinusoidal
terms. Indeed, one can express any "reasonable" function in such a form,
but the proof is not easy. In overcoming this difficulty to a specific problem,
we shall open the door to a general mathematical theory. But before we go
into this matter, let us examine some implications of the solution obtained.
INTERPRETATION; DIMENSIONLESS REPRESENTATION
Our solution (41) approaches the steady state temperature distribution

Os(x)given by (1), as expected. Of the various terms in the infinite series, the
fundamental(m = 1) decays much more slowly than the harmonics(m ~ 2).
Thus the deviation from the equilibrium distribution soon has a nearly
sinusoidal form, no matter what the initial temperature distribution may be.
The amplitude of this sinusoidal term decreases exponentially in time. The
Later we shall see how ( , ) can be regarded as a generalization of the scalar product of two
vectors, but this need not concern us at present. Only (38a) is needed to demonstrate (37), but
(38b) and (38c) are needed below.
exponent K(n/) 2 depends only on the length of the rod L and the thermal
diffusivity K of the material.*
For this problem, times should be compared to the time scale
L2
t0 = -, (42)
K
during which significant temporal change in the solution occurs. Significant

spatial changes take place in a fraction of the length scale L. Measuring
according to these scales is equivalent to introducing the variables
x t Kt
e=-
L
and t=-=2
t0 L
(43)
These variables are dimensionless in that their value is independent of the

unit of measurement. (For example, if x = IO cm and L = 100 cm, then
e = 110 , as it is if x = 0.1 m and L = 1 m.) Using these variables, we may take
our tentative solution (41) to be E>(e,t), where
co
E>(e,-r) = O(Le, tot)= (}.(Le)+ L Bmsin mne e-<mn) 2
', (44)
1
an equation that shows no explicit dependence on Land K. As a matter of fact,

this introduction of dimensionless parameters can be done from the beginning.
Then (9) becomes [Exercise 8(b)]
ae a2 e
~= ae 2 (45)
The equation is to be solved in the domain D' defined by

D': 0 < e< 1, 0< 't < 00. (46)
The initial and boundary conditions for the homogeneous problem are
0(0, T) = 8(1, T) = 0,
E>(e,0) = G(e), where G(e) = g(Le).
In the original formulation, the temperature at a point with coordinates
(x, t) seemingly could depend in an arbitrary way on the two parameters ,c and
L. The dimensionless formulation shows that explicit dependence on ,c and L
e
disappears if and t are used as independent variables. Such a reduction of
the number of parameters on which a solution explicitly depends generally
follows from the use of dimensionless variables.
The dimensionless formulation can be used for making predictions via the
use of experimental models. Thus, if we construct a scale model for testing
The most slowly decaying harmonic (m = 2) decays like exp (-41r 2Kt/L 2), a factor that
decreases to a fraction e- 4 (i.e., less than 2 per cent) of its initial value in L 2 /.,.2K ::::: 0.1 L 2 / K time
units. For a copper bar 10 cm long, this would be less than 10 sec.
Sec. 4.1] Conductionof Heat
129
a practical problem of heat conduction, using the same material, the time
scale in the actual case would be N 2 times as long if the model were 1/N of
the actual size. The reason is that the time t occurs only in the combination
-r = 1ct/L2 A given value of -r is reached in a model experiment, where L is
replaced by L/N, only if t is replaced by N 2 t. (See Section 6.2 for further
discussion concerning the use of models.)
These arguments can be applied straightforwardly to the three-dimensional
equation (12) and appropriate boundary conditions and initial conditions.
Consequently, for example, if we throw hot copper spheres into icy water,
it would take four times as long to cool down a sphere that is twice as large
in diameter (Exercise 4).
We emphasize that in dimensionless form the results are independent of the
units used for our measurements. The dimensionless formulation, therefore,
gives the true mathematical essence of the problem. For a fuller discussion of
dimensional analysis, see Section 6.2.
ESTIMATE OF THE TIME REQUIRED TO DIFFUSE A GIVEN

DISTANCE
To conclude, we deduce a fundamental fact concerning diffusion. This is a

consequence of the result that the slowest decaying mode, with spatial
dependence sin (1tx/L), decays like exp ( -TC 2 Kt/L2 ). Since the heat in this mode
is concentrated in the middle of the bar, we can regard the decay of the initial
heat distribution, roughly, as the passage of heat out of the bar through a
distanced= L/2. In terms of d, the decay factor is exp ( -TC 2 Ktj4,i2). In a time
t = d 2 /K the sinusoidal temperature distribution decays to exp (-1c2/4) ~ 8
per cent of its initial value. That is, diffusion through a distance d takes
roughly d 2/K time units. An alternative statement is in time t a substance
diffuses approximately (Kt)112 distance units.
EXERCISES
1. The general initial value problem for heat conduction is formulated for a
domain D bounded by a concave curve C and a line parallel to the x axis
(a characteristic curve), with the value of temperature specified on the
initial curve C (Figure 4.4). Explain why this formulation is reasonable,
using physical arguments.
2. (a) Check that the one-dimensional heat equation (8) holds for an
inhomogeneous rod.
(b) The equation of heat conduction in three dimensions may bederived
by using the same arguments employed in the text for deriving
the equation for the one-dimensional problem. One considers the
amount of heat flowing into a rectangular box through the six
sides. Carry out this derivation.
3. By using the method used in the text for proving the uniqueness theorem,
carry out the proof when the boundary condition is
8()
on+ h() = 0, h >0,
on the surface S enclosing the volume occupied by the conductor. Let

the initial temperature distribution be specified as in the text.
4. Formulate the problem of the cooling of a copper sphere thrown into
icy water. Show that it takes four times as long to cool down a sphere
that is twice as large in diameter.
5. (a) Verify (17).
(b) The expression in (34) could be preceded by the arbitrary constant
C. Show that this would introduce no new generality into the series
solution (35).
6. Solve the one-dimensional heat conduction problem by Fourier series
when the right end of the bar is insulated so that the second boundary
condition of (11) is replaced by the condition that 88/ox vanishes at
x=L.
7. Suppose that both ends of the bar are insulated. Then there is no longer
a unique steady state solution. Discuss this fact.
8. (a) Verify (37) and (38).
(b) Verify (45).
9. Consider the temperature () in a semi-infinite slab z ~ 0. Suppose that
() = 0 for time t < 0, but that there is an impulsive change in temperature
so that when z = 0, () = U (a positive constant) for t > 0. Find O(z,t).
Write the answer in terms of erfr, = 2n- 1 ' 2 Jg exp (-s 2 ) ds (so that
erf oo = I).
One expects the temperature to diffuse outward a distance (Kt)112 in
time t. Furthermore, (Kt)112 is the only other combination of variables
and parameters other than z, which has the dimension of a length. For
both these reasons, look for a solution to the diffusion equation of the
form
O(z, t) = f[z(Kt)- 112 ].
10. Make a change of variable of the form v = f(x, t)O that transforms the
differential equation
a, /3,y constants
into the standard form (9).
11. Consider the class C of functions that are continuous in the closed
rectangle x e [O,L], t e [O,T] and satisfy the heat equation (9) in the
interior of this rectangle.
(a) Using the maximum principle, prove uniqueness in this class for the
usual initial boundary value problem.
Sec. 4.2] Fourier's Theorem
131
(b) Prove that if two functions from C differ by less than e on the three
boundary lines x = 0, x = L, t = 0, then they differ by less than
e throughout the rectangle.
4.2 Fourier's Theorem

We have shown how the initial value problem for the heat equation can be
formally solved if any reasonable functionf(x) can be represented by a Fourier
sine series [Equation (1.36)] in the interval O < x < L. Since the series gives
zero as its value at the end points of the interval (0, L), it would appear natural
to expect a function like
to be so represented, but not a function like /{x) = 1. However, we shall see

that even the latter is possible in a suitable sense-indeed in more than one
sense.
We shall now justify the above statement about Equation (1.36); we shall
consider the limit of the sum of N terms of the series (1.36) and try to evaluate
the limit as N-+ oo. We shall see that this limit tends tof(x) in the open inter-
val (0, L), provided thatf(x) is piecewise smooth in the interval (0, L). This
smoothness requirement holds if the interval (0, L) can be divided into a
finite number of subintervals such that f(x) approaches a (finite) limit as x
approaches the end points of the subinterval and such thatf'(x) exists and is
continuous within each subinterval.
Since sines are odd functions, if a sine series representsf(x) on (0, L) it will
represent -f(x) on ( - L, 0). Sines and cosines must be used if a general
function is to be represented on the "full range" ( -L, L). As we shall see,
the convergence results can be readily extended to the full range situation.
SUMMATION OF THE FOURIER SINE SERIES

Let us denote the sum of the first Nterms of the series (I.36) by SN(x)so that
SN(x) = -2 J"KN(x, l;)f(l;) di;,

1l: 0
O<x<n, (1)
where
N
KN(x, ?;)= L sin (mx) sin (ml;), O<x,l; <n. (2)
m=l
We want to prove that SN(x)-+ f(x) as N-+ oo. For conveniencewe have set
L = n, and this will be done throughout the present section. This entails no
loss of generality, for it is equivalent to choosing n- 1L as a length scale in
nondimensionalization.
The summation of (2) can be carried out with reasonable ease if we use the
complex representation
eiw - e-iw
sinw=----
2i
The answer (Exercise 1) is
KN(x, ,) = ! {sin [(N + -!-)(x-

2 2 sin [t(x - e)]
rn_sin2 sin +[t(x})(x+ +e)]rn}. (3)
[(N
If x is a fixed point inside the interval (0, 1r), the first term within the curly
brackets on the right-hand side of (3) is well defined for O :::;;;
, :::;;;
1rexcept when
, = x; but even here, this term has a definite limit N + -!--The graph of this
term is shown schematically in Figure 4.5. In a very narrow interval Ix - eI <
n/(N + t), the function has a very high peak. Outside this interval, it shows a
highly oscillatory behavior.
The second term on the right-hand side of (3) also shows a highly oscil-
latory behavior for large N but no strong peak. Keeping the behavior of the
two terms making up KN in mind, we now consider SN(x), where
0<x < 7t, (4)
FIG URE 4.5. Schematic graph of the function

sin [(N + i)(x - g}]/2 sin [i(x - g)].
Sec. 4.2] Fourier's Theorem 133

with
s~P(x) =!I" sin [(1: + t)(x - rnf(f,,) df,,, (5)

1t O 2 sm t(x - f,,)
siAx)=_! J"sin [(1: + t)(x + rnJ<f.>

at.. (6)
n O 2 sm t(x + f,,)
Because of the highly oscillatory nature of the integrand in (6), we can expect
the integral to approach zero as N-+ oo. For when N is large, over each period
of the sinusoidal function in the numerator the other factors are practically
constant, and hence there is essentially no contribution to the integral from
each period. In (5) a similar argument holds except when f,,is near x, where
(roughly speaking) the large factor N + t is "neutralized" by the small
factor x - f,,. Thus the behavior of f(f,,) is important only when f,, is near
x. We state these conclusions in the form of the following lemmas, which
we shall subsequently prove.
Lemma1. If </>(f.)is piecewise smooth in the closed interval [a, b], then
b
,lf,,df,,= O(r
f </>(f.)c~s
a sm
1) ad-+ oo. (7)
Lemma2. If</>(?;,) is piecewise smooth in the closed interval [a, bJ,then at

a point x 0 where 4>is continuous,
a <X <b. (8)
Lemma3. At a point x 0 where <f>(x)has a jump discontinuity, (8) holds,

provided that we define (x 0 ) to be the average of the left- and right-hand
f,,)as f,,-+ x 0 :
limits of </>(
</>(x
0 ) = ![(x 0 - 0) + </>(x
0 + O)]. (9)
Here, by definition,
q,(x0 0) = lim q,(x0 &),
!0
where the downward-pointing arrow means that e decreases to zero; i.e., e
approaches zero through positive values.
According to Lemma I, si >(x)-+ 0.

2
Also S11)(x)-+ f(x), according to
Lemma 2, if we define
x-f,,
= f(f,,) 2 sin t(x - f,,)
</>(f.) (IO)
Furthermore, if there is a discontinuity of f(x) at x = x 0 , the sum of the sine

series at x 0 would be
f(xo) = H/(xo + 0) + f(xo - O)], (11)
in accordance with (9). This completes the process of summing the Fourier
sine series-except for proving the lemmas.
PROOF OF THE LEMMAS

To prove Lemma I, we break the integral into a finite number of similar
integrals over intervals in each of which the function cf>'(x)is continuous. Take
a typical interval (c, d); we integrate by parts, obtaining
f cf>(,)eu(de= [(U)-
c
d
1
- (iJr
</>(,)eu~]~ 1
f c/>'(,)ew;d,.
d
Since all the functions involved are bounded, Lemma I follows immediately. D
(We have used a complex exponential function in the integrand for con-
venience. Because the real and imaginary parts in an equation must be sepa-
rately equal to each other, both cosine and sine terms are included.)
To prove Lemma 2, let (c, d) denote an interval, containing x 0 , in whic11,
is smooth. We need only consider the integral (8) over this interval, since the
remaining contribution approaches zero as .A.-+oo, according to Lemma l.
Using the identity
</>(x)= (x 0 ) + [cf>(x)- </>(x
0 )],
and the definition

_ </>(x)- </>(x
0)
Fi(X,Xo ) - , (12)
X-Xo
we write the integral in question as
sin l(x - x0 ) (
I(l) = cf>(x
0) ( dx + F(x, x 0 ) sm l(x - x 0 ) dx. (13)
c X - Xo c
When A.-+ oo, the first term approaches mp(x0 ), since

oo sin v d
J
-oo
--
v
v= 1t. (14)
Formula (14) can be proved most easily by the method of contour integration
in the complex plane. Alternatively, it is not difficult to prove the convergence
of the integral, and the calculations in this section would assure us that the
value of the integral is 1t.
The second term in (13) can be shown to approach zero as A.-+oo with the
help of Lemma I. Given any preassigned positive number e, we can choose a
number c5(e)such that
Sec. 4.2] Fourier's Theorem 135
I
I[
o+6
F(x, x 0 ) sin A(x - x 0 ) dx < - ,
e
xo-6 2
since the integrand is bounded. In the intervals (c, x 0 - <>)and (x 0 + t>,d), the
contributions to the integral can be made as small as we wish by choosing A
sufficiently large, according to Lemma 1 : hence the desired result. D
To prove Lemma 3, we break the interval (a, b) into the intervals (a, x 0 )
and (x 0 , b), and repeat the arguments for Lemma 2. We are led to the inte-
grals
f00 sin v d
-- V=
f
O
sin v d
-- V,
O V -co V
whose values are n/2, according to (14). The rest of the derivation is easy
(Exercise 3). D
A FORMAL TRANSFORMATION
A formal method that is useful for quickly obtaining the right answer, but
which is not convenient for proving its correctness, proceeds as follows.
If we introduce into (8) the new variable u by the relation
A(~ - Xo) = U, (15)
we obtain an integral of the form
A(b-xo> ( u)
sin u
/(1i.)=
I A(a-xo)
<px 0 +-:; -du.
IL U
(16)
From a formal point of view it is clear that

.
hm J(A.)=
Joo<p(x) -sin u du = n<p(x). (17)
.i.-oo -oo 0 U 0
To justify the formal limiting process, however, we have to write
and use essentially the same arguments as above.
FOURIER SERIES IN THE FULL RANGE

Each term of the sine series is an odd function in x and hence the series
defines an odd functionf(x). That is, iff(x) is given by the series in (0, :n:),then
the series defines f(x) in the interval ( - :n:,0) by the identity
f(x) = - f(-x), -:n: <x<O. (18)
Beyond the interval (-:n:, :n:),the function defined by the Fourier series is
periodic:
f(x + 2:n:)= f(x). (19)
A periodic function satisfying (I 9) does not have to be odd. If we want to

represent a general function f(x) in an interval (-n, n), we can adopt the
"full range" Fourier series*
00
f(x),.., ,ta0 + L (ancos nx + bn sin nx); -1t < x < 1t. (20)
n=l
Formal use of Equation (1.38) yields (Exercise 4) the following expressions

for the coefficients :
1 "
an= -1t f_,,
f(x) cos nx dx, bn = -1
1t
f" f(x)
_,,
sin nx dx. (21)
We must now prove that if (21) holds, the series (20) converges to f(x) under
appropriate conditions.
The series (20) is also often used in the complex form
00
f(x) "' L cnein::c, (22)
-oo
where
(23)
Note that
ao
Co=z, (24)
Note also that if f(x) is real, then c_n = en, where the bar denotes complex
conjugate. For an application of complex Fourier series see Chapter 8 of II.
SUMMATION OF FOURIER SERIES
Consider the series in the form (22) and (23). Let
N
SN(x)= L Cnein::c. (25)
n=-N
Then, by using (23), we see that
(26)
where now
!<)_ ~ inC::c-(> _ sin (N + ,t)(x - e)
KN(X, 1:,- 1.., e - . (27)
-N sin !(X - e)
* The symbol - indicates that the series on the right of (20) "corresponds " to f. An equals
sign is not used (when one is being careful) because sometimes the series does not converge to the
function, at least not in the usual sense of"converge."
Sec. 4.3] On the Nature of Fourier Series
137
But (27) contains only the one significant term in (3), so that the lemmas
proved before lead immediately to the desired result. We thus .have the follow-
ing theorem.
Theorem. If f(x) is piecewise smooth in the interval ( - n, n), then its

Fourier series, as defined by (20) and (21), converges to
f(xo) = Hf(xo - 0) + f(xo + O)] (28)
at x = x 0 At the end points, the series converges to t[f( - 1t + 0) + f(n - 0)].
HALF-RANGE SERIES
The behavior of a Fourier sine series in the interval (0, n) can now be
understood. In view of the "oddness" relationship (18), the theorem above
shows that the series must converge to zero at the end points (0, n). (This is
not surprising, since each term of the series is zero at the end points.) Thus,
when a Fourier sine series in the interval (0, 1t) is regarded as a complete
Fourier series in the interval (-1t, n), it always represents an odd function
that vanishes at x = 0 or whose mean value is zero at x = 0.
EXERCISES
1. Verify (3).
2. (a)! Prove that the infinite integral of (14) is convergent.
(b) If you are familiar with contour integration, use this method to
demonstrate (14).
(c) Discuss the difficulties inherent in obtaining by numerical methods
a value for the integral in (14).
3. Complete the proof of Lemma 3.
4. Verify (21).
is. If a function of period 21t has a continuous third derivative in [ - n, n]
(including the end points), show that its Fourier coefficients are O(n- 3)
as n-+ co.
4.3 On the Natureof FourierSeries

The principal topics to be treated in this section are (i) integration and
differentiation of Fourier series; (ii) the Gibbs phenomenon, a concommitant
of the nonuniform convergence of the series; (iii) the relation between Fourier
series and least squares approximation; and (iv) Parseval's theorem on the sum
of the squares of the Fourier coefficients. But before we discuss the general
nature of Fourier series, let us examine a few simple examples. We consider,
in particular, the Fourier series of some simple powers and polynomials. In
each case we shall consider the three possible developments :
(i) The full range series in the complete interval (-n, n).
(ii) The sine series in the interval (0, n).
(iii) The cosine series in the interval (0, n).
The reader should verify all the necessary calculations that yield (I), (2), (4),
(5), (7), (8), and (9) below (Exercise 1).
FOURIER SERIES FOR THE CONST ANT FUNCTION
For the complete interval, the series for the constant function/(x) = l has
only one term, the constant term. All the higher terms vanish:
n ~I. (I)
The cosine series takes on the same form, since/(x) is even.

The sine series is much more interesting. It is
I = -4{.sm x + -I. sm 3x + -I. sm Sx + }, 0 <x < n. (2)

1t 3 5
Because the sines are odd, the above series represents - I for the interval
(-n, 0). We shall designate by S(x) the square wavefunction represented by the
series (2). Thus
-1 -1t < x < 0,
S(x)= ' (3)
{ +I, O<x<n;
and Sis periodic with period 2n. The function takes a step 2 upward at x = 0
and a step downward at x = n [Figure 4.6(a)].
S(x)
I
11 I -x
-2ir -lT 0 7T 2ir
(a)
C(x)
~
-2,r -ir 0
(b)
7T 2,r - x
FIG URE 4.6. The functions S(x) and C(x).

139
FOURIER SERIES FOR THE LINEAR FUNCTION

For the complete interval, the Fourier series for f(x) =xis
x = 2{sin x -t sin 2x + t sin 3x - }, -n < x < n. (4)
For the interval (0, n), the sine series is obviously still given by (4); but the
cosine series, which represents the function Ix I in [ - n, n], is
xi=~-~{~+
2 7t 12 cos323x+ } , -n::s;;x::s;;n. (5)
'
Let us designate the function represented by the series (5) by C(x). We note
that C(x) is continuous and has corners at x = 0, 2n, ... [Figure 4.6(b)]. We
also note that
C'(x) = S(x). (6)
This follows by comparing (2) with (5), or by a direct examination of the
functions C and S. Using (6), if we integrate the series in (2) from O to x, we
obtain
1t I I
cos x) + 32 (I - cos 3x) + 52 (I - cos 5x) + .
4 C(x) = (I - (7)
At x = n, this gives
n2 1 1
-=1+-+-+2 (8)
8 3 52
The reader should verify that one can obtain the same answer by formally
calculating the average of the square of the series (2) over the interval (O, n).
This is a special case of Parseval's theorem (to be considered below).
FOURIER SERIES FOR THE QUADRATIC FUNCTION
2
We can discussf(x) = x in the same spirit, but we know from our experi-
ence that its sine series is of primary interest. We proceed indirectly by formal
integration of (5), obtaining, not x 2 , but the polynomial
n (1tX
- - X
2) = SID
. x+ -I SID . 5X +
. 3X + -I SIO (9)
8 33 53
The sine series for x 2 can now be obtained by (4) and (9). By putting x = n/2
into this series, we can obtain a series for 1r3 [Exercise S(b)]. This series con-
verges faster than series (8) for 1t2 and can be used as a fairly practical way to
compute n.
INTEGRATION AND DIFFERENTIATION OF FOURIER SERIES
Let/(x) be a piecewise smooth function with the Fourier series
f(x)= ~o + J/ancos nx + bn
sin nx), (10)
convergent in the interval ( - TC,.re). Since the convergence is not uniform,

standard theorems give no assurance that the series can be integrated term
by term. Term-by-term integration is nevertheless permissible, as we shall
see. There is little reason to expect that (10) can be differentiated term by
term; indeed, series (2) and (4) are obvious cases in which the differentiated
series do not converge.*
Let us examine these questions a little more closely. The formal integration
of (I 0) from O to x leads to
( ) ao
F x - - x "' ~L.. - sm nx {a,.. + -bn [1 - cos nx l}, (11)
2 n=l n n
where
F(x) = r
0
f(t) dt. (12)
Formal differentiation of (10) yields

00
f'(x)"' 0+ I (-na. sin nx + nb. cos nx). (13)

n= 1
We shall now verify (11) and (13) directly by (i) calculating the Fourier series
of the relevant functions, and then (ii) by examining their convergence. For
(I I), the calculation is straightforward, and the reader can easily verify
(Exercise 3) that if we write
ao Ao co
2 x = 2 +.~1(A. cos nx + B. sin nx),

F(x)- (14)
then
-b
A=--"
n , nef;O. (15)
n
The convergence of the series (14) is assured, since Fis continuous (because
it is the integral of a piecewise smooth function). In particular, F(x) is con-
tinuous at x = 0, and F(O) = 0, so that
00
0=!A 0 + LA.
n= 1
(16)
Equation (11) then follows from (14) and (16).

Direct verification of (13) constitutes much more of a problem. There is no
assurance that the series converges at all, much less that it converges to
f'(x). To see what happens, let
a' oo
~ + L (a~cos nx + b~ sin nx) (17)
2 n= l
* The differentiated series can be "summed" in an appropriate manner, but this matter is
beyond the present discussion.
141
denote the Fourier series for f'(x). Upon partial integration of the defining
integrals of a~ and b~ and comparison with the corresponding integrals for an
and bn, we find (Exercise 3) that
na~ = nnbn + (-1?[/(n) - f(-n)], (18)
(19)
Reference to the formal series (13) shows that if
f(-n) =f(n), (20)

i.e., if the function is periodic and has no discontinuity anywhere, including
- n or + n, then formal differentiation is justified if the resultant series
converges. An example is afforded by the function C(x) with Fourier series
(5). Note that C(-n) = C( +n) = n. In this example, C'(x) can be shown to
have a convergent Fourier series (Exercise 4).
Considerations similar to the above may be made for sine expansions and
cosine expansions in the range (O, n). Referring to equations analogous to
(18) and (19), one can see that ifj'(x) has a sine expansion in (0, n), it can be
obtained by differentiating the cosine expansion of f(x). On the other hand,
if we start out with a sine expansion of f(x), formal differentiation is in
general not justified.
GIBBS PHENOMENON
Just before the turn of the century, the American physicist A. Michelson
constructed a machine by which the first 80 Fourier components of a graph-
ically given function could be determined. This analysis could be checked
by adding up the components obtained and verifying that the function so
synthesized was close to the original function. In most cases it was, but when
the square wave function S(x) was analyzed, at its point of discontinuity
"a peculiar protuberance appeared which was not present in the original
function. Michelson was puzzled and thought that perhaps a hidden mechan-
ical defect of the machine might cause the trouble. He wrote about his observa-
tion to Josiah Gibbs, the eminent mathematical physicist, asking for his
opinion. Gibbs investigated the phenomenon and explained it (in a letter to
Nature in 1899)."*
The unusual behavior of the partial sum near a discontinuity is called the
Gibbsphenomenon. Some such behavior should perhaps have been anticipated
from the outset, in view of the fact that the trigonometric functions are
infinitely differentiable, yet the Fourier series may converge to a discontinuous
function.
The quotation is from a book that is an excellent source of further reading on Fourier
series. The spirit is similar to that we have tried to convey, but much more material is presented.
The book is C. Lanczos, Discourse on Fourier Series (Edinburgh: Oliver & Boyd, 1966).
As a simple example of the Gibbs phenomenon, consider the square wave

function S(x) of (3) and the behavior of its Fourier series (2) near x = 0. The
series converges, partly as a result of the decrease in size of the successive
terms and partly as a result of the change of sign associated with the sine
function. However, for very small values of x, this oscillatory behavior is not
evident in the first k terms if (2k + l)x < n. All these terms are positive.
Indeed, we shall now show that as long as (2k + I)x is finite, there will
be a deviation from the final value. One would therefore expect the partial
sum of (2) to approach the limiting value only when (2k + I)x-+ oo.
Using formulas (2.26) and (2.27) for SN(x), particularized for the square
wave function S(x) of (3), we see that
SN(x) =_!_(-Jo+ J")sin(~+ t)(x - e) de. (21)

2n -n O sm t(x - e)
In the intervals ( - n, 0) and (O,n:),we introduce the transformations O= x - e
and O = e- x, respectively. Then (21) becomes
SN(x) =_!_
([
2n: x+n
+ r-x)sin sm(:4tO+ f)OdO.
-x
(22)
[The reader is asked in Exercise 6(a) to verify the above equation and also
equations (23), (25), (26), and (27) below.] Equation (22) may also be written
as
SN(x) =_!_
([
2n: -x
+
n+x
r-x)
sin(~ + t)O dO.
sm tO
(23)
One of the discontinuities of S(x) is at x = 0. Let us focus on the behavior

of the partial sum SN(x) in the neighborhood of this point.for positive values
of x. When x is small, the effect of the second term is small, since the de-
nominator in the integrand is near I. For the first term, we write
N+!=m, m0='1, (24)
with which
1
SN(x) ~ -n f
mx
O
.(
sin 'I
m sm 1'/12m
) d'I = IN(x). (25)
By writing
( )_ 2
1N x - ;
Jm.x
sin 'I ('1/2m) d
0
-,,- sin ('l/2m) 'I,
we see that for any fixed positive value of x, no matter how small, we have
lim IN(x) = -2 f co sin

- 'I
d'I = 1. (26)
N-+co 1t O 'I
On the other hand, for y = mx = (N + t)x fixed, we have
lim IN(x) = ~ Si[(N + t)x], (27a)

N-+co :n:
143
where the sine integral function Si is defined by*

1 sin r,
Si(y) =
f - r,
O
dYf. (27b)
The first maximum of Si(y) occurs at y = n [Exercise 5(a)]. Therefore

fix) and SN(x) reach approximately the peak value
2
Smax= -
J"'--sin YfdYf~ 1.179 (28)
1t O ,,
at x ~ n/(N + t), or Nx ~ n. This means that the maximum value of the

partial sum SN of the series (2) is reached if we include only the first set of
positive terms, in agreement with our general expectations.
R EM A R K . This comparison between the sum SN and the integral allows
us to deduce (27) by the following heuristic reasoning. If we denote 2x by h,
and introduce y = (N + t)x as before, we may write SN(x) as a sum approxi-
mating to the integral in (27). The reader is asked to carry out this reasoning
in Exercise 6(b).
Formula (27a) shows that, for N large, the values of x for which there is
deviation of I N(x) from its limiting value unity [for x > 0, (N + t)x - oo] is
restricted to x = O(N- 1). The maximum error remains finite at about 18 per
cent as N-... oo, but the range over which there is an appreciable error is squeezed
down to zero as N--. oo. This is typical of Gibbs phenomenon. It is a splendid
illustration of a series that converges but does not converge uniformly (Figure
4.7).
'--~~~~~...._x x x
0 0 0
(a) (b) (c)
FIGURE 4.7. Qualitative graphs of I,lx). The values of N are taken

successively larger in (a), (b) and (c), respectively. As N increases, the
peak continually narrows and passes to the left of any fixed x. Conse-
quently, IN(x) ~ l as N ~ co, for any fixed x such that O<x < l. But
the peak is always present; so for any N there are always small positive
values of x for which IIN(x)-11 is not small, and thus convergence is
nonuniform for O< x < 1.
For formulas, graphs, and tables of Si, see Abramovitz and Stegun (1964, pp. 231 ff.).
144
APPROXIMATION WITH LEAST SQUARED ERROR
While the study of the pointwise convergence of the Fourier series is
extremely interesting, one cannot help wishing for a simpler perspective of
the approximation procedure. A simple approach, which has much greater
generality, is the method of least squares, due originally to Gauss.
Consider a finite sum of trigonometric functions
N
SN(x) = L Yneinx, Y-n = Yn (29)
n=-N
If we want this sum to approximate a given function f(x), we would like to
make the difference
(30)
as small as possible in some sense. This may be accomplished by minimizing
the mean square error:
I "
M=-
21t
J -,r
sidx (31)
by a proper choice of the undetermined coefficients Yn.

The quantity M defined by (31) is a quadratic function of the coefficients
Yn, and hence a necessary condition for it to be a minimum is to require the
simultaneous linear equations
oM
-=0 (32)
oyn
to be satisfied. Equation (32) yields the condition
-I
27t
f" [f(x) - SN(x)]e'n"
-,r
- . dx = 0,
or the conditions
Yn= -
1 "
f
21t -,r
.
f(x)e-inx dx, n = 0, 1, 2, ... , N, (33)
as the reader can easily verify [Exercise 7(a)]. The quantities Yn are thus the
Fourier coefficientsdefined by (2.23) and denoted by en. In fact, a little calcula-
tion shows that
N N
M=(/
2
)- L lcn\2 + n=-N
n=-N
L \cn-Yn\ 2
, (34)
2
where (/ ) denotes the mean value of/2 (over the interval [-n, n]):
(/2)= ~ ("f2(x) dx.

2
Sec. 4.3] On the Nature of Fourier Series 145

If we note that only the last term is at our disposal when y,. is varied, it is
clear that we must necessarily have y,. = c..in order to minimize the error, and
that this condition is also sufficient.
Note that the coefficients Ynare finally decided by (33), in that they are not
to be changed if we choose to include more terms in the approximating sum
(29).
The method of least squares can be generalized in various ways. For
example, suppose that we have an infinite set of functions q,0 (x), ... , <f>
..(x), ... ,
which satisfy the following condition involving some given nonnegative
function w: J" (35)
_,.<l>m(x)<f,,.(x)w(x)
dx = ~-.
Before indicating the uses of such a set of functions, let us mention some
terminology that is very often employed. Condition (35) when m n is by
definition a property of functions that are orthogonal on the interval [ -11:, 11:]
with respect to the weighting function w. Condition (35) with m = n is said to
provide a normalization of the set of functions, for it specifies an otherwise
arbitrary constant multiple of the </>,..
Mutually orthogonal and normalized
functions are said to be orthonormal. (Subsequent discussion will justify the
use of the term " orthogonal.")
Given a set of functions that satisfy the orthonormality condition (35),
we may attempt to approximate an arbitrary function f(x) in the interval
(-11:, 1e)by a finite sum
(36)
by minimizing the weighted mean square error
Mw = J:.. ei(x)w(x) dx. (37)
Here sN(x)is defined by (30). Note that errors at values of x where wis relatively
large are weighted more heavily in the computation of Mw; hence the term
"weighted" error.
Let ( >w denote a weighted mean value so that
(J2)w = I: f 2(x)w(x) dx.

We now have the following generalization of (34) [Exercise 7(a)]:
N N
M.,..= (f 2 )w- L lcnl2 + L lcn- 'Ynl2 , (38)
n=O n=O
where
Cn = J:..f(x)<J>,,(x)w(x) dx. (39)

The error is minimized when the y/s are chosen to be the same as the en's
defined by (39).
The method for constructing the set of functions <J,n(x) for a given w(x) is
essentially that of the Gram-Schmidtprocess, to be discussed in Section 5.2.
We shall not go into this process here. Suffice it to say that it is possible to
proceed in a step-by-step fashion, even if we accept the restriction that the
<J>nCx)
are polynomials of degree n. As an illustration, we shall merely mention
the Legendre polynomials
-l<x<l. (40)
These can be regarded as polynomials defined on the interval (-1, 1), which
are mutually orthogonal in an unweighted sense (w = 1). The conventional
normalization requirement imposed on Legendre polynomials is not
f 1
-1
P:dx = I
but rather is Pn(l) = 1.
BESSEL'S INEQUALITY AND PARSEVAL'S THEOREM

If we put Yn= cn in (34), we obtain the Bessel inequality
(41a)
since M ~ 0. In terms of the real coefficients [compare (2.24)] this reads

N
fa~ + t L (a; + b;) ~ (j2). (41b)
n=l
If the Fourier series converges in a pointwise manner, with the exception

of a.finite number of points of discontinuity, then
lim M=O,
and hence
co
n=
I lc;I = <! 2 >
-<X>
(42a)
or
ao
{a~ + i L (a; + b;) = (/ 2 ).
11=1
(42b)
Either of these identities is known as Parseval's theorem. They can be formally

obtained from the Fourier series (2.20) or (2.22).
Even when the mean square error M does not approach zero, the inequalities
(41) guarantee the convergence of the infinite series in (42).
Sec. 4.3] On the Nature of Fourier Series 147

RIESZ-FISCHER THEOREM
Suppose that we now pose the following problem, which can be regarded
as a converse to Parseval's theorem. Given a set of real numbers a0 , a'", b'",
m = 1, 2, ... , oo such that the series
00
!a5 + L (a;, + b;,)

m=l
is convergent, is there a functionf(x) such that the series
00
!ao + L (ancos nx + bnsin nx)

n=l
(43)
is its Fourier series?

An affirmative answer to this question depends on the introduction of the
concepts of Lebesgue measure and Lebesgue integration. With these notions
introduced, we have the Riesz-Fischer theorem,* which states that (i) the
series (43) is indeed the Fourier series of a function!, which is square integrable,
and that (ii) the partial sums of the series converge in the mean to f.
The problem we posed is a very natural one from a mathematical point of
view. It appears that it might have a simple solution, but it is here that new
mathematical concepts and theories emerge. On the other hand, for physical
applications, such a mathematical question does not arise naturally.
APPLICATIONS OF PARSEVAL'S THEOREM
Let us now consider, in contrast to the above theoretical discussions, some
elementary applications of Parseval's theorem, which are useful for practical
purposes. In particular, we shall show how one can evaluate the integral that
occurs when one wants to calculate Stefan's radiation constant from Planck's
law of radiation. As indicated in Exercise 9, this integral is
I= foox3 dx.
0 <r- 1 (44)
First, we observe that from (2), by an application of the theorem of
Parseval, we obtain another derivation of (8). Second, if we apply Parseval's
theorem to (5), we obtain [Exercise 7(b)]
4
1 1 1t
1 + 34 + 54 + ... = 96. (45)
To evaluate (44), as requested in Exercise 9, one expands the denominator

of the integrand as a power series in e-x and integrates term by term. One
thereby obtains a constant multiple of
1 1 1
84= 14 + 24 + 34 +.. (46)
See E. C. Titchmarsh, Theory of Functions (Oxford: Oxford University Press, 2nd ed.,
1949), p. 423.
which differs from (45) in that all the integers, instead of just the odd ones,
are present. But S 4 can be related to the series (45) by noting that the even
terms themselves make u:o r 4 S 4 , so that [Exercise 7(b)]
(47)
If presented with the problem of summing S4 , few would have the ability
to obtain the answer (47) by making a seemingly roundabout detour through
Fourier series. Consequently, our calculations can be regarded as an illus-
tration of possible unexpected applicability of mathematical analysis.
EXERCISES
1. Verify the following Fourier series calculations: (1), (2), (4), (5), (7),
(8) [two ways-see the remark under (8)], (9).
2. Carry out the calculations outlined under (9).
3. Verify (15), (18), and (19).
4. Show that C'(x) has a convergent Fourier series.
5. (a) Verify that the first maximum of the sine integral function occurs
at y = rr.
(b) (Project) Without consulting the literature, derive as many prop-
erties as you can of the sine integral function.
6. (a) Verify (22), (23), (25), (26), and (27).
(b) Carry out the program of heuristic reasoning that is outlined
under (28).
7. (a) Verify (33), (34), and (38).
(b) Verify (45) and (47).
(c) Why would mean cubed error not be a useful concept?
8. With the aid of the results proved in this section show that
1t l 1 l
(a) 4 = l - 3+ 5 - 7 + ...
3
1t 1 1 1
(b) 32 = 1- 33 + 53 - 73 +" ..
4
1t 1 1 1
(c) 96 = 1 + 34 + 54 + 74 + ....
(d) Use series (c) to calculate an approximation to 1t.

9. (a) Using hints given in the text, show that
ae x3 dx = 1t4
J e" -
O J 15
(b) The integral of (a) occurs in the theory of thermal radiation. As
discussed in physics texts, according to Planck's law the density
149
ofradiation u(v) [per unit volume, per unit range of frequency v] at

temperature T is given by
8nhv3 /c 3
u(v)=-----
exp (hv/kT) - 1
Here h is Planck's constant and k is Boltzmann's constant.
Show that the total radiation density in a cavity at temperature
Tis
8n 5 k 4
15h3 c3 T4,
giving the famous proportionality to the fourth power of the
temperature of Stefan's law.
10. Calculate numerically the sum of a number of terms of series (2) to
observe the Gibbs phenomenon in practice. Graph your results. (Project.)
11. Calculate numerically the sum of a number of terms of the solution of
the heat equation when the initial temperature of the rod is uniform at
the beginning and the ends are suddenly brought to zero temperature.
Notice that there is, in general, no Gibbs phenomenon. Can you imagine
a situation in which the Gibbs phenomenon would be present? (Project.)
12. This exercise concerns the inhomogeneous boundary value problem
d2y
dx 2 +ky = J, 0 < x < n; y(O) = y(n) = 0. (48)
Herek is regarded as a given fixed constant, and/ is a given function of

x.
(a) Using integration by parts, show that if (48) is to have a solution,
then it is necessary that/be orthogonal [on the interval (0, n) with
weighting function unity] to all solutions cpof
d2cp
dx 2 + k<p= 0, 0 < x < n; (0) = </J(n)= 0.
(This necessary condition will be used in Exercise 7.2. I 1, which is

concerned with the application of perturbation theory to determine
eigenvalues.)
(b) Assume that
00 co
f = L Cn sin nx, y = L kn sin nx.
n=l n= 1
Since/is known, the enwill be known coefficients of a Fourier sine

series. The problem is to determine the kn. Do this by substituting
into (48). Show that there are zero, one, or an infinite number of
solutions, depending on the nature off (This is a special case of
the "Fredholm alternative.") Are the results compatible with
part (a)?
CHAPTER 5
Further Developments in
Fourier Analysis
IN CHAPTER 4we examined simple forms of the problem of heat conduction

and we presented some of the mathematical theory of Fourier series that was
inspired by attempts to solve this problem. Our treatment followed historical
developments closely and serves as a case study to illustrate the role of ap-
plied mathematics.
In this chapter we shall again make a study in this spirit but at a somewhat
deeper level. Once more starting with specific problems, we shall outline
theory and methods of wide-ranging applicability and implications.
We begin by considering some specific problems of heat conduction. A
study of the problem of heat conduction in inhomogeneous media will lead
to the discussion of eigenfunctions more general than the harmonic functions.
The case of an infinite homogeneous medium will introduce the topic of
Fourier integral. It will then be pointed out that the analysis of a time se-
quence, such as a weather record, may often be conveniently made in terms of
Fourier series or integral but that this is not sufficient. There exist time se-
quences that cannot be so represented. The theory of generalized harmonic
analysis is required, and this will be briefly described.
5.1 Other Aspects of Heat Conduction

This section begins by reducing the study of annual temperature variation
in the earth to a relatively simple heat conduction problem for a homogeneous
medium. A brief calculation brings out the main features of the solution.
The discussion then turns to numerical analysis of the heat equation. Finally,
advantages of a mixed numerical and analytical approach are illustrated
with reference to heat conduction in a nonuniform medium.
VARIATION OF TEMPERATURE UNDERGROUND
On the surface of the earth, the daily average temperature at time t, f(t),
may be regarded as varying with a period of one year. We may represent this
variation by
where Cn= C_n, T= 1 year. (I)
n=-oo
We want to know the consequent variation of temperature underground (say,

in a cellar). We shall neglect the curvature of the surface of the earth, since we
150
Sec. 5.JJ Other Aspects of Heat Conduction
151
shall be concerned with depths of the order of only a few meters. These two
approximations-the neglect of diurnal (daily) variation and the neglect of
curvature-are introduced to formulate a simplified model. Such approxi-
mations could in principle be justified by analyzing a model that contains
curvature and diurnal variation effects. Yet this may be quite complicated,
and is indeed hardly worthwhile, at least at this stage. But plausible argu-
ments to simplify the problem can be given in terms of the smallness of cer-
tain parameters. Adopting a procedure that we recommend for general use,
we shall sketch these arguments at once, before we attempt to carry out
the detailed calculations.
In the simplified model the equation governing the distribution of tem-
perature O(x, t) at a depth x and at time t is the usual equation of heat
conduction,
ae a2 e
-=,c-.-. (2)
ot ox2
Here " is the thermometric diffusivity of soil whose value is approximately
"= 2 x 10- 3
cm2 /sec. (3)
The relevant time scale T is a year :
T = 3.15 x 107 sec. (4)
It is a fundamental fact about diffusion (see the end of Section 4.1) that in
time T a substance diffusing with diffusivity " will spread a distance of order
of magnitude fe. In the present case this length scale is
J,cT~ 250 cm~ Sft. (5)
For diurnal variation the time scale is T/365 so that the length scale would be
smaller than (5) by a factor of J365 ~ 19. Thus neglect of diurnal variation
would be expected to produce errors of the order of 5 or 10 per cent, while the
curvature of the earth would produce entirely negligible effects on phenomena
limited to a few meters depth.
To find the temperature underground according to our simplified model,
we first take note of the boundary condition 0(0, t) = f(t), where f(t) is given
by (I). We try solutions of the form
(6)
n=-oo
where Cn= C_n so that (J is real. (The bar denotes a complex conjugate.) We
make the following stipulations:
(i) Each of the terms will satisfy (2).
(ii) wn(O)= l so that (1) will be satisfied.
(iii) wn(x) remains bounded and presumably approaches zero as x--+ oo
(for n #=0), since the temperature at great depths is not expected to be
sensitive to the variations of surface temperature.
152 Further Developments in Fourier Analysis [Ch. 5
Can all these stipulations be satisfied? We must carry out the calculations
to find out.
The first stipulation yields the following ordinary differential equation for
w,.(x):
2 2nin
p,. = KT. (7)
The possible values of Pn are

1 2
Pn = (1 i)qn, where qn = ( lnl1r)
KT ' > 0 (8)
and the sign is to be taken accordingly as n ~ 0 (Exercise I). The general

solution of (7) is then
(9)
Because of condition (iii), we must have An= 0. To satisfy condition (ii), we
must have Bn = 1. Thus the final form of solution (6) becomes
8(x, t) = L
00
Cne-(li)qnJCe2inrfT. (10)
""' -oo
To emphasize that 8 is real, we write the complex coefficients in polar form,
putting
c,.= Ic,,1ei'" (rt)
Equation (10) then becomes
8(x, t) =Co+ 2 ~ IC,.le-Qn"

cos (21rn~ + -y,. - q,,x). (12)
n= I
To interpret the solution, we first note that the cosine factor represents a
wave of frequency 21tn/T and wave number q,.. (If these matters are not
familiar, see Section 12.2.) Thus the nth "partial wave" propagates with
a speed
In I)1
(41tK 1 2
2nn (13)
q,,T= -T- .
In addition, there is exponential damping in the direction of propagation.

From (8) the damping coefficient q,. increases like In j 112
The constants IC,.I and 'l'ngive the amplitudes and phases of the various
damped waves. These can be determined from initial conditions, but this
would add little to the general understanding we seek.
Since damping increases with n, the most important contribution to the
solution comes from the term with n = 1. With the numerical values cited
above, for which fe ~ 250 cm, we find from (8) that
(14)
Sec. 5.1] Other Aspects of Heat Conduction 153
Thus, at the point x 1 = 4.4 m, where q1 x 1 = 11:,the temperature is opposite in

phase to the surface condition, but the amplitude is down by a factor of e - " =
0.0432. It is winter at a depth of 4.4 meters when it is summer on the surface.
The temperature variation, however, is only about 4 per cent of the surface
variation. This shows the usefulness of a deep cellar for wine and vegetable
storage.
As noted above, the diurnal variation has a much thinner penetration; all
the variations take place in a thin surface layer. This phenomenon is well
known in many processes, e.g., the skin effect in electromagnetism. Exercise
3.2.1 of II deals with the entirely analogous effect produced in a viscous fluid
by an oscillating plane boundary.
In all these cases the basic elements are the periodic variations in time of a
quantity whose distribution is governed by a diffusive process, with basic
terms exhibited by (2).
NUMERICAL INTEGRATION OF THE HEAT EQUATION

With the advent of large computers, there is a new balance between the
advantages of analytical methods versus numerical methods. So far, we have
concentrated on the use of analytical methods. We have nevertheless inadver-
tently discussed the basis for the numerical integration of the heat equation.
because the natural equation to be used for finding an approximate solution to
the heat equation by the method of finite differences is the difference equation
(3.3.1), which was the basis for our discussion of random walk.
Indeed, consider the difference equation
O(x, t + k) - O(x, t) = t[O(x - h, t) + O(x + h, t)J - O(x, t), (15)
[which is equivalent to (3.3.1)]. Approximating for small hand k, we find that
kO,(x, t) + O(k 2 ) = 1 [
O(x, t) - hOix, t) + h2 O,.;c(x,t) + O(h 3 )
2 2
+ O(x, t) + hO,.(x, t) + h2 O,.,.(x,t) + O(h 3 ) ] - O(x, t)
2
or
+ O(k) + o(~).
8,=:8,.,.
In the formal limit as h-+ 0, k-+ 0, with h2 /k fixed, we obtain the diffusion
equation
(16a, b)
Now if we start with the differential equation (16) and consider (15) as an
.. approximation" to it, we must insist that h = Ax and k = At are related by
At= (Ax)2
2D.
Thus the increment in t is smaller by an order of magnitude when compared

with the subdivisions in x. Indeed, books on numerical analysis prove an
extremely important related result, namely, that if
(Ax)2
At< 2D ' (17)
then the differenceequation method of analysis is stable. This means that if the
initial values are changed by a small amount, the solution is changed by an
amount of the same order. On the other hand, if the interval At is so large that
(17) is reversed, then the method is unstable.In that case any small error in
the initial condition (or inevitable small roundoff errors) would eventually
cause large errors in the subsequently calculated values.
If we refer to Figure 5.1, we see that (15) allows us to calculate the value of
u(x, t) at an interior point P' from the values at P and its neighboring points
P 1 and P 2 The same condition obtains for Q'. For boundary points R' and
S', the situation is different, but if u is given along the boundaries AA 1 and
BB 1 , the process of numerical integration can be carried on. [If the boundary
condition is that of insulation, then we take u(R') = u( Q') to conform to
ou/ox = O.] Thus the solution in the whole domain A 1ABB 1 can be found.*
A1"" - B1
Q' P'
R' s _l_
R s k
Q P1 T
M P\P 2
(x, t)
A B
FIGURE 5.1. Grid points involved in a numerical analysis of the heat
equation. In particular, to obtain the temperature 8 at P'(x, t + k), we
need to know 8 at P(x, t), P1(x - h, t), and Pl(x + h, t).
* Confidence that our problem is well posed is generated by verification that a straightforward
numerical scheme, in principle at least, can provide a solution to the partial differential equation
plus associated boundary and initial conditions. Another verification of such a fact, with a
numerical scheme for a more complicated problem, will be found in Appendix 3.1 of II.
Sec. 5.1] Other Aspects of Heat Conduction
155
In the problem of heat conduction along an infinite rod ( - oo < x < + oo),
we would expect the temperature distribution to approach constant values
with vanishing derivatives at infinity. If we take a sufficiently large interval
-A < x < A and impose a suitable condition at the end points, the method of
finite differences can be carried through.
A comparison of the analytical versus numerical methods shows clearly
that the former can give better insight into the nature of the problem. For
example, it is no doubt possible to discover by numerical methods the fact
that the deviation of the distribution of temperature from the steady state will
eventually become sinusoidal, but the analytical solution (Equation 4.1.41)
shows this immediately and clearly. Also, the source solution considered in
Chapter 3 is obviously very difficult to describe numerically.
The chief advantage of the numerical solution is its greater generality, when
the differential equation does not have a simple form. The analytical method
works beautifully, largely because the differential equation is linear and the
coefficients are constants. Indeed, the geometry has to be simple, too. Suppose
that we are dealing with a problem involving inhomogeneous material. We
shall see, in the next section, that analytical methods begin to run into com-
plications. Moreover, for nonlinear problems, analytical methods are often
very clumsy, if not impossible. But even for such problems, there are certain
situations where analytical methods have advantages over a purely numerical
approach, chiefly when singularities and other rapid variations are involved.
In general, a mixed approach, using both kinds of methods, is the most
sensible.
A mixed approach might begin with the analysis of a crude model to find
the most important effects and to determine interesting parameter ranges.
Then numerical analysis of a more complete model might provide needed
detail. In a problem area where understanding is minimal, it might be wise to
start with a few numerical experiments to gain a general understanding of the
phenomenon. Armed with such an understanding, one can often isolate the
central aspects of the phenomenon for analytical study. A return to the com-
puter may then be useful to obtain a detailed picture.
A three-dimensional problem containing a number of parameters will
probably continue to defy complete numerical analysis for a long time. And
even if such an analysis were available, the resulting surfeit of data would be
of little use to a person unless he had some analytically generated understand-
ing of the problem.
HEAT CONDUCTION IN A NONUNIFORM MEDIUM
We shall now study an example that illustrates some advantages of a mixed

approach and introduces an important subject called Sturm-Liouville theory.
Suppose that we are dealing with heat conduction in an alloy whose
composition is not homogeneous. The density p, specific heat c, and heat
conductivity k are then functions of position. As stated in Section 4.1, the

equation for heat conduction then becomes
(18)
or
pc
ao
ot= V (kVO).
This equation still allows simple solutions with the variable t separated
out:
lJ(x, y, z, t) = U(x, y, z)e-l'. (19)
In the one-dimensional case, U satisfies an ordinary differential equation
d ( dU)
dx k dx +ApcU=O, (20)
where k, p, and c are functions of the variable x.

In general, (20) cannot be solved explicitly in simple analytical form, but
we can still gain insight into the nature of its solutions by analogy with the
case of a uniform medium. For definiteness let us consider the specific prob-
lem where the end points of a rod lying along (0, L) are kept at temperature
zero, so that
U(O)= U(L) = O. (21)
Let us recapitulate our work on the homogeneous case. When k, p, and c
are constants, (20) has nontrivial solutions satisfying (21) only if A= n,
where
n= 1, 2, .... (22)
The corresponding solutions are constant multiples of
U.() . mr:x
nX =Sin, (23)
A solution corresponding to particular initial values is obtained by super-

position.
In the present case, by analogy with the development just presented and
with (6), we would expect a solution of the form
QC)
u = L Un{x)e-l,.r. (24)
n=l
Sec. 5.1] Other Aspects of Heat Conduction
157
Here each Un(x) is defined by the differential equation (20) (with 1 = 1,.)
d ( dU,.)+ ,l,.pcU.. =0.

dx k dx (25a)
The boundary conditions are

U,.(O)=0, U,.(L)=O. (25b)
The problem afforded by (25) has the trivial solution U = 0. Nontrivial solu-
tion functions U,.(x) generally exist only for certain constants 1,., n = 1, 2, ....
The U,. are termed eigenfunctions belonging to the eigenvalues 1,.. We note
that if U,.(x) is a solution, so is C,.UnCx),for any constant C,.. Ifwe now impose
the initial condition u(x, O)=f(x) on the series solution (24), we see that we
must have
00
f(x) = I C,.U,.(x) (26)
n=l
for some choice of the constants C,..

The solution of the heat conduction problem is now seen to depend on the
possibility of expressing (in an appropriate sense) an arbitrary function as an
infinite series of the form (26). The required body of theorems and tech-
niques is known as the Sturm-Liouville theory.* We shall not pursue it in
detail here, but in the next section we shall develop a few facts that can be
obtained from some relatively simple arguments. We shall show how these can
be useful for the explicit solution of our problem.
At this point, let us briefly compare the present approach and a purely
numerical approach to the solution of (18). If the domain in question is very
complicated, one might well decide to use a numerical approach right from
the beginning. However, even here, it might be simpler to use superpositions
of solutions of the form (I 9). Let us examine this in the one-dimensional case;
the remarks also hold in higher dimensions.
In the case of the one-dimensional medium, we know that for the homo-
geneous case,
k
n=l,2, ... ; K::-. (27)
pc
Since the corresponding nth spatial eigenfunction </,,.is multiplied by the
factor exp ( -,l,.t), only the terms for low values of n are important in the
/L is
2 2
series solution (24), except for a very short initial period wherein 1<.trc
not yet of magnitude unity. It is reasonable to expect that this qualitative
feature will not be changed by the introduction of inhomogeneity, with the
possible exception of cases where the inhomogeneity is "very large." Thus
the practical solution of most inhomogeneous problems should reduce to
See Courant and Hilbert {1953, Vol. 1, Chap. 5).

Further Developments in Fourier Analysis [Ch. 5
finding a few of the lower eigenfunctions and eigenvalues. These can be

determined by numerical integration of some ordinary differential equations,
in contrast to the partial differential equation (18). Thus we continue to expect
that the qualitative behavior of the solution is one in which various "modes"
are damped out at quite different rates. In the highly inhomogeneous case,
however, the mode shapes might be far from sinusoidal.
There is a further advantage to an analytic approach. The eigenfunctions
and eigenvalues thus found can be shown to be associated with other physical
problems. Consider an attempt to solve the equation
a2w a (k-aw)+- a (k-aw)+- a (k-aw)

pc-=- (28)
at2 ax ax oy oy oz oz
(which occurs for vibrations of a continuous medium), by a simple normal
mode solution of the form
w = W(x, y, z)ei"' 1 (29)
If w 2 = A, then the equation satisfied by W(x, y, z) is the same as that for

U(x, y, z). If the boundary conditions are also identical, we would be led to
the same eigenvalue problem with the same set of eigenfunctions. Two seem-
ingly unrelated problems, one involving diffusion and the other vibration,
are thus brought together by mathematical considerations.
EXE RCI SES

1. Verify (7) and (8).
2. For what physical reason do we expect all the eigenvalues of (25) to be
positive?
3. (Project.) Perform a numerical analysis of the problem offinding annual
temperature variation in the earth. Assume that the surface temperature
has only two values, a " summer" value for half the year and a
"winter" value for the other half. Compare your results with the
appropriate special case of (12).
4. Using Bessel functions, determine the eigenfunctions of (25) when k =
ax", pc= fix'; oc,P,m, and n are constants. Discuss.
S. Formulate the annual temperature variation problem for a spherical
earth. If u(r, t) is the temperature as a function of the radius r and the
time t, introduce a new dependent variable ru(r, t). Without actually
solving any equations, discuss the correction due to sphericity.
t6. The solution (12) shows that the annual temperature oscillates with an
amplitude that depends on depth. At one observation station in the
Soviet Union, the amplitude was measured at 11.5, 6.8, 4.2, and 2.6C
at 1, 2, 3, and 4 m. Use these data to determine a value for the thermo-
metric diffusivity " of the earth.
Sec. 5.2] Sturm-Liouville Systems
159
t7. Measurements 2 or 3 km below the surface of the earth have revealed

an increase in temperature with depth of about 3C every 100 m. As we
have seen, annual variations are insensible at these depths. A reasonable
explanation for the observations, however, is that the temperature grad-
ient is a result of the earth's cooling. Show that this is an unsatisfactory
hypothesis on the basis of the following data. Rocks melt at about 1200C,
the average thermometric diffusivity of granite and basalt is 6 x 10- 3
cm 2 /sec, and the age of the earth is about l 0 9 years.*
5.2 Sturm-Liouville Systems

We now consider a slight generalization of a mathematical problem that
arose in the previous section, the problem of determining certain solutions of
the secondorderSturm-Liouvilledifferentialequation
d [ p(x) dU]
ldx dx + [Jp(x)-q(x)]U=O. (I)
Here J is a parameter; p, p, and q are real-valued functions of x; and the func-

tions g and p are positive. (The function q was identically zero in the pre-
ceding example. For ordinary Fourier series,p = p = 1, q = 0.) We shall again
take the boundary conditions to be
U(a) =0, U(b) =0. (2)
It will become clear as we progress, however, that the gist of the discussion
will not be greatly changed when other homogeneoust boundary conditions
are introduced in place of (1). (See Exercise 7.)
We shall now show that
(i) Nontrivial solutions U; can exist only for certain real discrete values
A; of J.
(ii) These solutions are orthogonal to each other with weighting function
p(x) in the sense that
f p(x)U;(x)Uix)
b
a
dx = 0, (3)
This problem and the preceding are adapted from an excellent discussion by Tycbonov and
Samarski (1964, pp. 215-55) of geophysical applications of beat diffusion theory.
t The word "homogeneous" is overworked, but its various usages have become traditional.
Here are four of them: (a) A homogeneous condition on a function is one that is certainly satisfied
by the zero function. Thus (1) is a homogeneous equation but U'(O)- 5 = 0 is not a homogeneous
boundary condition. (b) A homogeneous material is one whose composition is uniform. (c) The
special differential equation dy/dx=f(y/x) is termed "homogeneous." (d) A function f of n
variables is said to be homogeneous of degree.m if
f(txi, tX2, .. . , tx.) = t"'/(x 1 , Xz, ... , x.).
16o Further Developments in Fourier Analysis [Ch. 5
By analogy with our experience with Fourier series, we would also expect (iii)
to have an infinite number of eigenfunctions and (iv) to be able to express a
broad class off unctions, in a suitable sense, as infinite series of eigenfunctions.
As mentioned before, these last two propositions are not easy to prove, and
we shall not attempt proofs here. However, we shall show (v) how the
higher eigenfunctions and eigenvalues can be approximately calculated, even
though we cannot give a rigorous proof of their existence. We shall also show
(vi) how the lower eigenfunctions can be calculated by mixed numerical and
analytical methods, or directly by numerical methods. [The higher (lower)
eigenvalues are the larger (smaller) ones. The corresponding eigenfunctions
are also called "higher" and "lower."]
PROPERTIES OF EIGENVALUES AND EIGENFUNCTIONS
We wish to show that the boundary conditions (2) can be satisfied only for
a discrete set of values of ,l. By this we mean that each eigenvalue ,l can be
made the center of a small circle which encloses no other eigenvalues.
Let Urz. and U6 form a fundamental set of solutions of (1). By definition this
means that any solution of (1) can be expressed in the form
,l) + c6 Up(x; ,l)
U(x, ,l) = crz.Urz.Cx; (4)
for certain constants c,. and cp. (Note that the explicit dependence of U,. and
Up on ,l has been exhibited.) The boundary conditions (2) require that
crz.U,la;,l) + c1p 6(a; ,l) = 0,
,l) + CpUp(b;,l) = 0.
Crz.Urz.Cb;
The solution to this pair of homogeneous equations for Crz.
and cp is nontrivial
only if
I
Urz.(a;..l)
Ur,.(b;,l)
Ug(a; ,l)
Up(b;,l)
I=0. (5)
We assert that this secular or characteristic equation (5) will yield a de-
numerable set of eigenvalues. (This gives the discrete spectrum, in contrast to
the continuous spectrum fase where there is a continuum of eigenvalues.)
A proof can be constructed by showing that U,.(x, ,l) and Up(x, ).) can be
chosen to be analytic functions of the parameter ..:l,regarded as a complex
variable. Those familiar with complex variable theory will recognize that the
conclusion follows from the fact that the zeros of an analytic function are
isolated.
It is not hard to show (once the idea of the proof is born-see Exercise I)
that the eigenvalues must be real numbers. [Also, the eigenfunctions can al-
ways be taken to be real. See Exercise I (b). ] By analogy with the case of the
Fourier functions, furthermore, we would anticipate that the sequence of
Sec. 5.2J Sturm-liouvi//e Systems
161
eigenvalues has no upper bound. We also expect the higher eigenfunctions (those
corresponding to larger values of l) to oscillate rapidly. These expectations
tum out to be correct. They form the basis of the asymptotic evaluation of
the eigenvalues and of their corresponding eigenfunctions.
ORTHOGONALITY AND NORMALIZATION
We shall now show that a pair of eigenfunctions corresponding to different

eigenvalues satisfy the orthogonality condition (3). Consider the two equations
-d ( p-'dU.) + (l-p -q)U = 0, (6a)

dx dx ' '
d ( p dU.)
dx ~ +().ip-q)Ui=O. (6b)
If we multiply (6a) by Ui and (6b) by Ui, and subtract, we obtain
U-- d ( p-'dU) - U-- d ( p-dU-)

1
+(k-J.)pU-U-=0
1 1
dx dx ' dx dx ' i
If we integrate this equation between the limits (a, b), and impose the bound-
ary conditions (2), we find using parts integration [Exercise 2(a)] that
so
r[
o
U.-
1
d ( p-'dU ) -Ui- d ( p-dU 1)] dx=O,
d:x dx dx dx
(7)
(li-l) I"pUiUjdx=O.
0
(8)
Thus, unless l; = ).i, the orthogonality condition (3) must be.satisfied.

If there are a finite number s of linearly independent eigenfunctions be-
longing to the same eigenvalue, we can construct (by the Gram-Sdunidt
process) a set of p mutually orthogonal eigenfunctions belonging to this
eigenvalue. To accomplish this, let us assume that Uni, Un2 , , Uns are
linearly independent eigenfunctions all belonging to the eigenvalue An.We
shall now specify, one by one, mutually orthogonal eigenfunctions Vi, ... , V3
For V1 , we shall take Uni For V1:,we shall take
I <k :::;;;s. (9)
We assume that all the previous V's have been chosen, and impose the re-
quirement that Vi shall be orthogonal to mutually orthogonal Vi, ... , Vk-i
We then find {Exercise 2(b)] that
(10)
where*
(U, V) = f p(x)U(x)V(x)
b
a
dx. (11)
Not all the ck, m's are zero, unless Unk is already orthogonal to all the Vm's.
We shall standardize our notation by adopting a set of eigenfunctions
{<f>nCx)} = I.
that are orthogonal to each other and normalized so that (<f>n,<f>n)
Thus
(12)
where
t'jmn= 0, m =I=n,
We should remember, however, that some of the eigenvalues may be the
same. If there are two linearly independent eigenfunctions with the same
eigenvalue, we speak of this eigenvalue as having a multiplicity of two, or
being a doublet, etc. An example of multiple eigenvalues is furnished by the
harmonic functions in the interval (-n, n):
0 (X) = 1/

1(x) = J; sinx, ef>z(x)=
1
J;. cos x, (13)
'P2n- 1 (x) =}. sin nx, 2 nCx)=

1
J-;rcos nx.
The functions {<f>

2 n_ 1 (x), q,2 nCx)}form a doublet with common eigenvalue
n 2 .t
EXPANSION IN TERMS OF EIGENFUNCTIONS

If we try to express an arbitrary function fin a series using the set of func-
tions {<f>n},
as was done with Fourier series, we may evaluate the coefficients in
a similar manner. Indeed, if
f(x) "' Cn<f>nCx),

=L (14)
n=O
The notation ( , ) emphasizes the connection with the scalar product of vectors. Indeed,
the Gram-Schmidt process is often first met in the purely algebraic task of converting a linearly
independent set of vectors to a mutually orthogonal set. As some readers may already know, the
connection hinted at here is one of basically identical formal structure. This is exploited in
Appendix 12.1 of II, where the ideas of Sturm-Liouville theory are generalized to self-adjoint
operators.
t The ,f,'s can be regarded as eigenfunctions of the problem
t/>"+>.2,f,=0, ef,(-1r)=,f,(1r), t/,'(-1r)=t/,'(1r)
On comparing with the Sturm-Liouville equation (1), we confirm that the weighting function p(x)
is identically equal to unity. Orthonormality has already been asserted in (4.1.38).
Sec. 5.2] Sturm-Liouville Systems 163
we have (formally)
00 ""
(</>;,/) = L cnC</>1,
= Len Jin= C;,
<f>n)
n=O n=O
i.e.,
,!)
en= (<f>n = f <J>ix)p(x)f(x) dx.
a
b
(15)
It is an exceptional function that requires only a finite number of terms in

(14). Thus we must have an infinite number of eigenfunctions before we can
hope that (14) will be valid for a general class of functions. Certainly, this
requirement alone is not sufficient. If by some accident we did not include
0 (x) in the list, for example, we could not expect to be successful. We thus
<J,
have the need to prove the completeness of the eigenfunctions with respect to
some class of functions-in the sense that a suitable series of eigenfunctions
will converge to any function of the class. This is by no means easy. We shall
therefore restrict ourselves to the statement that the set of eigenfunctions of
a Sturm-Liouville problem (with weak conditions on the coefficient functions)
is indeed complete with respect to a broad class of functions.
To enumerate the eigenfunctions, we may depend on the number of zeros
that a function has in the interval in question. For example, for the harmonic
functions (13), <l>m(x) has m zeros in the open interval (-n:, n:). Such behavior
of the eigenfunctions is guaranteed by certain oscillation theorems. The
reader is referred to books on differential equations for proofs of com-
pleteness theorems and oscillation theorems.
ASYMPTOTIC APPROXIMATIONS TO EIGENFUNCTIONS AND

EIGENVALUES
One can gain an insight into the nature of the eigenfunctions 'Pnand their
associated eigenvalues Anby an analysis of their behavior when n and Anare
large. For this purpose, it is convenient first to transform the original differ-
ential equation (I) into the normal form of Liouville by' introducing a new
independent variable t = t(x), and a new dependent variable w, where
U=y(x)w. (16)
We have two functions at our disposal, and we may aim at imposing two
desirable conditions on the equations. If we choose these conditions as (i) the
coefficient of dw/dt should be zero, and (ii) the coefficient of J..should be unity,
we must (Exercise 3) take the functions w and t so that
w
U(x) = [p(x)p(x)]lf4' t(x) = f p(x)]112
[ p(x) dx. (17a, b)
We then obtain Liouville'snormalform,

d2w
dt 2 + [)..- 4(t)]w = 0, (18)
where
4 = ~ + (pp)-1/4
2
d ((pp)l/4]. (19)
p dt 2
For an equation in the form (I) it is difficult to get any idea of how the
solution behaves when ;.. is large. For (18), however, this is an easy matter.*
Assuming that;.. is large compared to 4, we at once guess that (18) has the
approximate solutions exp (iA 112 t). Thus it is .. natural" to try to solve (18),
when ;..is large, by the series
w(t) = exp (iA 112
t)[w0
{t) + ;..- 112
w1(t) + ]. (20)
The reader should calculate a few terms of this series by direct substitution
(Exercise 4). He will see, for example, that w0 must be a constant.
Let us now impose the boundary conditions w = 0 at the end points,
(t,,, tb). We must make these points correspond to the original end points
x = a and x = b. This can be accomplished by writing {17b) in the form
t-
-I"" rem]1,2
LP@ di;,
and taking
(21)
We find (Exercise 5) that for large n the eigenvalues and eigenfunctions satisfy
2
A.!'
= [;~:~r'
nn{( dxr+ o(~),
2 1
(22)
dxI + 0 (1)]
dx}[
112
-1/4. { J:CP(x)/p(x)]
Un(x)= [p(x )p (x )] J!
sm mr (p(x)/p(x)] 112 ~ (23)
These results are very useful for practical purposes. Surprisingly perhaps, but
typically, they are very accurate even for fairly small values of n.
As an example, consider the Bessel equation
d (x dU)
dx dx + (k x - x
m 2
2
)
U = 0. (24)
We thus learn from Liouville that ifan equation contains a large parameter, we will probably
accomplish something useful if we introduce new variables so that certain terms are definitely
small relative to this parameter. These terms can be omitted to obtain a first approximation.
Sec. 5.2] Sturm-Liouville Systems 165
This is a special case of (I) with p = p = x, q = m2 /x. The Liouville trans-

formation (17) is U = wx- 112 and x = t, and the resultant equation (18) is
2
d w (
dx2 + k2 -
m
2
-i)w = 0.
--:xr- (25)
When k is large (and xis not near zero), the first approximation to the general
solution is
w(x) = c1 cos kx + c2 sin kx.
If the boundary conditions w(a) = w(b) = 0 are imposed, we are led to the
approximate eigenfunctions
wn(x) = sin kn(x - a), (26)
with
mr
kn= b--; n = I, 2, 3, ....
-a
If a = I and b = 2, the lowest approximate eigenvalue is k 1 = n. When
m = 0 (a favorable case to be sure) the actual lowest eigenvalue is between
3.1 and 3.2 (Exercise 6). Of course, when m = -!, our "approximation" gives
the exact solution.
Once the form of an approximate solution is known, it can often be
obtained by a more direct method than that originally used. In the present
case the calculations can be made after a simpler transformation U = vp- 112
112
followed by the formal substitution v = ei>- 4>{v 0 () + x- ' v 1 () + }.
12
With this, will automatically tum out to be the function t(x) defined by
( l 7b) (Exercise I 0). The series does not converge, but it is asymptotic-and
very useful. An asymptotic series of such a form is used to bridge classical
mechanics with quantum mechanics in the Schrodinger formulation. This
same form had been used long before in the theory of water waves. The
procedure is also found to be useful even when p(x) changes sign. In that
case the "JWKB theory" has to be invoked to resolve the difficulties occurring
near a zero of p(x). It is, however, beyond our scope to go further into these
matters. The reader is referred to such references as Cole ( 1968, Section 3.7).
OTHER METHODS OF CALCULATING EIGENFUNCTIONS AND
EIGENVALUES
Numerical techniques can often be used advantageously to calculate eigen-
values and eigenfunctions. One way to do this is to solve initial value problems
for the two linearly independent functions U1 (x, ).) and Ui(x, J), where, in
addition to the differential equation (I}, these functions respectively satisfy
the initial conditions
Ui(a, J) = 0, U1(b, I)= 1,
Ui(a, I) = I, Ui(b, I) = 0.
If the general solution U(x, l) = C1 U1 (x, l) + C2 U2 (x, l) is to satisfy the

boundary conditions (2), then C1 and C2 must satisfy the homogeneous
equations
C1 U1(a, l) + C2 Uia, l) = 0,
C1 U1 (b, ).) + C2 U2 (b, l) = 0,
which have a nontrivial solution if and only if
Roots of this transcendental equation can be determined numerically. This

is not an easy task, however, particularly in more complicated (nonself-
adjoint) problems with complex eigenvalues. Another difficulty arises when,
say, U1 increases much more rapidly with x than U2 Unless precautions are
taken, roundoff introduces a slight error, so instead of U2 one is calculating
eU1 + U2 Although e is very small, U1 becomes so large that eU1 + U2
becomes indistinguishable from U1 and therefore does not supply, as re-
quired, a solution that is independent of U1 In spite of these and other dif-
ficulties, however, numerical calculation of eigenvalues has been pursued
with great success in many investigations.
Another useful approximate method makes use of the fact that eigenvalues
can often be characterized as minima of certain expressions. This Rayleigh-
Ritz method in the calculus of variations is briefly illustrated in Section 12.4
and is treated more fully in Chapter 12 of II.
EXERCISES
1. (a) Prove in the following manner that the eigenvalues of(l) are real,
(i) Find an equation for U, the complex conjugate of U, by taking
the complex conjugate of (l). Remember that p, p, and q are
real-valued functions of x, although A may not be.
(ii) If l #=l, deduce a contradiction from the orthogonality of
eigenfunctions corresponding to distinct eigenvalues.
(b) Prove that if U is an eigenfunction of (I), then either Re U or Im U
is a real eigenfunction.
2. (a) Verify (7) and (8), formally.
(b) Fill in the details omitted in the text's discussion of the Gram-
Schmidt process.
3. Show that to obtain Liouville's normal form ( 18)the change of variables
should indeed be as given in (17).
4. Calculate at least two terms of the series (20) by formal substitution.
5. Verify (22) and (23).
6. Consider (25) with m = 0, subject to the boundary conditions w(I) =
w{2)= 0. If you are familiar with Bessel functions, find a transcendental
Sec. 5.3] Brief Introduction to Fourier Transform
equation for the eigenvalues k". By use of a table of Bessel functions,

show that k 1 lies between 3.1 and 3.2.
7. (a) Verify (3) when the boundary conditions (2) are replaced by
U'(a) = U'(b) = 0.
(b) Are there more general boundary conditions that will still permit
the derivation of (3)?
8. Verify (3) for the fourth order Sturm-Liouville problem
:: 2 [!(x) ~~ + ! [p(x) ~ + [2p(x) - q(x)]U =0;
U(a) = U'(a) = U(b) = U'(b) = 0.

9. (a) Show that the general second order equation
d2 U dU
a(x) dx 2 + b(x) dx + [c(x) + 2 d(x)]U = O
can be put into Sturm-Liouville form by an appropriate change of

variables.
(b) Can the general fourth order eigenvalue problem be put into the
form of Exercise 8?
10. Carry out the alternative approach to (l 7b) that is outlined below
(26).
11. Generalize the results of Exercise 4.3.12 so that they apply to the second
order Sturm-Liouville problem (I) and (2).
5.3 Brief Introductionto FourierTransform

This brief treatment of a large topic is motivated by a desire to find the
limit as L~ ex:, of the solution to a heat conduction problem in a rod of
length L. We are led from Fourier series to Fourier transform formulas. As a
check, we recover an earlier result for heat flow in the infinite rod. Further
application of the Fourier transform can be found in the treatment of dis-
persive waves in Chapter 8 of II.
FOURIER TRANSFORM FORMULAS AND THE FOURIER

IDENTITY
In Section 4.1 we considered the solution of the problem of heat conduction
along a rod of length L. In the limit when L~ oo, we should be able to re-
trieve the solution of the infinite interval heat conduction problem that we
obtained in Chapter 3 with the help of the source solution. Crucial here is
generalizing the representation of an arbitrary function
00
mnx
f(x) = L
m=l
Bm sin-
L
in the limit L-+ oo. (Compare Equation 4.1.36.) More generally, we shall
discuss the limit as L-+ oo of the complete Fourier series in complex form
00
f(x) = L eneinu/L, (I)
-co
where
en= 2~(L
f({)e-in~IL d{. (2)
There seems no obvious way to simplify (I) and (2) when Lis large. It is
true that for fixed L the integrand in en becomes highly oscillatory when
n-+ oo. Consequently, the variation in f during a single oscillation of
exp( - inn{/L) is hardly noticeable, so the positive contributions will almost
cancel the negative. On the other hand, this whole idea is invalid for those
values of n which are of the same magnitude as L.
From the study of the Liouville transformation we learned to attempt to
clarify the role of a large parameter by changing from the given variables in
the problem, in this case n and L. The previous paragraph suggests the im-
portance of n/L; therefore, we shall introduce a multiple of this quantity as
one variable:
(3a)
(The added factor ,r just makes the resulting formulas a little simpler.) It
turns out that the second variable (which we call /ik) should be taken as the
difference between adjacent values of k (as n increases by unity):
7t
lik=- (3b)
L
With these, (2) becomes
/1k
en= C(k) = -2
7t
f L
-L
f({)e-~
.
ae. (4)
We now introduce (4) into (1) and obtain
~ e io 21t
f(x) = t=.l.:.co Lf
/1k -LJ(,}e -at
a,,k = .. , -2/ik, -lllc,O,lllc, ... (5)
The behavior of the integral in (5) as L-+ oo depends crucially on the

behavior of f(f.) for IeI large. We have enough on our hands at the moment;
so let us bypass this matter by considering a function/(x) that vanishes iden-
tically outside an interval ( -A, A). If we take L > A, the integral in (5) may
be taken between the limits ( - oo, oo) and represents a definite function
g(k) =J 00
-oo
/(f.)e-iJ:~ a,. (6)
Sec. 5.3] Brief Introduction to Fourier Transform
With this (5) becomes

co 6k
f(x) = L
eikxg(k) - . (7)
k= -co 21t
The limit of the sum as L -+ oo and Ilk -+ 0 is
I co
f(x) = - J g(k)e;"" dk
21t -co
(8)
if the integral converges.*

In (6), the function g is known as the Fourier transform of the function f
Note that the inversion formula (8) from g back to /has substantially the same
form as the Fourier transform formula (6), the main differences being the
factor 21t and the change of sign of i. The numerical factor is sometimes redis-
tributed to give more symmetry to the formulas. It is convenient to combine
(6) and (8) into the Fourier identity
= ..!..
1<x> Jco dk fco1ce>ik(x-()
de. (9)
21t -co -co
So far we have proved (9) only for the case wherein f(x) = 0 when IxI > A.
We shall now verify (9) for a far more general class of functions by a limiting
process.
Consider the function f(x, K) defined by the integral
f(x, K) = -
1
27t
f K
-K
dk f-a,
a,
J(e)ik<:it-(>
de. (10)
If J~coIJ<e>I
deexists, then the order of integration may be interchanged, and
we obtain
f(x, K) = ..!_
2,r
f00
-a,
J(e) 2i si_nK(x - e) de.
i(x - e)
(11)
We now break up the interval of integration in (I I) into the subintervals

(-oo, a), (a, b), (b, oo); a< x < b. It can be shown that if/is of bounded
variation, then the contributions from the first and last of these subintervals
approach zero as K-+ oo.t If f is piecewise smooth, the treatment of the
Nontrivial limits involve such expressions as the quotient of two small terms, say, or (as
here) a sum of very many very small terms. We have succeeded in evaluating the present limit,
because by a change of variable we have cast our problem into a sum of many small terms whose
limit is readily recognized as a definite integral.
t See Jeffreys and Jeffreys (1962, p. 4S4). Recall that /(x) is of bounded variation if l:f.o
lf(x1+1)-/(x1)I < M for some M and for every possible subdivision - oo < x 0 < x1 <
<xN<XN+1<+00.
middle subinterval is precisely covered by the lemmas proved in Section 4.2.

In particular, according to (4.2.8),
limK_,00 f(x, K) = f(x).
Thus the identity (9) is established, subject to the same interpretations as before
at a discontinuity.
SOLUTION OF THE HEAT EQUATION BY FOURIER
TRANSFORM
Having established the Fourier transform relation, we can readily solve the
problem of heat conduction over an infinite rod. Again we look for solu-
tions to 8, = 1<:8""
that have the form X(x)T(t). We require X(x) to remain
finite as !xi- oo.The most general solutions of this form are multiples of
exp (ikx -k 2Kt), where k is any real number. From these simple solutions
we construct by superposition the general solution
0
O(x, t) = { A(k) exp (ikx - k 2Kt) dk. (12)

-<X)
Boundary conditions at x = 0 and L previously required us to select only a

denumerable family of wave numbers k (discrete spectrum), but now k can
be any real number (continuous spectrum). Previously, we took a weighted
sum of permissible eigenfunctions; now, with a continuous spectrum, this
sum becomes an integral weighted by the function A.
We can satisfy an arbitrary initial condition
O(x, 0) = f(x), (13)
by (6) and (8), if A(k) = g(k)f2n. If we introduce the form (6) into (12), the
solution is directly expressed in terms of the initial distributionf(x) by
1 J<X> f<X>
O(x,t) = 21t -<X>
dk _,,/<~)
exp [ik(x -
2
~) - k Kt] de. (14)
If we invert the order of integration and then carry out the integration with
respect to k, we find [Exercise l(a)] the solution in the form
O(x, t) = _1_ J(e)e-<x-1;)2/4KIde.

J<X> (15)
~-<X>
This was obtained before by the method of source solution in Section 3.4. We
note that (15) is a valid solution of the problem even when the infinite integral
of lf(x) I is not convergent, as the solution (15) can be directly verified.*
Here is an example that makes a strong case for heuristic reasoning and formal
Rigorous verification that a formal solution actually satisfies all the conditions of a problem
of course provides a (constructive) existence theorem. For such an approach to the heat equation
see Lecture 8 ofS. L. Sobolev, Partial Differential Equations of Mathematical Physics (New York:
Pergamon, 1964).
Sec. 5.4] Generalized Harmonic Analysis
171
manipulations, provided that we remember to check the validity of the final

answers thus obtained. In many practical situations, however, even this check
is too difficult.We must resort to other ways of assuring the correctness of the
final answer, by comparison with obse~vations, for example.
EXERCISES
1. (a) Verify (15).
(b) Consider (15) in the special case when/vanishes for negative x and
has the constant value 80 for positive x. Derive
x
11= 2(Kt)1f2
(c) Express the answer in the form of a series that is expected to be

valid when 17is large.
2. (a) Show that the Fourier identity can be written in the form
1
f(x) = - f dk f
7C O
Cl()
-oo
Cl()
/(f.) cos k(f. - x) de,.
(b) Find a form of the Fourier identity involving only cosine (sine)
functions when/ is even (odd).
(c) Use (b) to show that
1
D(x) = 2n- ( k- 1 sink cos kx dk,
where
lxl < 1,
D(x) = (~ for lxl
( lxl >= 1,
t 1.
3. If you are familiar with Bessel functions, see how far you can get in
generalizing the results of this section to a situation involving cylindrical
symmetry. Start with the problem of heat conduction in a cylinder of
radius R, and consider the limit R -too. Try to derive
5.4 Generalized Harmonic Analysis

Although the Fourier series arose out of eigenfunctions for the solution of
problems of heat conduction (and we have seen in Section 5.1 that these
eigenfunctions need not be harmonic functions), it has perhaps found its
greatest application for the analysis of time-dependent signals and periodic

structures. To glimpse the range of these applications, just consider a few
instances of periodic phenomena. The different colors perceived by man are
due to electromagnetic waves of different frequencies. The ear can detect
different notes when stimulated by mechanical waves of different frequencies.
Periodic structures occur in inanimate crystal lattices as well as in animate
millipedes. Periodic motions of planets, double stars, and variable stars are
well known. The diurnal and annual variation of weather is a consequence
of the periodic motion of the earth.
In spite of the prevalence of periodicity, however, once we decide to analyze
time-dependent signals (such as temperature records) into harmonic compon-
ents, we are introducing something artificial, since such sequences need not
really be periodic. To be sure, we can try to extract the periodic components
such as the diurnal and annual variations in weather, and the 11 year cycle in
sunspot activity. Yet there may still be a great deal of the signal left over. One
must therefore be prepared to deal with those nonperiodic time variations
that continue indefinitely with unabated (though bounded) amplitude. Such
problems fall outside the method of Fourier analysis discussed thus far. They
require what is known as generalized harmonic analysis.
REMARKS ON FUNCTIONS THAT CANNOT BE ANALYZED BY

STANDARD FOURIER METHODS
Although the general theory is too complicated to be presented here, one
can get some feeling for the nature of the problem by considering a function
that is the sum of a few periodic functions whose periods are not commen-
surable. An example is
(I)
where for all integers m and n
i, j = 1, 2, 3; i j.
-=I=
One may add to f(t) a function of the type
co0 > 0. (2)
But this function decays sufficiently rapidly as It I --+ oo so that it can be

resolved into harmonic components by the Fourier transform method, even
though it shows no resemblance to harmonic functions. On the other hand,
the functionf(t) of (I) does not show a singly periodic behavior and cannot
be represented by a Fourier series. Neither does it die off at infinity. It can
therefore not be represented by a proper Fourier integral. Of course, it is
already resolved into periodic components-but how shall we analyze more
general functions that do not have such simplicity?
Sec. 5.4] Generalized Harmonic Analysis 173
The first step is to take a sufficiently long record, hopefully lasting over
many of the periods of any significant periodic component that may be pres-
ent. For example, in the case of a meteorological record we need at least IO
years' observations. To see how each truly periodic component would show up
in a Fourier analysis based on such a record, let us examine the Fourier anal-
ysis of a single periodic component lasting over a finite period of time:
ltl 5. T, (3)
ltl > T.
We shall refer to f T, as a truncated sinusoidal function.
We shall now analyze the truncated sinusoidal function defined by (3),
using both Fourier series and Fourier integral. When we consider the limit as
T-+ oo,using the first method, we get discrete harmonic components. We may
even reproduce the function exactly if the interval 2T happens to be an exact
integral multiple of the period 21t/ro0 In the second method, we get a con-
tinuous spectrum, but with a peak presumably at ro = ro0 . Let us examine the
details.
FOURIER SERIES ANALYSIS OF A TRUNCATED SINUSOIDAL

FUNCTION
Let n = 21t/2T= n/T be the fundamental frequency. We wish to represent

(3) in the form
A 0 eiwot _ ,
-1-,
"" Cn eino.r , for -T<t<T. (4)
-00
Then [Exercise I (a)]

sin(ro0 - nO)T Wo
n:;!:-. (5)
Cn = Ao (ro0 - nO)T ' n
The important components are associated with small values of (ro0 - nO)T.
If ro0 happens to be an integral multiple of n, say ro0 = kn, then we have
only one nonvanishing coefficient C1cin the Fourier series. If k < ro0 /0 < k + 1,
then both Ck and Ck+1 would be important. In the worst situation, with
ro0/0 = k + 1, i.e., when the interval 2Thappens to be a half-integral multiple
of the basic period 21t/ro0 , we have [Exercise l(b)]
2A 0
Ck+1 = Ck=- (6)
7t
These are the dominant coefficients,each having an amplitude of about 0.64A0 .

Their frequencies do not differ greatly from ro0 , the fractional error being
only fk, which is just 5 per cent if the record covers a time interval of about
10 periods.
It might be puzzling that we have two fairly large components, the sum of
whose amplitudes exceeds A 0 Actually, we should really consider the measure
of "energy" E, where
CX)
E= L 1Cnl
n=-co
2

Then Parseval's theorem (4.3.42) guarantees that
E= IAl2 (7)
Further, Bessel's inequality (4.3.41) gives assurance that a finite number of

components contains only a fraction of the energy. In the case just considered,
the two dominant components contain about 80 per cent of the energy.
FOURIER INTEGRAL ANALYSIS OF A TRUNCATEJ)

SINUSOIDAL FUNCTION
The Fourier series approach just discussed has the somewhat disconcerting
feature of splitting one sinusoidal component into two. We shall show that
the method of Fourier integral yields only a "broadening of the spectral
line" -to borrow an expression from the experimental physicist. The general
Fourier transform relation
a(ro)
I
=-
21t
f CX)
-ex,
f(t)e-,o,,

dt
and its inverse give for the truncated sinusoid (3)

A Tsin (a>0 - a>)T
QT ( a>) =--7t0 (a>0 - w)T
(8)
and, noting the different placement of the factor 2,r compared to (3.6) and
(3.8),
f(t) = fa) eio,r a-,((1))d(I) (9)
-ex,
as the reader is asked to verify in Exercise 2. The square of the function

aT((/)) of (8) is plotted in Figure 5.2. The ratio of the central maximum to the
first subsidiary maximum is about (3n/2)2 :::::::22.2. Thus the energy is practically
all contained in the shaded central band, which has a spread in frequency of
/1(1)= 2n/T = 20. The fraction of energy contained is given by
!F _ J':,.,,-2 sin2 ,r d,r (10)

- J~a) ,,-2 sin 2 11en,'
which is about 90 per cent of the total energy (Exercise 3)':-
The advantage of the Fourier integral approach is the reproduction of the
original signal as a single component, with a broadening of its line spectrum
[aT(w)J 2
[AoT/,rJ2
FIGURE 5.2. Broadening of a spectral line of frequency w 0 , in the

Fourier integral representation of the truncated sinusoid (3).
but without a distortion of its frequency. This approach is also convenient for
considering the case T ~ oo. Ifwe consider the limiting value of (8), we obtain
a 00 (w) =r-oc
Jim ar(w) = { ~' .
mdetermmate,
W
W
= Wo;
"F Wo.
(11)
To obtain definite limiting values, we introduce the definition
Ar(w)= "' ar(w) dm= -A J" -sin 'I dtJ,

f-oc 0
(12)
1C -co ,,
where 'I = (w - w0 )T. When T ~ oo, Ar(w) ~ A 00 (w), where

Ao for w > W 0 ,
((I) ) = {0
AOC c (13)
1or w < w0
Following standard notation, we shall write (13) as
A 00 (w) = A 0 H(w - w0 ), (14)
where H(x) is the Heavisidefunctiondefined by

I for x > 0,
H(x) = { O (15)
for x < 0.
We can interpret our results adequately at this point. From (9) we see that
ar(w) is essentially the amplitude (per unit frequency) of the harmonic of
frequency ro. Definition (12) shows thatAr(w) gives the cumulative amplitude.
In the limit, Ar(m) suffers a jump as shown in (13) or (14). The entire contri-
bution to A 00 (m) comes at the single frequency w 0 Thus the Fourier integral
approach comes arbitrarily close to an exact reproduction of a signal of
period w 0 if the length of time over which this signal is observed increases
without bound.
More precise and usable results can be obtained upon introduction of the
Stieltjes integral.* With this, (9) can be written
fT(t) = (:~,cei"'tdAT(w) (16)
and the formal limit T -- oo has meaning, since the Stieltjes integral can be
defined for functions AT with discontinuities. Indeed, in the limit form with
A,c(w) given by (14), Equation (16) yields the original function A 0 ei"'0 ', for
- oo < t < co. To obtain this result from (9), we should write
(17)
in the limit T-+ oo. Here <5(x)is the Dirac delta function that satisfies
f <C
-ao
f(x)fJ(x - a)dx = f(a). (18)
Since the relation A7(w) = aT(w) is implied by the definition of AT(w) in

(12), (14) and (17) imply that
fJ(x) = H'(x). (19)
This is a valid and useful formal result.
The reader will find it worthwhile to justify all the formalism by referring
to the lemmas in Section 4.2.
GENERALIZATION TO STATIONARY TIME SEQUENCES
The discussion above can be formalized in the general case of a function
f(t) called a stationary time sequence.t To this end, consider a truncated func-
tion fT(t) where
fT(t) = f(t), II I $ T;
=0, itl > T.
Let F be continuous. The Riemann integral S:F(w) dw is defined by a familiar limit of
:E1F(wT)(w 1+ 1 -
w 1), w 1 $ wT $ w 1+i If A is a monotonically increasing function, the Stieltjes
integral IA= J!F{w) dA. is defined by a similar limit of :E1F(wT)[A.(w 1+ 1)- A.(w1)]. If A is a step
function [which we express in terms of the Heaviside function of (15)] then the integral reduces
to a sum:
ao ao
A(w) =n=l:E c.H(w- w,,) => IA= L
=l
c.F(w,,).
(We assume that :E:".1 c. is a convergent series of nonnegative constants.) If A is continuously

differentiable, we obtain an ordinary Riemann integral:
IA= J!F{w)A'(w) dw.
For proofs and further information, see texts on analysis, e.g., Chapter S ofD. Widder's Adrianced
Calculus (Englewood Cliffs, NJ.: Prentice-Hall, 1947).
t We shall not attempt a precise definition here. The nature of a stationary time sequence is
described at the beginning of the section, with the weather record as a typical example. Also see
the remark below (24).
If we define in analogy to (8)
(20)
then the cumulative contribution of ar(<.0)in the interval (co, m + .6.ro)is

o,+Ao, 1 Jco e-io,t _ e-i(o,+Ao,)t
.6.Ar(co)=
J
,.,,
ar(w) dw = -
2 7t
fr(t)
-co .
zt
dt, {21)
where we have interchanged the order of integration. [Convergence difficulties

prevent introduction of an integral from - oo to w, as in the corresponding
formula (12).] The appearance of the factor tin the denominator then facili-
tates the convergence of the integral in (21) as T-+ oo. If the limiting integral
does exist, one may then expect to be able to introduce a function A(<.0)such
thatf(t) may be expressed as the Stieltjes' integral
J(t) = (:~coeia>tdA(<.0). (22)
Theorem 1. Equation (22) holds if A(ro) is defined by
A(w) = -
1 fcof(t) 1 - e-ia>t
. dt. (23)
27t -co -It
Proof. We divide the infinite line - oo < a, < oo into intervals of length
.6.ro(with zero at one point of subdivision), and consider the sum
N
SN(t)= L eia>n' 1[A(<.0n+ .6.<.0)
-A(wn)l,
n=-N
where N(.6.(J))= n, a fixed number. It is expected that in the limit .6.w-+0,

n-+ oo, SN(t) will approachf(t ), as indicated by (22). In the above expression,
m~ may be taken to be any value in the interval (wn, Wn+ .6.w);it is convenient
+ .6,(J)/2.
to take it to be <.On
A little manipulation shows that
SN= 2..fco/(t')
2n -co
2 sin [(~w/2)t']
t n=-N
f i"'n'(t-1') dt'.
Using (4.2.27), we see that the sum within the above integral is
i(Aa>/Z)(r-n sin [(N + !)(t - t')(.6.c.t>)]

e .
sin [!(t - t')(.6.w)]
Thus, for fixed n, the limit of SN as N-+ oo is
S(O) = ! Jcof(t') sinQ(t ~ t') dt'.

7t -co t-t
The limiting process n- oo is already familiar to us (Lemma 2, Section 4.2).0
For a time sequence that combines the oscillatory behavior of ( 1) with an

additional contribution exhibiting the rapid decay of (2) with increasing It I,
the function A(w) would have three discontinuities, at w = w1, w2, w3 , re-
spectively, with a smooth variation in between, to take care of (2).
AUTOCORRELATION FUNCTION AND THE POWER SPECTRUM
It is clear that A(w) could be very complicated if there were an infinite
number of discrete components of the type (1) such that the infinite series
composed of the squares of the amplitude still converges. Moreover, a func-
tion representing noise or turbulence is generally not representable by a com-
bination of such an infinite sum together with a function that has a Fourier
transform. This is best seen by the following considerations.
We introduce the autocorrelation function R('r), which is employed for
analyzing noisy signals (using a bar for complex conjugate):
1 f.to+T _ _
R(r) = lim - f(t)f(t + r) dt = <J(t)f(t + r)),. (24)
T-+oo 2T to-T
(By definition, for a stationary time sequence, the autocorrelation function is

independent of the reference time t 0 .) In the case of the function given in (I),
the autocorrelation function is (Exercise 5)
(25)
which shows anharmonic periodic behavior. However, for noisy or turbulent

signals,
R(i-)-+O as r-+ oo,
since the random nature of the signal annihilates the autocorrelation for
large distances.
The theory of generalized harmonic analysis has been rigorously developed
by Norbert Wiener and other mathematicians. We cannot go into further
details here except to mention one principal result. The theory shows that
there exists a power spectrum defined by
1
-F(w)=
. !aAl2
hm --. (26)
2 6.w-+0 aw
Furthermore, it can be shown that this power spectrum stands in Fourier
cosine transform relation to R(i-):
R(i-) = J 0
00
F(w) cos wi- dw, (27a)
2
F(w) = -J
,r
00
O
R(i-) cos wi- di-. (27b)
Sec. 5.4] Generalized Harmonic Analysis
179
For = 0, (27a) and (24) combined yield the relation
<If(t) l2 )t = {' F(ro) dw.

0
(28)
Equation (28) is the equivalent of Parseval's theorem, which justifies the term
"power spectrum" [since F(ro) dro provides the contribution from the fre-
quency interval (ro, ro + dro) to the "energy" given by the mean of/ 2 ].
VERIFICATION OF THE COSINE TRANSFORM RELATION
BETWEEN THE POWER SPECTRUM AND THE
AUTOCORRELATION
The derivation of (27) is quite difficult. However, in a statistical sense, where
the correlation function R(.) is defined by the ensemble average (an average
over many repetitions of a phenomenon)
R('r) = <f(t)f(t + -r)>, (29)
then the result is relatively easy to obtain. We take/(t) to be real.
We start with (21) and consider its limiting form (as T- oo):
AA(ro) = -
1
f(t)
J" e-iwt - e-i(w+4w)t
. dt. (30)
2:n: -oo it
We form I AA( ro) 12 by multiplying (30) with its complex conjugate, using t' as
the variable of integration in the expression for AA. The double integral in
(t, t') may then be transformed into one using the variables (t, .), where
= t' - t. (31)
We then obtain
{l
2 IAA(ro)12 =
(2:n:) f 00
-oo
e;.,, d. f 00
-oo
dt - e
i(4o,)(t+s)}{l
(
tt+.
)- e
-i(4W)t}
f(t)f(t + .).
Let us take the statistical average. Then

{l _
(2:n:)2(IAA(ro)12 ) = f
co
-co
eiw, d. R(.) f-oo
oo
dt
ei(4w)(i+>}{t
(
tt+-r
)
_ e-i(4w)r}
and hence
lim
,1w-+o
<IAA(ro)12>=
Aro
Jco
-co
e1..,, d-rR(-r). ~ J"11
(2:n:) -oo
-t 1
x
2
dx,
where the variable x = (Aro)t is a new integration variable. It can be shown

(Exercise 7) that
11- 12 dx = 2 J"1-
J
oo ei" COS X
----,- 2- dx = 2:n:. (32)
-oo x -co x2
180 Further Developments in FourierAnalysis [Ch. 5
Hence
1
- F(ro) = -
2 21t
1
f (X)
-(X)

R('r:)e'ondt.
Now R(t) must be an even function of -r, since

R(t) = <f(t)f(t + t)) = <f(t - t)f(t)) = R( --r).
Thus the cosine transform relationship is established.
APPLICATION
Experimental verification of the Fourier transform relation (27) for turbulent
signals was first accomplished by G. I. Taylor. (See Figure 5.3.)
x
3 o~
i
0
x
2
I
x
ox
l
x
c
x 0
(a)
xo
~
4
~
x
2 xo
x ~ x X ,1
x (
0 100 200 300 400Hz

Frequency
(b)
FIGURE 5.3. One-dimensional spectra: o, direct measurement; x,

transformed double correlation. [Reproduced with permission from
R. W. Stewart and A. A. Townsend. Phil. Trans. Roy. Soc. (London),
A234, 359 (1951).]
In modem radio astronomical observatories, the spectral analysis of the

incoming radiation is done by the use of (27). The correlation function R( -r)of
the incoming signal is first computed, and the spectral function F(w) is then
obtained as its Fourier cosine transform. All this is done with automated
equipment. Knowledge of F gives clues to the nature of the signaling source.
EXERCISES
1. (a) Verify (5).
(b) Verify (6).
2. Verify (8).
3. Show that the fraction ff defined in (10) is about 190
4. (Project.) Justify various formal results of this section, as suggested under
(19).
5. Show that the autocorrelation function for the function given in (1) is
supplied in (25).
6. Fill in the details of the proof of Theorem 1.
7. If you are familiar with contour integration methods, use them to estab-
lish (32).
CHAPTER 6
Simplification,DimensionalAnalysis,
and Scaling
A NUMBER of procedures that are of value to the applied mathematician

will be described and illustrated in Part B of this volume. This chapter, the
first of six that comprise Part B, is concerned with steps that should be taken
to modify a preliminary formulation of a mathematical problem so that it
assumes a more manageable character. Section 6.1 deals with the advan-
tages and possible pitfalls of modifying a problem by deleting certain
terms, terms which seem to be relatively small and whose deletion makes the
problem easier to solve. In Section 2 we show that the use of dimensionless
variables reduces the number of parameters which determine the solution of a
problem. This is an aspect of dimensional analysis, a straightforward pro-
cedure but one that is more subtle and more useful than one first expects.
We demonstrate in Section 6.2 that dimensionless variables can be selected
in a number of ways, with the same relatively few dimensionless parameters
appearing in each case. In Section 6.3 we discuss a particular selection process
that gives rise to what are called "scaled" dimensionless variables. When
these variables are used in the dimensionless formulation of a problem, the
dimensionless parameters that appear acquire added significance: they provide
the relative magnitudes of various terms. Consequently, if correctly scaled
variables can be chosen (and this is not an easy matter), the problem is often
ripe for certain perturbation methods. There are two classes of such methods:
the simpler regular perturbation methods (the subject of Chapter 8) or the
more complicated singular perturbation methods (Chapter 9). Perturbation
methods provide one of the most useful sets of applied mathematical tools.
Thus it is well worth devoting some effort to understanding the procedure
discussed in Section 6.3, which sets up a problem so that perturbation
calculations become somewhat "natural."
Our pedagogical strategy in this chapter* is to present the various ideas
in the context of a particular problem, that of a point mass radially projected
from the earth's surface and subjected only to a gravitational force that
varies with distance. (All damping effects are neglected.) Since this" projectile
problem" is a very simple one, the reader should be able to focus all bis
* Sections 6.1 and 6.3 follow in outline, and in many cases word for word, the article" Simpli-
fication and Scaling" by L. A. Segel [SIA.M Rev. 14, 547-71 (1972)]. These are reprinted with
permission from SIAM Review (1972). Copyright 1972 by the Society for Industrial and Applied
Mathematics.
185
186 Simplification, Dimensional Analysis, and Scaling [Ch. 6
energies on mastering the new concepts and techniques. These, and the
material of the next chapter, are given nontrivial application in the treatment
of a physiological flow problem in Chapter 8. Mathematically, what is
required here is the solution of a boundary value problem for a pair of
nonlinear ordinary differential equations. Many of the ideas are used later
in the analysis of more complicated problems, involving partial differential
equations.
6.1 The Basic Simplification Procedure

Simplification of a given set of equations may make it possible to avoid
large machine calculations or massive analytic work and still to obtain useful
answers. Even if the equations can be solved exactly without undue effort,
simplified equations may yield a sufficiently accurate solution whose features
are more readily apparent than those of the exact solution. Drastic simplifi-
cation may allow rapid solution and immediate determination of whether one
is on the right track.
The following basic simplification procedure is used time and again by
applied mathematicians:
STEP A. Somehow identify terms that are relatively small. In practice the
smallness of a given term is generally gauged in relation to other terms in the
same equation.
STEP B. Delete relatively small terms and solve the resulting simplified
problem.
STEP c. Check for consistency. That is, use the approximate solution
just obtained to evaluate the neglected terms, to ensure that they are indeed
relatively small.
We shall now present some examples of the procedure. In doing so, we
shall use the symbol ,...,to mean "seems to be approximately equal to" and
~ to mean" is approximately equal to."
ILLUSTRATIONS OF THE PROCEDURE

Example 1. Suppose that the problem
x+10y=21, 5x+y= 7, (la, b)
is nontrivial. In (la) the coefficientof the xterm is small compared to the coefficient
of the y term, so it is tempting to assume that the x term may be neglected to first
approximation. Omitting this term, we obtain
y...., 2.1, x....,-!(7- 2.1) ~ 0.98. (2)
Approximating the unknowns by means of the values given in (2), we estimate that
the ratio of the first term on the left side of (la) to the second has magnitude
0.98/21 ~ 0.05. This number is small compared to unity, so that our approximation
appears consistent. Also, our approximate values of x and y are close to the true
values x = 1, y = 2.
Sec. 6.1] The Basic Simplification Procedure
Example 2. (The projectile problem). Consider a body of constant mass m that

is radially projected upward from the earth's surface with initial speed V. Let R
denote the radius of the earth and let x*(t*) denote radial distance from the earth's
surface at time t*. (We use starred variables here so that in later discussions of this
example we can introduce unstarred dimensionless variables.) If we neglect air
resistance, the governing equations and initial conditions are
d 2x* gR 2 dx*
d(t*) 2 = - (x* + R) 2 ' x*(O) = 0, dt* (0) = V. (3}
(We have merely equated acceleration with force per unit mass, using the inverse
square law in a form that correctly gives acceleration -g when x* = 0.)
If Vis small in some sense, the displacement x* should be small compared to R.
Instead of (3) we can then consider the simplified problem
d 2 x* dx*
-
dt* 2
= -g
'
x*(O)=O
'
dt*(O) = V, (4)
with solution
dx*
-
dt* = -gt*+ v, x* =- !gt* 2 + Vt*. (5)
At the time t* = Vg-1, when the speed vanishes, we find from (5) that x* reaches
a maximum of !V 2 /g. Thus x*/R is at most }V 2 /gR, and our approximation is
consistent when V 2 is small compared to gR.
There is a feeling of general well-being attached to the property of being

consistent, but precisely what does consistency imply in the present context?
To generate interest in the answer to this question, we now present some
examples that illustrate some possibly disturbing aspects of the basic simplifi-
cation procedure.
TWO CHASTENING EXAMPLES*

Example 3. Consider the equations
O.Olx + y = 0.1, x + lOly = 11. (6a, b)
In (6a) the coefficient of the xterm is small compared to the coefficient of they
term; so let us neglect the former term to obtain
y-0.1, x"' I I - 101(0.l) = 0.9. (7)
Using (7), we estimate that the ratio of the first term on the left side of (6a) to the
second has magnitude (0.01)(0.9)/0.1 = 0.09. Because this number is small compared
to unity, our approximation appears to be consistent. But the exact answer is
y = I, x = -90; therefore, our "approximation" for y is off by a factor of 10,
while our "approximation" for x is off by a factor of 100 and has the wrong sign
also.
See C.Lanczos, Applied Analysis (London: Pitman, 1957), pp. 198ff. and p. 276 for other
examples.
To see what went wrong, let us generalize (6) and consider the following
equations for x(t) and y(t):
ex+y=O.l, x+IOly=ll. (8)
Equations (8) reduce to (6) when t = 0.01. The approximation (7) corresponds
to taking t = 0. But is it true that x(e) :::::.x(O),y(e) :::::y(O) when e = 0.01?
As soon as we ask this question, it becomes clear that we shall be in trouble
if x(e) and y(e) change rapidly as t varies near e = 0. That this is indeed the
case can be seen from the exact solution to (8):
0.9 0.1 - Ile
x = I - !Ole' y = 1- IOle ; e #=I/IOI. (9)
Example 3 illustrates the fact that an approximation which appears

consistent may not actually be so. In approximating x(0.01) and y(0.01) by
x(O) and y(O), we simplified the equations by neglecting the ex(e) term. This
is a consistent procedure if IO.Olx(0.01)I is small compared to ly(0.01) I,
Assuming that we only knew the approximate solutions x(O) and y(O), we
checked for consistency as best we could by examining IO.Olx(O)/y(O) I
This ratio has the satisfactorily small value 0.09 so that the approximation
appeared consistent. We were deceived, however, since the exact result is
0.0lx(O.Ol)I = _
09
1 y(0.01)
We thus distinguish apparent consistency (the approximation to the neglected
term is small) from genuine consistency (the term neglected is truly small).
The third step in the basic simplification procedure should really be called
checking for apparent consistency.
It is not surprising that our approximation in Example 3 was a poor one,
because it was apparently consistent but not genuinely consistent. But it is
disheartening that an approximation can be apparently consistent but not
genuinely so. Even more disheartening is the fact that an approximation
that seems genuinely consistent may be very inaccurate, as is illustrated in
the next example.
Example4. (Wilkinson,1963,p. 41). The polynomial

(x - l)(x - 2)(x - 3) (x - 20) = x 20 - 210x19 + (10)
19
has as zeros the first 20 positive integers.If the coefficientof x is changed by the
addition of ex111, where e= -2- 23 ~ -1.19 x 10-s, then the smaller zeros are
almost unaltered. But the larger zeros are so radically altered that the changed
equation has five pairs of complex conjugate roots, given roughly by 10 0.64i,
12 1.7i, 14 2.5i, 17 2.Si, and 20 1.9i. Whether the added term ex19 is
evaluated using the approximate" zeros or the exact zeros, it is still small in
magnitude compared to the retained term -210x 19 Both apparent and genuine
consistencyseemverified,but the approximationis again a poor one.
CONDITIONING AND SENSITIVITY
We have established the existence of wretched consistent approximations.

It is clearly necessary to examine matters with more care. We shall find that
certain general remarks can be made, but that further understanding requires
a somewhat detailed analysis of various problem classes.
As was illustrated in Example 3, knowledge that a problem has been
altered by a small amount is significant only if the problem is well conditioned
in the sense that small changes in its statement are accompanied by small
changes in its solution. The problem is termed ill conditioned if this is not the
case. It will also be convenient here to use the terminology that the solution
to a problem is sensitive to neglect of a term T if such neglect cannot be ex-
pected to result in a small change in the solution even when the neglected term
is genuinely small. Thus Wilkinson (1963) reports that the smaller zeros of
(10) are not sensitive to alterations in any of the coefficients (these alterations
can be regarded as the neglect of certain terms in an obvious way), and even
the larger zeros are not sensitive to small changes in certain coefficients. But
the larger zeros are sensitive to small changes in some of the coefficients,
namely, those multiplying the higher powered terms in the polynomial.
The only blanket statement that it seems possible to make concerning the
value of consistency checks is that the lack of apparent consistency is almost
certainlyassociatedwith poor approximation.To show this, let x symbolize the
true solution to a problem and let x denote the approximate solution that
one obtains when the term T is neglected. Neglect of T is apparently or
genuinely consistent, respectively, if T(x) or T(x) is small. (The notation used
here is symbolic. x could be a vector function, for example, and T could
depend on the components of x and several of their derivatives.) If the
solution is sensitive to the neglect of T, then, by definition, such neglect
cannot be expected to lead to a good approximation. If the solution is not
sensitive to the neglect of T, then
[T(x) small] implies [x ~ x] implies [T(x) small],
assuming that T varies continuously. The finding of a large apparent error

T(x) is therefore logically inconsistent with the supposition that the term
neglected is genuinely small, and neglect of a genuinely large term cannot be
expected to lead to a good approximation.
The imprecise words " small" and "large" appear again and again,
but these words can only be given a definite meaning in the context of a
particular problem. Thus the numerical magnitude of an acceleration (say)
can be changed by altering the time scale, but this change can have no effect
on the question of whether the acceleration is negligible or not. To mention
a slightly different example, if a certain temperature should be 1C but was
calculated as 1.1C, the error is 10 percent. If the Fahrenheit scale is employed,
the percentage error is only about 0.5 per cent, but it is without significance
Simplification, Dimensional Analysis, and Scaling [Ch. 6
190
that the latter percentage is much smaller than the former. Whatever temper-
ature scale is used, the decision as to whether an error is acceptable or not
rests on scientific grounds and must be independent of the units employed.
N o T E The difference between the alterations of the time scale and the
temperature scale is that the former involves only the choice of a unit, while
the latter also involves the choice of a zero.
To make further progress, it is necessary to consider particular types of
problems. We shall discuss the determination of the zeros of a function,
and second order ordinary differential equations. See Segel (1972, op. cit.)
for similar remarks on linear algebraic equations.
ZEROS OF A FUNCTION
Let a zero of f(x, s) be x(s) so that
/[x(s), s] = 0. (11)
Suppose that e is a smaII parameter, and let.an approximation x<0 >to the zero
x(.:) be found by setting e = 0, so that /[x< 0 >, OJ = O.The equation error in
equating f(x, 0) to zero rather than /(x, s) is f(x, s) - f(x, 0). The error is
apparent or genuine, depending on whether this expression is evaluated when
x = x<0 >or x = x(e). Thus, for the genuine equation error g, we have
g =f[x(e), s] - f[x(e), O]= -ef~ 0 >x'(O)+ O(e2).

(We continue to use a superscript zero to indicate that e has been set equal to
zero.) The apparent equation error or residual satisfies
r =f[x<0 >,c:]- /[x<0 >,O] =/[x<0 >,c:]= c:/!0 > + O(e2 ).
We wish to relater to h(e) = x(e) - x<0 >,the error in the solution. We shall
be retaining only the lowest powers oft. To the lowest order,
h(e) = sx'(O).
But
O Ei/[x(s), s] - /[x< 0 >,O]= e{f~0 >x'(O)+ /! 0 >}+ O(s2).
The vanishing of the expression within the curly brackets implies two things.
(i) The genuine error and the apparent error are equal to first approximation.
(ii) To first order, the error is given by
r
h= -(0). (12)
f,,
As is generally true, the error depends not only on the residual r but also
on a measure of the condition of the problem. In this case the " condition"
1i
depends on 0 >.In particular, if /~ 0 > is small, then a small change inf is
associated with an 0(1) change in x; in other words, a small change in the
191
equation is associated with a nonsmall change in the solution. If such ill con-
ditioning is present, one expects a large error in spite of a small residual.
By contrast, for well-conditioned problems [!~0 > of order of magnitude unity],
(12) shows that a large residual r precludes a small error h.
If we write (12) in the form
(~) = - ftO)
(~)'

we see that when relative error hfx is to be estimated, it is the relative residual
r/x that is relevant, where xis the solution. Except for cases of ill conditioning,
r can be compared to the approximate solution. Consequently, in Example 4
one should not examine the ratio of the neglected term to a term retained but
the ratio of the neglected term to the estimated solution. The latter ratio is
ex.19
/x. Fore= -i- 23 , this ratio is small for x close to unity but becomes
large for x much bigger than 2. There is a clear warning of difficulty for the
larger roots. Actually, the warning is somewhat premature, for here the
conditioning factor acts to keep the errors reasonable for zeros up to about
10. [See Wilkinson (1963) for a full discussion of sensitivity in this problem.]
Normally, the various terms in a polynomial might be expected to be of the
same size with coefficients of unit magnitude; in such a case, comparison with
any term retained will give the magnitude of the desired ratio between the
residual and the solution. The present polynomial is not" normal," however.
Many of its coefficients are very large, but the terms nearly cancel because of
sign alternations.'' Abnormal" equations with such large but nearly canceling
terms are not a rarity; this must be kept in mind when comparing a term
neglected with an allegedly "typical" term retained.
SECOND ORDER DIFFERENTIAL EQUATIONS

Consider an initial value problem for x(t, e):
d
f(t, x, x, x,e) = O; x(O) = A, x(O) = B, lei~ 1, =di. (13)
(~ means "much less than.") The following reasoning closely parallels that
just used in the discussion of the equationf[x(e), eJ = 0.
An approximate solution x<0 >(t)is found by setting e = 0:
f(t, x<0 >,x<0 >,.x<
0 >,O)= 0.
In setting e = 0, one makes an equation error of

f(t, x, x,x,e) - f(t, x, x,x,0).
To find the residual, one evaluates the equation error when x = x<0 >:
r =f(t, x<0 >,x<
0
>,x<
0
>,e) = e[/ 5 ] ... 0 + O(e2 ).
The genuine equation error g satisfies

g = -f(t, x, x,x,0) = -e[f 2 x. + f 3 x. + / 4 x,J,=o + O(t 2 ).
(Subscripts denote partial derivatives.) But
o= f(t, x, x, x,e) - f(t, x<0 >,x<0 >,.x<
0
>,O)
= t[f2x. +/3i.. +f4x, +fsle=O + O(e2 ).
Thus, as before, using a superscript to remind us that all terms are evaluated
at e = 0, we have
(0) _x(O) +/(0) j(O) +/(0) X(O) _ -j<O) (14)
/ 4c 3 2,- s
We conclude from (14) that the residual and the genuine equation error are
the same to lowest order. Also, we can now obtain an expression for the error
in terms of the residual. To do this, we make the definitions
F(t) = (/~0)' /~O)' 110>), (15)
b(t) = (X - x<0>,X - j(O), X - 5c(O>)_
We note that
0>) + O(e 2).
h(t) = e(x~0 >,i.~0>,x~
Thus to lowest order, (14) implies that
Fh = -r. (16)
Now
!rl = !Fl !hi !cos (F,h)I ~ IF! !hi
so that
,b1~m
lrl
We again see that a large residual precludes a small error when the problem
is well conditioned, while ill conditioning breeds large errors.
Note that both the error vector h and the condition vector F involve not
only the behavior of the solution, but also of its first two derivatives. Thus,
even if a small change in the equation is associated with a small change in
the solution itself, ill conditioning arises if, for example, the solution's
second derivative is markedly affected.
It is difficult to make any general statements from (16) about whether a
small residual implies a small error. Ill conditioning must certainly be ex-
cluded. Further, one must somehow satisfy oneself that the smallness of r
does not arise from cancellation of nonsmall errors in .x1>, i.~0>,and x~ 0
>.
In particular cases one can often make stronger statements. To illustrate this,
consider
x+(t+ex)- 2 =0; x(O) = 0, x(O) = 1. (17)
193
As is demonstrated in Section 6.3, (17) is the governing equation of the

appropriately nondimensionalized version of the projectile problem of
Example 2 when the parameter e = V 2/Rg is small. In this case F = (0, o.I)
and r ~ -.2txC 0 >(t), so one can immediately deduce from (16) that
!ex!
0
)1S 2e max !xC0 >(t)I + O(a2). (18)
To first order in e, the above relation shows that the error in the acceleration
is small. Two integrations permit estimates of the error in x itself (Exercise I).
An important point can be introduced via considerauon of small vibrations
of a simple nonlinear pendulum. As is indicated in Section 7.1 the proper
formulation of this problem is
8 + e - 112 sin (e1 12 (J)= O; 6(0) = l, 8(0) = 0, 0 < e ~ I. (19)
Here O(t) is a measure of angular displacement at time t, and a is a measure
of the initial displacement, which is assumed to be small. The zeroth approxi-
mation is oco> =cost. Using (16) to express the error, we find that
e[0!0 >+ 8!0 >] = -r = e{k)[f1<
0 >]3. (20)
We seeillustrated here why it is difficult to bound the error from a knowledge
of the residual, for (20) is nothing less than a differential equation for B!0 >.
We are faced with the problem of obtaining an estimate of the solution
to a linear differential equation from a knowledge of the magnitude of its
forcing term. There is no universal formula here, but the following simple
calculation is typical of what can be done. Using the method of variation of
parameters and observing that the initial conditions 8~0 >(0)= ~ 0 >(0)= 0
hold. we write
8
Smee lrl S
6,
BT
le0!
0
>(t)IS 6 for OS t ST.
As is well known, there is indeed an increase of error with T.

From another point of view, the current example illustrates the fact that
perhaps the best way to obtain information about the validity of the first
approximation is to compute the second approximation. After all, 8!0 > is
nothing more than the O(a) coefficient in the power series expansion of
8(t, e). Equation (20), which governs 9!0 >, was obtained by what can be
regarded as the parametric differentiation approach to perturbation theory.*
See General Sensitioity Theory by R. Tomovic and M. Vukobratovic (N.Y.: Elsevier, 1972)
for a detailed discussion of the calculation of functions like 6, and of the role of such functions
in understanding the sensitivity of various systems to parameter variation.
194
RECOMMENDATIONS
In our analysis of the basic simplification procedure, we began a probe
into the relation between the residual and the error. In one sense, we only
scratched the surface of the subject, for an important part of numerical
analysis concerns itself with how to obtain successivelybetter approximate
solutions to various problems by successively decreasing residuals.* Our
main concern, however, is with problems which are so formidable that
successive approximations are out of the question-one can barely obtain
the first approximation. For such problems, which are frequently encountered
by the applied mathematician, our discussion leads us to make the following
recommendations.
(i) Use the basic simplification procedure in difficult problems. Not
much extra work is required to estimate the magnitude of neglected
terms, and you will at least learn where your simplification is almost
certainly invalid (when the residual is large).
(ii) Although it is difficult to bound the error associated with a given
residual, it may be helpful to regard the residual as an extraneous
forcing and to use physical intuition to estimate its effect. Another
possibility is to replace the residual by its mean or maximum value
and to evaluate the effect of this constant forcing. To spot hidden
ill conditioning, make small random modifications in the problem
and solve again (Lanczos, 1957, op. cit., p. 170).
(iii) Beware of comparing a term neglected with a term retained. To
estimate relative error, it is best to compare the neglected term with
your approximate solution.
(iv) An applied mathematician studies simplified models to gain an
understanding of complicated situations. In this spirit, regard a
deep study of the simpler classical problems of numerical analysis
and perturbation theory as not only of value in itself, but also of
value in its indication of what to expect in more complicated prob-
lems.
(v) If possible, compute an" extra" term and use it to estimate the error
in the first approximation. [For an illustration of this recommenda-
tion, see the discussion centered around Equation (7.1.25).]
EXERCISES
1. Use (18) to provide a bound for Iex~0 >(t)I-
2. A slab is in a steady state with temperature T0 at x = 0 and T1 at x = 1.
The thermal conductivity is given by k = k 0 exp (s.x), where lel ~ I,
* With the aid of Green's functions and Lipschitz conditions, the presence of a small residual
can often be used to demonstrate the excellence of approximations. For example, see N. Ferguson
and B. Finlayson's discussion of a mixed analytic-numerical attack on problems involving heat
and mass transfer of a multicomponent system undergoing chemical reaction in catalyst pellets
[Amer. Inst. Chem. Eng. J. 18, 1053-1059 (1972)]. In one case, 13-digit accuracy is proved.
Sec. 6.2] Dimensional Analysis
195
Obtain an approximation to the temperature distribution by replacing

k by its average value. Check for consistency.
3. Water flowing from a small circular hole in a container has a speed v
which is approximately given by v = 0.6J2gh, where g is the gravitational
acceleration and h is the height of the water above the hole. Let A(h)
be the area of the cross section at height h.
(a) Derive
dh = -0.6 A(O) J2gh.
dt A(h)
(b) Suppose that the actual shape of the container is approximated by
A(h) = hC,ca constant. Solve the initial value problem. Discuss the
apparent consistency of the approximation. Are there circumstances
where the approximation could be both apparently consistent and
wretched?
6.2 Dimensional Analysis

Dimensional analysis occupies an unusual place among the techniques
available to a theoretical scientist. The formal manipulations required are
easy, but the ideas involved are seen on close inspection to be rather deep.
(See Bridgman, 1931.) It is now known that dimensional analysis shares
with the theory of similarity solutions to partial differential equations a
common foundation in the theory of continuous groups. (See Birkhoff, 1960.)
It would take us too far afield to cover thoroughly the more theoretical
aspects of dimensional analysis. For these the reader is referred to the
literature. Rather, we shall introduce the practical aspect of the subject,
largely by illustrating general ideas using the simple projectile problem
formulated in Section 6.l. (At the end of Section 4.1 we discussed some
aspects of dimensional analysis, in connection with heat flow.)
We shall show how to nondimensionalize a differential equation. In
addition, we shall illustrate how nondimensionalization can be used even if
the equations governing a phenomenon are unknown, provided that the
relevant parameters can be listed. Included in our remarks on the advantages
of dimensional analysis is a brief discussion of the use of scale models. We
assume a rudimentary familiarity with the notion of physical dimension,
e.g., in the statement" acceleration, as indicated by its typical units cm/sec2 ,
has dimension length/time 2 ."
PUTTING A DIFFERENTIAL EQUATION INTO DIMENSIONLESS

FORM
We can illustrate the main ideas of dimensional analysis very simply on the
projectile problem, which was already referred to in Example 2 of Section
6.l. We repeat the equations that govern the displacement x*(t*) at time t* of
a particle which is radially projected at initial speed V:

d 2 x* gR 2 dx*
d(t*)2 =- (x* + R)2' x*(O) = 0, dt* (0) = V. (Ia, b, c)
Here R is the radius of the earth, g is the gravitational acceleration, and x*

is taken to be zero when the particle rests on the surface of the earth.
To put (1) into dimensionless form, we follow a two-step procedure for
nondimensionalization, which can be used for a large class of problems.
STEP A. List all parameters and variables, together with their dimensions
(in terms of fundamental units such as mass .,{{, length !i', time ff). In the
present instance, we have the following list:
Variables Dimension
Dependent variable x*
Independent variable t*
Parameters
Gravitational accelerationg
Initial speed V
Earth radius R
STEP B. Let v* be a variable. Form a combination ofparametersp* with

the same dimensions as v*. Introduce v*/p* as a new dimensionless variable.
Here R has the same dimensions as x* and Rv- 1 has the same dimensions
as t*. Thus we introduce the new variables y and -r defined by
x* t*
y= R' -r = Rv- 1
(2a, b)
These variables are dimensionless, because their numerical value is the same
whatever standard of measurement is used. For example, if R = 4000 miles
and x* = 8000 miles, then y = 2. If the kilometer is used as our standard
length, we have R = 6436 km, x* = 12,872 km, and still y = 2.
Substitution of the new variables of (2) into (1) requires the chain rule.
Thus (writing everything out in full)
2 2
d x* _ d (yR) _ R d (dy) _ R d (dy d-r) _ R d (dyV\
d(t*) 2 - d(t*) 2 - dt* dt* - dt* d-r dt* - dt* d-riJ
2 2
_ V d (dy) _ V d (dy) d-r _ V d y (3)
dt* d1: d1: d1: dt* R d1:2
A little further calculation (Exercise la) gives the final result:
d 2y I dy
e d1:2 =- (y + l) 2 , y(O) = 0, d1:(0) = l, (4)
Sec. 6.2] Dimensional Analysis 197
where
v2
=- (5)
gR
Note that the dimensions of the parameter t are
1 2
(!eff- ) =!e2-2s--2+2 =!eoffo
::es--2 !e .
That is, t is dimensionless.

REM AR Ks . (i) In (2), R is called an intrinsic reference length; Rf V is an
intrinsic reference time. Generally speaking, intrinsic reference quantities are
defined to be standards of measurement formed from the parameters of a
given problem. (ii) It can be shown that only dimensionless combinations of
parameters (like e), sometimes called dimensionless groups, appear in a
properly nondimensionalized problem. (iii) We have used starred quantities
for dimensional variables so that we could use more convenient unstarred
variables for the final dimensionless quantities. (iv) Computations like (3) are
continually required in nondimensionalization. The calculations are equi-
valent to a formal removal of constants so that they precede the derivative,
like this:
d 2x* d 2(yR) R d 2y V 2 d 2y
(6)
d(t*) = d(RV 't') = R v- 2 d't'2 = R d't'2
2 1 2 2
(v) The temperature 0 often has to be included as a fundamental unit, in

addition to mass JI, length !e, and time f7. Moreover, it can be of advantage
in certain special problems to assume the existence of additional fundamental
units. (See Exercise 7.)
Explicitly noting parameter dependence, we observe that (1) has a solution
of the form
x* = x*(t*; g, R, V), (7)
while the solution to (4) can be written
y =y('t'; t). (8)
The solution (7) depends on the three parameters g, R, and V, but in (8) only
the single dimensionless parameter t appears.
As an alternative to (2b), we could base our reference time on the accelera-
tion g, writing
x*
Z=- (9)
R'
with which (1) becomes
d 2z I dz (0) = "',,112.
d-r:~= - (z + 1)2 ' z (0) = 0 , (IO)
d't'1
Now we expect a solution of the form

z = z(r 1 , e). (11)
There are innumerable reference times, since h(e)Rv- 1 will serve, for any
function h. Whichever of these reference times is employed, it will still be the
case, as in (8) and (11), that the dimensionless distance from the earth
depends only on a dimensionless time and the single dimensionless parameter
e.
A priori it would seem advantageous in a given problem to compare
lengths, for example, with an intrinsic reference length rather than with an
arbitrary length such as the distance between two scratches on a certain metal
bar. As our example illustrates, this is the case-and the advantage is that
the number of parameters which appear in a problem is reduced when dimension-
less variables are employed. (To be more precise, no parameters disappear,
but they occur only in certain dimensionless combinations.) Our example
also illustrates the fact that in all but the simplest problems, nondimensional-
ization can be carried out in a variety of ways, each of which confers the same
reduction in the number of parameters that appear.
We stress that since all parameters appear in the dimensionless version of
the projectile problem only in the combination e, any dimensionless result
to be gleaned from the projectile problem depends only on the parameter e.
For example, suppose that we are interested in the dimensionless time rM
when the projectile achieves its maximum height. Using the dimensionless
formulation (4), we see that this time can be determined as the solution of
the equation
dy' - 0
dt t=<M -
But since y = y(r, e), the above equation shows that rM = f(e) for some
function f In terms of the original variables we have that the dimensional
time to maximum height, 1:,, is given by
1:,
Rv- 1 =f
(v
2
gR .
) (12)
NONDIMENSIONALIZATION OF A FUNCTIONAL
RELATIONSHIP
We shall now show that ( 12) can be deduced even if the governing differen-
tial equation of the problem is unknown, as long as all variables and para-
meters that should appear in such an equation are known. Again, we illustrate
a general procedure on the projectile problem. (We shall give the particular
example immediately below each of the six required steps.) The procedure is
based on the idea that since the parameters of the problem have dimensions
of the form .,{{0 ft'b_rc, one should attempt to find dimensionless parameter
combinations by taking products of parameters raised to a power.
Sec. 6.2] DimensionalAnalysis
199
STEP A. Assume that a dimensionless quantity of interest is a function <1>

of the dimensional parameters of the problem.*
ttt
R112 -112 = <1>(V,
g, R). (13)
9
STEP B. Consider a term n made up of a product of powers of the quan-
tities in the argument of <1>.
(14)
STEP c . Insert the dimensions of the various quantities
Dimension of n = (ft'ff- 1Y,-1(ft'ff- 2 Y,-
2ft'a 3
= fi'a' +a2+a3f7-a, -2a 2
STEP D. To obtain a dimensionless parameter n, require each exponent of

.,I{, ft', ff, etc., to vanish, and thereby obtain a system of linear algebraic
equations
STEP E . Find the general solution to the system of algebraic equations,

in terms of arbitrary constants c1 , c2 , , etc.
1X2 = c1 , c1 arbitrary;
(15)
1X1 = -2c1, IX3 = -(-2c1 + C1) = C1,
STEP F . Substitute back into n to obtain an expression of the form
n~1n 22 Any dimensionless combination of the original arguments of (f) can
be expressed as a function of then;. Thus <1> can be replaced by a function of
the n 1
In the present case, from (14) and (15) we have
, = (gR)c'
n = v-zc,gc,Rc, vi .
Since n is dimensionless for an arbitrary choice of c1 , in particular we obtain
a dimensionless parameter gRJV 2
=
n1 when c1 = I. Any other dimensionless
combination of V, g, and R must be a function of n 1 Since the left side of (13)
is dimensionless, the right side must be also. [Otherwise we could reach a
contradiction by evaluating <l>(V, g, R) for two different sets of measurement
standards.] Thus it must be that
R112 -112
tZt =
(gR)
</> y2 (16)
9
for some function </>.
The above result can be written
1: = R112
9 -112<t>(~~). (17)
* A parameter should be included if it would appear in a mathematical problem that would

lead in principle to the determination of a single-valued function Cl>.
200
It is perhaps not obvious that this is equivalent to
t:C= Rv- 1
1(;;), (18)
a form of (12) that we derived from the differential equation formulation of

the problem. But (16) can be written
1; = Rv-1R-112v
9 -112<1>(~~). (19)
If we can choose a function <psuch that (19) is true, we can choose a function
f(s) = s 112 ef>(l/e)to make (18) hold, and conversely.
We have validated our assertion that even without knowing the governing
equation and initial conditions, one is able to show that the quotient of
the time 1; and Rv- 1can depend only on the single combination V 2 /gRof
the three parameters V, g, and R. To see the value of this ability, imagine that
someone was collecting data on tZtfrom a number of experiments on various
planets. An ignorant person might try to organize the data by plotting ,;
versus V for fixed values of R and g. He would then choose several other
values of R and obtain several corresponding curves on the same page
(Exercise 2). He would repeat his plotting on another page, choosing another
value of g. The complete presentation of his data would require a book.
But if he knew about dimensional analysis, he could plot all his data on a
single curve of t:C/Rv-1 versus V 2 /gR.
Taking a slightly different point of view, we remark that it is a considerable
advance in understanding to know that, purely on dimensional grounds,
the dimensionless time to reach maximum height can only depend on V2 /gR.
To give another simple example, one can deduce by dimensional analysis
alone that the period during which a simple pendulum executes a small-
amplitude oscillation is independent of the mass of the bob, and in fact is a
constant multiple of jL{g. (See Exercise 3.) A less trivial illustration of the use
of dimensional analysis (which we shall merely mention) is provided by its
role in the universal equilibrium hypothesis put forward in 1941 by the
Soviet scientist Kolmogorov. This relatively simple hypothesis is a landmark
in the still elusive quest to understand turbulent fluid flow. See G. K.
Batchelor's Homogeneous Turbulence (New York: Cambridge U.P., 1960),
pp. 114ff.
USE OF SCALE MODELS

We have illustrated the fact that dimensional analysis is useful to an
experimenter because to determine the behavior of the system, he or she
need only make measurements of the relatively few dimensionless parameters
that characterize the system. There is another reason for the usefulness of
dimensional analysis to scientists and engineers. With it, one can show that
Sec. 6.2) Dimensional Analysis
201
scale models* will duplicate the behavior of the original system provided
that the governing dimensionless parameters have the same values in the two
systems. Since theoreticians must have an appreciation of the nature and
limitation of experiment, we shall devote a few paragraphs to a discussion
of scale models.
A scale model has the same shape as the device of primary interest but is
of a more convenient size. The object of the present remarks is to stress that
model performance which closely mimics that of the given device can only be
expected if not only size, but also other attributes, are proportioned so that
the important dimensionless parameters are identical in both device and
model.
Ship design provides a good example of the use of models. Suppose that
one is interested in the amount of work per unit time P (for power) which
must be applied to keep a certain ship of length L moving in a straight line
at constant speed U. Work must be supplied primarily to replace energy
that is either wasted in making waves on the water surface or that is dissipa-
ted because of the viscosity of the water. A hydrodynamicist would thus
expect that
P =f(U,L,g, p, v), (20)
where g is the acceleration due to gravity, p is the density of the water, and v
is the kinematic viscosity of water. [Material in II on water waves and on
viscous fluid motion gives the background required to make the assumption
(20) with confidence, but for present purposes it will do no harm to accept (20)
without scepticism.] By dimensional analysis (Exercise 4) one can deduce that
p
pL2U3 = ef,(Fr,Re), (21)
where Fr = Uf(Lg) 112is called the Froudenumberand Re = UL/vis called the

Reynolds number.These two dimensionless parameters are associated with
wave resistance and viscous resistance, respectively. This should not be
surprising, since only the first contains a measure of the force of gravity that
is overcome by the waves and only the second contains a measure of fluid
viscosity.
Suppose that one constructs a scale model of the ship in question, of
exactly the same shape but one-hundredth the size. (If primes are used for
model parameters, this means that L' = 10- 2.) In order that the model
experiment be associated with the same Froude number as the full scale ship,
one would have to move the model one-tenth as fast as the speed U. For then
U' 10- 1 u
Fr'= (L'g)112= (lo-2Lg)112 = Fr.
We emphasize that in this discussion the word "model" docs not refer to a mathematical
model but is used in the familiar sense, which applies, for example, to a small replica of an airplane.
To obtain the same Reynolds number, one would have to use a fluid whose
kinematic viscosity v is one-thousandth that of water. No such fluid exists.
(Air, for example, has a kinematic viscosity which is about 15 times greater
than that of water.) It thus seems that one cannot make a model that
reproduces the dimensionless parameters of the full scale ship.
Contrary, perhaps, to expectation, the fact that the Reynolds number for
the full scale ship motion is apt to be enormous does not save the situation.
[If a ship 100 m long moves 10 m/sec (about 20 miles/hr), since v = 10- 2
cm2 /sec for water, the Reynolds number is about 109 .] It is reasonable to
expect that if one knows how P behaves for Re= 107 and 108 , say, that a
simple extrapolation will give the behavior for Re = 109 . It is a central fact
of fluid mechanics, however, that "the fl.ow characteristics do not simply
flatten out when Re becomes large or small. They continue to vary and for
many geometries give big and unexpected 'kicks' as R passes through values
which are quite surprisingly high." (Lighthill, 1963) (The" kicks" arise from
phenomena such as the sudden transition from smooth laminar to eddying
turbulent flow or from other qualitative changes in the flow pattern.) In
assessing ship resistance, one finds "kicks" as the Froude number varies also.
An important reason for this is the sensitivity of the resistance to the nature
of the interaction between bow and stern waves.
Ship design, then, serves as an example of an area in which model experi-
ments cannot provide easy answers. Thus ship design remains more of an
art than a science, although a considerable contribution to the subject has
been made by workers in theoretical fluid mechanics.*
A possible source of error in model experiments is that the correctness of
deductions made by dimensional analysis obviously requires the correctness
of the basic assumption that the quantities of interest depend in essence only
on the parameters listed. If an important parameter is ignored, model
experiments cannot be expected to give accurate results. Waves from a
sufficiently small ship model, for example, may be strongly influenced by
surface tension T. The corresponding dimensionless parameter here can be
taken to be the Bond numberBo = pgL2 /T. If the model does not have the
same Bond number as the fu]l scale object, then errors will result if surface
tension is important in either experiment.
Excellent physical understanding is clearly required to ascertain all the
important parameters in complicated situations. The difficulties that arise
when such understanding is lacking are illustrated by the scale-up problems
which arise in chemical engineering. New chemical plants, although geo-
metrically similar to pilot plants and appropriately similar in many other
ways as well, sometimes do not work properly because the pilot plant did not
A good review article is G. Gadd, "Understanding Ship Resistance Mathematically,"

J. Inst. Math. Appl. 4, 43-57 (1968). Also see D. H. Peregrine," A Ship's Waves and Its Wake,"
J. Fluid Mech. 49, 353-60 (1971).
Sec. 6.2] Dimensional Analysis 203
reproduce correctly some dimensionless parameter which unexpectedly has a

major influence on plant operation. Chemical plants typically involve such
complicated interplay between flow and chemical reaction that it is no
surprise that model experiments are difficult to design.*
To cite another area of great current interest, physiological problems are
similar in many ways to the problems that face chemical engineers, but are
even more complicated, so that model experiments must be regarded with
even greater caution. Nevertheless, such experiments can be useful. (See
Section 8.2.)
SUMMARY
We have illustrated the fact that it is a routine matter to nondimensionalize

the governing equations of a problem. Furthermore, we have discussed the
use of dimensional analysis in situations where understanding is not sufficient
to allow formulation of the governing equations. What is necessary is enough
physical intuition to list the parameters that would appear in the governing
equations if these were known. From this list, a straightforward procedure
automatically yields a set of dimensionless parameters that characterize the
problem in question.
We have used physical intuition and specific examples in our discussion.
Our assertions and conclusions certainly bear further investigation. For
example, one should be able to prove in some sense that it is sufficient to
consider dimensionless groups which are formed by taking the products of
the various parameters, each raised to a power. Similarly, one should not
have to rely on "intuition" when one asserts that a dimensionless parameter
can only depend on dimensionless combinations of other parameters. One
approach to the required proofs is outlined in Exercises 9-12.
A detailed use of dimensional analysis on a nontrivial example will be found
in Chapter 8. For the use of dimensionless variables in a complicated system
of partial differential equations, initial conditions, and boundary conditions,
see the treatment of water waves in Chapter 7 of II. Nondimensionalization
of a less complicated system of partial differential equations is carried out
in Section 15.2.
It is important to remember that there are fewer, often very significantly
fewer, dimensionless parameters than dimensional (three as opposed to eight
in the example of Chapter 8). Whether they are obtained experimentally,
* Although not easy, model experiments are essential in chemical engineering. This is summar
ized in a famous phrase of L. H. Baekeland, "Commit your blunders on a small scale and make
your profits on a large scale." For an interesting treatment of the problems involved in predicting
large scale plant performance from small scale experiments, see the book by R. E. Johnstone
and M. W. Thring, Pilot Plants and Scale-up Methods in Chemical Engineering (New York:
McGraw-Hill, 1957).This book begins with the quotation," Vitruvius says that small models are
of no avail for ascertaining the effects of large ones; and I here propose to prove that this con-
clusion is a false one "-From the Notebooks of Leonardo da Vinci.
204
numerically, or analytically, results can thus be found with less effort and can
be displayed in a more compact and meaningful form. Nondimensionalization
should generally be employed on all but the simplest problems.
EXERCISES
1. (a) Complete the derivation of (4).
(b) Exhibit explicitly the type of contradiction that is mentioned
in the sentence in square brackets above (16).
2. Use your qualitative feeling for the answer and plot t1 versus V for
several values of R. Imagine that g is fixed. Refer all your plots to a
single set of axes.
3. (a) A pendulum is executing small vibrations. Show that it is impossible
on dimensional grounds that the period T depends only on the
length L of the pendulum and the mass m of the bob.
(b) Show by dimensional analysis that the assumption that T depends
on m, L, and g, the acceleration due to gravity, leads to the relation
T = k(L/g) 1' 2 , k a constant. [As in (2.2.18), by solving a simple
differential equation, one can show that k = 21t.]
t(c) Suppose that the pendulum is pulled out farther, so that its ampli-
tude of oscillation is no longer small. What can you conclude by
dimensional analysis?
4. Show that (20) implies (21). (The dimensions ofv are 9' 2 ff- 1 .)
S. The thrust T developed by a ship propeller in deep water (dimensions
of T = J{ f ff- 2 ) depends on the radius a of the propeller, the number
of revolutions per minute n, the velocity V with which the ship advances,
the gravitational constant g, the .density p, and the kinematic viscosity v
of the water (dimensions of v =f 2 ff- 1). Show by using the formal
approach of the text that
T (an aV
pa 2 V 2 = <p v' -;-Vag) 2
[See E.G. Keller and R. E. Doherty, Mathematics of Modern Engi-

neering, Vol. I (New York: Dover, 1961).]
6. Consider D, the diffusion constant for spherical particles of radius a
in Brownian motion. Here D depends on a, on the temperature T and
the viscosity of the surrounding gas, and on the Boltzmann constant
k = 1.38 x 10- 6 erg/degree. Show by dimensional analysis that Dis a
constant multiple of kTfa. (The dimensions of D and are f 2 fi- 1
and J{ !:- 1ff- 1, respectively.)
7. For a small sphere falling under gravity in a viscous fluid, it is observed
that the speed of fall is (after a short time) a constant v. Let a, p 1 , p 2 , ,
and g represent the radius and density of the sphere, the density and
viscosity of the liquid, and the gravitational acceleration, respectively.
205
(a)Using dimensional analysis with fundamental units of mass,

length, and time, show that v = (/p 1a)f[pi- 2 a 3g, p 2 /pi]. Use
the fact that viscous stress (force/unit area) equals the product of
and the velocity gradient (derivative of velocity with respect to
length).
(b) Since motion here is unaccelerated, we need not make use of the pro-
portionality of acceleration to force, and force can be treated as a
separate fundamental unit. Hence show that v = a 2 p 1g- 1 ef,(p2 / p 1).
[Stokes derived the formula ef,(x)= (j)(l - x), a result of great
utility. For example, it was used in the Millikan oil drop experiment.
Stokes's result is obtained in Exercise 3.6.6 of II.]
8. Consider steady nonturbulent incompressible flow of a liquid in
a circular pipe. The pressure difference !:J.pbetween the two ends
of the pipe should depend only on its length L, radius R, and the
maximum speed U of the fluid of viscosity and density p. Show
that according to dimensional analysis the situation can be des-
cribed equally well by either of the following equations:
~=f(pUR,pUL) ~=f(pUR~)-
!pU2 ' -tpU2 R
[It is reasonable to suppose that when L/R is large-say L/R > 20
-changing L/R should have little effect on the answer. In this case
the relation ll.p/-tpU2 = f(pUR/) should hold to good accuracy.
This illustrates the fact that physical reasoning sometimes shows
how to choose the dimensionless parameters so that one of them
can be neglected. For more examples of this type of reasoning, see
the valuable book of Kline'(l965), from which the present problem
was adapted. See Exercise 6 for dimensions of.]
Exercises 9-12 guide the reader through proofs of some of the principal
results of dimensional analysis. The approach is analytic, as in Bridgman
(1931). For an algebraic approach see Birkhoff (1960).
9. Given a fundamental unit of length like the meter, we know how to

measure the length between two given marks. Length satisfies the
change of units rule: If a new fundamental unit is chosen, 1/x times as
large as the old, then the new measurement gives a number that is x times
as large as the old. (Example: A centimeter is 1 50 times as large as the
meter; so, to change meter measurements to centimeter measurements,
multiply by 100.) Quantities that satisfy the change of units rule, such as
length, mass, and time, are called primary quantities.
(a) Derive the requirement of absolute significance of relative magni-
tude (ASRM): The ratio of two measurements of a primary quantity
is independent of the fundamental unit of measurement chosen.
2o6 Simplification, Dimensional Analysis, and Scaling [Ch. 6
(Example: Mass of hydrogen atom M 8 equals 1.67 x 10- 24 g.

Mass of electron Me equals 9.11 x 10- 28 g, M 8 fMe= 1800. Or
MH = 3.69 x 10- 21 lb, Me= 2.01 x 10- 30 lb,andMH/Me = 1800.)
(b) Give another example of ASRM.
10. Secondary quantities are measured by combining the measurements of
primary quantities in a prescribed way. (Example: Measurement of the
secondary quantity speed equals measurement of length divided by
measurement of time.)
(a) Let (pi, p 2 , , Pn)and (q 1, q2 , , qn) denote two sets of measure-
ments of primary quantities. (Example: Pi and q 1 are distances
transversed by cars p andq in timesp2 and q2 .)Letf(Pi,P 2 , ... , Pn)
and f(q 1 , q 2 , , qn) denote corresponding measurements of a
certain secondary quantity. (Example: For average car speed,
/(pi, p 2) = p 1/P2.) Explain why we should postulate
f(Pi, ,Pn)
f(X1P1> , XnPn)= f( ) f(x1q1, , Xnqn) (22)
qi,, qn
for any X;, p 1, q 1 ; i =I, ... , n. (Note that the distinction between
primary and secondary quantities is somewhat arbitrary.)
;(b) Differentiate (22) with respect to x 1-keeping x 2 , , Xn fixed-and
then set x 1 =I; i = 1, ... , n. Show that
f(Pi, P2, , Pn) = g(pz, , Pn)P~',
where a 1 is a constant and g is an arbitrary function. Deduce
f(P1, Pz, , Pn)= Cp~'Pi2 P:",
where C is a constant. Usually, we take C = 1. Thus the measure-
ment of a secondary quantity is expressible as a product of powers
of measurements of primary quantities. The secondary quantity
is said to have dimensions p~'pi2 p:". (We use the same symbol
for a primary quantity and its measurement.) [Example: Speed has
dimensions (length) 1 (time)- 1 .] A secondary quantity is dimension-
less if all its dimensional exponents a 1 are zero.
(c) Show that if the fundamental units of measurement for primary
quantities are decreased by the factors Yi, ... , Yn, then S' and S,
the new and old measurements of a secondary quantity, are related
by
S' = ~ J:"S.
11. Let M be the net mass of water that has leaked out of a certain container,
since the container was opened at time t = 0. Suppose that if M is
carefully measured in grams and the time t in minutes, the results fit the
following law:
M = 5 - 5 exp (-t). (23)
207
(a) How will (23) be altered if kilograms are used as a mass scale and
seconds as a time scale?
If, like (23), an equation is a true description of a situation for
some given system of fundamental units, the equation is said to be
correct (in that system of units).
{b) Show that no matter what mass and length scales are used, the
measurements will be described by the formula
C1 M = 5 - 5 exp (-C 2 t), (24)
1
where the constants C1 and C2 have the dimensions (mass)- and
(time)- 1 , respectively.
Equation (24) is called complete because it is true whatever
fundamental units are used (as long as the dimensional constants
C1 and C2 are chosen correctly). Although not complete, (23) is a
correct description of the facts. As in the transition from (23) to
(24), any correct equation can be made complete by introducing a
dimensional constant in front of each observed quantity.
12. This problem outlines a proof of the Buckingham pi theorem, which
states (roughly speaking) that any complete equation implies an equation
in which all variables appear in dimensionless combinations.
(a) Let R 1 , .. , Rm be measurements of m (primary or secondary)
quantities. Suppose that there is one and only one functional
relationship
(25)
connecting these measurements and that (25) is complete ..Suppose
further that there are n fundamental units and that each is decreased
by a factor xi, i = 1, ... , n. Show that there must be a relationship
of the form
(x~ 11 x~12 x: 1 "R 1 , , x~m1 x:m"Rm)= 0. (26)
(As the proof unfolds, it may be helpful to work a particular
example that illustrates the general results for which the exercise
calls.)
(b) Deduce from (26) that
m
L ak1RkiR1, .. ,Rm)= 0,
k=l
where a subscript i on </>denotes the partial derivative with respect

to the ith argument.
(c) Assuming that ak1 # 0, introduce the new independent variables Tk,
k= l, ... ,m, (27)
so that
where
Show that
where the subscript k on Cl>

denotes a derivative with respect to the
kth argument. Show also that all quantities Tk have dimensional
exponent unity with respect to the first primary quantity.
(d) Introduce another new set of independent variables Uk:
Tk
U:k=- k=l, ... ,m-1, Um=Tm,
T'm
so that
x(Ui, ... , Um)= 0, (28)
where
x(U1 , ... , Um)= Cl>(U
1 Um, U2 Um, ... , Um).
Show that ox/oUm= 0. Show also that the quantities Uk are

dimensionless with respect to the first primary quantity. Thus (25)
implies (28), a relation involving m - I quantities that have a
dimensional exponent zero with respect to the first primary
quantity.
(e) Show that the above conclusion remains valid even if some or all
of the quantities ak 1 are zero.
(f) Deduce finally the relation
F(1r:1, 7r:z' . ) = 0,
where the quantities 1r:k are dimensionless. Noting the change of

variables involved, show also that the 1r:k are products of powers of
R 1 , R 2 , , Rm, Show that "usually" there will be m - n such
dimensionless products.
REM A R K Suppose that F = 0 can be solved for n 1 and that
the expression for n 1 actually contains R 1 If one solves for R 1 ,
then each term in the resulting expression must be of the same
dimension. This result is called the principle of dimensional homo-
geneity.
13. Some cookbooks say that a roast should be cooked x minutes per
pound; others say x 1 minutes per pound for small roasts and x 2 minutes
per pound for large roasts. Discuss. [See M. S. Klamkin, " On Cooking
a Roast," SIAM Review 3, 167-69 (1961).]
Sec. 6.3] Scaling 209
6.3 Scaling
Many mathematical problems contain small or large parameters. We need
only concern ourselves with small parameters, for if J is a large parameter,
then we can introduce a small parameters= l/J.
Taking advantage of the smallness of a parameter is not as simple as it may
seem. We saw some instances of this in Section 6.1; these were connected
with ill conditioning and insensitivity.
To see another sort of difficulty that one has to contend with, consider a
function of a single independent variable and a single small parameter s.
Presumably, a first approximation to the function can be obtained by letting
s tend to zero with the independent variable fixed. But the results depend on
how the independent variable is chosen. As an illustration consider
u(x, s) =x + e-:x:t, for O < x ~ I, s > 0. (la)
Here
Jim u(x, s) = x. (lb)
t-+O
If we switch to the new independent variable ~ = x/e, we have

v(e, s) = u(ee,s) = se + e-~. lim v(e, s) = e-( (2)
....o
On the other hand, if we introduce 17= x/e2,we obtain
lim w(17,s) = I. (3)
-+0
Which of the three limits
x, exp (-e) = exp ( ~x),or I, (4a, b, c)
if any, gives the correct first approximation to u? This is not too difficult a
question, but let us proceed slowly.
Section 6.2 revealed an important class of situations where there is a
choice of independent variable. These stem from a degree of arbitrariness in
selecting the intrinsic reference quantities that are used as denominators
when defining dimensionless variables. This was iIJustrated using the projectile
problem (Equation 2.1):
d 2x* gR 2 dx*
d(t*)2 =- (x* + R)2' x*(O)= O, dt* (0) = v. (5)
Employing a reference length Rand a reference time Rv- 1, we introduced

the dimensionless variables
t*
T=-- (6)
RV-1
and obtained (2.4):

d2 y 1 dy
e d-.2 = - (y + 1)2' y(O) = 0, a/O)= 1. (7)
Another choice,
x*
z=""ii, (8)
led to (2.10):
d 2z I
d-ci = - (z + 1)2' z(O) = 0, (9)
Here
v2
e=- (10)
gR
If e is known to be very small compared to unity, then a naive person might
merely delete terms preceded by e. But this cannot be right, for from (7) one
would obtain
dy
-(y + 1)- 2 = 0, y(O) = 0, d-. (0) = I, (11)
while from (9) one would obtain

d 2z
-
d-.21
=- (z+I) 2, z(O)= 0, ::1 (0) = 0. (12)
These "approximate" problems are different; so at least one is inappro-

priate. In fact, both are inappropriate. Problem (11) has no solution, while
problem (12) has a negative solution (Exercise I) and can therefore only apply
beneath the surface of the earth. It is apparent that a term cannot be neglected
merely because it is preceded by a small parameter.
Enough seeds of doubt have been sown. It is time to weed out false ideas
and to reveal flourishing techniques.
In the last section we showed how to choose dimensionless variables, and
we pointed out that this generally can be done in a number of ways. In this
section we propose that dimensionless variables should be selected so that
if a term is preceded by a small dimensionless parameter, then that term
should be negligible to a first approximation.*
The process of choosing this particular set of dimensionless variables is
called "scaling." We shall demonstrate that scaling is far from a routine
matter, but it is nevertheless helpful to be clearly aware of the goal of this
process and of the pitfalls that remain even if the goal can be attained.
It does little good to know that a dimensional parameter is small, for such a parameter can
be made to take an arbitrary value by an appropriate choice of units.
DEFINITION OF SCALING
Since dimensionless variables can be chosen in a variety of ways, one
cannot expect that the appearance of a small dimensionless parameter will
inevitably signify the presence of a relatively small term. Scaling amounts to
nondimensionalizing so that the relative magnitude of each term is indicated
by a dimensionless factor preceding that term. More formally, in the process
of scaling one attempts to select intrinsic reference quantities so that each term
in the dimensional equations transforms into the product of a constant dimen-
sional factor which closely estimates the term's order of magnitude and a
dimensionlessfactor of unit order of magnitude. (For the time being we shall
use the phrase" order of magnitude" in the sense of" approximate size." We
shall shortly specify the meaning of this phrase more precisely.) Intrinsic
reference quantities that are selected by this process are called scales. Gener-
ally, scales differ for different parameter ranges. Also, we shall see that for
a given range of parameters, it may be necessary to choose different scales
for different ranges of independent variables.
SCALING THE PROJECTILE PROBLEM
To illustrate the scaling procedure, let us consider the projectile problem (5)
in situations where the projectile's distance from the earth's surface x* is
always small compared to the earth's radius R. Such limited motion occurs
only when the initial speed Vis "sufficiently small." On dimensional grounds
we can assert that V must be small compared to a multiple of (Rg) 1' 2 , the
only combination of parameters other than V with dimension length/time.
When x* 4; R, it is clear that the acceleration has order of magnitude g,
the gravitational acceleration at the earth's surface. Now if a projectile is
launched with initial speed V and is then acted on by a uniform deceleration
of magnitude g, the projectile will momentarily come to rest (at its maximum
elevation) in time V/g. Taking the average of its initial and final speeds, we
estimate that in time V/g it will move a distance equal to (tV)(V/g) = f V 2 /g.
To keep the factor tin this expression might imply more accuracy than we
can legitimately claim-remember that we have replaced the continuously
changing speed of the projectile by the average of its initial and final values,
and we have ignored the change of the force of gravity with distance. We
thus take V 2 fg as our estimate of the order of magnitude of x*.
REM ARK . Our initial assumption that the parameters are such that x*
is always small compared to R is now seen to require that V 2 /g be small
compared to R. Because we have thought more deeply about the problem,
it is not surprising that this is a more precise statement than the requirement
"V must be small compared to a multiple of (Rg)112 ," which is all that one
can deduce on dimensional grounds.
We have completed the most difficult part of the scaling procedure, esti-
mation of the size of various terms in the special circumstances under con-
sideration. We now show how to take formal advantage of the above
estimates, which show (i) that the displacement x has order of magnitude
V 2 /g and (ii) that the acceleration d 2 x*/dt* 2 has order of magnitude g.
Using (i), we assert that the dimensionless displacement x should be defined
by
x
x = (V2g-1). (13)
With this change of variable, x will be replaced according to the equation

v2
X*=-X. (14)
g
The first factor on the right side of (14), (V 2/g), correctly and explicitly shows
the order of magnitude of x*. As x* is oforder of magnitude V 2 /g,(13) shows
that the dimensionless factor x must have order of magnitude unity, as
required by the scaling procedure.
Because of estimate (ii) we must choose a time scale T such that if the
dimensionless time t, defined by
t*
t =-, (15)
T
is introduced, then the term d 2 x*/dt* 2 is transformed into the product of a
dimensionless term and the constant g. But, from (13) and (15),
d 2 x* V2 d 2x
d(t*) = gT 2
2 dt 2
The requirement V 2 /(gT 2 ) = g gives T = Vfg as an equation defining the time

scale.
With the appropriate scaled dimensionless variables
x* t*
x = vzg-1 and I=--, (16)
Vg-1
the governing equation (5) becomes

d 2x gR 2 dx
g df =- (xV 2g- 1 + R) 2 '
x(O) = 0, dt (O) = 1. (17a,b,c)
To simplify, we divide both sides of (l 7a) by g and divide the numerator and
denominator of its right side by R 2 This gives
..
X=- (1 +a )-2 , x(O) = 0, x(O) = 1. (18)
(Note that because of this division, it is relative orders of magnitude that are
now explicitly displayed. In particular, the factor e gives the order of magni-
tude of x divided by R.)
Since the projectile problem (5) has been nondimensionalized in the

particular way called for by the scaling procedure, we are confident that when
e is small compared to unity, the term ex is small compared to I. If the lowest
approximation to x(t) is denoted by x< 0 >(t), we have, from (18),
x<0 > = -1, x<0 >(0)= o, x<0 >(0)= I. (19)

Thus
x<0 > = t - !t 2 and Os; x<0 >s; t for O s; t s; 2.
Particularly in its dimensional form,
d 2 x* dx*
d(t*)2 = -g, x*(O) = 0, dt* (0) = V,
everyone knows that (19) provides the zeroth approximation to the solution
of the projectile problem. But few inexperienced individuals can without
hesitation arrive at (18), a form of the problem that permits ready deter-
mination of higher approximations. (See Section 7.2.)
The inappropriateness of the previously obtained" approximate" problems
(11) and (12) when e ~ I can now be ascribed to the fact that in both (6)
and (8) the choice of dimensionless variables was not in accordance with the
scaling procedure. That these incorrect " approximate" problems do not
have a sensible solution is comforting and typical but, unfortunately, not
inevitable. (Recall Examples 3 and 4 of Section 6.1.)
We must emphasize that the choice of scales depends on the parameter range
under consideration. As an illustration of this, observe that we have scaled the
projectile problem when e ~ 1, but if e- 1 is a small parameter, then new
scaling is required. To mention one facet of the new situation, when e- 1 1s
small the initial speed is so large that the particle soon passes many earth
radii from the earth's surface, and it becomes wrong to estimate that d 2 x*/dt* 2
has magnitude g.
Another point can now be illustrated. Full understanding of a problem
requires one to have in mind a physical interpretation of each dimensionless
parameter which appears in its dimensionless formulation. In the present case,
the only such parameter is e = V 2/gR. As was mentioned under (18), e can be
interpreted as the ratio of two lengths. In the denominator is the earth radius
R. In the numerator is V2 jg, an estimate of the maximum height achieved
by a projectile shot from the surface of the earth with a nonlarge initial
speed V.
ORDER OF MAGNITUDE
It is helpful to make a precise definition of the phrase "order of magni-
tude." A number A will be said to have order of magnitude 10",nan integer, if
310..- 1
< IA! s;310".
Simplification,DimensionalAnalysis, and Scaling [Ch. 6
214
Since log 103 ~ !, a virtually equivalent characterization of a number A with

order of magnitude 10" is
n-t<los10IAI Sn+t.
By the orderof magnitude of a function/defined over a certain region, we mean
the order of magnitude t0f the number M, where M is the maximum (or
perhaps the least upper bound) of 1/1over the given region. Note the distinc-
tion between the order of a function defined over a certain domain, denoted
by the O symbol (Appendix 3.1), and the numerical orderof magnitudeof a
function as defined above.
SCALING KNOWN FUNCTIONS

The scaling of known functions is now a straightforward matter. Considera-
tion of a few examples will generate some intuition.
Let us at first consider a phenomenon that is governed by the first order
ordinary differential equation
F(u*, ~::) = 0. (20)
Suppose that the independent variable x* is restricted to an interval/* (which

may be infinite). It will be convenient to refer to u* as a velocity component
and to imagine that x* is a spatial variable, although our discussion applies
to any dimensional dependent and independent variables u* and x*.
Let L be the length scale and Uthe velocity scale. Introduce dimensionless
variables by
u*
U=- (21)
u
We have
du* U du(x)I
u*(x*) = Uu {~), dx* = L --;I;-x=x*/L. (22)
If U and L are indeed appropriate scales, then the combinations of U and L

that appear on the right side of the equations in (22) must be reasonable
estimates for the maximum absolute values of the terms on the left. Now
ordinary scales can usefully be regarded as estimates of exact scales in which
U and U/L actually equal these maximum absolute values. For exact scales
U = max lu*(x*)I (23)
x in I*
and
u
-= max du*I (24)
L x* in J* l dx*
From (24), using (23),

_ lu*lmax
(25)
L= ldu*/dx*lmax
Equations (23) and (25) give explicit expressions for the exact velocity and
length scales in problems governed by (20). Note from Figure 6.1 that L can
be interpreted as the base of a right triangle whose altitude is Iu* Imax and
whose hypotenuse has the slope Idu*/dx* Imax.Moreover, Figure 6.1
supports the qualitative statement that the length scale is an estimate of the
shortest distance over which the function undergoes a significant change in
magnitude.
u* (x*)
!slope I= ldu*/dx* lmax
FIGURE 6.1. The length scale L according to (25).
Example 1. Find velocity and length scales U and L when

u*(x*) = A sin AX*, -oo <x* < co,
where A and ,\ are positive constants.
Solution. Obviously, U = A. Since
ldu*I
dx* max = IA,\ COSAX*Imax = A,\,
(25) impliesthat L = ,\ - 1
Example 2. Find U and L when
u*(x*) = A [x+ exp (-;*)],Os;;x* s;; 1,

where A and e are positive constants, e ~ 1.
Solution
U= iu*lmu~ A,
ldu:1
=IA
-Ae-
dXmax
1
exp (-x*)I~ Ae-
Emu
1
,
so L = e. See Figure 6. 2.
u*
u* (x 0 ) = A (x + e-"'1')
!slope I= ldu*/dx* 1.,.,
O E
FIGURE 6.2. The length scale is approximately e for the function u,

where u(x*) =A[x + exp (-x/e)]; A and e are positive constants,
0::,;;x::,;;1, and e ~ 1.
Suppose that the problem is governed by

du* dNu*)
F ( u*' dx* ' ... ' dx*N = 0, (26)
generalizing (20).t Equation (23) is still an appropriate definition of the

velocity scale U, but in choosing the length scale L we have to consider each
of the derivatives
diu*(x_*)= u_[diu(~)] i = l, 2, ... , N.

d X *' L' dX I x=xJL '
We must choose L so that
i = I, 2, ... , N. (27)
t Generalization to problems in which there is more than one independent variable or more
than one dependent variable is also possible. See Exercises 3 and 4.
Sec. 6.3] Scaling
217
To make our estimate as sharp as possible, we must select the largest value of
L such that (27) is satisfied. It follows that L is the largest constant such that
for i = 1, 2, ... , N; or
U [ U ] 1/2 [ U ] 1/N
L = smallest of Idu*Jdx*Imax' Id2u*/dx*2 Imax ' ' IdNu*/dx*NImax
Example 3, If (26) is the governing equation, find the length scale L for the
sinusoid of Example 1.
Solution
11
[ld'u*/~*lmaJ =
1 [A~,]'" =A-i
so that L = ,\ - 1, regardless of how many derivatives are involved in (26). (See
Figure 6.3.)
u*
ISJopel= ldu*/dx* lmax
1
FIGURE 6.3. The length scale ,\- of a sinusoid is approximately
- 1
one-sixth of its period 2'1T1\
Note that x* does not explicitly appear in the governing equation

F(u*, du*fdx*) = 0 of (26). This is characteristic of a wide class of spatially
homogeneousproblems. For such problems "one place is just as good as an-
other." Formally, spatial homogeneity manifests itself in the fact that the gov-
erning equation, having no explicit appearance of the independent variable
x*, retains its form when subject to the axis translation x* --+ x* + constant.
As would be expected, length scales for problems not characterized by
spatial homogeneitymust take into accountthe variationsof the inhomo-
geneity. For example, even if the magnitudes of u* and du*/dx* are correctly
assessed, without further thought one cannot be sure of the magnitudes of
such terms as u* exp (x* I>..)and ( l Ix*)( du* I dx*).
The function considered in Example 3 has a length scale that is independent
of N, the number of derivatives that must be taken into account. Here is an
example of a function whose length scale does depend on N.
Example 4. If (26) is the governing equation, find U and L if

u*(x*) = M + A sin ,\x*, -oo <x* < oo,
where M, A, and A are positive constants.
Solution
U=M+A,
The dependence on N is weak providing M/A is not large compared to unity.

When M is large, the length scale is large.
ORTHODOXY
Two matters of concern remain even after the process of scaling enables
one to put a problem into a form that explicitly reveals the presence of terms
(if any) having relatively small orders of magnitude. The first matter is that
neglect of relatively small terms may have a large effect. This has been dis-
cussed in Section 6.1. In what follows, we shall assume that relatively small
terms are negligible.
The second matter of concern stems from the fact that the order of magni-
tude of a term estimates that term's maximum magnitude. If the magnitudes
of the various terms in an equation stray too far from their maximum values,
then the order of magnitude estimates may give misleading impressions of
their relative sizes over much of their domain of definition. We say that a
term in an equation satisfies the orthodoxy requirement in a given domain if
the term's absolute value does not differ drastically from its maximum
absolute value except_perhaps for a negligible portion of the domain in
question. Suppose that in a certain equation the order of magnitude of a
term T1 is much greater than that of a second term T 2 If the first term fails
to satisfy the orthodoxy requirement, then it may be smaller in absolute value
than the second term for much of the domain in question, even though its
maximum absolute value is much greater than the maximum absolute value
of the second term. (See Figure 6.4.)
Both unorthodoxy and difficulties due to it are less widespread than one
might think. Unorthodoxy might seem almost inevitable, e.g., when the same
length scale must serve to characterize the change of several dependent
variables; but this is not the case. The reason seems to be that the dependent
variables are bound to combine so that certain differential equations are
satisfied. It is difficult to imagine how such variables can differ wildly in
behavior. Presumably, this is why it commonly occurs in practice that a
length scale selected by concentrating on just one of several dependent
variables frequently serves well for all of them.
"Harmless" unorthodox functions occur in the study of small amplitude
water waves. (See Section 8.1 of II.) According to linear theory the velocity
components and the (dynamic) pressure, and their derivatives, are unorthodox
Sec. 6.3] Scaling
219
10 10
(a) (b)
FIGURE 6.4. Two terms Ti and T2 are defined on the interoal I.

Their orders of magnitude are 10 and 1, respectively. (a) Ti satisfies
the orthodoxy requirement; T2 is small compared to Ti throughout I.
(b) Ti does not satisfy the orthodoxy requirement; T2 appears to be
negligible only in the subinterval Ii, Ti appears to be negligible in ]3,
and both functions are of about the same size in the intermediate
subinteroa/ 12.
functions because they decay rapidly with depth. But all decay at exactly the
same exponential rate so that the relative orders of magnitude of the various
terms are correctly estimated by examining what goes on near the water
surface. Such behavior is to be expected whenever the problem can be re-
garded as being governed by a system of linear equations with constant
coefficients.
Another common situation in which a degree of unorthodoxy is tolerable
involves oscillatory terms of relatively large maximum amplitude. Thus one
would say that terms of unit order of magnitude in an equation can probably
be neglected to first approximation if other terms are known to behave like
sinusoids of large amplitude. This is so in spite of the fact that the sinusoidal
terms are small near their zeros. Because there is almost no room for inter-
esting behavior in a function that must pass smoothly from values specified
on one side of a narrow region to nearby values which are specified on the
other side, it would appear that these small regions of unorthodoxy can be
ignored. (Compare Exercise 6.)
An important failure of orthodoxy occurs when dependent variables behave
like the function
u*(x*) = A[x* +exp(-:*)]. 0 < x* :5: 1, 0 < e ~ I, (28)

220
whose graph appears in Figure 6.2. Suppose that the term du*fdx* must be
assessed. Because t is small, the rapid change of the exponential in (28) means
that an estimate of Idu*/dx* I in an interval containing points near x* = 0 is
a gross overestimate for an interval not containing such points. For example,
when 3t ~ x* ~ 1,
du*I -1 -3
dx* max
::::;
At e .
l
Consequently L:::::e3t, a length scale more than 10 times as large as c:,the
length scale appropriate for O ~ x* ~ I. To satisfy the orthodoxy requirement,
we must split [O,I] into two parts and choose a different length scale in each
part. In the outer region, more than a few multiples oft from x* = 0, we have
Iu* Imax
:::::
A and I:: Lax:::::
A, so U = A and L = 1.
Introducing these scales, we obtain
u*
u= A, x=x*, so u(x, t) =A- 1
u*(x, t) = x + e-:xf.
In the inner region, within a few multiples oft from x* = 0, we find that
lu*lmax:::::
A and 1::1
....
,.::::: At-
1
, so U = A and L = t.
Using eand v for scaled variables in the inner region, we have
u* x*
v=-,
A
e=-,
t
To obtain a first approximation in the two regions, we let t-+ 0, keeping the
respective independent variables fixed. We obtain in the outer region
u(x, s):::: x, (29)
and in the inner region
v(e, t) ::::e-(, so u(x, t):::: e-xf. (30)
Our simple example illustrates that lack of orthodoxy in a given domain
can be remedied by introducing subdomains in each of which orthodoxy is
present. Different scales will be required in the different subdomains. Thus,
in the presence of unorthodoxy, one must not seek a first approximation with a
single scale but rather must seek different approximations in different sub-
domains. An approximation that is valid throughout the domain can actually
be found, but it must vary on two or more scales at the same time.
By our last point, as illustrated in (29) and (30), we have essentially answered
the question posed at the beginning of the section, as to which of the three
Sec. 6.3] Scaling
221
expressions in (4) is the correct first approximation to u. It remains only to

state that the approximation u ~ I of (4c) is only valid very near x = 0. This
is to be expected, for lengths are measured with respect to e2 , a quantity that
is small, even compared to the width of the inner region. The limit e -+ 0,
Y/= e- 2 x fixed, gives an approximation that is valid only within a few multi-
ples of e2 from x = 0; u ~ I is clearly the appropriate form for such an
approximation. Nothing of interest is gained here by considering an" inner-
inner" region of width O(e2 ).
In the present instance we have advocated the introduction of different
scales in different regions to overcome unorthodoxy, while in our earlier
discussion of large amplitude sinusoids we suggested that the unorthodoxy be
ignored. The region of unorthodoxy is narrow in both cases, but only in the
present case does the function change rapidly and so accumulate considerable
alteration.
An example of a situation in which more than one scale must be introduced
is provided by the projectile problem when s- 1 is large. The scaling of (16) is
still appropriate when the projectile is near the earth, but another scaling is
required when the projectile is far from the earth. As above, there is an inner
region (thickness ~ V 2 fg) and a much larger outer region (projectile more
than a few radii from the earth's surface).
It can happen that two scales coexist. An example is e-x sin x which
oscillates on an 0( I) scale with an amplitude which varies on a relatively
long O(e- 1) scale. Under some circumstances it is profitable to recognize the
simultaneous existence of more than one scale, e.g., in the slow change in
the frequency of a pendulum. (Compare Section 2.2.) Under other circum-
stances, it may be best to ignore the rapid variation or to average it out, even
though this means that large errors will be made in approximating the
derivatives of the answer. Averaging small scale irregularities in order to
reveal major trends is the strategy of many successful phenomenological
theories.*
SCALING AND PERTURBATION THEORY
Once a problem has been correctly scaled, one can in principle derive
arbitrarily accurate approximations by systematic exploitation, via pertur-
bation theory, of the presence in the equations of a small parameter. From the
present point of view, "regular" perturbation theory deals with orthodox
situations where one set of scales suffices." Singular" perturbation theory has
been developed to handle unorthodox situations in which more than one scale
must be introduced in order to obtain an approximation that is uniformly
valid throughout the domain. The elements of regular and singular pertur-
bation theory will be discussed in Chapters 7 and 9, respectively.
For an example see D. Drew's use of averaging methods to obtain field equations for two-
phase problems in mechanics [Stud. Appl. Math. 50, 133-66 (1971).]
SCALING UNKNOWN FUNCTIONS

Order of magnitude estimation and scaling require a knowledge of the main
features of the solution to the very problem one is trying to solve. How can one
obtain the information necessary for the required order of magnitude esti-
mates? We list six possibilities.
(i) Utilize experimental or observational evidence concerning the pheno-
menon in question. Often this indicates that a phenomenon is
primarily due to a balancing of two effects, and a scaling can be
obtained by balancing the terms that correspond to these effects.
(The derivation of the boundary layer equations in Section 3.3 of II
provides a good example of such balancing.)
(ii) Obtain hints from experience of related problems.
(iii) Solve highly simplified versions of the given problem. (This was
illustrated in our discussion of the projectile problem, where we used
knowledge of the answer obtained when variation of gravitational
force with distance is neglected. Knowing the zeroth approximation,
we employed scaling to adapt the problem for efficient determination
of higher approximations.)
(iv) (Inverse procedure.) Make certain order of magnitude assumptions
merely because the concomitant neglect of terms renders the problem
tractable. (For example, assume that nonlinear terms are negligible.)
Evaluate the neglected terms, perhaps using the approximate solution
to do so. If possible, select parameter ranges so that these terms do
indeed appear to be negligible. Hope that these parameter ranges are
characteristic of interesting phenomena and that the demonstrated
lack of inconsistency signals a justified neglect of small terms.
(This is illustrated in the discussion of linearized water wave theory,
in Section 7.3 of II.)
(v) Use a trial and error approach. Assume a certain scaling; solve the
resulting simplified dimensionless problem; and then check to see
whether the dimensionless terms are of unit order of magnitude.
(vi) Employ the results of computer calculations for particular but repre-
sentative values of the various parameters involved in the problem.
The last item forms the basis for a joint computational analytical inves-
tigation. For problems involving ordinary differential equations, analog
computers* are an invaluable tool; parameter variation is accomplished
merely by twisting a dial. The use of such a computer often provides enough
" feel " for the problem so that a simple and revealing analytic approximation
can be determined.
A form of scaling is required in analog computation. Units must be chosen so that various
magnitudes do not exceed the capacity of the machine. Here too, one must somehow obtain
information about the solution to the problem one is trying to solve. See, for example M. L. James,
G M. Smith, and J.C. Wolford, Analog Computer Simulation of Engineering Systems (New York:
International Textbook, 1966).
Models for natural phenomena often embody a system of partial differential

equations containing many parameters and involving two or three spatial
dimensions. To follow the temporal evolution of the solution even once by
numerical methods may be extremely demanding of computer time. There
may thus be no possibility of thoroughly understanding the effect of para-
meter variation by using a direct numerical attack. But if a few numerical
solutions are available, order of magnitude estimates can often be made
which provide the key to the simplifications required in an analytic investi-
gation. If successful, such an investigation will exhibit the dependence of the
solution on the various parameters and will also reveal the general features
of the solution that form the basis for physical understanding.
EXERCISES
1. Show that the solution to (12) is negative for -r1 > 0. (It is not necessary
to solve the problem in order to do this.)
t2. Find the "velocity scale" U and the "length scale" L for the function
x* < oo, when the governing equation is of
u* = A exp ( - ax*), b ::;;;;
the form (20). Comment on the fact that Lis independent of b.
3. Generalize the text's characterization of the length scale to the case
where the governing equation is F(u*, ou*/ox*, ou*/oy*) = 0.
4. Generalize the characterization of the length scale to the case where
the problem requires determination of two dependent variables u* and
v*. Suppose that the governing equations are
ou* ou* ov* ov*)

F ( u*, v*, ox*' oy*' ox*' oy* = O,
ou* ou* ov* ov*)

G ( u*,v*'ox*'oy*'ox*'oy* =O.
5. Concoct a specificexample that shows an effect of spatial inhomogeneity

on the length scale.
6. Solve for y(x; e):
d2y
dx2+ n2y = sin x + e, y(O) = I, y'(O) = 0.
Compare y(x, 0) with y(x, e) when e ~ I. Sketch. Thereby verify the

text's remarks (concerning harmless unorthodox functions) in the
paragraph preceding (28).
7. Why is the "harmless unorthodoxy" of small amplitude water waves
typical of problems governed by a system of linear equations with
constant coefficients?
224
8. This exercise concerns the projectile problem when air resistance is

considered but changing gravitational attraction is not taken into account.
The equation and initial conditions are
d 2 x* dx* dx*
m dt* 2 +k dt* = -mg, x*(O) = 0, dt* (0) = V.
(a) When Vis sufficiently small, scaled variables are still given by (16).
Why?
(b) Introduce the variables of (16) and show that the problem becomes
one in which the only parameter is p = kV/mg. Give a physical
interpretation of p.
9. (a) Oscillations of a pendulum released from rest are governed by
Equation (7.1.1) (with n = 0). Show that (7.1.3) defines properly
scaled dimensionless variables, assuming that the initial amplitude
a is not too large.
t{b) What choice of scaled variables should be made if a= 0 and n is
small?
t(c) Discuss scaling when both a and n are nonzero and small.
10. Show that the limits in (I b), (2), and (3) are not uniform for, respectively,
x e (0, I] ee(O, oo), '1e(O, oo).
CHAPTER 7
Regular PerturbationTheory
GI v EN the solution to a certain mathematical problem, how is the

solution determined when the conditions of the problem are slightly altered?
The systematic answer to this question forms the subject of perturbation
theory.
Section 2.2 contains some discussion of perturbation theory. In keeping
with the broad viewpoint of Part A, the treatment there is not detailed. Its
principal purpose is to illustrate how the approach of Poincare can remove
bothersome secular terms that appear in a nai:veapproach to including the
effects of other planets upon the motion of a selected planet about the
sun. This is done indirectly, by discussing the effect of small nonlinearity
upon the motion of a simple pendulum.
We are now in a more favorable situation regarding the study of perturba-
tion theory. This arises through the discussion in Chapter 6 of ways to
formulate a mathematical model so that a meaningful small parameter ap-
peared in it (if this is permitted by the conditions of the problem). In the
present short chapter, our object is to provide some detailed examination
of regular perturbation theory. Roughly speaking, this theory is appropriate
when a reasonable initial approximation is obtained by means of a straight-
forward attempt at simplifying a problem that involves a small parameter.
Section 7.1 provides details that were omitted in the treatment of the
pendulum problem given in Chapter 2. Section 7.2 illustrates various
aspects of regular perturbation calculations on the projectile problem.
(The mathematical formulation of this problem was the subject of much
detailed discussion in the previous chapter.) In all, three approaches to the
calculations will be presented: series, parametric differentiation, and suc-
cessive approximation.
Further application of regular perturbation theory, this time on a more
difficult problem involving a pair of ordinary differential equations, is made
in Chapter 8. Moreover, Chapter l O in II contains a lengthy analysis of the
use of this theory (with some modifications) in a water wave problem gov-
erned by a system of nonlinear partial differential equations. Taken together,
all these examplesprovide a foundation for the use of an important technique.
7.1 The SeriesMethodAppliedto the SimplePendulum

In this section we discuss the power series approach to perturbation calcu-
lations. The calculations are illustrated on the problem of the simple pendu-
lum. Except toward the end of the discussion, we have avoided essential
225
Regular Perturbation Theory [Ch. 7
226
reference to Chapter 6. Thus the material to be presented here could well be

read in conjunction with Section 2.2, where we presented an introduction
to perturbation calculations for the pendulum.
PRELIMINARIES
In Section 2.2 we introduced the simple pendulum, a weightless rigid rod
of fixed length L with a point of mass m at one end; the mass is free to rotate
about a horizontal axis at the point of suspension. There, or in introductory
physics texts, the governing mathematical problem is shown to consist of
the following equations of motion and initial conditions for the angular
displacement 8* as a function of time t*:
d2(J* 2 ll* 0
dt*2 + Wo sm u = t>O.
(1)
d8*
at t*= 0, 8* = a and dt* = n.
(Here 8* = 0 when the pendulum hangs vertically downward.) With the
"usual" assumption of small displacement, one replaces sin 8* by 8* and
obtains (Section 2.2)
8*(t*) = Q COS Wot* + fiwo
1
sin CL>ot*. (2)
We wish to improve this approximation in a systematic manner. For sim-
plicity we shall restrict our discussion to the case n = 0 (pendulum released
from rest).
The first step is to introduce dimensionless variables 0 and t:
8*
t = Wot*, 0=- (3)
a
The problem now takes the form
d20 sin (a0)
2 + = 0, 0(0) = l, d0 (0) =0. (4)
dt a dt
The new dependent variable 0 measures angular displacement in units of
the initial angular displacement a. The new independent variable t measures
time in units of ro01 , where 21tro0 1 is the period of small oscillations. The
introduction of these dimensionless variables has provided us with a problem
in which there appears a single parameter a, and we wish to develop an
approximate solution under the assumption that a is small compared to
unity.*
Those who have studied Section 6.3 should m:ognu.ethat the new variables are scaled.
The reference angle a is the largest value of B(t). The reference time is w 0 1 , in accord with
Example 3 of Section 6.3.
Sec. 7.1] The Series Method Applied to the Simple Pendulum 227
A basic tool in our analysis is the Maclaurin expansion

00
F(a) = L C;ai;
i=O
(5a, b)
Assuming convergence, the successivepartial sums of the above series provide

ever more accurate approximations to F for small a. Even to write down the
series, however, it is necessary that the function in question be infinitely
differentiable at the origin. On the other hand, if only a few derivatives are
known to exist, we can use the Maclaurin formula in the form of a finite
sum plus remainder:
[d"+lp ] a"+l
F(a) = Jo
n
C;a'
.
+ dd"+i (e) (n + l)!; ebetween Oand a. (6)
No convergence questions arise here, and bounds on the (n + l)st derivative

permit estimates of the error.
In the material to foJlow, for definiteness we shall proceed under the
assumption that various functions have infinite series of the Maclaurin type.
If these functions have only a finite number of derivatives, however, there
is a tacit understanding that the series is appropriately terminated.
To make sure that there is no confusion on the matter, we note that if a
function F depends on variables other than a, partial derivatives must be
used in the Maclaurin expansion, and the coefficientsare no longer constant.
Thus, if F = F(t, a), then (5) becomes
00 1 oiF
F(t, a)= L C;(t)a;; C;(t) = -:-,-;---'"
(t, 0).
l. ua'
(7a, b)
i=O
It is usually only feasible to obtain the first few terms in the various series.
To make sure this is done systematically, we must keep track of the terms
we have neglected. To do this we use the O notation, where by O(ai) we
mean a collection of terms containing jth or higher powers of a. For example,
3
sin a = a - ; ! + O(a5), cos a = 1 + O(a 2 ).
Assuming that our formal calculations are valid, this is consistent with the
precise use of the O symbol as defined in Appendix 3.1.
We now outline the series approach to perturbation theory as it is applied
to a differential equation. The approach is broken down into five steps. Each
step is immediately illustrated as it applies to the pendulum problem (4).
SERIES METHOD
Perturbation Method I. Assume that the dependent variable can be expanded
in a series of powers of the small parameter.
228 Regular Perturbation Theory [Ch. 7
STEP A. Substitute the power series into the differential equation. Here
we assume
00
0(t, a) = L 0;(t)ai. (8)
i=O
The substitution gives
f E>;ai+ a-
i=O
1
sin(a I:0iai)
i=O
= 0 (9)
where the validity of term-by-term differentiation has been taken for granted.
STEP B. Expand all quantities so that each term is written as a power series.
There is nothing further to do at present with the first term in (9); it is already
written as a power series. Treatment of the second term requires that we use
the Maclaurin series for the sine function to write
(10)
We aim to retain terms only through O(a 2 ), but during intermediate calcu-
lations we retain some smaller terms, to make sure that nothing crucial has
been neglected. We find
"' ) = 0o
a- 1 sin ( ai'2;,00ia' + a01 + a202 + O(a3 ) -
a2
6 [0~ + O(a)] + O(a4 )
(11)
STEP c. Collect all terms in the equation and equate to zero the success-
ive coefficients in the series. Using (11), we write (9) as
(E>o+ 0o) + a(0 1 + 0 1) + a 2(02 + 0 2 - i 0~) + O(a 3 ) = 0. (12)

Equation (12) has the form of (7a) with F = 0. From (7b), this implies
that Ci= 0 for all i. Assuming the necessary convergence, we are led to
require the vanishing of each coefficient in (12):
0 0 + 0 0 =0,
01 + 01 = 0, (13)
02 + 02 = i-0~.
Our original nonlinear differential equation has been replaced by a sequence
of linear differential equations. This is a characteristic feature of the per-
turbation method.
Sec. 7.1] The Series Method Applied to the Simple Pendulum
229
Sr E P D . Substitute the power series into the original initial (or boundary)
conditions, expand, equate the coefficients to zero, and obtain a set of initial
(or boundary) conditions to supplement the sequence of differential equations
obtained in Step (c). In the present case, 0(0) = 1 implies that
ex,
1= L E>iCO)a; or O = [0 0 (0) - l] + a0 1(0) + a2 E>i(O)+ O(a3). (14)
i=O
Hence we require that

0 0 (0) = l, 0 1(0) = 0, E>i{O)= 0, etc. (15)
Similarly, we require that
E>o(O)= 0, 0 1(0) = 0, 0 2 (0) = 0, etc. (16)
STEP E . Successively solve the sequence of differential equations and
boundary conditions obtained by Steps (A}-(o ). In the present case, we first
have to solve
0o + E>o= 0, 0 0 (0) = 1, 0 0 (0) = 0. (17)
The differential equation has the general solution
0 0 = A cost+ B sin t; A, B constants. (18)
The particular solution satisfying the initial conditions is
0 0 =cost. (19)
This is the usual approximation, in dimensionless form.
From (13), (15), and (16), E>1(t) satisfies
E>1
+ E>1= 0, E>1(0)= 0, E>1(0)= 0.
The solution is
=
E>1(t) 0. (20)
There is no correction at this stage.
Similarly, E>i{t)satisfies
02 + E>2 = !E>~= ! cos3 t, E>i{O)= 0, E>i{O)= 0. (21)
Since the solution of the homogeneous equation is known, we could use the
method of variation of parameters to obtain a particular solution of the
inhomogeneous equation.* The method of undetermined coefficients is
frequently easier if it is applicable. It can be used here if we perform the
often helpful step of using a trigonometric identity to transform an expression
involvingpowers of cost and sin t into an expression involving linear combinations
of these trigonometricfunctions. In the present case, the appropriate identity is
cos3 t = -!(cos 3t + 3 cost).
This is a general feature of regular perturbationcalculationsforordinarydiff'erentialequations.
Using this to transform the differential equation of (21) into
02 + 02 = 214 cos 3t + i cos t, (22)
we can now apply the method of undetermined coefficients. In doing so,

we must take note of the fact that the forcing term k cos t is a solution of
the homogeneous equation. Consequently, the usual form of the correspond-
ing particular solution must be multiplied by the independent variable t.
We find (Exercise 1) for the general solution to (22),
0i(t) = A cost+ B sin t - 1 J2 cos 3t + / 6 t sin t. (23)
Imposing the initial conditions, we obtain A = 1 ~ 2 , B = 0. *

From (19), (20), and (23), we see that the complete approximate solution,
taking terms of O(a2 ) or lower into account, is
0(t, a)~ cost+ a 2( 1 i 2 cost -Th cos 3t + 1\t sin t). (24)
Noteworthy is the higher frequency term cos 3t.
DISCUSSION OF RESULTS SO FAR
Knowing a correction to the usual approximation cos t, we can now de-

termine when this correction is small. We see that the terms proportional to
the bounded quantities cos t and cos 3t in the correction are small for all
time provided only that a is small. It is otherwise with the term proportional
to t sin t: the latter is small compared to unity if and only if -tt;a 2
t ~ 1.
It is becoming clear that we have different situations, depending upon whether
or not we examine the approximation (i) for a fixed finite time interval
[O,T], or (ii) for the unbounded interval [O, oo). Let us consider case (i) first.
For the time interval [O,T], the correction is small and the approximation
0(t, a)~ cost appears to be consistentt provided that -tt;a 2
T ~ 1. We
would expect, in fact, that the approximation is asymptotically valid as a! 0,
uniformly for t E [O,T]. By this we mean that we anticipate that the exact
solution and the approximation can be made to differ by an arbitrarily small
amount for the entire range of the independent variable provided the parameter
is taken sufficiently small. Symbolically, in the present case uniform asymptotic
validity of the approximation cos t requires that, given any e > 0, there
exists a positive constant a0 , which may depend on e and T but must be
independent of t, such that
I0(a, t) - cos t I < e whenever O ::;;;t ::;;;T,
Note that a multiple of the homogeneous solution, cos t, appears in the final answer t23).
Beginners sometimes write down just a particular solution to equations analogous to (22),
forgetting that they have no justification for discarding the homogeneous solution, which in any
case is needed to satisfy the initial conditions.
t We are using the terminology of Section 6.1.
Sec. 7.1] The Series Method Applied to the Simple Pendulum 231
Now consider the unbounded interval [O,oo). For a given initial angular
displacement a, the term -h(z2tsin t oscillates with ever-increasing amplitude.
Such behavior can be rejected on physical grounds. Thus our perturbation
scheme has apparently not provided an approximation that is asymptotically
valid for t E [O,oo). Methods for obtaining improved expansions will be
discussed in Section 9.2. For now, let us be content with the fact that all
appears well in [O,T].
Let us look at our approximations from a slightly different point of view.
At the beginning of our discussion we estimated the validity of the usual
approximation cos t by examining the next term. in the perturbation ex-
pansion. It is important to realize the advantage of this approach compared
to the basic simplification procedure of Section 6.1. For the pendulum prob-
lem, that procedure would entail making the approximation a- 1 sin a0 ~ 0,
finding the approximate solution 0 0 = cos t, and checking for consistency by
determining whether a- 1 sin a0 0 ~ 0 0 . Now
. a- 1 sin a0 0
l1m = I uniformly int, (25)
a--+O 0o
since j 0 0 I ~ I for all t, so that the approximation is apparently consistent
when a is small. But the true solution also satisfies j 0(t) I ~ I. (This is obvious
physically-a proof is requested in Exercise 4.) Thus(25) remains true when 0
is substituted for 0 0 and the approximation is genuinely consistent. In the
terminology of Section 6.1 the approximation cos tis nevertheless "wretched"
in that after sufficient time has elapsed its misestimate of the period of oscilla-
tion will bring about grave errors in its prediction of the true angular dis-
placement 0. The deficiency in the approximation was not signaled by the basic
simplification procedure but it was revealed by the appearance of the next term
in the perturbation expansion. This fact serves as an antidote to the examples
of wretched consistent approximations given in Section 6.1 and given again
here in our discussion of the pendulum.
It is hard to think of an instance in which the failure of an approximation
is not indicated by the appearance of the next term. Be that as it may, the
antidote is not universal because in complicated problems a forbidding amount
of additional work may be required to compute one more term. The simple
substitutionrequiredby the consistencycheckin the basicsimplification
pro-
cedure may be all that it is feasible to do. In spite of its hazards, a check for
consistency is well worth carrying out in such instances, for reasons that have
been mentioned in Section 6.1.
HIGHER ORDER TERMS
We can obtain higher terms in our perturbation series by continuing our

calculations. No new questions of principle arise, but the computations
eventually appear forbiddingly difficult.
232 Regular PerturbationTheory [Ch. 7
It is helpful to take advantage of the fact that the identical vanishing of

0 1(t) is not accidental. Indeed 0;(t) = 0 for all odd i. To see this, observe
that the governing differential equation (4) is unchanged if -a is substituted
for a. Assuming that (4) has a unique solution, this means that 0(t, a) =
0(t, -a). In other words, 0 is an even function of its second argument.
(What is the physical reason behind this fact?) From this it is not difficult
to show (Exercise 2) that all odd-ordered derivatives of 0(t, a) with respect
to a vanish when a = 0. Thus coefficients in the series (8) satisfy
0;(t) = ;
z.
[aie(t;
oa
a)] = 0,
a=O
i odd.
We conclude that there are no odd powers of a in the series for 0(t, a).
To obtain higher corrections, it is best to take advantage of this and assume
from the outset that 0(t, a) has an expansion in powers of a2
REM AR K The considerations of the previous paragraph show that
before starting perturbation calculations, you should see if symmetry arguments
can be used to prove that the proper expansion variable is the square of the
naturally occurring small parameter. If you miss such symmetry you will not
blunder because of your error of omission; you will only have to work
harder to obtain a given degree of accuracy.
By calculating more terms, we expect to improve the validity of our approx-
imation. This is true for the bounded interval [O,T]. In fact, it can be shown
that the series (8) is convergent when Ia I s; a0 , for some sufficiently small
a0 The convergence is uniform for t in [O,T]. That is, given any e > 0
there exists a nonnegative integer N 0 (which is independent oft) such that
Ie(t,a) - it 0;(t)a
1
I< e (26)
whenever N > N 0 , Ial s; ao, t e [O,T].

Implied by this convergence (see Section 3.2) is the fact that the series (8)
is asymptotic as a -+ 0, uniformly in t for t e [O,T]. This means that for
any e > 0 and any fixed N, there exists a positive number a 1 (which is inde-
pendent oft) such that
a-N \ 0(t, a) - J I
0
0;(t)a; <e (27)
whenever Ia I s; a 1, t e [O,T]. In terms of the notation of asymptotic analysis

(Appendix 3.1),
N
0(t, a) = L 0 1(t)a; + o(aN) uniformly for t e [O,T].* (28)
i=O
In our formal calculations we assumed that the remainder in (28) was O(a1'+1), a stronger
requirement than o(d').
Sec. 7.2] Projectile Problem Solved by Perturbation Theory 233
Since it is difficult to calculate very many terms of the series, the asymptotic
result is apt to be more useful than the convergence proof. But remember
the distinction. In a convergent approximation one considers the parameter
as fixed and thinks of improving accuracy by taking more terms. In an asymp-
totic approximation one considers a fixed number of terms and thinks of
improving accuracy by letting the parameter approach a "favorable" value.
Taking more terms does not improve the situation as far as the unbounded
interval is concerned. The O(d') terms contain contributions proportional
to tn, so all "correction" terms are 0(1) when t = O(a- 1). In Section 2.2
we briefly considered Poincare's method of obtaining an improved approxi-
mation. Another approach to this matter will be discussed in Chapter 11.
EXERCISES
1. Use the method of undetermined coefficientsto derive (23).
2. t(a) Show that the derivative of an even function is an odd function and
that the derivative of an odd function is an even function.
(b) Show that the second derivative of an even function is an even
function.
(c) Show that all odd-ordered derivatives of an even function vanish
at the origin.
3. (Lengthy.) Assume that (4) has the series solution
and calculate 0 2
t4. Prove that (4) implies that l0(t)I ~ 1.
5. Calculate a correction to the lowest order approximation when the
initial conditions of (1) are taken with a = 0, n " small."
7.2 Projectile Problem Solved by Perturbation Theory

In this section we treat by three different perturbation methods the pro-
jectile problem that was finally formulated in Equation (6.3.18). First, we
employ the series method that was introduced in the previous section.
By means of worked examples, we show how the somewhat lengthy calcula-
tions proceed. The student should not readthese examples, of course, but
should use them as a check on his own work.
The remainder of the section illustrates, on the same problem, the method
of parametric differentiation and the method of successive approximation.
SERIES METHOD
The dimensionless distance x of a projectile from the earth at dimension-
less time t is governed by the following equations and initial conditions:
x= -(l + ex)- 2 , x(O) = 0, x(O) = 1. (la, b, c)
Regular PerturbationTheory [Ch. 7
234
This was shown in Section 6.3 (compare 6.3.18). Recall that e = V 2 /gR,
where V is the initial speed upward, g is the gravitational acceleration, and
R is the earth's radius. We are interested in the effect of the variation of
gravitational attraction with distance in situations where the projectile does
not stray too far from earth. As discussed in Section 6.3, we should therefore
examine the solution to (1) when e is small. Following the perturbation
procedure outlined above, we assume a power series expansion
x(t, e) = x 0(t) + ex1(t) + e2xi(t) + (2)
and substitute it into the governing equation. Next, we use (2) to write the
term (1 + ex)- 2 as a power series in e. In principle this could be done by
employing the Taylor series formula but experience teaches that in con-
structing power series one should use the binomial expansion whenever possible.
It is best to perform preliminary manipulations if necessary so that the first
term in the binomial is unity, in which case the expansion takes the simple
form
I) I)(n - 2)
( 1 + J)n = 1 + J
n
+ n(n 2!- J2 + n(n - 3!
J3 + ... ' 111< 1. (3)
In the present case we have

(1 + ex)- 2 = 1 - 2ex + i-(-2)(-3)e 2 x 2 + O(e3 )
= 1 - 2e[x 0 + ex 1 + O(e2 )] + 3e2 [x 0 + O(e)] 2 + O(e3 )
= 1 + e(-2x 0 ) + e2 (-2x 1 + 3x5) + O(e 3 ).
Equating the coefficientsof various powers of e to zero, we find the following

sequence of problems.
x0 = -1, x 0 (0) = 0, x0 (0) = 1. (4a)
X 1(0) = 0, x1(0) = 0. (4b)
xi(O) =0, x2(0) = 0. (4c)
Example 1. Solve the system (4).

Solution. Equation (4a) yields upon integration
xo(t)= - t+ I, Xo(t) = t - !t 2. (5)
With this (4b) becomes

i1(t) = 2t- t 2, X1(0) =0,
so
(6)

From (4c)
= -3t 2
+ :!J-t 3
- H-t4,
so that
xi(t) = _,3 + gt4 _ -a,s, (7)
Thus, the solution is

x= t -it 2 + e(!t 3 - ii.-1
4
) +e 2
(- tt 4 + Ht 5 - .'.1.t6 ) + . (8)
Of interest is the time tmat which the projectile reaches its maximum height.
At this time the speed vanishes; i.e.,
(9)
We can find a first approximation to tm by solving x0 (t) = 0. From (5), this

I. To do better, we shall utilize the solution to the accuracy we
gives tm:::::::
have obtained. Combining (9) and (8), we thus derive the following equation
for tm:
1 - tm + 8[t;, - it!]+ 82[ -t! + ~ ~t!- ~~t;,] + 0(8 3) = 0. (10)

This is a quintic equation, neglecting 0(8 3 ) terms. But we are only interested
in the root that is approximately unity. Furthermore, we expect that this
root, like all the other quantities we have been seeking, has a series expansion
in 8. Since Equation (10) for tm is correct only through 0(8 2 ), we assume an
expansion correct to the same order:
(II)
It is now a straightforward calculation to obtain the coefficients in this

expansion.
Example 2. Calculate a1 and a2 in (11).

Solution. Inserting (11) into (10), we obtain
+ e[l + 2a 1e + O(e HI + 3a1e + O(e

0 = -a 1e - a 2 e 2 2
) -
2
))]
+ e [-1 + H - H + O(e)]
2
= e(-a1 + 1 - t) + e2(-a2 + 2a1 -a1 -!.) + O(e 3).

Upon equating each power of e to zero, we find
so
tm = 1+ fe +te 2 + O(e 3 ). (12)
Example 3. Determine the maximum height x .. =x(t .,).

Solution
x,,. = x(t,.) = t,. - it;+ e[tt!- 11.t!] + e 2[-tt! + tttii - fl;;t!] + O(e
3
)
= 1+ te + te - U1+ te + te + O(e )]2

2 2 3
+ eH(l + te + O(e2)) 3 - i'.(1 + te + O(e2)) 4 ]

+ e2[-t + tt-1-o + O(e)] + O(e3)
= 1 + te + te 1 - U1+ te 1 + te + fe 1 + O(e3)]
+ e[O(l + 2e + O(e2)) - i\{1 + te + O(e2))]
- :.e 1
+ O(e3 ),
so
x.,(e) =i +te +te 1
+ O(e3). (13)
The answers that we have obtained on carrying through the exercise of

treating the trajectory problem by perturbation theory are, in retrospect,
easy to understand qualitatively. The correction terms to tm and Xmare posi-
tive. This is to be expected. When the weakening of gravitational attraction
with distance is taken into account, the computed time for the projectile to
reach its maximum distance from the earth will be longer, and the farthest
excursion achieved will be larger, compared to the case when gravitational
attraction is deemed invariant with distance.
In the present problem the solution is of interest only during the finite time
interval [O,2tmlUnlike the pendulum problem, there is no difficulty caused
by the inevitable buildup of small effects over a sufficiently long interval.
We expect our approximations to be uniformly valid for O ::;;;t ::;;;2tm. In-
deed, using the exact solution, one can see that, for example, the series (13)
converges to the correct value of xm(e) provided that O ~ e < 2. (See Exercise
2.)
Recall that in Section 6.3 we showed that only if dimensionless variables
were chosen correctly (by scaling) did the "naive approximation" of neglect-
ing terms multiplied by a small parameter yield an appropriate mathematical
problem. From the present point of view, the "naive approximation" is
the first term in a series expansion of the solution. We draw the conclusion
that correctscalingis a prerequisiteto successfuluse of the regularperturbation
method.
Until now, we have confined our discussion to a single method, that of
series. We now treat briefly two closely related methods: parametric differ-
entiation and successive approximations.
PARAMETRIC DIFFERENTIATION
The idea of the parametric differentiation method is that equations for the
coefficients xi(t) in the expansion
co
x(t, e) = L xi(t)e; (14)
i=o
Sec. 7.2] ProjectileProblem Solved by PerturbationTheory
237
can be obtained by successively differentiating the problem for x(t, 8) with

respect to 8. After differentiating, one sets e = 0. The reason that this method
works is simply that the coefficients X; satisfy
xlt) = ;,
Z,
s)] .
[oix(t;
08 e=O
(15)
In the series method, of course, equations for the X;(t) are obtained by sub-
stitution and collecting. The two methods are equivalent in essence, but
sometimes the differentiation method yields slightly easier calculations.
We now present the method, illustrating it on the projectile problem.
PerturbationMethodIl
STEP A . Differentiate equations and boundary conditions with respect to
the small parameter. Consider the projectile problem of (6.3.18):
x(t, s) = -[l + ex(t, s)r 2
, x(O,s) = 0, x(O, e) = I. (16)
We obtain the following upon differentiating the differential equation with
respect to 8:
a:=+
o" 2(1 8X)-
3
a:+)
(8 a x . (17)
Note that we have had to keep in mind that x depends on s. It is convenient

to introduce the notation
(il( ) _ aix(t, 8). i = 0, l, ....
x t, 8 = 08i ' (18a}
In particular,
x<0 >(t,e) = x(t, 8). (18b)
With this (17) can be written
jCll = 2[1 + 8X(O)r3[SX(1) + x<O>], (19)
and a second differentiation (Exercise 3a) gives
j:(2) = 2[1 + ex<O)r3[8,t2) + 2.tt>] - 6[1 + 8,tO>r4[8X(l) + X(Ol]2. (20)
STEP B. Set the parameter equal to zero in the various equations. We
introduce another notation :
y<il(t) = x<0 (t, 0). (21)
[Comparison with (15) shows that they(i>differ only by a constant factor i!
from the X; that are used in the series approach to regular perturbations.]
Using (21), we obtain from (16), (19), and (20),
(22a, b, c)
Regular PerturbationTheory (Ch. 7
If we repeat Steps (A) and (B) for the boundary conditions, we obtain
y<0 >(0) = 0, y<0 >(0) = I. (23a)
y'l)(O) = 0, y'O(O) = 0. (23b)
y<2 >(0) = 0, y<2 >(0) = 0. (23c)
As the reader can easily verify, the equations for the Y; differ only by the
appearance of appropriate constants from the corresponding equations
for the X;, which were obtained above (Exercise 4). The same final result
is attained by a different route.
SUCCESSIVE APPROXIMATIONS (METHOD OF ITERATION)
Another approach to perturbation theory can be motivated in the following
way. Suppose that some problem has the form
F(z) = G(z). (24)
(One can think of both F and Gas polynomials for definiteness.) Suppose
further that the equation F = 0 provides a rough approximation z 0 to the
desired solution z. To obtain a second approximation, rather than ignore
G(z) entirely, we can approximate it by G(z0 ). We can then solve for a new
approximation Zi, which satisfies
F(z 1) = G(z0 ).
The procedure can be repeated by means of
n=2,3, .... (25)
Example. If
z - 2 = O.Olz3,
then an initial approximationto a solution is zo = 2. A better approximationis
z = 2 + 0.01(8) = 2.08.
The iteration method was briefly mentioned in Section 2.2, as a way to
improve upon an initial approximation in the pendulum problem. We now
present a more general description of this method, using the projectile
problem as an illustration.
Perturbation Method m. (For a problem involving a single ordinary

differential equation for an unknown function x.)
STEP A. Write the differential equation in the form F(x, x, x, ... , t) =
G(x, x, x, ... , t), where F = 0 will provide an initial approximation to the
desired solution. The initial approximation to the projectile problem (})
comes from solving x = -1. We therefore write (la) in the form
x+ I = -(I + ex)- 2 + l
or
.. e(2x + ex2 )
x+ 1 = . (26)
(1 + ex)2
REM ARK. In proceeding, we shall use the notation zi for the ith approx-
imation. This is to distinguish the present approximation from xi and y<O,
which were introduced in our discussions of Methods I and II.
STEP B Successively solve the equations
F(z 0 , z0 , z0 , , t) = O;
F(zn, Zn, Zn, ... , t) = G(Zn-l Zn-l Zn-l ... , t); (27)
n = 1, 2, ....
Usually, each approximation zi is made to satisfy the full initial conditions or
boundary conditions, but this requirement may sometimes be relaxed with
advantage. For the projectile problem we have the iteration scheme
z0 + 1 = 0, z 0 (0) = 0, z0 (0) = 1. (28)
.. s(2zn-l + sz;_ 1)
Zn+ 1 = = 0,
znCO) = I.
znCO) (29)
(1 + BZn-1)2 '
An example will show a useful consideration in problems that explicitly
involve a small parameter. Consider the quadratic equation
-ex 2
+x - I = 0, (30)
The iteration scheme for approximations to the root near 1 is
x0 - I =0, (3la)
or Xn = 1 + sx;_i, n = I, 2, .... (3lb)
Thus
x1 = I + e, x 2 = I + e(l + e)2 = I + e + 282 + e3,
x3 = I + 6(1 + s + 262 + 63 ) 2 = 1 + s + 262 + 563 + . (32)
What is noteworthy here is that the coefficient of 6 3 in x 3 is not the same as
the corresponding coefficient in x 2 The latter coefficient is wrong. This is
because x 1 is correct only through terms of order 6, so that x 2 is only correct
through terms of order 6 2 We can take account of this fact and save much
profitless calculation of higher order terms ifwe replace (31) by
Xo = l, Xn + 0(6"+ 1)
= 1 + BX;-1 (33)
and do not calculate the O(s"+1) terms.
*The symbol~ means "much less than."

Regular Perturbation Theory [Ch. 7
240
Exactly the same type of considerations can be used in the analysis of (28)
and (29). From (28) we find that z 0 = t - !t 2 Substitution into (29) yields
.. 2e(t - !t 2) + e2 (t - !t 2) 2
z1 + I = [I + e(t - !t2)J2 . (34)
With no ultimate loss of accuracy, we write this complicated equation as

z1 + 1 = 2e(t - !t 2) + O(e2 ). (35)
Together with the initial conditions of (29) this yields
z1 = t - !t 2 + e(!t 3 - 1\t 4 ) + O(e2 ). (36)
It is not difficult to show [Exercise 7(a)] that the next approximant z2 satisfies
4
z2 + I = 2e(t - ft 2
) + e2 [-jt3 - !1 - 3(t - ft 2) 2 ] + O(e3). (37)
Solving this equation, we find on comparison with (8) that the second
approximation z2 is correct through terms of order e2 as anticipated [Exercise
7{b)].
An advantage of the successiveapproximation method is that it does not
require identification of a small parameter. As the method is iterative, a
minor mistake in calculation may only delay, not destroy, close approxi-
mation to the correct answer. Moreover, there are times when the form of the
answer cannot readily be guessed, but yet it emerges from several iterations
of the successive approximation method. Finally, this method sometimes
leads to relatively easy proofs of convergence, using induction.
The chief disadvantage of the successiveapproximation method is that gen-
erally many hard-won terms in a new iterate only give the illusion of providing
better accuracy. As we have seen, however, sometimes this disadvantage
can be circumvented by a modification of the procedure.
GENERAL REMARKS ON REGULAR PERTURBATION THEORY
We have in essence been considering determination of a function/ which
depends on a parameter e as well as on an independent variable t, and which
satisfies a differential equation with appropriate initial or boundary condi-
tions. When we set e = 0, we obtained a solvable problem whose solution
was a function/ 0 We can categorize regular perturbation theory more pre-
cisely than we have, but still somewhat roughly, by saying that it consists
of methods for systematically improving on the approximation/ 0 , when that
approximation seems to be uniformly valid in the entire domain of the inde-
pendent variable t. An equation lies in the province of singular perturbation
theory (the subject of Chapter 9) when setting e = 0 yields a problem that
(i) has no solution, or (ii) has a solution that does not provide a uniformly
valid approximation over the entire domain of interest, or (iii) has many
solutions.
We have considered here only the case when f satisfies a single ordinary
Sec. 7.2} Projectile Problem Solved by Perturbation Theory .241
differential equation, but the method of perturbation is substantially the

same when f satisfies an algebraic equation, a difference equation, a partial
differential equation, an integral equation, systems of such equations, etc.
Any of the three methods-series, differentiation, or successive approxima-
tions-can be used. Each yields a sequence of problems to be solved. The
nature of the individual problems and the techniques required to solve them
will be peculiar to the type of equations one works with, but the perturbation
method provides a common framework.
The calculations we have employed are formal in that various limit opera-
tions have been interchanged without justification. For example, upon sub-
stituting an infinite series into a differential equation, we have interchanged
the operations of summation and differentiation. In the differentiation
approach, we exchanged the order of differentiation with respect to the pa-
rameter and differentiation with respect to the independent variable. In
all cases we assumed, without rigorous justification, the existence of solutions
having suitable dependence on a parameter.
The theoretically inclined mathematician might have acquired motivation
for the study of theorems justifying our formal manipulations and motivation
for the invention of further-reaching theorems than are currently available.
Others will be content to proceed formally, knowing that the class of problems
for which formal calculations can be justified is considerably narrower than
the class for which it is imperative that such calculations be performed if
progress is to be made.
A difficulty with perturbation methods is that the complexity of the cal-
culations increases rapidly with the number of terms retained. Often, however,
one or two terms suffices. (We see an example of this in Chapter 8.) In other
cases, a computer can be used to perform the required manipulation. See,
for example, "Viscous Incompressible Flow Between Concentric Rotating
Spheres, Part 1, Basic Flow" by B. R. Munson and D. D. Joseph, J. Fluid
Mech. 49, 289-303 (1971). Here terms through the seventh power of an
appropriate small parameter are retained, which requires calculation of
about 1000 coefficients.
EXERCISES
fl. x,
Find the exact solution for in the projectile problem (I).
2. Use Exercise 1 to show that Xm = !/(1 - fe). Use this to check (13).
3. Verify the following equations.
(a) Equation (20).
(b) Equations (35) and (36).
4. Show that the equations of (4) are compatible with those of (22) and (23).
5. Construct an example for Perturbation Method ill in which it is
advantageous not to make all iterates satisfy the original initial con-
ditions.
6. (a) Verify (32).

(b) Show that the scheme (33) also yields (32).
7, (a) Verify (37).
(b) Solve (37) and thereby demonstrate agreement with (8).
8. Treat the pendulum problem of Section l according to the following
methods.
(a) By parametric differentiation.
(b) By successive approximations.
9. Consider the quadratic equation m2 + em - 4 = 0, 0 < e ~ 1.
(a) Solve by the quadratic formula and expand the results as series
in e. Then, pretending that the problem m2 - 4 = 0 is solvable but
the original problem is not, develop approximate solutions.
(b) Solve by the series method.
(c) Solve by parametric differentiation.
(d) Solve by successive approximations.
10. In the projectile problem, when air resistance is taken into account but
changing gravitational attraction is neglected, the equations are
d 2 x* dx* dx*
m dt* 2 +k dt* = - mg, x*(O) = 0, dt* (0) = v.
For small damping, the change of variables
x* t*
x = v2g-1' t= -,
Vg-1
is still appropriate. (Compare Exercise 6.3.8.) Assume the power series

expansion
x(t, /3)= x 0 (t) + {3xi(t) + , /3= kV.

mg
Proceed as in the corresponding problem treated in the text to show that
tm = l - tP + t/32 + , Xm = t- f{3 + !f32 +
11. Consider the eigenvalue problem
y" + A.(l+ ex)y = 0, 0 < x < n; y(O) = y(n) = 0, 0 < e ~ l.
t(a) By perturbation theory, using the result of Exercise 4.3.12(a),
determine a first correction eA.~n>to the eigenvalues A.~>= n2
which correspond toe= O; n = 1, 2, ....
(b) Use Exercise 4.3.12(b) to find a series expansion for the first
correction to the corresponding eigenfunctions y<n>.
To obtain a
unique answer, impose the normalization
r0
[y<n>]2
dx = 1.
12. Find approximate solutions to

36x3 + (162 + 4e)x2 - 24ex - 9e = 0
that are valid for small Ie I
13. Of interest in a certain calculation are the two roots of
e+ 247 = x-1(x + 1)3
that are near x =!,a root fore= 0. Find the first term or two in series
expansions for these roots.
CHAPTER 8
mustrationof Techniqueson a
PhysiologicalFlow Problem
THIS
chapter is concerned with an analysis of" standing gradient fl.ow."
The fl.owis osmoticin nature, being driven by solute concentration differences.
The problem will be treated on a sufficiently simple level so that our dis-
cussion can be self-contained.
In our analysis we shall put to good use a number of the ideas we have
studied in simpler contexts-dimensional analysis, scaling, and regular
perturbation theory. We shall see how dimensional analysis reduces the
number of relevant parameters from eight to three. Part of the scaling is
rather challenging.Moreover, the perturbation theory is not entirely straight-
forward, for there are two important parameters in the problem whose
relative magnitudes must be ascertained. In summary, while only involving
a pair of first order ordinary differential equations, the model for standing
gradient flow that we shall consider is sophisticated enough to require a
panoply of our methods for its thorough analysis.
8.1 PhysicalFormulationand DimensionalAnalysisof a

Model for "Standing Gradient" OsmoticallyDriven
Flow
In this section we outline the various physiological and biophysical con-
siderations that are relevant to the problem we wish to investigate. We then
use dimensional analysis to draw a few preliminary conclusions about the
phenomenon under consideration.
SOME PHYSIOLOGICAL FACTS
The relevant physiologicalfacts are these.
(a) Solute is secreted by various tissues into the fluid that is adjacent
to them, or bathes" them. For example, several different salts are
secreted by the tissues of the kidney into the fluid that flows in
kidney tubules.
(b) In the cases of interest, the solute concentration in the bathing fluid
is approximately equal to, or perhaps somewhat greater than, the
average concentration of solute in the secretingtissue.
(c) One possible mechanism for solute secretion is diffusion, but this
cannot be the principal mechanism here. The reason is that solute
244
Sec. 8.1] Physical Formulation and Dimensional Analysis 245
is entering a region with equal or higher solute concentration than

the region it is leaving.
(d) Another possible mechanism is active transport by the membrane
that forms the boundary between the tissues and the bathing fluid.
Here chemical "pumps" can expend energy to move solute across a
membrane in a given direction, more or less independently of the
concentrations of solute in the fluids bathing the membrane. In
particular, an actively transporting membrane can pump solute
from a region of low concentration to a region of high concentration.
It can be shown, however, that active transport cannot be directly involved
in the phenomenon of interest, i.e., the solute cannot be pumped directly
into the bathing fluid. The rate at which the solute is entering the bathing
fluid, per unit area of membrane, has been measured. Knowing this, one
can compute the dependence of solute concentration on distance normal
to the membrane. (Simple diffusion theory is appropriate here.) But measure-
ments show different results from those predicted in this calculation.*
The physiologist J. Diamond and his associates have put forth a hypothesis
that reconciles the various facts. t As of this writing, the hypothesis can be
said to be well regarded but not fully accepted. We shall now explore this
hypothesis in some detail.
The physiologists proceed from the observation that long narrow channels
consistently are a structural feature of tissues that transport fluid. Examples
are the channels between adjacent cells in the wall of the gall bladder, in-
foldings in kidney tubules, the bile canaliculi (small canals) of the liver, and
canaliculi in the cells lining the walls of the stomach. It is conjectured that
solute is actively transported into the "far" end of the channels, deep in the
secreting tissue (Figure 8.1). The relatively high concentration thus established
in the far end would then be progressively diluted as the fluid passes toward
the opening of the channel, for water will be drawn into the channel by
osmosis due to this concentration. A steady or standing gradient of solute
concentration will soon be set up. The concentration will be high at the closed
"inward " end of the channel and will decrease to a relatively low value at
the other " secreting" end.
We emphasize that the mechanism we are treating is postulated for a
variety of secreting and absorbing tissues. For definiteness we shall confine
our discussion to the secreting case. Absorbing tissues can be handled
merely by running the whole mechanism backward, although a few extra
complications arise (Diamond and Bossert, 1968, op. cit.).
* See J. Tormey and J. Diamond, "The Ultrastructural Route of Fluid Transport in Rabbit
Gall Bladder," J. Gen. Physiol. 50, 2032 (1967). Also see Exercise 2.6 and the reference cited there.
t In addition to the reference just cited, see J. Diamond and W. Bossert, "Standing-gradient
Osmotic Flow, a Mechanism for Coupling of Water and Solute Transport in Epithelia" J. Gen.
Physio/. 50, 2061-2083 (1967), and, by the same authors, "Functional Consequences of Ultra-
structural Geometry in Backwards' Fluid-transporting Epithelia," J. Cell. Biol. 37, 694-702
(1968).
Illustration of Techniques on a Physiological Flow Problem [Ch. 8
Secreting
tissue
. :. .:: .
Bathing"..
: :,fluid
Water
FIGURE 8.1. Long, thin channels, as are consistently found in secreting

tissues. A higher magnification view of a channel, showing the conjectured
transport mechanism, appears in the insert. The solute is actively pumped
into the channel near the closed end; water is drawn in by osmosis so that
the solution that emerges at the open end of the channel can be considerably
diluted.
It would be out of place to discuss in depth the direct and indirect physio-
logical evidence that supports the standing gradient hypothesis. For this
we must refer the interested reader to the references cited, and to the papers
quoted there. We shall confine ourselves to a study of the illumination cast
on this hypothesis by the mathematical model that was devised and numeri-
cally analyzed by Diamond and Bossert in 1967, and treated analytically
by L. Segel in 1970. The Diamond-Bossert paper has been cited above.
We shall refer to it frequently below, often abbreviated by the initials DB.
The paper by Segel* will be referred to as S; much of our discussion follows
this paper very closely.
"Standing-gradient Flows Driven by Active Solute Transport,'' J. Theoret. Biol. 29, 233-50
(1970). Material is reproduced with permission of the publisher of the journal, Academic Press
Inc. (London). In particular, Figures 8.2, 8.4, and 8.5 have been redrawn from figures in the paper
just cited.
OSMOSIS AND THE OSMOL

To make sure that readers feel secure about the biophysical prerequisites
necessary to an understanding of the subsequent discussion, let us briefly
state some important facts about osmosis.
The boundary of the tissue under consideration can be regarded as a semi-
permeable membrane, through which water but not salt can flow. Thus
solute (e.g., salt) cannot flow through the membrane but solvent (e.g.,
water) can. If the number of solute particles per unit volume is different
on the two sides of such a membrane, then water flows across the membrane
to equalize the solute concentration. The water thus flows from a region of
relatively high water concentration to a region of relatively low concentration,
which is not surprising.
The rate of osmotic flow across a unit area of a semipermeable membrane
turns out to depend primarily on the ratio of the number of solute particles
per unit volume on the two sides of the membrane. In discussions of phe-
nomena involving osmosis, it is therefore natural to use as a measure of
concentration the number of particles present in a given volume of fluid.
A unit of such a measure is the osmol, * which we now define.
Recall that a gram molecular weight or mole of a substance is an amount
of that substance whose mass in grams is numerically equal to its molecular
mass. A mole contains Avogadro's number 6.02 X 1023 molecules.
Certain molecules break apart or dissociate when put into solution; others
do not. If a mole of a nondissociating substance like glucose is put into
water, we say that the water contains l osmol of solute. A milliosmol is one-
thousandth of an osmol.
A salt molecule in water dissociates into a sodium ion and a chloride ion.
If a mole of salt is placed in an effectivelyinfinite amount of water, the water
is said to contain 2 osmols of solute. (" Effectively infinite" means "large
enough to permit essentially complete dissociation.")
Because of" interference" effects, doubling the number of solute particles
in a.finite volume does not double the osmotic effect. Similarly, to give an ex-
ample, adding 1 mole of salt to a solution having the concentration of
human body fluid is found to increase the effective concentration of salt in
that fluid by 1.85 osmols.t Note that the correction is only a few per cent.
We need not concern ourselves with such a small factor; therefore we shall
Introduction of this term is not strictly necessary, but it points up the lesson that an applied
mathematician must learn the terminology of a new field when he starts to work in it, so that he
can communicate easily with the experts. Once one has decided to make the effort, it is surprisingly
easy to master a new technical vocabulary and thereby to have large areas of a new field revealed
as aspects of science with which one is more or less familiar. Some further effort, also illustrated
to some extent by the following discussion, shows other areas of the new field to be understandable
with the aid of new but relatively simple scientific principles.
t See Chapter II of R. F. Pitts, Physiology of the Kidney and Body Fluids (Chicago: Yearbook
Medical, 1963).
assume that the rate of solvent flow across a unit area of semi-permeable
membrane is proportional to the difference of solute concentrations (mass
per unit volume) across the membrane.
FACTORS THAT AFFECT STANDING GRADIENT FLOW
As a preliminary to setting up the mathematical model, let us list the
physical factors that are presumed to govern the phenomenon. We must
emphasize that the model building has already started, for a number of factors
will be entirely ignored, their influence being considered secondary. Examples
of such factors are the changing cross-sectional shape of the channel, radial
effects, and the permeability of channel wall to solute. Similarly, factors that
are considered will be taken into account in a way which is more or less
approximate. For example, a linear relation will be assumed to connect
osmotic flow of water across the channel wall and solute concentration
difference across the wall, so that the permeability of the wall to water can
be characterized by a single, constant, parameter. As we have mentioned,
such a linear relation is probably adequate, but it should be kept in mind
that a possible source of error has been introduced.
Here, then, are the assumptionswe shall make.
(i) The channel will be i<;lealizedas a right cylinder with cross-sectional
area a, circumference c, and length L.
(ii) Suppose there is a solute concentration difference across the wall
of the channel. Then the rate of flow of water across a unit area of
channel wall will be assumed to be directly proportional to this
concentration difference. The constant of proportionality is the
permeability and is denoted by the letter P.
(iii) The concentration of solute in the "bathing fluid" outside the chan-
nel is assumed to have the constant value C0 Not only will this be
assumed to be true along the sides of the channel, but contrary
tendencies at the open end will be neglected so that there, too, the
concentration will be required to have the value C0 .
(iv) In the channel, the solute will be carried along by a varying fluid
velocity which depends only on the distance along the channel.
In addition to this convectionof solute, solute diffusionwill take place.
The diffusion constant will be denoted by D.
(v) The active solute transport will be assumed to take place only in a
portion of the tube wall extending a distance 5 from the closed end.
The rate of solute transport per unit area of channel wall will be
assumed to have the constant value N 0 The assumptions and no-
menclature are summarized in Figure 8.2.
The quantity that is of primary physiological interest is the rate at which
solute emerges from the open end of the channel. It is useful to think in
terms of the emergent osmolarity Os*, which is defined as that concentration
of solute which, when multiplied by the volume of solvent that leaves the
N 0 =Rate
O~

Circumference =c
of active C,, =.'\.mbient solute concentration

transport
FIGURE 8.2. The simplified standing gradient flow system.
tube per unit time, gives the rate at which solute leaves the end of the tube
per unit time.
The definition of Os* will be clearer if we introduce some notation. Let
us select an x* axis that is parallel to the generators of the tube so that the
closed end of the tube corresponds to x* = 0 and the open end to x* = L.
Let v*(x*) denote the speed of the fluid at position x*, in the direction of
increasing x*. Then the volume of fluid that emerges from the open end of
the tube per unit time is av*(L), since a is the cross-section area. We denote
the flux density of solute at x* by F*(x*). t By definition, the flux density
F*(x~) is the rate at which solute crosses a unit area of the plane x = x~
in the direction of increasing x.Thus solute emerges from the open end of the
tube at a rate aF*(L). By definition, the emergent osmolarity Os* satisfies
av*(L)Os* = aF*(L), so
0 *- F*(L) (])
s - v*(L).
We now turn to a discussion of what one can learn from dimensional
analysis at this point. The reader can pass at once to Section 8.2 if he wishes,
for results obtained in the remainder of the present section will not be used
in what follows.
DIMENSIONAL ANALYSIS OF A FUNCTIONAL RELATIONSHIP
If assumptions (i)-(v) above are correct, then a standing gradient flow is
completely determined by the eight parameters fJ,L, c, D, C0 , P, N 0 , and a.
Once these parameters are given, the idealized problem is completely specified,
assuming that we have not left out anything essential. This is a dangerous
assumption, for unerring characterization of a physical problem requires
t We are making an exception to our usual practice of employing J to denote flux density.
250 Illustration of Techniques on a Physiological Flow Problem [Ch. 8
well-developed intuition. Granting this assumption, however, let us see what

the methods of dimensional analysis can tell us.
If the above-mentioned eight parameters completely characterize the stand-
ing gradient flow system, then, in particular, they must uniquely determine
the quantity Os*. Symbolically, we write
Os*
Co = f(l,, L, c, D, C0 , P, N 0 , a). (2)
Note that we have written (2) so that its left side is dimensionless, being the
quotient of two concentrations.
To proceed with a dimensional analysis, we must know the dimensions of
each of the various parameters. The required material is given in Table 8.1.
TABLE 8. 1. Parameters and their dimensions

(!' = length, vii= mass, !I = time).
Parameter Dimension
8,L, c ft'
D !'2 !J-1
Co, Os* .,I( ft'-3
p ,,11-1fi''!J-1
No J(fC-29"-
a !'2
The most difficult entry, that for the permeability P, is obtained from the
following considerations. To measure P, one must set up a known solute
concentration difference across a membrane and measure the volume of
solvent that crosses a given area in a given time. This volume is proportional
to the concentration difference via the permeability constant, so that
.li _ solvent volume/(area x time)
permea b 1 ty - d"fli . . ,
1 erence m so1ute mass per umt volume
giving the dimensions .,11- 1!' 4!/- 1 of Table 8.1.

Since Os*/C 0 is dimensionless, the value of the function on the right
side of (2) must also be independent of units, so that/must depend on dimen-
sionless combinations of its arguments. Since .,It, ft', and !I appear as powers,
only powers of the arguments will permit their cancellation. Thus f must
involve its variables in combinations of the form
l,a1r2ca3D""'C't/P""6Ng1d', (3a)
where the a; are constants. This combination has dimensions
fi'""lfi'""2fi'a3(fi'2!J-lt(.,11 fi'-3)a'(.,lt-1 fi'4ff-lt'(..lt fi'-2 !J-lt7 fi'2aa (3b)
Sec. 8.1) Physical Formulation and Dimensional Analysis
251
or
and will be dimensionless if and only if

a1 + a 2 + a 3 + 2a4 - 3a5 + 4a6 - 2a 7 + 2a8 = 0,
a5 - a 6 + a 1 = 0, (4a)
-a 4 -a 6 - a1 =0.
The solution to the above equations can be written (Exercise I)
a6 = oc, a 7 = P, a 4 = -(oc + P), a 5 = ex- p,
(4b)
a2 = y, a 3 = 0, a 8 = e, a 1 = oc+ p - y - (} - 26,
where oc,P,y, 0, and s are arbitrary. With this, (3a) becomes
[J+P-1-9-2.L 1c9D--Pc~-PpNga
or
(~~:r
ePior (~r(i)
9
(: 2)
(4c)
Thus, as discussed in Section 6.2, we see that the only dimensionless groupings
of the eight parameters are
fJPC0 a
7t1=v, 1ts = [J2' (5)
or some combination thereof. Consequently,

O*
do = f(1t1, 7t2, 7t3, 7t4, 1ts)
From the assumption that the emergent osmolarity depends on the eight
parameters of Table 8.1, we have deduced that the ratio of emergent to
ambient osmolarity depends on the five dimensionless groups of (5). With
further thought we can do better. It is possible to convince oneself that the
permeability P must always occur in the combination cP, for with one-
dimensional flow (or even just axisymmetric flow) only the permeability
per unit length matters. Similarly, N O and D must occur only in the combina-
tions cN O and aD. Moreover, the area a and the circumference c do not enter
the problem except in the above combinations. (This last assertion is not
easy to justify at the present stage, but let us accept it for the moment.)
With the definitions
P1 = cP, N 1 = cN 0 , D 1 = aD,
we have, for some function g,
O*
_s = g(fJ,L, D1, Co' P1, N1). (6)
Co
Illustration of Techniqueson a PhysiologicalFlow Problem [Ch. 8
As above, we can therefore assert that the dimensionless emergent osmo-

larity must depend on dimensionless combinations of the form
(7)
This term has the dimensions
which (Exercise 2) can be shown to be dimensionless if and only if, for arbi-
trary a, p, and y,
b1 = 2oc+ 2P - y, b2 = y, b3 = - ex- P, b4 = a - P, b5 = a, b6 = p.
(8)
With this, (7) becomes
2 0 1)"(t5
t5 C P
2N 1)P(~)
7.
( D1 D1Co t5
Returning to our original notation, we deduce that for some function/ 1,
C:=/1(1t6, 1e,,
O*
1ta), (9)
where
cozNo L
1t, = aDC 0 ' 1Cg = -g (10)
Instead of the eight dimensional parameters listed on the right side of (2)
we need only consider the three combinations of (9) !
Other combinations than those in (9) can be used. For example, a function
of 1t6 , 1t7 , and rc8 is also a (different) function of rc7 /n 6 , l/n 7 , and 1t8
Making the abbreviations
1 aC0 D L
'7=-=--2, A.= 1Cg = -g, (11)
1C7 cNot5
we can therefore write, for some function / 2,
O*
do = /:i{v,'1, ..1.). (12)
The same basic result, that Os*/C 0 depends only on three dimensionless
parameters, will be deduced below from the governing differential equations
of the problem .. There we shall see some theoretical consequences that can
be drawn from this result. Now let us examine some possible consequences
for experimenters.
Sec. 8.2] A Mathematical Model and Its Dimensional Analysis
253
POSSIBILITY OF A SCALE MODEL FOR ST ANDING GRADIENT

FLOW
Our dimensional analysis reveals the possibility of making scale models
for standing gradient flows. In contrast with airplane and ship models, here
the model would be much larger than the true system, for the tubes in which
standing gradient flow would actually take place are so tiny that it is only
with great difficulty and ingenuity that measurements can be made.
Suppose that it was deemed worthwhile to construct a model secreting
tube that was 100 times as long as the genuine tube. U:sing primes to denote
the parameters for the model system, this means that L' = 100. If we also
make the region of solute pumping 100 times as long (c5'= 1006), then
,r 8 =L' W = L/6 = ,r 8 , so that one of the three parameters governing the
original system remains the same in the model system. If we make the area
to circumference ratio 104 times as large (a'/c' = l04a/c) and retain the
same values of the remaining dimensional parameters (C0= C0 , P' = P,
Nf, = N 0 , D' = D), then the other two governing dimensionless parameters
in (9) retain their values (,r 6 = ,r 6, ,r 7 = ,r; ). If (6) is valid, then (9) is valid
and the ratio of emergent to ambient osmolarity will be the same in the two
systems (Os*'/Cf, = Os*/C 0 ). Measurements in the model system will yield
information about behavior of the original system, but only if we have not
omitted any important parameters. When dealing with living organisms,
it is very difficult to be sure that all essential factors are included in a model.
Further progress requires formulation of a definite mathematical problem.
To this we now tum.
EXERCISES
1. Use the theory of linear equations to verify that the solution to (4a)
contains five arbitrary parameters. (Students often feel that systems of
linear equations containing fewer equations than unknowns are "excep-
tional," but dimensional analysis is an area where such equations occur
all the time.)
2. Verify that (7) is dimensionless if and only if (8) holds.
8.2 A Mathematical Model and Its Dimensional Analysis

In this section we formulate a mathematical model for standing gradient
flow and then subject it to a dimensional analysis. The required equations
could be obtained by specializing general results obtained elsewhere, but
we shall proceed essentially from first principles, if only in order to make the
present discussion nearly self-contained. There are two governing equations,
one for conservation of fluid mass and one for conservation of solute mass.
CONSERVATION OF FLUID MASS

We can assume that water has a uniform density so that it will suffice
to require conservation of fluid volume. Consider a test cylinder extending
from x* to x* + Ax*. Here Ax* stands for a positive number, not necessarily
a small one. Let v*(x*) denote the fluid velocity at x*, taken positive in the
direction x* increasing. Then the rate at which fluid enters the test cylinder
through its ends is
a[v*(x*) - v*(x* + Ax*)], (I)
since a is the cross-section area of the cylinder. By the assumption of propor-

tionality between rate of water flow through the lateral walls and concentra-
tion difference, the rate at which water enters the test cylinder through its
lateral walls is
+.u:
Pc [ ,,. [C*(s) - C 0 ] ds. (2)
[As a check on signs, if C* > C 0 , so solute concentration is higher inside

the tube, then the term (2) is positive, corresponding correctly to the fact
that water will enter the tube to dilute the relatively high solute concentration
there.]
Since no water is created inside the test cylinder, the terms in (1) and (2)
must add to zero. Thus, using the integral mean value theorem on the second
term,
a[v*(x*) - v*(x* +Ax*)]+ Pc[C*(x* + 8 Ax*) - C0 ) tu*= 0,
where O < (}< 1. Dividing by Ax* and taking the limit as Ax* -+ 0, we obtain
the first equation:
(3)
CONSERVATION OF SOLUTE MASS

Let N*(x*) denote the rate at which solute is actively transported inward
across a unit area of lateral boundary located at x*.t As above, let F*(x*)
be the amount of solute per unit time that crosses a unit area which is located
at x*, and which is oriented normally to the generators of the cylinder.
Recall that F* is positive if solute flows in the direction x* increasing. Setting
to zero the net rate at which solute flows into the test cylinder, we obtain
+.u:
c[ N*(s) ds + a[F*(x*) - F*(x* +Ax*)]= 0. (4)
,,.
t For the moment we consider the general case. Later we return to the special assumption
that N is constant for O::;;x ::;;3 and zero elsewhere.
255
As in the previous equation, by using the integral mean value theorem,
dividing by Ax*, and then letting Ax* -+ 0, we obtain
dF*
cN*(x*)- - a dx* = 0. (5)
The solute flux per unit area F* is the sum of two parts, (a) flux due to
convection Fc!nvand (b) flux due to diffusion Fd~rr.First, let us ascertain the
convective contribution, that due to the bodily transport of the solute by fluid
motion. We observe that if the fluid at x* has speed v*(x*) then, per unit time,
the volume of fluid which crosses a unit area at x* (in the direction of increas-
ing x*) is v*(x*). (Here the area is assumed to be oriented normally to the
generators of the cylinder.) The mass of solute that crosses this unit area in
a unit time is therefore C*(x*)v*(x*), since C* is the mass of solute per unit
fluid volume. Hence
(6)
The diffusive flux is given by
dC*) (7)
F:m = -D ( dx* '
where D, a constant, is the solute diffusivity. Equation (7) is sometimes

called Fick's law; it states the solute flux is proportional to the spatial deri-
vative in solute concentration, just as the Newton-Fourier Law of cooling
says that heat flux (or temperature flux) is proportional to the spatial
derivate of temperature. With the negative sign in (7) the constant D will be
positive, since solute flows away from regions ofrelatively high concentration.
A justification of assumption (7) can be provided by an argument identical
with that used to justify Equation (1.3.7). Or the analogy with heat flux can
be accepted, in which case (7) is essentially the same as Equation (4.1.4).
At any rate, combining (6) and (7) we have
F* = v*C* - D ( dC*)
-- . (8)
dx*
BOUNDARY CONDITIONS
The differential equations necessary for our problem are supplied by (3),
(5), and (8). It remains to state boundary conditions. These follow.
No fluid is to flow across the closed end of the channel, so that
at x* =0, v* =0. (9)
No solute is to flow across the closed end of the channel so that F* = 0
at x* = 0. Using (8) and (9), we can write this condition as
dC*
at x* = 0, dx* = 0. (10)
As stated earlier, we assume that the solute concentration at the open end
of the channel equals C0 , the assumed concentration of bathing fluid. Thus
at x* =L, C* = C0 (11)
To determine as simply as possible the effect of a solute "pumping"
confined to the closed end of the tube, we take N* to have the constant value
N O near the closed end of the channel and to be zero elsewhere:
N* = N 0 for O =::;;x* =::;;<5; N* = 0 for <5< x* L.
=::;; (12)
Since N* has been assumed to be discontinuous at x* = b, our problem
has now been effectively split up into two problems, one for O < x* < <5,
the other for <5< x* < L. We must assume further conditions to "link"
these problems together appropriately. We shall derive the conditions from
naive physical considerations. Other derivations are requested in the exercises.
No solute is being created at x* = {);therefore the solute that enters from
the left must equal the solute leaving from the right. In symbols,
F*(a-) = F*(b+), where F*(a) = lim F*(b e).t (13)
10
Similarly, no fluid is being created at x* = b, so the fluid volume entering
from the left must equal the fluid volume leaving from the right. Thus
v*(b-) = v*(a+). (14)
We adopt as an additional physical hypothesis, because its denial is contrary
to our intuition concerning molecular behavior, that the solute concentration
itself is a continuous function of x* at x* = b. If C* and F* are continuous
at b, then dC*/dx* must be continuous at x* = <J,or
(15)
With the form of active solute transport assumed in (12), the solute mass
conservation equation (5) can easily be integrated once. We find
aF* = cN0 x* + Q 1, 0 S x* < <5; aF* = Q2 , {) < x* SL. (16)
F*(O)= 0 implies that Q 1 = 0. The continuity of F* at x* = /j implies that
Q2 = cN 0 (). Thus
aF*=cN 0 x*, 0Sx*<c5; aF*=cN 0 '5, a<x*=s;;L. (17)
Since the lateral walls have been assumed impermeable to solute, (17)
could almost have been written down at once. It merely states that solute
flux increases linearly with x* in the region of the channel where there is
uniform active transport per unit length, and that the flux remains constant
in the region of the channel where no further transport takes place.
t lim/(e)
aio
= ,lim/(e).
...o
1>0
Sec. 8.2] A Mathematical Model and Its Dimensional Analysis 257

Of central interest is the emergent osmolarity Os*, which was defined
in Equation (1.1). Summarizing our previous discussion, we see that Os* is
determined by the following mathematical problem:
for O < x* < '5 and '5 < x* < L.
v
* * _ D dC* _
c d *-
{a-cN x* 1
0 for O < x* < '5,
x a -1 .lt.T
CH 0 us: for '5 < x* < L.
(18)
v*(O)= 0, C*(L) = C0 , v* and C* continuous at x* = '5.
Determine
* F*(L)
Os=--=--
cN0 '5
v*(L) av*(L)
[In obtaining the last line, we have used (17).)

In the above, we have eliminated the condition that dC*/dx* must vanish
at x* = 0 and the condition that dC*/dx* must be continuous at x* = '5.
These are implied by the conditions already required. This conforms to our
expectation that having integrated an equation once, fewer boundary con-
ditions are necessary.
As a check on the above equations we can test for dimensionalhomogeneity:
in a given equation each term must have the same dimension; otherwise,
one could render an equation false by changing one of the fundamental
units of measurement. (For a formal discussion of this matter, see Exercise
6.2.12.) To give an example, in the first equation the dimensions of the left
side are those of velocity/length or, ff- 1. From Table 8.1 the dimensions of
the right side are
(.,((-1 fi'4ff-l)(!t')(!i'2)-1(.,(( fi'-3) = ff-1
so that the equation is indeed dimensionally homogeneous.

The boxed equations (18) comprise the mathematical formulation of the
standing gradient problem. We are faced with a system of two first order
nonlinear differential equations for the two unknown functions v*(x*) and
C*(x*). As expected, there are two boundary conditions. There is also a
pair of continuity conditions, to " stitch together " the variables at the point
x* = ~.where the specification of the problem changes.
INTRODUCTION OF DIMENSIONLESS VARIABLES
The next step is to switch to dimensionless variables. In doing so, we shall
refer distance x* to the length '5 and concentration C* to the referenee con-
centration C0 There is no obvious choice for a velocity scale. We can find
one by trial and error, but the following procedure is a little more rational.
Illustration of Techniques on a Physiological Flow Problem [ Ch. 8
From (6) the convective flux density equals the product of velocity and
concentration. Thus velocity must have the dimensions of flux density
divided by concentration. From (17) the combination of parameters cN0 l,/a
has the dimensions of flux density, for this is the total solute flux density
at the open end of the tube. A typical concentration is C0 Thus we take
cN0 lJ/(aC0 ) as our reference velocity. To check, we consult Table 8.1,
which shows that
. . cN or, (!t')( JI !t'- 2ff- i )!t' !t'
d1mens1onsof aCo = !t' 2 (Jt!L'- 3) = 9".
This is, of course, the correct dimension for a velocity.

We thus introduce the dimensionless variables x, v, and C, defined by
x* C* v*
x=~ C=-, V= (19)
Co cN 0 lJ/(aC0 )
We obtain [Exercise S(a)]
C-1 =vv' for O < x < l, (20a)
for O < x < l,
vc- ,,c= {;
for 1 < x < l,
(20b)
v(O)= 0, C(l) = l, (20c, d)

C and v continuous at x = 1. (20e)
The dimensionless emergent osmolarity Os is given by
Os* 1
Os= C0 = v(l)' (21)
Here
No L
v=-2 l=-, (22)
PCo ()
just as in (1.11).
In the original dimensionalproblem (18), the eightparametersr,,L, c, D, C0 ,
P, N 0 , a appear so that any feature of the" answer," e.g., the value of Os*,
can only depend on these eight parameters. In the dimensionlessproblem (20),
only the threeparameters v, '7, and A appear. Any feature of the answer to
the dimensionless problem, e.g., the value of Os= Os*/C0 = 1/v(l), can de-
pend only on these three parameters.
Suppose that we had used L as a reference length rather than r,,What then?
Withy= x*/L replacing x = x*/{J,we would have obtained [Exercise S(b)]
C- 1= V1 ii', for O < y < 1 (23a)

259
-- "''= {ly
vC-11 1 v
1
for O<y<l-
for i- <y < l. 1
1
,
(23b)
v(O)=0, C(l) = l. (23c, d)

- 1
C and v continuous at y = i- 1
, Os=- (23e)
v(I)
Here a bar indicates that the dependent variables are a function of y, and
aC 0 D No"
111= cN {)L V1 =--2-
0 PC 0 L
Again, only three parameters are involved, this time ''It, v1 , and .k As expected,
the new parameters are functions of the old, for 111 = i- 1,,, v1 = i- 1 v.
This iJJustrates the fact that although the form of the equations and boundary
conditions is changed when different nondimensionalizations are used (the
dimensionless parameters are different and are located in different places),
this inessential change does not alter the number of dimensionless parameters
on which the answer depends.
COMP ARIS ON OF PHYSICAL AND MATHEMATICAL

APPROACHES TO DIMENSIONAL ANALYSIS
The same result, that only three dimensionless parameters govern the
standing gradient problem, was obtained from a rather general physical
formulation of the problem in Section 8.1, and a precise mathematical
formulation here. (The same sort of result was obtained in Section 6.2, for the
relatively trivial projectile problem.) Let us compare the two approaches.
It would appear that formulating the equations which govern a phenomenon
has great value, even if the equations a-ppear entirely intractable. Thus,
murky depths of physical intuition must seemingly be plumbed if the para-
metric dependence of Os* given in (1.2) is to be "intuited." Deriving the
governing equations required a comparatively less difficult selection and
simplification of known physical laws. And once the equations are derived,
the relevant dimensional parameters appear automatically. Again, rather
considerable physical intuition is required to be sure of our earlier statement
[leading to (1.6)] that the parameters a, c, N 0 , P, and D enter our stand-
ing gradient flow problem only in the combinations cP, cN 0 , and aD.
From the governing equations (18), one sees at once that N 0 , P, and D
are always coupled with a and c as in the just-mentioned combinations.
The fact that a and c appear in no other combinations is not obvious but
becomes so if one replaces the speed v* by the volume flux w*, where w* = av*
(Exercise 4). Thus introduction of dimensionless variables into a set of equations
isolates the dimensionless parameters in a way that is computationally easy
and does not require the somewhat deep justification of the pi theorem.
26o Illustration of Techniques on a Physiological Flow Problem [Ch. 8
Knowing dimensionless parameters on which the solution to a problem

must depend, one can often obtain without further effort some significant
aspects of the phenomenon under investigation. For example, we see from
the parameter list in (22) that the shape of the cross section affects standing
gradient flow only through the parameter a/c. It is natural to define an effec-
tive radius r by a/c = r/2. If the cross section is circular, then the effective
radius equals the actual radius. Since r and D occur only in the parameter 'I,
we can be certain that halving the effective radius (say by squeezing a
channel of given circumference) has precisely the same effect as halving the
diffusivity D. To give another example, the flow is unaffected if N 0 , P, and r
are all doubled.
EXERCISES
1. (a) Deduce (13) by integrating (5) from x* - e to x* + e, using (12),
and then letting e approach zero.
(b) Using the same process, deduce (14) from (3). [A pointed out, e.g.,
by Friedman (1956, p. 176), for this formal process for finding
jump conditions in the presence of discontinuities to be valid, it is
necessary that the differential equations be written in a form that
holds for an inhomogeneous medium, where all " parameters., are
functions of the independent variables. For example, the heat
equation must be written in the form (4.1.13) rather than (4.1.12).
Another approach to the derivation of jump conditions, is given in
Exercise 2.]
2. (a) Deduce from (4) that if N* is bounded, then F* is continuous.
(b) Similarly, deduce from the fact that (I) and (2) add to zero that v*
is continuous if C* is bounded. (This use of the fundamental integral
equations of conservation is the best way to derive jump conditions.
For more complicated examples, see Section 12.3 and Exercises
14.4.6-9.) N.B. We have assumed that integrals exist in (4)
and (2).
3. Check all the equations of (18) for dimensional homogeneity.
4. Show that if w* = av* is introduced into (18), then a and c appear only
in the combinations cP, cN0 , and aD, thereby verifying the assumptions
leading up to (1.6). Can you see why w*, not v*, is the appropriate depen-
dent variable?
S. (a) Derive (20), the dimensionless version of (18).
(b) Use y = x*/L, instead of x = x*/{J,to derive (23).
6. The purpose of this exercise is to guide the reader through certain cal-
culations that led to the realization that a simple view of solute transport
in membranes had to be abandoned. As pointed out at the beginning of
Section 8.1, the standing gradient hypothesis is an attempt to remedy
the situation.
For definiteness, let us focus our attention on frog skin. It is known that
Sec. 8.3] Obtaining the Final Scaled Dimensionless Form 261
an isolated piece of skin, immersed in a salt solution of concentration

Cb, can transport salt and water from outside to inside. Consider the
steady state situation. Suppose that there is a uniform salt flux Fs through
the skin, leading to a salt concentration C0 on the outside of the skin
and a concentration Ci on the inside. Fluid layers are assumed to exist
on the outside and inside of the skin, of thicknesses c50 and c5i,respectively.
Let C(x) denote the salt concentration in these layers, where x is a
coordinate normal to the (parallel) bounding planes of the skin.
t(a) It is asserted that C satisfies the following differential equation
dC
D dx -FwC= -Fs. (24)
Here the constant Fwis the water flux that arises due to the osmotic
effect of salt gradients, and Fs is the salt flux. What is the constant
D? What has been assumed in (24)? Why is it legitimate to assume
that Fs and Fw are constants?
(b) Solve the differential equation in the inner layer; impose boundary
conditions; and obtain the result
(c) Correspondingly obtain the following result for the outer layer:
C0 = (Cb - F) exp ( Fwc5o)

~ + F.
(d) Justify the assumption

Fw = L(C; - Co).
Use this assumption to obtain
F5 L(c5o+ c5;)
Fw= (25)
D + CbL(c5o+ c5;)'
For the frog skin, the various parameters on the right side of (25) have
been measured. The resulting prediction for Fw is one to two orders of
magnitude below measured values. See "Unstirred Layers in Frog Skin"
by J. Dainty and C.R. House, J. Physiol. 182, 66-78 (1966).
8.3 Obtainingthe FinalScaledDimensionlessFormof the

MathematicalModel
The next step in our analysis is to choose scaled dimensionless variables so
that the various dimensionless parameters which appear \\'ill correctly give
the relative orders of magnitudes of the various terms. This we shall do.
We then show that there exists a small parameter v in the problem. It appears
that straightforward application of the regular perturbation method can be
made, but difficulty arises when this is attempted. To avoid the difficulty,
the relation between a second dimensionless parameter, 17,and v must be estab-
lished. When this is done (in the second half of the section), at last the problem
is in final form.
SCALING
As was mentioned in the discussion of scaling in Section 6.3, we are faced

with estimating the size of various terms that will only be known when
the problem is solved. But we are not operating in a vacuum-we have
computer calculations for some typical values of the various parameters
(Figure 8.3).
Determining a scale for the concentration C*(x*) is easy. The obvious
guess that C*(x*) is always about the magnitude of the concentration C0
of bathing fluid is reinforced by the results displayed in Figure 8.3. We
0.80
0.70
0
..
...
!c 0.60
c
.9
'E
u 0.50
8
0.40
- ...., ....4
_____
,, ......... _
10 20 30 40 50 60 70 80 90 100
Distance from closed end of channel ()
FIG URE 8.3. Examples of concentration variation in a standing gradient

flow system, from Figure 3 of DB (reproduced with permission). In the
numerical calculations, whose results are given here, the parameters had
thefollowingvalues: L = 100 = 10- 2 cm, 8 = 10, r =0.05, Co =0.3
m0sm/cm 3 , D = 10-s cm 2 /sec, No= 10-s (curve 1), 5 x 10- 6 (curve 3),
or 10- 6 (curves 2 and 4) mOsm/cm 2 sec, P = 10- 5 (curves l, 3, and 4)
or 10- 6 (curve 2) cm 4 /sec mOsm. The calculated values of Os* for curves
1-4 were 0.342, 0.803, 0.318, and0.304.
Sec. 8.3] Obtaining the Final Scaled Dimensionless Form
therefore take C0 as the concentration scale and introduce a dimensionless

concentration C by
C*
C=- (1)
Co
Direct estimate of the magnitude of v* seems difficult. After some thought,
however, an indirect approach suggests itself, based on the fact that it is
concentration differences that drive the flow. We first note that the maximum
velocity (which the scaling procedure requires us to estimate) is expected to
be at the open end of the tube. This is because the cumulative amount of
water drawn in by osmosis increases with x*. We therefore focus our attention
onx* =L.
From (2.8), the flux density at x* = L satisfies
F*(L) = v*(L)C*(L) - D ~~: (L). (2)
The two terms on the right side reflect the effects of convection and diffusion,
respectively, in discharging solute from the tube. We expect that dC*/dx*
will be negative at x* = L so that we can write
* F*(L)
F*(L) 2:::v*(L)C*(L) or v (L) ~ C*(L). (3)
Using the boundary condition C*(L) = C0 , and the expression in (2.17)

for F*(L), we can write the last inequality as
v*(L) <--
cNo<>
- aC
(4)
0
We anticipate that diffusion effects at their most intense will still not
greatly outweigh convection in discharging solute.* The two sides of the
inequalities in (3) should therefore not differ by an enormous amount.
Consequently, the right side of (4) can serve as an estimate of the largest
value of v*, and this is precisely what a velocity scale should do.
In determining a scale for the independent variable x*, we should focus
our attention on the region where the dependent variables C* and v* are
changing most rapidly. The scale should be a constant that estimates the
distance over which a significant part of that change takes place. From
Figure 8.3 it appears that the dependent variables change appreciably over
distances considerably smaller than the channel length, and that <> is the
appropriate length scale.
Using (2.18), we see that F*(L)/[Co v*(L)]= Os*/C0 In the DB calculations this ratio of
total to convective flux ranged from around unity in three cases to about t in one case. This bears
out our contention that diffusive flux will never greatly exceed convective flux.
To summarize, concentration, velocity, and length scales are estimated to

be C0 , cN0 lJ/aC0 , and lJ,respectively. Thus we should introduce the scaled
dimensionless variables
C* v* x*
C = Co' v = cN bjaCo'
0
x = -g- (S)
But this choice of variables is one of those considered [in (2.19)] when our
only concern was to put the problem in dimensionless form. The resulting
equations are given in (2.20).
ESTIMATING THE SIZE OF THE DIMENSIONLESS PARAMETERS

The equations of (2.20) form a nonlinear system; there is no apparent way
to solve them exactly. It is natural in such circumstances to look for a sim-
plification. Because scaled variables were employed, the dimensionless para-
meters v and I'/ that appear in (2.20) should provide good estimates for
the relative magnitudes of the various terms. If one of these parameters is
relatively small, the corresponding term should be negligible to a first approx-
imation. In such a case, perturbation theory might provide a method of
treating the problem.
To estimate the parameters, we must turn to the physiologists for quanti-
tative information. Using data compiled and analyzed by DB, we find the
estimates for minimum, typical, and maximum values of the various di-
mensional and dimensionless parameters listed in Table 8.2. This table
TABLE 8. 2. Estimated values of the various parameters.
Parameter Units Minimumvalue Typicalvalue Maximumvalue
r cm 10-6 5 x 10- 6 10- 4

L cm 4 x 10- 4 10-2 2 x 10- 2
cS cm 4 X 10-s 10-3 2 x 10- 3
D cm2 /sec 10- 6 10-s 5 X 10-s
No m0sm/cm 2 sec 10-10 10-7 10-s
p cm4 /sec mOsm 10- 6 2 X 10- 5 2 x 10- 4
Co m0sm/cm 3 3 x 10- 1
v Dimensionless 10-s 5 x 10- 2 102
TJ Dimensionless 4 x 10- 3 75 1010
K Dimensionless* 4 x 10- 3 5 200
K will be defined below.
shows that v is typically small and can be very small. (The maximum and
minimum values of the dimensionless parameters are exaggerated because
all the constituent dimensional parameters are unlikely to be of extreme size
in a single system.)
Sec. 8.3] Obtaining the Final Scaled Dimensionless Form 265
AN UNSUCCESSFUL REGULAR PERTURBATION CALCULATION

Since we have employed scaled dimensionless variables, we can deduce
from (2.20a) that C(x) ::::::l; i.e., C*(x*) ::::::
C0 To obtain more information,
we assume that the dependent variables can be expanded as a power series
in the small parameter v:
C(x) = c<0 >(x)+ vcCl>(x)+ v2 c<2 >(x)+ , (6a)
v(x) = v<0 >(x) + vvCl>(x)+ v2 v<2 >(x)+ . (6b)
We now substitute into (2.20). From (2.20b), for example, we obtain
[v<0> + vv<O+ .. J[c<0 >+ vc<l) + .. ] - ,,[c<0 > + vc 0 >+ .. ]'= (7)
(Here we introduce a two-element column vector, with the convention that
the upper element is to be used for O ~ x < 1 and the lower for 1 < x ~ l.)
The next step is to equate to zero the coefficients of each power of v. In do-
ing so, we must decide on the role of '1 A natural assumption is that r,
is independent ofv, in which case we find
v0 : C(O) = }, V(O)C(O) - '7C(O)i = (~) (7a, b)
(Sa, b)
It is readily seen that the solutions to (7) and (8) which satisfy the boundary
conditions (2.20c, d) are (Exercise 1)
c<0 > = 1, v< > =

0
(l),c< > 1
= (~). v<l) = - (~) (9)
The functions v<0 and c<0 are not continuous at x = l, which violates
(2.20e). There appear to be no solutions of the type sought.
RELATION BETWEEN PARAMETERS
The difficulty arises from inadequate handling of the two parameters

v and r, that appear in the differential equations of the problem.* (It turns
out that the parameter ,l. in the boundary conditions does not play a crucial
role.) Our basic simplification of (2.20a, b) for small v was to omit the term
vv' in (2.20a). We retained all the terms in (2.20b). But we may set 'I= v,,1,
where, from the definitions of r, and v in (2.22), r,1 = vf'I = cN~ o2 /aPCJ D.
In Segel's work, the assumption that 'f/ is independent of~ was rejected on the grounds that
a discontinuous solution is unacceptable. But the discontinuity can be "fixed up" by the methods
of singular perturbation theory (Exercise 9.2.9). This type of solution can be ruled out, however,
on the grounds that no trace of such behavior is found in the numerical results provided by DB.
That is, although a solution with a rapidly varying pan near x = 1 is certainly possible in principle,
it does not appear to be appropriate in the parameter domain of biological interest.
266 Illustration of Techniques on a Physiological Flow Problem. [Ch. 8
If it is assumed that 711 remains fixed as v approaches zero, then the term
proportional to C' would be omitted in the basic simplification of (2.20b).
Still another possibility is to write
aD
,, =-,'h where 172=v11= 2 (10)
v PC 0 cJ
Suppose that 172 remains fixed as v goes to zero. Making the substitution
2 /v, multiplying by v, and then letting v-+ 0, we see that the leading
71= 17
term in (2.20b) is now C' = 0. It has become clear that the presence of two
parameters allows contemplation of a number of different limits. Which one
is appropriate to our problem?
From (9) we know that
lim C(x; v, 11,l)
v-o
11,A fixed
has the acceptable value unity. We run into trouble, however, in looking at
the next approximation, since
1
lim v- 1 [C(x; v, 11,l) - I]= c 0 > = ( ) (11)
v-o O
11,A fixed
is not continuous. It turns out, as we shall see, that the appropriate limit
keeps fixed the quantity 172 defined in (10). For
lim C(x; v, 71,l) = 1, and lim v- 1 [C(x; v, 11,l) - I] (12)
..-o v-o
112,A fixed 112,A fixed
is continuous.
Before proceeding, it is worth noting that the limit used in (12) can be
written more fully as
lim C(x; v, 71,l). (13)
v-+O, ,r-+oo
IIV=ll2
x, 112,A fixed
An example of a function of 11,l, and v that does not permit the limit
(11) but does permit the limit (13) is k" sin (l/17v).For
lim kv sin
v-o 11V
(2..)
11,J.fixed
does not exist but
lim lev sin (2..)= A sin (2-).

v-O 11V 112
A, 1111=112
fixed
There is no automatic procedure that will lead inevitably to success in

problems, such as the present one, which involve several parameters. Try
variouspossibilities is one bit of advice. Many a confident invitation to the
Sec. 8.3] Obtaining the Final Scaled Dimensionless Form
reader to " consider the following dimensionless variables" leaves unmen-

tioned the trial and error leading to the correct choice. Somewhat better
advice is: Use your physical intuition. Because one of the authors provided
the analysis of the present problem it can be stated that the correct approach
here was originally arrived at (after a time) by a combination of trial and
error and the following physical reasoning.
When v is small, we see from the volume conservation equation (2.20a)
that the dimensionless solute concentration C is approximately unity. In
dimensional terms, this means that the solute concentration C* is approxi-
mately equal to the ambient solute concentration C0 . If C = 1, then of
course C' = 0, but making the approximation C' = 0 and eliminating the
term t1C' in (2.20b) is dangerous. One reason for this is that although C
may remain nearly equal to unity, it may fluctuate rapidly around unity,
so C' could be rather large. Even if C' is not too large, it is multiplied by a
parameter which might very well be large (see Table 8.2) so that the term
t1C' could be important.
On the other hand, to use the fact that C ::::::
1 to justify the approximation
v seems very sensible. The term vC represents the convection of solute.
vC ::::::
To assume that vC:::::: v is tantamount to replacing convection of solute at
the actual spatially varying concentration C(x) by convection of solute at
the approximate (dimensionless) concentration unity. This should be appro-
priate either if the actual concentration C*(x*) does not stray too far from
C0 or if solute convection is not too important.
FINAL FORMULATION
Just a little directed trial and error now leads to the final answer. If the sub-
stitution (10) is made, the lowest order approximation gi".es c<0 > = I, while
the next approximation gives an O(v) correction c 0 > to C and an 0(1)
velocity v<0 >.*The quantities c<l) and v<0 >satisfy a (linear) conservation of
solute equation in which vC is replaced by v<0 >,i.e., in which fluid carries
along solute of concentration unity.
One other consideration enters the final formulation of the problem,
but it is a trivial one. Formulas look a little simpler if one writes the pa-
rameter t72 of (10) in the form
(14)
Although the argument leading to it may have seemed circuitous, the pro-
cedure now is straightforward. Using (10) and (14), we replace the parameter
'1 by means of
cPCoL2)112
SO K=
( aD (15)
* It is encouraging at this point to note, using Table 8.2, that 712 , unlike 711 , is not typically
large. It has a typical value of about 4.
The problem of (2.20) now becomes
C- 1 = vv', v,c2 vC - t1.C' = vK

2
(7). (16a, b)
v(O)= 0, C(J) = 1, C and v continuous at x = 1. (16c, d, e)

0 < v ~ l, K,.A= 0(1).
Determine
1
Os=-
v(l)
This is the final scaled dimensionless form of the problem. It was quite a
struggle to obtain the proper formulation. But it is a straightforward matter
now, as we shall see in the next section, to solve the problem by the methods
of regular perturbation theory.
We stated in Section 1.1 that the solution of a mathematical problem
may not be the hardest job facing an applied mathematician. Our discussion
here provides a good illustration of this point.
EXERCISES
1. Verify (7) and (8), and obtain the solutions of (9). Also, verify (16).
8.4 Solution and Interpretation

We are now ready to solve our problem by a regular perturbation. We
shall find that the very first approximation to Os, the dimensionless quantity
of primary interest, is adequate for our purposes. The analytic expressions
obtained in our approximate solution are used to provide a deeper under-
standing of standing gradient flow. A noteworthy point here is the physical
interpretation of the dimensionless parameters.
A FIRST APPROXIMATION TO THE SOLUTION

Upon substituting the series (3.6a) and (3.6b) into (3.16a, b), we obtain
the following equations on collecting powers of v (Exercise l):
v0 : c<0 >= 1, i 2 c<0 >= o; (la, b)
Vl : C(l) = v<O\ )c2C(l) = ,c2[-(~) + V(O)C(O)]; (2a, b)
v2 : c<2 > = v<0 , )c2C(2) = ,c2[v(O)C(l) + V(l)C(O)]. (3a, b)

Sec. 8.4] Solution and Interpretation 269
Equations (la) and (lb) hold if and only if c<0 > = I. Using this, on com-
bining (2a) and (2b) we find that
(4)
The solution to this equation and the relevant boundary conditions can be
found by standard methods (Exercise 2). Thus, for O :::;;x :::;;1,
v<0 > = x - K 1 sinh KA.- ix,
and for 1 :::;;x :::;;A,
v<0 > = 1 - K 2 cosh K(I - A.- 1x),
Here
K _ A cosh K(l - r 1
) _ A sinh KA.- 1
1 K2- (7)
- K cosh K ' Kcosh"
The first approximation to the emergent osmolarity is
Os*
Os= - = [v(A.)r1 ~ [v<0 >(A.)r1 = (1 -K2)- 1
. (8)
Co
Since A.typically has the value 10, KA.- i will often not exceed unity. In this
case we can make the approximation sinh (KA.- i) ~ KA.- 1, which enables
us to write the strikingly simple formula
cosh K
O s~----
cosh 1 K -
(9)
COMPARISON WITH NUMERICAL CALCULATIONS

Let us compare our approximate results with numerical calculations done
for a number of particular cases by DB. Figure 8.4 gives a plot of Os as
determined by (9). Almost all of DB's numerical results, displayed by them
in five graphs, fall on this single curve. To show the extent of agreement and
disagreement we have also plotted a number of points from their Figure 7,
a graph of emergent osmolarity as a function of permeability in which L =
100 m, b = 10 m, r = 0.05 m, C0 = 0.3 m0sm/cm 3 , D = 10- 5 cm2 /sec,
and N0 = 10-5 (()), 10-6 (0), 10-7 (\7), 10-8 ( +),or 10- 9 (0) m0sm/cm 2 sec.
We have used a semilog plot of Os versus 5,c2 /6 to facilitate the comparison.
It will be seen that agreement ranges from good to excellent except for the
largest values of N 0 For such values the parameter v is no longer small
enough to permit neglect of the higher order terms in (6). The O(v 2 ) terms
in the series expansions are given in Exercise 3. As will be discussed further,
these terms improve the agreement at large v.
A clearer presentation of our approximate results is obtained by plotting
Os versus K as in Figure 8.5. We use (9) and do not employ a logarithmic
270
Os
t
l
1.5
\ 0
"' o~
0--
5K2f6
1 2 s IO 20 so
FIGURE 8.4. A plot of Os, the ratio of the concentration of fluid emerg-
ing from the open end of the channel (Os*) to the ambient concentration
(Co). According to (9), the approximate theoretical expression used, Os
depends only on the parameter K that was defined in (3.15). The various
points were obtained numerically from the "exact" solution by DB,
using various parameter values as detailed in the text.
scale on the abscissa as we did in Figure 8.4. To see the effect of solute trans-
port site, we also specialize (8) to plot Os versus K when A = 2, i.e., when solute
is actively transported over the first half of the channel.
We shall discuss our results further in a moment. First, let us pause to
make explicit what we have achieved from the point of view of perturbation
theory. Our problem was a second-order nonlinear inhomogeneous system
of ordinary differential equations. There was in addition one boundary
condition at each end of the interval. The inhomogeneous terms were differ-
ent in each of two different subintervals. Continuity conditions linked the
solutions appropriate for each of the subintervals. Thus the problem was not
a particularly simple one.
Diamond and Bossert treated the problem numerically for a number of
cases. Their analysis involved setting an extra boundary condition at x = 0,
" shooting" the solution toward x = l, and adjusting the extra boundary
Sec. 8.4] Solution and Interpretation
271
Os
FIGURE 8.5. Plots of the dimensionless emergent concentration Os

versus K. ( The parameter K measures the relative importance of convection
and diffusion; it is independent of the rate at which the solute is actively
pumped into the tube.) Bottom curve: theoretical prediction when A-+ oo,
where ,\ is the ratio of the channel length L to the length 6 of the region of
solute pumping. The results for,\= 10 are indistinguishable from those
plotted here. Top curve: theoretical prediction for ,\ = 2.
condition at x = 0 again and again (according to a certain scheme) until

the solution behaved properly at x = l.
By a regular perturbation method we were able to fit virtually all the
results of DB with the simple expression (9). The straightforward calculations
necessary to derive this expression could be repeated in an hour or two.
But the reader should not lose sight of what was required to set up these
calculations. In particular, he should recall the intuitive thinking (which
was aided by perusal of some numerical solutions) that went into the process
of scaling. By this process we were assured that the size of the dimensional
parameters represented the relative magnitudes of the various terms. The
reader should also remember that experimental information (borrowed from
the compilation of DB) enabled us to assert that one of these parameters was
small for many problems of physiological interest and that the other was
rather large. Finally, we should keep in mind that physical intuition coupled
with some trial and error was required to exploit the information on pa-
rameter size by means of the perturbation method.
Perhaps it appears discouraging at first that straightforward perturbation
calculations must be supplemented by considerable additional reasoning in
some problems. In the authors' view, although there are many problems that
can in fact be treated by essentially routine application of regular perturba-

tion theory, it is the more difficult applications that challenge a working
applied mathematician.
INTERPRETATION: PHYSICAL MEANING OF THE
DIMENSIONLESS PARAMETERS
To conclude this section, let us attempt to use the results we have obtained
in our study of standing gradient flow in tiny tubules to obtain a better under-
standing of how and why such flows take _place.
To give a simple but worthwhile example of what we have learned, our
analysis shows that since only ,c appears in the final result (9), doubling the
circumference/area ratio has approximately the same effect as doubling the
permeability. This effect may be important in relatively short channels.
More generally, identification of the important dimensionless parameters
governing a phenomenon is one of the chief fruits yielded by consideration
of simplified models. For the model under consideration here, dimensional
analysis showed that Os, the ratio of the emergent osmolarity Os* to the
ambient osmolarity C0 , is determined by three dimensionless parameters,
whatever the value of the eight dimensional parameters governing the
problem. Approximate solution of the governing equations by means of
perturbation theory led to the realization that Os is often only a function of
the single parameter K. In determining the meaning of this parameter, it is
more convenient to consider tK 2 , which can be written
I 2 [c(C - C0 )PL]C 0
2'' = aD(C-C 0 )/(L/2) (IO)
In the above expression C is a typical concentration in the channel. (For
definiteness, C can be regarded as the concentration halfway along the
channel.) The square-bracketed factor in the numerator is thus an estimate
of the total water flow per unit time into the channel across its boundaries.
This of course equals the total volume of water that flows out of the open
end of the channel per unit time. The entire numerator is therefore an estimate
of aF!nv, the convective flux of solute out of the channel. The denominator
is an estimate of aF!rr, the diffusive flux of solute out of the channel, since
(C - C0 )/(L/2) should approximate the concentration gradient at the end
of the channel. t Thus
1 2 F!nv
-K ~--. (11)
2 F!rr
To be compared with (11) is
Os* Fc!nv+ Fd~rr F!rr
OS=-= * = 1 +-.-. (12)
Co Fconv Fconv
t Our estimate of the concentration gradient is obtained by dividing the concentration differ-
ence between mid-channel and the open end by the half-length of the channel.
This is an exact result, which can be obtained from the defining relationship
for Os* in (2.18). Combining (11) and (12), we obtain
(13)
Thus, when the solution (9) is expanded for small values of K, we find that
Os~ 2,c- 2 , in agreement with the corresponding expansion of (13).
We have seen that the parameter " provides an estimate of the relative
importance of convection and diffusion that is closely related to Os. When
" is small, diffusion is dominant so that the graph of concentration vs.
distance very nearly represents the constant gradient, which is typical of
diffusion problems. (Compare curve 2 of Figure 8.3.) It is no wonder that
(13) is nearly exact in this case-for C*(L/2) is almost exactly the average
concentration in the tube, and [C*(L/2) - C0 ]/[L/2] is an excellent estimate
for the concentration gradient.
When" becomes large, then the interpretation (II) of2,c- 2 loses its quan-
titative accuracy, but still (13) provides a rough approximation. As we shall
now show, the approximate interpretation (11) of Equation (10) for ,c2 can
be exploited in order to understand a fundamental aspect of the phenomenon
under investigation.
A striking feature of flow from the salt glands of birds and the kidney tu-
bules of crocodiles is that the concentration of the secretion seems virtually
unaffected by changes in the rate of active transport into the tubule. DB's
numerical calculations showed that the standing gradient system is able to
reproduce this feature. In present terms we are speaking of the fact that
Os is independent of v unless v is large. How can this be? Formula (10)
gives the answer. If N 0 {J decreases, then the typical solute concentration C
decreases, less water is drawn into the channel, and the importance of con-
vection in sweeping out solute is lessened. According to (10), however, the
decrease of C results in a decrease of the concentration gradient at the open
end. This restores the balance between convection and diffusion, and thereby
keeps the emergent osmolarity at its original value. It appears that only when
the solute concentration in the fluid that issues from the active region is
rather large do changes in this concentration affect the proportions of solute
removed from the open end of the channel by convection and diffusion.
Determination of the next term in the perturbation series (Exercise 3) gives
the expected qualitative result, originally illustrated in DB's numerical work,
that when v is large, (9) provides an underestimate of the emergent osmo-
larity
Note that it is essentially the first term in the series that provides almost all
the interesting information. This occurs frequently.
To get an idea of what values of v signal the inappropriateness of the lowest
order approximation (9) in physiological situations, we observe from Figure
8.4 that agreement with (9) is good for the point () when 5,c2 /6 = 20 but poor
when 5,c2 /6 = 2. In the former case v = 5 but in the latter v = 50. From Table
Illustration of Techniqueson a PhysiologicalFlow Problem [Ch. 8
274
8.2, even values of v as high as 5 would be unusual. We conclude that the

lowest order approximation is expected to furnish an adequate description
of standing gradient flows in all but highly exceptional physiological situa-
tions.
It is apparent from its definition (2.22), with N 0 in the numerator, that v
is a dimensionless measure of the importance of active transport. If desired,
a more accurate interpretation of v can be obtained by writing
v = [ Pc~(C
2
:~:~C*(l)cJ[
C*(l)
][C Co]' 2
~o (14)
where C2 is a typical concentration toward the closed end of the tube. (Again
it facilitates interpretation to introduce the same factor into the numerator
and the denominator.) From (14) one obtains
V l F:iff]
= [ + [c*(l)]
Fconv
[C2
-Co]
Co
' (15)
where F!rr and F.~nvare evaluated at x = l, the end of the active transport
region.
The last of the three parameters governing Os isl= L/lJ, which obviously
measures the degree of concentration of the actively transporting sites at the
closed end of the tube. Also, ). does not enter the simplified formula (9), while
Figure 8.5 shows that use of (9) makes very little change even when l is as
small as 2. (When l is decreased, the pumping rate N 0 must be decreased
if the total solute transported is to remain the same, but for the range of
parameters of primary interest such a decrease in N 0 has no effect on Os.)
This very weak dependence of Os on l gives confidence in the model; it
shows that the features of standing gradient flow will be found as long as
active transport is confined " to a reasonable extent" toward the closed end
of the tube. To give one more observation on the effect of A, we note from
Exercise 3 that the larger l is, the higher v must be before effects of high trans-
port rate becomes evident. It is in accord with intuition that even high trans-
port rates have little effect when the active site only occupies a very small
fraction of the tube length.
FINAL REMARKS
In closing this chapter let us recall that the problem of standing gradient
osmotic fl.owhas served to illustrate a number of basic techniques in applied
mathematics such as nondimensionalization, scaling, and perturbation theory.
More important, the problem serves as a paradigm of how theory can be
used to take a step forward in scientificunderstanding.
As we have seen, the standing gradient problem was originally formulated
to make precise a suggested answer to a physiological riddle. The numerical
calculations of DB showed that the suggestion was a feasible one: active

solute pumping at the far end of a channel was demonstrated to be capable
of doing the job required of it for reasonable values of the various parameters
involved. The numerical calculations also helped in ascertaining the typical
values of various terms, as was required in the present analytic treatment
of the problem. The analytic treatment yielded a simple formula (9), which
summed up the operation of the standing gradient model.
There is some value in being able to consult this formula rather than a
number of tables or graphs for the "answer," given a particular set of para-
meters. A more important result of the analysis is the understanding of the
flow process, for the model is doubtless oversimplified so that precise pre-
dictions cannot be trusted. The understanding resides (i) in the isolation, from
eight dimensional parameters, of the dimensionless parameter K as the govern-
ing parameter of the flow, (ii) in the physical interpretation of K, (iii) in the
prediction that the independence of emergent osmolarity Os* on the solute
pumping rate N 0 is to be expected in physiological systems, and (iv) in the
analysis of why this prediction emerges. To see the virtue of theoretical work,
compare what you now know of standing gradient flow to the ingenious
but unsupported idea that started the whole train of reasoning: " secreting
tissues may work by active pumping deep in their ubiquitous infoldings. "*
EXERCISES
1. Derive (1), (2), and (3).

2. Solve (4), subject to appropriate boundary conditions. Thereby obtain
(5) and (6).
3. (a) Use (5) and (6) to show that v<I) satisfies the following equations:
O < x < I:(-; )\<>"-

1
vC1> = ~ KJKA-1 sinh 2KA-1X
- K1Asinh KA-x - K 1 KA-1X cosh KA-1X + x.
I<x :s; A: (; rvCI),,- v<IJ= - ~ K~ K/A sinh 2K( I - A-1x)
+ K2 KA- 1 sinh K( 1 - A- 1x).
Refinements and extensions of the standing gradient model have been made and analyzed.
See, for example, S. Weinbaum and J. R. Goldgraben, "On the Movement of Water and Solute
in Extracellular Channels with Filtration, Osmosis, and Active Transport," J. Fluid Mech. 53,
481-512 (1972). Weinbaum and Goldgraben stress that C*(L) can take values other than Co,
depending on conditions at the exit. They draw interesting conclusions from considering a
=
one-parameter family of solutions wherein C*(L) C, C arbitrary. Another important part of
their paper concerns etfects of an imposed pressure gradient.
(b) Show that solutions are

0~x ~ 1: v<1> ==A1 sinh ,c>.- +
1x B1 cosh ,c>.-1x
+ iK~,c>.- 1x - lK1,ccosh ,c>.-Ix
1 sinh 2,c>.-
-lKI,c2>.-2x2sinh ,c>.-1x - x, c<2> ==v<l)';

1~ x ~ >.:v<1> ==A2 sinh ,c(l - >.-1x) + B2 cosh ,c(l - >.-Ix)
-!K2K2>.-2xcosh ,c(l - >,-Ix)
-iK:,c>.,-I sinh 2,c(l - >.-ix), cc2>==v',
where the four constants A 1 , BI, A 2 , and B2 are to be determined
by the requirements v<0 (0) = 0, v<01 (.t) = 0, and vU>and v<1 >
continuous at x = 1.
(c) Simplify otherwise unwieldy expressions by neglecting terms of
order unity compared to terms of order .t. Obtain
= -5 sinh 2,c- 41 sinh ,c1- 6,ccosh2 ,c
vO>(>.)
12,c-). cosh ,c
(d) Make a further simplification by expanding the hyperbolic functions
in powers of K and retain the first nonvanishing term. Obtain
7,c8
= - 120>.cosh8 ,c
v(>.)
The correction is thus seen to decrease the dimensionless exit speed

v(.t) (at least when ,t - i and K are not too large) and therefore to
increase Os.
CHAPTER 9
Introductionto Singular
Perturbation Theory
SING u LA R perturbation theory has grown out of attempts to deal with

physical problems, the most important of which involves the boundary
layer phenomenon in fluid mechanics. The classical instance of this phenom-
enon occurs in the flow past an object such as an airplane wing. As will be
discussed with some care in Chapter 3 ofll, the appropriately scaled equations
in this case contain the effects of viscosity only in terms multiplied by an
extremely small parameter. Yet if these terms are neglected, computations
show that the air exerts no drag on the wing. This is D'Alembert's paradox
(see Section 15.4). The paradox was resolved early in this century when
Ludwig Prandtl showed experimentally and theoretically that however small
the viscosity or stickiness" of the fluid, viscous effects are important in a
thin layer near the wing. Prandtl's discovery not only was a breakthrough in
the understanding of aerodynamics, but it opened up the investigation, via
singular perturbation theory, of how small causes (e.g., small fluid viscosity)
can lead to large effects (appreciable drag).
It might seem that the practitioner of singular perturbation theory would
find relatively few important problems to work on, but this is not at all the
case. The abundance of such problems is at least partially explained by the
fact that singular perturbation techniques seem to be appropriate for examin-
ing the behavior of solutions near their worst" singularities. Taking the
bull by the horns" and examining functions near their worst* singularities is
the best way to obtain information about qualitative behavior, and it is
elucidation of this behavior that is very often the goal of a theoretician.t
In this lies the point of a remark once made about a distinguished applied
mathematician, "All he can do is singular perturbation problems. But of
course he can turn all problems into singular perturbation problems! "
Our introduction to singular perturbation theory begins with a discussion
of algebraic equations having a small parameter multiplying the term of
highest degree. First, introductory concepts of singular perturbation theory
are used to find the behavior of certain roots which are missed" by regular
perturbation theory. Then successive approximation procedures are applied
to find better approximations to these roots, or to suggest appropriate series
In the language of complex variable theory, the worst singularities are essential.
t See L. A. Segel, "The Importance of Asymptotic Analysis in Applied Mathematics,"
Amer. Math. Monthly 73, 7-14 (1966).
277
Introduction to Singular Perturbation Theory [Ch. 9
assumptions. After this we turn to an extensive treatment of a particular

ordinary differential equation which serves as a model equation for singular
perturbation phenomena. With it we illustrate a goodly amount of the analysis
required in applying singular perturbation theory to simple boundary value
problems.
We are using the word" model" here in a somewhat different sense than
we have heretofore. Most of the time we use descriptive models, which are
designed to caricature definite natural phenomena. Now we are speaking of a
conceptual model, which is "constructed to elucidate delicate and difficult
points of a theory. Not always is it easy (or even desirable) to draw a sharp
dividing line between the two types, but the extremes on both sides are clearly
recognizable."*
The model equation that we shall study is a linear second order equation
with constant coefficients and a small parameter multiplying the highest
derivative. We carefully examine the exact solution to this equation and its
associated boundary conditions. This enables us to develop the main lines of a
singular perturbation procedure that works for a large class of problems in
which a small parameter multiplies the highest derivative in a differential
equation. The variety of singular perturbation problems that one is likely to
meet is so large as to preclude the hope of a comprehensive recipe. Certainly,
none is offered here; indeed, none is available. Instead, stress is laid on the
" spirit of singular perturbations" with the hope that the reader will be
spurred to further study of this fascinating and valuable subject.
Opportunity for further study is offered in the next chapter, where singular
perturbation theory is used to analyze enzyme-catalyzed chemical reactions.
Then in Section 11.2 we employ the powerful "multiple-scale" approach to
singular perturbation calculations on the now-familiar pendulum problem.
All these illustrations concern problems that are governed by ordinary
differential equations. In Chapter 3 of II we discuss the application of
singular perturbation techniques to viscous flow at high Reynolds number.
This is the boundary layer problem which, as mentioned above, provided
the primary stimulus to the development of singular perturbation methods.
9.1 Roots of Polynomial Equations

A SIMPLE PROBLEM
A very simple problem which gives insight into singular perturbation theory
is the quadratic equation
em2 + 2m + 1 = 0, where O < e ~ 1. (1)
* See "Some Mathematical Models in Science" by M. Kac, Science 166, 695-99 (1969).
Sec. 9.1] Roots of Polynomial Equations 279

To obtain an approximate solution, we neglect the seemingly small term em 2
and find
2m + 1 ~o, m~ -t.
We can improve this "nai:ve" approximation by any of the three methods
discussed in Chapter 7 (Exercise 1). For example, if we assume a series
solution
(2)
we find that
m= -t - ie - / 6 e2 + . (3)
We can thus obtain an arbitrarily accurate approximation to one solution of
the quadratic (1). But according to the fundamental theorem of algebra,
there are two such solutions. Here, then, is a perturbation problem in which
a straightforward approach does not produce a completely satisfactory
answer. This is the kind of problem in which the singular perturbation method
can be used with advantage.
The singular nature of (l) is connected with the fact that the full equation
is of second degree (and therefore has two roots), while the approximate
equation 2m + 1 = 0 has but a single root. By comparison, the equation
m 2 + em - 4 = 0 (treated in Exercise 7.2.9) is second degree, and so is the
approximate equation m 2 - 4 = 0; this is an example of a regular pertur-
bation problem. It is a characteristic feature of singular perturbations that the
naively approximated problem has a different qualitative character from the
original problem.
In searching for an approximation to the second root m of the quadratic,
we begin with the knowledge that neglect of the term em2 is not justified.
We have already found the root to the first degree equation that results when
this term is neglected.
If m were 0(1), i.e., if m approached a (finite) limit as e-+ 0, then the
neglect of the term em2 presumably would be justified for small &-because
then this term would be small compared to the term 1. We are presumably
dealing with a root that becomes very large when e-+ 0 (or, possibly, with a
root that does not approach a limit when e -+ 0).
Ase-+ 0, each of the terms em2, 2m,and l has a certain magnitude. If all
three terms have the same magnitude, the task of obtaining mis not lightened
by the fact that e is small: all three terms in (1) are important and we are back
to the original equation. Perhaps two of the terms are of the same magnitude
and the third is relatively small. We have agreed that em2 must be retained,
so the possibilities are em2 and 1 of the same magnitude, 2mrelatively small;
em 2 and 2m of the same magnitude, l relatively small. If the first possibility
obtains, then from (l) we see that em2 + 1 ~ 0 or m~ ie- 112 In this case,
however, 2m ~ 2ie- 1' 2 and the" relatively small" term is in fact of relatively
280 Introduction to Singular Perturbation Theory [Ch. 9
large magnitude (12ml ~ 1). The second possibility is that em2 + 2m ~ 0.

If we discard the alternative m~ 0 for the moment, we find that m ~ - 2e- 1
Here the term omitted is, indeed, relatively small compared to the terms
retained:
Thus our approximation appears to be consistent. The discarded alternative

should be ignored, for if m~ 0, then the neglected term is not small compared
to the terms retained, which is an inconsistency.
To improve on our first estimate ofm, let us employ the successiveapproxi-
mations procedure. As discussed in Section 7 .2, the idea is to write the equa-
tion in the form/(m) = g(m), where the initial approximation m<0 >is provided
by the solution to f(m<0>)= 0 and higher approximations by the solutions to
!(m<O) = g(m<i- 1>),i = l, 2, ....
It might appear that the way to do this is to choose/(m) = em2 + 2m, but
then setting/= 0 yields the spurious first approximation m<0 > ~ 0 as well as
the true first approximation m<0 > ~ -2e- 1 To avoid this, we divide (I) by m.
Thus a scheme to obtain successive approximations mto m is given by
m<0 > = -2e- 1 and
em+2= --!-
m<-1>
or
mCi) = - ~- __ 1_-, i = 1, 2, .... (4)

e em<-1 >
With this
and
m<2>= -~+-1-.
e 2-te
Note that if we expand the second term in powers of e, we obtain
m~ m<2> = -2e- 1 + t + ie + .
Inspection of (4) shows that a similar series can be obtained for m(J), m<4 >,etc.
Here, as in other problems, it is probably best to abandon the successive
approximations method once the form of the solution becomes clear. We have
seen from the above calculations that the root m can be written in the form
(5)
Sec. 9.JJ Roots of Polynomial Equations 281
The coefficients m 1 can be obtained with less labor by the series method,
starting with the assumption (5), than by successive approximation [Exercise
2(a)]. Alternatively, one can begin with the assumption that em can be expanded
in powers of e and use the parametric differentiation procedure [Exercise 2(b) ].
Let us summarize. From (3) and (5) we see that the quadratic equation
em 2 + 2m + I = 0 has the roots
m= -!+, m= -2e-1 + .... (6)
An arbitrarily accurate expression for the first root can be obtained by the
methods of regular perturbation theory, since that root has a power series
expansion. The second root becomes infinite as e --+ 0. The leading term in its
approximate expression can be determined by finding and equating to zero
a pair of terms with matching and relatively large magnitudes. One of these
must be the term em2 whose neglect, on the "naive" grounds that it has a
small parameter as a factor, led to an approximation only for the 0(1) root
m= -!+.
Our first example (I) is readily solved by the quadratic formula. The
expressions for the two roots may be expanded to provide a check on our
calculations (Exercise 3). It should already be apparent, however, that the
approach just illustrated can be used on any algebraic equation with a small
parameter multiplying the term of highest power.
A MORE COMPLICATED PROBLEM
Consider the quartic equation
ex 4 + ex 3 - x2 + 2x - I = 0, O<e~l. (7)
Roots that approach a nonzero limit as e--+0 can be approximated by neglect-

ing the terms ex 4 and ex3 This gives the quadratic x 2 - 2x + I = 0 with the
double root of unity.
Example 1. Find higher approximations to the roots of (7) that are approxi-
mately equal to unity.
Partial Solution. We assume
Upon substitution into (7) we obtain

e[2 + O(e)] - xie 2 + O(e 3 ) = O.
Collecting terms of O(e), we obtain 2 = 0, a contradiction. The form we assumed
for the solution must be incorrect. To find the correct form, we set up a successive
approximations scheme. Putting the dominant portion of (7) on the left, we consider
(x- 1)2 = ex4 + ex 3
This leads to the scheme

Xo = 1; x. = 1 +Ve~ -1 + ex!_i, n= 1,2, ... ,
for one of the two roots and
x 0 = 1; x. = 1-V ex!-1 + ex!-i. n=l,2, ... ,

for the other. A couple of iterations indicate that the roots can be expanded in
powers of e1 ' 2 Let us introduce a= e1 ' 2 so that (7) becomes
a 2x4 +a x 2 3
- (x - 1)2 = 0.
Then the problem reduces to a straightforward determination of the coefficients in

the expansion
The reader is asked to carry out the calculations [Exercise 4(a).J
To obtain first approximations for the two remaining roots, we again

employ the very frequently useful balancingprocedurewhich supposes a pair
of terms to be of the same magnitude and the remaining terms to be negligible.
We have already considered the case in which the first two terms of (7) are
negligible. Thus the alternatives that remain for balancing terms are (a) ex4 ~
- ex3, (b) ex 4 ~ x 2 , (c) ex4 ~ -2x, (d) ex 4 ~ 1, (e) ex 3 ~ x 2 , (f) ex 3 ~ -2x,
(g) ex 3 ~ 1, with the remaining terms negligible.
The zero roots that arise from alternative (a) can be excluded, as above.
The remaining root x ~ -1 also leads to an inconsistency, for the "large"
terms ex4 and -ex 3 are in fact O(e).
Alternative (b) gives the nonzero roots
1
~
(O) __
P (8)
-
[In (8) we have denoted the positive and negative roots by p and n.] Using the
approximations (8) to x, we see that ex4 and x 2 are O(e- i ), while the remaining
terms behave as follows:
1 = 0(1). (9)
These are indeed relatively small so that (8) gives initial approximations to
the two roots we are investigating. It can be shown, as expected, that
alternatives (c)-(g) are all inconsistent [Exercise 4(b)].
To obtain more accurate expressions for the O(e- 1' 2) roots by successive
approximations, we mimic our earlier procedure and, dividing through by x 2 ,
rewrite (7) as
ex2 - 1 = x- 2
- 2x- 1
- ex,
Sec. 9.1] Roots of Polynomial Equations
where the left side is the binomial whose two roots are the zeroth approxima-
tions (8). The iterates therefore satisfy
a[x<o>]2- I ==0, a[x<1c>]2
- I = [x<1c-1>r2- 2[x<1c-1>r1- ax<1c-1>;
k= 1,2, ....
Solving for x<">,
we obtain the final iteration scheme
(IOa)
n<O)__ -1
-~
(lOb)
Here p<k> denotes the kth approximation to the positive root, p, of magnitude
a - 112 and n<k>is the kth approximation to the negative root, n, of this
magnitude. Note that each of the two roots p and n is associated with a
uniquely determined sequence of iterates. Beginners often make the mistake
of sprinkling signs throughout formulas that correspond to (10). This
results in a proliferation of "approximations" that cannot be correct, since
the total number of roots must, of course, equal the degree of the equation.
For each root there must be an iteration scheme that yields a unique answer at
every stage.
From (10) we find
(11)
We can expand these expressions about a= 0. As is common in successive

approximation schemes, only the leading term in the correction is trustworthy,
so we obtain finally
(12)
Determination
of thenextapproximation
confirms
whatis alreadyindicated
by (I I) and (12): p and n have series expansions in powers of Je.
The reader should fill in the details of the above example and carry out
some exercises to gain facility in exploiting the presence of a small parameter
in algebraic equations. Our discussion has illustrated the only relatively
difficult features: (i) the need to be careful of multivalued roots, and (ii) the
possible occurrence of fractional powers of e in series expansions for the roots.
The appearance of fractional powers can be anticipated by application of the
successive approximations procedure.
THE USE. OF SCALING

Those who are familiar with the scaling ideas of Section 6.3 might find it
worthwhile to apply those ideas here. Thus, in (1) only one of the two roots
is 0(1), and only for this root is the em2 term expected to have a small effect
on the answer. The other root is O(e- 1), so that the change of variable
m = e- 11r1is necessary to obtain a problem with a correctly scaled 0(1)
variable 1r1.In terms of 1r1,(1) becomes
or 2
1r1 + 21r1+ e = 0,
which yields the approximation 1r1~ -2 in accord with earlier results. Note
that in this example, and in the other example of this section, no single scale
can be used to characterize the complete solution of the problem. This will be
found to be a distinctive property of singular perturbation problems.
EXERCISES
1. (a) Assume a solution to (1) of the form (2) and thereby verify (3).
(b) Obtain (3) by the parametric differentiation procedure.
(c) Obtain (3) by successive approximations.
2. (a) Find a series expansion to the second root in of(l) by assuming (5).
At the least, find mo and m 1
t(b) Verify the above calculations by using the parametric differentiation
method.
t3. Solve (1) by the quadratic formula; expand the expressions for the two
roots in series; and compare with the results of Exercises 1 and 2.
4. (a) To a first approximation, as the text showed, (7) has a pair of
equal roots. Find a three-term approximation to these roots by
completing Example 1. In particular, show that the roots are not
equal.
(b) Show that alternatives (c)-(g) that were proposed in the discussion
of (7) are inconsistent.
(c) Graph the four roots of (7) as a function of e, for small e.
(d) Verify (12) and find one more term in the series using successive
approximations.
(e) Obtain the result of (d) by using series.
t(f) Obtain the result of (d) by a modified parametric differentiation
approach.
5. Find a first approximation, valid for small positive e, to all roots of the
following equations.
(a) ex4 - x 2 + 3x - 2 = 0.
(b) ex 5 + x 3 - 1 = 0.
(c) x + ex3 = 2.
(d) ex 3 + ex 2 + x - 1 = 0.
Sec. 9.2] Boundary Value Problems for Ordinary Differential Equations 285
6. Find a second approximation to the roots estimated in Exercise 4.

7. Discuss in detail the character of the roots of
epn(x) + Pm(x) = 0, n>m,
where P; denotes a polynomial of degree i.
8. This problem outlines a more deductive approach to the estimate of the
root m of (1). Let m(e) = E(e) + R(e). Because we wish the remainder R
to be small compared to the estimate, we require that lim,_ 0 (R/E) = 0.
From (1),
e(E + R) 2 + 2( + R) + 1 = 0. (13)
t(a) Suppose that IeE\ - co. Obtain a contradiction.

t(b) Suppose that IeEI - 0. Obtain a contradiction.
(c) From (a) and (b), if eE has a limit, then it must approach a non-
zero constant. Determine this constant.
9.2 Boundary Value Problems for Ordinary Differential

Equations
In this section we shall study the use of singular perturbation methods to
obtain approximate solutions for second order ordinary differential equations
with a small parameter multiplying the highest derivative. Most of the
discussion concerns a simple "model" problem. Some of the results found in
the previous section will prove useful here.
EXAMINATION OF THE EXACT SOLUTION TO A MODEL

PROBLEM
Consider the following equation for y(x):
d 2y dy
e dx 2 + 2 dx + y = O; 0 < x < 1, 0 < e < 1, (1)
with the boundary conditions

y(O) = 0, y(l) = 1. (2a, b)
Note that e is positive.
Since e is very small, we naively suppose that the first term in (1) may be
neglected. This gives the equation
dy
2 dx + y=O (3)
with the general solution

y = K exp ( -!x),
where K is an arbitrary constant. A dilemma at once appears, for if K is

determined by (2a) the "approximation"
y~o ~
results, while if K is determined by (2b), the" approximation" is

y = exp (t - !x). (5)
Neither "approximation" can satisfy both boundary conditions, so that
as e -+ 0, neither can approach the true solution for all x in the closed interval
[O,I]. Is either of the two "approximations" evera good one? If so, in what
part of the interval? How does the solution behave in the remainder of the
interval? Can one obtain an approximation that is good throughout the
interval?
To answer these questions, we shall solve {l) and (2) exactly and then
approximate the answer for small positive e. Then we shall reconsider the
problem, pretending that we do not know the exact solution but keeping in
mind its qualitative behavior. We shall be able to develop an approximation
procedure that will work for a large class of problems whose exact solutions
are impossible to attain.
Before embarking on this program, let us note that the present problem
possesses " singular" features that are similar to those of the algebraic
equations discussed above. For the second degree polynomial equation {1.1),
the approximate equation was of first degree and had a single root. Here the
approximate equation (3) is first order, so the two boundary conditions (2)
cannot both be satisfied. In Equation (1.1) the term em2 could not be negligible
for all roots. Here the term ed 2y/dx 2 must be important at least for portions
of (0, 1). It must be that d 2y/dx 2 is large in those portions; otherwise e(d 2y/dx 2 )
would indeed be negligible for small enough e.
Returning to the analysis of (1), we first observe that since this equation
is linear and has constant coefficients, it has the general solution
y = C1 exp (m1x) + C2 exp (m2 x), (6)
where m 1 and m2 are the two roots of
em2 + 2m + 1 = 0. (7)
When the boundary conditions of (2) are imposed, we obtain the final
solution,
(8)
Equation (7) is the same as Equation (1.1), so that when e is small, we may
employ our earlier analysis, as summarized in (1.6), and approximate m 1
and m 2 by
1 2
m2 = --. (9)
e
Using (9), since e is positive, we see that we can neglect exp (m 2) compared
with exp (m1) in the denominator of (8). We thus write the approximate
solution to our problem as
O<e~I. (10)
What is the qualitative behavior of the approximate solution (10)? As
required by boundary condition (2a), this approximation has the value zero
at x = 0, since the second term in the square bracket has exactly the same
magnitude as the first and is of opposite sign. As x increases, however, the
second term decreases rapidly. When x = e this term is only -e- 2 , about
one-seventh of its value at x = 0. When x = 2e it is less than 2 per cent of
this value. Thus, except when x is within a few times e of x = 0, the second
term is negligible and (IO) may be further approximated to yield
y = elf2e-xf2 = ll/2)(1-x)_ (II)
The graph of (IO) may then be drawn, as in the solid line of Figure 9.1. We
see that the solution changes rapidly in a layer near x = 0 whose thickness
is O(e). Such a region of rapid change near a boundary is called a boundary
layer.
Notice that (I I) is the same as (5). Thus, for small positive e, the solution
to the differential equation (1) and the boundary conditions (2) rises rapidly
from its assigned value of zero at x = 0 until it merges with (5), the solution
y
e=O.l
e
FIGURE 9.1. Thegraph of (10), a good approximation to the solution
of (1) and (2) when e is small andpositive (solid line). The "outer ap-
proximation" [(11), (14), or (S)) is depicted by a dashed line.
2.88 Introduction to Singular Perturbation Theory [Ch. 9
to the naive equation (3) that satisfies boundary condition (2b) at x = 1.

This "naive" solution is depicted by a dashed line in Figure 9.1.
It is instructive to consider the case when e is small and negative.Approxi-
mate roots of (7) are still given by (9), but now it is the first term of the
denominator of the exact solution (8) that is negligible. We have
lei~ 1, e < 0. (12)
Moreover, the first term of this expression is exponentially small for all x in
[O, l] so that we can write, with negligible loss of accuracy,
y(x, e) ~ e<2/t)(1-x), lel~I,e<O. (13)
The sole term remaining in (13) is virtually zero except in a boundary layer
where xis within a few times e of x = 1. (See Figure 9.2.) For small negative e,
e=-0.J
FIGURE 9.2. The solution to (1) and (2) when e is small and negative.
the solution of (1) and (2) is thus well approximated by the (identically zero)
solution of the naive equation that satisfies boundary condition (2a). The
boundary layer still has a thickness that is O(e), but in contrast to the case
where e is small and positive, it is now located near x = 1.
If this example is typical of a class of second order boundary value problems
(and it is), it appears that an appropriate "naive" equation serves to
furnish a good approximation to the solution of the full equation except near
one of the boundaries. There, the solution changes rapidly to conform with
the boundary condition that cannot be satisfied by the solution to the naive
equation. Which boundary condition should be applied to the naive equation,
and which left to be satisfied with the aid of a boundary layer, depends on the
relative signs of the various terms. (It.is not merely the sign of e that determines
this, for the equation can always be multiplied by -1 if necessary, to make
Sec. 9.2) Boundary Value Problems for Ordinary Differential Equations 289
the small parameter positive.) If we assume that the boundary layer will
either be near x = 0 or x = 1, its location can presumably be ascertained
fairly efficiently by trial and error.
So that the reader will not entertain false notions, we must warn him that
although the generalizations of the previous paragraph apply to a large class
of important problems, they do not, for example, apply to all boundary
value problems for linear second order ordinary differential equations with
a small parameter multiplying the highest derivative. We shall return to the
matter of delineating the problems for which singular perturbation methods
are appropriate. For the present, a mood of cautious optimism is appropriate.
It is profitable to examine more carefully the relation between the solution
to the approximate equation (3) and the solution to the full problem (1).
Because the solution to (3) and the appropriate boundary condition provides
an approximation that is valid outside a boundary layer, it is customary to
associate the word outer with that approximation. Until further notice we
shall consider the case e > 0; therefore the appropriate outer approximation
will be that of (I I). We use O for outer and write
Yo(x) =exp [!(I - x)]. (14)
We know that y 0 (x) is a good approximation to the exact solution y(x, e)
except near x = 0. Indeed, it is not difficult to show [Exercise l(a)] that
lim y(x, e) = y 0 (x) for x in (0, 1].* (15)
alO
The limit is uniform for a :::;;x :::;;l, a positive. That is, given any E > 0,
there exists a positive number D that depends on E but not on x, such
that ly(x, e)- y 0 (x)I < E whenever O < e < D for all x in [a, I].
Equation (15) contains the first of several limits in this section that
involve a function of more than one variable. Whenever we take a limit with
respect to one of the stated variables, it will be tacitly understood that the
remaining variables are fixed during the limiting process. Thus, in (15) the
dependence of y on x and e is explicitly stated. In the limit. x is considered to
be fixed as e decreases toward zero.
Returning to our discussion of y 0 (x), we note that, using (10} and (14),
lim [lim y(x,

x+O a+O
e)]= y (0) = e
0
112
(16)
But
lim [lim y(x,

i+O x+O
e)]= lim O = 0,
a+O
(17)
so that the operations lim,,1.0 and limx,1.ocannot be interchanged. Such an

interchange would be permitted if the limit of (15) were uniform over (0, b)
The downward arrow in (IS) signifies that a approaches zero through positive values.
290
for some positive b.* Thusy(x, 8) does not approachy 0 (x) uniformly on such
an interval, as can also be deduced directly.
To sum up, the outer solution y 0 (x) does not satisfy the boundary condition
at x = 0. However, y 0 (x) is the limit as 8 LOof y(x, 8) in (0, l], but the limit
is not uniform. We say that the approximation y 0 (x) is not uniformlyvalid
in (0, 1].
However close a fixed x is to the origin, that x becomes farther and
farther outside the boundary layer as 8 L0. Thus, from (15) the outer solution
y 0 (x) is aptly named, for it approximates the true solution only outside the
boundary layer. But what is an appropriate approximation for small 8 inside
the boundary layer? We now turn to this question.
It turns out that there is little to be gained in arbitrarily pinpointing a
precise" edge" to the boundary layer. On the other hand, it seems clear, for
example, that x = 0.18 is well inside the boundary layer, while x = 48 is far
toward the edge of the boundary layer. This is true regardlessof the value of 8,
as long as 8 is small. To find a typical value of the solution well inside and far
toward the edge of the boundary layer, then, one could consider, respectively,
lim [y(x, 8)lx=o.1e] and lim [y(x, 8) lx=4 .].

.,1.0 t,1.0
More generally,
=
lim [y(x, 8)1x=~,] yi(e),
!0
efixed, (18)
will provide an approximation yi(i;) to the solution at an unchanging pro-

portional distance i; into a boundary layer whose absolute thickness shrinks
continually as 8 L0. We call this the inner or boundary layer approximation,
where I stands for" inner." Using alternative notation, we define the function
Yby
Y(i;, 8) s y(i;8, 8) (19)
with which
YJW= lim Y(i;, 8). (20)
!0
We emphasize that i; is held fixed in the limits of (18) and (20).

In our present example, from (8),
y(x, 8) = K[exp (m1x) - exp (m2 x)],
so
* See, e.g., Franklin (1964), p. 395.

From (9), or more fully, from our earlier discussion concerning the roots of
the quadratic equation (7),
lim em 1(e) = 0, lim emi(e) = -2 .
o iO
Hence
yi(~) = lim Y(~, e) = e1;2(1 - e-2~). (21)
o
Figure 9.3 shows that yi(~) does indeed provide a good approximation to the
exact solution within the boundary layer .
I
.,,.--
.....
',
I '
I ',,
I .._
I ',
I
II '
', ,...._
I ,..._
I
I
I
I
I
I
I
I
I
I
E =0.1
FIGURE 9.3. Comparison between the exact solution to (1) and (2) for
small positives (dashed line) and the" inner" boundary layer approxima-
tion (21) (solid line). Agreement is good near x = 0.
FINDING AN APPROXIMATE SOLUTION BY SINGULAR

PERTURBATION METHODS
Let us return to the original problem of (1) and (2):
d 2y dy
e- 2 + 2- + y = O; y(O) = 0, y(l) = 1, O< e <O. (22)
dx dx
We shall now derive an approximate solution to this problem, pretending that
we do not know its exact solution. We shall take from our previous discussion
the supposition that the solution to the outer equation well approximates the
exact solution except in a thin layer near one of the boundaries. For the
moment we shall also take as given that the boundary layer is located near
x=O.
292
The outer approximation does not vary rapidly, so it should be permissible

to neglect the term ed 2y/dx 2 Using the advance knowledge that the boundary
layer is near x = 0, we require the outer approximation Yo to satisfy the
boundary condition at x = 1. Thus
dyo
2 dx + Yo =0, YoO) = 1, (23a, b)
so
Yo(x) = exp [HI - x)J, (24)
in agreement with (14).
Let the boundary layer near x = 0 have a thickness whose order of magni-
tude is given by <5(e).The layer thickness approaches zero as e ! 0, so we
require that <5! 0 as e ! 0. To ascertain how the solution behaves in the
boundary layer, we introduce
x
~ =-,() (25)
which permits us to measure proportional distance into the layer. [We know
from (18) that <5(e)= e, but we are temporarily "forgetting" this.]
Strong motivation for the change of variable (25) is provided by the
scaling concept (Section 6.3). The length scale (x scale) of the outer solution
(24) is of order unity. This scale is not appropriate in the boundary layer,
however, for there the solution changes rapidly in a distance of magnitude <5.
Thus (25) can be regarded as a rescaling appropriate for the boundary layer
region. Two scales, <5inside the boundary layer and unity outside, must
thus be used to specify the variation of the solution to (22). Such double
scales are a telltale sign of a singular perturbation problem, as we have
mentioned in the discussion of .. unorthodoxy" in Section 6.3.
Let us introduce the change of variable (25) into the differential equation
of (22). Replacing the function y by Yasin (19), we obtain
e d2 Y 2dY
152 d,2 + J de + Y = o. (26)
We now balance terms pairwise, just as we did in determining the magnitude

of the second root m of the quadratic equation (1.1). Since <5is the length
scale in the boundary layer, the expressions d 2 Y/de2, dY/de, and Yare all 0(1)
there. The orders of magnitudes of the various terms are
e
<52' ~ 1. (27)
Hopefully, two of these balance and the third is negligible (for then the
problem can be simplified by neglecting a term). One term of the balancing
pair must be s/o2 , since we have already seen that an approximation which
is valid throughout [O, 1] cannot be obtained if the second derivative term is
never taken into account.
If the first expression in (27) balances the third, then o = s 1 ' 2 Hence 1/lJ
is large compared to the other expressions (1/o= s- 112 ~ I), not negligible.
If the first expression in (27) balances the second, then
s I
2b =- b or o= s (28)
Now the remaining expression is negligible (1 ~ e- 1

), so (28) provides the
scaling that we seek. With (28), (26) becomes
d2 Y 2dY
dez + de + s y = 0. (29)
This is the innerequation;it has been obtained with a scaling appropriate to

the inner or boundary layer. As a first approximation, we neglect the last
term and take as an approximate equation describing behavior in the bound-
ary layer
d 2YI 2dy1
(30)
de2 + de = 0
e
[In passing from (29) to (30), we have let s ! 0 with fixed so that the replace-
ment of Y by y 1 is called for, by (20).] This second order equation requires
two boundary conditions. One stems from y(O) = 0. In the present variables
this is
yi(O) = 0. (31)
The second boundary condition on y, y(l) = I, will not supply a second
restriction on y 1 , since x = 1 is far outside the boundary layer, while all the
above discussion has been aimed at approximating a function that describes
the solution inside this layer of rapid change. For the moment, then, let us
put aside the question of the second boundary condition and solve (30) and
(31). We obtain
yr(e) = C(l - e- 2 ~), (32)
where C is a constant.
MATCHING
It turns out, not surprisingly with hindsight, that to specify C we must
match the inner solution at its "farthest extremity" with the outer solution
at its" nearest extremity," for both these extremities correspond to the same
"edge" of the boundary layer. One hopes for the existence in some sense of
such an "edge" (or, better, an overlap or intermediateregion), where both
the inner and outer solutions are reasonable approximations to the true
solution. The two approximations may reasonably be required to agree in an
overlap region.
The matching is a subtler matter than it might seem at first and is an
object of continuing intensive study at this writing. It appears that the best
way to proceed begins by recognizing that when x = 0(()), then xis within
the boundary layer, whereas when x = 0(1), then xis outside the boundary
layer. For x to be in the intermediate region, we should let x = O[E>(a)]
where O(E>)lies between 0('5) and 0(1); i.e.,
E>
Iim 7 =oo, lim E>=0. (33a, b)
+0 U
Now, when x was within the boundary layer [x = O(fJ)],it was appropriate to
e e
introduce the variable = x/fJ and to let s ! 0 with fixed. Analogously,
when x is in the intermediate region it is appropriate to introduce 11,where
x
,, =- (34)
E>
and to let s ! 0 with 11fixed. Thus the inner and outer approximations are
said to match if they have a common limit when the new variable 11is
introduced and held fixed as s tends to zero. Hence matching requires that
11fixed. (35)
In the present case, the left side of (35) is, from (24) and {33b),
lim [e<112><1-x>lx=11e]= lim e<1t2)(1-11e)= e112. (36)
+O +0
The right side of (35) is, from (32) and (33a),
lim C( 1 - e- 2 ..et~)= C. (37)
lO
Thus (35) requires that C = e 1' 2 so, from (32),

Y1(e) = e112 (1 - e- 2 ~), (38)
and this agrees with (21).
For our particular problem, both from the exact solution and from a
fairly comprehensive-seeming approximation method, we have derived an
inner approximation yi(e) and an outer approximation y 0 (x). It would be
useful to have a uniformapproximationYu, which would be valid throughout
[O, 1]. This can be obtained by adding y 1 and y 0 and subtracting their common
part. Thus
(39)
In the inner region the sum of the first and third terms on the right is negligible,
leavingtheinnerapproximation. In the outer region the sum of the second and
third terms is negligible, as can be seen from (35), which leaves the outer
approximation. Thus Yu reduces to the appropriate approximation in each
region. For the example under study we have, from (24), (38), and (36),
Yu(x) = e<l/2)(1-,c) + el/2(1 _ e-2x/') _ et/2 = el/2(e-x/2 _ e-2:xf),
which agrees with (10), our initial approximation to the exact solution.
FURTHER EXAMPLES
In our study of the illustrative problem (22) we took as given that the
boundary layer was located near x = 0. We know that this is the case when
e J.0, but the boundary layer is near x = 1 when e j 0. Fortunately, if we make
an incorrect assumption concerning the location of the boundary layer, it
shows up, since it leads to a breakdown in our procedure. We illustrate this
and, at the same time, provide more familiarity with the singular perturbation
procedure in the following examples.
Example 1. Show that the boundary layer procedure breaks down when e to
if a boundary layer is assumed to exist near x = 0.
Solution. All steps of the analysis go through just as before except that at the
very last stage, (37) is replaced by
lim C(l - e - 2 ' 8 '~) = oo
ro
and no matching is possible.
Example 2. Carry through the calculations assuming that a boundary layer is

present near x = 1.
Solution. We should now require the outer approximation to satisfy the
boundary condition at x = 0. The outer approximation still satisfies (23a) then,
but (23b) is replaced by Yo(O)= 0, giving Yo(x) = 0.
For x to be within the boundary layer, it is necessary that 1 - x be O(o) so that
the appropriate boundary layer variable is (1 - x)/o =[.
(Note that l is positive.)
Introducing t we find that
(40)
Since (40) differs from (26) only in a sign change, term balancing is accomplished
just as before and again o(e) = e. Neglecting the last term in (40) and requiring Yto
have the value unity at g= 0 (corresponding to x = 1), we find
yi(f) = C(e 2 i - 1) + 1.
To specify the constant C, we introduce the intermediate variable ii = (1 - x)/0,
where 0 is subject to the restrictions (33) as before. We then impose the matching
requirement (35) (with appropriate small modifications):
lim
10
Yr(ij~)
= lim yo(l -
0 10
0,j) = 0.
If e approaches zero through positive values, the limit of y 1 does not exist (as
expected), but if it approaches zero through negative values, the limit is - C + 1.
In the latter case C = 1, y 1(f) = exp (2t), and
yv(x) = 0 + e 2 <1 -">1 - 0 = e 211 -,,>1,
in agreement with (13).
Generalizing, we propose the following basic procedure for singular

perturbation boundary value problems for second order ordinary differential
equations. This procedure will not always work. However, it will work for a
large class of problems, and will work for a still larger class of problems with
modifications that should be within reach of the reader.
Consider the boundary value problem
,_ d
ey" + f(x, y, y') = 0, y(a) = A, y(b) = B; (41)
= dx'
Suppose (with no loss of generality) that e is positive. Suppose also that the
problem has a unique solution* with a boundary layer at x = a. This supposi-
tion does not commit us to whether the boundary layer is to the right or to
the left, for we have not specified which of a orb is the larger.
(a) Determine the outer approximation y 0 by solving
f(x, Yo , Yo) = 0, Yo(b) = B. (42)
(b) Introduce into (41) the boundary layer variable
x-a
e= 8(e) ' (43)
where the sign is chosen to make e

positive in the interior of the
interval over which the differential equation holds. (Note that
x = a be.) With the notation Y(e, e) = y(a be,e), (41) can be
written
2
ed Y ( 1 dY)
b2 de2 + f a be, Y, ~ de = 0. (44)
Suppose that for small e (and fixed e) the dominant contribution to

the second term in (44) has the form
(45)
The reader should be aware of the fact that even if (41) is linear, it may have an infinite
number of solutions or it may have no solutions at all. See, for example, Section l.S of I. Stak.gold's
Boundary Value Problems of Mathematical Physics (N.Y.: Macmillan, 1967).
for some function Fand some constants. [Note that F(e, Y, dY/de) is
independent of <>.JBalance the second derivative term with (45) by
choosing
or r,= 8 11<2+>. (46)
The problem must be such that s > - 2 if f>is to approach zero with e
as required.
(c) Determine y 1(e) by solving
2
d y1 ( dy 1 )
de2 + F e,Yr, de = 0, yi(O) = A. (47)
(d) To find the arbitrary constant in the solution to (47), introduce the
intermediate variable
x-a
1/ = 0(e)' (48)
where 0 is a function that is unspecified except for the requirements

. 0(e)
lim 0(e) =0 and hm i:() = oo. (49)
e!O t!O u ll
Impose the matchingrequirement

Jim [Yo(x)lx=ae~]= Iim [yi(e)l~=~e,a], 1ffixed. (50)
+O +0
REM ARK s . (a) Further restrictions on 0 may be imposed to obtain
matching. (b) The above procedure should not be slavishly plugged into"
but should rather be regarded as a summary and a guide. This will result
in a more efficient attack on easier problems and will increase the reader's
chance of developing the feeling" for singular perturbations which is
necessary for generalizing the method to harder problems. (c) Nonlinear
problems frequently require considerable adaptation of the above procedure.
Typically, the nonlinear reduced equation (42) will have more than one
solution. Different solutions serve as the outer approximation in different
regions. The choice of outer solution is governed by the possibility of matching
with "boundary layers" located near the boundaries and in the interior.
[See Cole (1968), Chap. II.]
Example 3. Find outer and inner approximations for small positive e to the
solution of
d 2y dy
e- 2 + cx(x)- + {J(x)y = O; y(O) = 0, y(l) = 1, cx(O)= a0 > 0, ,8(0) finite.
dx dx
To save trial and error, the information that the boundary layer is near x = 0 is
given.
Solution. The outer approximation satisfies cxy'+ f3y= 0, y(l) = 1, and so is

given by
Yo(x) = exp (J 1
ds].
/3(s)
" cx(s)
=
We introduce g x/S, where S(e)-+O as e i 0. We obtain
e d2 Y cx{Se)dY
S2 df2 + -S- df + /3(Se)y = O.
Since cx(Se}-+ao when eio for fixed f, the magnitudes of the various terms again
are as in (27); hence, by the same argument as was used there, S = e. The inner
approximation satisfiesd 2y,/de2 + aody,Jdf = 0, y,(0) = 0, so YzW = C(l - e-of).
Matching gives C = exp {J~[/3(s)/cx(s)]ds}.
Other interesting features of singular perturbation problems are illustrated

in the exercises, e.g., problems
(a) of higher order than second (Exercise 7),
(b) where it is not true that c5(e)= e [Exercise 2(c)], and
(c) where the outer solution in a second order equation does not satisfy
either boundary condition (Exercise 6).
As we have mentioned, singular perturbation theory is very much an art at
present. The procedures have been rigorously justified only on a fraction of
the problems upon which they have been used. As is illustrated in the books
of Van Dyke (1964), Cole (1968), and Nayfeh (1973), ingenuity in applying
singular perturbation techniques, particularly to nonlinear ordinary differen-
tial equations and to partial differential equations, leads to elegant and satisfy-
ing results. Traps threaten the unwary, but as has been remarked,* "the
inalienable right to think while using any technique provides a very real
degree of protection against the acceptance of ... spurious constructions."
EXERCISES
1. (a) Verify (15).
(b) Verify that the limit of ( 15) is uniform in [a, 1] as stated in the text.
(c) Show directly that the limit is not uniform in (0, l].
2. Find inner, outer, and uniform approximations to the solutions of the
following problems. Assume that e is small and positive and that there
is a boundary layer at x = 0. Proceed from first principles in each case;
do not "plug" into formulas.('= d/dx.)
(a) ey" + (1 + x)y' + y = O; y(O) = 0, y(l) = 1.
(b) ey" + y' + y 2 = O; y(O) = !, y(l) = ,t.
t(c) ey" + x 113y' + y = O; y(O) = 0, y(l) = e- 312
G. F. Carrier, "Singular Perturbation Theory and Geophysics," SIAM Rev. 12, 175-93
(1970).
Sec. 9.2] Boundary Value Problems for Ordinary Differential Equations
299
3. Show that contradictions result in the problems of Exercise 2 if the

boundary layer is assumed to be near x = 1.
4. A " small" mass m hangs from a weightless spring with internal damping
proportional to speed. A vertical impulse I (instantaneous momentum
change) is imparted to the mass by striking it with a hammer. Initial
conditions on the vertical deflection y* at time t* can thus be taken as
dy*
at t* = 0: y* = 0, m dt* = I.
The governing equation is
m(::~:)
+(it:)+ ky* = 0,
where and k are the damping and spring constants, respectively. Since
the mass is small, we have a strongly "overdamped" situation wherein
the mass will quickly return to rest after the impulse is expended in
stretching the spring. This exercise requests a perturbation analysis of
the situation.
(a) Show that a certain choice of dimensionless variables reduces the
problem to
ey" + y' +y = 0, y(O) = 0, ey'(O) = 1. (51)
(b) Solve the problem by singular perturbation techniques, using the
following hints. Do not impose an initial condition on the outer
approximation. Find an inner approximation that satisfies both
initial conditions. Then complete determination of the outer
approximation.
;(c) Is your answer plausible? Discuss briefly. In particular, explain
why (51) is correctly scaled, provided that tis not too small.
(d) Find a composite solution valid for the entire range t > 0.
(e) Check your answer by examining the exact solution when e ~ 1.
t5. Use singular perturbation theory to obtain outer, inner, and composite
expansions to the solution of the problem (Carrier, op. cit.)
eu" - (2 - x 2)u = -1, u(-1) = u(l) = 0.
R EM AR K It is sufficient to solve the differential equation on (0, I)
subject to the boundary conditions u'(O) = 0, u(l) = 0. Why?
6. (a) Use singular perturbation techniques to approximate the solution
to
e2y" -y = 0, y(O) = 1, y(l) = 2, 0 < e ~ 1.
(b) Show that singular perturbation techniques fail on the problem
e2y" +y = O; y(O) = 1, y(l) = 2, 0 < e ~ 1.
Use the exact solution to show why things go wrong.
7. For a slightly stiff string with fixed ends, the modes of vibration have a
shape y(x) given by the eigenvalue problem
ey<iv)-y" = .ity, y(O) = y'(O) = y(l) = y'(l) = 0.
Here e is a measure of the stiffness; 0 < t ~ 1. The eigenvalue .it is a
dimensionless frequency of vibration. Find outer and inner approxima-
tions, assuming that .itis 0(1),and conclude that the stiffness has no effect
on the eigenvalues to lowest order. [For a discussion of corrections due
to stiffness see Sec. 3 of R. E. O'Malley's ''Topics in Singular Pertur-
bations," Advan. Math. 2, 365-40 (1968). Reprinted in Lectures on
Ordinary DifferentialEquations (New York: Academic, 1970).]
is. Find the exact solution to
y"-t 2
y=0; y(O) = 1, y'(x) -o as x-+ oo,
and take its limit as e ! 0. Show that there are three different types of
limits, depending on the range of x under consideration. Thereby
demonstrate that this problem has a singular perturbation character for
small e in spite of the fact that the order of the equation does not change
when e is set equal to zero.*
9. The object of this exercise is to show how the "unacceptable" dis-
continuity that appeared in (8.3.9) can be "fixed" with the aid of
singular perturbation theory.
(a) In (8.4.5), (8.4.6), and (8.4.7), reintroduce the parameter fl, using
K = .il(flv)-112. Show that when flV~ 1,
c- 1 {~ 1 - te<x-1)/('lv)l/2' x s; 1,
..., .1. -(x-1)/('lv)l/2
v ...,-ze , x;;:: 1.
Sketch. Also sketch the discontinuous function
C-1
lim --.
v-+O V
., fixed
(b)It should now be apparent that (8.3.6) provides an outer expansion

for the problem of (8.2.20). Pretend that you do not know the
results of part (a) and "fix" the discontinuity obtained in (8.3.9)
with a boundary layer.
10. In studying viscous fluid flow past an infinite plane, one is led to the
following problem for y(x; e) on Os; x < oo, 0 < e:
e(y"' + yy") + 1 - (y') 2

= 0, y(O) = y'(O) = 0, lim y'(x) = 1.
x-+c:o
This exercise is taken from Sec. 1 of the O'Malley paper cited in Exercise 7.
Sec. 9.2] Boundary Value Problems for Ordinary Differential Equations
301
We are interested in the case e ! 0, and you may assume that there is a
boundary layer at x = 0. Determine the first term in the outer solution.
Also determine the boundary layer equation, boundary conditions,
and matching conditions. Do not solve. [To do this, you will have to
introduce boundary layer variables for both x and y, and also make use
of the requirement that in passing from the inner to the outer region
y'(x) must be continuous as e!O.]
CHAPTER 10
SingularPerturbationTheory Appliedto
a Problemin BiochemicalKinetics
As AN example of a relatively simple but important application of the

techniques introduced in Chapter 9, we shall now examine a problem in
biochemical kinetics. We shall be led to consider singular perturbation
theory for an initial value problem involving a certain system of ordinary
differential equations. Some details of the discussion of boundary value
problems in Chapter 9 are inapplicable, but the arguments are the same in
spirit. The reader's understanding should thus be reinforced.
In Section 10.1 we formulate a mathematical problem that describes the
course of a chemical reaction catalyzed by an enzyme. We shall consider the
frequently occurring case when the initial enzyme concentration is small
(in a sense to be made more precise). The problem can then be profitably
analyzed with the techniques of singular perturbation theory. This we do
in Section 10.2. In particular, we make a technical advance over Chapter 9
by indicating how higher approximations can be obtained.
10.1 Formulationof an Initial Valoe Problemfor a

One Enzyme-One SubstrateChemicalReaction
THE LAW OF MASS ACTION
We shall be mainly interested in the time course of reactions that are cata-
lyzed by enzymes, but first we must consider some simple cases of the law
of massaction, which forms the foundation for quantitative studies of chemical
reactions.
For definiteness let us start by considering a situation wherein molecules
a and b combine to give molecule c. If a and b are to react, they must undergo
an effective collision. Not every collision may be effective because juxta-
position of active or important parts of the molecules is probably required
for the reaction (Figure IO.I).
If we hold external factors (such as temperature) fixed, and also keep the
concentration of b fixed, then it is natural to suppose that the frequency
of effectivecollisions (and hence the rate of reaction) is proportional to the
concentration of a. For if we double the amount of a involved, there should
be twice as many effective collisions per unit time. (This assumes that the
concentration of a is not too high; otherwise doubling this concentration
302
Sec. 10.1] Formulation of an Initial Value Problem 303
(a) (b)
FIGURE 10.1. Schematic illustration of molecular collisions. The heavily

bordered portions of the "molecules" Mand N must join for a successful
reaction. (a) An effective collision. (b) An ineffective collision.
may lead to interference effects.) Using the same symbols for the chemicals
and for their concentrations (number of molecules per unit volume), we
may write dc/dt,.., a. By the same token, dc/dt,.., b. This leads us to the
equation
de
dt =k1ah,
where k 1 is the rate constant.Chemists use the symbolism

k1
a+b - c
to denote the reaction in question. As should be apparent, the equations
governing the concentrations of a and b in this case are
da
- = -k 1ab
dt '
An important special case is the situation where a and b are identical,

and combine to form the two element chain or dimera2 We then have
k, da2 2
a+a - a2, dt=k 1a.
This illustrates the fact that when just two molecules of a substance are
involved in a reaction, the rate is proportional to the square of the corres-
ponding concentration.
The above discussion, while plausible, is speculative. The acceptability
of the ideas we have put forth dates from 1867 when their validity was
experimentallyestablished by the Norwegians Guldberg and Waage.
It is only a slight extension of our basic ideas to consider a reversible
reaction
a+b (I)
Singular Perturbation Theory Applied to Biochemical Kinetics [Ch.JO
304
wherein the chemical c may spontaneously break up into its original com-
ponents a and b. The reader should satisfy himself that the governing equations
now are
From these equations we can at once deduce that

d(a + c) = 0
d(b + c) = 0
dt ' dt '
so that a+ c = constant and b + c = constant. This is to be expected,
because each new molecule of c results from a combination of a and b,
and hence from a disappearance of a molecule of a and a molecule of b.
ENZYME CATALYSIS
Certain chemicals have been found to be of vital importance in speeding
up and controlling biochemical processes. These enzymestare now recognized
to be large protein molecules that somehow can catalyze chemical reactions,
most often a specific reaction for each enzyme. The precise details of enzyme
action remain a topic of intensive research. But, as was suggested in the late
nineteenth century by the Swedish chemist Svante Arrhenius, the essence
of the matter is that the enzyme (concentration e*) combines with the
"input" molecule or substrate (concentrations*). This produces a combined
enzyme-substrate molecule, or complex (concentration c*). Bound to the
enzyme molecule in this complex, the substrate molecule is "activated"
and far more likely to form the "output" or product (concentration p*)
of the reaction. When this happens (in the simple case under consideration),
it is presumed that the complex breaks up into the product molecule and the
original "free" enzyme molecule. One must also allow for the possibility
that the complex breaks apart without the reaction having occurred. Schema-
tically, then, one can write
k, k2
s* + e* k-1
c* ---+ p*+e*. (2)
Let us assume that originally only the substrate and enzyme are present,
at concentrations s and e, respectively. Then the concentration of chemicals
at time t* is determined by the following equations and initial conditions:
ds*
- = -k 1s*e* + k -1 c*'
dt* (3a)
de*
-
dt* = -k 1s*e* + k -1 c* + k 2 c*, (3b)
t A classical reference, which includes theoretical discussions, is The Enzymes by M. Dixon

and E. C. Webb (New York: Academic, 2nd ed., 1964).
Sec. JO.I] Formulation of an initial Value Problem
305
de*
-
dt* = k 1s*e* - k -1 e* - k 2 e*, (3c)
dp*
dt* = k2e*, (3d)
at t* =0: s* = s, e* = e, e* = 0, p* = 0. (3e)
Adding (3b) and (3c), and using (3e), we conclude that
e*(t*) + e*(t*) =constant= e. (4)
We can thus eliminate e* and consider the problem
ds*
dt* = -k 1es* + (k 1s* + k_ 1)e*, (Sa)
de* - * - (k 1S * + k -1
dt* = k 1eS + k 2 )C*' (Sb)
s*(O)= s, c*(O)= 0. (Sc)

Once s*(t*) and c*(t*) have been determined by the above equations, e*(t*)
and p*(t*) can be determined from the easily deduced equation
s* + c* + p* = s. (6)
Setting up the equations for the time course or kinetics of the reaction is
not an idle exercise, for one cannot merely take a snapshot of the molecular
interaction to check on whether or not Arrhenius' speculations are correct.
But experimenters have been able to follow the time course of reactions, for
example by taking advantage of certain changes in light absorption properties.
Comparison of such observations with theoretical kinetic results form the
classical evidence for the existence of the enzyme-complex mechanism.
Certain simplifying assumptions have enabled clear comparison with
experiment. A number of different people are given credit here. The names
most often mentioned are those of the Berlin scientists Leonor Michaelis
and Maude Menten, but others prefer to cite the Frenchman Henri or the
Englishmen Briggs and Haldane.t Be that as it may, the main idea is as
follows.
It often occurs, or can be arranged, that the initial concentration of sub-
strate is much larger than that of enzyme. (" Concentration" means number
of molecules per unit volume.) In such a case there will be a short initial
t Later analyses from a singular perturbation point of view oc:c:urin articles by J. R. Bowen,
A. Acrivos, and A. K. Oppenheim, "Singular Perturbation Refinement to Quasi-steady State
Approximation in Chemical Kinetics," Chem. Eng. Sci. 18, 177-88 (1963); and F. G. Heineken,
H. M. Tsuchiya, and R. Aris, "On the Mathematical Status of the Pseudo-steady State Hypoth-
esis of Biochemical Kinetics," Math. Biosci. 1, 95-113 (1967).
306 Singular Perturbation Theory Applied to Biochemical Kinetics [Ch. JO
period when the enzyme will quickly "load up" with substrate so that the
concentration of complex increases rapidly. During most of the reaction time,
however, the concentration of complex will remain approximately constant.
The reason for this is that as long as there are many substrate molecules
around for each enzyme molecule, the speed of the reaction will only be limited
by how fast the enzyme can work, not by the availability of substrate to work
on. The time for an enzyme molecule that becomes "free" to bind to a new
substrate molecule is always about the same. Of course, when most of the
substrate is used up, this time will become longer and longer. In the final
stage of the reaction, therefore, complex concentration should noticeably
decrease. Ultimately, all the substrate will have been converted to product by
reaction (2), the enzyme will be entirely in its " free" form once again, and no
complex will be present.
In comparing their observations with theory, biochemists therefore make
the hypothesis that complex concentration can be regarded as constant to
first approximation. Equation (Sb) with dc*/dt* = 0 can then be used to
express c* in terms of s*. Upon substitution of this result into (Sa), the prob-
lem reduces to a single ordinary differential equation for ds*/dt* in terms of
s*. This gives the "Michaelis-Menten kinetics." In this chapter we shall
give a detailed treatment of the problem by singular perturbation methods.
In particular, we shall show that the Michaelis-Menten kinetic equation
corresponds to the "outer" equation in the terminology of the previous
chapter. Our analysis follows rather closely that of Heineken et al. (op. cit.).
SCALING AND FINAL FORMULATION
The first step is to introduce dimensionless variables in such a way that the
maximum magnitude of each term is correctly estimated by the parameter
that precedes it (scaling). In doing this, we focus our attention on the situation
after the initial "loading-up" phase and before the final stage. For the
period in question, sis an excellent estimate for s*, since our entire approach
is based on the fact that s* will not decrease much from its initial value s.
Further, in the presence of an abundance of substrate, most of the enzyme
initially present should soon join with substrate to form complex. We thus
take eas an estimate for c*. The choice of time scale is perhaps not immediately
obvious; therefore, let us simply denote it by i. We thus introduce dimension-
less substrate and complex concentrations and a dimensionless time by
s* c* t*
S=-, c = -::-, t=- (7)
s e t
Upon substituting into (Sa), we obtain

sds
-kiess(l - c) + L ec. (8)
1dt = 1
Sec. JO.I] Formulation of an Initial Value Problem 307
The first term on the right side of (8) gives the rate of decrease of substrate
concentration caused by conversion into complex. The second term on the
right gives the rate of increase of substrate concentration caused by spontane-
ous breakdown of complex. If the latter effect were large, there would be
considerable formation of complex which splits apart without accomplishing
anything. Such dominance of the "back reaction" apparently occurs on
occasion and performs a controlling function. But we shall consider more
usual circumstances in which there is a rough balance between the terms
(s/i) ds/dt and -k 1 ess(l - c) of (8). This requires that the time scale be
- I
t=- (9)
k 1e
Introducing (7) and (9) into (5), we see that the dimensionless scaled equations
are
s = -s + (s + 1e- l)c, (IOa)

ec= s - (s + 1e)c, (=!) (IOb)
s(O)= 1, c(O)= 0, (lOc)
where
e k_1 + k2 A=-,
k2
t = ::, K=
s k 1s k1s
are parameters.
The boxed equations (10) comprise the mathematical formulation of our
problem. We are faced with an initial value problem for a pair of first order
nonlinear ordinary differential equations. Three parameters occur, of which
one is small. If the small parameter is set equal to zero, we can reduce the prob-
lem to a single first order differential equation, but now at least one of the
initial conditions cannot be satisfied. Singular perturbation methods seem
called for. These are the subject of the next section.
EXERCISES
1. Verify (4), (5), and (6).

2. Suppose that the second reaction in (2) is reversible, with rate constant
k-2
(a) How is (3) changed?
(b) How are (4), (5), and (6) changed?
(c) Begin to trace the consequences of this reversibility in the remaining
calculations of this chapter.
308 Singular Perturbation Theory Applied to Biochemical Kinetics [Ch.JO
10.2 Approximate Solution by Singular Perturbation

Methods
In this section we use singular perturbation methods to obtain an approxi-
mate solution to the problem posed in (1.10).
MICHAELIS-MENTEN KINETICS AS AN OUTER SOLUTION
In (1.10) the dimensionless parameter e is small by hypothesis. Since we
have introduced scaled dimensionless variables, when the scaling is valid
the term multiplied bye may be neglected, yielding the following equations
for the approximations s0 and c0 :
so= -so+ (so+ JC -A)c 0, (la)
So
s0 - (s0 + JC)c0 = 0 or Co=-- (lb)
So +K
Substituting (lb) into (la), we obtain the single first order equation
A.So
So=---. (2)
So+ 'IC
We shall refer to the solutions of (I) and (2) as the Michaelis-Menten

approximation. With this approximation, data from chemical experiments
can be organized to permit ready determination of the parameters K and A
(Exercise 8).
It should be evident that the Michaelis-Menten approximation can be
regarded as the outer approximation of a singular perturbation problem.
The first order equation (2) cannot be expected to have a solution that satisfies
both initial conditions of (l.lOc). This is no surprise, for we have reduced
the order of the system by neglecting ec.Furthermore, our scaling was made for
a period after the initial moments when the complex concentration builds up
rapidly from zero. Solving (2), we find easily [Exercise l(a)] that
s0 + 'ICIn s0 = -At+ Q, (3)
where Q is a constant. This equation and (lb) provide the outer solution,
where Q must be determined by matching.
INNER SOLUTION
To find the inner equations, we postulate the existence of an inner or
initial layer of width" c5(e).On introducing the appropriate change of vari-
able t = t/b(e), we see that the inner solutions S(t, e) = s(c5t, e) and C{t, e) =
c(bt, e) satisfy
dS
r 1
dt = -S + (S + " - A)C, (4)
Sec. 10.2] Approximate Solution by Singular Perturbation Methods
In order to retain the term involving dCfdr:, we must take

a(s) = 8 so T=t/s.* (5)
Thus the inner equations are
S' = s[ - S + (S + 1e- .t)C], C' =S - (S + 1e)C, (6)
Because we want the inner solutions to be appropriate to the first few mo-
ments of the reaction, the original initial conditions must be satisfied. Thus
S(O,s) = l, C(O,s) = 0. (7)
Let S0 (r:) and Co(r:) denote the first approximation to the inner solutions.
From (6) and (7) these satisfy
S 0= 0, C0 = So - (So+ 1e)Co; S 0 (0) = I, Co(O)= 0.
Hence [Exercise l(b)]
S0 ('r) = l, (8a, b)
and the inner approximations are completely determined.

To match, we introduce an intermediate variable T; such that
. 'l'(s)
lim 'l'(s) = 0, lim ~() =co, (9a, b, c)
e.j.O e.j.O u 8
in parallel with (9.2.49). Recall from (5) that a(s) = s. (Note also that
-. = T;'fl/a.) In parallel with (9.2.50), the matching condition on the sub-
strate concentration is that, for fixed;,
or s0 (0) = 1, (10)
since S 0 = 1. Hence Q = I in (3), so that

s0 + 1' In s0 = -).t + 1. (11)
This completesthe (implicit)determinationof the outer approximationto the

substrate concentration.
The corresponding matching condition on the concentration of the com-
plex requires that
c0 (0) = lim {(,c+ 1)- 1 [1 - = (,c + 1)- 1 ,
e-<"+ 1 >,'l'/cl]} (12)
e.j.O
We can regard the introduction of.,. rather than t as a required rescaling in an initial layer
where, far from being negligible compared to s and (s + K)c, c is comparable to these terms as
complex concentration rapidly builds. Here, then, is another instance of an unorthodox function
whose proper description requires the introduction of more than one scale. (Compare Section 6.3.)
310 Singular Perturbation Theory Applied to Biochemical Kinetics [Ch.JO
using (Sb). There are no remaining constants which can be chosen so that
this condition will be satisfied. From (lb) and (10), however, we have
cO(O) = s0 (0) = _I_ (13)

s0(0) + K 1+ K
so that the condition is automatically satisfied. This state of affairs is not a
coincidence. It comes about because of our correct approach to the problem.
A UNIFORM APPROXIMATION
The procedure for obtaining a uniform approximation is always the same:
add the inner and outer approximationsand subtract their commonpart. The
justification given for this procedure after its first introduction in (9.2.39),
is not restricted to the particular context in which it appeared. In the present
case the initial term s&">in the uniform approximation sM to s is given by
s~">(t)= so(t) + so(;)

- 1 or (14)
Here we have used (10) and the fact that S 0 = 1. Correspondingly, for the
complex concentration we have, using (12),
so cM(t) = so(t) (15)

O So(t) + K
All our various results are expressed in terms of so(t), which is given
implicitly by the transcendental equation (11). For fixed values of the para-
meters K and J, s0 can be plotted as a function of t using various numerical
methods [e.g., see Ralston (1965)]. Furthermore, as we shall now show,
useful analytic approximations can be obtained for small and large values
oft.
For small t a power series expansion is appropriate. To obtain this ex-
pansion, we require the derivatives of s0 (t) evaluated at t = 0. But the initial
condition on s0 is s 0 (0) = 1. Also, from the differential equation (l.lOa) and
the initial condition (13), we find at once that
l+K-A
s0(0) = -1 +
1 +K
= -(K + 1)- 1i. (16)
Higher order terms can be found either by successively differentiating

(l.lOa) or by substituting a power series into (11) and equating coefficients
(Exercise 2).
For our purposes it will be sufficient to record the result
s0 (t) = 1 - (K + l)- 1A.t+ O(t 2 ). (17)
When t ~ A.- 1, (11) can be approximated by
s0 + K ln s0 R: -A.t. (18)
311
Since s0 is positive, the right side equals a large negative number when t is
large. The only way the left side of (18) can yield such a number is for s 0
to approach zero as t --1- oo.* When s0 is small and positive, we have IIn s 0 I ~
s0 Hence (18) can be approximated by" lns 0 ::::: -At, giving
-At)
s 0 (t) :=:::exp (7 , t large. (19)
Using the above result, our approximation !In s 0 l ~ s 0 seems consistent

when At/,c ~ exp (-Atf,c). Since exp (-2)::::: 0.14, we can probably regard
this requirement as fulfilled when ).t/,c > 2. Another requirement, made at
the beginning, is that t ~ ). - l. Since ). and " are typically of order unity,
we can thus expect the exponential substrate decay of (19) after a few units
of dimensionless time t, i.e., when the dimensional time t* is a few multiples
of(k 1e)- 1 .
COMMENTS ON RESULTS SO FAR

Until fairly recently, biochemists have been content with an informal
derivation of the Michaelis-Mentert relation (lb) between complex and sub-
strate. Interpreted as in Exercise 8, this relation has proved very useful in
understanding the action of enzymes in biological systems. As one would
expect, however, eventually problems arise in which deeper analytically based
understanding is required in order that the fundamental ideas can be applied
to more complicated situations. One example that shows this is some work of
S. Rubinow and J. Lebowitzt concerning the interaction with a substrate
and an inhibitor of the enzyme L-asparagine amidohydrolase. (This chemical
shows promise in cancer therapy.)
From our detailed analysis we can obtain a quantitative estimate of how
long an initial interval must pass before the Michaelis-Menten approximation
can be used. From (15) we see that the approximation s 0 /(s 0 + ,c) to c0
[see ( 1b)] will be accurate only after an initial interval of dimensionless
duration equal to a few multiples of s(,c + 1)- 1 . During this interval, the
complex concentration c&"Xt) increases rapidly from its initial value of zero
to a value approximately equal to (" + 1)- 1 In practice the duration of the
initial interval is typically a second or less.
HIGHER APPROXIMATIONS
Until now we have found only the simplest kinds of inner, outer, and
uniform approximations to the solutions of singular perturbation problems.
It is possible to improve on these approximations in a systematic manner,
as we now illustrate. The main idea is to posit series expansions for the
We certainly expect on physical grounds that the substrate will all be used up eventually.
t "Time-dependent Michaelis-Menten Kinetics for an Enzyme-Inhibitor-Substrate System,"
J. Amer. Chem. Soc. 92, 3888-3893 (1970).
Singular Perturbation Theory Applied to Biochemical Kinetics [Ch.JO
312
solutions to the inner and outer equations. (The approximations obtained

thus far are the first terms in these expansions.) Additional conditions neces-
sary for the unique specification of higher approximations can be obtained
by introducing an intermediate variable as before, but requiring a more accur-
ate match.
To illustrate the procedure for finding higher approximations, we return
to the enzyme kinetics problem. We now deal with the inner and outer
equations by assuming series expansions. For the inner equations (6) this
requires the assumption
S(-r, e) = S0 (-r)+ eS1(-r)+ , C(-r,e) = Co(-r)+ eC1 (-r)+ . (20)
Upon the usual substitution and collection of terms (Exercise 3), we find,
of course, that the lowest approximations S 0 and C0 are given as before by (8).
The next higher order approximations satisfy
S~ = -S 0 +(IC-A+ S0)C 0 , C~ =-(IC+ S0)C 1 + (1- C0)S 1. (21)
Upon substitution of the series into the boundary conditions (7), we find that
the higher approximations satisfy the homogeneous boundary conditions
S1 (0) =0, C1 (0) =0. (22)
By elementary but somewhat lengthy calculations, the solutions of (21) and
(22) are found to be
S1 = -(IC+ 1)- 2 [).-r
1 + (1 + IC- )..)(1- e-' )], -r1 = (K + 1)-r,
1
4
C1 = -(IC+ 1)- {AICT 2 1
1 + 1C(l+ IC- 2)..)+ (1 + IC- )..)e- ' (23)
2
+ [ f).-r~+ (1 + IC- J)(l - 1C)T1- (1 + 21C+ K - ).. - 2,c)..)]e-' }.
1
For the outer solutions we also assume expansions in powers of e:

s(t, e) = s0 (t) + es1(t) + , c(t, e) = c0 (t) + ec1(t) + . (24)
Note that here the independent variable is the outer variable t. In (20) we
used the inner variable -r. Again, the lowest order approximations s 0 and c0
are given, as before, by (lb) and (11). Exercise 4(a) shows that the next
approximations satisfy the equations
We lack initial conditions. These must be determined by matching.

For higher order matching, the reader will do well if he introduces the
intermediate variable and matches "as accurately as is reasonable." To know
what is "reasonable" requires experience (some of which will now be pro-
vided) and varying degrees of ingenuity.
In the present initial value problem the inner solution is completely
determined. Let us construct an intermediate expansion for the inner substrate
concentration. To do this, we express the inner variable in terms of the inter-
Sec. 10.2] Approximate Solution by Singular Perturbation Methods 313
mediate variable -r; = t/'P of (9). We find from the expressions for S 0(-r)
and S 1 ( -r)in (8) and (23) that
-
')+e s ('r;eq,)+
S(-'r;eq,)= so('r;
e
1 -
= 1 - e(K + 1)-
2
+ l)C;e'P) + (1 +" -
[ A.(K A.!+
TST] +
=1- (K + 1)- 1A.-r;' - (K + 1)- 2 (1 + " - l)e + TST + . (26)
Following Cole (1968), we use the abbreviation TST for transcendentally
small terms. In S1(-r;'/e) such a term is
which decreases exponentially to zero as e ! 0 with -r; fixed, since ' /e -+ + oo

by (9c). This is to be contrasted with the algebraic decrease of a term such as
-(K + 1)- 1M; ', which is explicitly written in (26).
The above expression must be matched with the outer solution s(t) when
t is expressed in terms of the intermediate variable -r;. But how can the term
proportional to ' in (26) be matched ? With the insight provided by the early
masters of singular perturbation theory, we observe that at lowest order a
match was afforded by proper choice of s0 (0), the "innermost" value of
the outer approximation s0 (t). Now we must examine this approximation
more closely near t = 0. In fact we utilize not just s0 (0) but the first two
terms of the expansion of s0 (t) about t = 0. From (17)
so(t) = 1 -(K + l)- 1A.t+ O(t 2 ) ' + 0(' 2).
so s0(?; ') = 1 - (K + l)- 1A.-r;
Upon introduction of the intermediate variable, the O(t) term in so(t) pro-
vides just what is necessary to match the 0(') term in (26).
Not only must our more accurate match take account of two terms in
s 0 , but it must also introduce one term from s 1 Thus we should match (26)
with
so(-r,') + u 1(-r1') + O(e2) = 1- (K + l)- 1
'
A.-r; + 0(' 2)
+ es1(0) + O(e'I') + O(e2). (27)
This requires that
(28)
which is the desired initial condition for s 1

Let us discuss the matching a little further. We ask: What will match the
0(' 2) term in (27)--which came from the O(t 2) term in so(t)? The matching
term must come from the neglected e2 S2 term in (26). Just as a term pro-
portional to e-rin eS1(-r)gave rise to an 0(') term in the intermediate region,
so we expect a term proportional to e2 r 2 in e2 Sir) to give rise to an 0('1' 2 )

term in the intermediate region. In a very satisfactory way, then, the leading
terms that are neglected when the approximation is carried to a certain
point will be taken into consideration when the approximation is carried
one more step.
Having made the specification (28), we expect from (26) and (27) that
[S 0 (r) + eS 1(r)],=,,.,,. - [so(O) + ts 0(0) + es1(0)]r='1',

= 0('1' 2) + O(e'P) + O(e2).
We see in retrospect that we can impose the two-term matching condition
Equation (29) holds if '1' 2 /e-+0 as e ! 0. This is a restriction on 'I' in addition

to the earlier requirement that 'I' /e-+ oo as e ! 0. It means that our matching
is imposed in a layer of O('I') thickness, where e ~ 'I' ~ Je.
Example. By appropriate matching, find the initial condition that c 1(t) must
satisfy. Verify that this condition is implied by (28).
Solution. Using (Sb) and (23), we find that
-(K + l)- 4K(l + K - 2,\)e+TST.

From (lb) and (17)
So(t) 1 - (K + 1
n-
,\t + O(t 2)
Co(t) = s 0(t) + K = K +l - (K + J)- 1,\t + O(t 2)
= (K + 0- 1
- (K + l}- 3 ,\Kt + O(t 2 }
so that
[co(t) + eC1(t)]ra11'1'= (K + 1)- 1 -(K + l)- 3,\K'T1'f" + 0(' 2) + ec1(0) + O(e'lj.
Matching requires the initial condition

(30)
But from (25b)
co+s1Co-S1 (31)
C1= -
K+so
Differentiating (lb), we find that

Co= KSo(So + K)- 2 (32)
Using this, (2), and (28), we recover (30).
315
We can now complete the determination of the second approximation.
From (25b) we can derive (31), which, upon substitution into (25a), gives
an equation for s1(t). After some computation (Exercise 5), it is seen that
this equation, together with initial condition (28), has the solution
So [1C - ..1.I ( 1C+ So ) +

1C- ..1. So]
(33)
S1 = So+ " -1C- n (1 + K)so - +
1C So
so that from (31) and (32)
C1 =( 1 )3 [2"so..l. - 'KSo+ so('K- )..)In (c1'K+ S)o)] (34)

'K + So 1C+ So + 'K So
As in the case of the lowest approximation, the uniform solution s(u)equals
the inner solution plus the outer solution minus the common part, but now
the solutions are taken to higher order. For example, let sf>be the approxi-
mation to s(u)in which O(s) terms are retained. Then
s\u>(t,s) = [so(t) + es1(t)] + [ So (;) + tS1 (D]

- [l - (1e + 1)- 111- t(K + 1)- 2(1 + " - 1)]
) sso(t) 1 [" - ).. ( " + s0 (t) ) " - )..+ s0 (t)]

= So(t + s0(t) + 1e -"- n ,(" + l)s 0 (t) - "+ so(t)
s(l + 1C_ ..l.)e-<K+l)r/
+ (1C + 1)2 . (35)
FURTHER ANALYSIS FOR LARGE TIMES

We have investigated a problem in enzyme kinetics by developing and
explaining principles that are of wide application in singular perturbation
theory. The present problem, however, is one of the widening (but by no means
all-inclusive) class that can be treated by a rigorously based "prescription."
As applied by Heineken et al. (op. cit.), the prescription comes from work by
the Soviet mathematician A. B. Vasil'eva.* This work proves, for example,
that the above expression for s~">(t, s) is an asymptotic approximation to the
exact solution s(e)(t, s) in the sense that
when O ~ t ~ T, (36)
for sufficient small e. The quantity a is independent of s and t but it may
depend on the constant T. In other words the error is uniformly O(s2)
throughout a time interval of fixed length T.
In contemplating a result such as the one just stated, one naturally is
curious as to whether the restriction to a fixed time interval is made because
* Russ. Math. Surv. 18, 13 (1963).

316 Singular Perturbation Theory Applied to Biochemical Kinetics [Ch. 10
of technical difficulties in constructing the proof or whether the restriction

is required by the nature of the problem. That the former is likely is seen
by calculating the ratio es1(t)/s 0 (t), using the "large time" approximation
s0 {t) ~ exp (-J..t/K)of (19).We find
s 1(t) ~ ! [" - J..ln " _" - J..]~ J..(K- l)t. (37)
s0 (t) " " (1 + ,c) exp ( -kr/K) " "3
For our approximation to be apparently consistent (in the terminology
of Section 6.1), the correction es1(t) should be small compared with s0 (t).
From the above calculation we see that this is the case (if).. K) unless et =
0(1), t = 0(1/e). But when t = 0(1/e) then, by (19), s0 ~ exp ( - l/,ce), a trans-
cendentally small quantity. Thus the inconsistency affects the solution only
when it is completely negligible. Indeed, direct calculation of the solution for
very large times produces results that are the same as those of (19) and
(lb). (See Exercise 7.)
Our remarks lead to the conjecture that inequality (36) holds for all non-
negative t. Proofs of this type of conjecture have in fact been given. [See
F. C. Hoppensteadt, "Singular Perturbations on the Infinite Interval,"
Trans. Amer. Math. Soc. 123, 521-35 (1966).]
0.95
Zeroth order approximation s'\r
0.90
0.85
0.80 First order

approximation st>
0.75 O
0.2 0.4 0.6 0.8 1.0
FIGURE 10.2. Approximations to the dimensionless substrate con-

centration when e = 0.1. The "exact" solution was obtained numerically.
[Redrawn with permission, from F. G. Heineken, H. M. Tsuchiya, and
R. Aris," On the Mathematical Status of the Pseudo-steady State Hypo-
thesis of Biochemical Kinetics," Mathematical Biosciences 1, 95-113
(1967).]
Sec. 10.2) Approximate Solution by Singular Perturbation Methods 317
FURTHER DISCUSSION OF THE APPROXIMATE SOLUTIONS

To gauge the accuracy of the approximations they obtained, Heineken
et al. (op. cit.) found by numerical calculations some "exact" solutions to
the governing equations (1.10). Employing data relevant to the hydrolysis
of benzoyl-L-arginine ethyl ester catalyzed by the enzyme trypsin, they took
JC= 1, l = 0.375. Their results are shown in Figures 10.2-10.4. These figures
show excellent agreement with e = 0.1, "a value that the biochemist would
regard as quite large," and good agreement even when B = 1.0.
The more analytic information that is available concerning the time course
of the chemical reaction, the easier it is to deduce the values of the various
parameters from concentration measurements. Analytic solutions lend
themselves readily to investigations of parameter variation. But perhaps the
largest recompense for an investigation such as the one just concluded is the
thorough understanding it gives of the scientific phenomenon.
From the point of view of presenting important ideas in applied mathe-
matics, the present example has served to illustrate the technique for finding
higher order approximations in singular perturbation theory. The reader
Zeroth order
0.5 approximation cJu>
Exact solution
0.4 and first order
approximation cl">
0.3
c
0.2
0.1
OL-~......L.~~-'-~--'~~......1...~~"-~--'
0 0.2 0.4 0.6 0.8 1.0
FIGURE 10.3. Approximations to the dimensionless complex concen-

tration when e = 0.1. [Redrawn, with permission, from F. G. Heineken,
H. M. Tsuchiya, and R. Aris, "On the Mathematical Status of the
Pseudo-steady State Hypothesis of Biochemical Kinetics," Mathematical
Biosciences 1, 95-113 (1967).]
1.0 .------------------,
0.9
0.8
Zeroth order approximations:>

0.7
s 0.6
0.5
0.4
0.3 First order

approximation s f"'
0.2 .___...._ _ __.__ __J.___.....__ ..........
_ __._ ___,
0 2 3 4 5 6
FIGURE 10.4. Approximations to the dimensionless substrate con-

centration whene = 1.0. [Redrawn, withpermission,from F. G. Heineken,
H. M. Tsuchiya, and R. Aris, .. On the Mathematical Status of the
Pseudo-steady State Hypothesis of Biochemical Kinetics," Mathe-
matical Biosciences 1, 95-113 (1967).]
should obtain a good initial grasp of this matter if he fills in the details of
the present discussion, particularly as in Exercises 3, 4, and 6, and also com-
pletes Exercise 9, which involves higher approximations for a boundary
value problem.
EXERCISES
1. (a) Verify (3).
(b) Verify that (8) gives the inner approximations S0 and C0
2. (a) Find .s0 (0) and s0 (0) by successivelydifferentiating (l.lOa).
(b) Assume that s0 (t) = 1 + u 1 t + u 2 t 2 + u 3 t 3 + ; find u 1 , u 2 , and
u 3 by substituting into ( 11) and collecting terms. Check your answer
by comparing with part (a).
Sec. 10.2) Approximate Solution by Singular Perturbation Methods
319
3. (a) Show that S1 and C1 , the O(e) coefficients in the inner expansions,
satisfy (21) and (22).
(b) Find the equations that are satisfied by S2 and C 2
t(c) Obtain S 1 and C1 , as in (23).
4. (a) Show that s 1 and c1 , the O(e) coefficients in the outer expansions,
satisfy (25).
(b) Find the equations that are satisfied by s2 and c2
S. t(a) Use (31) to obtain a nonhomogeneous first order differential
equation for s 1 . Show that this equation and the initial condition
(28) has the solution (33).
(b) Find the solution for c1 , and thereby verify (34).
6. Find in a manner similar to that used in (35), the uniform approximation
c~u>.
7. For very large times we expect that s=s*/s~ 1 and c=c*/c~ 1.
Subject to this expectation, find an approximate solution to (1.10).Show
that retention of the dominant term in the solution permits a match with
(19) and (lb), and thereby show that s0 and c0 continue to give first
approximations to s and c even for very large times.
8. The object of this exercise is to indicate the way that biochemists utilize
the Michaelis-Menten approximation.
(a) Define the rate of product formation V by V = dp*/dt*. Let Vmax
be the maximum rate of product formation at any time for any
initial substrate concentrations. Use (1.4) to show that Vmax =k 2 e.
(b) Using the Michaelis-Menten equation (lb), show that whens= K,
then V = ! Vmax. Equivalently, show that V = ! Vmax when s* = Km,
where Km= (k_ 1 + k 2 )/k 1 (This gives an interpretation of " or,
as biochemists prefer, of Km, the Michaelis constant.)
(c) Show that the results can be presented as shown in Figure 10.5.
This double-reciprocal plot is deservedly popular with biochemists
because (if applicable) one can fit data with a straight line.
I
v
(-I/Km, 0)
FIGURE I 0.5. Lineweaver-Burk double reciprocalplot relating velocity

V (rate of product formation) with substrate concentration s*.
[Ch. 10
320 Singular Perturbation Theory Applied to Biochemical Kinetics
t(d)
Part {b) gives an interpretation of ,c. Rather than use the other
dimensionless constant, l, it is perhaps preferable to present the
results in terms of P= k_ifk 2 [A= ,c/(l + P)J.The quantity P
is the rate at which the complex dissociates back into substrate
and enzyme, divided by the rate at which it yields enzyme plus pro-
duct. But A, or other combinations of K and A., are also directly
interpretable. What is the interpretation of ).? Of some other
combinations?
9. Consider
sy" + (1 + x)y' + y = 0, y(O) = 0,. y(l) = 1, 0 < e ~ 1.
This is a special caseof an illustrative problem discussed by Cole ( 1968).
(a) Show that an outer approximation is Yo(x) = 2/( I + x) and an inner
approximation is Y,W = 2 - 2 exp(-~). where~= x/e. Show that
a uniform approximation is 2/(1 + x) - 2 exp (-x/e).
(b) Show that the O(e) contribution to the outer solution is
Y1(x) = -t(l + x)- 1 + 2(1 + x)- 3,
while the O(e) contribution to the inner solution is
Y1(~) = C(l - e-() - 2(e-( + e- 1 -t, 2 e-().
(c) Verify that to determine the constant C, one should set 'f(e) = x/E>(s.)
and match
and
Y0 (;) + eY1 (~") =2 - 2011+ 0(0 2 11

2
) + 2 + Ce+ TST,
giving C = -t. (Recall that TST indicates transcendentally small
terms.)
(d) Find the uniform approximation that retains terms through O(s).
CHAPTER 11
Three Techniques Applied to the
Simple Pendulum
TH Is chapter contains some brief discussions of certain useful techniques

and concepts. All the discussions are centered on the pendulum problem. We
have already treated this problem in some detail, but nevertheless, as will
be seen, much can be learned from further investigation. From this we draw
the general moral that a surprising amount of scientific understanding can be
gleaned from an intensive study of a properly selected special problem.
In Section 11.1, we investigate the fate of small perturbations to the two
possible equilibrium positions of the pendulum, bob upward and bob down-
ward. This provides a simple illustration of the concept of stability, which was
briefly introduced in our earlier investigation of ameba aggregation (Section
1.3) and which is taken up again in more detail when we study the stability of
density-stratified fluids (Section 15.2).
Section 11.2 illustrates the power of the " multiple scale " perturbation
method. For the pendulum this method recognizes that there are two time
scales, one during which the pendulum executes an oscillation, and a much
longer one during which the period of the oscillation changes significantly.
The multiple scale or "two-timing" method has recently found application
in many contexts.
Section 11.3 shows how qualitative understanding of all possible motions
of a simple pendulum can be obtained by an analysis of the displacement-
angular velocity plane. The damped and undamped pendulum are contrasted.
Another method of wide applicability is thereby illustrated, phase plane
analysis.
The three sections are mutually independent and can be read in any order.
11.1 Stability of Normal and Inverted Equilibrium of the

Pendulum
Very often, in subjects ranging from mechanics to economics, the equations
that govern a phenomenon have time-independent steady state or equilibrium
solutions. A simple pendulum, for example, has two such solutions-one
with the bob downward (normal equilibrium)and one with the bob upward
(invertedequilibrium).
Why is it that we never see a pendulum in inverted equilibrium? It is
because the equilibrium is unstable, in the sense that the slightest disturbance
321
322 Three Techniques Applied to the Simple Pendulum [Ch. 11
will result in a radically different situation. This rather simple problem

will be examined here, to provide an elementary example of a general tech-
nique that will be developed further in Section 15.2.
DETERMINING STABILITY OF EQUILIBRIUM

The equation governing the motion of the pendulum contains the time
variable t only in t derivatives. This occurs in many problems, indeed in all
problems in which the choice of a time origin is immaterial. Using the pen-
dulum as an example for such time-translation-invariant situations, we shall
outline a procedure for investigating the stability of equilibrium. We shall
restrict ourselves to small departures from equilibrium or perturbations.
As a consequence we need only consider linear equations. Nevertheless, much
useful information can be obtained.
STEP A. Determine all equilibrium solutions. For the pendulum it is con-
venient to write the equation for the angular displacement 9* in the form
d29*
dt 2 + sin 9* = 0. (I)
[Here t is time, measured in units (L/g) 112 The derivation of (I) was given
in Section 2.2.]
By definition an equilibrium state does not change with time; hence we
search for a constant 0 such that (}* = 0. Thus sin 0 = 0 and we have two
cases.
Case (i): 0 = 0, normal equilibrium.
Case (ii): 0 = n, inverted equilibrium.
STEP B . Introduce a variable that measures departure from equilibrium.
For the pendulum, such a variable is 9', where
9' = 9* - 0. (2)
In terms of 9', the governing equation is
d 2 9'
dt 2 + sin (0 + 9') = 0. (3)
STEP c. Assume that the departure from equilibrium is just a small pertur-
bation and retain only linear terms in the governing equations. To obtain an
approximate equation for 9', we use the Taylor formula
2
f(9') = f(O) +f'(0)9' +f"(e) (9') ,
2!
where ~ is between O and 9'. We neglect the remainder term, since it is
proportional to (9')2, a nonlinear expression composed of the square of a
small quantity. This gives the linear approximation
sin (0 + 9') = sin 0 + 9' cos 0.
Sec. 11.1] Stability of Normal and Inverted Equilibrium of the Pendulum 323
Substituting for 0, we find in case (i) that

d2()'
dt2 + ()'= 0, (4a)
and in case (ii) that
(4b)
STEP D. Determine the behavior of the perturbations asfunctions of time.

We find that (4a) has the solution
()' = ()0 cos t + e1 sin t, (5)

where the constants ()0 and e1 are determined by the initial conditions
()'(O)= ()o' (6)
According to (5), if one slightly perturbs a pendulum hanging in normal

equilibrium, small oscillations result. The smaller the initial perturbation
displacement ()0 and velocity ()1 , the smaller the amplitude of the resulting
oscillations.
To investigate the stability of the inverted pendulum, we solve (4b) and
find, using the initial conditions (6),
(7a)
or equivalently,
()' = ()0 cosh t + e1 sinh t. (To)
DISCUSSION OF RESULTS
The inverted pendulum is an example of an equilibrium state wherein
small perturbations almost certainly increase with time. Since slight disturb-
ances are inevitable, such unstable equilibrium points should not be observed
in practice.*
An exceptional situation occurs when 01 = - 00 Here (7a) shows that
(J'(t) decays exponentially to zero. Evidently, for an initial angular displace-
mentfromequilibriumB0 , an initialangularvelocityof -B 0 is just exactly
enough to restore the pendulum, finally, to the inverted equilibrium position.
While theoretically interesting, this exceptional situation is of little practical
significance.
The pendulum can persist in an inverted position if the pivot oscillates appropriately.
See, for example, J. P. Den Hartog, Mechanical Vibrations (New York: McGraw-Hill, 4th ed.,
1956), pp. 348-50.
Equilibrium with bob upward is said to be unstable because there exists

a small perturbation that grows with time. This definition takes into account
the fact that only unusual initial conditions would be such as to exclude this
growing solution.
Equilibrium with bob downward is termed neutrally stable to the class of
perturbations considered, because these small disturbances neither grow
nor decay. Another term is also used to describe this situation, Liapunov
stable. This term is applicable if (as here) the perturbed solution will remain
arbitrarily close to the equilibrium solution, provided that the initial dis-
turbance is sufficiently small.
The effect of damping is considered in Section 11.3. There we see (from
a somewhat wider point of view) that small perturbations from normal
equilibrium eventually die out completely, so that the equilibrium is again
attained. In such a case we speak of asymptotic stability. As a matter of fact
the pendulum eventually comes to rest in the bob-down position, whatever
the initial condition, when damping is present. This position is therefore
said to be globally asymptotically stable. The large number of definitions,
and there are more, give a hint of the subtleties involved in stability theory.
We remind the reader that we have deleted all nonlinear terms, a procedure
that is reasonable in a study of small perturbations, but one that has not been
rigorously justified. On this matter we comment that our results are expected
to be correct except perhaps for the borderline case of neutral stability, where
a tiny change in the problem could nudge the state of the system either way.
A more extensive study of stability theory will be found in Section 15.2.
There, the stability of a layer of fluid whose density varies is analyzed. As
one would expect, there are close parallels between the behavior of a top-heavy
pendulum and a top-heavy fluid layer. But the latter situation is considerably
more complicated; its description involves partial differential equations.
EXERCISES
1. In biological applications the population P of certain organisms at time t

is sometimes assumed to obey the logistic or Verhulst-Pearl equation
dP _1
dt = aP(I - PE ), t = time, Ea constant. (8)
(a) Determine the equilibrium population levels.

(b) Examine their stability.
(c) With the information obtained from (a) and (b), discuss the quali-
tative behavior of the population. Reinforce your assertions by
examining the levels of P at which Pis decreasing and increasing,
respectively.
(d) Solve (8) exactly and compare the results with (a), (b), and (c).
Sec. 11.2] A Multiple Scale Expansion
325
2. Another population equation in common use, associated with the name
of Gompertz, is
dP
dt = aP(ln E- lnP).
Carry out the instructions of Exercise I (a), (b ), (c) for this model.
3. Still another population model is
a;= aP(E1p-c - E 2 ),
where the "competition constant" c satisfies O < c :::;;1. Carry out the
instructions of Exercise I(a), (q), (c) for this model.
11.2 A Multiple Scale Expansion

The stability analysis of Section 11.1 revealed important distinctions
between equilibrium points, and, by virtue of linearization, this was easily
accomplished. But our approach is of limited validity. The limitations are
most manifest in our treatment of the inverted pendulum. The solution (1.7)
predicts a displacement that continues to increase exponentially, but if we
tap an inverted pendulum we do not expect it to whirl with ever-increasing
speed. After a time, however, {I.7) cannot be expected to hold, because there
is an inconsistency between the large predicted values of ()' and "small
perturbation theory." There are no apparent inconsistencies connected with
{1.5) but, as we have discussed in Section 7.1, this solution, too, has defici-
encies that become more and more pronounced as time passes. In Chapter 2
we have briefly indicated one method of remedying the deficiency. This is
Poincare's introduction of a perturbed independent variable designed to
take into account the fact that small oscillations have a period that depends
mildly on the amplitude of oscillation. We now discuss another approach.
In our discussion of singular perturbation theory we pointed out that the
solutions to a number of problems possessed more than one scale. For ex-
ample, in the illustrative boundary value problem (Equation 9.2.1)
d 2y dy
e dx 2 + 2 dx + y = O; y(O) = 0, y(l) = I, 0 < e <0.,
the solution varied rapidly with scale 1/e in the boundary layer near x = 0
and then varied more slowly with an 0(1) scale in the rest of the interval.
Similar variation occurred in the enzyme kinetics problem of Chapter 10.
In the pendulum problem, we saw in Section 7 .1 that the regular perturbation
solution fails because on a long 0(1/a 2) time scale there is a significant devia-
tion in the pendulum's position compared to the prediction made by linearized
theory, with its slightly erroneous amplitude-independent 0(1) period.
It has become increasingly evident in recent years that explicit formal

recognition of the presence of multiple scales leads to approximation methods
of great power. For example, consider the simple problem
y + 2ey+ (1 + e2 )y = 0, y(O) = 0, Ji(O)= 1 (1)
The exact solution is
y(t, e) = exp (-it) sin t. (2)
Suppose that one applied regular perturbation theory to obtain an expansion

in powers of e, assuming that
y(t, e) = y 0 (t) + eYt(t) + e2yit) + (3)
One would then obtain the following expansion of (2):
y(t, e) = sin t - et sin t + !e 2 t 2 sin t + . (4)
As expected, terms containing powers of t appear and limit the usefulness

of the expansion. On the other hand, suppose that a solution of the form
y =f(t, et, e) were sought, where
f(t, T, e) = f( 0>(t,T) + ej<ll(t, -r) + ,;21<
2>(t, T) + . . . . (5)
Here the existence of two scales is explicitly recognized. By comparing with

the exact answer, we know thatJ< 0 (t, -r) = 0, i ~ 1, and that
so f(t, et, e) = e-u sin t. (6)
The first term of the expansion would yield the exact solution in this specially
designed problem.
We shall now illustrate the multiple scale method on the pendulum prob-
lem written in the scaled form of (7.1.4):
d2 0 sin (a0) d0
-d2
t
+ a =0; at t = 0, 0= 1, -=0.
dt
(7a)
We shall show that if an expansion like (5) is assumed, then the successive
terms in the series can be systematically obtained. Approximations result
that are valid for successively longer time intervals.
The solution 0 of (7a) is actually a function of a 2 (see Section 7.1). To
bring this out, we introduce a2 = e and expand sin (a0) = sin (e1' 2 0) in a
Taylor series. Equation (7a) is thus replaced by
d2 0
dt2 + 0 - i,-e03+ ... = 0. (7b)
Sec. 11.2] A Multiple Scale Expansion 327

SUBSTITUTION OF A TWO-SCALE SERIES INTO THE
PENDULUM EQUATION
To obtain an approximate solution to {To) we assume that E>(t,e) =
f(t, et, e) and that an expansion of the form (5) is valid in some sense. In
computing derivatives, we must use the chain rule to find, for example,
that if/< 0 >(t,-r) = exp (--r) sin t, then
!!..
j<0 >(t,et)=! (e- sin t)l,=.,dtdt + !._(e- sin t)I d-r
dt ot a-. ,=., dt
= e-u cost - ee_., sin t.
More generally, it is convenient to write

d dt d-r:
- j<o>(tet)= po> - + po>
2 - = po> + epo>
2
dt ' l dt dt l '
where the subscript i indicates partial differentiation with respect to the ith
argument. Thus if we assume that
f(t, et, e) = J<0 >(t,et)+ ef<1 >(t,et)+ e2J<2 >(t,et) + , (8)
then
and
d2 0 0>+J< 0>+JO>]
-J(t
dt2 ' et ' e) =J<11 > + e[J<
12 21 11
+ e2 UW+Jg>+ 1n> + JH>J+ , (to)

where all functions on the right side of (9) and (IO) are to be evaluated at
(t, et). Assuming continuity of second partial derivatives, !1?=ff, so that
we can write (IO) as
2
d f~t:t, e) =ff~>+ e[2ff~>+ ff Pl+ e2 [JJ~>+ 2Jg>+Jg>]+ (11)
We are now ready for the standard step of substituting the assumed series
into the governingequation.Using (11), we thus find that if 0 is given by the
seriesf of (8), then (To) implies that
0 = [ff~>(t,et)+ j<0 >(t,et)]
+ e[ff P(t, et)+ J0 >(t,et)+ 2ff~>(t,et) - iu< 0 >(t,et)) 3] + ....
Since the terms in the square brackets depend on e, we have here a situation
which is not that of a power series identically equal to zero. It is not necessary
328 Three Techniques Applied to the Simple Pendulum [Ch. JI
to set equal to zero the factors multiplying successive powers of e. It is, how-
ever, sufficient and convenient, so we impose the conditions
ff~>+jCO) = 0, (12)
ff P+ J<1> = 1u<0 >]3 - 211~>. (13)
The initial conditions of (7) require, using (9), that
p 0 >(o,0) = 1, Jf0 >(o,O)= o; (14)

p 1>(0,o)= o, /11}(0,0) = -J4>(o,
0). (15)
SOLVING LOWEST ORDER EQUATIONS

We are faced with the partial differential equation (12). If we regard f
as a function of the variables t and T, this equation can be written
a21co>
at2 +f=O.
We see that T appears only as a parameter; so the partial differential equation

is, in fact, no more difficult to solve than the simple ordinary differential
equation one meets in finding an approximate solution to the pendulum prob-
lem by regular perturbation theory. Thus
J<0 >(t,-r) = A(-r)cost+ B(-r)sin t, (16)
where the "arbitrary constants" A and B must be regarded as functions of -r.
The initial conditions (14) lead to the requirement
A(O) = 1, B(O) = 0. (17)
At this stage we have reached essentially the same conclusions as we did
in our regular perturbation calculation, where the lowest approximation was
0 ~ 1 cost+ 0 sin t. (18)
We do not know what A(-r) and B(r) are, but we do know that they are
initially unity and zero, respectively. We also know that it is A(et) and
B(et) which appear in the solution so that these quantities will remain close
to their initial values until t becomes comparable with 1/e. Equation (18)
gives no fuller information than (16) and (17), for we know that (18) becomes
a poor approximation just when t = 0(1/e).
HIGHER APPROXIMATIONS; REMOVING RESONANT TERMS

To proceed with the calculation, we substitute the results we have obtained
so far into (13), obtaining
a2p1> 1 t
--;-r + j< > = - (A cos t + B sin t) 3 - 2( -A' sin t + B' cos t),
ut 6
where ' = d/d-r. Using trigonometric identities [as in Exercise 2(a)], we can
rewrite this equation in the form
Sec. 11.2) A Multiple Scale Expansion
329
a--J<0
1
I
+JU>=[-2B' + -(A 3 + AB 1 )] cost
a, 1
s
+ [2A' + l(.83+ A1 B)] sin t
+ / 4 (A3 - 3AB2 ) cos 3t - -h(.B 3 - 3A1 B) sin 3t. (19)
Since (19) is "essentially" an ordinary differential equation, we can employ
the method of undetermined coefficients. According to this method, since
cos t and sin t are solutions of the homogeneous equation, the corresponding
particular solutions will contain terms proportional to t cos t and t sin t,
respectively.
The presence of cos t and sin t on the right side of (19) will thus force
the term ef<l) in the series (8) to contain terms proportional to et cos t and
et sin t. However small e is, these terms will eventually cease to become
negligible, compared to the lowest order approximation j< 0 >. We wish to
avoid this. Indeed we shall require that the successive approximations
l:f=oej<i) are of order eN uniformly in time. To this end we require that each
f(t,et) is bounded for O~ t <oo. Methodical application of this boundedness
requirement forms the heart of multiple scale methods.
In the present instance the boundedness requirement demands the deletion
of the "resonant" forcing terms proportional to cos t and sin t in (19);
therefore we must have
-2B' + t(A 3 + AB 1) = 0, 2A' + t(.83+ A 2B) = 0. (20a, b)
At first this nonlinear system might seem formidable, but soon we notice
that since A(r) :;,!:0, at least near r = 0 [since A(O) = 1], we can use (20a)
to find A 1 + B 1 = l6B'/A. Substitution into (20b) then gives
or
Using the initial condition (17), we at once infer that A 1 + B 1 = I. Substitut-

ing back into (20), we find that -16B' +A= 0, 16A' + B = 0. Again
employing (17), we find that [Exercise 2(b)]
't
A(r) = cos , B(r) = sin..:...
16
(21)
16
We use (16) and recall that e = a1 , and so we conclude that the pendulum
amplitude can be approximated to lowest order by
a1 t a2 t
f<0 >(t,a1 t) = cos cost+ sin
16 16sin t
=cos(1-;;), (22)
in agreement with the result of (2.2.35) and (2.2.37).

With (20), (19) becomes [Exercise 2(c)]

iJ2/(1) 1
-;-r + /<1> = -(A 3
- 3AB 2 ) cos 3t - 2 ~B3 - 3A2 B) sin 3t. (23)
vi 24
This has the solution [Exercise 2(d)]
j<1 > = C(t) cost+ D(r) sin t - 1 J2 (A3 - 3AB2) cos 3t
+ 1 ~ 2 (B
3
-3A 2
B)sin3t. (24)
At this stage, all that is known about C and D are the initial conditions
required by (15). The calculations can in principle be continued. We shall not
carry them further, however, for the-main ideas have already been illustrated.
We note the appearance in (24) of higher frequency harmonics proportional
to cos 3t and sin 3t. It should be apparent that other harmonics will continue
to appear if the calculations are continued. This harmonic generation is a
characteristic nonlinear phenomenon.
The simple linearized theory of the pendulum makes an O(a 2 ) error in its
prediction of the period. This means, as we have seen, that predictions become
unreliable when t = O(a- 2 ). The improved result (22) provides an O(a 2 )
correction to the period so th.at predictions are expected to be reliable until t
becomes O(a- 4 ). Exercises 4 and 5 show that the next approximation should
hold until t = O(a- 6 ). It is typical of a class of solution methods for nonlinear
problems defined on unbounded intervals that successive approximations
adequately cover successively larger portions of these intervals.
SUMMARY AND DISCUSSION
Let us sum up the basic elements of the procedure. Consider an ordinary
differential equation for /(t, s), where sis a small parameter. Suppose that
failure of a regular perturbation expansion to be uniformly valid, or some
other reason, leads to the supposition that/varies on two scales, one 0(1)
and the other O(s). Calculations can then proceed as follows.
(a) Assume the expansion
f(t, s) = J<0 >(t,i, s) + sf< 1 >(t,-r,s) + s 2J<2 >(1,-c,s) + , t = st.
(b) Substitute into the original equation and obtain a sequence of equa-
tions for the /Ci>by equating to zero the coefficients of successively
higher powers of s.
(c) Remove indeterminancy by requiring that the /Ci>remain bounded
for all time.
(d) Solve the equations for the j,successively.
REM AR Ks . (i) Higher approximations often require that the variables
t and t = = =
st be supplemented by -r2 s2 t, -r3 s 3 t, etc. (See Exercises 5 and 6.)
(ii) The equations for the J<i) are partial differential equations, but they involve
both variables only in the nonhomogeneous terms. Thus there is little added
difficulty on this score, since "extra" variables in many respects behave like
parameters.
We have provided a guide to the use of multiple scale techniques in a
class of problems, but this is certainly not a universal recipe. As usual, a
flexible attitude will help the reader to widen the class of problems with
which he can deal. So will practice, as encouraged by the exercises.
We can now usefully characterize, more precisely than we have heretofore,
problems for which singular perturbation methods are appropriate. These
problems have the property that an expansion in powers of a suitable small
parameter, with coefficients depending only on the independent variable,
is not uniformly valid for all relevant values of the independent variable.
We have seen three types of singular perturbation methods that can be used
advantageously on such problems. These involve (i) introduction of a slightly
distorted independent variable (Chapter 2), (ii) use of differently scaled
variables in different domains, with the separate expansions joined by the
matching technique (Section 9.2), and (iii) application of multiple scale
assumptions (this section). For a given simple problem it is often possible to
decide which of these methods is" best." More often than not, however, even
on simple problems this decision involves value judgments such as whether or
not it is " better" to sacrifice some brevity in calculation to gain a more
automatic approach.
EXERCISES
1. Assume a solution of (1) of the form (5). Verify that (6) is obtained by
substituting into the equation.
2. (a) Verify that j< 1> satisfies (19).
(b) Verify (21).
(c) Verify (23).
(d) Verify (24).
t3. Consider
ji + ey3 + y = O; y(O) = 0, y(O) = I.
Assume that
y(t, e) =J<0 >(t,t) + ej<1 >(t,r) + , where t = et,
and obtain the approximate solution
(1 +43st)-sin t + O(s).
1 2
'
y ~
Give a physical interpretation of the problem and of the solution.

4. Poincare noticed that introduction of a slightly distorted time scale -r,
where
(25)
332 Three Techniques Applied to the Simple Pendulum [Ch. II
will lead to a uniformly valid perturbation solution of the pendulum

problem. (Do not confuse this t with the abbreviation for tt used
elsewhere.) In Section 2.2 the successive coefficients were chosen
to remove " resonant" terms in the various differential equations.
M. Pritulo* suggested that in such problems, one should carry out a
regular perturbation calculation and then introduce the new time
scale chosen in such a way that secular (growing) terms are eliminated.
This exercise applies Pritulo's approach to the pendulum problem.
(a) To find a first correction to the period in the pendulum problem,
replace t by -r(l + h2 a2 + ) in (7.1.24) and show that no secular
terms remain when h2 = / 6 In doing so, prove and use the
approximation cos (-r + h2 a 2 -r) ~cos. - h2 .a 2 sin t.
(b) As a start toward finding the coefficient h4 , show that if (25)
holds, then
cost= cost - (h2 sin .)a 2
- !(h~ t 2 cost+ 2h4 t sin t)a 4 + .
(c) Given (25), show that the secular terms in the regular perturbation
solution 0 0 (t) + a 2 0i(t) + a 4 0 4 (t) of (7.1.24) and Exercise 7.1.3
are
- (h 16I) a -rsm
2 -
2
T -
2-
(h~ h2 +
16 512
1)a 4 2
-r cos -r
- ( h4 + h2 - a4 t sm

-r + h2)
- 1 ) a4 T sm
(h2
3-r.
192 16 64 1024
Show that these terms vanish if
t = t[l +-ha2 + Ji~2a4 +. ).
(That a single choice of h2 makes three expressions vanish at first
seems miraculous but is merely a consequence of the fact that we
have picked the correct form for the solution, so that everything
must work out.)
(d) With (c) the solution
cost+ (!!_)(cost - cos 3-r)+

192
has period 21t in -r and
21r[l +-ha2 + J~~2a4+ ]
int. Verify this answer by expanding the integrand of (2.2.19) in
powers of a2
J. Appl. Math. Mech. 26, 661-67 (1962).

REMARK. Pritulo's approach offers an easy way to get the first

correction. It seems to offer no advantage at higher orders, however.
Then, in contrast to Poincare's original method, one must actually com-
pute the secular terms, which are later discarded.
5. From Exercise 4 it should be apparent that determination of higher
corrections to the period of the pendulum by the multiple scale method
would require an assumption of the form
e = J<0 >(t,et, e2 t) + ef<0 (t, et, e2 t) + e2J<2 >(t,et, e2 t) + O(e3 ) (e = a2 ).
It turns out that the calculations are very formidable.

(a) To get some idea of how things go, show that the method fails if
(8) is assumed and the calculations are carried to O(e2 ).
(b) Assume that thej<il are also a function of e2 t and carry the calcula-
tions at least through O(e).
6. For a relatively simple calculation involving three scales consider
y(O) = 0, j(O) = 1.
Assume that
y =J<0 >(t,'t 1 , 't 2 ) + ef<1>(t,'ti, 't 2 ) + e2J<

2
>(t,'t 1 , 't 2 ) + O(e3 ),
(a) Show that

+ f (O) = 0,
JJg>
JJA>
+ j<ll = -2JJ?>
- 2JJO>,
+J<>= -2JJ~>
JJ~> 2
- ff?> - 2JJ
0
>- 2JJP-2JJ
1
>,
where subscripts 0, 1, 2 denote partial derivatives with respect to
t, 't 1 and 't 2 What are the initial conditions?
(b) Show thatj< 0 > = A('t 1 , 't 2 ) cost+ B('t 1 , 't 2 ) sin t.
(c) Show that
J<1> = C(r1 , t 2 ) cost+ D('t 1 , t 2 ) sin t.

(d) Find oc('t2 ) and P('t 2 ) and compare /<0 > with the exact solution.
7. To show that boundary layer problems can be treated by the multiple
scale method, consider again the example [Equation (9.2.1)]
d 2y dy
e dx2 + 2 dx + y = O; y(O) = 0, y(l) = 1, 0 < e ~ 1.
Assume that
0
y = J<- >(x,X) + ef 0 >ex,X) + ... + transcendentally small terms,
where X = x/e.
(a) Show that on substituting the series into the differential equation,
the vanishing of 0( e- i) terms leads to
J<0 >(x, X) = C1(x) + Cix) exp (- 2X).

t(b) Show that the vanishing of 0(1) terms requires
!Bl+ 2JJ1>= -[2C~ + C1] + [2C2- C2 ] exp (-2X).

Why must we require that 2C~ + C1 = O? Why is it not necessary
to require that 2C 2 - C2 = O? Show that
C1(x) = exp [t{l - x)],
while C2 (x) is unspecified except for the condition CiO) = -e 1' 2
Compare your answer with that obtained by other approaches
to the problem.
(c) What happens when e is small and negative?
11.3 The Phase Plane

Our previous discussions of the pendulum have all been concerned with
what happens when it is slightly deflected from an equilibrium position.
All the perturbation methods we employed in these discussions are useful
in a wide variety of problems, e.g., problems governed by systems of non-
linear partial differential equations. (The methods must be extended, but
their principles remain the same.)
The pendulum problem can, in fact, be solved in closed form with the
aid of elliptic integrals. We have not followed this approach because it is
rarely the case that solutions can be expressed exactly in terms of known
functions, so that discussion of the exact solution would not lead to ideas
with the desired breadth of applicability. We now take a point of view midway
between those of perturbation theory and exact solution. This will enable us
to illustrate the important phase plane approach to nonlinear problems.
We wish to consider in one" glance" all possible initial conditions; there-
fore, precise scaling is not helpful. Nondimensionalizing merely to reduce the
number of parameters that appear, we thus start with the pendulum equation
in the standard dimensional form [Equation (7.1.1)]:
d2(J* g
d(t*)2 + L sin 8* = 0, 8*(0) = a, d8* (O)= n. (1)
dt*
Sec. 11.3] The Phase Plane
335
We introduce the variables
(}= (}*, L)-112

t = t* (- . (2)
g
Here time is referred to the scale (L/g) 112 and angular displacement to the
radian. With (2) our problem becomes
0 + sin O= 0, (J(O)= a, oco)

= n(
L)112
= h. (3)
9
We now take advantage of the fact that (3) is a second order differential
equation in which the independent variable does not appear explicitly.
Using a" trick" that has become part of the elementary literature of differ-
ential equations (Boyce and DiPrima, 1969, p. 89), we make the abbreviation
ro = (Jand regard ro as a function of e.Thus
d2 8 dro dro d() dro
dt 2 = df = d() dt - (1) d()'
and our equation becomes

, dro
roro' + sin () = 0, ros-. (4)
d()
This can be integrated once, yielding

tro2 - cos () = constant = tb2 - cos a (5)
or
ro2 = 2 cos () - 2 cos a + b2 (6)
We have used the fact that ro = b when e = a, from (3).
THE PHASE PORTRAIT OF AN UNDAMPED SIMPLE

PENDULUM
Let us sketch the family of curves given by (6) when aJl possible values
of a and bare considered. To do this, we make the following observations.
(i) Both ro and e can be replaced by their negatives without altering (6)
so that the curves are symmetric in both the e and ro axes. It is therefore
sufficient to restrict our attention to the first quadrant. (ii) The right side of
e
(6) has period 2n: in so that we can further restrict attention to the strip
ro ~ 0, 0 ::;;e ::;;
2n:. (iii) If ro is positive, it follows from (4) that ro' = 0 at
e e
= 0 (horizontal tangent), ro' < 0 for O < < n (ro decreasing), ro' = 0 at
e = n (horizontal tangent), and ro' > 0 for n < ()< 2n (ro increasing). (iv) If
e :p 0, n, then Iro' I -+ oo as ro-+ 0 (vertical tangent on () axis). We have de-
duced the qualitative behavior shown in Figure 11.1.
~~..._..___.1--~---~__,~--.--1~.,q;.9
FIGURE 11.1. Portion of the phase plane for the undamped pendulum:
schematic drawing.
Of particular interest are the points (0, 0) and (n, 0). Here a tangential
slope 0) 1 is undefined by (4). But for 8 near zero, (4) gives approximately
(J)d(J)+ 8 d8 = 0, (7)
so that 0) + 8 = constant, and in the neighborhood of (0, 0) the curves are
2 2
circles. For 8 near 2t we write 8 = n + lJ and assume that lJ is small. We

obtain
ro dro - 9 dlJ = 0, 0)
2
- lJ = constant.
2
(8)
Near (n, 0) the curves are thus hyperbolas, of which the degenerate hyperbola
(J) = lJ, i.e., ro = (8 - n),
passes through (n, 0).
e
One more observation will be useful. From ro = it follows that 8 increases
with time if and only if ro is positive. This enables us to indicate with an
arrow how the curves are traced as time passes. Combining all this, we obtain
Figure 11.2.
To each initial condition 8(0) = a, 0(0) = b, there corresponds a point
(a, b) in Figure 11.2. With the exception of the points (mt, 0), n = 0, 1, ... ,
from each initial point one can follow in the direction of the arrow a unique
curve that traces out the subsequent angular displacement 8 and angular
speed 9 = (J)of the pendulum. These curves are called trajectories,and the
plane itself is called the phaseplane.The collection of trajectories in the phase
plane is aptly said to comprise the phase portrait of the system, for this
collection depicts the behavior of which the system is capable.
It must be kept in mind that values of 8 that differ by 2n correspond to
the same position. One way to take account of this fact is to imagine that
the phase plane strip O :S 8 :S 211:is wound on a right circular cylinder with
8 = 0 identified with 8 = 211:.Thus all trajectories correspond to periodic
motions.
337
FIGURE 11.2. Schematic phase portrait for the undamped pendulum. The
separatrices S, and S2 (heavy lines) bound a domain consisting of closed
trajectories about the center (0, 0). These trajectories represent periodic
oscillations of bounded extent-i.e., the pendulum never overturns. A
trajectory such as T represents a continually overturning motion.
SEP ARA TRICES

Trajectories that start near (0, 0) satisfy I81< 1t (pendulum never over-
turns). By contrast, on all trajectories starting at (1t, b) for nonzero b, the
angle 8 periodically passes through x. (An inverted pendulum, given a non-
zero initial speed, will continually circle its support.)
Consider now the special trajectories passing through odd integer multiples
of x. These separate the trajectories that correspond to paths circling the
support from trajectories that correspond to nonoverturning paths. Such a
special curve, which separates regions of different qualitative behavior,
is called a separatrix.
Let the separatrix passing through (1t,0) intersect the w axis at (0, b0 ).
Then w2 = 2 cos 8 - 2 + b~, from (6). In order that w-+ 0 as 6-+ x, we must
have b5 = 4 or b0 = 2. Hence a pendulum whose bob is instantaneously
vertically below its support will circle its support if and only if its instantaneous
angular speed has magnitude greater than 2.
Example. Give a physical explanation of the condition just obtained.

Sohdion. An initial dimensionlessangular speed of 2 corresponds by (3) to an
initial dimensionalangular speed of n = U..L/g)-1 12 Let the bob have mass m, and
take the zero of potential energy at the bob-down position. Then the initial energy
is all kinetic and hasthe magnitude !m(f!L)2 = 2mLg, sincethe bob has linear speed
QL if the pendulum is of length L. Energy is conserved so that this initial kinetic
energy is just enough to elevate the bob a distance 2L to a motionless inverted
position. If the initial energy is any larger, the bob will reach the inverted position
with nonzero speed and will continue to rotate.
CRITICAL POINTS
e
Consider the points at which OJ= = 0. From (4), at such points we also
have sin O = 0. These points play an important dual role. From the point
of view of the original differential equation (3), these are the equilibrium
points that represent solutions which can persist through time without
change. When the problem is cast in the form (4), these points represent
the criticalpoints.
To see the role of critical points, consider first a point P which is not
such a point. Then, from (4), one has
, sin fJ
OJ=---,
(J)
OJ"'0.
This gives a unique value to the slope of the trajectory OJ= OJ(fJ)that
passes through P. By contrast, the above equation cannot be used as it stands
at a critical point (w 0 , fJ0 ). The only way to obtain a slope at such a point is
to consider
sin fJ
lim
(I)
and this limit generally has different values depending on how the approach
to (OJ0 , fJ0 ) is made. (See Exercise 1.)
Because trajectories have a unique tangent except at critical points, it
follows that curves composed of trajectories cannot cross except at critical
points. Consequently, examination of trajectories in the neighborhood of
critical points is particularly revealing. But since critical points and equi-
librium points are identical, such an examination is equivalent to a study of
the stability of equilibrium. Thus the small perturbation solutions given in
(1.5) and (1.7) describe behavior in the neighborhood of normal and inverted
equilibrium. Equivalent descriptions are provided by (7) and (8). In particular,
the exceptional perturbation of the inverted pendulum that decays to zero
[see the material below (1.7)] is represented in the phase plane as a perturb-
ation along a separatrix. A solution starting precisely along the appropriate
separatrix will follow that trajectory into equilibrium at (n, 0).
The wide applicability of our approach can be appreciated if it is realized
that for any dynamical system whose position can be described by a single
coordinate, say x, Newton's second law takes the form x = g(x, x, t). Most
Sec. 11.3] The Phase Plane 339
important are autonomoussystems in which time does not explicitly appear;

i.e., g = g(x, x). For these, if we introduce the speed v = x,we have v(dv/dx) =
g(x, v}, and consideration of the v-x phase plane is again appropriate.
More generally, one can consider the first order equation for y(x):
dy
f(x, y) dx =g(x, y). (9)
The slope dy/dx is uniquely defined except at critical points wheref(x, y) = 0,

g(x, y) = 0. Let trajectories in the x-y (phase) plane be parametrized by t, so
dy/dx = (dy/dt)/(dx/dt). Then the trajectory through (x 0 , y 0) can be described
by x = x(t),y = y(t), where
dx dy
dt = f(x, y), dt = g(x, y); x(t 0 ) = x 0 , y(t 0 ) =y0 (IO}
LIMIT CYCLES
By the relevant existence and uniqueness theorems (assuming sufficient
smoothness-see Chapter 2), precisely one trajectory passes through each
point. The further fact that the slope is uniquely defined except at critical
points makes plausible the following classification of possible trajectories
emerging from a point (x 0 , y 0 ). *
(i) The trajectory through (x 0 , y 0 ) consists entirely of the point (x 0 , y 0 ).
Then this point is an equilibrium point (i.e., a critical point).
(ii) The trajectory through (x 0 , y 0 } returns to (x 0 , y 0 ). The resulting
simple closed curve is called a limit cycle and is the phase plane path
of a periodic solution.
(iii) The trajectory through (x 0 , y 0 ) neither remains at this point nor
returns to it, but either passes to infinity, or approaches some other
equilibrium point, or approaches a closed curve. Although the closed
curve could pass through several critical points, in practice it is
usually a limit cycle (and therefore passes through no critical points).
The location of limit cycles is generally not a simple matter and will not
be discussed further. It is frequently easy, however, and is always valuable,
to determine behavior in the neighborhood of critical points.
BEHAVIOR OF TRAJECTORIES NEAR CRITICAL POINTS
If (X, Y) is a critical point of (9) or (10), then by definition
f(X, Y) =0, g(X, Y) = 0. (11)
Introducing deviations from equilibrium x and y by
x(t) = X + x(t), y(t) = y + y(t),
S. Lefschetz, Differential Equations: Geometric Theory (New York: Wiley-lnterscience,

1957), Chap. 10.
Three Techniques Applied to the Simple Pendulum [Ch. 11
we have by formal application of the two-variable Taylor formula
dx _ b- d-y - d-
dt =ax+ y, dt =ex+ y, (12)
where the constants a, b, c, and dare given by

a=f,e(X, Y), b = f.,(X, Y), c = g,.(X, Y), d =g1(X, Y),
and where remainder terms have been neglected. The linear system (12) has
solutions of the form
if
For a nontrivial solution the determinant of the coefficient matrix must

vanish so that
m2 +/Jm+y=O where fl= -(a+ d), y =ad-be. (13)
We assume that 'l' ::!=

0 and P2
- 4y #=0. There are then these possibilities.
Equation (13) has the real roots m1 and m2 , where

2m 1 = -P +JP 2 -4y, 2m 2 = -fl-Jp 2
-4y.
l(a): y < 0, so m1 > 0, m2 < 0. Equilibrium is unstable and trajectories
near (X, Y) appear like the level lines near a mountain
pass. Hence (X, Y) is called a saddlepoint or col. An
example is the equilibrium point (,r, 0) in Figure 11.2.
l(b): y > 0. Both roots are positive (negative) and equilibrium is
unstable (asymptotically stable) if fl is negative (positive).
The trajectories can be shown to form a nodeas illustrated
in Figure 1I .3.
II: /J2 - 4y < 0.
Now (13) has the complex conjugate roots --!P fiJ 4y - p2 and (12) has
solutions proportional to
341
FIGURE 11.3. Some trajectories near a typical stable node. The arrows
are reversed when the node is unstable.
Il(a): P= 0. Perturbations are periodic and trajectories near the equi-

librium point are thus closed. Such an equilibrium point is
called a center. Equilibrium is Liapunov stable and neu-
trally stable. An example of a center is the point (0, 0) in
Figure 11.2.
Il(b): p "F0. The equilibrium point is a focus for spirals. If /J< 0, as

time passes the phase point spirals farther and farther
from the focus, which is unstable. If /J> 0, on the other
band, the focus is asymptotically stable.
The different types of qualitative behavior exhibited by solutions to (12)

in the various situations can be quickly ascertained from Figure 11.4. But is
this behavior appropriate for the nonlinear equations (10) in the neighbor-
hood of the equilibrium point (X, Y)? That is, can one justify the linearization
leading to (12)? It turns out that a sufficient condition for an affirmative
answer is the continuity of f and g and their first partial derivatives in a
neighborhood of (X, Y). An exception,not surprisingly, is the center, for the
slightest inaccuracy can prevent the perfect return of the state point to a
given position that is required for a periodic solution.
The case of the center, and the cases y = 0, p2 - 4y = 0, which were not
included in the possibilities discussed above, certainly require more dis-
cussion. This is not within our scope here.
As a worked example, we now consider the dampedpendulum. The device
used above to obtain an explicit expression for the trajectories no longer can
be employed. Thus general phase plane methods are essential in this case.
Three TechniquesApplied to the Simple Pendulum [Ch. 11
342
ff -4-r<O,(j>O,-y>O
Stable focus
1< 0: saddle
P2 - 4'y < 0, /3< 0, 'Y> 0

Unstable focus
Unstable node
FIGURE 11.4. Qualitative behavior of the solutions to (12): dx/dt =

ax+ by, dy/dt =ex+ dy. Abbreviations: f3= -(a+ d), y =ad-be.
Example. Analyze phase plane behavior for the equation

8 + v8+ sin 8 = 0, (14)
governing a damped pendulum.
Solution. We write
x=8, .i=y,
so that
.i=y, y= -vy-sinx .
Linearizing about the equilibrium point (0, 0), we have
.i=y, y= -vy-x,
so that, from (12) and (13),
a=O, b= I, c=-l, d=-v;
fJ=v, y= 1,
Therefore (0, 0) is a stable node if v 2 - 4 > 0 and a stable focus if v 2 - 4 < 0.
The two possibilities are identical with the over- and under-damped situations
encountered in elementary treatments of the (linearized) pendulum problem.
Consider the equilibrium point x = rr,y = 0. Introducing perturbations x and y by
x=rr+x, y=y,
Sec. 11.3] The Phase Plane 343
and linearizing, we obtain

x=y, y= -vy+x.
Hence fJ= v, y = -1, fJ2- 4y = v 2 + 4, and (1r, 0) is a saddle point (as it was for
the undamped pendulum). Physical intuition rules out limit cycles, for there can
be no periodic solutions with no supply of energy to replace that which is dissipated
through damping. With slight damping, trajectories like T in Figure 11.2 now
spiral slowly into an equilibrium point that represents a motionless pendulum with
bob down. It can be seen that in the underdamped case O < v < 2, the phase portrait
is as given in Figure 11.5.
Our treatment of the phase plane provides the means to obtain some useful
results. Still, it only scratches the surface. For an exposition that is simple
and brief but a good deal more comprehensive than that given here, see
(/
'--'----x=e
FIGURE 11.5. Phase portrait for the slightly damped pendulum. The
"domain of attraction" of (0, 0) is shaded, and heavy lines denote separ-
atrices. Along trajectory T, the angle increases by several multiples of21r
(several overturnings) before the trajectory finally approaches an asymp-
totically stable focus. Just above the separatrix S2, such "overturning"
trajectories are headingfor the stable focus at (0, 0). Just below S 2, the tra-
jectories represent decaying oscillations of bounded extent. The reader
should try to convince himself that the qualitative behavior depicted here
follows from the nature of the singular points (alternating saddles and
foci), from the absence of limit cycles and the fact that trajectories never
cross but terminate at singular points, and from some physical considera-
tions. Compare Figure 11.2.
Chapter 9 of Boyce and DiPrima (1969). Much richer treatment with an

eye to applications can be found, for example, in books by J. Stoker* and by
A. Andronov, A. Vitt, and S. Khaikin.t For theoretical results, start with the
classical text of Coddington and Levinson (1955).
The phase plane concept is capable of wide generalization. For example,
consider a system of N particles. Initial conditions require specification of
three position components and three velocity components for each particle.
Instead of the phase plane, we must consider a phase space of 6N dimensions,
half for the position coordinates and half for the velocity coordinates. One
speaks of a system of 6N degrees of freedom.In continuum mechanics, par-
ticles lose their identity, but it has proved useful to introduce infinite dimen-
sional phase spaces, with a single dimension being identified with a Fourier
component of some variable.
Observe that what is of interest in phase plane analysis is the qualitative
behavior of trajectories, not their precise course. Modem research on differ-
ential equations is largely concerned with elucidating such qualitative be-
havior, and this is just what often concerns the applied mathematician.
EXERCISES
t. (a) Use (7) to show that if one approaches the origin along the line
ro = KO, then the trajectories one crosses have the slope - K- 1
(Thus the limiting slope depends on the mode of approach.) Interpret
geometrically.
(b) Harmonize (1.5) and (1.7) with (7) and (8).
2. (a) Eliminate e* and c* from (10.1.3) to obtain a pair of equations for
s* andp*.
(b) Show that there is only one physically significant equilibrium point,
and determine its stability.
(c) Find the four regions in the first quadrant of the s* - p* plane where
ds*/dt* and dp*/dt* are both positive, both negative, or have the
two combinations of opposite signs. Use this information to sketch
the possible trajectories.
(d) Compare the present approach to the problem with that of Chapter
10. [See I. G. Darvey and R. F. Matlak, "An Investigation of a
Basic Assumption in Enzyme Kinetics Using Results of the Geo-
metric Theory of Differential Equations," Bull. Math. Biophys.
29, 335-41 (1967).]
3. (a) The equations
dx dy
dt = (a - bx - cy)x, -=(e-fx-gy)y,
dt
Nonlinear Vibrations (New York: Wiley-Interscience, 1950).

t Theory of Oscillations (Elmsford, N.Y.: Pergamon, 1966).
Sec. 11.3] ThePhase Plane 345
are a simple model for the competition between two species of

organisms. (Here a, b, c, e, f, and g are constants.) Write a brief
essay on what is assumed in this model, and on what some of its
limitations are expected to be.
(b) By examining the phase plane, show that if (a/c) > (e/g) and (a/b) >
(elf), then species x wins. Is this reasonable?
(c) Discuss the case (a/c) < (e/g), (elf)< (a/b). (See Section 3D of
J. M. Smith, Mathematical Ideas in Biology, Cambridge: Cambridge
U.P., 1968.)
4. Use the phase plane approach on Exercise 15.2.7(which is self-contained).
CHAPTER 12
Longitudinal Motion of a Bar
WE NOW begin the third part of the present volume, a part devoted to the
study of basic continuum mechanics. This chapter will be devoted to a
discussion of the one-dimensional longitudinal motion of a bar, a physical
problem that arises in many different contexts. As we shall see, it is possible
to elucidate a number of significant aspects of the phenomena involved, with
relatively elementary mathematical techniques. To give one example, an in-
coming elastic wave doubles its force at a plane where it is reflected (Exercise
12.2.9). There could be dire consequences if this were not anticipated.
We have based the development of the basic equations on the same ap-
proach that is used in deriving similar results for the three-dimensional
continuum. We feel that the general approach is clearer if it is first presented
without the complications of three-dimensional geometry. Thus we rec-
ommend that at least Section 12.1 be studied before the general material,
which begins in Chapter 13. (On the other hand, some will prefer to regard the
present chapter as providing a particular illustration of the general theory
of continuum mechanics, and this is certainly a feasible procedure.) Many of
the specific questions that will be considered are prototypes of similar
problems which will be tackled later with more complicated geometries.
Such items as wave propagation, reflection and transmission, dispersion,
propagation of discontinuities, and free vibrations arise in a natural manner.
The mathematical techniques used in studying these can be generalized,
although with difficulties at times, so that they apply in two and three dimen-
sions.
Our general strategy will be to adopt an optimistic attitude toward ob-
taining a simple but meaningful one-dimensional formulation of elastic*
motion. Certain subtleties that a deep discussion would require are not
treated here.
12.1 Derivationof the GoverningEquations

GEOMETRY
The bar shown in Figure 12.1 is assumed to be slender. That is, the average
diameter of a cross section must be small when compared with the length of
the bar. Although the cross section may be variable, we shall require that its
"The property of recovery of an original size and shape is the property that is termed
elasticity" (Love, 1944).
349
Longitudinal Motion of a Bar [Ch. 12
350
-f--------~----r-
FIGURE 12.1. A bar that probably can be regarded as "slender."
Cross sections normal to the unit vector i are assumed to remain plane
and to retain their shape as they move.
variation along the bar be small. Because of these geometrical restrictions it

is not unreasonable to assume(as we shall) that each crosssection remainsplane
throughout the motion and moves longitudinallywithout distortion. It is this
assumption, that the only variation is longitudinal, which gives the problem
its one-dimensional character.
Throughout this chapter we assume that the bar can be regarded as
composed of continuously distributed matter in smooth motion (with possible
exceptions at special locations). This is perfectly natural, but some discussions
of the deeper implications of the continuum assumption will be found in
Section 13.1.
We shall switch back and forth between two independent variables. The
spatial coordinate x picks out a particular location in space, while the
material coordinate A picks out a particular cross section. We shall always
specify a cross section by stating its distance A from a reference point (the
origin). This measurement will be made at some fixed reference time (usually
time t = 0). Distance to the right (left) of the origin will be taken positive
(negative), as usual.
In our one-dimensional problem, complete information about the motion
of the bar requires knowledge of the position of every section at all times. We
shall write the required functional description of the cross-section paths as
follows:
x = x(A, t), x(A, 0) = A. (la, b)
In words, these equations state "x is the position now (at time t) of the
section that was initially (at time zero) located at A; A is the position at time
zero of the section initially located at A."
We shall assume that (1) is uniquely invertible so that we can write
A= A(x, t) [with A(A, 0) = A]. (2)
In words, "A is the initial position of the particle now at x." Using common
language, (2) is obtained from (1) by solving for A in terms of x. By definition,
this implies the identities
x[A(x, t), t] = x, A[x(A, t), t] = A. (3a, b)
Example 1. If x= A+ t, then A= x- t and, indeed, (x- t) +t = x, (A+ t)

-t=A.
Sec. 12.1) Derivation of the Governing Equations 351

As stated, sometimes we use A as the independent variable [in (l)] and
sometimes x [in (2)]. But note that we also sometimes use A as a dependent
variable [in (2)] and we sometimes use x in this role [in {1)].
Usually, we shall use lowercase letters for functions of the spatial variable
and uppercase letters for functions of the material variable. Thus we shall
write
f[x(A, t), t] = F(A, t), f(x, t) = F[A(x, t), t]. (4a, b)
Example 2. Let 8(x, t) denote the temperature at time t at position x. Let

0(A, t) denote the temperature at time t of the section which was initiallylocated
at A. Then
8[x(A, t), t] = 0(A, t), 8(x, t) = 0[A(x, t), t]. (Sa, b)
Note that (Sa) can be converted into (Sb) by substituting for A in terms of x and t
and then using (3a). Similarly,(3b) can be used to convert (Sb) into (Sa).
We shall frequently have to convert from spatial to material variables.

Rather than use the somewhat bulky notation of (4a), or the even bulkier
f(X, t) lx=x(A, I)'
we shall sometimes make the abbreviation

f[x(A, t), t] =f(x, t) IA or f[x(A, t), t] =!IA (6)
This should be read"/ evaluated at A." Similarly, we write
F[A(x, t), t] =Fix (7)
To see this notation displayed in useful formulas, consider the differ-
entiation of (4a) by means of the chain rule.* We find [Exercise l(a)]
oF of I
ox
oA = ox .. oA'
oF
ai =
of I
ox
ox .. ot
of
+ ot
I... (Sa, b)
Focus attention on a section that is initially located at A and now, at time t,

is to be found at position x. The vector joining the initial position to the
current position is called the displacement vector and is denoted by
U(A, t) = U(A, t)i.
* A detailed discussion of the chain rule will be found in Appendix 13.1. We note here only
that it might be helpful to think of x and r as spatial variables, and A and r' as material variables,
with r' = t. Then chain rule applied to (4a) gives
u lfl b lfl ~ u lflh lfl ~

oA- ox " oA+ or ....oA ot'- ox ....ot'+ or ....ot'
These equations reduce to (8) when t' is identified with t.
352 Longitudinal Motion of a Bar [Ch. 12
Here i is a unit vector in the positive A direction. The component function

U(A, t) will be known as the displacement (in material coordinates). We see
from Figure 12.2 that
x(A, t) = A + U(A, t). (9)
---m----------;-
TimeO Time t
---ffi----
A
U(A, <)i.
FIGURE 12.2. The displacement U(A, t) takes place when a section

moves from initial position A to final position x(A, t) =A+ U(A, t).
We define the displacement in spatial coordinates, u(x, t), by

u(x, t) = U[A(x, t), t]. (10)
If we substitute for A in terms of x, we find from (9) that

x = A(x, t) + u(x, t). (11)
This has the interpretation "the initial position of the section now at x plus
that section's displacement equals its present position."
A section's longitudinal velocity component, or velocity for short, is
defined in material coordinates by
u( ) _ ox(A, t)
,,,A, t - (12)
a,
From (9) we also have
V(A, t) = oU(A, t). (13)

at
The spatial velocity component v is related to V by
v(x, t) = V[A(x, t), t], v[x(A, t), t] = V(A, t). (14a, b)
The velocity vector is then given by v(x, t)i = v(x, t) and V(A, t)i = V(A, t) in
spatial and material coordinates, respectively.
THE MATERIAL DERIVATIVE AND THE JACOBIAN

In present notation the chain rule result (Sb) can be written
(15)
Sec. 12.1] Derivation of the Governing Equations 353
We shall define the material derivative DJ/ Dt of the function f(x, t) by the
important equation
Df(x, t) = oF(A, t) I . (16)

Dt ot ,t=,t(x, I)
By making the substitution A = A(x, t) in (15), we thereupon find
Df(x, t) = of(x, t) + v(x t) of(x, t).

(17)
Dt ot ' ox
The material derivative Df(x, t)/ Dt represents the rate of change of the
function f following a material section. It can be thought of as giving the
change with time off, as seen by an observer who is riding with the section
now located at x.
Example 3. Consider a motionless bar (v =0) of uneven temperature 9(x, t).

Here
D9(x, t) 89{x, t)
Dt 8t
for the only reason that the temperature can change at x is because of local heating
or cooling processes which occur at that fixed point. But if the bar moves, we have
DB a(J 89
-=-+v-.
Dt 8t 8x
(18)
There is an additional term because material is being bodily carried or convected

past the position x at velocity v.
Example 4. Consider the case of an infinite bar that steadily moves through a
roller which thins it. (See Figure 12.3.) Let some time be designated t = 0, and
label with the coordinate A some section far to the left of the roller. The velocity
of this section will be approximately constant until the section reaches the roller.
Then it will speed up to the faster velocity that is required to move the thinner bar
out of the way. Thus V(A, t) will have a graph like that in Figure 12.4. On the other
hand, if one fixes one's gaze at a particular point x, the velocity there will never
FIG URE 12.3. A very long bar moving steadily through a roller. The
thinner material moves faster than the thicker.
354
V(A,t)
F10 URE 12.4. The velocity V of a fixed section of the bar of Figure 12.3
increases as the bar is stretched because of passage under the roller.
change, for the process has been assumed to proceed steadily. Thus at any fixed
value of x, v(x, t) is constant; i.e., ov/ot = O.
For the velocity, (17) generally takes the form
Dv(x, t) = ov(x, t) + v(x, t) ov(x, t). (19)

Dt ot ox
The left side of this equation is the rate of change with time of the velocity of a
certain cross section, namely, that cross section now located at x. In other words,
(19) is an expression for the acceleration of the cross section now located at x.
As we have seen, in the present example the first term on the right side of (19) is
zero, for the velocity at a given position is always the same. But we expect from
Figure 12.4 that the acceleration of any given section is nonzero. And indeed, the
second term in (19) gives a nonzero contribution to the acceleration. This arises
from the fact that the section which is at x at time tis not the section that was at x at
time t-At.
It is hard for some people to believe that ov(x, t)/ot is not the acceleration of the
section located at x. But note that
ov(x, t) . v(x, t + At) - v(x, t)

--=lun .
ot M-+o At
The numerator gives the difference in velocities of sections located at x at times

t + At and t, respectively. Generally, these are different sections, but to find the
acceleration we must compute the time rate of change of a single section's velocity.
We have derived some extremely useful results by manipulating (Sb).

Somewhat less useful formulas, but still worthwhile ones, can be derived from
(Sa). Indeed, (Sa) and (9) imply that
aF= (i+ au) aJI

aA aA ax A
Sec. 12.JJ Derivation of the Governing Equations
355
or
of
ox
I..t
oF/oA
= 1 +oU/iJA
0 (20)
Moreover [Exercise l(b)], one can show that

of/ox
oFI (21)
oA x 1- au/ax
The derivative iJx/iJA that appears in the preceding calculations is the
Jacobian of the one-dimensional transformation between the x coordinates
and the A coordinates. With this in mind, let us write ox/oA= J(A, t). We
shall have need in the near future for a simple expression for iJJ/iJt.Applying
the chain rule, we see that
oJ a (ax) a (ax) av avj ox

ot ot oA - oA ot - iJA - ox ..t oA.
We thus obtain the Eulerexpansionformulain material coordinates,
iJJ(A, t) =
ot
J(A, t) ov(x,
t)
ox ...
I. (22)
In spatial coordinates, (22) becomes

DJ[A(x, t), t] [ ( ] ov(x, t)
Dt = J Ax, t), t ~. (23)
The extension of this expansion formula to higher dimensions is found in

Section 13.4, where it will be seen that the results are essentially the same as
(22) and (23), but the calculations are more elaborate.
CONSERVATION OF MASS
As will be the case for the three-dimensional continuum, we shall take as

basic assumptions generalizations of the various fundamental laws that are
postulated in the study of the mechanics of systems of particles.
We shall begin by considering the conservation of mass,by which we mean
that the mass of an arbitrary material portion of the bar is unchanged
throughout the motion. In order to see the analytic implications of this
statement, let us designate the mass per unit volume of the bar at a section
occupying the position x at time t by p(x, t). The mass density,or simply
density,c5(A,t) in material coordinates is therefore given by
c5(A,t) = p[x(A, t), t]. (24)
Similarly, the area of a cross section occupying the position x at the time t
will be denoted by o'(x, t). In material coordinates the cross-section area is
denoted by S, where
S(A, t) = a[x(A, t), t]. (25)
Recall that we have assumed that each cross section moves as a whole in
the longitudinal direction only. Consequently, the area of a given cross
section will not change in time; i.e.,
S(A, t) = S(A, 0). (26)
Thus
a(x, t) = S[A(x, t), O]. (27)
Furthermore, since A(x, t) = x - u(x, t), we may write

o{x, t) = S[x - u(x, t), O]. (28)
Since the distribution of area at time t = 0, S(A, 0), is assumed known, we

observe that o'(x, t) is a known function of x - u(x, t), although the dis-
placement u(x, t) must itself be found.
We shall follow an arbitrary portion of the bar in time. At t = 0, this
portion is designated by the interval
Ms;, As;, N,
while at t it is given by
m(t) s;,x s;,n(t),
where
x(M, t) = m(t), x(N, t) = n(t). (29)
The assumption of mass conservation can now be written in the form

d n(t)
dt Im(t)
p(x, t)a(x, t) dx = 0. (30)
In order to avoid differentiating the limits of integration, we change the

variable of integration to material coordinates with the result
d
dt
I"p(x, t)a(x, t) dx = dtd Jr1'b(A, t)S(A, t) OX
m M oA dA.
Taking the derivative of the integrand, and replacing ox/oA by J, we obtain
t [aot
N a]
ot
(c5S)J+ c5S J dA. (31)
Using the Euler expansion formula (22), we may write the last integral as
Returning this last result to spatial coordinates, we find that the conservation
of mass requirement takes the form
av]
I [D(pu)
n
m Dt + pu ax dx = 0. (32)
Since the interval m ::;;x ::;;n is taken over an arbitrary portion of the bar,
the integrand in the above equation must itself vanish if (as we assume) it is
continuous.* We thus obtain the differential equation of mass conservation
D(pu) + pu ov= O. (33)

Dt ax
This equation is frequently known as the continuity equation. A form that may
look more familiar to some can be written if we define the mass per unit length
=
p pu with which (33) becomes
Dp _av
Dt +pox =O. (34)
The continuity equation can be simplified in the one-dimensional situation

we have been considering. To see this, recall that S(A, t) = S(A, 0). Thus
Du
so-=0 (35)
Dt
and (33) becomes
Dp av
u-+pu-=0. (36)
Dt ax
In the event that the material is incompressible,
b(A, t) = b(A, 0). (37)
This implies that
aJ = 0 (38)
at '
Formally, we have used the Dubois-Reymond lemma. [See the remarks following Equation
(4.1.15).] It might be preferable to use this lemma before making the final change to spatial coordi-
nates, for then the arbitrary nature of the integration interval is clearer.
and the continuity equation reduces to
av=0. (39)
ox
Consequently, the bar does not deform as it moves (Exercise 11).
FORCE AND STRESS

In order to consider the hypothesis of balance of linear momentum, we
shall have to say a word or two about the types of forces that will be con-
sidered. These will be divided into two classes, body forces and surface
tractions. Bodyforces (like gravity) will be defined as forces acting on a portion
of the bar which are functions of the volume or the mass of that portion.
Surface forces are forces that are transmitted across a cross section from one
portion of the bar to the other. It is assumed that they depend only on the
position of the section, and whether the portion of the bar being acted upon
lies on the left or the right of the section. (For the corresponding analysis in
three dimensions, the reader is referred to Section 14.2.)
The forces and other vectors we shall be considering are all parallel to the
i direction, and hence the distinction between the vector and its one component
is not too significant. Nevertheless, it will be useful for the moment at least
to preserve this difference. The body force per unit mass will be denoted in
spatial coordinates by f(x, t). In material coordinates the body force is
F(A, t), where
F[A(x, t), t] = f(x, t).
We shall also write
f(x, t) =f(x, t)i.
The surface traction or stress vector will be denoted by t(x, t; + ). This
designates the force per unit area across a section located at x at time t.
The force is exerted by material on the positive side on material on the negative
side.* Similarly, t(x, t; - ) is the force per unit area across a section at x
at time t exerted by material on the negative side acting on material on the
positive side.
We emphasize that if we consider a section of area <1at the location x at
time t [Figure 12.S(a)], we cannot specify the traction until we agree which
portion of the bar is considered to exert the stress. Thus, in Figure 12.S(b),we
consider the stress produced by the material to the right of the section. The
force so produced is <1(x,t)t(x, t; + ). Similarly, in Figure 12.S(c) we have
Remember, by convention the positive side is the right side. Another notation, which is
more cumbersome but which is required in Section 14.2, uses t(x, t; i) instead of t(x, t; +).
Here i is the unit normal pointing outward from the material on which the stress is acting.
Similarly, t(x, t, -i) is used instead of t(x, t, - ). Again, the indicated vector (-i in this case)
points outward from the material on which the stress is acting.
Time t
-x axis
Jo(,n
<(x,-):z{]
(bl
x
(a)
(c)
FIG URE 12.5. (a) The stress in the bar at the section x is undefined.
(b) The product of the stress vector t(x, t, +) and the cross-section area
a(x, t) gives the surface force exerted by the unshaded portion of the bar
on the shaded portion . (c) In contrast to (b), here the surface force repre-
sents the effect exerted by the shaded portion on the unshaded.
indicated the force u(x, t)t(x, t; - ) produced by the material to the left of the
section on the material on the right. We have drawn the vectors t(x, t; +)
and t(x, t; - ) oppositely directed in anticipation of some type of law of
"action and reaction." Such a law will be proved below.
We must point out that it is not known at this stage whether the force in
Figure 12.5(b) points to the right (as drawn) or to the left, i.e., whether the
material to the right of the section pulls or pushes on the remainder of the
bar. When a problem is completely analyzed, this question can be settled.
In the absence of this analysis the direction of the arrow is arbitrarily chosen,
and the lack of significance of this direction must not be overlooked.
By contrast, if we know that a portion of the bar is in a state of tension,
then the forces applied at the end sections will be as shown in Figure 12.6(a).
If the section of the bar is in a state of compression, the forces are as shown
in Figure 12.6(b).
A state of tension is sometimes referred to as one of "positive" stress,
whereas a state of compression is considered to be one of "negative" stress.
If we use this nomenclature, we see that there are two sign conventions at
work. Considering tension as positive means that we are using a sign con-
vention intrinsic to the body with positive stresses pointing outward and
(a) (b)
FIG URE 12.6. (a) A bar under tension (positive stress). (b) A bar under
compression (negative stress).
negative stresses pointing inward. On the other hand, there is a sign con-
vention for the components induced by the orientation of the A axis. Thus
we see from Figure 12.6(a) that in a state of tension, of" positive" stress, the
stress t(m, t; - ) has a negative component, whereas t(n, t; +) has a positive
component. The opposite is, of course, true for a state of compression. Both
types of sign conventions are useful and can be effective, provided that one
is careful to state which is being employed.
BALANCE OF LINEAR MOMENTUM
As in our discussion of the conservation of mass, let us follow an arbitrary
portion of the bar described in spatial coordinates by m(t) ::s;x ::s;n(t). In
mechanics, (linear) momentum is the product of mass and velocity. Thus the
linear momentum of this section is defined by the integral
('n(r)
J_ puv dx. (40)
m(r)
The hypothesis of balance of linear momentum* states that the time rate of
change of linear momentum of any portion of the bar is equal to the sum of
all the external forces acting on it. Thus
dt In
d /uv dx = /uf dx In + t(m, t; - )u(m, t) + t(n, t; +)u(n, t). (41)
The calculation of the left-hand side of (41) can be handled in exactly the
same manner as that used in our discussion of mass conservation. One shifts
to material coordinates, carries out the differentiation using the Euler expansion
formula, and then returns to spatial coordinates. After observing that certain
terms vanish because of the mass conservation formula (33), one obtains
[Exercise l(c)] the result
d f f Dv
-d J_PUVdx = J_PU- dx. (42)
t m m Dt
With this, our expression (41) for the balance of linear momentum becomes
([pu(~;-c)]dx=t(m,t;-)u(m,t)+t(n,t; +)u(n,t). (43)
The right-hand side of (43) is in somewhat awkward form, since it does

not appear as an integral and thus precludes use of the Dubois-Reymond
lemma. This difficulty can be alleviated by the following procedure. Let us
consider a generic section x = x 0 , in the arbitrary portion of the bar, which
divides it into two pieces I and II, satisfying
I: m ::s;x ::s;x 0 , II: x 0 ::s;x ::s;n.
If the linear momentum of the bar were truly conserved, then the time derivative of (40)
would be zero.
Sec. 12.1] Derivation of the Governing Equations
II
1-------------.\----------+------ x
m n
FIG URE 12.7. Section x = Xo divides the bar into pieces I and ll.
(See Figure 12.7.) We shall now apply formula (43) to each of these parts,
with the results
f pu (DV
m
xo )
- - f dx = t(x
Dt
0, t; +)u(x 0 , t) + t(m, t; - )u(m, t), (44a)
f pu (Dv
n
XO
- - f ) dx = t(x
Dt
0, t; - )u(x 0 , t) + t(n, t; + )u(n, t). (44b)
Adding (44a) and (44b) and subtracting (43), we find that
t(x 0 , t; + )u(x 0 , t) + t(x 0 , t; - )u(x 0 , t) = 0.

Since x = x 0 is a generic section, we may then conclude* for any section
It(x, t; - ) = - t(x, t; + ).1 (45)
It should be noted that the result derived in (45) can be considered as a form
of Newton's third law in ordinary particle mechanics (action equals reaction),
although in this case it is found as a consequence of the conservation of linear
momentum. Equation (45) also tells us that we predicted correctly in drawing
the oppositely directed vectors t(x, t, +) and t(x, t, - ) in Figures 12.5(b)
and 12.5(c).
We are now in a position to simplify (43) through the use of the action-
reaction result (45). To this end, we introduce the following definition: the
stress component, T(x, t), is definedt by
It(x, t; +) = T(x, t)i. J (46)
The derivation just completed, save for notation, is precisely the same as one that can be
applied in three dimensions. See Exercise 14.2.1.
t The reader will note that we have been inconsistent in our notation in using a capital letter
to refer to a quantity evaluated in spatial coordinates. We shall find it convenient, however, to
make this deviation to match more closely with the symbolism conventionally used in three
dimensions.
We thus see from this definition and (45) that
t(x, t; -) = -T(x, t)i. (47)
In a few instances, we shall want to refer to the stress component in material

coordinates and shall use the notation '(A, t) for this purpose; i.e.,
f?(A, t) = T[x(A, t), tJ or T(x, t) = 9""[A(x, t), tJ.

We are now in a position to wpte the right-hand side of(43) in the form of
an integral, for
t(m, t; - )a"(m, t) + t(n, t; +)u(n, t) = i[T(n, t)u(n, t) - T(m, t)u(m, t)J
= i ( a(Tu) dx. (48)

m ax
Dropping the vector form of (43), since there is only one component present,
we can then write the balance of linear momentum as
Dv ) a(Tu)] (49)
[m [ pu ( Dt-f -~ dx=O.
Since the portion of the bar has been arbitrarily chosen, if the above integrand
is continuous, then the Dubois-Reymond lemma allows us to conclude the
following final form for the law of balance of linear momentum:
Dv a(Ta)
pu- =puf +--. (50)
Dt ax
STRAIN AND STRESS-STRAIN RELATIONS
In analyzing the motion of a bar, we normally consider as known (i) the

cross-section area u, (ii) the body force/, and (iii) appropriate initial data on
all the other variables. We have three unknown functions, the density p,
the displacement u (or equally well the velocity v = Du/Dt), and the stress
component T. On the other hand, we have developed but two basic differential
equations (33) and (50). Thus it would appear that our mathematical descrip-
tion of the motion is incomplete.
At the same time, our discussion of the physics of the situation is also
deficient. We have derived field equationsthat describe the universal physical
requirements of mass conservation and momentum balance. But we have in no
way incorporated any properties of the particular material from which the bar
has been made. Common experience shows us that two bars which are
identical in geometry and subject to the same loadings can vary considerably
in their subsequent motion if they are composed of different materials. This
deficiency in our analysis can be repaired through the introduction of a

constitutive relation, which in this case describes the particular nature of the
material by relating a measure of the distortion of the bar to the amount of
stress required to produce this distortion.
We must first assign a precise meaning to the notion of" distortion." To do
this, consider two sections that originally (at t = 0) were located at the nearby
locations A and A + AA. At time t these sections occupy the positions x(A, t)
and x(A + ~' t). The sections in question originally enclosed an element of
material having length~- At time t this material has a new length
x(A + AA, t) - x(A, t),
which can be approximated by
[We have used the relationship x(A, t) = A + U(A, t) of (9).]

A reasonable measure of distortion is
new length - original length

(51)
original length
In the present instance, to lowest approximation this measure is
(1 + auJoA)
~ - ~ au
S=
AA =-oA
Thus the axial strain, or simply the strain (ass is called), is given by
s(A, t) = oU(A, t). (52)

oA
Since higher powers of ~ have been neglected in approximating our
measure of distortion, one often speaks of the strain as a change in infinitesimal
length per unit original infinitesimal length. It will be observed later that the
task of defining an appropriate measure of strain in three dimensions is
more elaborate.
In forming a constitutive equation, we shall assume that the stress compo-
nent $"(A, t) is a function of the strain s(A, t). This makes concrete our
feeling that the more one wants the bar to elongate, the harder one has to pull
it. We shall limit ourselves to situations in which the initial state of the bar is
free of stress. Thus the graph of stress component versus strain might have
the appearance of the heavy line in Figure 12.8. In particular, the graph must
go through the origin.
I
Slope&
I
I
I
I
I
I
I
I
I
I
I
,.I
FIGURE 12.8. Plot of a "typical" relation between stress !i and

strain e.
We can make some progress without reference to experiments.* We reason

that for sufficiently small strain e, the graph of Figure 12.8 can be approxi-
mated by its tangentline at the origin. If the slope of this line is denoted by t!,
we thereby hypothesize that
!i(A, t) = tl(A, t) e(A, t). (53a)
More simply, we find that
I~=8=1~-1 (53b)
In Hooke's law (53b), the proportionality factor 8 is called Young'smodulus.

By the way, Robert Hooke (a seventeenth-century contemporary of Newton)
postulated the proportionality of stress and strain only for large scale or
global situations. In our pointwise or local version of Hooke's law, Young's
modulus may be a function of A and t-for the nature of the bar may vary
with these variables. We shall assume, however, that
e'(A, t) = 8(A, 0), (54)
where the initial composition of the bar, and hence the initial distribution
of the stiffness factor 8, is known.t
This is a dangerous procedure, but it is sometimes a necessary one. Perhaps only by guessing
a reasonable constitutive equation and then making some plausible and interesting theoretical
deductions can one convince an experimenter that it is worth his while to investigate certain
phenomena.
t Aging effects could be included by allowing 8 to be a known function of t as well as of A.
Sec. 12.J] Derivation of the Governing Equations
In spatial variables, Young's modulus is designated by E. That is, E(x, t) =

= 3 x 107 lb/in. 2 (2 x 1011 dynes/cm2) for steel and
G[A(x, t), t]. Typically, E
8 2 12
10 lb/in. (7 x 10 dynes/cm2) for aluminum.
We note that, by (21) with F= U and f= u, in spatial coordinates our
constitutive assumption is
ou/ax
T(x, t) = E(x, t) a /a (55)
I - u x
There is no doubt that it is a good approximation to replace the actual
stress-strain curve by its tangent line at the origin, provided that strains are
sufficiently small. The question is: Are there significant problems in which
such small strains are found? It turns out that there is a wide range of such
problems. To give one indication of this, consider the change in length of a
standard 39-ft railroad track. During the temperature change from winter to
summer an expansion of i in. could easily occur. The corresponding strain
is 0.00053. [If no provision were made by a gap between rails to allow for
this expansion, the resultant axial stress (approximately 16,000lb/in. 2) would
be more than sufficient to cause the rail to buckle out of its straight position.]
The above remarks should provide sufficient motivation to pursue the
consequences of our constitutive assumption. Using Hooke's law, we can
write the momentum balance requirement (50) as
p<1Dv = p<1f+ ~ (88 au'). (56)

Dt ax oAx
Alternatively, using (55) and the result v = Du/Dt, we have
D2u a E<1ou/ax
p<1Dt2 = p<1f+ ax 1 - au/ox. (57)
At this point, there are no obvious omissions in our basic equations.

Let us therefore review the situation to see whether the number of equations
appears sufficient.
In the momentum equation (57) we regard <1,E, and fas known, while
u and p are unknown. Thus from the initial area distribution, one can
determine <1as a function of u from (28). Similarly, from (54), Young's
modulus is a function of u determined from the original distribution of 8:
E(x, t) = 8[x - u(x, t), O]. (58)
Finally, the body force F(A, t) (typically gravity) will be regarded as equal
to its given initial distribution F(A, 0) so that
f(x, t) = F[x - u(x, t), O] (59)
can be regarded as known.
Again recalling that v = Du/ Dt, we see that the mass conservation equation
(33),
D(pt1) + pt1ov= O
Dt OX '
also contains the two unknown functions u and p. We are thus faced with two
equations in as many unknowns. Even for linear algebraic equations, one can
draw no definitive conclusions from the fact that there are as many equations
as unknowns. Nevertheless, we can draw the tentativeconclusion that we have
done a satisfactory job in formulating the governing differential equations.
INITIAL AND BOUNDARY CONDITIONS

One expects that a properly formulated problem will have a unique
solution. But in the simplest problems of mechanics, and here, too, uniqueness
cannot be expected unless initial conditions are prescribed. Different dis-
placements and density variations must certainly be expected from bars that
have different initial configurations. Thus we make the stipulation
u(x, 0) = u0 (x), v(x, 0) = v0 (x), p(x, 0) = Po(x), (60)
where u0 , v0 , and Po are given functions.
We must also make some statement about what happens at the ends of the
bar. This is certainly reasonable in the case of a bar of finite length or at
the finite end of a semiinfinite bar. The situation is perhaps not as clear in the
case of an infinite bar; but we have an intuitive feeling that if nothing is
said about what happens at "infinity," there may be serious difficulties in
trying to find a unique solution. We shall make no attempt to discuss this
uniqueness question at this time. Rather, the case of the finite end will be
examined in a few examples. We shall formulate boundary conditions in
material coordinates and shall leave to the reader the task of translation into
spatial coordinates.
Let us assume that the bar has a finite terminus at A = L. First consider
the case in which the bar is built in at the end, so that it is constrained from
movement. In such a case we refer to the boundary as a fixed end. The
boundary condition can then be stated as
U(L, t) = 0. (61)
[It is understood that (61) must hold for all positive t.] In contrast with the
case just treated, we can have the situation where no constraint whatsoever
is applied, a free end. Under these conditions, no force is applied at the end
to restrict the motion. At the end A = L, this requirement can be written,
using Hooke's law (53), as
8(L, t)S(L, t)U.iL, t) = 0, (62a)

where we have used subscript A to denote partial differentiation.
In most cases neither tS nor S will vanish at A= L, and (62a) can be

replaced by the simpler form
U,1.(L,t) = 0. (62b)
But S will vanish if the bar comes to a point at A = L. It must not be con-
cluded, however, that (62a) is vacuous in such an instance. We find frequently,
when S vanishes at A = L, that possible solutions of the equation of motion
can have singularities at that point. Equation (62a) then tells us that the
singularity cannot be so strong that the indicated product will not vanish.
For example, if S behaves like (A - L) near A= L, a purported solution U
that behaves like log (A - L) is ruled out by (62a).
The fixed end and the free end are, by all odds, the most common end
conditions met in practice. There are, however, a number of others that do
arise; and we shall consider two of them. These boundary conditions are a bit
more subtle in derivation, primarily because one must be careful with the
algebraic signs.
Let us consider a rigid mass fixed to the free end of the bar at A = L, as
shown in Figure 12.9. The mass Mwill undergo the same displacement U(L, t)
--A
FIGURE 12.9. A semiinfinite bar joined to a mass M.
as the end of the bar. Now let us isolate the mass itself. If we ignore the
effect of gravity, the force acting on the mass due to the bar is S(L, t)t[x(L, t),
t; - ]. In terms of the stress component,
t[x(L, t), t; - ] = -iT[x(L, t), t].
Applying Newton's second law to the mass M and replacing the stress com-
ponent through Hooke's law, we obtain the following boundary condition:
-S(L, t)C(L, t)U,t(L, t) = MU,r(L, t). (63)
As it should be, (62a) is recovered in the limit M-+ 0.

As our final example, we shall consider a linear spring attached to the free
end of the bar on one side and rigidly fixed at the other end (Figure 12.10).
As the bar undergoes various displacements, the spring will follow. But there
will be a lag, due to the inertia of the spring. In general, one would be faced
with a separate, difficult, problem for the spring. One would have to find a
x=O
FIGURE 12.10. Bar with one end built in and one end constrained by
a linear spring.
solution that matches the solution of the bar at the bar-spring junction.
To avoid this difficulty, we shall assume that the spring is so light that its
inertia is negligible. Under such circumstances, a displacement of the bar will
instantly cause a corresponding uniform stretching of the spring.
Consider a slice at the end of the bar, corresponding to initial coordinates
between L - fl and L. There is a surface stress t[x{L - fl., t), t; - ] acting on
the left face of this slice. In obtaining the force on the right face, we assume
that the spring is initially in an unstretched position. We denote the spring
constant by k 2 Then the magnitude of the spring force is Ik 2 U(L, t) j. We
assert that the actual force on the bar due to the spring is - k 2 U(L, t )i.
Indeed, if U(L, t) is positive, this force is leftward-which is the appropriate
direction for the force due to a compressed spring. On the other hand, if
U(L, t) is negative, the force on the end of the bar is rightward-which is
appropriate when the spring is extended.
The final step is to equate the rate of change of the slice's momentum with
the sum of the surface forces just discussed and the body force. Upon taking
the limit as fl -+ 0, one sees that the inertia terms and the body force are
negligible. Thus the boundary condition
2
S(L, t)8(L, t)U,4.{L,t) = -k U(L, t) (64)
emerges as the required balance between surface stress and spring force
{Exercise 15). Boundary conditions, such as (64), that involve a linear com-
bination of the unknown function and its first derivative are frequently
referred to as impedance boundary conditions.
LINEARIZATION
A well-formulated problem seems to be offered by the governing equations
(57) and (33), the initial conditions (60), and a pair of boundary conditions
of the type we have just derived. But the equations are nonlinear, so that, in
general, numerical or advanced analytical techniques are required to obtain
useful exact or approximate solutions. Such techniques are out of place in
this chapter, which is supposed to serve as a relatively simple introduction
to continuum mechanics. Without further ado, therefore, we shall linearize
the equations and boundary conditions.
The basic fact which makes linearization worthwhile is that in metallic

materials, even a very large stress will produce but a small strain. We shall
therefore systematically simplify all expressions under the assumption that
the displacement u and its partial derivatives are small compared to unity.
We shall also assume that the density at a particular location x does not
vary much from its initial value.
Consider u(x, t), the cross-section area of the section now (at time t) located
at position x. Recall that we assumed that the area of any given section does
not change with time; i.e., S(A, t) = S(A, 0). From this we deduced in (28)
that
u(x, t) = S[x - u(x, t), O].
Since u has been assumed to be small, we should be able to make the approxi-
mation
u(x, t) ~ S(x, 0). (65)
The effect of this approximation is to replace the area of the section that
is now x units from the origin by the area of the section which was initially x
units from the origin. When displacements are small, such an approximation
indeed seems reasonable.
A glance at the difference between the exact and approximate cross-
sectional areas clarifies the nature of our approximation. By the mean value
theorem we can write
IS(x,0)-S(x-u,O)I = luSx(~,O)I, (66)
where ~ is between x - u and x. Assuming that the derivative S" is bounded
in the neighborhood of x, the right-hand side of (66) can be made as small as
desired by taking u sufficiently small. The bigger the derivative, the smaller
one has to take u to achieve a given level of approximation. This is to be
expected; the more rapidly the cross-section area changes, the bigger is the
error made in misestimating the correct location to measure this area.
With exact and approximate values of the same quantity both present,
some notational difficulty is experienced. It seems best to use the function
of a single variable u(x) to denote our approximation to u(x, t). A "hat"
will be employed to signify the difference between the exact and approximate
values. Thus we shall write
u(x) = S(x, 0), h(x, t) = cr(x, t) - u(x). (67a, b)
With this notation, our approximation (65) is written

o'(x, t) ~ cr(x) [i.e., 1h(x, t) I ~ Iu(x) I]. (68)
Remember-cr(x) is a known function, the initial area of a section x units

from the origin.
In (58) and (59) we have exact analogues of (28) so that we can treat the
Young's modulus E(x, t) and the body forcef(x, t) in exactly the same way
as we treated u(x, t). We approximate the present value of each quantity at x
by the known initial value of that quantity at x. * Thus we define
E(x) =8(x, 0), (x, t) =E(x, t) - E(x); (69)

f(x) =F(x, 0), ](x, t) =f(x, t) - f(x). (70)
We assume that
12{x, t)l ~ IE(x, t)I, lf(x,t)I ~ lf(x,t)I, (71)

so that
E(x, t) ~ E(x), f(x, t) ~ f(x). (72)
We have said all along that a(x, t), E(x, t), and f(x, t) were regarded as
known functions, although (before approximation) the argument of the
functions contained the unknown displacement u. Previously, p(x, t) was
regarded as unknown, but this function, too, must be approximated by its
known initial value. Otherwise, the term p(x, t)u(x, t) D 2 u/Dt 2 in the momen-
tum equation (57) will remain nonlinear. Approximation of the density
requires extra care, because the assumption
c5(A,t) = c5(A,0)
is inappropriate. If this assumption holds, the material is incompressible
and the only one-dimensional motions are uninteresting rigid motions
(Exercise 11).
Although c5(A,t) =/:-c5(A,0), we still write
p(x) =c5(x,0), P(x, t) =p(x, t) - p(x), (73)
and assume that
IP(x, t)I ~ lp(x)I so p(x, t) ~ p(x). (74)
If we approximate the difference P by the lowest order terms in the appro-

priate Taylor series, we find [Exercise 18(a)] that
IP(x,t)I z luc5,.-tc5,I. (75)
Knowledge that the displacement u is small, then, is not sufficient to guarantee

that IP I is small. It could be that the action of small displacements produces
continual compression or rarefaction that would eventually lead to a significant
Roughly speaking. we ignore the difference between material and spatial coordinates.
Sec. 12.1) Derivation of the Governing Equations 371
deviation of the density of a section from its initial value. As we see from
(75), this will not happen if the motion has been proceeding for a sufficiently
short time (t sufficiently small).
As the final link in our chain of approximations, we recall that the velocity v
is given by
Du
V = Dt = Ut + VUx.
Solving for v, we find that
ut
v=-- (76)
1- Ux
Since we are assuming that u and its derivatives are small compared to unity,
we can make the approximation
au(x, t)
v(x, t ) :::::-a--. (77)
1
Note that the assumed smallness of u, implies
lv(x, t)I ~ 1. (78)
As is easily verified (Exercise 19), if we retain only lowest order terms we

obtain the following linearized version of the momentum balance equation
(57):
2
a u = p(x)a(x)f(x)
p(x)a(x) at a [ E(x)a(x) au]
+ ax ax , (79)
2
The continuity equation (33) can be written in the form
a
at [p(x, t)a(x, t)]
a
+ v ax [p(x, t)a(x, t)] + p(x, t)a(x, t) ax = 0.
&
The first term can be rewritten as follows:
f,[p(x, t)a(x, t)] = :t {[p(x) + P(x, t)][u(x) + lt(x, t)]}.
Again retaining the lowest order terms, we obtain the following linearized
version of the continuity equation :
a
ot [u(x)P(x, t) + p(x)lt(x, t)]
a
+ v ox [p(x)u(x)] + p(x)u(x) ox=
av 0. (80)
Equation (79) contains but a single unknown function, the displacement

u(x, t). Appropriate boundary conditions are linearized versions of (61)-(64).
These versions are as follows [Exercise I 6(b)]:
Fixed end: u=O. (Sia)

Free end: uEu:i:= 0. (Sib)
Mass-loaded end: Mu,, + uEu:i:= 0. (Sic)
End constrained by light spring: uEu:i:+ k 2 u = 0. (Sid)
Each boundary condition is to be imposed at the (known) initial position of

the appropriate end.
No TE. In the above equations, and in the remaining equations of this
chapter, if u, E, and fare written without arguments, the known functions
u(x), E(x), and/(x) are implied.
It would appear that a self-contained linearized mathematical model for
one-dimensional deflection of a bar is composed of the differential equation
(79), a boundary condition such as those of (81) at each end of the bar, and
the initial conditions
Iu(x, 0) = u0 (x), u,(x, 0) = v0 (x), I (82)
What, then, is the role of the continuity equation (80)? The answer to this
question is a little easier to see if we introduce the mass per unit length
p(x, t) and its approximation p(x):
p(x, t) = p(x, t)u(x, t), p(x) = p(x)u(x). (83)
To lowest order
p(x, t) - p(x) = [p(x) + fJ(x, t)J[u(x) + h(x, t)] - p(x)u(x)
:::::p(x)h(x, t) + fJ(x, t)u(x)
so that the linearized continuity equation (80) can be regarded as stating
o_ _ o
ot [p(x, t) - p(x)] :::::- ox [v(x, t)p(x)u(x)]. (84)
We see, then, that in linear theory the continuity equation provides a first
correction to the difference between the actual and initial values of the mass
per unit length. This correction is easily computed once the velocity v :::::
u, is
found. The correction is not of major interest in most applications, however,
and the continuity equation is usually ignored in linear theory. Nevertheless,
a careful worker would at least check to see that the predicted correction
is indeed small, so that the approximate analysis has the appearance of
consistency.
373
EXERCISES
1. (a) Verify (8).
(b) Verify (21).
(c) Verify (42).
2. Consider a motion described by
x=A + t(A - t),
(a) Describe the motion in qualitative terms.
(b) Find U(A, t). Find u(x, t) from the formula u = x - A. Verify
that ul...= U.
(c) Find
oV(A, t) ov(x, t)
V(A, t), v(x, t),
ot at'
(d) Make the substitution x = x(A, t) in the second and fourth
expressions. Compare the first pair; the second pair. Discuss fully
in terms of the material derivative.
3. Consider a motion described by
x =tA +tAe',
(a) Find the material variable A as a function of x and t.
(b) Calculate U(A, t) and u(x, t).
(c) Find the velocity V(A, t) in material coordinates and v(x, t) in
spatial coordinates.
(d) Check the Euler expansion formula for this case by direct compu-
tation.
4. Establish the following differentiation formulas:
(a) D(f g)/Dt = Df/Dt Dg/Dt.
(b) D(fg)/Dt =f(Dg/Dt) + g(Df/Dt).
(c) Does the analogue of the usual formula for the derivative of a
quotient hold?
5. Consider a bar in the form of a truncated right circular cone. It is of
length 1 and has radius r 1 at one end and r 2 at the other. The bar is
assumed to undergo the motion described in Exercise 2.
(a) Find the cross-section area S(A, 0) at time t = 0.
(b) Let u(x, t) denote the cross-section area at the position x, attime t.
Find u(x, t) and calculate Da/ Dt. How does this compare with
oS(A, t)/01?
l6, Derive the equation for mass conservation by a direct attack on the
integral
d [(N,t)
-d p(x, t) dx, P= pu.
X x(M,t)
7. (a) Show by use of the procedure and notation of the text's discussion
of mass conservation that
~ r"<r>
g dx = ("<'>
[Dg +g ov]dx.
dt J111(1) )111(1)
Dt OX
(b) Use integration by parts to derive the one-dimensional version
of the Reynoldstransporttheorem:
~ [(t)g dX = [(I)09 dX + gvln(I)
dt 111<1> ar
111<1> 111<1>
(c) Interpret the result of (b). (In three dimensions the interpretation
is the same; see Section 14.1.)
8. If v(x, t) is independent oft, the motion is called steady.
(a) If you are familiar with first order partial differential equations,
show that the most general steady motion has a displacement that
can be written in the form u(x, t) = x - F[t - g(x)] for some
functions F and g.
(b) What can be said about the other quantities of interest?
9. Show that the mass conservation equation in material coordinates can
be quickly obtained in the form
a [c')(A,t)S(A, t)J(A,
ot t)] = 0.
Deduce the alternative form
c')(A,t)S(A, t)J(A, t) = c'){A,O)S(A, 0).
=
[Since S(A, t) S(A, 0) the cross-section area factor can be canceled.]
10. In the text's discussion of mass conservation, it is Mand N, not m(t)
and n(t), which are arbitrary. Nevertheless, (33) can be concluded
directly from (32). Show this.
11. Show that (39) implies that x(A, t) =A+ f(t) for some function f
Interpret this result. Is it reasonable?
t12. Find the " action-reaction " equation by the " thin-slice " approach.
Consider a section of the bar of width /u with the end forces as
indicated in Figure 12.l l. Apply the momentum conservation equation
and then let /:,.x-+ 0.
x x+Ll.x
FIGURE 12.11. Surface forces on a thin slice of bar with cross-

section area er.
375
13. If we wish to be cavalier in our derivation of the approximate equation

of motion (79), we can use a "thin-slice" approach and linearize as we
go along. Carry out this procedure in the following manner. The forces
acting on a small slice of the bar of width .6.x are as shown in Figure
12.12.
(a) Apply Newton's second law of motion to the slice.
(b) Find the limiting form of the equation when .6.x~ 0.
(c) Use Hooke's law.
---~fpal:!.x
---Ta+ o(Ta) l:!.x
ox
x x +l:!.x
FIGURE 12.12. Body and surface forces on a thin slice of bar.
14. The Eulerianstrain*is defined as the lowest order approximation to the

quotient
present length - original length
present length
Find an expression for the Eulerian strain in spatial coordinates. Use
two different approaches.
(a) Consider a section of bar that is now (at time t) located between
x and x + .6.x.
(b) Consider a section of bar that was initially located between A ana
A+ L\A. Approximate the quotient; then change to spatial
coordinates.
15. Complete the derivation of (64).
16. (a) Transform the boundary conditions (62a), (63), and (64) into
conditions on u.
(b) Linearize to obtain (81b-d).
17. Gravity was neglected in obtaining boundary condition (63). Under
what circumstances is this justified?
18. (a) Verify (75).
(b) What is the relation between the function p 0 (x) of (60) and the
function p(x) of (73)?
Euler's name is associated with spatial coordinates, Lagrange's with material. Consequently,
Euler's name is invoked when the change in length is compared to the present length, and Lagrange's
when comparison is made with the original length.
19. Show that our various assumptions give (79) as a first approximation
to the momentum balance equation (57).
20. Show that (84) has a very reasonable physical interpretation. You may
find this easier if you integrate between fixed limits x = a and x = b.
21. Show that a -!-in. elongation of a 39-ft steel rail is equivalent to a
strain of 0.00053. Show that the corresponding stress is about 16,000
lb/in. 2 (This verifies a statement in the text.)
22. For a certain spatial region, and in a certain time interval (0, t 1), the
one-dimensional motion of a bar is described by x = (A - t)/(1 + At).
(a) Find A(x, t) and verify that A[x(A, t), t) = A.
(b) Find a formula for the velocity, as a function of time, of the cross
section that was initially located 3 units to the right of the origin.
(c) The density is given in material coordinates by the formula
'5(A, t) = A 2 Find an expression for the density that would be
measured at a fixed point located one unit to the right of the origin.
23. Let us employ J(A, t) and j(x, t) to denote the Jacobian in material
and spatial coordinates, respectively. Recall that
oJ(A,
t) = J [ov],
Ot OX ..t
Dj =j
Dt
(av)
ox
The goal here is to find a formula for D2j/Dt 2 Do this in two ways,
and show that the answers are the same.
(a) Proceed in spatial coordinates and use the "natural" formula for
the material derivative of a product.
(b) Find o2J/ot2 in material coordinates and then switch to spatial
coordinates at the end. Use very careful and explicit notation.
12.2 One-dimensional
Elastic WavePropagation
In this section and the following, we shall exclusively treat the linearized
momentum balance equation (1.79). We shall consider, primarily, the case of
constant physical properties. Consequently, the governing equation is the
one-dimensionalwave equation. Many of our conclusions concerning this
special problem will be generalized to more complicated situations throughout
the course of the text. It should be recalled that the linearization process
told us essentially that to the order of magnitude that has been retained,
there is no distinction between material coordinates and spatial coordinates.
Furthermore, the density p, cross-section area u, and Young's modulus E
can be taken as known functions of position only. For reference, we repeat
the linearized momentum equation (1.79):
[u(x)E(x)ux(x, t)]x - p(x)u(x)u 1t(x, t) + p(x)u(x)f(x, t) = 0. (1)

Sec. 12.2) One-dimensional Elastic Wave Propagation
377
THE WAVE EQUATION
We shall begin by considering the important case that occurs when the
material and geometric properties, u, E, and p, are independent of x and
the body force,/, vanishes. Equation (1) then reduces to
p
u,.,.- E Utt= o. (2a)
It is frequently convenient to introduce the constant c, the speed of sound, given

by
2 E
c =-
p
Equation (2a) can then be written as
u,.,.- c- 2 utt = 0 or Utt= c2u,.,., (2b)
a differential equation that is normally called the wave equation. We shall see
shortly that these descriptions of c and (2a) or (2b) are indeed apt.
GENERAL SOLUTION OF THE WAVE EQUATION
In order to find the solutions of (2b), we shall try to ascertain an appropriate
change of variables. Let us examine what we can do with what is perhaps
the simplest such change, namely, a linear transformation of x and t. We
shall set
and 'I= yx + ()t,
where the constants, oc,p, ')',and o will be determined so as to simplify the
resulting differential equation. Direct application of the chain rule of partial
differentiation then yields (Exercise 1)
(oc2 - p 2 c- 2 )u(( + 2(ocy- ()pc- 2 )u(., + (12 - ()
2
c- 2
)u,,,,= 0. (3)
As we shall see, the most helpful selection of the constants in the linear
transformation is one for which the coefficients of the terms containing u((
and u.,,,
vanish. One such selection is to set
oc= pc- 1 and 1 = -()c- 1

(There are other choices that will accomplish the same end, but the examination
of them is left as an exercise.) The transformation of variables now becomes
~ = cc(x+ ct), 'I= y(x - ct).
We note that neither ccnor 1 can vanish if the transformation is to be non-
degenerate. The coefficient of the u('l term can now be simplified, since
2
ccyc
ccy- ()pc- 2 = ccy+ -c2 = 2ccy.
Thus (3) can be reduced to
or, more simply, to*

~=Q ~
The general solution to (4) can now be readily obtained, for the equation
states that u~ is independent of 'I Consequently, we may integrate directly
with respect to 'I, with the result
u~= F(e),
where Fis an arbitrary function. Integrating once more, we find
u= JF(e) de + G(,,),
where G is an arbitrary function.
We have tacitly assumed that Fis integrable. In fact, for future purposes
the arbitrary functions that appear in the solution should be twice differenti-
able but otherwise arbitrary. Since F can be chosen at liberty, subject to
these restrictions, we can equally well write
u = H(e) + G(f/). (5)
In terms of the original variables, x and t, (5) becomes
u(x, t) = H[a.(x + ct)] + G[y(x - ct)].
Since H and G are arbitrary, we may write finally
u(x, t) = f(x - ct) + g(x + ct), (6)
where f and g are arbitrary functions. Equation (6) is the general solution
of (2b).
PHYSICAL SIGNIFICANCE OF THE SOLUTION
Before considering special forms off and g, let us examine the physical
significance of (6), concentrating on the first termf(x - ct). Figure 12.13 is
a typical plot of f(x - ct) as a function of x with t considered as a parameter.
The solid curve corresponds to t = t 0 and the dashed curve tot= t 1 (Only a
portion of the curves are graphed.) We see that the second curve is obtained
from the first simply by translation to the right through a distance c(t 1 - t 0 ).
Thus f can be considered as a fixed pattern or wave that moves to the right
a distance c(t 1 - t 0) in a time interval t 1 - t 0 In other words, the wave moves
to the right with the.fixed speed c.
* The reader who has some acquaintance with the theory of partial differential equations will
recognize that what we have just done is to reduce the original equation to canonical form by
introducing the characteristic coordinates g and T/
Sec. 12.2] One-dimensional Elastic Wave Propagation
379
FIGURE 12.13 A wave traveling to the right with speed c.
An important and useful special case occurs when the functions f and g
are taken to be sine and cosine functions. With such a choice, a solution could
be written in the form
u(x, t) = A sin a(x - ct)+ B cos a(x - ct)
+ E sin a(x + ct) + F cos a(x + ct). (7)
SOLUTIONS IN COMPLEX FORM

It is frequently more convenient to replace the trigonometric expression
by the complex exponential through the use of the relation
exp (i1:)= cos -r + i sin -r,
1 2
where i = (-1) ' Equation (7) can then be rewritten as
u(x, t) = .Silexp [ia(x - ct)] + t?Jexp [ia(x + ct)], (8)
where the coefficients .Siland t?Jare generally complex (Exercise 4).
We shall digress for a moment to discuss complex solutions. First of all,
note that any linear partial differential equation in x and t, say
OU o2 u
OU
L(u) = a + a1 ox+ a ot + a3 ox2 + = 0
0 2 (a; are real constants)
has solutions of the form

u = f(x, t), f(x, t) = .Silexp (ax) exp (/3t).
(Here any of the constants .Siland a and Pmay be complex.) For
L{f} = .Silexp (ixx) exp (Pt)[a0 + a 1ix + a2 {3+ a31X2 + ]
and a solution will result if the polynomial in the square brackets vanishes.
Second, since L is linear and has real coefficients,
L(Re f) = Re L(f) = 0, L(Imf) = Im L(f) = 0.
(In the above equation, Re denotes "real part of" and Im "imaginary part
of," as usual.) We therefore see that both Ref and Im/provide real-valued
solutions of L(u) = 0.
Not all constant coefficient linear equations have solutions that are purely
sines or purely cosines, but all have solutions of exponential form. One can
always work with linear combinations of sines and cosines (or sinh and
cosh) instead of exponentials, but the latter often provide more compact
calculations.
ANALYSIS OF SINUSOIDAL WAVES

Let us consider one of the terms in the solution (7) in some detail. In
particular, let
2n(x - ct)
u(x, t) = a cos A , a>O. (9)
Since the cosine is bounded between + 1 and -1, the coefficient a is the
maximum of the absolute value of u(x, t) and is called the amplitude.Further-
more,
2n(x + A.- ct)
u(x + A, t) = a cos A.
2n(x - ct) } 2n(x - ct)

= a cos { A. + 21t = a cos A
= u(x, t).
We can interpret this result in the following manner. If we regard t as fixed,
then u(x, t) as a function of x repeats itself in a distance A..In other words,
u(x, t) is a periodic function of x of period A. This "spatial period," A, is
called the wavelength; and its reciprocal, k = .1.-1, measures the number of
waves in a unit distance and is known as the wavenumber.
Similarly, we can regard x as :fixedand ask for the time 't' at which the wave
will repeat itself. This requires that
2 2
u(x, t + -r)= a cos ( ; (x - ct) - :C't']
= u(x, t);
from which it follows that
2nct = 27t
A
or
A
t =-
c
(10)
We call -rthe period and its reciprocal n, n = c/l, the frequency. The frequency
thus measures the number of waves passing a given point x in a unit time.
Now let us consider two solutions, u1 and u2 , given by

u1 = a cos 21t(kx - nt),
u2 = a cos [21t(kx - nt) + e]
= a cos 2n[k(x +2 :J- nt].
It is seen that u2 is the same as u1 except for a shift in x in the amount of

e/21tk = el/2n. The parameter e is known as the phase of u2 relative to u1
In the particular case in which e = n, u2 = -u 1 ; we refer to this situation
by saying that u1 and u2 are 1t radians (or 180 degrees) out of phase.
We would like to point out finally that solutions of the form
u(x, t) = exp [21ti(kx - nt)]
= exp (2nikx) exp (-21tint)
are not as particular as it might first appear. More general time dependence
can be synthesized as a series or integral through Fourier methods. For
example, consider the sum of sinusoidal waves which are congruent but
which travel in opposite directions
sin 21t(kx - nt) + sin 21t(kx + nt) = 2 sin (21tkx) cos (21tn~).
The result is a pattern that rises and falls between fixed nodes at x =m/k,
m = 0, l, .... This is a stancling wave. Rather general functions of x can
be similarly expressed in terms of complex exponential functions. Although
the general line of attack is rather straightforward, there are numerous
situations where the resulting mathematical problems can be both intriguing
and challenging. See Chapter 4.
EFFECTS OF A DISCONTINUITY IN PROPERTIES
Let us now consider a very simple example that makes use only of expo-
nential-type solutions and yet has some interesting conclusions. We take as
our configuration two semiinfinite bars, I ( - oo < x < 0) and II (0 < x < oo).
Both have constant Young's modulus and density, the first with values E1
and p1 and the second with values E2 and p 2 Both parts have the same shape
and cross-section area. The velocity of sound in I will be denoted by c1 and
in II by c2 The two semiinfinite bars are assumed to be welded together at
x = 0 (Figure 12.14).
Considering a right-moving incidentwave
exp [i(x - c1t)J
in bar I, we shall try to determine both the ensuing transmitted wavein bar
II and any additional reflected wavethat might be produced by the presence
of the interface.
FIGURE 12.14. Two bars welded together. The bars differ in their
values of Young's modulus E and density p, and therefore in their sound
speed c=(E/p) 112
In order to carry out the required calculations, we must first translate the
conditions at the interface into analytic terms. Let us denote the displacement
in I by u1 (x, t) and in II by ui(x, t). Since the two bars are welded together,
the displacement must be continuous across x = 0. Thus
ui(O-, t) = ui(O+, t), (11)
where O-- implies a limit where x approaches the origin but always remains
to its left, whereas O+ implies a limit where x approaches the origin but
always remains to the right.
In addition, the force must be continuous across the interface (Exercise 13).
Noting that the cross-section area of the two portions are the same, this
requirement implies that
Ei [oui(~:, t)] = 2
[ OU2~:,t)]. (12)
We shall look for a solution in which u1 is composed of the incoming wave

together with a reflected left-moving wave and u2 is a right-moving wave.*
In other words, we try for solutions
ui = exp [i(x - Cit)]+ A exp [icx(x+ cit)],
u2 = B exp [iP(x - c2 t)].
The constants A, B, ex, and P will now be determined from the interface
conditions (11) and (12).
We derive from (11) the requirement that
exp (-icit) +A exp (icxcit)= B exp (-iPc 2 t)
for all values of t. In order for this to be true, the arguments of all the ex-
ponentials must be the same. Thus
(X = -1 and
An effective way of solving simple problems is, as here, to guess the form of the answer.
Expertise in this solution method, as with others, comes with practice.
In addition, the coefficients must satisfy
1 + A =B. (13a)
Taking into account the conditions on the exponentials, we may write
Introducing these results into boundary condition (12), we find that
{13b)
Equations (13a) and (13b) can now be solved for the amplitudes with the
result [Exercise 5(a)]:
1-R 2
A--- B=-- (14)
-1 +R' 1 +R'
The expression for the amplitude A of the reflected wave shows us that there
will be a phase shift in this wave of 1t if R is greater than unity.* As should be,
the case when R = 1 (uniform bar), the transmitted wave is identical with
the incident and there is no reflection.
The contemplation of limiting cases almost always increases one's under-
standing of a solution's implications. Often one can view the general case as
some sort of compromise between comparatively understandable extremes.
In the present instance we note first that as R ___. oo, A___.-1 and B ___.0.
For R ~ 1, the situation is that of a wave in a relatively light flexible bar
impinging on a relatively heavy stiff bar. In the limit, the latter bar remains
motionless. Taking the real part of the limiting incident solution
ui = exp [i(x - cit)] - exp [-i(x + cit)], (15)
we obtain
cos (x - cit) - cos (x +cit)= 2 sin x sin Cit. (16)
This is a standing wave. The displacement vanishes at the discontinuity x = 0

just as if the light flexible bar were rigidly fixed at that point. Indeed, the
limiting solution for x < 0 is identical to that for a wave in a semiinfinite bar
Since A and B are always real, there is no advantage in using complex notation. Cosine
solutions or sine solutions could have been assumed from the outset. See Exercise 16 for an
instance when this is not the case.
which is built in at one end [Exercise 7(a)J. This is no surprise, at least in

retrospect.
Note that the imaginary part of (15) is
(17)
which differs from (16) only in phase.

In the other limiting case R -+ 0, the reflected wave is the same as would
be obtained if the bar along x < 0 had a free end at x = 0. The transmitted
wave has the maximum possible amplitude. Details of the discussion are left
to the reader [Exercise 7(c)].
It is illuminating to consider the distribution of energy in this problem.
The required generalizations of energy concepts from particle mechanics
are provided in Section 12.4. With these one can compute the energy per unit
length in the various waves. Since no dissipative mechanisms are present,
one expects that the energy in the incoming wave will just balance the sum of
the energies in the transmitted and reflected waves. This is, in fact, the case
(Exercise 4.6).
One more observation is required. In Section 12.1 we referred several
times to a mathematical problem in which the governing equation (or
equations) was supplemented by appropriate initial conditions (at time t = 0)
and boundary conditions (at the twoends of the bar). The problem we have
just solved, however, was not of this initial-boundary type. Rather, it was of
the scattering type, which is commonly encountered in wave motion problems.
Here an incoming signal is prescribed for all time, and the resultant effect is
sought. To make sure that only the prescribed incoming signal is "causing"
the phenomenon, one must (as here) require a radiation condition. Roughly
speaking, this is
total solution minus prescribed incoming solution is entirely outgoing.
(18)
EXERCISES
1. Carry out the detailed calculations required to derive (3).
2. In the discussion below (3), we made one selection for the parameters ex
and y. Consider all the other choices that would make the appropriate
coefficientsin (3) vanish and determine which ones will lead to meaning-
ful solutions. Do any of these additional choices yield essentially new
solutions?
3. In the text we analyzed the geometrical significance of the solution
f(x - ct). Carry out a similar analysis for g(x + ct). What would have
happened if we had chosen the solutionf(ct - x)?
4. (a) Writing .!II = M + iN, fJI = P + iQ, determine M, N, P, and Q in
terms of A, B, E, and F so that (7) and (8) are equivalent.
(b) Also show that a solution of the form

u(x, t) = A sin cx(x- ct) + B cos cx(x- ct)
can be written in the form
u(x, t) = H sin cx(x- ct + e)
as well as
u(x, t) = K cos cx(x- ct + <5).
What is the corresponding exponential form?
5. (a) Derive the expressions for the coefficients A and B given in (14).
(b) Redo the problem, using dimensionless variables from the outset.
6. In imposing the displacement continuity requirement (11), we asserted
that all the exponentials must have the same argument if the appropriate
equation was to be valid for all t. Supply a formal proof for this asser-
tion.
7. (a) Suppose that a semiinfinite uniform bar occupying - oo < x ~ 0 is
subject to the incident wave exp [i(x - ct)]. Find the reflected wave
when the bar has a fixed end at x = 0.
(b) Repeat part (a) in the case when the bar has a free end at x = 0.
(c) Provide a full discussion of the solution given by (14) when R ~ I.
8. In the case of the bar composed of two media, suppose that an incoming
wave is given by the known function
f(c 1t- x).
Find the reflected wave g(c 1 t + x) and the transmitted wave h(c 2 t - x).
How does the behavior depend on the parameter R?
9. Consider a semiinfinite uniform bar occupying - oo < x ~ 0, carrying
a mass Mat the finite end x = 0. It is subject to a right-moving incident
wave f(ct - x), which has the property that
f(z) = 0 whenever z ~ 0.
Writing the displacement in the form
u(x, t) = f(ct - x) + g(ct + x),

find the reflected wave g(ct + x). After obtaining the general solution,
examine the special cases that occur when M is allowed to approach
zero and when Mis allowed to approach infinity. In the latter case,
calculate the force at x = 0 that would be produced by the incident wave
f(ct - x) alone, and compare it with the force produced at this location
by the complete solution. (You will find that the force is doubled.
Failure to anticipate this effect could lead to structural failure.)
10. If a bar of uniform properties is immersed in a viscous medium that

imparts a resisting force proportional to the velocity Ur , the equation
of longitudinal motion of the bar is
1
Uxx = 2 (uu + kur), (19)
c
Derive (19), using the approach of Exercise 1.13. Explain the constant
k and determine its units.
11. Show that if we look for a solution of (19) of the form
u(x, t) = exp ( -tkt)v(x, t),
then v(x, t) will satisfy
1(
v"" = 2 V 11 - !k 2 v). (20)
c
Furthermore, if k 2 ~ 1 in (20), approximate solutions can be found by
dropping the last term. Show that such approximate solutions can be
written in the form
u(x, t) = exp (-tkt)[f(x - ct)+ g(x + ct)]. (21)
Solution (21) is known as an attenuated wave.
12. Study the appropriateness of the approximation described in Exercise 11
by solving the following specific initial value problem
x
u(x, 0) = 0, ui(x, 0) = cos - ,
c
first using (20) as it stands and, second, approximating (20) by
(22)
How do the two solutions compare? Proceed by using separation of

variables (Sections 4.1 and 15.4); i.e., write
v(x, t) = h(x)m(t),
from which
where ,l..is a constant. Use the same technique on (22).

13. The derivation of boundary condition (12) was omitted. Supply an
argument for this condition based on a "thin-slice" analysis similar
to that used in the Exercise 1.12.
14. Consider the longitudinal motion of a uniform bar governed by
u(x, 0) = f(x), u,(x, 0) = g(x), 0 < x <L,

u(O,t) = u(L, t) = 0, t > 0.
(a) Describe the physical meaning of these initial and boundary

conditions.
:j:(b) Derive the solution
1 ( 1)
u(x, t) = 2 [f(x +ct)+ f(x - ct)]+
2
c J,,-ct
x+ct
g(s) ds
of the wave equation subject to these boundary conditions.

(c) The functions f(x) and g(x) are defined only for O < x < L. How-
ever, using/ and gin the solution written above requires that/ and
g be defined for all values of their arguments. Use the boundary
conditions to argue that/ and g should be extended as odd functions
of period 2L.
15. (a) Use the results of Exercise 14 to examine the following case. The
initial velocity g(x) = 0 and the initial displacement is given by the
graph shown in Figure 12.15. That is, f = 0 except for a small
neighborhood about x 0 Examine how this " bump" propagates.
Draw a sketch in the (t, x) plane of the regions where the bump inf
has an effect. For example, what is the significance of the line
x- ct= x 0 ?
(b) Perform a similar analysis when the material in a small region
around some point x 0 is given an initial velocity, but no point of
the bar is given an initial displacement.
FIG URE 12.15. Graph of a bump in displacement.

16. Consider a semiinfinite uniform bar occupying - oo < x s 0. Suppose

that a light spring, with spring constant k 2 , attaches the bar to x = 0.
We wish to examine the reflection of an incoming sinusoidal wave and
thereby to explore the possible advantages of complex notation. We note
that our problem is to determine a solution of the wave equation which
satisfies
k2
u' + KU =0 at x = 0, K=-
uE
t(a) Look for a solution of the form
u= ei(ct-x) + Qei(ct+x)
and determine Q. Take the real part of the final answer, and write
it in the two ways
u = cos (ct - x) + excos (ct+ x) + p sin (ct+ x) (23)
and
u = cos (ct - x) + B cos (ct+ x + <>).
(b) Look for a solution of the form (23) and compare the amount of
work with that required in (a) to obtain such a solution indirectly.
17. Show that a particular solution of the forced wave equation
u""'- (: 2 )urt = F(x, t)
is given by
1
u(x, t) =-
2C
f J_
t f+c(t-i)
O x-c(t-i)
F(x, l) dx di.
18. (a) Use the concepts discussed in Chapter 6 to provide a more careful
treatment of Exercise 11.
(b) Use perturbation theory to compute a correction to the approximate
solution found in Exercise 11.
12.3 DiscontinuousSolutions*
We have tacitly assumed in our discussion of particular examples in the
preceding section that the solutions under study had all the analytic prop-
erties-continuity, differentiability, and the like-that one might desire.
We shall now briefly consider what can be said if some of these restrictions
Sections 12.3 and 12.4 can be read in either order.

Sec. 12.3] Discontinuous Solutions
are removed. In particular, we shall look at "mild" discontinuities, which

will be characterized more explicitly in the course of the discussion. It should
be noted at the very outset that the displacement, u(x, t), will be assumed to be
continuous, for we shall rule out motions that characterize tearing or folding
of the material. Discontinuities of certain types will be allowed in the deriv-
atives, however.
Investigations of discontinuous solutions arise naturaJly in the study of
the motion of a bar. For example, one can ask what happens when the bar is
suddenly struck at one end. This will no doubt produce a jump in the stress
in the bar and very likely a jump in the velocity. Will an impulse imparted at
one section be felt over the entire length of the bar? If so, will this occur
instantaneously or will it take a finite length of time for the section at a
particular location to "hear" the disturbance? Can the disturbance propagate
at any speed and how will the properties of the bar affect the disturbance as
it moves along? We shall see that it is possible to answer some of these
questions with rather elementary methods.
The reader who is familiar with the elements of the theory of second order
partial differential equations will recognize that what we shall be doing is,
in disguised form, analyzing the characteristics and the equations that hold
along them. The language to be used, however, is somewhat different from
that normally employed in purely mathematical discussions.
It should be noted that we have already studied one example that can be
thought of as having a form of discontinuous solution. We were concerned
in Section 12.2 with two semiinfinite bars, with different properties, welded
together at a fixed position, x = 0. This situation may also be interpreted as
one infinite bar having a jump in section properties at x = 0. Boundary
condition (2.12) states that the stress is continuous across the fixed material
discontinuity. Since E jumps from E 1 to E 2 across the section x = 0, it
follows that the displacement u(x, t) has a jump in the derivative ux(x, t).
Consequently, this problem may be considered as one in which the solution
displays some form of discontinuity. We shall be concerned in the following,
however, with somewhat more general situations in which the location of the
discontinuity is not fixed in space and the material properties can vary from
point to point.
One final remark should be made before proceeding with the specific
analysis. We have blithely talked about waves and discontinuities propagating
along the bar; and yet we have taken, as a paramount assumption, that the
motions of the sections, or particles, are small. There is no inconsistency,
however, for it is not the particles constituting the cross sections that propa-
gate along the bar but rather certain properties, or functions of space and
time, which do so. For example, if we note the distribution of stress in the
bar at one instant of time and then at a later instant, we may find that the
graph at the later instant appears to be more or less the same as that at
the earlier time save for a shift in position. Thus we would conclude that
the stress distribution has propagated along the bar with perhaps some distor-
tion in shape. Similar descriptions could be made for the velocity as well as
other properties of the bar. At the same time the particles that compose
the bar have only moved within a small range.
MOTION OF THE DISCONTINUITY SURFACE

We shall now consider a bar in which the section properties u(x), E(x),
and p(x) are twice continuously differentiable. We shall assume that the
longitudinal displacement is also twice continuously differentiable and satis-
fies the equation of motion (2.1) with body forces neglected,
(uEu,,)x = puu 1r- (1)
However, (1) is not required to hold across a section that may move along
the bar in time. This section is a movingdiscontinuitysurface. The displacement
u(x, t) is assumed to be continuous across the surface, but the first derivatives
(hence the stress and the velocity) may undergo finite jumps.
We shall assume that the jumps in ux and Ur are themselves continuous
functions along the discontinuity surface. This means that if we consider the
jump in ux which occurs when the discontinuity surface occupies the position
x 1 at time t 1 and the jump when the discontinuity surface occupies the
position x 2 at time t 2 , then
lim jump in ux at (x 2 , t 2 ) = jump in u" at (x 1 , t 1 ),
x2~x1
r2~ti
where the limiting operation follows the motion of the discontinuity surface.
The purpose of our analysis will be to find equations from which we can
determine the location of the discontinuity surface and the magnitude of the
jumps for t > t 0 if we are given these quantities at an initial time t 0 We
shall begin by finding an expression for the speed at which the discontinuity
surface propagates.
The position x of the moving discontinuity surf ace is assumed to be given
implicitly by
cj,(x,t) = 0 (2)
or, explicitly, either by
x = x(t) or t = t/l(x).
Differentiating (2), we find that
or
(3)
391
Equation (3) relates the speed of the discontinuity surface, dx/dt, to the
partial derivatives of the function </J(x,t). One of the questions that we shall
study is whether there are any restrictions on this speed.
We shall denote the jump in a function/(x, t) across such a discontinuity
surface by the operator [f(x, t)], which is defined by
[f(x, t)] = f(x - , t) - f(x+, t),
where
f(x-, t) = Iim/(e, t),
~-+X
e<x,
and
f(x+, t) = lim/(e, t),.
Recalling that the displacement u(x, t) has been assumed to be continuous

across the discontinuity surface, we may write
[u] = 0. (4)
Rather than regard the discontinuity surface as a moving section in
one-dimensional x space, it will 'be more convenient to consider the surf ace
as being represented by a fixed curve in an (x, t) space, as indicated in Figure
12.16. We shall designate the arc length along the curve <J>(x, t) = 0 by s.
Integrating along the curve between two arbitrary points (x 0 , t 0 ) and (x 1, t 1),
we have
(x1-,r1)
J
(xo-, to)
u,(x-, t) ds = u(x 1 -, 11 )- u(x 0 -, 10 ).
Similarly,
J
(X1 +,ti)
(xo+, to)
u.(x+, t) ds = u(x 1 +, t 1) - u(x 0 +, t 0 ).
FIG URE 12.16. Path of the discontinuity surface plotted in the xt plane.
392
Subtracting the two integrals and using (4), we can then conclude that
(X1,l1)
J (xo,ro)
[us(x, t)] ds = 0. (5)
From the chain rule

dx dt
us(x,t) = ux(x, t) ds + ur(x, t) ds,
and
dx dt
[us(x, t)] = [ux(x, t)] ds + [u,(x, t)] ds
Thus [us(x, t)] is continuous along the discontinuity surface. Since the
points (x0 , t0 ) and (x 1 , t 1) have been arbitrarily chosen along the curve,
the Dubois-Reymond lemma may be applied to (5), yielding
[us(x, t)] = 0. (6)
Equation (6) states that the tangential derivative of u is also continuous

across the discontinuity surface.
We may exploit (6) by rewriting it in terms of the derivatives Ux and u,,
with the result that
dx
[ux] dt + [u,] = 0.
Therefore,
dx <Pr [u,]
dt = --= --,
tf>x [u:x]
or
(7)
Equation (7) shows that the jumps in the first derivatives are related to one
another, and to the shape of the discontinuity surface.
We have yet to examine the implications that u is the solution of (1) on
either side of the discontinuity surface. Let C be an arbitrary closed curve in
the (x, t) plane bounding an interior G. The unit exterior normal n will be
written in component form as
where the unit vectors i and j are shown in Figure 12.17. The boundary curve
C ranges from t = t 1 tot = t 2 , the right-hand side of C is denoted by x = n(t),
Sec. 12.3) DiscontinuousSolutions
393
FIGURE 12.17. The curve C is the boundary of a region G with unit

exterior normal n.
and the left-hand side by x = m(t), as indicated. For a given value of t, let
us consider a portion of the bar given by (but see p. 411)
m(t) s x s n(t).
We apply the integrated balance of linear momentum equation, in linearized
form, to this portion of the bar. Combining (1.43), (1.46), and (1.47), we
obtain
f
n(r)
m(r)
p<1u
11 dx = T(n, t)a(n, t) - T(m, t)<1(m,
t).
Integrating this relation over t 1 St S t2 , we have

12
[(I) f2 J.12
J. p<1u dx dt = J. T(n, t)<1(n,t) dt - 11
t1 m(I)
11
t1
T(m, t)a(m, t) dt
= f
t,
T(n, t)<1(n,t) dt + f
t2
T(m, t)a(m, t) dt.
It should be noted that we have assumed that the integrals involving the
acceleration, Urr, and the stress, T, exist.
Let us introduce as a parameter the arc length, s, along the curve C.

Referring to Figure 12.17, we see that on x = n(t)
dt = cos (J ds = n 1 ds.
Similarly, on x = m(t), we have
dt = -cos (J ds = n1 ds.
Consequently, our integral relation may now be written as
ffpuurt dx dt = f Tun ds = f Euu" n

G C
1
C
1 ds.
Applying Green's theorem to the left-hand side, we obtain
fc (Euu" n1 - puut n2 ) ds = 0. (8)
We now consider the region shown in Figure 12.18, where r is an arbitrary

trajectory of the discontinuity surface. We see that G1 u G2 is bounded by
r 1 u r 2 , G1 is bounded by r 1 u r, and G2 is bounded by r 2 u r. Applying(8)
to each of these three regions in turn, we obtain
f
r1vr2
[Euuxn1 - puurn 2 ] ds = 0,
and
f r2
(Euu"'n1 - puurn 2 ) ds - f r(+ >
(Euu"'n1 - puurn 2 ) ds = 0,
where r( - ) indicates that the discontinuous functions are evaluated as

limits from the G1 side of r. Similarly, .r( +) indicates that the limit is taken
</>(x, t) =0
FIGURE 12.18. Trajectory r of the discontinuity surface divides an arbi-

trary region into two portions, G1 and G2.
Sec. 12.3] Discontinuous Solutions 395
from the G2 side of r. Furthermore, the normal n = n 1i + n 2 j appearing in

the integrals over r is taken as directed from G1 into G2 Consequently, the
second integral in the last equation must be prefixed by a minus sign, as
indicated.
Subtracting the first integral from the sum of the remaining integrals, we
then obtain
fT(-)
(Euu"n 1 - puu,n 2 ) ds - fT(+)
(Euu"n 1 - puu,n 2 ) ds = 0.
Since E, u, and p are continuous across r, the last equation may be written
in the form
t (Eu[u"]n 1 - pu[u,]n 2 ) ds = 0. (9)
We recall that the jumps [u"] and [u,] have been assumed to be continuous
along the curve r. Since the arc r has been arbitrarily chosen along q,(x, t) = 0,
the Dubois-Reymond lemma may be applied to (9) with the result
Eu[u"]n 1 - pu[u,]n 2 = 0
t) = 0. Now the unit normal vector n is proportional to
along <J>(x, q,"i+ q,,j.
Consequently, we may conclude that
(10)
Equations (7) and (10) constitute a pair of linear, algebraic, homogeneous

equations in the unknowns [u"] and [u1]. These will possess a nontrivial
solution provided the determinant of the coefficients vanishes. Thus
- Eq,;+ pq,;= 0,
or
Applying (3), we may then conclude that
or
dx = (~)112 (11)
dt p
Thus we see that the discontinuity surface or wave front moves either to the
left or to the right at a determined speed. We recall that in the case of constant
properties this speed, (/p) 112
,
was designated by c. We shall use the same
notation for variable properties, noting that c now depends upon x. Thus (11)
can be written as
dx
dt = c(x).
We see that weak discontinuities propagate at the local speed of sound.

(The discontinuities must be small or "weak"; otherwise the basic lineariza-
tion would be invalid.) This result is a generalization of that found in the
previous section, where we showed that in a uniformbar all smooth small
longitudinal disturbances propagate with the speed of sound. The generaliza-
tion might have been anticipated if we had realized that a discontinuity can
be regarded as a local disturbance which can only sense the local properties
of the bar.
No TE. Propagation of disturbances at the local sound speed is also dis-
cussed in the treatment of one-dimensional gas dynamics found in Section
15.3.In contrast to a solid, however, a gas generally possesses bulk convective
motion. Consequently, disturbance propagation must be measured with
respect to the bulk motion.
BEHAVIOR OF THE DISCONTINUITY
Now that we have determined the behavior of the moving discontinuity
surface, let us see whether something can be said about the growth of the
jump itself as time progresses. We have written the equation of the dis-
continuity surface in implicit form, <J,(x,t) = 0. Solving this equation for t,
we can write
t = i/J(x),
i.e.,
<J,(x,t) = i/J(x)- t = 0.
The function i/J(x)is known as the wave function. Since
dx
<Pxdt + <Pt= 0,
it is seen that
"'=-,
I dt
dx
or
(12)
which is known as the eiconal equation.There is a three-dimensional counter-

part to (12), but the corresponding derivation is much more tedious.
397
Let us now return to the question of how the discontinuity itself behaves
as the wave front propagates. Knowing the jump in stress or velocity at a
section x 0 at time t 0 , we would like to find out whether one can determine,
from the properties of the bar, the subsequent jumps that will occur when
the wave front has reached the position x at the time t. In particular, we shall
consider a wave front propagating to the right with the material to the right
of the front being undisturbed. The case in which the material to the right of
the front is not quiescent will be left as an exercise.
We shall designate the velocity by
u,(x, t) = v(x, t).
It is convenient to introduce a special notation for the jumps in displacement
and velocity at the discontinuity. (Of course, the first of these is identically
z.ero.) Thus we make the definitions
U(x) =[u(x, t)],=~Cx)
and
V(x) = [v(x, t)],=<x>.
Talcing into account the quiescent state to the right of the wave front, we have
U(x) = u(x, tj,(x)) (13)
and
V(x) = v(x, 1/J(x)). (14)
Since the displacement is assumed to be continuous across the front,
0 = U(x). (15)
Differentiating (15), we then find that
0 = U'(x) = Uz[x, v,(x)] + u,[x, v,(x)] !,

which can be rewritten through the use of (14) and the wave function 1/1as
0 = u.,[x, if,(x)] + V(x)i/J'(x).
Thus
u.,[x, if,(x}] =- V(x)if,'(x). (16)
Since Un(x) also vanishes, we can carry out a similar calculation with the
result
u"""[x,1/J(x)]= - Vzl/1' - (VI/I')'. (17)
Performing a similar computation on V(x), we find
V'(x) = v., + v, ,fl(x),
which can be transformed to yield

v, = c2 (V'1// - Vx//). (18)
Finally, we evaluate the equation of motion (I) on the left-hand side of the
front. Thus
pav, = (Eaux)x = (Ea)'ux + Eauxx,
or
ux(Ea)'
V, = -- + C
2
Uxx. (19)
pa
Before continuing with the calculations, let us review the situation briefly.
Our aim is to find an equation for V = [u,] and for [ux] in terms of the
material properties of the bar. Equations (16), (17), and (18) express the
pertinent derivatives in terms of V; these expressions can then be used in (19).
The only possible complication appears to be terms containing v.x, which will
have to be eliminated to carry out the program.
Introducing (18) into (19), we can show that
',t,' ,t,' (Ea)'

V 'I' - Vx'I' = Ea Ux+ Uxx.
Inserting (16) and (17), we find that
V'lf,' - Vxlf,' = - (.:;' Vi//- Vxlf,' - (Vi//)'.
We observe that the terms containing v" cancel and the resulting expression
can be reduced to
(20)
Let us denote by x 0 the location of the discontinuity surface at time t = t0

Then (20) may be integrated to yield
V(x) [E(x 0)a(x 0)tf,'(x 0)] 1' 2
V(x ) = [E(x)a(x),f/(x)] 112
0
2
Since if,'= l/c and E = c p, we may rewrite this as
V(x) = V(xo) (p(xo)a(xo)c(xo)l1/2.

(21)
(p(x)u(x)c(x)] 112
Equation (21) then states that the jump in velocity satisfies
p(x 0 )a(x 0 )c(x 0 )] 112
[v(x, t)]r=l/t(x) = [v(xo' to)]ro=f/1(.xo)
[ p(x)a(x)c(x) . (22)
Sec. 12.3] DiscontinuousSolutions 399
Since T = Eux, (16) yields

[T(x, t)],=o/l(x)= -E(x)V(x)1J/(x).
Applying (21), we can then show that
[T(x, t)],=o/l<x>
r
p(x)c(x)a(xo)]
= [T(xo' to)]ro=l/l(xo)l_p(xo)c(xo)a(x)
112
(23)
Equations (22) and (23) are the propagation equations we have been seeking.
Consider, for example, a bar of uniform material; i.e., p(x) = p(x 0 ),
c(x) = c(x 0 ). Suppose that the bar has the shape of a truncated wedge
of unit thickness, and of half-height a at x = 0 and half-height bat x = L. Since
c(x) = c(x 0 ) = c, a constant, we reach the conclusion that the eiconal equa-
tion (12)
can be immediately integrated to yield

x
t = if,(x) = -,
c
where we have agreed that x = 0 when t = 0. In addition,
11(x)=2[a-(a~b)x].
Thus the jump in velocity will satisfy
[v(x, t)],=x/c = [v(O,O)] [1- (1 ~ b/a)x/L] 112
and the jump in stress obeys
[T(x, t)]r=x/c = [T{O,O)][1- (1~ b/a)x/L] 1/2.
We observe that in this special case, the jump in velocity increases as the
area decreases and the jump in stress behaves similarly. This is not un-
reasonable, since one would expect the stress to increase as the cross-section
area diminishes and the corresponding jump in velocity to increase as well.*
For varying density p and local sound speed c, the situation is somewhat
more complicated. We note from (22) and (23) that the jump in velocity will
increase if p(x)c(x)11(x) < p(x 0 )c(x 0 )11(x0 ); whereas the jump in stress will
increase if p(x)c(x)/11(x) > p(x 0 )c(x 0)/a(x 0 ).
The fact that the intensity of the jump in stress produced by a hammer blow would increase
as the area diminished could be of practical importance, for a high stress level can lead to failure
of the material.
EXERCISES
1. In the analysis just concluded, we have assumed that the jumps were
continuous along the discontinuity surface. Reexamine the arguments
to determine what could be said if this requirement were lightened to
piecewise continuity.
2. Carry out an analysis of. the propagation of a discontinuity for the
following equations: .
(a) The heat equation, uu = cur, where c is constant.
(b) Laplace's equation, u,.,.+ u11 = O.*
3. The preceding analysis has been carried out under the hypothesis that the
material properties E, u, and p are continuous. A different form of dis-
continuity was considered in the previous section. The purpose of this
exercise will be to show that the approach just used can be applied to the
previously considered problem. Assume that we have a discontinuity sur-
face in the bar which is fixed for all time. Furthermore, E, u, and p may
have jump discontinuities across this fixed surface. Apply the method of
analysis of this section to the new problem. In particular, show that the
result equivalent to (7) states that the velocity is continuous across the
discontinuity surface, and that the equivalent to (10) yields the fact that
the force is also continuous across the surface.
4. The detailed calculations that yield (17) through (23) have not been
included in the text. Verify these equations.
S. Carry out the same analysis as that of the example at the end of the
chapter, with the bar reversed.
6. Modify the discussion under (23) so that it applies to a truncated cone
of basic radii a and b (see Figure 12.19). What is the difference in the
behavior of the jumps?
7. Define an appropriate " difference displacement" so as to remove the
text's restriction to wave fronts moving into undisturbed material.
Ta
l_ 0
FIGURE 12.19. Side view of a truncated cone whose ends are circular
disks of radii a and b.
Imaginary propagation velocities should be rejected, as they have no direct physical meaning.
Yet extensions into the complex domain have proved of considerable value. See Garabedian
(1964), especially Chap. 16.
Sec. 12.4] Work, Energy, and Vibrations 401
12.4 Work, Energy, and Vibrations

We shall consider briefly in this section the relation between work and
energy. In addition, we shall examine some special solutions of the wave
equation known as free vibrations. In the course of our discussion, we shall
see how the balance between work and energy plays a role in the vibration
problem that is both physically interesting and computationally valuable.
WORK AND ENERGY
Let us consider a bar of length L which has varying sectional properties
and which is subject to end forces F(L, t) and F(O, t) only. (See Figure 12.20.)
Using Hooke's law, we observe that
F(L, t) =E(L)u(L)u,;(L, t)i
and
F(O,t) = -E(O)u(O)u,,(0, t)i.
F (0, t) F (L,t)
FIGURE 12.20. Bar subject to endforces.
In analogy with ordinary particle mechanics, we define the kinetic energy

density (per unit length) of the bar as
fp(x)u(x)u;(x, t) (1)
and the total kinetic energy of the bar at time t by
K(t) = -l t
0
p(x)u(x)u;(x, t) dx. (2)
Assuming that there are no distributed forces along the length of the bar,
we define the workof the externalforces,We, in the time interval O5: t 5: t 1 as*
We = t0
[F(L, t) u1(L, t)i + F(O,t) ui(O,t)i] dt. (3)
Thus
We= f0
[E(L)u(L)u,;(L, t)ui(L, t) - E(O)u(O)u,;(O,
t)ui(O, t)] dt. (4)
Recall that work is done at a rate equal to the scalar product of the force and velocity
vectors. We are extending this definition of point mechanics with the goal of obtaining some
meaningful generalizations of classical results.
If -r denotes the internal force per unit length, the internal work, W;, will
be given by
tt fL
W;= J 0
wrdxdt.
0
To evaluate -r, consider an arbitrary portion of the bar defined by

e~ x ~ e+ 1.
The force exerted on this portion by the remaining parts of the bar is given by
T(e + l, t)u(e + l) - T(e, t)u(e) = f~+I
~
[T(x, t)u(x)l~ dx.
Thus the internal force, -r,per unit length ate is given by
-r(e,1) = lim (/)-

1-+0
1
I
~+I
~
[T(x, t)u(x)l~ dx
= [T(e, t)u(e)]~. (5)

Therefore,
W;= .c
0
l(Tu)xut dx dt
0
= r 0
l(Euux)xur dx dt.
0
(6)
We shall define the change in potential energy in the time interval O ~ 1 ~ 11

as the difference between the work of the external forces and the internal
forces in that time interval. In order to find an analytic expression for the
change in potential energy, let us return to the formula for W;. Integrating
by parts, we find that
W;= 1ft Euux Ur ILdt - JLEu fltUx Uxr dt dx

0 0 0 0
= JEuux
ft
Ur
ILdt - JLtEuu; 1!1dx
0 0 0 0
= We- t t E(x)u(x)u;(x, t)t dx.
Thus we see that the change in potential energy in the time interval O ~ 1 ~ 11
is given by
t t
0
E(x)u(x)u;(x, t 1) dx - t(
0
E(x)u(x)u;(x, 0) dx.
Consequently, the potential energy P at time t is given by

L
P(t) = t IoE(x)u(x)u;(x, t) dx + C. (7)
Sec. 12.4] Work, Energy, and Vibrations
We shall take C = 0 so that the arbitrary zero of potential energy is identified

with the strain-free state (ux = 0).
Let us return now to the equation of motion
(Euux)x = puu 11
If we multiply by u, and integrate over both x and t, we then have

1
[' t(Euux},,u, dx dt = t { puu 11u 1 dt dx.
0 0 0 0
Integrating, we obtain
1r''Euuxu, ILdt- 1r'' JLEuuxUrx dx dt = t JL puu; !'1dx.

0 0 0 0 0 0
The second integral on the left-hand side can be simplified as
l f 0
Eu
0
UxUx1 dt dx = t (E(x)u(x)u;(x,
0
11
t)l dx.
0
Thus we may conclude that
1r''Euu,. u, ILdt = t JLpuu; ,,,dx + t LEuu; 1'1

0 0 0 0
dx. f
0 0
(8)
We note that (8) states that the work done by the external forces equals the
change in kinetic energy plus the change in potential energy. This result is
referred to as the work-energy principle.
A VIBRATION PROBLEM
We now tum our attention to a special problem that can serve as a prototype
for a whole class of questions that appear frequently in this text. Let us
consider a bar of length L, fixed at the end x = 0, and attached to a linear
spring at the other end. We assume that there are no external forces. Thus the
displacement satisfies the following differential equation and boundary
conditions :
(Euux)x = pO'Urr, 0 < x < L,

u(O,t) = 0, (9)
E(L)u(L)ux(L, t) + k2u(L, t) = 0.
We ask whether it is possible to obtain standing wave solutions with harmonic
time dependence; i.e., are there solutions to (9) of the form
u(x, t) = w(x) exp (iwt)?

Substituting this assumed solution into the boundary value problem (9) and
dropping the exponential factor that is common to every term, we find
(.Eow')' = -aipuw, 0 < x < L,

w(O)= 0, (10)
E(L)u(L)w'(L) + k 2 w(L) = 0,
where ' denotes differentiation with respect to x.

Equations (10) define a Sturm-Liouville problem, which, in general, has
only the trivial solution w = 0, corresponding to the uninteresting case of
no motion. There are, however, certain values of the frequency a, for which
there are nontrivial solutions. Such values of (J) are eigenvalues and the corre-
sponding solutions are eigenfunctions.A number of general results for Sturm-
Liouville problems have been given in Section 5.2. Our discussion here is
largely independent. We concentrate on a particular method for practical
estimation of the eigenvalues.
THE RAYLEIGH QUOTIENT

Let us examine some of the properties enjoyed by the eigenvalue and the
corresponding eigenfunction w(x). We begin by multiplying the differential
equation in the system (10) by w(x) and integrating from Oto L, with the
result
2
( (Eaw')'w dx + a, ( puw 2 dx = 0.
Integrating the first integral by parts, we obtain
Euw'w IL-
o o
l
Eaw'2 dx + a, 2 f puw 2 dx = O.
Jo
Now
=0
E(O)o'(O)w'(O)w(O)
from the first of the boundary conditions in (10) and
2
E(L)u(L)w'(L)w(L) = -k w2(L)
from the second of the boundary conditions. Thus
-k 2
w2(L)- lo .Eow'2 dx + (J) 2 f puw2 dx = 0,
Jo
or
w2 = R(w), (11)
405
where
JL0Euw' dx + k2w (L)
2 2
R(w) = "- ---=--~---
J~puw 2 dx
is known as the Rayleigh quotient. Equation (I I) will be seen to be a very
useful relation connecting the eigenfunction and the eigenvalue.
The Rayleigh quotient R(w) has an interesting physical interpretation.
To see this, multiply both the numerator and the denominator of R by !-,
Then upon substitution of u = w exp (icot), one can see at once that R is
equal to the potential energy of internal strain (7) plus a potential energy
equal to one-half the force* k 2 w(L) on the stretched spring times the distance
w(L) stretched, all divided by the kinetic energy (2). In short, the Rayleigh
quotient is the ratio of total potential to kinetic energy for a given mode of
vibration.
PROPERTIES OF EIGENVALUES AND EIGENFUNCTIONS

Let us return to the eigenvalue problem (10) and assume that we have a
second eigenvalue , corresponding to the eigenfunction y(x). We also
assume that w 2 and 2 are distinct from one another. Thus and y(x)
satisfy
(Euy')' + 2 puy = 0,
y(O) = 0, (12)
E(L)u(L)y'(L) = -k 2y(L).
By an argument virtually the same as that used to prove (5.2.3), one can show
that eigenfunctions corresponding to distinct eigenvalues are orthogonal, in
the sense that
( puwy dx = 0. (13)
Also by reasoning which differs only in detail from that used in Chapter 5,
one can show that the eigenvalues co2 of (10) are necessarily real and that the
eigenfunctions can always be taken to be real. But is it possible for co2 to be
negative, in which event co would be purely imaginary? Should this prove
to be the case, the temporal behavior would no longer be oscillatory but
=
would rather exhibit exponential growth or decay. [If co iu, then exp (icot) =
exp (-ut) grows (decays) if a is negative (positive).] This question can be
settled by noting that the expression (l l), which equates ro2 to the Rayleigh
quotient, demonstrates that since the eigenfunction corresponding to the
eigenvalue w2 is real, then w 2 is positive.
Finally, we note that the eigenfunctions corresponding to a single eigenvalue
are unique to within a multiplicative constant (see Exercise 1).
* For a spring, the potential energy can be regarded as the average force required to stretch it
a given amount, times the amount stretched. See Exercise 5.

AN EXACT SOLUTION WHEN PROPERTIES ARE CONSTANT
Next, we shall consider the very simple example that arises when one
considers a bar with constant properties which is built in at x = 0 and free
at x = L. Here, u, and pare constant and k 2 = 0. Boundary value problem
(10) becomes
w(O)= w'(L) = 0,
where, as usual, c2 = E/p. The general solution of the differential equation is
w(x) = M sin(; x)+ N cos(~ x).

The first boundary condition implies that N = 0, and the second boundary
condition then requires that
~cos (;L)=0.
Noting that ro = 0 yields a trivial solution, we see that
~ L = (2n + I ),r '

c 2
where n = 0, I, 2, .... Thus there are a denumerable number of eigenvalues
ronand a denumerable number of corresponding eigenfunctions wn(x) given by
(2n + l),rc
ron = 2L
. [(2n + l),rx]
wn(x ) = SID 2L ,
n = 0, I, 2, ....
CHARACTERIZATION OF THE LOWEST EIGENVALUE
AS THE MINIMUM OF THE RAYLEIGH QUOTIENT
The preceding example illustrates the fact that the exact calculation of the
eigenvalues and eigenfunctions depends on finding explicit solutions to the
governing differential equation, a task that is not always easily accomplished.
A number of techniques have been developed for finding eigenvalues approxi-
mately, and we shall close this chapter by considering one of these methods
briefly.
In Chapter 12 of II we shall consider a number of minimum properties
that are related to eigenvalue problem (10). One result, which we shall state
without proof, is the following. Let Ube the class of functions u(x) that are
twice differentiable and vanish at x = 0. Designate the lowest eigenvalue of

(10) by rof. Let
=min R(u) (14)
ueU
and let w(x) denote the corresponding minimizing function. Then = wi

and w(x) is the corresponding eigenfunction.
The minimum principle has the following physical interpretation. Functions
in class U are called virtual displacements (perhaps "conceivable displace
ments" would be a better phrase). They are smooth and satisfy the boundary
condition at x = 0. The corresponding Rayleigh quotient gives the ratio of
virtual potential energy to virtual kinetic energy. The lowest eigenvalue is the
minimum of this ratio.
It is interesting to note that the functions which belong to the class U are
not required to satisfy the second of the boundary conditions in (IO). But it
can be shown that the minimizing function will automatically satisfy this
boundary condition. Such a boundary condition is termed natural and
appears in a wide variety of problems associated with minimum principles.
Further discussion, which includes a proof of the characterization of ro2,
as the minimum of the Rayleigh quotient (if it has one), will be found in
Section 12.2 of II.
The minimum principle stated above, as well as the similar theorems for
the higher eigenvalues, gives rise to a very effective technique, the Rayleigh-
Ritz method, for obtaining approximate values for the eigenvalues. Here we
shall only give an indication of the power of the technique by finding an
estimate for the lowest eigenvalue in a particular problem. The idea is
simple. If the lowest eigenvalue rof is the smallest value of R(u) among all
u e U, then an upper bound for rof can be computed by evaluating R for any
function u whatever which belongs to U. Presumably, the cleverer we are in
guessing a u that approximates the correct eigenfunction w, the closer our
upper bound will be to the exact eigenvalue.
ESTIMATE OF THE LOWEST EIGENVALUE FOR A WEDGE

We now consider the special example of a bar that is of unit thickness but
wedge-shaped in plan form, as shown in Figure 12.21. The bar is free at the
pointed end and clamped at its base. The area a is given by
bx
a=-.
L
The eigenvalue problem for this configuration is
2
EbLxw')'= _ w pLbxw for O < x <L,
(
w(L) = 0, xw' -+0 as x-+0.

FIGURE 12.21. A wedge that is built in at its broad end and is free at
its pointed end.
The second boundary condition is by no means a trivial one, for it states

that the solution w cannot be as singular as log x in the neighborhood of
the origin. Since E, b, p, and L are constants, the boundary value problem
reduces to
w2p
(xw')' +E xw = 0 for O <x <L,
w(L) = 0, xw' -o as x-O.
The solution to the differential equation satisfying the second boundary

condition is
where. 10 is the Bessel function of zero order. The first boundary condition
requires that
There are a denumerable infinity of solutions ro to the above equation, one

for each zero of the Bessel function 10 Taking the smallest zero from a
table in one of the various collections, we find that
)1/2
W1 = (p 2.4048 L- 1

Let us pretend that we do not know the above solution and try to estimate
w1 by means of the minimum principle. The Rayleigh quotient corresponding
to this eigenvalue problem is found to be
J~Exu' 2 dx
R(u) = JL
O pxu
2dx '
with the required boundary condition that u(L) = 0. If we take as our trial
function
u=L-x
and denote the corresponding value of the Rayleigh quotient by A 2 , a simple
calculation shows that
A2 = 6E,
pL2
or
A= (~f
' 2
2.45 L-
1

We thus see that our very simple approximation provides an upper bound
which is in error by less than two per cent.
It should be quickly pointed out that the example just cited is somewhat
too successful. It will be found, in practice, that such simple trial functions
rarely lead to such close estimates. Nevertheless, it is hoped that this case
will convince the reader that the minimum characterization is an extremely
valuable one and that comparatively elementary calculations can lead to
quite decent estimates of the eigenvalues. Experience shows that if the trial
function reflects a physically reasonable displacement,the estimated eigenvalue
will be much closer than one resulting from an arbitrarily chosen function.
There is an art in using the Rayleigh-Ritz method efficiently, and prior
reflection frequently produces rewards in accuracy and ease of computation.
EXERCISES
1. Establish the fact that the eigenfunctions of (10) corresponding to a
given eigenvalue are unique to within a multiplicative constant by the
following argument. Suppose that there are two linearly independent
eigenfunctions corresponding to the same eigenvalue. Show that there
is a linear combination, u(x), with the property u(O)= u'(O) = 0. Apply
the uniqueness theorem for second order equations (see, e.g., Boyce and
DiPrima, 1969, p. 99).
2. The eigenvalue problem for a bar fixed at x =0 and carrying a mass M
at x = L is given by
(Euv')' + oipuv = 0, 0 < x < L,
v(O)=0,
E(L)u(L)v'(L) = a>2Mv(L).
What is the Rayleigh quotient for this problem? What is the orthog-
onality relation that eigenfunctions corresponding to distinct eigenvalues
satisfy?
410
3. Consider the eigenvalue problem discussed in the text for a bar of con-
stant properties. The eigenvalue problem takes the form
v" + A2 v = 0, 0 < x < !n,
v(O) = 0,
v'(!1t) + a2v(!n) = 0.
(a) Find the transcendental equation satisfied by l. What is the lowest
nontrivial eigenvalue if a2 = t?
(b) Find an approximate eigenvalue by taking
ii=x.
(c) Find an improved eigenvalue by taking
3
v=x+ax ,
where a is determined so that R[vJ is as small as possible. What are

the percentage errors when compared with the previous exact calcu-
lation? (The calculations here are somewhat lengthy. Excellent re-
sults can frequently be obtained by the Rayleigh-Ritz method if a
number of parameters are employed, but then use of a computer
is essential.)
4. The transverse vibrations of a simply supported beam of length L are
governed by the eigenvalue problem
[E(x)I(x)w"]" = aipu(x)w,
w(O)= w"(O)= W(L) = w"(L) = 0.
(The boundary conditions correspond to simply supported ends.)
(a) Show that the corresponding Rayleigh quotient is
JtE(x)l(x)w"(x) dx 2
R(w) = J~pu(x)w(x) dx 2
Begin, as in the text, by multiplying the governing differential

equation by w.
(b) What is the orthogonality relation that eigenfunctions corresponding
to distinct eigenvalues satisfy?
5. The displacement x(t) of a mass m restrained by an undamped spring
satisfies md 2x/dt 2 = -k 2 x. [Here x = 0 when the spring is unstretched.]
Multiply the equation by dx/dt and integrate. Interpret in terms of a
kinetic energy !m(dx/dt) 2 and a potential energy ,tk 2 x 2 Compare these
definitions of energy with the continuum versions of kinetic and potential
energy defined in the text. [Note that the potential energy can be inter-
preted as stored work, where this work is the product of the distance the
spring is extended times the average of the initial force (uro) and the
final force in the extension process.]
Sec. 12.4] Work, Energy, and Vibrations 411
6. At the end of Section 12.2 we obtained a complex solution to a wave

reflection problem. The real part of this solution was
u = cos (x - c1 t) + A cos (x + c1 t), x < 0;
u = B cos [.B(x - c2 t)], x > 0.
[The constants A and Bare given in (2.14).] Provide a full discussion of
the distribution of energy.
(a) In particular, find the total energy per period for x < 0 and for x > 0.
(b) Your results in (a) should be independent of time; show from general
considerations that this must be so.
(c) Making allowances for the fact that the transmitted wave does not
have the same period as the incident and the reflected waves, show
that the energy in the incoming wave balances the sum of the energies
in the transmitted and reflected waves.
Correction to pp. 393-94:

The region G in Figure 12.17 should be a slightly distorted rectangle,
corresponding to a bar that at time t 1 is located between x= a and x =band
whose points never deviate significantly from x = a and x = b. (This is the
general case, given our assumption of small deformation, and invoking a
rigid translation if necessary.) The same remark holds for Figure 12.18; in
addition the curve r should be a slight perturbation from a vertical line.
Equation (8) cannot be derived from Green's theorem, since ut is dis-
continuous. One can obtain this equation as follows. Define
=(aT, - pau,) =(Eau",
v - paut).
Note that on the top and bottom of C =r 1 u r 2 , n = (0, 1) and (l, 0)
respectively. The left and right sides of Care nearly vertical lines so that n2 is
negligible and the normals can be approximated by ( - 1, 0) and (l, 0) respec-
tively. If one now integrates the appropriate linearized version of (41) with
respect to t, one obtains
0= fti
t2 { d
T(n, t)a(n, t) - T(m, t)a(m, t) - d
t
f n(t)
m(t)
}
pau 1 dx dt.
This yields c vini = 0, which is equivalent to (8).

CHAPTER 13
The ContinuousMedium
HISTORIC moments usually culminate one long series of events and

begin another. One such moment came in 1687 with the publication of
Newton's Principia. Dramatically positioned at the head of Book 1 were the
three laws.
I. Every body continues in its state of rest, or of uniform motion straight
ahead, unless it be compelled to change that state by forces impressed
upon it.
II. The change of motion is proportional to the motive force impressed,
and it takes place along the right line in which that force is impressed.
III. To an action there is always a contrary and equal reaction, or, the
mutual actions of two bodies upon each other are always equal and
directed to contrary parts.
By taut and clear deductions from these laws Newton summed up a
century's work on the motion of one or two bodies in a vacuum. "What
Newton writes is correct, clear, and short. In earlier works the brilliant
diamonds of discovery lie concealed in an opaque matrix of wordy special
cases, laborious details, metaphysics, confusion, and error; while Newton
follows a vein of pure gold."*
Unlike his work on the one and two body problems, Newton's work on
the three body problem was not centered on the formulation and solution of
the governing differential equations. Indeed "he shows no sign of an attempt
to set up equations of motion. That he obtained some correct inequalities
and groped his way to the major approximate results under this handicap
is one more tribute to his peerless grasp of the physical essence of mechanics
and to the might of his brain. It does not show, however, that his formulation
of the general laws of mechanics was adequate. History proves the contrary. "t
Further progress on the three body problem required formulation of the
general equations of celestial mechanics and these equations were not pub-
lished until 1749 (by Euler).
Newton's definitive and lasting work concerned bodies that could be ideal-
ized as rigid bodies or as point masses. In Book II Newton struggled with
motion in a resisting medium, but most of his work has since been discarded.
There is no disgrace in this. It took at least a century to achieve even a
general understanding of the problems Newton grappled with, and many
important questions remain to this day.
C. Truesdell, "A Program Toward Rediscovering the Rational Mechanics of the Age of
Reason," Arch. Hist. Exact Sci. 1: 1, S (1960).
t Ibid., p. 7.
412
Sec. 13.1] The ContinuumModel 413
Motion in a resisting medium is a major problem in the subject that is

now termed continuum mechanics. A turning point in the history of contin-
uum mechanics occurred in 1776, nearly a century after the publication of
Principia, with the publication of Euler's laws of mechanics. These state that
for every body or system of bodies, and for every part of every body, whether
the bodies are conceived of as points or as filling space, whether deformable
or rigid:
1. The total force acting upon the body equals the rate of change of the total
momentum.
2. The total torque acting upon the body equalsthe rate of change of the total
momentof momentum,whereboth the torque and the momentare taken with
respect to the same fixedpoint.*
Euler's laws form the core of the restrictions governing the motion of
continuous media, but they are not the only ones. It is the purpose of this
chapter and the next to present a mathematical formulation of all such
restrictions that need to be considered, whatever the constitution of the
medium, if one ignores electric, magnetic, chemical, and certain other effects.
The few historical remarks we have made indicate that formulation of these
restrictions, or laws, required more than a century of endeavor by some of
the most brilliant men of their age. Further evidence for this assertion is
provided in the historical discussion of Appendix 14.2.
As this book is being written, the equations of continuum mechanics are
an active subject for research. Older equations are being clarified and unified
by the invention of deeper and further-reaching concepts, while new equa-
tions are being formulated to encompass ever-widening classes of important
scientific phenomena. Thus, it is our hope that from the discussion to follow
the reader will not only gain a good understanding of " laws of nature "
that have great historic importance and great present importance, but that
he will also gain an understanding of the process of formulating the equations
that embody such laws which may aid him in a future attempt to formulate
new equations.
We begin the present chapter with a discussion of the continuum model.
We then introduce the material and spatial descriptions of the motion of a
continuous medium and some related mathematical concepts and formulas.
With this background, the equations that embody the basic laws can be effi-
ciently presented. This is done in Chapter 14.
13.1 The Continuum

Model
In earlier discussions we have used models of nature in which a substance
or quality is regarded as continuously distributed in space. For example,
temperature and mass have been regarded as smoothly varying functions
* Ibid., p. 31.
414 The Continuous Medium [Ch. 13
of position. In this section we shall subject the concept of a continuum to

closer scrutiny, particularly as it is used in mechanics.
Particle mechanicsdeals with the equilibrium and motion of systems of
point masses. In continuum mechanics one deals with a continuousmedium
wherein smoothly varying properties like density and velocity are assigned
to each point in space that is instantaneously occupied by the body under
consideration.
Most of us probably view materials like water or steel as being composed
of assemblages of molecules. This corpuscular (corpus= body) viewpoint
and the continuum viewpoint are both illuminated by considering how de-
pendent variables in the second can be regarded as a suitable average of
corresponding variables in the first. To fix ideas, let us start by considering
how the density of a fluid is related to its molecular structure.
MOLECULAR AVERAGES
At time t consider a cube C, centered about the point (x0 , y 0 , z0 ), whose

sides have length h.
C: Ix - x0 I ::;th, IY-Yol :::;;;

th, Iz - Zo I :::;;;
th.
Within C there is a certain amount of mass Mh and hence a mean density
Ph = Mh/h 3 In order to define a density p(x 0 , Yo, z 0 , t) at the point
(x0 , Yo, z 0 ) and at the time t, we examine what happens as h becomes smaller
and smaller. We obtain the graph of Ph as a function of h that is depicted
in Figure 13.1. This graph does not record the results of actual experiments;
it is impossible to make a sequence of measurements at a fixed time t. Rather,
II III
P <xo,Yo,zo,t)
"----~---------'----h
7
(cm)
10- 102.
FIG URE 13.1. Graph of mean density in a box of side h, versus h.

The box is centered on a fixed point (xo, y 0 , zo), and in a "thought experi-
ment" all measurements are made at a certain time t. The scale on the
abscissa is logarithmic.
Sec. 13.1] The Continuum Model
the figure presents the conclusions of a "thought experiment" based on the

following general considerations.
The density varies in region III because of the nonuniform nature of the
fluid. For example, suppose that (x 0 , Yo, z 0 ) were a point in the Gulf Stream.
If C were very large, it would contain some relatively warm Gulf Stream
water and some relatively cool ordinary ocean water, and would have a higher
density than a smaller cube containing only Gulf Stream water. But one
would anticipate only very small density fluctuations ash decreased from 10
to I cm. The transition between regions III and II might be expected to
occur when his of the order of magnitude of hundreds of meters.*
In region II the graph of Ph is horizontal because one would certainly not
expect fluctuations in Ph as h decreased, say, from about 10 cm to about
10-s cm.
If h were so small that there were only a few molecules in the cube C,
then a small decrease in h might well cause a large change in the mean density
Ph. This change would occur when a previously counted molecule was ex-
cluded by a slight shrinkage of the box. To gain an idea of when this effect
is significant, note that 1 cm 3 of water contains about 3 x 1022 molecules.
Consequently, when h is approximately 10- 7 cm one expects about 30
molecules in C. For values of h less than about 10- 7 cm, one therefore
expects the rapid variations of region I.
Thus, as h decreases from large values, fluctuations in Ph will only become
negligible at values of h small compared to L, the length scale of density
nonuniformities. (This scale is a typical distance over which significant
changes in density take place. Compare Section 6.3.) Rapid fluctuations
again appear when h is of the order of a length scale t describing the typical
distance between molecules. Region II will exist providing that t ~ L.
The quantity p(x 0 , Yo, z0 , t) should be taken to be the value of Ph in region II.
From the present molecular average point of view, it thus seems possible to
consider mass to be continuously distributed if the scale of distance between
molecules is small compared to the scale of density nonuniformities.
Similar considerations lead to the identification of the continuum velocity
v(x0 , y 0 , z0 , t) with the average velocity of the molecules that at time t are
located in a small cube centered about (x0 , y 0 , z0 ). The cube should contain
many molecules, but it should be small compared to the scale of velocity
non uniformities. (Incidentally, "the" scale of non uniformity actually de-
pends on what types of changes are of interest in a particular problem.
For example, the scale of velocity variation in the study of global air circu-
lation would be far different than the scale that is relevant to an analysis
of insect flight.)
We have in mind the grossest density variations. Certain small variations would be masked
by coarse averaging. If these are of interest, the transition between regions II and Ill would occur
at a much smaller value of h.
MASS DISTRIBUTION FUNCTIONS

A more fundamental approach to averaging out molecular detail involves
the mass distributionfunctiont/J(x,y, z, u, v, w, t). This function is defined by
the requirement that the integral
fv2[2[2r.2[2l/l(x, y, z, u, v, w, t,) dx dy dz du dv dw
fw2 (1)
w1 Ut u1 %1 Yl Xt
gives the mass of material at time t that satisfies the following restrictions
on its location and its velocity components:
X1 $ x $ X2' Yi $ y$ Y2' Zi $ z $ Z2'

(2)
U1 $ U $ Uz, Vi $ V $ Vz, W1 $ W $ Wz.
To simplify our brief discussion of the mass distribution function, let us

restrict our consideration to the case of a single spatial component x and a
corresponding velocity component u. In this case an informal version of the
above definition is that for very small ll.x and au, t/J(x,u, t) ll.x au gives the
mass located between x and x + ll.x which has a velocity component between
u and u + au. Note that the total mass located between x = a and x = b
is given by
f' -co
r a
t/l(x, u, t) dx du.
Consider a system composed of a single particle of mass m 1 and velocity

ui, located at x 1 Since there is no material with positions other than x 1
and velocities other than u 1 , the distribution function must satisfy
t/J(x,u, t) = 0
In addition, since the total mass of the system is m 1 , we have
Jfl/J(x, u, t) dx du= m1 ,
R1
where R 1 is any region of the xu plane that contains (x 1 , u 1). A "function,"

like the one just considered, which vanishes everywhere except at a point
yet has a nonzero and finite integral, is called a deltafunction.
Now consider a system composed of N particles having masses m;,
velocities u;, and locations X;, respectively; i = l, 2, ... , N. The correspond-
ing distribution function must vanish except at the points (x;, u;), where it
can be regarded as having infinitely tall, slim peaks such that
Jft/J(x, u, t) dxdu
R,
= m;.
Here R; is a region that contains (x;, u;) but not other points (xk, uk), k "F i.
Sec. 13.1] The Continuum Model 417
In many situations N is very large, but there is a distance increment fix

and a velocity increment Au such that the following two requirements are
satisfied: (i) Each rectangle of dimension fix by Au in xu space contains
many peaks of t/1;and (ii) the increments fix and Au are small compared to
interesting length and velocity changes. Since details are not of interest,
it is appropriate to define an average mass distribution function 'l'(x, u, t). One
way to do this is to require that the mass corresponding to ' is correct in an
appropriate fix by Au region:
'l'(x, u, t) Ax Au= J
u+(l/2).4.u
u-(1/2)
f x+(l/2).4.x
.4.u x-(1/2) .4.%

1/J(e,
r,, t) de dr,. (3)
So defined, the function 'I' no longer has infinite peaks, but it does have
small discontinuities when the boundary of the fix by Au rectangle about
(x, u) crosses one of the (xi, ui). Since the integral of a piecewise continuous
function is continuous, however, another averaging would remove these
discontinuities. The first derivative of the new averaged function would still be
discontinuous, but this defect can be removed by another averaging, etc.
The details are not important here. What should be realized is that for large
scale phenomena one can achieve considerable simplification by focusing
attention on a smooth averaged distribution function. In Chapter 1 we dis-
cussed the use of such an average, ', in stellar dynamics. Also see Exercises
14.1.12 and 14.1.13.
Further simplification results if it is also true that the distance that a particle
travels before it is sensibly deflected by another is small compared to the scale
of interesting phenomena. Then, for example, one can write equations for
p(x, y, z, t) and v(x, y, z, t), rather than 'l'(x, y, z, u, v, w, t). This matter
will be discussed further in Section 14.5.
THE CONTINUUM AS AN INDEPENDENT MODEL
Another approach to the continuum, although superficially naive, is no
less valid than the molecular-average approach just discussed. This approach
starts by questioning the assumptions underlying the justification of a con-
tinuum model by reference to the "real" molecular structure of nature.
Indeed, are molecules "real"? Are those assemblages of spheres and rods
which the chemists use "real"? Do not the physicists tell us that molecules
are "really" complicated structures of atoms, which are themselves tiny
solar systems of electrons in orbit about a nucleus? When pressed a little,
physicists in fact caution us that the tiny solar system is a model for what is
more accurately described by certain clouds of probability densities which
give the statistical behavior of fundamental particles. When pressed still
further, they aver that there may be something more fundamental than
fundamental particles but urge that the questioner return in a few years
when the present stimulating confusion may have given way to a clearer
understanding of the foundations of physics.
418 The ContinuousMedium [Ch. 13
From this point of view the continuum is that one of a number of models used
in science which should obviously be the first employed in attempting to explain
what "appear to be" macroscopic phenomena. This model is neither inferior
nor superior to other models of nature. Like all of them, it is to be judged
by experimental examination of the deductions made from it. In particular,
there is no necessity to refer to the molecular model in developing the con-
tinuum model. The latter can stand on its own.
Nevertheless, one might assert that the molecular model is superior to the
continuum model in the same way that quantum mechanics or relativistic
mechanics is superior to classical mechanics, because the latter can be derived
as an average or a special case of the former. This is partially true. By taking
appropriate averages of the equations used in the kinetic theory of perfect
gases, for example, one can derive the corresponding continuum equations.
"In theory" one could also do this for all materials, but in practice this is
not at present possible; the molecular details of the liquid and solid states
are imperfectly understood and the required mathematical techniques are not
fully developed. Indeed, a great part of the utility of the continuum model
lies in its ability to disregard molecular details.
A more subtle point is this. The measurements needed for quantum me-
chanics and molecular mechanics are made with instruments. To interpret
the needle deflections or oscilloscope traces, one must use the laws of contin-
uum mechanics. From this point of view the roles of "superior" and "in-
ferior" in the previous paragraph are reversed. It appears that the various
models of nature are probably better conceived as lying on a circle, equally
near and far from "reality," than as being linearly ordered by their degree
of reality.
Most scientists on most occasions do not concern themselves with the
thorny philosophical questions that emerge from a searching examination
of what lies at the foundation of their endeavors. Our brieflook at the founda-
tions of continuum mechanics has revealed that they are no less secure than
the foundations of other branches of science. The reader is probably ready
to cover over what defects there are in the foundations and, hoping for the
best, to commence building.
EXERCISE
1. Demonstrate that the value for p(x 0 , y 0 , z 0 , t) would be unchanged if
spheres, rather than cubes, were used in the computation of Ph. Gener-
alize.
13.2 Kinematics of Deformable Media

We tum to the study of the motion of a continuum without regard to the
forces that produce the motion. This is called kinematics as opposed to
dynamics,the study of the influence of forces on motion.
Sec. 13.2] Kinematics of Deformable Media 419

POINTS AND PARTICLES
As a continuum moves, it is natural to try to specify the position of each
point as a function of the time t. To identify a given point, we specify its
position A at a certain initial time t 0 If x is the position vector of this point
at t, we describe the motion of the medium by the equations
x = x(A, t), x(A, t 0 ) =A. (la, b)
In words, xis the position at time t of a point that was initially (at t = t 0 )
located at A. The word particle is used to denote the point that starts at a
fixed initial position A and moves according to (1). For fixed A the curve
x = x(A, t), t ~ t0 ,
is called a particle path.
We emphasize that in the continuum model a particle is not a molecule.
An example gives an idea of the magnitudes involved. Recall from the
conclusion of Section 4.1 that in time ta substance diffuses roughly (Dt) 1f 2
units, where Dis the diffusivity. The observed value of D for carbon dioxide
diffusing in air is 0.14 cm2 /sec at 0C. Thus molecules of carbon dioxide
that were initially confined to a very small volume in 1 min would diffuse
so as to be located mainly within a sphere of radius about 3 cm. Diffusion
into a sphere of radius 1 cm would take place in just 7 sec.
Let A(x, t) be the function of x and t obtained by solving (1) for A. The
result
A= A(x, t)
states that A is the initial position of the particle now (at time t) located at x.
Since A(x, t) and x(A, t) are a pair of inverse functions,
A[x(A, t), t] =A, x[A(x, t), t] = x. (2a, b)
We shall postulate impenetrability of matter: A(x, t) and x(A, t) will be
assumed to be single-valued invertible functions so that, at time t, for each A
there exists a unique x and vice versa. We shall also postulate smoothness of
motion: A(x, t) and x(A, t) will be assumed to have as many continuous
partial derivatives as required for the various operations we shall perform
on these functions.
In some contexts it is necessary to relax these assumptions on certain
curves and surfaces to allow for the splitting or coalescence of material.
Moreover, consideration of discontinuous shock waves or the formation
and collapse of bubbles require that the invertibility assumption be dropped
for certain domains.
As one example of the many dependent variables that will enter our dis-
cussion, consider the density. Let ci(A,t) denote the mass per unit volume at
time t of the particle initially located at A. Let p(x, t) denote the mass per
unit volume at time t, at the fixed point x. We assert that at time t the value
of a dependent variable at the point x equals the value of the same dependent
variable for the particle located at x at time t. This assertion could be called
point-particle interchangeability. It is really part of the definition of" particle."
In the case of density, point-particle interchangeability gives
p(x, t) = c>[A(x,t), t]. (3)
To see this, observe that, by the definition of c>,the expression c>[A(x,t), t]
describes the density now associated with a particle that was initially located
at the initial location of the particle now at x. ("Now" means "at time t. ")
This is the same as saying that c>[A(x,t), t] is the density associated with a
particle now at x.
MATERIAL AND SPATIAL DESCRIPTIONS
Substituting x = x(A, t) into (3) and using (2a), we find
p[x(A, t), t] = c>(A,t). (4)
Equations (3) and (4) illustrate the transformation between the spatial
description and material description, where these terms* are defined as
follows.
Spatial description: x, t independent variables.
Material description: A, t independent variables.
To obtain a spatial description of air motion, one would measure (as time
progresses) velocity, pressure, temperature, etc., at various fixed stations
on the ground and in towers. To obtain a material description, one would
make such measurements in various freely traveling balloons.
Let us consider the relation between material and spatial descriptions of
velocity. In analogy with elementary point mechanics, we define the velocity
of a particle by the rate of change of that particle's position with time:
V(A t) = ax(A, t). (5)
' - a,
To preserve point-particle interchangeability, we must define v(x, t), the
velocity in spatial coordinates, by
v(x, t) = V[A(x, t), t).t (6a)
Instead of" spatial "and "material" description the terms Eulerian and Lagrangian descrip-
tion are often used. Both descriptions were originated by Euler, but Lagrange's contributions to
the development of particle mechanics were sufficient to justify his name being used.
t Equation (6a) must hold for the same reasons as those used to confirm Equation (3). Because
the argument is involved, we repeat it in somewhat different terms. (The required verbalization
deepens understanding of material and spatial variables.) Note first that V(A, t) is the velocity at
time t of the particle initially located at A. Also, in the equation A= A(x, t), the expression
A(x, t) denotes the initial position of the particle that now (at time t) is located at x. Therefore,
V[A(x, t), tJis the velocity at time t of the particle whose initial position is that of the particle now
at x. In other words, V[A(x, t), t] is the velocity at time t of the particle that is located at x at time t.
By point-particle interchangeability, this must be the velocity at time t at the fixed spatial point x.
Equations (6a) and (2a) imply that

v[x(A, t), t] = V(A, t). (6b)
When all dependent variables, like v(x, t) and p(x, t), are independent oft,
the flow is called steady.* Note that the term "steady" implies time inde-
pendence in the spatial description. Thus in steady flow of water through a
pipe of varying cross section, the velocity measured by a fixed probe will
always be the same. But the velocity of a given fluid particle will vary as the
particle speeds up through constrictions and slows down at wide sections
of the pipe.
Throughout this chapter (and most of this book) we shall determine the
components of vectors by referring to a fixed right-handed cartesian co-
ordinate system. For theoretical purposes numerical subscripts are best,
but for particular problems it is often convenient to adopt a simpler notation.
The two notations that we shall use are summarized in Table. 13.1.
TABLE 13.1. Two notations.
Theoretical discussions Particular problems
Unit vectors i, j, k
along coordinate
axes
Position vector
components
Material A,B,C
Spatial x,y,z
Velocity vector
components
Material V1, V2, V3 u, v,w
Spatial Vi, V2, V3 u, v, w
Theorem 1. If the particle paths are known, the spatial velocity field
v(x, t) can be determined, and vice versa.
Proof. Given x(A, t), we can find V(A, t) from (5). If we express A in
terms of x and t and substitute into V(A, t), then we obtain v(x, t), by (6a).
On the other hand, suppose that v(x, t) is known. Combining (5) and (6b),
we obtain
ox(A, t) = [ (A )
v x 't 't. 1 (7)
at
* The notion of a steady flow is used mainly in fluid mechanics. Displacement rather
than velocity is the central dependent variable in solid mechanics, and the displacement generally
depends on time even when the velocity is steady. Compare Exercise 12.1.8.
If we no longer explicitly indicate the dependence of x on its initial position

A and on the time t, we can rewrite (7) more simply as
OX
ot= ~(x, t). (8)
Since v(x, t) is assumed known, (8) can be regarded as a first order nonlinear
differential equation for x (or as a system of three equations for the three
components of x). The required initial conditions are simply
at t = t0 , X=A. (9)
Provided these equations have a solution (as we shall assume), we can find x. O
In dealing with differential equations, one usually does not explicitly
indicate the dependence of the unknown function on the initial conditions.
For this reason, in many texts (8) is written
dx
dt = v(x, t). (10)
Here x is regarded as a function of t only, with "fixed" initial point A.

The partial derivative notation of (8), which says "keep A constant," is
therefore not deemed necessary.
STREAMLINES AND PARTICLE PATHS

At time t a streamline is defined as a curve whose tangent vector at each
point is parallel to the instantaneous velocity vector at that point. Let s
be the arc length of the streamline x = x(s, t), which emanates from a point
x 0 at time t. Choose the sign of s so that dx/ds points in the same direction as
the velocity vector, not in the opposite direction. Lets= 0 at x 0 Then x(s, t)
is determined by the following equation and initial condition:*
dx(s) = v[x(s), t] , x(O) = x

ds Iv[x(s), t] I O (!vi ~O). (11)
(We have used the fact that dx/ds is a unit tangent vector.) If written in
subscript notation, the differential equations of (I I) become
3
dx, v,.
-=-, i=l,2,3; lvl2 = Iv;.
ds lvl i= 1
As above, we shall often use t = 0, rather than t = t 0 , for the initial time.
Similarly, if v2 0
-::/= or v3 -::/=0, one can write
v1 dx 1 v3 dx 3
-=-, -=- or
Vz dx 2 v2 dx 2
Such equations are often compactly summarized by
dx 1 dx 2 dx 3
-=-=- (12)
Vz
It can be shown that in steady flow, streamlines and particle paths coincide
(Exercise 5).
Example 1. Suppose the material description of a certain two-dimensional

motion is given by
t~O. (13)
Find the spatial description of the motion and show that the motion is steady.
Reverse the process, and find the material description from the spatial. Find
particle paths and streamlines.
Solution
(14)
Inverting (13) is trivial:

(15)
Substitution of (15) into (14) yields the spatial description

V[A(x, t), t] =v(x, t) = (x1e-)e'i- (x2e')e-j = Xii- x2i- (16)
Since vis independent oft, the motion is a steady one.

Working backward, given the spatial description
V= x,i- X2j, (17)
we find that
dx
-=v or (18)
dt
Solving (18) subject to the initial condition x(O) = A, we obtain
(19)
We thereby recover (13). Ast increases from Oto oo, one can thus determine from
(19) the path of the particle that was located at x = A when t = 0. In this case the
parameter t can easily be eliminated giving for the particle path a portion of the
hyperbola
(20)
The Continuous Medium [Ch. 13
Suppose that Ai> 0, A2 < 0. Then, for t > 0, Xi= Aie' implies that xi> Ai,
while x2 = A2e- implies that A2 < x2 < O; therefore the particle path is the
"upper" portion of that part of xix2 = A1A2 which lies in the fourth quadrant.
See Figure 13.2. (This example illustrates the care that must always be taken in
passing from a parametric representation of a curve to a nonparametric represen-
tation.)
FIGURE 13.2. The particle path determined in Example 1. As t in-

creases, x1 increases, so the arrow indicates the direction of motion.
The streamlines are found by using (12). Thus

dxi dx2
Hence
Inlxil = -Inlx2I +constant or X1X2= constant. (21)
Comparison of (21) and (20) shows that the coincidence of particle paths and
streamlines for steady flows is verified in this particular case.
The passage from particle paths to spatial velocity field requires as a key
step the inversion
x = x(A, t)--+A = A(x, t).
In the very simple example just discussed, this was easily accomplished; but
can it be accomplished in general? Another difficult mathematical question
arises in attempting to compute particle paths from spatial velocity fields.
Under what conditions is there a solution to the differential equation (8)
subject to the initial conditions {9)? We cannot delve into these matters
here, but we refer the reader to Chapter 2, where a start was made on presenting
some of the mathematics involved.
There are concrete problems that are easier to solve if the material descrip-
tion is employed. Examples can be found in the nonlinear theory of one-
dimensional compressible fl.ow phenomena. (See Courant and Friedrichs,
1948.) Such problems are the exception, however, since there is usually not
sufficient compensation for the extra effort required to ascertain the fate of
each particle. But although the spatial description is generally more conve-
nient than the material* when it comes to investigating specific phenomena,
the material description offers the more natural framework for the formula-
tion of the basic physical laws. It is therefore necessary to transform certain
expressions from material to spatial coordinates. Consideration of such
transformations begins in the next section.
A SIMPLE KINEMATIC BOUNDARY CONDITION
Materials are often subject to exterior constraints of a kinematic nature.
A simple but important example occurs when the material is in contact
with an impermeable boundary (wall). Since no material can penetrate such
a boundary, the velocity component normal to the wall must be zero. (No
penetration is caused by relative tangential velocity.) Ifn(x, t) is a unit normal
to the wall, the boundary condition for x on the wall is
v(x, t) n(x, t) = prescribed normal velocity of wall. (22)
When the wall is stationary, of course, v n = 0. Other kinematic boundary
conditions arise (e.g., at the interface between two media), but at this point
it is sufficient to indicate the existence of this type of constraint.
EXERCISES
tl. Suppose that the spatial coordinates x; of a particle at time t are given by
Find the velocity vector v(x, t); i.e., transform from the material to the
spatial description.
2. In a two-dimensional steady flow the velocity components are given by
u=y, v=x.
(a) Show that the streamline passing through (x 0 , y 0 ) is a hyperbola.
t(b) Find the particle path that commences at (x 0 , y 0 ). Note the coinci-
dence of the particle path and the streamline.
t3. In a two-dimensional unsteady flow the velocity components are given by
u = t, v = x +I (tis the time).
(a) Find the streamline passing through (x 0 , y 0 ).
(b) Find the particle path that commences at (x 0 , y 0 ).
For certain classes of models, notably those of linear elasticity, the difference between
material and spatial variables is of the order of terms already neglected. Consequently, the
distinction between these variables is essentially negligible.
t4. In a two-dimensional flow, the x and y velocity components are given by

u=e-'cost, v = I.
(a) Find the streamlines at time t.

(b) Find the particle paths, expressing the constants of integration in
terms of the initial coordinates A and B.
S. Give a physical argument showing that particle paths and streamlines
coincide in steady flow.
6. Interpret (6b)in the manner of the text's interpretationof (6a).
13.3 The Material Derivative

To provide physical motivation for the mathematical content of this sec-
tion, consider the fact that fishing boats in Norwegian fjords sometimes act
sluggishly because they generate internal waves on the underwater interface
between fresh fjord water and heavier sea water beneath. The time in which
significant diffusion of salt takes place is very large compared to the period
of the waves generated. As a consequence, one can assume that each packet
of fluid maintains its density as it moves. Formally, this incompressibility
assumptionis
a~(A, t) = O
(1)
ot '
where 8(A, t) is the density at time t of a water particle that was initially at A.
The reader should contrast (1) with the assumption of uniform density
8(A, t) = constant, (2)
which is not appropriate for this problem.

In Section 15.2 we discuss waves in layers of incompressible fluid whose
density varies with height. Here we wish only to examine how equations like
(1) can be transformed from material to spatial variables. We shall obtain a
formula that is fundamental not only for the problem of internal density
waves but for the entire formulation of continuum mechanics.
Before proceeding, the reader should review the chain rule (Appendix
13.1). Use of the careful notation discussed in this Appendix is essential for
secure understanding of what follows.
Recall that in spatial variables, we denote the density by p(x, t) so that
p(x, t) = o[A(x, t), t], o(A, t) = p[x(A, t), t], (3)
as in (2.3) and (2.4). In material variables the density is described by

8(A, t). When we compute a partial derivative of <5 with respect to t, we
regard the initial position A as fixed. On the other hand, when we compute
Sec. 13.3] The Material Derivative
427
such a partial derivative of p(x, t), the density described in spatial variables,
we regard the spatial position x as fixed. Since A and x are related, we can
relate the two derivatives by the chain rule [compare (5), Appendix 13.IJ:
ob(A, t)
ot =
op I
OX1 x=x(A,t)
OX1
at+
Op I
OXz x=x(A,t)
OXz
at
+-op I
OX3x=x(A,r)
ox3
-+-
ot
op
ot
l
x=x(A,t)
(4)
As in (2.5), ox;{ot evaluated at (A, t) is simply V;(A, t), the ith velocity
component expressed in material variables; i = I, 2, 3. Our goal is an expres-
sion for the left-hand side of (4) in spatial coordinates. To obtain this, we
must make the substitution A= A(x, t), which yields
ob(A, t)
Ot
IA=A(x,t)
= I
i=1
vi(x, t) op(x, t) + op(x, t)'
OXi Ot
(5)
where we have used the relation
V;(A, t) IA=A(x, r) = v;(x, t). (6)
The notation D/ Dt is frequently employed for the material or substantial

derivative which appears on the left-hand side of (5). With this we have
ob(A, t) I = Dp(x, t). (7)

ot A=A(x, I) Dt
Dp(x, t) _ op(x, t) ( ) .V ( )
Dt - ot + v x, t p x, t . (8)
It is important that the reader fully understand the meaning of (7) and
(8). Remember that the function ob(A, t)/ot gives the rate of change with
time of the density of the particle whose initial position was A. If the substitu-
tion A = A(x, t) is made in the afore-mentioned function of A and t, then a
function of x and t results. This gives the rate of change with respect to time
of the density of the particle now located at x. With (8) the right side of (7)
is a calculable function of x and t. Thus (7) and (8) allow computation of the
rate of change of particle density when all variables are referred to spatial
coordinates. We see that this computation requires knowledge of the spatial
description of density p(x, t) and of the spatial description of velocity v(x, t).
We remark that in spatial variables, the condition for incompressibility is
Dp/ Dt = 0. A motion satisfying this condition is sometimes called isochoric.
We have used the density throughout our discussion only for concreteness.
For any quantity, the material derivative tells the time rate of change of that
quantity for a fixed particle. Thus, to give an important example, if v(x, t)
428 The ContinuousMedium [Ch. 13
denotes the velocity of the particle at position :x and time t, then Dv(:x,t)/ Dt
denotes the acceleration of the particle located at position x at time t. That is,
oV(A, t)
ot
I
A=A(s, r)
= Dv(x, t) ,
Dt
Dv av
-=-+(v
Dt ot
V)v. (9)
The ith component of Dv/Dt is given by the formula
Dvi OV1 ~ ovi

-=-+ ~VJ-
Dt ot 1=i ox1
We have derived the important formulas (7) and (8) by a careful application of
rather standard mathematics. Having accomplished this, we shall reexamine the
material derivative in a way that is less careful but which perhaps sheds more
light on its physical meaning.
The time rate of change of the density of a particle is the result of the following
two effects:
(a) At the spatial position x the density changes with time.
(b) The particle at x at time t arrived there from another part of the material,
so there may be a density change even in a steady motion for which
ap(~,t) ===O.
:;Jt
Effect (a) is clearly given by the first term of (8). We now show that effect (b) is
given by the second tenn of this equation.
In the very short time interval (t, t + At), the particle located at x at time t moves
a distance approximately given by lv(x, t) I At. This motion is (approximately) in
the direction of the unit vector v(x, t)/ Iv(x, t) I The change of p per unit length in
this direction (at a fixed time) is (v/lvl) 'ilp. The change of p due to motion
through the distance IvI Jl.t is the product of the distance traveled and the change
per unit distance, i.e.,
. (lvl At)(i:
1
vp) =(v, Vp)At.
This is the change in time Jl.t.To find the change per unit time, we must divide by il.t,
which gives the desired result.
EXERCISES
tt. Suppose that p(:x,t) =Xi+ x 2 sin t; x 1 =Ai+ t, x 2 = A 2 + t 2 , x 3 = 0.
Verify (8) in this case by explicitly computing both sides of the equations.
[Find ~(A, t); compute o~(A, t)/ot; etc.]
2. (a) Show that the flow given by
x = A + t sin wt, y= B (t and a, constants)
is not a steady one. Describe it.
(b) If the density is given by p(x, t) = p 0 /(l + x 2 ), calculate Dp/Dt.
Recall that the directional derivative of a function is obtained by taking the scalar product
of the gradient of the function with a unit vector in the appropriate direction.
Sec. 13.4] The Jacobian and Its Material Derivative 429
3. Consider the two-dimensional motion whose material description is,

for some constant w,
x 1 = A 1 cos wt - A 2 sin rot, x 2 = A 1 sin rot+ A 2 cos rot.
(a) Show that in spatial coordinates, the velocity v(x, t) is given by
v = -wx 2i + roxd
so that the flow is steady.
(b) Show that the particle paths are circles about the origin.
(c) Calculate the streamlines at time t and thus show that they coincide
with the particle paths.
(d) Find V(A, t) and verify (9).
4. In the Eulerian description, a flow is given by the steady velocity
field
Vz = X1 -Xz,
(a) Find the corresponding Lagrangian description.

(b) Plot the path of particles that pass through the point x 1 = 2,
x 2 = 0, sometime during the motion. Note that all such particles
have the same path, since the motion is steady.
(c) Find the acceleration of the particle that is instantaneously located
at (3, 4). Comment on the fact that the answer is nonzero, although
the flow is steady.
5. Let x 1 = A 1 exp t, x 2 = A 2 exp (-t) (as in Example 1 of Section
13.2). Show that p(x, t) = x 1x 2 is the density of an incompressible
motion in two ways:
(a) By computing Dp(x, t)/ Dt.
(b) By computing oJ(A, t)/ot.
13.4 The Jacobian and Its Material Derivative

Having introduced via x = x(A, t) the material description of a moving
continuum, we can make precise the notion that a specified portion of material
occupies the region R(t) at time t. We specify the portion of material by giving
its location at some instant of time, say t = 0. Denote by R(O) the region
occupied at t = 0 by the designated portion of material. Let the subsequent
motion of any point x bedescribed by x = x(A,t). Thenthe region R(t) occu-
pied by the designated material at time t is in mathematical language the
image of R(O) at time t under the mapping x = x(A, t). More informally,
a point xis in R(t) if, at time t, x = x(A, t) for a point A in R(O).
Let .Y(t) denote the volume of R(t). Then
?"(O) = fff dA 1 dA 2 dA 3 and .Y(t) = fJfdx 1 dx 2 dx 3 (la, b)

R(O) R(r)
Although any letters could have been used for the three dummy variables of
integration in (1), it is suggestive to use the "initial" variables Ai, A 2 , and
A 3 in (Ia) and to use in (lb) the variables Xi, x 2 , and x 3 associated with the
time t. In (lb) we can change the variables of integration to A 1 , A 2 , and A 3 ,
where A= A(x, t). According to the Jacobian rule
"Y(t) = fff J(A

R(O)
1, A 2 , A 3 , t) dA 1 dA 2 dA 3 , (2)
where the Jacobian is defined by the following determinant:
OX1 OX1 OX1

0A 1 oA 2 oA 3
O(X1,X2, X3) OX2 OX2 OX2
J(A 1, A 2, A 3 , t) =o(A A2,1,
A )=
3 oA 1 oA 2 c3A3
OX3 OX3 OX3
oA 1 oA 2 oA 3
No absolute value sign is needed on J, since the components of x and A

are both determined in right-handed coordinate systems.
Using the integral mean value theorem (Appendix 13.2), we see that (2)
implies that
'1'"(t) =J(A1, A2, A3, t) IffdA1 dA2 dA3 =J(A1, A2, A3, t).Y(O)
R(O)
(3)
for some point (A 1 , A 2 , A 3 ) in R(O).

We now wish to consider (3) as R continues to "shrink" about a fixed
point (Ai, A 2 , A 3 ). To make this idea precise, let us apply (3) to each of a
sequence of similar regions R,,(0) whose volume '1'"n(O)approaches zero as
n-+ oo. (Roughly speaking, "similar" means "having the same shape,"
but the idea of similarity, one that we shall use several times, is made precise
in Appendix 13.3.) Every region R,, is required to contain the fixed point
(A1 , A 2 , A3 ). According to (3),
n = 1, 2, 3, ....
But, since the R,,are similar and have ever-decreasing volume,
lim (A~">,A~">,A~">)= (A 1, A 2 , A 3),

..... 00
as is proved in Appendix 13.3. Thus, if J is continuous at (A 1, A 2 , A 3 ),
(4)
Sec. 13.4] The Jacobian and Its Material Derivative 431
More geometrically but less precisely, one can say that the Jacobian at
point A and time t represents the dilatationof an infinitesimal volume initially
at A, where the dilatation is the ratio of volume occupied by the infinitesimal
material region at time t to its initial volume. The importance of the Jacobian
is confirmed by our ability to describe it, as in (4), in a manner that does not
depend on the use of a particular coordinate system.
In later work we shall frequently need an expression for oJ(A, t)/ot, the
rate of change of the dilatation.* We shall now find this expression in a
straightforward manner for the two-dimensional case. The three-dimensional
case will be left as an exercise. A compact and elegant derivation is afforded
by the machinery of cartesian tensors (Example 3 in Section 1.2 of II).
We shall employ the following formula for the derivative of a determinant:
dAdB AB
!!_'A(t) B(t),_ dt dt (5)
dt C(t) D(t) - + dC dD '
. C D dt dt
which the reader can easily verify. Using (5), we have the following expression
for the partial derivative of the Jacobian with respect to time, keeping the
initial coordinates A 1 and A2 fixed :
OX1 OX1 oVi av1 OX1 OX1

oJ(A1, A2, t) a oA1 oA2 0A1 0A2 oA1 oA2
=- = + (6)
at at OX2 OX2 OX2 OX2 oV2 oV2
aA1 aA2 0A1 0A2 aA1 aA2
We have employed the result ax/at = V of (2.5). We have also assumed x 1 and
x 2 to be twice continuously differentiable so that orders of differentiation
can be interchanged. For example,
By the chain rule, the first of the two determinants on the right-hand side
of (6) can be written
I
OV1axl + OV1OX2
OX10A1 OX20A1
I I
OV1OX1+ OV1OX2
OX10A2 OX2 0A2
I
(7)
OX2 OX2
oA 1 oA2
* Some authors define the dilatation as the limit of ['f'".(t)- 'f'".(O)]/"l'".(O).Whether this
definition or (4) is used, 8J/8t still gives the rate of change of the dilatation.
432 TheContinuous Medium [Ch. 13
In (7) we have indicated by a vertical line that x must be expressed in terms

of A and tin ov1/ox1 and ov1 /ox2
To obtain the final simplification in a generalizable way, we notice that
(7) can be written as the sum of two determinants:
I
OV1 OX1
OX1 0A1
I
OV1 OX1
OX1 0A2
I
OV1 OX2
OX2 0A1
I
OV1 OX2
OX2 0A2
+ (8)
OX2 OX2 OX2 OX2
oA1 oA2 oA1 oA2
The second determinant in (8) is zero, since it equals ovtfox2 times a deter-
minant with two identical rows. The first determinant on the right side of (6)
thus reduces to (ov1/ox 1)J. Similarly, the second determinant on the right
side of (6) reduces to (ov2 /ox 2 )J, giving
oJ(A1 , A 2 , t)
ot
= (ov 1 + ov2 )
OX1 OX2
I
x=x(A, f)
J(Ai, A 2 , t). (9a)
Analogous calculations (Exercise 3) yield the three-dimensional result
-------A2' A3' t) - (OV1+ OV2+ OV3)

0J(A1,
at OX1 OX2 OX3 x=x(A,f)
I
J(A 1, A 2, A 3,t. ) (9b)
lfwe make the substitution A = A(x, t) [i.e., if we write (9) in spatial variables]
and use the definition of V v, we have the Euler expansion formula,
DJ[A~~ t), t] = (V v)J[A(x, t), t]. (10)
Equations (9) and (IO) look simpler when written as follows.
a1
In material variables: ot= [(V v) lx=xCA, r)]J. (11)
DJ
In spatial variables: -=(Vv)J . (12)
Dt
We emphasize that the derivatives required by the divergence V are with

respect to spatial variables.
The reader may choose to remember the Euler expansion formula as (11)
or as (12) or as: "The derivative of the Jacobian is the divergence of the
velocity times the Jacobian." When actually using the Euler formula, how-
ever, he risks trouble unless he keeps the independent variables straight, as
in (9) and (10).
Appendix 13.J] On the Chain Rule of Partial Differentiation 433
EXERCISES
1. Give an example of a sequence of regions (R,J with the following prop-
erties:
(i) A and B are distinct fixed points (independent of n) interior
w
each~egion Rn, n = l, 2, ....
(ii) The volume of Rn approaches zero as n -+ oo.
2. Use induction to generalize (5) to a result for an n x n determinant.
3. Generalize the proof given in the text so that (10) holds for the three-
dimensional case J = J(A 1, A 2 , A 3 , t).
t4. Calculate the Jacobian of the two-dimensional motion and verify (12)
in the following cases:
(a) x 1 = A 1 exp t, x 2 = A2 exp (-t), x 3 = A3
(b) x 1 =Ai 12 expt, x 2 =A 2 exp(-t), x 3 =A 3
S. By direct calculation show that o(x 1 , x 2 )/o(A 1 , A2 ) = l for the motion
of Exercise 3.3. What is the significance of this result?
Appendix 13.1 On the Chain Rule of Partial Differentiation

By means of an example we illustrate the careful notation that is essential
if confusion is to be avoided in all but simple applications of the chain rule.
Suppose that a certain quantity is given as a function of the Cartesian
coordinates x and y. For definiteness let us consider the temperature of acer-
tain body B. In the present discussion we shall use the notation T = T(x, y).
Suppose that polar coordinates (r, 0) are introduced so that
x = r cos 0, y = r sin 0.
Denoting the temperature of B as described in polar coordinates by -r(r, 0),
we have
-r(r, 0) = T[x(r, 0), y(r, 0)], T(x, y) ==-r[r(x, y), O(x, y)]. (I)
For example, if
T(x, y) ==x 2 + sin y, (2)
then on introducing polar coordinates, we find
-r(r, 0) = r 2 cos 2 0 + sin (r sin 0). (3)
Note that
T(r, 0) = r 2 + sin O -r(r, 0). (4)
If we wished to find o-r/06,
the chain rule could be used as follows:
o-r ar
oo= ox
Ix=rcos9
ox or
08+ oy
I
x=rcos9
oy
08
(5)
7=rsin9 1=rsin9
The Continuous Medium
434 [Ch. 13
The notation is unambiguous. For example, since Tis a function of x and y,

y is held constant in computing aTJox. The equations associated with the
vertical line indicate that after computing aTJox,one should substituter cos 8
for x and r sin 8 for y. Thus (5) becomes
~:=2xl (-rsinll)+cosyl (rcos8)

uc, :r=,cos 9 :r=,cos 9
:,,=,sin 9 :,,=rsin 9
= - 2r 2 sin 8 cos 8 + cos (r sin 8) r cos 8.

The same result can, of course, be obtained directly from (3).
A proof of the chain rule is given in calculus books, but sometimes care
is not taken to distinguish between expressions such as T(r, 8) and -r(r, 8).
In these books an implicit change in the usual functional notation is made in
discussing certain topics such as the chain rule. When r and 8 are used
as arguments of T, then T refers to the function in (3) that we call -r. When
x and y are used as arguments, the original function (2) is meant. Thus the
symbol Tis ambiguous. In our work such ambiguity would lead to confusion.
EXERCISES
1. Derivation of the following equation is requested in Exercise 2 of
Appendix 14.1:
a0 aP a0
-+ 0--P-=0 (V, E independent). (6)
av aE aE
Here V is the volume of a system, 0 is its temperature, E its internal
energy, and P its pressure. Any two of these variables can be regarded
as dependent and the other two as independent. Show that for the five
other choices of independent variables the above relation implies that
0 _ P(0) a(0, E) = O (V, P independent), (7)

ap + a(V,P)
aE aP
--0-+P=O (V, 0 independent), (8)
av a0
aE av av
-+0-+P-=0 (0, P independent), (9)
aP a0 ap
a(P, V) av
9 8(0, E) - p oE - l = O (0, E independent), (10)
a0 _ av_ o(V, 0) _
0 0 (P, E independent). (11)
oP aE p o(E,P) -
Appendix 13.2] The Integral Mean Value Theorem
435
A good way to proceed is as follows. First consider a general variable

change and specialize later to all the particular cases. To rewrite (6)
using new independent variables x and y, define 0(x, y) by
E>[E(x,y),V(x,y)] = 0(x,y).
Recall the chain rule results
ex= eEEx + 0y vx, e,.= eEE)I + E>vv)I.
Express eE as the following quotient of Jacobians:
0E = a(0, V)/a(x, y) .
a(E, V)/a(x,y)
Find similar expressions for 0v and PE. Proceed by successivelyidenti-
fying (x, y) with the various independent variable pairs of (7)-(ll).
N o T E Equations (7)-(11) have been written in the " usual " nota-
tion in which, to give an example, no distinction is made between
the internal energy regarded as a function of P and V, or as a function
of p and e.
t2. Using the notation of (I), by chain rule express
air air]
[ -+-
axi ayi .,.=,cosll,:,=rsinll
in terms of -r(r, 8) and its derivatives.

3. Exercise 2 requested the transformation of the Laplacian into polar
coordinates. Make the transformation into spherical coordinates by
expressing
air air air]
[ -+-+-
axi ayi azi x=psin4>cosll,:,=psin4>sinll,:z=pcos4>
in terms of -r(p, <J,,
8) and its derivatives, using one of the following
possibilities.
t(a) Do the problem in full generality (a rather long calculation).
(b) Assume that aJa<J,= o.
(c) Assume that a/a<J,= a/a8 = 0.
Appendix13.2 TheIntegralMeanValueTheorem
1beorem. Let R be a connected closed bounded region of three-dimen-
sional space with volume V, 0 < V < oo. Letfbe continuous on Rand denote
by M and m the maximum and minimum values off (x) for x in R. Then
there exists a point i in R such that
ffff(x) d-r=f(i)
R
V. (1)
The Continuous Medium [Ch. 13
Proof. From the definition of the triple integral as a limit of sums, it

follows that
JfImd-c~ IfIf(x)
R R
d-c~ JfIMd-c,
R
so
mVS IIIf(x) d-c~ MV

R
or mS v- 1 IIIf(x)d-c SM.
R
(2a, b)
Since a continuous function defined over a closed bounded region is bounded

and takes on its upper and lower bounds, there exist points x<111> and x<M>
in R that satisfy
The hypothesis of connectedness assures the existence of a space curve

x = x(t) that joins x<111>and x<M>
and lies within R:
x(t) in R for 1< > S t ~ r<M>.
111
Sincef[x(t)] is a continuous function of the parameter t, it must take on every

value between m and M, by the intermediate value theorem. In particular,
from (2b), there exists a parameter value i between tC>and tCM>such that
f[x(l)] = V'""1 fJff(x) d-c. (3)

R
The conclusion (1) follows immediately from (3) and the definition i = x(i). O
REM ARK s . The integral mean value theorem holds for n-dimensional
connected closed bounded regions R in m-dimensional space; n ~ m,
m = 1, 2, .... Only slight changes in the wording of the proof are necessary.
If F(x) is a vector function of x, then (1) obviously holds for each compo-
nent F; of F. There will generally be a different point i for each component.
Appendix 13.3 Similar Regions

In this appendix we shall sketch the derivation of certain useful properties
of "star-shaped" regions. We then generalize to unions of such regions.
Definition1. A closed region of space R is called star-shaped if there exists

an interior point of R (called 0) that can be connected to each point of R
by a straight line lying within R. Figure 13.3 illustrates a planar star-shaped
region and a possible point 0.
Appendix 13.3] Similar Regions
437
FIGURE 13.3. Planar star-shaped region. The inner region is similar

to the outer.
If we choose O as the origin, then R can be described using spherical

coordinates (p, , 8) by the relations
0 ::s;p ::s;f(4>,
8),
(The coordinate p merely provides the distance from point O to a given point
in R, along the line joining the two points.) We shall assume that f is con-
tinuous and piecewise continuously differentiable with/{, 0) = f(</J,21t).
Definition 2. Two star-shaped regions R<1 > and R<2 > are called similar
if, after possible rotation and translation, they can be described by the equa-
tions
i = 1, 2, (1)
where 1<2 > = Lj< 1 >for a positive constant L (Figure 13.3).
We now show that the volumes V[R]and the surface areas S[R]of
the similar star-shaped regions Rare in the ratios L 3 and L 2 , respectively.
Proof. Using the formula for volume in spherical coordinates, we have
V[R<2 >]=f f f p sin dp d<pd8

2n
0
,.
0
Jl2l(q,, 9)
0
2
= t f f [j< >(,8)]3 sin q,dq,d8.

2n ,.
2 (2)
0 0
The result follows at once from the relation 1<2 >= LJ<1 >. D
Theorem 2. S[R<2 >]=L 2 S[R<1 >].

Proof. Using (1), if a point x = (x 1, x 2 , x 3) is on the boundary of R<i),
it can be described by the relation x = x, where
x~>= J<i)(q,, 8) sin</>cos 8,
x~>= j<i)(</>,8) sin</>sin 8, (3)
x~>= jCi)(q,, 8) cos q>.
We regard (3) as a parametric representation of the bounding surfaces and
use the appropriate formula for surface area to write*
S[R(i)] = tz..Ia..Iu;</>/\ u;8 Idq,d8.

!l (I) !l (i)
(4)
The result follows from the relation x<2 >= Lx<1 >. D
Another result that will prove useful involves a sequence of mutually

similar star-shaped regions R<n>.
R(nl: p 5,f(n>(q,,8), J<n>=Lnf< 1 >; n = 1, 2, 3, .. ,. (5a, b)
Let Dn be the maximum distance between two points in RC">.
Theorem3
As n-+ co, V[R<n>]-+0 if and only if D"-+ 0. (6)
Proof. Using Theorem 1, V[R<n>] = L!V[R< 1 >], so V[R<">]-+0 if and
only if Ln-+ 0. By (5b), L"-+ 0 if and only ifj<">-+ 0. This implies that
V[R<">]-+0 if and only if p~Jx-+0, (7)
where
P:::lx= max J<">(q,,8).
0:S<l>:S!t
0 :$8:S z,,
[The existence of p~Jxis guaranteed, since/Cn>(q,,8) is continuous in the closed
bounded region O 5. </>5. n, 0 5. 8 5. 2,r.] It is not difficult to show (Exercise 2)
that
and (Sa, b)
The result follows from (7) and (8). D
'lbeorem 4. The three previous theorems remain true for regions that are
the union of a finite number of star-shaped regions.
Proof. Left to the reader.
The letter x is employed so frequently that, to avoid confusion, we use a wedge to denote
the cross product.
Appendix 13.3] Similar Regions
439
EXERCISES
1. (a) Use polar coordinates to define planar star-shaped regions.
(b) Show that " similar triangles " are similar in the sense used here.
(c) State and prove versions of the above theorems which are appro-
priate for planar star-shaped regions.
t2. (a) Prove (8b).
(b) Prove (8a).
3. Show that for the family of mutually similar regions R<n>,
V[R(n)]= ARL~, S[R(n)]=RL;,
where AR and R are independent of n. (This result is used in Section
14.2.)
4. Formulate and prove for regions that are the unions of star-shaped
regions:
(a) Theorem 1.
(b) Theorem 2 (note that corresponding regions must be "similarly
located," in a sense that should be explained).
(c) Theorem 3.
CHAPTER 14
Field Equationsof ContinuumMechanics
IN THIS chapter we shall provide careful derivations of the field equations

that restrict the behavior of any continuous medium. These equations
express conservation of mass, and balance of linear momentum, of angular
momentum, and of energy. In addition there are Gibbs relations, which
couple heat flow with entropy and temperature, and couple work done with
pressure and volume. Finally, there is an entropy inequality.
Only " ordinary " mechanical effects will be considered-electrical, mag-
netic, and chemical effects will be ignored.
In the concluding section of the chapter we give a short discussion of a
topic to be taken up much more fully in the remainder of the book, the differ-
ent constitutive equations that distinguish different media. It is important
to bear in mind that such equations (indeed, all physically meaningful
equations) must not differ essentially if a new coordinate system is employed.
This matter is discussed briefly, although it is only taken up fully in II. We
conclude with a few additional remarks on the validity of the continuum
model.
14.1 Conservation of Mass

Derivation of the equations governing continuum mechanics is not an
easy matter. Hard study is necessary before real understanding is gained.
Although it may initially be confusing to contemplate more than one approach
to a topic, nevertheless we shall present four ways to derive the equation of
mass conservation. Consideration of the advantages and deficiencies of all
four approaches will lead the reader to a deeper understanding of the issues
involved in formulating equations that describe natural processes. We aim
not to give an excuse for using a "well-known" equation but rather to
attempt a derivation that will satisfy a critical person. We hope thereby to
aid those readers who, sometime in the future, might attempt to derive
equations governing phenomena not previously formulated in mathematical
terms.
The first method begins with a statement of mass conservation for an arbi-
trary material region. The central technical content of the derivation is a
computation of the time derivative of an integral over a moving material
region. Ultimately, a volume integral over an arbitrary region is shown to
vanish. The desired differential equation follows by the Dubois-Reymond
lemma.
440
Sec. 14.JJ Conservationof Mass
441
The second method, an application of the divergence theorem, is based

on a statement of mass conservation for an arbitrary spatial region. Botti.
the third and fourth methods are based on invoking mass conservation for a
fixed rectangular box. In the third approach the box is infinitesimal; in the
fourth, suggestive but dangerous manipulation of infinitesimals is avoided
by means of the integral mean value theorem.
INTEGRAL METHOD: ARBITRARY MATERIAL REGION

Section 13.4 gave a precise characterization of the idea that a specified
portion of material occupies the region R(t) at time t. That the mass of a
portion of material does not vary as time increasesis one way to state what is
meant by conservationof mass. Thus we hypothesize that
d
dI
fffp(X1, X2,
R(t)
X3, t) dx1 dx2 dx3 = 0, R(t) arbitrary, (1)
where p(x, t) denotes the density at point x, time t.

Since (1) holds over an arbitrary material region R(t), we shall be able to
transform it into the more tractable form of a differential equation by means
of the Dubois-Reymond lemma, which follows.
Suppose that
(2)
for every regionR containedin a domainD. /ff(x) is continuousfor x in D, then

/(x) = 0 for x in D.
The simple proof is given in Section 4.1.
To take the derivative inside the integral in (1), as required for (2), we
shall transform to initial coordinates A via the material description :x =
x(A, t). The region of integration will thereby no longer depend on time.
Thus, denoting the integral in (1) by /(t) (the integral is a function only of
time), we have
di
dt
=~ff
dt
I p(:x,t) dx
R(t)
1 dx 2 dx 3 = ddff I p[:x(A,t), t]J(A, t) dA
t R(O)
1 dA 2 dA 3
(3)
Here R(O)is the region occupied by the material when t = 0. We have changed
the element of volume by using the Jacobian rule.
We can write (3) more simply as
df= ~ ffI c5(A,t)J(A, t) dt

ut ut R(O)
by employing the material density c5(A,t) = p[x(A, t), t] of (13.3.3). We now

take the time differential inside the integral and use partial derivative notation
Field Equations of Continuum Mechanics [Ch. 14
to indicate that A is held constant during the differentiations with respect

tot:
di
-d =
fff[oc5(A,
O
t) oJ(A, t)]
J(A, t) + c5(A,t) O dA1 dA2 dA3, (4)
t R(O) t t
Using the Euler expansion formula (13.4.11), we obtain
di = t) + c5(A,t)(V v) I
Jff {oc5~A, }J(A, t) dA1 dA2 dA3.
dt R(O) vt x=x(A, t)
Making the variable change A= A(x, t) and using (13.3.7), (13.3.3), and the
relation between x and A given in (13.2.2b), we then obtain
ddl
t
=fff {DpDt (x, t) + p(x,
R(t)
t)[V v{x, t)]} dx 1 dx 2 dx 3 (5)
But dl/dt = 0, by (1). From (5) we thereby conclude that
Dp
-+pVv=O (6)
Dt '
where we have assumed that the integrand in (5) is continuous and therefore
have used the Dubois-Reymond lemma. Equation (6), often called the
continuity equation, is the desired differential equation form of mass con-
servation. Since Dp/ Dt = op/ot + v Vp, an alternative form of (6) is
op
ot + V (pv) = 0. (7)
We have seen how the time derivative of an integral over the moving
material region R(t) can be calculated by the device of shifting from R(t)
to R(O) and back. It is appropriate here to present two useful formulas
that essentially record the results of using this device in commonly encoun-
tered contexts. Proofs are left to the reader (Exercise 1).
For a sufficiently smooth function F and material region R(t), assuming
conservation of mass, it is not difficult to show that
:t fffFp dT =fffr;:
R(t) t
p dT.
R(t)
(8)
From a formal point of view the above result can be regarded as permitting one to
take d/dt inside the integral provided that one recognizes the "moving" variables
of integration by changing this operator to the material derivative DI Dt. If this
were permitted, we would have
~t fffFp d-r= fff~F

R(<) R(r) /
p d-r+ fff
R(<)
F:
t
(p d-r). (9)
Sec. 14.1] Conservationof Mass
443
Comparison of (8) with (9) shows that the validity of the well-established former
equation requires D(p dT)/Dt = 0. But this is reasonable, as D(p dT)/Dt can be
interpreted as giving the change with time of the mass of a moving material element.
A result closely related to (8) is the following. For any sufficientlysmooth

function G and material region R(t),
d
dt
fffG dt = fff0:iGdt + ffGv n du.*
R(t) R(t) vt oR(t)
(10)
Here, as usual, n is the unit exterior normal to R, dr is a volume element,

du is a surface element, and oR denotes the boundary of R. Both (8) and
(10) are valid if F and Gare replaced by vector fields F and G [Exercise l(d)].
Equation (10) is the Reynolds transport theorem. Its interpretation is
illuminating. Let Rt be a region fixed in space that coincides with R(t) at
time t. Certainly,
:tfff R,
G dr = fffaa~dr = fffaa~dr.
R, R(t)
(11)
(A partial derivative notation must be used under the integral sign, to show
that the integration variables are held constant during the t differentiation.)
But in computing the left side of (10), the contribution of (11) must be
supplemented by
ffGv n du =ffGv n du.

DR(t) DR,
(12)
We assert that the surface integral (12) can be regarded as giving the rate
at which a substance of concentration G units per unit volume is being
carried by the material as it flows through the boundary of Rt. To see this,
observe that v n provides the component of the velocity v that points
outward from the boundary. Suppose, for the moment, that v is independent
of time. In U units of time, a column of material Uv n units long and du
in cross-sectional area passes through surface element of area du. (Tangential
flow does not cause material to pass through the element.) As a consequence,
material passes through the element at the rate v n du volume units per
unit time. Even if v depends on time, v n gives the instantaneous volume
flux density through the surface element. (Flux density means flow rate per
unitarea.)If the materialcontainsa substanceof concentration
G unitsper
unit volume, then the flux of this substance through the surface element is
Gv n du. The total flux of G is given by (12), a result that we shall often use.
Putting the pieces of our discussion of (10) together, we see that the
derivative with respect to time of JJJR<tl G dt is equal to. the derivative with
As in (10), we frequently employ the traditional notation to indicate that an integral is taken
over a closedsurface. But we have not attempted to be completely consistent in this practice.
444 Field Equations of Continuum Mechanics [Ch. 14
respect to time of the integralof G over the statio11aryregion R.,that instanta-

neously coincideswith the material region R(t), plus the flux of G through the
boundaryof R.,.
INTEGRAL METHOD: ARBITRARY SPATIAL REGION
We start with the statement that conservation of mass requires that for
a regionfixed in space the rate of increaseof the mass containedin this region
must equalthe net massflux into the region.Only flow of mass across the bound-
ary can cause a change in the mass of the material contained in a fixed region,
since mass is neither created nor destroyed. In mathematical terms, then,
the italicized statement is
~ fJfp(x, t) dr: = - ffn(x, t) v(x, t)p(x, t) da. (13)

ut R IIR
Here R is an arbitrary region, fixed in space; oR is its boundary; and n

is a unit exterior normal to oR. Using the divergence theorem, we can write
(13) as
J[f{op~, t) + V [v(x, t)p(x, t)]}

dt = 0. (14)
(Partial derivative notation is used in the first term, because the variable of
integration, x, must be kept fixed during the time differentiation.) Since R
is arbitrary, if the integrand is continuous, then (7) follows by the Dubois-
Reymond lemma.
SMALL BOX METHOD

The ideas we wish to illustrate are more clearly seen if we confine ourselves
to two-dimensional situations, with velocity components u and v in the direc-
tions of increasing' x and increasing y, respectively. We shall assume that
there is no velocity component in the z direction, and that u, v, and the den-
sity p are independent of z.
We again apply the idea that the rate at which mass in a region increases
must equal the mass influx across its boundary. In contrast to the previous
method, here we .consider a special region. This is a " small box " of unit
thickness in the z direction but of small dimensions liX and .ll.yin the x and y
directions (Figure 14.1). The box is an imaginary one; fluid freely passes
through faces F 1, F 2 , F 3 , and F4 shown in Fig. 14.1. Because of the assump-
tion of two-dimensionality, no fluid passes through the faces that are parallel
to the xy plane.
We shall reason informally. The reader is asked to abandon skepticism
for the moment. The mass flux across face F1 into the box is (pu)0 .ll.y,where
the subscript indicates that pu is to be evaluated at (x 0 , y 0 ). (The z coordinate
can be ignored, since there is no variation in the z direction.) Since .ll.yis
Sec. 14.1] Conservation of Mass
445
t:.y
L.
z
FIGURE 14.1. Imaginary box used for the derivation of the continuity
equation in cartesian coordinates.
very small, we can probably ignore the fact that pu is not strictly constant
over F 1
The mass flux into the box across F3 is
o(pu)] }
- { (pu)o + [ ax"" o /u Ay. (15)
The second term in the square brackets is a correction to account for the
fact that pu is to be evaluated on F 3 , not F 1 Since o(pu)/ox gives the change
in pu per unit length, we should get the required correction by multiplying
by /u, the actual distance between F1 and F 3 As /u is small, it is probably
permissible to evaluate o(pu)/ox at x = x 0 The negative sign in (15) comes
from the fact that at F3 a positive u velocity component means net outward
mass flux.
Adding the contributions from F 1 and F 3 , we find a net mass influx of
-[o(pu)] /u l!y. (16)

ox o
Similarly, or by symmetry, the net mass influx across F2 and F4 is
-[o(pv)]/u l!y. (17)

oy o
The sum of (16) and (17) must equal the net rate of mass increase as expressed
by means of a change in density, i.e.,
{-[o(pu)]
_[o(pv)]}/u l!y
ox o oy o
= (op)/u !J.y.
. aro
(lS)
Since (x 0 , y 0 ) was arbitrary, we have at any point

op o o
ot + ox (pu) + ay(pv) = o. {19)
Equation {19) is the two-dimensional version of the continuity equation (7).

Extension of the argument just given to three dimensions should present no
difficulty.
This type of derivation was used by the early masters of continuum mech-
anics, yet a critical person cannot be satisfied with it. These early masters
were certainly geniuses, and a genius might be defined as a person who gets
the right answer for inadequate reasons, or the wrong reasons. Not having a
genius's almost unerring instinct for truth, and the self-confidence that goes
with it, the ordinary person is filled with doubt. Why can we ignore the
variation of pu across F 1 when it is essential to take into account the variation
of pu as one passes from F 1 to F 3 ? How can we be sure that we are not
neglecting something important when we evaluate op/otat (x 0 , y0 )? What
confidence can we have in our ability to discern negligible terms after learning
that in polar coordinates one cannot neglect Ar in the expression (r0 + Ar) A9
for the length of one side of the appropriate box? (See Exercise 10.) Our final
derivation retains the spirit of the previous method but overcomes our doubts
by a more rigorous procedure.
LARGE BOX METHOD

We still consider two-dimensional motion through a box such as that in
Figure 14.1, but we no longer require Ax and Ay to be "very small." In fact,
to avoid unwarranted simplifications, it is best to think of Ax and Ay as
large.
Consider again the mass flux inward through F 1 At time t this is
Yo+Ay
J l'O
p(x 0 , y, t)u(x 0 , y, t) dy, (20)
where we have used the fact that x = x 0 on F 1 Similarly, the mass flux
inward through F 3 is
fo+Ay
- J.. p(x 0 + Ax, y, t)u(x 0 + Ax, y, t) dy.
Yo
The net inward mass flux contribution from F 1 and F 3 is, using the integral
mean value theorem,
f N+~
l'O
M(y) dy = M(y 1) f
N+~
)10
dy = AyM(Ji). (21)
Here
M(y) = p(x 0 , y, t)u(x 0 , y, t) - p(x 0 + Ax, y, t)u(x 0 + Ax, y, t), (22)
447
and y 1 is a constant that satisfies Yo S y 1 S Yo+ ay. In employing the nota-

tion M(y) above, we have not explicitly indicated the dependence on x 0 and t
in order to emphasize they dependence relevant to application of the integral
mean value theorem.
In like manner, the mass flux inward through faces F4 and F 2 is
o+4X
[ [p(x, Yo, t)v(x, Yo, t) - p(x, Yo + Ay, t)v(x, Yo + .6.y,t)] dx
xo
= .6.x[p(x1, Yo, t)v(xi, Yo, t) - p(xi, Yo + .6.y,t)v(xi, Yo + ay, t)]. (23)
On the other hand, the rate of change of mass contained in the "big box" is
d
-d
t
f 10+.<1y
)IO
[o+4X
XO
p(x, y, t) dx dy
=
ro+.<1y[o+.<1xop
Yo xo ut
The second equality in (24) follows from the integral mean value theorem
op
-;- (x, y, t) dx dy =-;- (x 2 , y 2 , t) ax ay.
ut
(24)
for areas, so x 2 and Y2 are constants that satisfy

~s~s~+~ hSY2Sh+~ ~~
Contemplation of the factor .6.xay in (24) makes one wish that there
were a similar factor in (22) and (23). This can be arranged by applying the
derivative mean value theorem, which states that if f' exists in an interval
containing [y 0 , Yo+ Ay], then
f(yo + Ay) - /(yo) =/'(e) Ay, (26)
for some ~ satisfying
Yo < e< Yo + ay.
To apply this to (23), we identify f(y 0) with p(xi, Yo, t)v(xi, Yo, t). The right
hand side of (23) can then be written
A- A o[p(X1,Y3, f)V(X1,Y3, t)]
-~uy , (27)
oy
where y 0 < y 3 < Yo + Ay. Similarly, from (22) the contribution from F 1
and F3 can be written
A- A o[p(X3,YJ,t)U(X3,Y1,t)]
-L.L.1.L>J' ' (28)
ox
where x0 < x 3 < x 0 + .6.x.Conservation of mass requires that the sum of the
expressions in (27) and (28) equals the right-hand side of (24). Canceling
.6.xAy, we find that for any .6.xand ay, however large or small,
-
o(pu)
--
ox
I
:c=x3 -
)1=)11
o(pv)
--
Ix=:c1
op
oy ,=n = -at ,=n
I
:c=x2 (29)
for some x 1 and Yi satisfying
Yo ::;;Yi::;; Yo + ll.y, i = I, 2, 3.
The desired equation (19) now follows for continuous u, v, and p by taking
equations (29) in the limit as ll.x -+ 0, ll.y -+ 0.
In performing our derivations, we require differing degrees of smoothness
in order that we can be guaranteed that various theorems hold. It is inappro-
priate to ask whether the velocity components, density, etc., " really" have
the required number of derivatives. What we actually would like to know
is whether a model in which, say, thrice-differentiable velocity components
are assumed can accurately describe the phenomena in which we are interested.
Only experience in solving problems can give the answer, so for the present,
we merely need to keep in mind that our derivations require " sufficient"
smoothness. Looking ahead, we can remark that experience indicates that
no smoothness difficulties arise for most solutions, but that in certain very
important problems discontinuous solutions definitely must be allowed
and appropriate modifications made in our derivations.
EXERCISES
1. (a) Derive (8) from first principles.
(b) Derive (10) from first principles.
(c) Show that (8) follows from (10) and vice versa (assuming mass
conservation).
(d) Show that (8) and (10) remain true if F and G are replaced by
vector fields F and G.
2. Carry through the following alternative approach to the manipulations
of the large box method. Show that (21) can be written as
....
fo+Ay
[o+A.x
a[p(x, y, t)u(x, y, t)] dx dy.
70 XO ax
Combine with a similar expression for the first line of (23), and with the
left-hand side of (24), to obtain a double integral that vanishes over the
arbitrary rectangle of Figure 14.1. Deduce (19).
3. Here is another approach to conservation of mass that is analytically
neat but somewhat unsatisfying physically.
(a) Show that in one dimension mass conservation requires that
xo+~
(pu)(xo, Yo, t) - (pu)(xo + ,, Yo, t) =
J
XO
Pr dx.
Differentiate with respect to e.

Obtain the final equation by
e e
setting = 0 or by noting that Xo + and Yo are arbitrary.
(b) Repeat for the two-dimensional case.
Sec. 14.1] Conservation of Mass 449
4. Derive the three-dimensional version of (19) using the following methods.

(a) The small box method.
(b) The large box method.
5. According to a certain cosmological theory, matter is continually being
created in the universe. Assuming that the material in the universe
can be taken as a continuum, write a modified mass conservation equa-
tion.
6. (a) Interpret
JJJo(A, 0) dA1 dA2 dA3 = JJJ

R(O)
p(x, t) dx1 dx2 dx3
R(I)
(30)
as a form of the mass conservation requirement. Deduce the

mass conservation equationin material coordinates.
p[x(A, t), tV(A, t) = c5(A,0) or c5(A,t)J(A, t) = c5(A,0).

(31)
(b) Show that for an incompressible motion J(A, t) = 1.

(c) Show that J(A, 0) = 1.
7. Derive the spatial continuity equation from the material, and vice
versa, by direct change of variables. Start with
0
at[c5(A,t)J(A, t)] = 0.
Cartesian tensors, the subject of Chapter 2 of II, provide an efficient

framework for formulating physical laws that must be invariant in some
sense, regardless of which cartesian coordinate system is used. Rules for
changing from one cartesian system to another are formulated in this
subject. General tensors permit manipulation of general coordinate
changes. The next four exercises give a hint of what is involved.
8. Show by direct calculation that the two-dimensional mass conservation
equation (19) has the same form in variables appropriate to new axes
rotated counterclockwise through an angle 6 with respect to the original
axes. Follow these steps.
(a) Denote the original unit vectors by i and j and the new unit vectors
by i' and j'. Show that
i i' = cos 6, i j' = - sin 6, j i' = sin 9, j j' = cos 6.

(32)
(b) Let
v(x,y) = u(x,y}i + v(x,y)j = u'(x,y)i' + v'(x,y)j'. (33)
450
Thus u(x,y) and u'(x,y) are the velocity components at (x,y)

in the directions of i and i', respectively. Show that
u(x, y) = u'(x, y) cos 8 - v'(x, y) sin 8, (34a)
v(x, y) = u'(x, y) sin 8 + v'(x, y) cos 8, (34b)
u'(x, y) = u(x, y) cos 8 + v(x, y) sin 8, (34c)
v'(x, y) = -u(x, y) sin 8 + v(x, y) cos 8. (34d)
(c) Let (x', y') be the coordinates in the primed system of a point
whose coordinates in the unprimed system are (x, y). Show that
x = x' cos 8 - y' sin 8, y = x' sin 8 + y' cos 8, (35a, b)
x' = x cos 8 + y sin 8, y' = -x sin 8 + y cos 8. (35c, d)
(d) Use a bar to denote the use of primed coordinates so that, for
example,
u(x, y) = u[x'(x, y), y'(x, y)],
u(x',y') = u[x(x',y'),y(x',y')].
Employ the chain rule to establish that (19) implies
(36)
where p(x', y') gives the density in primed coordinates.

9. Consider a two-dimensional flow (w = o/oz= 0) of a compressible
fluid of density p(x, y). By passing from the known cartesian equation,
using changes of variables and the chain rule, show that in polar co-
ordinates
aR 1 a 1 a
-+--(rUR) +--(VR) =0. (37)
ot r or r 08
Here
x = rcos 8, y = r sin 8, R(r, 8) = p(r cos 8, r sin 8),
and U(r, 8) and V(r, 8) are the velocities in the directions of r increasing
and 8 increasing, respectively. If you need a hint, refer to Appendix I 5.2.
10. Derive (37) by direct application of the box argument to the element
of area pictured in Figure 14.2. (After the derivation has been completed,
note that the difference in length between the outer and inner curved
boundaries can be ignored in taking the area of the box but not in
computing the flux across these boundaries.)
451
FIGURE 14.2. Front view of an imaginary box used for the derivation
of the continuity equation in polar coordinates. (The box should be
regarded as extending a unit distance into the paper.)
11. By considering a moving point to have coordinates [x(t), y(t)] or

alternatively [r(t), O(t)], derive (34a) and (34b) from x = r cos 0,
y = r sine.
12. The object of this exercise is to explore how the concepts of continuum
and conservation can be applied to stellar dynamics. These concepts
are applied to the average mass distribution function 'P. (See the dis-
cussion in Section 13.1.) What we wish to do is to derive the funda-
mental conservation equation (1.2.4). No knowledge of Section 1.2
is required to do the exercise.
For simplicity we shall restrict consideration to a function that
depends on spatial coordinate x, the corresponding velocity component
u, and the time t. Thus 'P(x, u, t) is a smooth function with the property
that to sufficient accuracy
2 2
[ [ 'P(x, u, t) dx du
1'1 Xt
gives the mass of material which is located between x 1 and x 2 , and

which has a velocity between u1 and u2
(a) Consider the "big box" in (x, u) space pictured in Figure 14.3.
Explain fully what is meant by "the mass of the fluid occupying
the box." That is, what can one say about the position and velocity
of suchfluid?Thisis an easyquestion.
(b) Write a double integral giving the mass of fluid in the box at time t.
(c) Show that the time rate of change of (b) is 'P,(x*, u*, t) Ax ~u.
where
x 0 :s;x* :s;x 0 + Ax, u0 :s;u* ::;;u0 + ~.
Remember that u is an independent variable.
452
S4
T
.6u S1 S2
1 (xo,uo>i S3
I
I
I
FIGURE 14.3. Imaginary box used to derive an equation for the

distribution/unction 'F(x, u, t).
(d) What is meant by "the net amount of fluid mass passing inward
through the side labeled S1 in the time interval (t, t + At)'"?
(e) Explain why (d) is given by
(+Au
[[:_YA,
'l'(x, u, t) dx] du.
(f) From (e) show that the rate of mass flow inward through side S 1
at time tis
o+4u
[ u'l'(x 0 , u, t) du.
uo
(g) Derive a similar expression valid for side S 2

(h) Combine (f) and (g) to form an expression proportional to Ax Au.
(i) Same as (d), only substitute S 3 for S 1
(j) Explain why (i) is given by
xo+Ax
[f.uo ir(x, u, t) du ] dx,
iXO uo- J,r+
..a(x. s)ds
where a(x, t) is the acceleration in the direction of x increasing.

As only gravitational forces need be considered, the acceleration
(force per unit mass) is independent of u.
(k) Use the methods of (f)-(h) to find an expression for the mass flux
inward through S 3 and S4
(l) Since collisions between stars can be ignored, the final answer is
obtained by equating (c) with (h) + (k) and taking the appropriate
limit.
Sec. 14.1) Conservationof Mass
453
13. The object of this problem is to derive the same final differential equation
(l.2.4) that was obtained in Exercise 12, but in a different manner. Here
we shall imitate the first method that was used above for deriving the
ordinary mass conservation equation.
We shall use the notation
x = x(X, U, t), u = u(X, U, t), (38a, b)
to describe motion in xu space. We shall assume that (38a) and (38b)

can be inverted to yield
X = X(x, u, t), U = U(x, u, t). (39a, b)
(a) What is the physical meaning of (38a) and (39b)'?

(b) We make the definition
A(X, U, t) = ou(X, U, t).

at
What is the physical meaning of A'? How should we define a(x, u, t),
the " spatial " version of A '?
(c) Let
'li(X, U, t) ='P[x(X, U, t), u(X, U, t), t]. (40)
Define the generalized material derivative ~/~t by

~'i'(x, u, t)
!iJt
= o'li(X, u, t)
iJt
I= .
X X(.x, 11,r)
(41)
U=U(.x,11,t)
Show (using careful and explicit notation) that

~' aq, o'i' o'I'
-=-+ u-+A-. (42)
~t at ax OU
(d) Define
ox ax
ax au
,l(X, U, t) = (43)
au au
ax au
Using careful and explicit notation, derive
a,1 _ ,I . aa1 (44)

ot - OU.x=z(X, U, t).
u=u(X, U,r)
(e) In the absence of "collisions," a final differential equation can

now be obtained by using the procedure that we followed in deriv-
ing the usual equation for mass conservation. That is, we define
a suitably generalized "material region" R(t) and assume that
ff
_dd '(x, u, t)dxdu
t R(t)
= 0. (45)
The final differential equation is obtained by manipulating (45).

Perform this manipulation.
14.2 Balanceof LinearMomentum

We shall now apply a version of Newton's second law to the motion of a
continuum.
AN INTEGRAL FORM OF NEWTON'S SECOND LAW

In particle mechanics a particle P that has mass m and moves with velocity v
is said to possess (linear) momentum mv. According to Newton's second law,
the rate of change of the linear momentum of P is equal to the net force
acting on P. To generalize this law to continuum mechanics, we define the
linear momentum at time t possessed by material of density p(x, t) that occu-
pies a region R(t) as
Jff p(x, t)v(x, t) d1:,

R(t)
(I)
where vis the velocity vector. Again, the time rate of change of the integral
in (1) is hypothesized to be equal to the net forces acting on the material in R.
What kind of forces are expected? The force of gravity is the most familiar
example of body forces, which in particle mechanics are proportional to the
mass of a point particle. In continuum mechanics it is natural to pass from
points to regions by means of an integral. We thus consider body forces on
the material in R that are of the form
ffJp(x, t)f(x, t) d1:,

R(t)
(2)
where f(x, t) is the body force vector per unit mass. We also consider surface
forces. These have the form
fJ t(x, t, n) da,
DR(t)
(3)
where x is the position vector of the surface element, t is the time, and n
is the unit exterior normal to 8R at x. The dependent variable t is called the
stress vector.
Sec. 14.2] Balance of Linear Momentum
455
In writing (3), we have implicitly assumed that force per unit area (stress)
approaches a limit as area approaches zero. To see this, let Sk (k = I, 2, 3, ... )
denote a sequence of smooth similar surfaces of decreasing area, each a por-
tion of oR and each containing the point x<0 >. Let n<0 > be the unit exterior
normal to oR at x<0 >.Then if Ak is the area of Sk, the ith component of the
force per unit area on Sk is
.!_
A k
fft '(X'' t n) da = t '-(x<k>
sk
t
',,
n<k>)

'
i = 1, 2, 3,
where we have used the integral mean value theorem. Here n<k> is the unit
exterior normal at x<k>,
a point of Sk. If t; is a continuous function of its
variables,
(4)
since x must approach x<0 > as Ak-+ 0. [Note that n is determined at every
point x on a given surface so that in this situation t(x, t, n) is actually a function
only of x and t.]
Why ought one to assume the existence of a stress? Historically, this
assumption emerged as a generalization of what were originally special
assumptions concerning special materials. In stretching a metal bar, for
example, it has been assumed since Hooke that the relative elongation is
proportional to the applied force per unit area. The randomly colliding
elastic spheres model of a perfect gas leads to the introduction of a normal
force per unit area (pressure), as the reader doubtless recalls from his intro-
ductory physics course. Some historical remarks can be found in Appendix
14.2, but an adequate appreciation of the long and interesting history of
continuum mechanics requires further reading.
The definitions of f and t are given fully by the assumptions that forces
of the form (2) and (3) act on R. But it may be useful to provide more verbal
descriptions of these important quantities.
Body force. At time t, a force per unit mass f{x, t) is assumed to act
at each point x of R.
Surface str~. At time t, consider a point x on the boundary oR of

region R. Let n be the unit exterior normal* to Rat x. Then t{x, t, n) is the
force per unit area exerted at the point x of oR by the material exterior to R
on the material interior to R.
Whenever we employ the unit exterior normal n, it virtually goes without saying that we
assume that n exists. That is, we assume that the boundary BR of R dhides space into two regions,
a bounded connected set of points called the interior of R, and the complement of this region
called the exterior. We generally assume that n varies smoothly as x moves about the surface.
Combining our speculations thus far, we put forth the hypothesis that the
rate of change of linear momentum in a material region R(t) equals the con-
tribution of the body forces plus the contribution of the surface forces:
d Jff
d t R(t)
pv d1:= Jffpf d1:+ fJt du.
R(t) 8R(t)
(5)
This fundamental equation, due to Cauchy, can be viewed in the following

way. The effect of the rest of the universe on the linear momentum of material
occupying R(t) can be replaced, it is hypothesized, by the effect of a long
range body force per unit mass f, acting throughout R(t), plus the effect of
stresses t, acting on the boundary of R(t) and representing the additional
effect (contact force) of material adjacent to R(t). In other words, as far as
linear momentum is concerned, our assumption is that we can replace the
effect of the universe outside R(t) by appropriate body forces f and stresses t.
Alternatively, (5) simply states that linear momentum increases because
of a momentum source per unit volume pf and a momentum influx per unit
area t. Such a lofty point of view has a certain appeal, but a descent is prob-
ably necessary when one wishes to identify momentum sources and fluxes
with the pushes and pulls of everyday experience.
To attain an appropriate degree of respect for (5), the reader might (a) mull
it over until he is convinced that it is obviously correct, and then (b) consult
Section 7.1 in II which shows that (5) is incorrect under certain circumstances
(when surface tension must be taken into account).
LOCAL STRESS EQUILIBRIUM
By the vector version of (1.8) we can take the time derivative in (5) inside
the integral:
d fJJ
dt R(t)
pv d1:=
R(t)
fJfp(Dv)
Dt
d1:. (6)
With this, (5) can be rewritten as
fJJ d1:- fffpf d1:= fft du.

p(Dv) (7)
Dt
R(r) R(r) 8R(t)
Important consequences can be derived from (7) even without a clear idea
of the forces involved. To do this, consider a family of similar regions RL
characterized by the length L. (An example is provided by. the family of
rectangular parallelepipeds with sides of lengths aL, bL, and cL.) As men-
tioned in Exercise 3 of Appendix 13.3, such regions have the same shape and
their volume ~ L and area 9' L satisfy
457
for some constants lR and R, which depend only on the shape of the regions.
Hence for any continuous function G, the integral mean value theorem
implies that
(8)
Let us apply (8) to the right side of (7), component by component. Since all
volume integrals are proportional to L 3 , by dividing by L 2 and then letting
L -+ 0, we find that
lim
L ... o L
ff t du = 0.
~ oRL(r> (9)
Equation (9) is an important one. Insight into its meaning can be acquired
by recognizing that if the material in the region RL were subject to surface
stresses and to no other forces, and if this material were in equilibrium, then
necessarily the total force on the material would have to vanish. That is,
ff tda=O.
ilRL
Actually, each region RL is moving and is subject to body forces. But as

L -+ 0, the effect of the surface stress term becomes more and more dominant
compared to the volume integrals
and fff
RL
pf dt,
which represent the effects of motion (inertia) and body forces. In the limit
one obtains (9), which states, roughly speaking, that the surface force on an
infinitesimal region, divided by its area RL2, vanishes-just as if the region
were in equilibrium. Equation (9) is thus known as the principle of local
stress equilibrium.
ACTION AND REACTION
We shall derive consequences of the principle of local stress equilibrium

by applying it to various regions. We first consider a rectangular paralleli
piped Fi, containing a fixed point x on one corner and having dimensions
eL by L by L, e ~ I (Figure 14.4). Let n, -n, nT, -nT, nRi , and -nR; be the
unit exterior normals to the front, rear, top, bottom, right, and left faces of
this "flake," respectively. (See Figure 14.4.) By the integral mean value
theorem, the ith component of the surface integral in (9) can be written as
i = I, 2, 3, (10)
where
Qi= t;(XT, t, nT) + ti(Xs, t, -nT) + t;(XR;, t, DR;)+ l;(XL, t, -DR;).
-n
FIGURE 14.4. The "flake" FL. Application of the principle of local

stress equilibrium to this family of regions gives the "action-reaction"
law (12).
Here xF, xR, Xr, x 8 , XR;, and xL are points in the front, rear, top, bottom,
right, and left faces of the flake, respectively. Equation (10) comes from
multiplying ti at some point on each face by the area of that face. The area
is L 2 for the front and rear faces, sL2 for the remaining faces.
Combining (9) and (10), we obtain
tlx, t, n) + t;(x, t, -n) + sQ; = O; i = l, 2, 3.
The first argument in each t; is x, since xF, x 8 , etc., all approach x as L -+ 0,
and since we assume that t; is continuous in a domain containing x. Using
the continuity and hence the boundedness of Q;, we observe that sQ; can
be made arbitrarily small so that
t;(x, t, n) + tlx, t, -n)t= O; i = I, 2, 3; (11)
or, suppressing the dependence on x and t,
t(-n) = -t(n). (12)
We have thus deduced the counterpart of Newton's third (action-reaction)
law for continua.
Consider Figure 14.5. At time t the stress at x representing the action of
adjacent material exterior to R 1 on R1 is t(x, t, n). We have shown that this
stress is equal in magnitude and opposite in direction to t(x, t, -n), which
represents the effect at x on R 2 of adjacent material exterior to R 2
THE STRESS TENSOR
We now show that t(n) can be expressed as a linear combination of t(e<1>),
t(e<2 >), and t(e<3 >). To do this, we apply (9) to the tetrahedron of Figure
14.6. The mutually perpendicular faces of area S; are normal to the ortho-
Sec. 14.2] Balance of Linear Momentum 459
FIGURE 14.5. At x the stress on Ri, due to the material of R2, is equal
and opposite to the stress on R2 , due to the material of R1.
FIGURE 14.6. A tetrahedron whose slanting face has area S =L 2

Application of the principle of local stress equilibrium as L - 0 gives the
relation (17) between the stress vector and the stress tensor.
normal basis vectors e<i).These faces meet at x. The "slanting" face has
=
area S. We defineL by L S 1 ' 2 .
The jth component of the force acting on the face of area S; is, by the
integral mean value theorem,
Sit-(x<il t - e) i,j=I,2,3,

J ' ' '
where x<i) is a point in this face. Let us denote by n the unit exterior normal
to the slanting face of area S. Then, for some point x<0 > in the slanting face,
Stj(x<0 >,t, n) is the jth component of the force acting on this face. Since
S; is the projection of Son the ith coordinate plane,
S-I = [n e<i)JS= n.S

I '
(13)
where n; is the ith component ofn. Applying (9), withL 2

= S, we obtain
3
" n; tj(x, t, -e
L.. (i) ) + t/x, t, n) -- 0. (14)
i=l
We have used the fact that all points x<'ll selected by the mean value theorem
must approach x; q = 0, l, 2, 3. Using (12), we can rewrite (14) as
3
tj(x, t, n) = I. n; tj(x, t, eCi>), j = l, 2, 3. (15)
i= l
We define T;i by
i,j = l, 2, 3. (16)
(See Figure 14.7.) Now we rewrite (15) finally as
3
tj(x, t, n) = I, n; T;j(x, t), j= l, 2, 3. (17)
i= 1
Using what can be regarded as a generalization of dot-product notation, we

sometimes write (17) as
t(x, t, n) = n T(x, t).
(This direct notation will be used extensively for tensors in II.)
x,
l j Timer I
1
/
FIGURE 14.7. Some components of the stress tensor. The shaded area
represents a surface element at point x with unit normal e<3 >. The effect
at time t of material above the element on material below it is given by
the stress vector t(x, t, e<3 >). The three components of this vector are
T3i, T32, and T33.
The very important relation (17) shows that the dependence of the stress
vector t on the exterior unit normal vector o is necessarily of a special linear
form. This dependence is a consequence of the principle of local stress equi-
librium. The nine quantities T;i; i,j = l, 2, 3 are called components of the
stress tensor T. The significance of the term tensor will be discussed in detail
in Chapter 2 of II. It will, however, be worth remembering that the first
component of Tii refers to the unit vector that is normal to the face on which
the stress, denoted by the second component of T;i, is acting. [Refer to defini-
tion (16).] A good mnemonic is
irst econd
F ace S tress .
For example, T13 is the third component of the stress vector acting on the
face (area element) whose exterior normal points in the direction of
increasing x 1
NEWTON'S SECOND LAW IN DIFFERENTIAL EQUATION FORM
As a first consequence of (17) we observe that the jth component of the
surface term in the momentum conservation equation (7) can be written
ff
oR(r)
ti du= ff f n; Tii du= ffJf ,,,
oR(r) i= 1
0
(T;) d1:,
R(r) i= 1 uX;
(18)
where in the last equation we have used the divergence theorem. With this
(7) can be written in component form as
fff [pDvi - Pfi - I !___

(T;)] d1:= 0, j = l, 2, 3.
R(r) Dt i= 1 OX;
Since R(t) is arbitrary, if we assume that the integrand is continuous we can
write
Dv- 3 o
p DtJ = pfj + L -;----
i=l ux,
(T;), j=l,2,3. (19)
In direct notation, (19) is written

Dv
p Dt = pf + V . T. (20)
Relations (19) or (20) are Cauchy's differential equations expressing for

any continuum the balance*of linear momentum. The only hypotheses used in
obtaining these equations were sufficient smoothness of the functions in-
volved and mass conservation [employed to obtain (6)].
EXERCISES
1. Let S denote the common material boundary of the regions R 1 and R2
depicted in Figure 14.5. Write the integral equations of linear momen-
tum balance for R 1 , for R 2 , and for R 1 + R 2 Infer that
ffs [t(x, t, -n) + t(x, t, nJ] du= 0,
* If momentum were conserved, a particle's momentum would change only as a result of

momentum flux across its boundary. The body force term in (19) provides an additional source of
momentum. Hence it has become customary to speak of" balance" rather than "conservation."
and hence obtain an alternative derivation of (11). [Compare the deri-

vation of Equation (12.1.45).]
t2. In a certain two-dimensional problem, the stress tensor has four non-
zero components: T11 , T12, T21 , and T22 . In terms of these four quanti-
ties, find the force exerted by the material in the direction of the vector
i - j on the surface element of area dA whose exterior normal has the
direction of the vector 2i + j.
3. (a) Show that if no body forces are acting, then
'.fi1 = T22 = T33 = T12= T21 = 0,
and
T32 = T23 = !(xi + 2x1 - x~)
can be the components of the stress tensor for a material at rest.
(b) Consider an infinite equilateral triangular cylinder whose cross
section is depicted in Figure 14.8. Show that the net force on each
of the bounding planes is zero, if the stress tensor is that of part (a).
---~x1
FIGURE 14.8. Cross section of a triangular cylinder.
t(c) Find the net force (per unit length in the x 3 direction) exerted by
the lower shaded region on the upper unshaded region.
4. (a) Using conservation of mass, show that the ith component of
pDv/Dt is
t
j=l
o(v;pv)
oxj
+ o(pv;).
ot
Deduce from (19) the linear momentum transfer equation
d
dt
ffJ dr: = fJf dr: + fJ[t -
R
pv
R
pf
iJR
pv(v n)] du. (21)
Here Risa region.fixed in space.

Sec. 14.2) Balance of Linear Momentum
(b) The ith component of pv(v n) is I,j= 1 pv;vi ni. The nine quantities
pv; vi (i,j = I, 2, 3) are called the components of the momentum
flux tensor. Why is this an appropriate name? (Concentrate on
the words "momentum flux." The significance of "tensor"
will be discussed later. Also see Exercise 4.5.)
5. Suppose that fluid is in steady motion past a bounded obstacle. Let S
be an imaginary surface enclosing the obstacle (Figure 14.9). Suppose
that f = 0. Let K be the force acting on the obstacle. Use the linear
momentum transfer equation to deduce K = s [t - pv(v n)] d<1.(The
force on an obstacle immersed in a steady fl.ow can thus be deduced
from a knowledge of what is happening far from the obstacle. See
Exercise 6.)
- -
-
- -
FIGURE
- -
14.9. Knowledge of the stress and velocity on the control
surface Sallows computation of the force on the obstacle.
!6. Consider a uniform steady fl.owpast an obstacle. Suppose that the fluid
is inviscid so that (Section 15.1)t = - po, where p is the pressure. Suppose
further that
V = V + O(jxj-
00
3
), (22)
where the constants v00 andp 00 represent the constant velocity and pres-
sure at jxj = oo. [As stated in Appendix 3.1, g(:x:)= O(lxl-n) (with
"!xi-+ oo" understood) means that lim 1% 1...00 lxl"lg(x)I= constant.]
Use the momentum transfer equation to prove that the force on the
obstacle is zero. This result is essential in the proof of the D' Alembert
paradox. (See Appendix 15.1.)
7. (a) For various purposes it is worthwhile writing the Cauchy equations
(19) in material variables. We obtain
~ a:~i
vt
= ~Fi
+[ t ::.o
i= 1 vX;
(!!";)]
x=s(A. r)
. (23)
The body force and stress tensor terms on the right side of (23)
are defined by
Fj(A, t) = Jj[x(A, t), t], Y-ii(A, t) = Tii[x(A, t), t].
What equations justify the transformation of the left side of (19)

to the left side of (23)?
(b) A weakness in (23) is that it contains a differentiation with respect
to the dependent variables xi, while the independent variables are
Ai. Make the necessary adjustment using the chain rule. To do
this, use the chain rule to deduce expressions for the quantities
oAj{oxi from the result oA;foAi = Jii
(c) For inviscid fluids
0
becomes --p
oxi '
where p is the pressure, and (23) becomes
o2 xi op
p :it2
u
= pF. - -;;-
uXi
Deduce
k = 1, 2, 3. (24)
8. Suppose that a primed coordinate system is rotated by an angle 8

compared to an unprimed system so that
x 1 = x~ cos 8 - x 2sin 8
and
X1= Xi cos()+ X2 sin 8,
X2 = X1sin 8 + X2cos 8 X2= -Xi sin (J + X2 cos 8.
(a) To obtain a relation between components of the two-dimensional
stress tensor in the primed and unprimed coordinates, verify
the following reasoning.
T22 = stress component in 2' direction on face
with normal in 2' direction.
The ith component of stress of face with exterior normal in 2'
direction is given (in unprimed system) by
ti= -sin 8 Tli + cos 8 T2 i
Therefore
T 22 = T11 sin2 8 - (T12 + T21 ) sin 8 cos 9 + T22 cos2 9.
(b) Similarly, show that
T{2 = -T 11 sin 9 cos 8 + T12 cos2 9 - T21 sin2 9 + T22 sin 9 cos 8,
and find expressions for T21 and T{1
Sec. 14.3] Balance of Angular Momentum
14.3 Balance of Angular Momentum

In this section we shall establish an equation for the balance of angular
momentum. To begin, let us review some material studied in elementary
particle mechanics.
TORQUE AND ANGULAR MOMENTUM

The torque M (or moment of force) about the origin O due to a force f
acting at point x is defined by the equation
M = XAf, (1)
where /\ denotes vector product. If
f =f1i +/2j +/3k, X = X1i+ X2j + X3k,

M = M 1i + M 2 j + M 3 k, (2)
then, in particular,
M3 = xif2 - X2f1 (3)
The third component of M (the moment about the x 3 axis) can thus be
computed by multiplying the force components / 1 and / 2 by their respective
"lever arms" x 2 and x 1, and adopting the convention that counterclockwise
and clockwise contributions are positive and negative, respectively. (See
Figure 14.10.*) Analogous results hold for M 1 and M 2
FIGURE 14.10. The lever arms of the force components J;. and f2 are
x 2 and x 1 Counterclockwise contributions to torque about the origin
are regarded as positive.
The sign can also be remembered by the right-hand rule. Point your thumb along the
x 3 axis (out of the paper). Then the fingers of your right hand will point in the direction of a
positive contribution to the moment about this axis.
466 Field Equationsof ContinuumMechanics [Ch. 14
If a mass point at x possesses momentum P, its moment of momentum or

angular momentum about O is defined to be x AP. Suppose a system of mass
points is subject to mutual central forces (like gravity) which are directed
along the lines joining the interacting particles. It is then a theorem of point
mechanics that the rate of change of each particle's moment of momentum
is equal to the sum of the torques acting on the particle.
In continuum mechanics we deal not only with central acting body forces
like gravity, but also with contact forces that differ greatly from material to
material. Now we must make the assumption that in any "nonmoving"
or inertial coordinate system, the rate of change of moment of momentum is
equal to the torques exerted by body forces and surface stresses. To formulate
this assumption more precisely, we must generalize the concepts of moment of
momentum and torque used in point mechanics, but this is straightforward.
We thus postulate
d
dt
III(xApv)d-r=IfJ(xApf)d-.+
R(t) R(r)
ff (xAt)da.
i1R(I)
(4)
POLAR FLUIDS
Assumption (4) is not obviously true. In fact, it is sometimes not true at all.
For polarfluids, which typically contain long molecules, it is necessary to introduce
an internal angular momentum per unit mass I. Thus the total angular momentum
contains a contribution from internal angular momentum in addition to the usual
contribution from the moment of linear momentum. We assumed in (4) that the
moment of momentum was changed only by the same body forces and surface
stresses which change linear momentum. Now we allow the presence of a body
torque per unit mass g and a couple stress per unit area c. Instead of (4), then, for
polar fluids one assumes that
i
dt~
III(x AV+ l)p III(x Af +g)p +II(x At +c)
dT =
~
dT
u~
da. (5)
Manipulation of (5) is not difficult once cartesian tensors can be used (see Exercise
2.3.12 of II). It was introduced here to counteract complacent acceptance of (4).
We shall therefore not consider (5) further in this section. References that show
the possible usefulness of (5), to analyze experiments involving liquid crystals, are
articles by H. Gasparoux and J. Proust [J. Phys. 32, 953 (1971)] and by H. Tseng,
D. Silver, and B. Finlayson [Phys. Fluids 15, 1213 (1972)].
SYMMETRY OF THE STRESS TENSOR

In our studies of mass conservation and linear momentum balance, we
started with an integral statement of the appropriate law and derived a differ-
ential equation as a consequence. From the angular momentum balance
relation (4), an even more striking consequence can be derived, namely,
(6a)
Sec. 14.3] Balance of Angular Momentum
In words, "the stress tensor is symmetric when the moment of momentum

equals the sum of torques due to body forces and surface stresses." In direct
notation the symmetry is indicated by the equation
(6b)
where Tr indicates transpose.
Once the machinery of cartesian tensors has been mastered, proof of (6)
can be relegated to an exercise (Exercise 2.3.11 of II). Some w1ll therefore
prefer to take the symmetry of T for granted temporarily.
An alternative proof of (6) proceeds in a way that is analogous to the
development in the previous section. A principle of local moment equilibrium
is established. Application of the principle to a cube yields the desired result.
The argument is not particularly short, but it confers the advantage of gen-
erating familiarity with the notions of " stress tensor" and "moment." This
argument is given in the remainder of this section.
THE PRINCIPLE OF LOCAL MOMENT EQUILIBRIUM
In analogy with our treatment of linear momentum, we first apply the
fundamental integral assumption (4) to a family of similar regions RL
characterized by the length L. For the body force term in (4) one can obtain
the estimate
Ifff (x" pf) I~ Ix "

RL(t)
dr pf Imax volume of RL(t), (7)
where IP(x) Imax

denotes the largest value of IPI for x in RL. [See Exercise
3(a).]
As we have mentioned, the volume of RL is ).RL3, where ).R depends only
on the shape of the regions RL, not on their size. Let 1/1denote the angle
between x and f. Since Isin 1/1I ~ I,
~ jxlmax
jx /\ pflmax= [IXI Ipfj Isin IPI1max !Pflmax
Thus
Ifff
RL(I)
(x /\ pf) dr I~ ).RL
3
!Pflmax
lxlmax

By Exercise 3(b), we can transform the inertia term in (4) as follows:
d Jff
dt R(t)
(x " pv) d'C =fff (x" p DDv)
R(t) t
d'C. (8)
Hence the inertia term in (4) can be treated just like the body force term,
and it is also bounded by an expression proportional to L 3 j x Imax.
We thus
obtain the principle of local moment equilibrium
Iim L 2 I 1I
L~o x
fJ(x"
max BRL(I)
t) do-= 0. (9)

LOCAL MOMENT EQUILIBRIUM FOR A CUBE
Let us now apply (9) to a cube C of side L. We choose our cartesian co-
ordinate system so that the origin is at one comer of the cube and the axes
point along its edges (Figure 14.11). The faces of the cube have exterior nor-
mals i, j, k so that the components of the stress vectors acting on these
/
/
/
JI
FIGURE 14.11. A cube for application of the principle of local moment

equilibrium. The three components of the force on the front face are
equal in magnitude to the indicated components of the stress tensor.
forces are appropriately signed components of the stress tensor T1i. For C,
IxImax = J3 L, so the principle of local moment equilibrium (9) becomes
lim L\
L--o
ff(x"
ac
t) dO'= O. (10)
We shall discuss the moment about the x 3 axis in detail, but we shall leave
the analogous treatment of the moments about the other axes to the reader
(Exercise 5). Since we consider the situation at time t throughout, we shall not
explicitly indicate the time dependence of the various quantities until we
obtain our final result.
On the face of C in the plane x = L, the stress vector t(L, y, z) has the
components 7i1(L, y, z), i = 1, 2, 3. Because of (3) the contributions to the
moment about k of these three components are
-f f T
0
L L
0
11 (L, y, z)y dy dz, Lt
0
t T 12 (L, y, z) dy dz,
0
0, (11)
respectively. Of course, T13 makes no contribution to the moment about the

Sec. 14.3] Balance of Angular Momentum 469

k axis. The contributions due to stresses acting on the face in the plane
x=Oare
ff
L
0
L
0
Tu (0, y, z)y dy dz, 0, 0. (12)
The T12 contribution to (12) is zero because its lever arm has zero length.
The sign of the T11 contribution requires careful consideration. Since t(i) =
-t( -i), t(O, y, z) has the components -Tu(O, y, z), i = 1, 2, 3. The first
component, - Tu, points in the direction of i, so the product of - T11 and
its lever arm y must be preceded by a minus sign.
Adding (11) and (12) and then applying the integral mean value theorem,
we find a net contribution of
where
0 ~Y;, Z; ~ L, i = l, 2. (14)
Using the mean value theorem for derivatives, we find that

0T11
T11 (L, Yi, zi) - Tu(O, Yi, Zi) = ox (xi, Yi, Zi)L;
Hence the absolute value of the first term in (13) is less than
I
OT11(xi, Yi, Zi)
-;--
vX
I
max
L4 .
When the limit of (10) is applied, this O(L 4 ) term will vanish. (We assume
that oTufox and similar quantities are bounded.)
Let us now consider the contributions to the moment about k caused by
the stresses acting on the faces in y = 0 and y = L. Arguments of the kind
we have just presented show that the T2 i contributions are
(15)
In like manner the contributions from the faces in z = 0 and z = L are all
O(L 4 ). Applying (IO) to the cube C, we find that
(16)
Similarly (Exercise 5),
T3 i(O, 0, 0) = TiiO, 0, 0) and T23 (0, 0, 0) = T32 (0, 0, 0). (17)
Since the origin was taken at the arbitrary point that we initially selected
for consideration, (16) and (17) hold for any point. Thus (6) has been
demonstrated.
EXERCISES
1. Let f = fii +/ 2 j be a force that acts at x = x 1i + x 2 j. Show that the
formula IM1=
3 =
Jfld, where Jfl (jf +/;) 112 , is an alternative to (3).
Here d is the length of the perpendicular from the origin to the "line
of action" off, as pictured in Figure 14.12.
-----ri
d\
I
I
L-------
FIGURE 14.12. For calculation of the moment about the origin, the
" lever arm " off has length d.
2. Show that (4) remains true if the coordinates are translated so that x
becomes x + k, where k is a constant vector.
3. t(a) Prove (7).
(b) Derive (8).
4. (a) Verify the results stated in the text concerning the contributions
to the moment about k caused by the faces of the cube C in the
planes y = 0, y = L, z = 0, z = L.
(b) Using the data of Exercise 2.3 (Figure 14.8), find the torque about
the origin which is exerted on the material below x 3 = 0 by material
above x 3 = 0.
5. Show that
(a) T3 1 = T13.
(b) T23 = T32
14.4 Energy and Entropy

Completion of the set offield equations that are satisfied for any continuous
medium requires introduction of thermodynamic concepts. To do this
with any degree of thoroughness requires considerable effort, and the reader
may be wearying of general considerations. For those who feel refreshed by
frequent contact with particular problems, we begin this section with a dis-
cussion of the special case of ideal gases.* Indeed, except for certain important
This material could be read after Section 14.2.

Sec. 14.4] Energy and Entropy 471
general results that are the subject of exercises, the material on perfect fluids
is the only thermodynamical material that is used subsequently (in discussions
of sound waves, Sections 15.3 and 16.3).
Energy and entropy considerations must be included for completeness,
and such considerations play a major role in many areas of applied mathe-
matics. Those who wish to have a strong foundation in continuum mechanics
must at some time return to the main part of this section, beginning with the
portion labeled " Equilibrium Thermodynamics." There they will find dis-
cussions of the general differential equation of energy balance, Gibb's
relations, the Clausius-Duhem law of entropy increase, etc.
IDEAL GASES
It is an experimental fact that at low pressures the product of the pressure
P and vol~me Vn of n moles* of a gas is equal to the product of n, the
temperature 0, and a constant fJl whose value is the same for all gases.
Moreover, at low pressures the internal energy Eis almost completely deter-
mined by the temperature. With support from a simple kinetic theory model
to which the reader probably has been exposed, one is thereby led to define
an ideal gas, a "model" material with the equations of state
PV,, = nfJl0, E=E(0) (fJl = 8.3143J/mole deg). (la, b)

Common gases can be regarded as ideal at pressures below about 2 atm if
errors of a few per cent are tolerable.
In particular problems (la) takes a more convenient form if it is divided
by Mn, the mass of n moles of gas. We may therefore write
(2a)
where R = nfJlfMn is a constant for any particular gas. We also note that it
can be shown that (I b) can be replaced by the relation
E = Cv 0 + constant, (2b)
where Cv is the specific heat at constant volume [Exercise l(a)].

As is usual in elementary treatments of thermodynamics, our discussion
thus far has concerned a container filled with gas under uniform conditions.
We are aiming at a theory for a gas wherein conditions vary, but we shall
proceed in the simplest way possible by assuming that (2a) and (2b) still
hold locally. Thus we write
p = pRO, e= Cv O + constant, (3a, b)
* Recall that a mole of a gas has a mass in grams that is numerically equal to the molecular
weight of the gas. It contains a number of molecules equal to Avogadro's number: 6.0225 x 1023
where p = p(x, t) is the local mass per unit volume, (} = O(x, t) is the local
temperature, p = p(x, t) is the local pressure, e = e(x, t) is the local internal
energy per unit mass, and Cv is the local specific heat.*
Elementary physics texts give arguments which show that the assumption
t(x, t, n) = - p(x, t)n (4)
for the stress vector tis a reasonable one for an ideal gas. We shall accept
this assumption here. It is discussed at some length in Section 15.1.
With (4), the equation of momentumbalance (2.20)plus the equation of
mass conservation (1.6) provide four equations for five unknowns: three
components of velocity, the density, and the pressure. The introduction of
the equation of state (3a) introduces one more relation, but also one more
variable, the temperature. According to (3b), the temperature is related to
the internal energy of the system. Hence we need to introduce a new physical
law: the principle of energy balance, or the first law of thermodynamics.t
Consider a unit mass of gas of volume V. The first law states that
dQ = dE + P dV, (5)
where dQ is the heat introduced to the unit mass and Eis its internal energy.
Before we attempt to transform (5) into a partial differential equation,
let us note an important relationship. If we consider the ratio dQ/0, we have
(using dE = Cv de and P = R0/V)
dQ d0 dV
-=Cv-+R-. (6)
e e v
Assume that Cv is constant. Then the right-hand side of (6) equals
d(Cv loge+ R log V). Thus dQ/0 is a total differential in terms of the state
variables e and V. (Q itself is not a state variable.) Equation (6) can be
regarded as a very special case of the second law of thermodynamics, which
states that the quantity dQ/0 is a total differential of the state variables for
all reversible (infinitesimally slow) exchanges of heat. Thus, if we follow a
reversible process and restore a medium to its original state, we have
,CdQ = 0
re , (7)
where the integral is taken, for example, along a closed curve in the (V, 0)-
state plane. For reversible processes we may consequently define a new func-
tion S of any state I, the entropy, by
'E.dQ
S - So = { , (8)
0
* Often the specific heats can be considered constants, so there is no difference between
Cv and c.
t In many cases the density of a fluid particle, or even of the whole fluid, can be regarded as
constant. (See Section 15.2 for an example.) Then there is no need to introduce thermodynamic
concepts.
where the integral is taken along any path in a state plane, starting from a
reference state O and ending at 1:. In particular, for an ideal gas, (8) yields
S = Cy log 0 + R log V + constant (9)
if the specificheat at constant volume is independent of temperature. From (9)
it follows that the equation of state P = R0/ V may be rewritten in the form
p = Ke5fCvv-.,. (10)
Here
Cp R+ Cy
y=-=-- (11)
Cy Cy
is the ratio of specific heats (at constant pressure and constant volume,
respectively), and K is a constant characteristic of the gas. (See Exercise I.)
We now tum to the reformulation of (5) as a partial differential equation
for the flow field. We regard (5) as referring to a given fluid particle, and we
invoke the substantial derivative. Since V, the volume of a unit mass, is the
reciprocal of the density, we are led to write
De + p E_{ = Dq.!..) (12)

Dt Dt \p Dt
Here we have used local variables. In particular, e is the internal energy
per unit mass. Also Dq/ Dt is the rate at which heat is introduced to the gas
per unit mass. From our previous experience with heat conduction, we expect
that (in the absence of radiation effects)
Dq = ! V (kVO). (13)
Dt p
If we introduce the entropy per unit mass s, and if we assume that local
processes are reversible, we may also reformulate (8) as
(14)
Moreover, (IO) becomes

(15)
Since e is a function of the state variables, a combination of (12) and (13),
!
De+ p D(l/p) = V. (kV(}) (16)
Dt Dt p '
together with the equations of continuity and motion, form a system of
five differential equations for the three components of velocity and two of
the state variables. The three equations {3a), (3b), and ()5) relate the five
state variables p, p, 9, e, ands.
Hopefully, the above discussion will make it easier for the reader to under-
stand the situation for a general medium. The energy equation and, conse-
quently, other thermodynamic relations must be introduced to complete the
formulation of the basic equations. To show that we can put the full set of
equations to practical use, we point out (as mentioned above) that they
can be used to study acoustic waves, to give just one example. If we consider
small disturbances in a homogeneous gas originally at density p0 so that
and
(I~ l,
then one can show that u satisfies, in the linear approximation, the partial
differential equation of wave motion
o--c2-
2
u o2 <J
ot2 - O ox 2
Here c0 is the wave speed, a constant in the medium. The detailed discussions,
which the reader may wish to study at this point, are given in Section 15.3.
Having completed our treatment of an important special case, we now turn
to a general discussion of thermodynamical considerations in mechanics.
Some of the points made above will be repeated, but in a fuller and more
general context.
EQUILIBRIUM THERMODYNAMICS
Elementary treatments of thermodynamics normally consider only materials
whose properties are uniform in space and time. The term thermostatics has
been introduced to emphasize that only infinites.imally slow changes can be
directly handled by the elementary theory. We shall retain the older termin-
ology equilibrium thermodynamics.
We wish to study materials whose properties change from moment to
moment and from point to point, so we cannot restrict ourselves to equi-
librium concepts. On the other hand, these simpler concepts must provide the
foundation of our discussion. Although we have generally assumed a know-
ledge of basic physical principles, it is our experience that thermodynamic
concepts such as entropy are far less familiar than corresponding mechanical
concepts such as force, work and kinetic energy. Consequently, equilibrium
thermodynamics is briefly reviewed in Appendix 14.l. *
Some salient ideas of equilibrium thermodynamics are these: (i) Energy
is conserved in the sense that an increase in a system's internal energy E
is equal to the heat added to the system minus the work done by the system.
Energy and entropy considerations must be included for completeness, but further use of
the results we obtain is restricted to our discussion of compression waves and to various exercises.
Sec. 14.4] Energy and Entropy
475
(ii) For a reversible process, excluding electrical and chemical effects (as we
shall throughout), an element of work dW done by the system on a unit
area can be computed by forming the product of the pressure P and the
element of distance ds moved by the substance on which the pressure acts.
(Summing such contributions, one obtains the formula dW = P dV, where
dV is a change in volume.) (iii) If heat Q is reversibly added to a system,
the corresponding entropy change S2 - S1 is given by the integral Je- 1 dQ,
where e is the absolute temperature. If the change in entropy is irreversible,
then S2 - S1 exceeds J0- 1 dQ. (iv) The entropy of an isolated system does
not decrease, and only in reversible processes does it remain the same.
EFFECTS OF INHOMOGENEITY AND MOTION
Situations in which properties change from point to point are handled
in a familiar manner. One defines point functions (intensive variables) whose
integral provides the corresponding overall (extensive) quantities. Thus, if
R is a region with boundary oR, at some point x in R or on oR, and at a
time t, e(x, t) = internal energy density per unit mass, at point x and time t,
and h{x, t, n) = heat efflux defined as follows. At time t, consider a point x
on the boundary oR of region R. Let n be the unit exterior normal to R at x.
Then h{x, t, n) is the rate of heat fl.ow per unit area through oR at x, from
the interior to the exterior.
The net internal energy in R at time t is JJJR e{x, t) pdt; the net heat efflux
is given by the surface integral .,Rh(x, t, n) du.
In contrast to equilibrium thermodynamics we must consider the effects
of motion. One effect is that the internal energy must be supplemented by
an energy of motion. This is the familiar kinetic energy with a density per
unit mass of !v v, where v is the velocity vector. Motion must also be
considered when net work is computed. Extending (ii) above, we assume
that the rate of doing work at a point is the product of the force at that point
times the local velocity. There will thus be contributions to the net work
from an integral over R of the body force pf times the velocity, and an integral
over the boundary oR of the stress t times the velocity.
We cannot limit ourselves to the homogeneous states of equilibrium thermo-
dynamics, where, for example, a single temperature characterizes the entire
domain under consideration. But we shall assume that the various dependent
variables have a sufficiently weak dependence on both space and time that we
can regard our domain as composed of a number of volume elements, each
in equilibrium but changing reversibly due to slight property differences
between adjacent elements. If this localequilibriumassumption is justified, we
can postulate with some confidence that the various classical equilibrium
relations hold for a material particle. Presumably, our results will be valid
if elements that are small compared to length scales of interest will reach
equilibrium in times that are short compared to time scales of interest.
Deeper study is certainly needed to build confidence in the equations, but
the reader can be assured that they have been successfully used in very many
instances. On the other hand, the search for a further-ranging thermodyna-
mics continues to the present day, spurred particularly by the "far-from
equilibrium" behavior that occurs in biology.
ENERGY BALANCE
Taking into account nonhomogeneity and motion, then, we postulate
the following extension of the energy balance* condition (i), for any material
region R(t):
!!
dt
fff(!v2
R(I)
v + e)p d-r= fff(f v)p d-r+ ff (-h
R(I) oR(t)
+ t v) da. (17)
The heat efflux h is preceded by a negative sign because inward fl.ow of heat
increases the energy in R.
It is instructive to contemplate the analogy between (17) and the angular momen-
tum balance equation for polar fluids (3.5):
i Jff
t R(I)
(x Av+ l)p dr
Both equations have the form of a general law of balance, namely,

rate of change of quantity
= rate of volume creation + rate of flux through surface.
But further, each termt is composed of a macroscopic part and a microscopic part.
For example, the energy is divided into kinetic and internal. Kinetic energy can be
thought of as being due to the gross macroscopic motion of the material, while
internal energy can be regarded as being due to microscopic internal degrees of
freedom. The distinction between macroscopic and microscopic seems to blur on
more intense scrutiny, however, for is it not true that all continuum quantities are
macroscopic manifestations of microscopic behavior? A deeper view of the distinc-
tion we seek seems to involve the notion of observer invariance. For example,
internal energy looks the same to two observers moving with constant velocity
relative to one another; kinetic energy does not. We shall not pursue these matters
further, except to note the remarkable fact that all the fundamental balance laws of
mechanics can be deduced from the energy balance postulate and suitable require-
ments of observer invariance. (See Exercise 2.3.13 of II.)
We assume a nonpolar fluid; otherwise there would have to be additional terms to represent,
e.g., the rate of working of the couple stresses c.
t In (17) r v could be supplemented by a heat generation contribution r. Such contributions,
however, will not considered here.
As in previous manipulations of balance laws we wish to derive a differ-

ential equation from the integral equation (17). To accomplish this, we must
first parallel the procedure used to obtain (2.18) and write
=fJJV (T v) d-c.
R
(18)
In the usual way (as in (1.8)] we can bring the time derivative on the left
side of (17) inside the volume integral. We can then consider (17) for a se-
quence of similar regions RL and deduce that
(19)
Applying (19) to a flake and a tetrahedron, we can show that there exists a
beat flux density vector q(x, t) such that
h(x, t, n) = n q(x, t). (20)
Using (18) and (20) in (17) and assuming continuity of the integrand, we
can finally obtain the differential equation of energy balance:
p (~v v + e)
!:... = p(f v) - V (q - T v). (21)
Dt 2
The detailed derivations of (19), (20), and (21) are so similar to the deriva-
tions of (2.9), (2.17), and (2.19) that the reader should have no trouble in
supplying them, as he is requested to do in Exercise 2(a).
ENTROPY, TEMPERATURE, AND PRESSURE

Roughly speaking, (21) is a generalization of the equilibrium version of the
first law
dE dQ dW
for reversible changes.
dt = dt -dt
We must also generalize the relations that connect heat input to entropy
through temperature, and work to volume change through pressure.
In the reversible case, these Gibbs relations are
dQ = E>dS' dW = p dV for reversible changes.

dt dt dt dt
Thus the first law can be written

dE dS dV
-=0--P-. (22)
dt dt dt
Temperature 0 and pressure P were regarded as known concepts in the
informal discussion from which (22) was generated as (9b) of Appendix 14.l.
(Other notation: E = internal energy, S = entropy, V = volume.) An alternate
approach is to arrive at (22) by starting with the definitions
0= (oE)
as v'
P=_(oE)
av s
, (23)
where the subscripts indicate the variable that is held constant in the partial
differentiation. (It is not without intuitive appeal, for example, that the partial
derivative of internal energy with respect to volume V is the negative of the
pressure.) Now two state functions determine any third state function. In
particular, if E = E(S, V), then
dE (aE) dS
di= as v Tt + av s Tt'
(aE)
dV (24)
and (22) follows with the aid of (23).

To generalize, let us now use capital letters to denote quantities that are
regarded as functions of material variables A and t. Functions of thermo-
dynamic variables will be written with an overbar. For example, E(A, t)
gives the internal energy at time t of the particle which was initially at A.
On the other hand, E(S, V) provides the internal energy of a particle whose
entropy and specific volume are Sand V, respectively. We assume that if
two state variables are known for a particle, then any third state variable is
uniquely determined. In particular,
E(A, t) = E[S(A, t), V(A, t)].
Paralleling (24), we have by chain rule that*
aE aEas aEav
(25)
ai = asai + av at
We can use the following counterpart of (23) to define temperature 0 and
pressure P:
a.E aE
0=-
- as' P=-- av (26)
With this, (25) becomes

aE as av
-=0--P-. (27)
at at a,
Since the dependence on S, V, A, and t has been exhibited, there is no need to make explicit
the variables that are held constant in the partial differentiations.
479
Let us return to spatial coordinates, making a corresponding shift to small

letters for the various variables. Entropy per unit mass, for example, satisfies
s(x, t) = S[A(x. t), t].
A partial derivative with respect to time keeping A constant becomes a
material derivative. In spatial coordinates, then, (27) is
(28)
In (28), p denotes density as usual. Mass per unit volume turns out to be a
more natural variable in continuum mechanics than its reciprocal, the volume
per unit mass.
The energy balance requirement (21) and the Gibbs relation (28) complete
the list of field equations that express fundamental physical laws that are
valid for any continuum for which chemical, electrical, and relativistic
effects are negligible. There is another fundamental law, however, an in-
equality which generalizes the thermostatic idea that entropy increase equals
or exceeds heat addition divided by absolute temperature. This is,
d!:t fffpsdt~ -
R(t)
fJ9- q n du,
oR(t)
1
(29)
where as usual n is a unit exterior normal to the material region R(t). In

differential equation form, assuming appropriate continuity, we have [Exer-
cise 2(c)]
(30)
Equations (29) and (30) are two versions of what is sometimes referred to
as the aausius-Duhem inequality.
INTERNAL ENERGY AND DEFORMATION RATE
Some manipulations of the basic energy and thermodynamic equations

provide insight into their implications and also provide alternative and often
more convenient forms. These manipulations give rise to expressions that
contain quantities D ii defined by the following combination of derivatives
of velocity components:
D .. =- l (OV
-' +-1 ov'). (31)
IJ 2 oxj OX;
The quantities D ii are called the componentsof the rate of deformation tensor
D. As their name indicates, these components give information on how and
at what rate the material is being deformed. Our later study of cartesian ten-
sors provides a background against which the meaning of the Du is more
clearly seen. This meaning is discussed in detail in Section 3.1. of II. The
exercises for that section request proofs of several results which elucidate
the relation between rate of deformation and energy charges. One of these
results will be noted in the next paragraph.
A relation involving kinetic energy is implied by the momentum balance
requirement (2.19)
Dv.
P _i
3
= pfj + i=lL OXj
a
- (Tu)
Dt
By taking the scalar product of this equation with the velocity vector v
and integrating over a material volume R, one finds that
d
dt
fff\~~2
R
vp d-c= Jfffvp d-c+ fJt v du - Jff T:D d-c,
R oR R
3
T:D =i,j=lL Tii Du. (32)
(A proof is requested in Exercise 3.1.7 of II). Thus the kinetic energy in a given
piece of material is increased by the rate of working of body forces and
surface stresses and decreased by a term involving the interaction of stress
and deformation. This result can be converted into the following differential
equation by a process used several times:
pit GvvJ=pf v+ V ( T v)- T:D. (33)
Upon subtracting (33) from the energy balance law (21), one obtains
De
P Dt = -V q + T:D. (34)
This equation shows how_internal energy results from heat flux and from the
stress-deformation term T:D that appeared in (33). This term is thus seen
to represent the conversion of kinetic to internal energy.
ENERGY AND ENTROPY IN FLUIDS
Thermodynamic relations are particularly important in fluid mechanics,
so it will be worthwhile to abandon complete generality and derive a few more
relations concerning energy and entropy in fluids. This will require antici-
pation of the characterization of fluids to be given later, but this should not
cause difficulty.
We split the components of the stress tensor into a part proportional
to the Kronecker delta and a "leftover" part with components Vii:
(35)
Sec.14.4] Energy and Entropy
For compressible fluids, p can be identified with the pressure. When viscous
effects are ignored, Vii= 0. Substitution of (35) into (34) yields
De
p Dt = -pVv-Vq+ , = V:D. (36)
From this, one can deduce [Exercise 3(b)]
(37)
which shows how the entropy of a material particle changes. Both terms on
the right side presumably represent entropy change due to heat flow divided
by temperature. In the second must therefore represent heat generation
by deformation. Indeed, the dissipation function Cl>represents the rate per unit
volume at which mechanical energy is dissipated into heat.
Comparison of (37) with the second law of thermodynamics in the differ-
ential equation form (30) leads to the requirement
(38)
Sufficient conditions for satisfaction of this requirement, and therefore
sufficient conditions for the validity of the second law, are
-q VO::: 0, (39)
The first condition states that heat always flows against a temperature gradi-
ent, the second that deformation never converts heat into mechanical energy.
EXERCISES
1. (a) In general, the specificheats Cv and Cp are defined as the rate of
change of heat with temperature at constant volume and pressure,
respectively. That is,
Cp = (aQ)
ae '
p
where we use subscripted parentheses to remind us of the variable

held constant in the partial differentiation. Use the first law to
justify Cv = (oE/o0)v. Deduce that an ideal gas satisfies (2b).
(b) Show that consideration of changing temperature at constant
pressure gives the result dQ = Cp d0; hence Cp = Cv + R.
[Use (a).]
2, (a) Supply detailed derivations of equations (19), (20), and (21).
(b) Verify (33) and (34).
(c) Derive (30) by extending the Dubois-Reymond lemma.
3. (a) Verify (36).

(b) Use the continuity equation and (28) to derive (37).
(c) Verify (38).
t4. If the velocity v of a motion satisfies v = V<f,for a scalar function ef>,
then the motion is said to be irrotational. Suppose that a material has
constant density and occupies a simply connected region R. Let K be
the kinetic energy of the irrotational motion of such a material and K*
the kinetic energy of any other motion having the same normal velocity
at the boundary. Prove that K::::; K*.
5. Note that in their differential equation versions the law of mass con-
servation, and the laws of momentum and energy balance, all have the
form
DF
p-=pQ-VJ. (40a)
Dt
(a) What is the general interpretation of equations with this form?

In particular, why is the Q term missing in the mass conservation
equation but not in the other two balanceequations?
(b) Why can v be interpreted as a mass flux vector, and -T v as
an energy flux vector? (The flux of a scalar is a vector. The flux
of the momentum vector must be a "higher" quantity in some
sense, and analogy with the other equations shows this flux to be
given by the negative stress tensor -T.)
(c) Explain why
d Jff
dt R(t)
pF d-r =fff pQ dt + ff j(n) du
R(t) oR(I)
(40b)
is the integral form of a general balance law; in particular, what.is

j(n) =j(x, t, n)? Show thatj(-n) = -j(n), and hence thatj(x, t, n)
= n J(x, t) for some vector J. Deduce that a continuity assumption
allows one to derive (40a) from (40b). [Exercise 8 shows what
happens in the presence of discontinuities.]
The field equations were initially formulated as integral equations
involving arbitrary material regions. When the integrands are continuous,
differential equations can be deduced. When the integrands are dis-
continuous (as may be the case), one can derive relations between the
various discontinuities. This is done in the next four exercises.
6. Let Q(t) be the set of points x that were in a given region Q(t 0 ) at
time t 0 and whose movements are governed by a certain function y.
That is,
Q(t) = {x Ix = y( Y, t) for every Y in Q(t 0 }, where y(Y, t 0) = Y}.

Show that if F and p are smooth functions of space and time, and if
f is defined by
f(t) = fff Fp d,,

Q(r)
then
f = Jff ~
t Q(r) ut
(Fp) d, + ff Fpwndu,
oQ(r)
(41)
= =
where w oy/ot,wn w. n, and n is the unit exterior normal to oQ.
This result is similar to the Reynolds transport theorem (1.10). The
present problem deals with the change of a quantity that is integrated
over a region moving according to a context-free rule given by the
abstract function y, whereas (1.10) deals with the case y = x, Y = A,
so that the region moves according to the transformation x=x(A, t).
Recall that when this particular notation is used we speak of Q as
a material region and think of the motion of a continuum with velocity
v = ox/ot.
The interpretation of (41) is as follows. The instantaneous rate of
change at time t of a quantity obtained by integrating Fp over a moving
region Q(t) is caused by two factors:
(a) The change of Fp with time at all points that are included in the
region Q at time t.
(b) The net distortion of Q due to the motion of its boundaries with
=
velocity oy/ot w. Locally tangential motion of the boundary
adds no new points to the region of integration; hence the appear-
ance of the normal component wnin (41).
7. Let R be a material region that is divided into subregions R- and R+
by an arbitrarily moving surface I. Dependent variables, including
density and velocity, may be discontinuous across I. Let s-cs+)
be the portion of oR, the boundary of R, which is also a boundary of
R-(R+). Thus
oR=s- +s+,
Let n be the exterior normal to R and let n1:be the normal to I which
points into R+. Lef n-(n+) be the exterior normal to R-(R+). (See
Figure 14.13.)
If a is a point on I, we define for any scalar or vector quantity E
E+(a) = lim E(x), -(a)= lim E(x),

x-a
x in R+
FIGURE 14.13. Diagram/or the calculation of relations between discon-

tinuities across the surface :E.
(a) Apply the result of Exercise 6 to R+ and R-, then add. (Be careful
of the normals.) Derive
d fff
dt R(I)
Fp th= Iff aat (Fp )d-r+ ffFpvn du - ff[pF]wn du.
R(I) BR :t
(42)
Here vn = v n where v is the velocity of a particle, wn = w Dt.

=
(b) Let v;' = v n. For a on I:, define s,;'(a) [w(a) - v(a)] D:2:,
the relative normal speeds of the discontinuity surface. Derive
d fff
d t R(r)
Fp d-r = Iffa0t (Fp) d-r+ ffp+ p+v: du+ ffp-p-v;
R(t) BR+ BR
du
- If[snpF]du.
I:
(43)
(c) Assume that there are no discontinuities except across I:. Derive
finally, now identifying p with the density, that
d Jff
d t R(r)
Fp dT =
R(r)
fff
DDFp dT -
t
ff[pFsJ du.
t
(44)
8. (a) Define [j] by

LJ(x, t, n)] =r(x, t, nt) - F(x, t, nJ, {45)
where we have used notation from Exercise 7. Deduce that
[pFsn] + LJ] = 0 (46)
at each point of a surface of discontinuity.
{b) Deduce the following "jump conditions."
Conservation of mass: [psJ =0. {47a)
Balance of linear momentum: [pvsn + t] = 0. (47b}
Balance of energy: [p(e + tv v)sJ + [t v - h] = 0. {47c)
Sec. 14.5] Constitutive Equations, Covariance, and Continuum Model 485
(c) Show that for a nonpolar material, balance of moment of mo-

mentum adds no new restrictions.
9. (a) Suppose that the discontinuity surface is actually a material surface.
Show that the jump condition derived from the mass conservation
equation in the previous exercise is no longer a restriction and that
an arbitrary discontinuity in density is allowable.
(b) Suppose that a material surface divides two inviscid fluids, where
t = -po. (Here pis the pressure; see Section 15.1.) Show that the
momentum balance jump condition leads to the requirement that
the pressure be continuous across the discontinuity surface.
For more applications of the jump conditions, see the exercises for Section
15.3.
14.5 On ConstitutiveEquations,Covariance,andthe
ContinuumModel
A wide view of continuum mechanics is provided in this section. We begin
by recapitulating the field equations so far derived. We see that there are
many fewer equations than unknowns, and we give a few simple examples
of the constitutive equations that are needed to fill out the picture by pro-
viding the nature of a particular material's behavior.
When formulating constitutive equations, perhaps the only general re-
striction is the requirement that the equations not depend essentially on the
choice of a number of possible, equivalent, coordinate systems. This is an
instance of the principle of covariance. In a brief discussion of this principle,
we give as a primary example the " derivation" of the Lorentz transformation
from the postulate that the velocity oflight be the same in systems that display
uniform relative motion.
It is appropriate to conclude the chapter by reviewing the appropriateness
of the continuum model. We shall see that this model is only applicable
if the typical distance between molecular collisions is small compared to
length scales of interest.
RECAPITULATION OF FIELD EQUATIONS
Let us recapitulate the equations we have obtained for a nonpolar single-

phase continuous medium, which does not exhibit electric, magnetic, or
chemical effects.
Conservation of mass:
Dp
-+
Dt
pVv=O. (1.6)
Balance of linear momentum:

Dv
p-=pf+V T. (2.20)
Dt
Balance of moment of momentum:
T= TT'. (3.6)
Balance of energy:
P ~t [ Gv v + e)]= p(f v) - V (q - T v). (4.21)
Gibbs relations:
(4.28)
() = oe(s, v) . -oe(s, v)
p= .
OS ' av
Second Jaw of thermodynamics:
Ds
p Dt ~ -V. (O-lq). (4.30)
Kinematic boundary condition:

vn prescribed on boundary. (13.2.22)
The unknowns are density p, velocity vector v, pressure p, specificvolume v,
heat flux vector q, temperature (), specific internal energy e, specificentropy
s, and stress tensor T. Regarded as given is the specific body force f.
There are 21 unknown scalar functions, counting three components for
the vectors v and q and nine components for the stress tensor T. There are
the equivalent of 12 scalar equations, one inequality, and one boundary
condition. We must reconcile the disparity between the number of equations
and the number of unknowns with our feeling that if the initial state of a
medium is known, then normally the laws of mechanics should uniquely
determine how the medium moves and changes.
Predictions of existence and uniqueness of solutions from tallys of the
number of equations and unknowns is dangerous, even for systems of linear
algebraic equations. Our system is vastly more complicated-but an explana-
tion of the excess number of unknowns does not lie in this direction.
INTRODUCTION TO CONSTITUTIVE EQUATIONS
It is as obvious as any statement about the material world ever is that
additional equations are necessary to describe fully the restrictions on the
motion of a given material. The deformation of a railroad bridge as a train
passes over it and the meandering of the Gulf Stream are both limited by
the common necessity to conserve mass, balance momentum and energy, etc.
But there must be further different constraints which describe the fact that
throughout the deformation steel remains steel and water remains water.
Constitutive equations is the term given to the mathematical descriptions of
the various kinds of solidity of solids, fluidity of fluids, and the exotic quality
of exotic materials.
Here are some examples of constitutive equations. In a rigid body, the
distance between any two particles remains constant. In an incompressible
medium the density of each particle is constant, so Dp/ Dt = 0. For a perfect
gas (with constant specific heats), the pressure p and the density are related
by p = KpY, where K and y are constants for a given gas. For ordinary
materials the heat flux vector q is related to the temperature () by the Newton-
Fourier law of cooling q = - KVO, where the scalar thermal conductivity K
depends on(), and to a small extent (usually) on the pressure p. In Chapter 15
we consider an inviscidfluid for which the stress vector t, the normal vector n,
and the pressure p are related by t = - pn.
Some constitutive equations, such as those for rigid bodies and incom-
pressible materials, are reasonably obvious idealizations. Others, such as
Newton's law of cooling, are generally regarded as obtained by experiment.*
But to obtain the constitutive equations such as those for ordinary elastic
solids and viscous fluids, one starts with assumptions that hardly seem obvious
even in retrospect, and then simplifies by means of some reasonably so-
phisticated mathematical manipulations. These manipulations are most easily
performed with the apparatus of tensors. That this apparatus is not necessary
is demonstrated by the fact that it was not used by the men who originally
formulated the constitutive equations for elastic solids and viscous fluids.
Tensors do seem a necessary conceptual tool, however, in modern attempts
to find constitutive equations for more exotic materials. We therefore post-
pone our major discussions of constitutive equations until II, after tensors
have been introduced.
THE PRINCIPLE OF COVARIANCE
The study of constitutive equations perhaps brings into the sharpest focus
the need to stress the principle of covariance. Indeed, many of the modern
attempts to find constitutive equations for more exotic materials are based
on this principle.
The principle of covariance is most conveniently formulated in terms of
tensor calculus. A detailed discussion of covariance and its relation to car-
tesian tensors is provided in II. At this point, let us explain the principle only
briefly.
* This is often an oversimplified view. For example, see J. A. Ruffner's article, "Reinterpre-
tation of the Genesis of Newton's Law of Cooling." Arch. Hist. Exact Sci. 2, 138-52 (1964).
(Note that, as above, the name Fourier is often also associated with this cooling law.)
Basically, the principle of covariance (or invariance) states that all physical
laws must take on the same form when stated in "equivalent" frames of
reference of the observers. Thus two observers who choose two different sets
of coordinate axes, stationary relative to each other, should certainly find the
same physical laws stated in the same form. Examples are provided in Exer-
cises 1.8 through 1.11. To give another example, consider the equation of
beat conduction in two spatial dimensions:
(1)
If a second observer employs another Cartesian coordinate system, he will

find the equation
ar = "(ar2 + a T').
2 2
2
(2)
at ax' ay'
However, there is also a mathematical connection x = x(x') and x' = x'(x)

between the coordinate systems (x, y) and (x', y'). Since 1e- 1 aT/at is inde-
pendent of the coordinate system chosen, it must follow that
2 2 2 2
a T a T [a T' a T']
-a 2
X
+ -aY 2 = -a ,2
X
+ -aY ,2 :w:'=:w:'(x)
, (3)
as can inl> ' he verified (Exercise 1).

When vecto.ial quantities are involved, the situation becomes much more
complicated. The formalism of vector analysis avoids this issue. However,
this is only superficial. One must make sure that a certain quantity, such as a
velocity, is a vector. Not every quantity with a direction and a magnitude
is a vector. For example, an infinitesimal rotation or an angular velocity
is a vector, but a finite rotation is not. (It is a tensor, as we shall see later.)
Besides tensor analysis, the principle of covariance can underly other
theories. Thus the variational formulation of mechanics provides an illus-
tration of an extensive branch of applied mathematics that is developed in a
manner independent of a particular choice of coordinates. This leads to the
development of the subject in terms of generalized and canonical coordinates.
The classical developments of the theories of Lagrange, Hamilton, Jacobi,
and others may be found in treatises of mechanics.
As an example of the usefulness of the principle of covariance, consider
the Michelson-Morley experiment, which shows that light propagates at a
constant speed whether the frame of reference is moving with constant velo-
city or not. We shall now demonstrate that the Lorentz transformation is
the only linear transformation which leaves the propagation speed of waves
invariant in two (relativistically equivalent) frames of reference that are in
uniform relative motion.
The wave equation
(4)
describes light propagation in the frame of reference of the observer (x, t).
Now, if another observer, moving at a speed v relative to the first, attempts
to describe the same phenomenon with his own frame of reference (x', t'),
he finds that the light waves are also propagating with the same speed c
(not c v). We wish to show that this can be true if the two frames (x', t')
and (x, t) are related by the Lorentz transformation
, x-vt
x= ' (Sa)
J1- v /c 2 2
, t - vx/c 2
t = . (Sb)
J1- v /c 2 2
We observe that in the new frame of reference, the equation of wave pro-
pagation for u'(x', t') = u[x(x', t'), t(x', t')] takes on the standard form
(6)
One can readily verify that the linear transformation (5) does transform (4)
into (6) (Exercise 2). However, we wish to "discover" (5).
Let us recall, as we just mentioned above in connection with the heat equa-
tion, that the Laplace operator remains invariant under a rotation of axes.
Thus, if we define
y = ict, y' = ict', (7)
(4) and (6) would transform into each other if the variables are related by a
rotation between the (x, y) system and the (x', y') system; i.e., if for some angle
(),
x' = x cos () + y sin (), (Sa)
y' = -x sin()+ y cos(). (Sb)
If the equations of (8) are to represent real relations between (x, t) and (x', t'),
then () must be complex. Indeed, if () = i</>, where </>is real, we derive from
(8) the real relationships
x' = x cosh </>- ct sinh </>, (9a)
ct'= -x sinh </>+ct
cosh <J>. (9b)
To obtain a physical interpretation for </>,
we note that to an observer using
the (x, t) system, a point x' = constant fixed in the (x', t') system satisfies
dx
dt = c tanh <J>.
Thus
v
tanh = -,
c
where v is the velocity of the frame (x', t') as measured in the frame (x, t).
Conversely, for x fixed, we can verify from (9) that
dx'
-= -ctanh ,1. = -v
dt' 'I' '
which confirms the symmetry of the two systems. Equations (9) can now be
easily rewritten into the form (5) (Exercise 3).
We are all well informed of the revolutionary changes of our fundamental
concepts caused by the special theory of relativity, whose essential mathe-
matical basis is the Lorentz transformation. Here is a powerful demonstration
that physical issues must be examined and appreciated beyond pure mathe-
matical formalism. At the same time, it is a demonstration that the seemingly
innocuous principle of covariance can lead to important consequences.
VALIDITY OF CLASSICAL CONTINUUM MECHANICS

Now that the foundations of continuum mechanics have been laid, it is
appropriate to reconsider the relationship between the continuum and the
particulate models of a material. Complete understanding of this relationship
is still an unattained goal; here we confine ourselves to some qualitative
remarks.
As we pointed out in Section 13.l, continuum variables such as density
and velocity can be defined, a priori or by an averaging process, providing
that a typical interparticle distance is small compared to a typical length scale
of the phenomenon under investigation. Other continuum variables can be
defined in the same fashion. For example, the components of the stress tensor
at a point can be related to average values of momentum flux. But for a con-
tinuum description to be of use, not only must continuum variables be
definable, but it must be possible to relate these variables by equations that
determine them completely.
In classical continuum theories the values of the various variables at a given
point of space and time are determined by a set of partial differential equa-
tions. These equations relate the values of the unknowns and their derivatives
at each of the various points of the spatial and temporal domain of interest.
Implicit in these equations, then, is the assumption that the average variables
are locally determined.For if the change in average quantities at a point
were influenced by faroff configurations, then no set of equations at that
point could completely determine these quantities.
Local determinism requires that particles do not "remember" their past
for long. To be more precise, the "forgetting time" must be short compared
Appendix 14.J] Thermodynamics of Spatially Homogeneous Substances 491
to time scales of interest if ordinary continuum theories are to be employed.

An example illustrating a situation when this is not the case is provided by the
very high frequency waves which make up ultrasonic phenomena.*
The process of effectively "forgetting" the details of past configurations
is accomplished by the interactions between particles. One. can imagine a
group of people who are blindfolded and set out into a crowded room.
After a few bumps into others, each blindfolded person will become dis-
oriented. Similarly, after a few interparticle interactions (which are usually
called collisions),the fate of a given set of particles will scarcely be predictable
from their initial configuration. t
The term mean free path is used to denote a typical distance between
collisions. Another way of saying that "forgetting time" must be short
compared to time scales of interest is that the mean free path must be short
compared to distance scales of interest.
To summarize our all-too-brief discussion, for classical continuum theories
to be applicable both the typical interparticle distance and the mean free path
must be small compared to length scales of interest. When the first of these
conditions holds, but not the second, a modified continuum approach can
be used-as in the treatment of stellar dynamics in Section 1.2.
EXERCISES
1. Use (8) to verify (3).
2. Verify that if a change of variables is described by (5), then (4) trans-
forms into (6).
3. Show that (5) is an equivalent form of (9).
Appendix 14.1 Thermodynamics of Spatially Homogeneous

Substances
We discuss here the thermodynamics of continuous substances that are
spatially homogeneous. By a spatially homogeneoussubstance we mean one
that can be regarded as characterized by a single value of such variables as
temperature, pressure, etc. (Thus we can speak of the temperature of the
* Considerable effort is being devoted to formulating generalizations of classical continuum

theories that will provide improved predictions in situations such as those found in ultrasonics. A
good illustration is provided by the papers of Ph. Selwyn and I. Oppenheim, Physica 54, 161-94,
195-222 (1971).
t "In principle" the future configuration of a set of particles may be completely determined by
initial conditions. After a few collisions, however, inevitable slight errors in specification of the
initial conditions will have brought about such a large cumulative error that further calculation
would be meaningless.
Our presentation has been considerably influenced by H. C. Van Ness's excellent elementary
explanation of thermodynamic concepts in the slim paperback Understanding Thermodynamics
(New York: McGraw-Hill, 1969).
substance and the pressure.) In the present context, homogeneity will also
be regarded as implying lack of a preferred direction of molecular motion-
thus only motionless substances will be discussed. A substance that remains
in the same homogeneous state is said to be at equilibrium.
An example of the type of system to be considered is provided in Figure
14.14. There a motionless quantity of gas is contained between thermally
insulating walls and an insulated piston. With the gas can be associated various
quantities, such as its volume V, the pressure Pit exerts, its temperature 0
(we reserve the letter T for stress), and its internal energy . It is an experi-
mental fact that substances having the same values of any pair of these
state variables will always have the same value of the remaining variables. By
contrast, there are measurable quantities, such as the heat Q, which are not
determined by the values of a pair of state variables. These quantities depend
on the history of the manipulations to which the material was subjected.
FIGURE 14.14. Gas contained by a piston that is weighted down by a

container filled with sand. The system is regarded as perfectly insulated
so that no heat can flow into or out of the gas.
EXPERIMENTS WITH A PISTON
Consider some thought experiments on the configuration of Figure 14.14.

There the weight pushing down on the piston is partly composed of fine sand
located in the container pictured in the figure. Suppose that a single speck
of sand is removed. (We do not wish to complicate our discussion by con-
siderations of work against gravity, so we imagine that the sand is trans-
ferred, without changing its height, to a shelf.) On removal of the speck, the
piston would move upward a very small amount (assuming a frictionless
piston). There would only be a tiny amount of nonuniformity introduced
into the system, and that for a very short time. If the remaining grains of
Appendix 14.1] Thermodynamics of Spatially Homogeneous Substances 493
sand are removed, one at a time, the piston would gradually become elevated.
This elevation is a result of the work W done by the gas on the piston and the
sand. Now W = JdW, where, by definition, dW = F ds. Here Fis the force
and ds is the element of distance moved against that force. Because each
tiny step of the process barely disturbs the uniformity of the system, with
little error we can continue to speak of the pressure P in the system at each
volume V. It is this pressure that is responsible for the force on the piston.
If the area of the piston is A, the magnitude of the force is PA. Suppose that
the initial gas volume was Yi and the final volume V2 Now F ds = PA ds
and A ds = dV, the element of volume. Therefore,
work d?ne by the gas on the piston in slowly= fv 2 P(V) dV. (1)
expandmg from volume Yi to volume V2 v ,.
If the specks of sand were placed back on the piston, one by one, we assert
that the piston would be found to return to a position that would be indis-
tinguishable from its original position. Although the temperature of the gas
would be found to drop as it expanded, the temperature would also return
to its original value when the sand specks were replaced.
Suppose, in contrast, that all the sand were suddenly removed. The
piston would shoot upward, fall back, rebound upward a little, oscillate a
bit more perhaps, and finally settle to a new level. During the course of this
process the pressure and temperature would not remain uniform throughout
the gas. (Remember, heat fl.ow is prevented by insulation.) In particular,
since one could not speak of the pressure P at a given volume V, one could
not use (1) to evaluate the work done. One would observe in addition the
qualitative result that the piston would not rise as far as it did in the very slow
process considered in the previous paragraph. Also, the temperature would
not decrease as much. If the sand were suddenly dumped back onto the piston,
the apparatus would not return to the position it was in before the experiment
started. Neither would the temperature return to its initial value. It would,
in fact, be higher.
We shall return later to a discussion of the irreversible processes, such as
that which occurs when all the sand is suddenly removed. For the present
it is sufficient to note that the course of such processes cannot be described
by a formalism restricted to homogeneous states. On the other hand, the
final result obtained by carrying out an irreversible process and then waiting
a time of sufficient duration is a homogeneous state and can be described.
Furthermore, we have seen that a very slow process virtually does preserve
homogeneity at all times, and that such processes are virtually reversible.
We shall be able to describe the course of limiting infinitely slow reversible
processes. " Limiting" here is used essentially in its mathematical sense,
so there should be no cavils that an infinitely slow process will never achieve
a change. The slower an actual process is, the more nearly it will be approxi-
mated by the limiting reversible case.
For simplicity we considered an insulated cylinder and hence an adiabatic

process in which no heat was added to the system or removed from it. But
one can also consider essentially reversible processes in which heat is slowly
added by means of tiny temperature differences between the inside of the
container and the outside.
We shall now formulate the first of the two main postulates of equilibrium
thermodynamics. Like all physical laws, these postulates cannot be "proved"
by experiment. Rather, they conjecture that a certain amount of experiment
and experience has revealed results of a universal character. The validity
of the conjectures is established by comparing deductions from them with
experiment. Consequently, it is not profitable to engage in exacting scrutiny
of the heuristic reasoning that leads to the conjectures. Some feeling for the
heurism must be acquired, however, if only because "laws" that are posited
without supporting discussion have little psychological appeal.
Two versions will be given for each of the thermodynamic " laws." The
first will be a comparison of two homogeneous states, the second a statement
about the behavior of slowly varying systems.
FIRST LAW OF EQUILIBRIUM THERMODYNAMICS
First Form. Let two homogeneous states of a system have internal

energies 1 and 2 , respectively. Then if llE = 2 - 1 ,
llE=Q-W,
where Q is the heat added to the system and Wis the work done by the system
in passing from state 1 to state 2. (Note that nothing is said about the process
by which state 2 is attained from state 1. This process may be irreversible,
so that nonhomogeneities are present in the course of the process.)
Second Form. For reversible processes, a shorthand* for the computa-

tion required to compute changes in internal energy from the amount of heat
transferred and work done is
dE=dQ-dW. (2)
Equation (2) reminds one of the line integral that is required for calculation
of the change in . In the present discussion, let us regard P and V as in-
dependent state variables. Suppose that a system passes from a state Li
with internal energy 1 , pressure P 1 , and volume V1 to a state L2 with internal
energy 2 , pressure P 2 , and volume V2 Suppose that Lz
is reached from
* The "shorthand" is analogous to the fonnula ds 2 = dx2+ dy 2 for the arc length of a plane
curve. If the curve is given by x = x(1), y = y(r), 11 ~ t ~ 12 , then of course the "long-hand"
formula is s = f ds = f!i [(dx/dr)2 + (dy/dr)2}1' 2 dt.
Appendix 14.J] Thermodynamics of Spatially Homogeneous Substances
495
along a state curve L that specifies the P-Vvalues

}: 1 traversed. (See Figure
14.15.) The first law (2) tells us that
E2 - E1 = I -aQ aQ
dP + - dV -
f -aw dP + -aw dV (3)
1: aP av aP1: av
As usual, a line integral can be converted to an ordinary single integral
if a parametrization of the curve L is available. Suppose that such a para.
metrization is given by
P = P(t), v= V(t), where t is time, t1 ~ t ~ t2 (4)
Then
E2 _ E 1 = J'(aQ
,,
dP + aQav) dt _ J'(aw dP + aw av) dt
2
aP dt av dt ,, aP dt av dt
2

(5)
[In (5), P(t) and V(t) are substituted for P and V, the arguments in the deriva
tives of Q(P, V) and W(P, V).)
'------------+-V
FIGURE 14.15. The P-V state plane. The passage of the system from
state ~1 to state ~2 is fully described by specification of the state curve ~-
Since, for example,

a _ _ aQaP aQav
dt Q[P(t),V(t)] = aP dt + av Tt'
in (5) we can replace t 2 by a variable upper limit t, differentiate with respect
tot, and obtain the first law as the differential equation
(6)
This equation is valid only in the limit of very slow changes.

In the piston system of Figure 14.14 we have seen that dW = P dV. Gen
eralization of our earlier argument establishes the same expression for the
work of expansion in such systems as balloons of gas. More general systems

can accomplish chemical and electrical work, but we shall not introduce
such generality here.
A very important point is that neither the heat Q nor the work W is a
state variable. As can be verified by experiment, Q and W in a reversible
process depend on the process, not just on the initial and final states. In
particular, it is profitless to try to ascertain "the heat content" of a piece
of material; for, suppose that to establish a reference point the material were
regarded as having zero heat at some moment. A later value of the material's
"heat content" would presumably be equal to the measurable quantity
"net heat added." But if the exact details of the entire history were ever lost,
then the value of its heat content could not be checked-since the present
state of a system gives no clue to the way heat was added to it in the past.
The same argument shows that it is profitless to try to assign "work content"
to a piece of material.
Since W and Q are not state variables, the line integrals Ji:dW and J:i:dQ
would give different values for different state curves I. [But the first law
guarantees that the difference of these two integrals, Ji:(dW - dQ) = J:i:dE,
depends only on the initial and final points of I, for Eis a state variable.]
In other words, neither dW nor dQ is an exact differential. Note, however,
that (in a reversible process) multiplication of dW by the integrating factor
p-t provides an exact differential, for p-t dW = dV. * Empirical establish-
ment of the fact that p- i dW is the differential of a state function might be
accomplished by forcing some substance to pass in various ways from state
I, 1 to state I, 2 by doing work very slowly. Suppose that the pressure was
measured frequently so that plots of p- 1 versus W could be constructed, as
in Figure 14.16. Then the area under these curves would all be the same, for
f p-i
I
dW= f dV=
I
V2 - V1 , (7)
and the difference between volumes does not depend on the process by which
the system passes from its initial state to its final state.
ENTROPY
An important observation is not yet reflected in our theoretical construc-
tion, namely that there is a "natural" direction to processes. For example,
if a cold metal bar is brought into a room, heat always flows into the metal
until the temperature of the bar and the room are equalized. No energy has
been lost, it has just been spread uniformly around. But uniformly distributed
energy is not available to flow and do useful work. Indeed work has to be
done (say by means of a refrigerator) to restore the original nonuniform
state. The Nobel physicist Richard Feynman sums up the observations in a
Since dW = P dV for a reversible process, only a change in V causes an alteration in W.

Thus 8W/8P = 0 in (3) and (5).
Appendix 14.1] Thermodynamics of Spatially Homogeneous Substances
497
p-1
State l: 2
p-1
2
p-1
1
--~~~~~~~~~~~W
FIGURE 14.16. Results from two hypothetical reversible experiments

plotted as curves of 1/P versus W. Although two different values of W
are found at the same state l:z, the area under the two curves is the same
(up to their respective termination points). But if f i p- 1 dW depends only
on the initial and final points of l:, then p- i dW is an exact differential of a
state function (the volume V).
picturesque way* by comparing energy with water that initially is located

all over a man who has just emerged from a bath. Two handy but tiny towels
are dry. The man pats himself first with one towel and then the other, but
after awhile no further transfer of water can take place (in this idealized
model). The man and the towels are all equally damp. All the water (energy)
is still there, but no further natural "flow" will take place.
None of the state functions that we have discussed changes in a "one-
way" fashion. With hindsight, at least, it is natural to speculate that there
is a missing state function, and that this function will indeed change in a
directional manner. We have pointed the way toward discovery of this func-
tion in an earlier paragraph. Multiplication of dW by the integrating factor
p- 1 converted it into an exact differential, the differential of the state function
volume. In addition to dW, the first law contains another nonexact differ-
ential, dQ. And indeed, experiments for reversible processes show that an
exact differential results if dQ is multiplied by an integrating factor, in this
case 0- 1 (Figure 14.17). The state function whose differential is given by
e- 1 dQ is called the entropyS. Thus
dS = 0- 1 dQ (reversible process). (8)
No TE. For (8) to hold, temperature must be measured with respect to

a zero value at the lowest possible temperature. Thus in all calculations in-
volving thermodynamics an absolute scale (such as the Kelvin scale) is used
in ascribing a value to 0. We emphasize that only for reversible processes
does (8) provide a shorthand for a method to calculate entropy changes by a line
integral.
In The Character of Physical Law (Cambridge, Mass.: MIT Press, 1967).

9-1
State l: 2
e-1
2
'-------------Q
FIGURE 14.17. Results from two hypothetical reversible experiments
plotted as curves of 0- 1 versus Q. The area under the two curves is the
same. Thus 0- 1 dQ is an exact differential of a state function (the
entropyS).
With (8) and the expression dW = P dV, which holds for reversible ex-
pansion and compression processes, the first law can be written
dE= 0dS-PdV, (9a)
or for slow changes,
dE dS dV
-=0--P-. (9b)
dt dt dt
Is there a directionality to the change of S? To obtain some insight, con-
sider two experiments with the insulated piston of Figure 14.14. In both
cases we remove the sand from the piston so that the gas expands from its
Li
initial state to its final state :Z:
2 Since the piston is insulated, the expansion
is adiabatic.
CA s E l . Slow reversible adiabatic process. Here
S2 - S1 = fdQ.
0
The system is insulated; so, certainly, no heat is added. Thus S 2 = S1 :
the entropy does not change. (Consequently, the process is called isentropic.)
CA s E 2. Irreversible adiabatic process. Consider the P-0 diagram
of Figure 14.18. For reference, the reversible expansion from :[ 1 to some
state :Z:
3 , caused by slow removal of the sand, is represented by a solid curve.
The course of the irreversible process cannot be wholly characterized on the
diagram, because at a given time no single values of P and 0 characterize
the system. It is helpful to consider average values of P and 0, represented
by the dotted curve, but only the end point :Z: 2 is of importance. The P
coordinate of this point is the same as the corresponding coordinate in the
reversible case, because the same amount of sand has been removed. But
Appendix 14.1] Thermodynamics of Spatially Homogeneous Substances 499
I:1
p
FIGURE 14.18. The curve from ~1 to ~3 represents a reversible expan-

sion of a piston. The dashed curve represents the average course of an
irreversible expansion that ultimately results in a gas state ~ 2 with the
same pressure as ~3. Properties of 'J:.2can be deduced by imagining the
effect of the reversible cooling necessary to bring the system from :I:.2
to ~3 (dotted curve).
experiment shows that there is a "wasteful" extra heating in the fast ir-
reversible expansion, hence the higher temperature coordinate at point L2
To determine information about the state of the system at L2 , we adopt the
frequently used stratagem of imagining a reversible process that would
bring the system to a state with known properties. The state of the system
could be brought from I; 2 to I; 3 by a slow reversible cooling. Since heat
would have to be removed from the system, dQ is negative in the formula
dS = dQ/0 and the cooling would result in a decrease in entropy relative
to the uniform value of the entropy which is characteristic of the states along
the reversible curve between Li and L3 This reasoning shows that the
entropy increases in the irreversible adiabatic expansion from Li to Lz.
The same reasoning applies to adiabatic compression. If it is reversible
.it is isentropic, but irreversible adiabatic compression increases entropy
(Exercise 1). Similar results are obtained in other experiments and thought
experiments with systems, like the piston, which are insulated from their
surroundings so that there is no heat transfer across their boundaries.
Indeed, presumably one must excludethe transfer of anything at all across the
boundaries. In isolated systems, then, one is led to the hypothesis that the
entropy never decreases, and only stays the same in reversible processes.
A related hypothesis is that the entropy in the universe never decreases,
for the universe is " isolated" in the sense that nothing can enter it from
"elsewhere." But we must beware of generalizing too far from our mundane
experiences. In most heat transfer experiments, it is hard to trace any in-
fluence of surroundings farther away than the building, or perhaps the city
soo [Ch. 14
Field Equations of Continuum Mechanics
in which the experiment is being conducted. It is perhaps not surprising,

therefore, that there are unanswered questions involved in truly cosmological
thermodynamics.* To evade such questions, let us hypothesize that the
entropy of a systemplus its surroundingsnever decreases,and leave somewhat
vague the definition of " surroundings."
A useful property of entropy is implied by the statement in italics that
concludes the last paragraph. This property applies to systems that change
irreversibly but slowly enough so that they can still be characterized by a
singletemperature0. (An exampleis providedby a systemcomposedof two
constituents which are allowed to mix, or by a system containing a slowly
moving but "nonfrictionless" piston.) We know that if the change were
reversible, then the entropy change would be given by f e- 1 dQ. We assert
that in the irreversible case the entropy change exceeds this quantity.
To back up our assertion, we begin with the observation that the heat ff.ow
into the system should have an effect that is independent of the type of
surrounding. We therefore imagine surroundings that have a uniform tem-
perature ~ which is equal at any time to the temperature of the system.t
No irreversible aspects will be allowed in our imagined surroundings, so the
slow temperature change of the surroundings will be reversible.
Let S1 and S 2 denote the values of entropy in the initial and final states of
the system, and let S1 and S2 denote the corresponding entropies of the
surroundings. Using the fact that the heat flowing into the system is ft.owing
reversibly out of the surroundings, we have
S2 - Si= J~-tdQ.
But by the italicized statement above
S2 + S2 - (S1 + Si)~ 0.
Consequently, since e = ~.
S2-S1 f
~ e- 1 dQ.
Our discussion can be summarized by the postulation of another law of
nature.
SECOND LAW OF EQUILIBRIUM THERMODYNAMICS
Fmt Version. Suppose that an isolated system changes from a homo-
geneous state with entropy S 1 to a homogeneous state with entropy S 2
Then S2 ~ S1 , with equality holding only for a reversible change.
See, for example, B. Gal-or, "The Crisis About the Origin of Irreversibility and Time
Anisotropy," Science 176, 11-17 (1972).
t No heat fl.ows in the absence of a temperature difference. Thus it would be more accurate
to speak of the limit as the heat conductivity of the surroundings increases without bound, so that
a vanishingly small temperature difference between system and surroundings is required to
generate dQ.
Appendix 14.2] Some Historical Remarks 501
Second Version. Suppose that a nonisolated system changes slowly from a

homogeneous state with entropy S1 to a homogeneous state with entropy S2
(The change must be slow enough so that at any moment the system is
characterized by a single temperature 0.) Then
S2 - S 1 ;;;:: fe- 1
dQ,
with equality holding only for a reversible change.
As for the connection between entropy and the availability of energy for
work, we mention a revealing deduction by Lord Kelvin. Kelvin considered
processes that operate among various heat reservoirs at temperatures 0 0 ,
0 1 , 0 2 , . , where 0 0 ~ 0 1 ~ 0 2 He showed that such a process
causes an amount of energy to become unavailable for further work that is
equal in magnitude to 0 0 times the change in entropy brought about by
the process.
EXERCISES
1. The text discusses entropy changes in reversible and irreversible expansion
experiments that use the insulated piston apparatus depicted in Figure
14.14. Provide a similar discussion for the corresponding compression
experiments.
2. If S = S(E, V), deduce the following equation from the facts (i) that
dS = e- 1p dV + e- 1 dE is an exact differential, and (ii) that the order
of second partial derhatives is immaterial (if these derivatives are con-
tinuous):
ae oP ae
-+0--P-=0 (V, E independent).
av oE oE
(Similar relations are valid when other variables are regarded as inde-
pendent. Derivation of some of these provides good practice in the
technique of partial differentiation. See Exercise 1 of Appendix 13.1.)
Appendix 14.2 Some Historical Remarks

In the next paragraphs we shall sketch very briefly some points in the early
historical development of continuum mechanics so that the reader may begin
to see that it was developed only after years of effort by brilliant men. In
doing so, we lean heavily on a characteristically penetrating and eloquent
article by Clifford Truesdell,* to which we refer the reader for much more
detail and also for assessment of the role of experiment in the development
of mechanics.
"A Program Toward Rediscovering the Rational Mechanics of the Age of Reason,"
Arch. Hist. Exact Sci. J(l), 3-36 (1'960).
502
That Newton's laws are only the vital first step in the development of
continuum mechanics is apparent from the fact that Newton himself could
not solve a number of the major problems of the subject. Indeed, many of
these problems are not solved today.
For example, Newton tried to find the force on a body immersed in flowing
air by considering the motion of an elastic solid moving through a lattice of
equally spaced elastic particles. The resulting theory gave a resistance pro-
portional to the density of the air, the cross-section area of the body, and
the square of the speed. It accounts poorly for the force on the front of the
body and not at all for the relatively low pressures at the rear that are re-
sponsible for the major part of the resistance. Late in the nineteenth century,
Newton's theory was used to "prove" that airplane flight was impossible
because of the large resistance that would be encountered. Its influence
lingered into the twentieth century in wind pressure calculations required
by building codes.* This is not surprising, since it was work at the beginning
of the twentieth century by Prandtl that opened the way for the good under-
standing of the resistance of streamlined bodies that we have today. Even
this understanding is incomplete, while bluff body resistance remains largely
an unsolved problem. (See Chapter 3 in II.)
Newton did not try to ascertain the resistance of a body by attempting
to find an approximate solution to certain partial differential equations that
followed from his laws. Instead he tried to solve the problem by means of
various particular assumptions and special insights. This is typical. Only
after years of experience with particular problems did it become clear which
assumptions were valid and which invalid. More study was necessary before
it was possible to say which of the valid assumptions were fundamental
and which could be deduced from fundamental assumptions. Most difficult
of all was the distillation of "ways of looking at things" into clearly defined
principles.
An instance of a "failure" to recognize what later turned out to be a vital
general principle is provided by some work on the vibrating string by the
English mathematician Brook Taylor. In 1713 Taylor applied Newton's
law to a differential element of the string, but he did not see that equations
of motion can be derived by applying Newton's second law to each infini-
tesimal part of a body. "He did not recognize the result as a differential
equation of motion, and his (further) work rests on confusing and partially
erroneous assumptions" (Truesdell).
The first explicit derivation of an equation of motion by application of
Newton's laws to an infinitesimal element appeared in John Bernoulli's
Hydraulics (1739). In this treatment of the one-dimensional motion of in-
compressible inviscid fluid, Bernoulli introduced as a dependent variable
P. F. Nemenyi, "The Main Concepts and Ideas of Fluid Dynamics in Their Historical
Development," Arch. Hist. Exact Sci., 2(1), 52-86 (1960).
Appendix 14.2] Some Historical Remarks
503
what he called internal force and what we would now call pressure. The pres-
sure exerted by a wall or by a column of liquid in hydrostatics had been used
before. Nevertheless, although he explained it badly and hardly exploited it,
John Bernoulli made a major conceptual advance by introducing the unknown
internal force.
We come now to the titan Leonard Euler (1707-1783), who "calculated
without apparent effort, as men breathe, or as eagles sustain themselves in
the wind" (Arago). Euler's works, comprising numerous volumes, codified
great quantities of earlier material and made profound contributions to
every kind of mathematics known in his time. "His notation is almost
modern-or perhaps we had better say that our notation is almost Euler's"
(Struik). Even total blindness during the last 17 years of his life seemed only
to heighten the quality of Euler's work.*
We can only touch upon a few of Euler's most important discoveries.
Two vital contributions appeared during the 1750s. First, a sharpened
appreciation of the difference between kinematics and dynamics led Euler
to an understanding of mass conservation. Second, he applied John Ber-
noulli's barely exploited concept of internal force for one-dimensional flows
to fl.owin pumps and turbines, and then generalized to our modem notion
of the internal normal pressure in three-dimensional fl.owof inviscid fluid.
For years Euler tried to reconcile the theory of flexible lines with that of
elastic bands. The former uses equilibrium of forces, the latter equilibrium
of moments. In 1771 he saw that a shear as well as a tension is necessary to
account for the force exerted on an isolated element of an elastic line. This
was the first use of a general stress vector.
It appears that Euler was led to the postulation of the conservation of
moment of momentum by the work just mentioned and by work in 1744 on
many linked bars in a plane. Thus decades of effort culminated in 1776 with
the publication of Euler's laws of mechanics.
Although Euler introduced an unknown stress vector in his study of
elasticity, the vital general concept of stress vector was not introduced for
56 more years. This was accomplished by no less a person than Cauchy, who
also did fundamental work in the theory of complex functions, formulated
the calculus as we now learn it, and made a number of important early
investigations of finite groups.
Even the few paragraphs we have devoted to some historical points suffice
to show that the development of continuum mechanics required years of
work by men of the highest ability. One more point remains to be made.
It is that although the principles of continuum mechanics can now be reduced
to a few lines, thanks to a distillation by men of genius, nevertheless, secure
understanding requires much study.
* But even geniuses are not flawless. For example, Newton and Euler made an incorrect
physical assumption when they tried to calculate the speed of sound. See Section 15.3.
By contrast, calculational methods are an example of ingenious discoveries

that are readily appreciated. For instance, consider Euler's discovery that
linear ordinary differential equations with constant coefficients have expo-
nential solutions. In a few hours an able college sophomore can learn all
there is to know about this matter. It is usually otherwise with a new physical
concept. To a student, at first the new way of looking at things often seems
simple and natural. Experience indicates, however, that if an individual is to
gain a thorough understanding of a hard-won concept he must personally
experience some of the painful intellectual struggle that led to it. A lecture
or a text can do no more than point the way.
The phrase "intuitively obvious" therefore does not belong in a discussion
of the basic principles of continuum mechanics, or in a discussion of the basic
principles of any branch of modern science. A more appropriate watchword
is that found on one of the humorous cards that are sometimes displayed in
officesand laboratories: "Anyone who remains calm amidst all this confusion
simply does not understand what is going on."
CHAPTER 15
InviscidFluid Flow
A WAS discussed in Section 14.5, the field equations describe properties,

such as conservation of mass, which are common to all continua, whereas
constitutive equations describe the qualities that distinguish one particular
kind of continuum from another. We begin this chapter with a discussion
of one of the simplest constitutive equations, the one for an inviscid or
perfect fluid. (As we shall see, this equation holds for all fluids when there
is no motion.) We then turn to the formulation and solution of some specific
problems and thus begin to understand the power and defects of the inviscid
fluid model.
The first specific class of problems treated deals with the fate of small
disturbances to a quiescent horizontal fluid layer whose density varies with
height. The important concept of stability is introduced, the technique of
linearization is used, and the resulting eigenvalue problems are examined for
qualitative behavior or are solved in special cases. We see that the inviscid
model provides reasonable predictions for situations of physical interest.
We next consider compression waves in gases. Our discussion touches on
small-amplitude sound waves and large-amplitude shock waves. We encounter
solutions to the linearized wave equation which propagate without changing
shape, and we examine deforming simple wave solutions to the appropriate
system of nonlinear partial differential equations. Solutions of the latter type
develop such rapid variation that virtually discontinuous shock waves result.
All these predicted flows are well verified experimentally.
Finally, we employ separation of variables to find a two-dimensional flow
of a uniform stream past a circular cylinder. This flow does not seem in accord
with experience, for a calculation indicates that the net drag force on the
cylinder is zero. Further investigation shows that the same result holds for
the flow of a uniform stream past quite general bodies. For this problem, then,
the inviscid model does not seem satisfactory. A discussion of how viscous
forces cause drag is a central motif of Chapter 3 in II.
15.1 Stressin MotionlessandInviscidFluids

The central object of the present section is to characterize the stress vector
in a motionless fluid and to give evidence that the result obtained could also
be useful for a class of fluids in motion. A molecular model will provide the
basis for our initial inference.
505
506 Inviscid Fluid Flow [Ch. 15
MOLECULAR POINT OF VIEW

Consider an assemblage of many tiny mass points moving rapidly about
with no preferred direction and colliding elastically. This is a reasonable and
familiar "perfect gas" molecular model of a gas at rest. (Molecules have a
preferred direction in a moving gas.) From this model, as in elementary
physics texts, we deduce that the molecular bombardment of a surface element
is responsible for a purely normal stress on that element. The tangential force
components add to zero because there is no preferred direction. For the same
reason, the magnitude of the pressure (normal force per unit area) on an
element should be independent of the element's orientation. Thus we assume
that the stress vector t satisfies
t(x, t, n) = -p(x, t)n, (1)
where p is the pressure.

The elastic mass point molecular model of the previous paragraph is
obviously inappropriate for liquids. A summary of progress in the study of
more appropriate molecular models can be found, for example, in Seeger
and Temple (1965). But without delving into the kinetic theory of liquids,
we can guess that some sort of jostling will replace the elastic collisions of
the perfect gas. There will still be no pref erred direction so that (1) can be
expected to hold in a motionless liquid.
Although solids are not the object of our present investigation, we should
perhaps mention the very different behavior enforced by their molecular struc-
ture. Molecules that are normally constrained to more or less preferred
locations can generally adjust to any time-independent force by a fixed dis-
placement, with a concommitant alteration of the various molecular inter-
actions. Consequently, just because the solid is motionless, there is no reason
to expect a simple stress distribution within it.
It is also appropriate here to recall that many materials are not easily
classified as "solids," "liquids," or "gases." An example is the children's
toy Silly Putty, which bounces if a ball of it is thrown on the floor but slowly
flows into the shape of its container if it is left undisturbed. We shall not
concern ourselves here with such borderline cases. We shall use the words
"solid, liquid, gas" according to their meaning in ordinary discourse. (Thus
a "fluid" is either a liquid or a gas.) Our constitutive relations will apply to
typical representatives of each class of material: the solid steel, the liquid
water, and the gas air.
CONTINUUM POINT OF VIEW

Let us reconsider stress in a motionless fluid, this time using common
experience of fluid as a continuum. In doing so, we shall find it useful to
distinguish between shear stresses that act on an element in a direction
perpendicular to its normal n and normal stressesthat act along n.
Sec. 15.JJ Stress in Motionless and Inviscid Fluids
501
It is noteworthy that shear stresses are particularly apparent in liquids like
cold molasses, which we call "very viscous." As an illustration, think of how
difficult it is to slice through cold molasses with a knife. The edge of the knife
blade is so thin that this resistance must be due to shear stresses acting on the
flat part of the blade (Figure IS.I). The faster you try to slice, the harder it is.
K K
I J JI
I J jl
(a) (b)
FIGURE 15.1. Knife K slicing downward through cold molasses.

(a) S denotes the resisting shear stress exerted on the knife by the fluid.
(b) Arrows indicate typical fluid velocities.
If you tum the knife sideways and try to "spread" the molasses, the faster you
want to make one layer of molasses flow over another, the harder you have to
push (Figure 15.2). These -observations could be summed up in the qualitative
graph of Figure 15.3. Note that small relative speeds are associated with small
shear stresses. One expects that the shear stresses will vanish for a motionless
fluid, wherein relative velocities of adjacent material are always zero.
We are led to assume that/or a motionless fluid, stresses are purely normal.
That is, t must be in the direction of n, so
t(x, t, n) = - p(x, t, n)n, (2)
for a scalar function p. The present continuum argument thus leads to an as-
sumption that is the same as (I) except that at the moment we cannot rule out
- K
FIGURE 15.2. Knife K moving to the right and spreading molasses.

Arrows in the fluid indicate typical velocity vectors.
1'
'--~~~~~~~~--dv
FIGURE 15.3. Qualitative plot, deduced from ordinary experience, of

the shear stress .,.versus dv, the relative velocity of adjacent layers of fluid.
a dependence of p on n. But from the relation between the stress vector and
the stress tensor
3
ti(x, t, n) = L n1TIJ(x, t)
1=1
we find using (2) that

3
-p(x, t, n)n1 = I: n1T 1i(x, t),
1=1
which implies that

(3)
where ~11is the Kronecker delta. Since the left side of (3) is independent of n,
it must be that p(x, t, n) is actually independent of n. Thus we find from a
continuum viewpoint that the pressure p on an element of area located at
point x must be independent of the normal vector n-just as we inferred
above from the lack of a preferred direction of molecular motion.
N o T E Tiny gas bubbles in carbonated beverages are spherical. This
confirms our view that the magnitude of the pressure in a motionless fluid
is independent of the orientation of the surface element over which the stress
acts.
HYDROSTATICS
Since p does not depend on n, the constitutive equation for a motionless
fluid takes the final form
T1i(x, t) ==-p(x, t)~11 (4)
For a motionless fluid we, of course, have
v=O. (5)
Sec. 15.JJ Stress in Motionless and Inviscid Fluids
With (4) and (5), Equation (14.2.19) for linear momentum conservation
becomes simply
pf-Vp=O, (6)
which is the hydrostatic equation. It is easily seen that unless there are heat
effects, all the other field equations are automatically satisfied.
As a simple application of hydrostatics let us check that our results are
compatible with the fact that one never sees a block of fluid resting motionless
on a solid surface. Suppose that there were such a motionless block of fluid, of
height H and width W, and with a coordinate origin located in the center of
its base (Figure 15.4).We attempt to solve (6), where the body force is due to
gravity,
f= -gk, (7)
and the fluid is subject to a uniform atmospheric pressure p,.:
p=p. when Os; z s; H, x = tW (on the sides); (Sa)
p=p. when z = H, -tW s; x S tW (on the top). (Sb)
FIG URE 15.4. Postulated stationary block of fluid resting on a solid.
As is customary, when solving specific problems we use x, y, and z as spatial

coordinates rather than the more cumbersome x1, x 2 , and x 3
Ifwe use (7), (6) becomes
-py = 0, -pg- Pz = 0. (9)
Equations (9) and (8b) require that
p = -pg(z - H) +Pa,
but we are unable to satisfy (Sa) with this pressure. As expected, our equations
show that Figure 15.4 cannot depict a fluid at rest. On the other hand, the
hydrostatic equations and boundary conditions would be satisfied if a wall
were present at x = tW, in which case (Sa) should be deleted.
InviscidFluidFlow [Ch. 15
510
Suppose that a block of material is placed on a table and does not deform.
Our argument would seem to allow us to rule out the hypothesis that the
block is a fluid. But it could be that the block would undergo fluid-like
deformation if observed for a longer period. We should say that" the block
does not behave like a fluid when subjected to gravity-like forces over a
period of hours." The more precise terminology reminds us that the ab-
straction " fluid" is useful for understanding a certain range of phenomena.
It diminishes our surprise on learning that "solid rock" can usefully be
regarded as fluid in the study of deformations that occur over geological time
scales.*
INVISCID FLUIDS
Our discussion of slicing cold molasses associated large shear stresses with
what are commonly called "fluids of high viscosity." On the other hand, it is
easy to slice a knife through water or air. It therefore appears reasonable to
assume, as a first approximation at any rate, that shear stresses are always
zero in these important fluids. Thus we define a class of inviscid ftuids, for
which, even when they are in motion, the stress tensor satisfies the constitutive
equation(4). The equation for conservation oflinear momentum (14.2.19)now
becomes the famous Euler momentUJbequation for inviscid ftow:
(10)
The remainder of this chapter contains some examples that show successes
and failures of the inviscidt assumption. We shall use as governing equations
the Euler momentum equation just derived, an equation for conservation of
mass, and (in our study of compression waves) appropriate equations of
In a note in Physics Today (Vol. 17, p. 62, 1964) M. Reiner ascribes the knowledge that
rock flows to the Prophetess Deborah. "In her famous song after the victory over the Philistines,
she sang, 'The mountains fl.owed before the Lord.' When, over 300 years ago, the Bible was
translated into English, the translators ... translated the passage as 'The mountains melted
before the Lord '-and so it stands in the authorized version. But Deborah knew two things.
First, that the mountains flow, as everthing flows. But second, that they fl.owed before the
Lord, and not before man, for the simple reason that man in his short lifetime cannot sec them
tiowing, while the time of observation of God is infinite. We may therefore well define as a
nondimensional number the Deborah number
D = time of relaxation
time of observation
The difference between solids and fluids is then defined by the magnitude of D.''
From a scientific point of view, then, one should not ask whether a given piece of material is a
solid or a fluid. If the Deborah number is large, the material should be regarded as a solid. Other-
wise it should be regarded as a fluid.
t The term "perfect" used to be applied to fluids that satisfy (10). This term perhaps reflects
an optimistic nineteenth-century view that only failures of its Maker could cause nature to
deviate from behavior that is relativelr accessible to mathematical analysis.
Sec. 15.1) Stress in Motionless and Inviscid Fluids 511

state. We shall require that no fluid penetrate the boundary. But in the
absence of shear stresses, we have no reason to postulate any restriction on
how our inviscid fluid can slip along the boundary.
EXERCISES
tl. Show that the magnitude of the net pressure force on a body that is
completely submerged in a liquid is equal to the weight of the liquid
displaced by the body. Show that the direction of the force is upward
(i.e., opposite to the direction of gravity).
2. Assuming that gravity acts in the direction opposite to that of the
z axis, derive in detail from first principles the hydrostatic equation
dp
dz+ pg=O.
Do this by equating to zero the net vertical forces on a horizontal fluid

slice.
The next two exercises deal with the useful concept of the stream
function.
3. (a) Consider two-dimensional steady fluid flow. Show that the

continuity equation implies that V /\ ( - pvi + puj) = 0. Thus there
exists a function 1/1such that pu = 1/1,, pv = - 1/1" Why'? Show
that the same relations hold for unsteady flow when p = constant.
(b) Justify the name .. stream function" for 1/1by showing that the
curves 1/J= constant are streamlines. (Thus determination of 1/1is
tantamount to obtaining streamlines for the flow: also pu and pv
can be found merely by differentiation.)
t4. (a) Let C be a curve joining a pair of streamlines 1/1= v,0 and 1/1= 1/1
1
of a two-dimensional flow. (See Figure 15.5.). Let the arc length s
FIG URE 15.5. Two streamlines if,= Y,1and if,= if,o, Computation of
an integral along an arbitrary curve C (with unit normal n) provides a
physical interpretation for the difference Y,1
- Y,o,
of C have the values s0 and s 1 when C intersects these two curves,

s 1 > s0 Show that
St
Jpv n ds
so
= V/1- 'Po,
where v is the velocity vector and
Interpret physically.
(b) Consider two-dimensional motion of an inviscid fluid of constant
density. Let a conservative body force f be given in terms of a
potential <l>by f = V<l>.By taking the curl of the momentum
equation, show that
o(v2 .;,) + o(v2 .;,, 1")= o,

ot o(x, y)
(The three components of a vector equation are thus replaced

by a single scalar equation.)
is. A barotropicfluid is defined to be a fluid in which the pressure p depends
only on the density: p = p(p). (An example is provided by an ideal gas in
isentropic flow.) In such a fluid show that V" (p- 1Vp) = 0. IfV "q =
0, then q = Vr for some r. Advanced calculus books give methods for
determining r. Use such a method to determine the scalar function
of which p- 1vp is the gradient in a barotropic fluid.
6. This exercise shows that a velocity potential continues to exist for a
portion of inviscid barotropic fluid subject to a conservative body force
if it exists initially for that portion of fluid. This justifies the assumed
existence of the velocity potential in many problems, in particular
those in which the motion can be regarded as commencing from a
motionless state. (Another derivation of this important result, using
tensor calculus, is requested in Exercise 2.3.10 in II.)
(a) The starting point is the following equation of motion for an
inviscid fluid in material variablesthat was obtained in Exercise
14.2.7:
k = 1, 2, 3.
Suppose that external forces are conservative, so that the force

vector is the gradient of some scalar potential '2 (Fi= -0'2/oA).
Suppose further that the fluid is barotropic [p = p(p)]. By inte-
Sec. 15.1] Stress in Motionless and Inviscid Fluids 513
grating the above equations with respect to time between zero

and t, show that they can be written
k = 1, 2, 3,
where
x= {[ft+ n-!(Vi + v~+ v~)]dt,
To do this, use integration by parts and show that
I
ox" -1
oA",=o-
directly from the definition of the partial derivative.
(b) Suppose a velocity potential exists at t = 0 for a finite portion of
the fluid:
3
dq,0 = L ~<o>dX;; v(O)
r;
_
-
vi
r; t=O
i=l
Derive
and thereby show that a velocity potential exists for the same
portion of fluid at all other times.
7. This exercise shows that under certain circumstances the momentum
equation (10) for inviscid flows can be integrated, yielding the Bernoulli
equation. As later exercises indicate, the Bernoulli equation provides
a very useful relation between flow speed and pressure.
(a) Verify the identity
by comparing the components of each side of the equation.

(This result is Theorem 7 of Section 2.3 in II. There a more elegant
proof is furnished by the methods of tensor calculus.)
(b) A fluid motion satisfying V " v = 0 is termed irrotational.(A
discussion at the beginning of Section 3.1 of II shows why the
term is appropriate.) For irrotational fluids, v = Ve/>for some
velocity potential cf,.Also, if the force field F is conservative,
then F = -Vfor a force potential .
For the irrotational motion of an inviscid barotropic fluid

subject to a conservative body force, deduce the unsteady Bernoulli
equation:
~~ + ~ lvl2 + Jp- 1 dp +fl>= a function of time.
No TE. An important special case is steady flow of fluid

with uniform density, in which case the Bernoulli equation clearly
becomes
!2 Iv12 + fp +fl>= constant.

is. An inviscid fluid of uniform density is in steady irrotational flow along
a horizontal pipe of variable cross section. Assume that the speed is
uniform across any given cross section. If .tip is the pressure difference
indicated by a differential pressure gauge connected at two locations
having cross sections s1 and s2 , find a formula for the mass m of liquid
flowing per second. The answer is the square root of a certain quantity.
Can this quantity be negative so that the mass flux is imaginary?
9. Consider some water in a tin can (right circular cylinder) rotating with
constant angular velocity c.o.The velocity field for a fluid rotating as a
solid body is
u = -c.oy, V = <.OX, w=O.
(This is in ordinary cartesian coordinates-not polar coordinates.)
The above velocity field satisfies the boundary conditions on the tin
x ....-~-----------,
\
\
\
\
\
~-_._
I
!
I
___
FIGURE 15.6. Vessel in the form of a figure of revolution. At each

value of y, the cross section is a circle of radius rand area S= 7rr 2 Liquid
flows out through a small hole in the bottom. The vessel is to be designed
so that its surf ace descends at a constant speed V.
Sec. 15.2] Stability of a Stratified Fluid
515
can. Show that the inviscid incompressible equations of motion, with
gravity, are also satisfied. Find the form of the free surface, assuming
that the pressure at the fluid surface equals the air pressure A (A a
constant).
UO. An axisymmetric hourglass is to be designed so that the surface of the
incompressible inviscid liquid it contains descends at a constant rate V.
(See Figure 15.6.) Show that y "'x 4 and find the proportionality
constant.
15.2 Stability of a Stratified Fluid

Slime mold amebae, after exhaustion of their food supply, suddenly begin
to aggregate into more or less regularly spaced collection points. Fluid flowing
smoothly through a pipe develops violent random eddies if the driving
pressure gradient is increased too far. A metallic structure collapses if a slight
additional load is placed on it. All these phenomena and many more can be
regarded as exemplifying an instability of equilibrium. By a state of equilibrium
we mean a steady state that is permitted by the governing equations of the
problem. We say that a system is perturbed from equilibrium if it is put in a
state that differs somewhat from the equilibrium state. The perturbation is
what must be added to the equilibrium state to achieve the perturbed state.
If an initial perturbation of equilibrium becomes larger in some sense as the
system evolves, we say that the equilibriumstate is unstable to that perturbation.
An instability problem of biological interest was briefly discussed in
Section 1.3. It will help in motivation if that discussion has at least been
skimmed. More important, before proceeding the reader should be familiar
with the pendulum instability calculations presented (independently of
surrounding material) in Section 11.1, for here we shall examine material
that is related to, but is more complicated than, stability theory for the
pendulum-namely, the study of the stability of bottom-heavy and top-heavy
layers of fluid.
The simple pendulum is a mechanical system with but a single degree of
freedom. Its state can be described in terms of a single variable, the angular
deviation of the rod from a reference value. Thus the evolution of the pen-
dulum in time is governed by a single ordinary differential equation. By
contrast, a layer of fluid can be said to have an infinite number of degrees of
freedom. In the present case, description of the state of the fluid at a given
moment requires specification at every point in the layer of three velocity
components, a pressure, and a density. The evolution of the system in time
is governed by a set of partial differential equations. But in spite of these
profound differences, one suspects (correctly) that there are many similarities
in the behavior of a top-heavy fluid and an inverted pendulum, and in the
behavior of a bottom-heavy fluid and a pendulum in "normal" position.
The motion of a pendulum is studied so carefully not because of its direct

technical importance, but because it provides a relatively familiar and readily
understandable prototype of many physical situations. Similarly, the inviscid
fluid motion to be studied below, although of no small interest in itself,
acquires considerable added importance when it is regarded as a prototype
of the natural phenomena that are manifestations of instability in continuously
distributed systems.
An extreme case of the sort of problem that we shall consider is a situation
that involves two immiscible fluids, say oil and water. Each of the two fluids
separately can be regarded as possessing a uniform density, but this is not
true of the combinationof fluids.During the courseof motion, the density
at a particular point in space might first be that of water, then later that of
oil, again that of water, etc. In spite of this, the density associated with a given
small packet of fluid particles remains constant as that packet moves about.
Generalizing, we shall consider in this section a fluid that is incompressible
but not of uniform density.
We shall see that if gravity acts vertically downward, then in principle it is
possible for a horizontal layer of incompressible fluid to remain motionless
for any distribution of density that depends solely on the vertical coordinate z.
Hence we speak in the title of this section of a stratified fluid, a fluid "formed
in layers" (of different densities). Assuming that the fluid is inviscid, we
turn to an investigation of the stability of a stratified horizontal layer. We
shall obtain some general results for classes of stratified layers and some
particular results for special cases. The first goal of our analysis is to obtain
a set of linear partial differential equations for the perturbations in velocity
and density. Since the equilibrium solution depends only on the vertical
coordinate z, we can then reduce the problem to a single ordinary differential
equation of a function of z.
GOVERNING EQUATIONS AND THEIR EXACT EQUILIBRIUM

SOLUTION
Consider, then, an incompressible inviscid fluid of varying density that is
contained between two rigid horizontal planes located at z* = 0 and z* = d.
(Here z* is the vertical coordinate; x* and y* will be the horizontal coordi-
nates.) The governing equations are the following:
Conservation of mass:
Dp*
Dt* + p*Vv* = 0. (la)
Balance of linear momentum for inviscid fluid:

Dv*
p* Dt* = -Vp* - p*gk. (lb)
Sec. 15.2] Stability of a Stratified Fluid 517
Incompressibility:
Dp*
-=0. (le)
Dt*
No penetration of horizontal planes:
v* k = 0 for z* = 0, d. (ld)
All variables have their conventional meaning, except that an asterisk has
been added to emphasize that variables are not dimensionless. As usual, k is
the "vertical" coordinate vector and g is the acceleration due to gravity,
so - p*gk is the correct term to represent a constant gravitational force that
acts vertically downward.
We introduce dimensionless variables in order to reduce the numbers of
parameters that occur explicitly. (See Section 6.2 for a discussion of dimen-
sional analysis.) At our disposal we have the length d and the gravitational
acceleration g (dimension: length/time 2 ). In order to be able to obtain a
quantity with the dimensions of mass, we employ a "typical" density R 0
(dimension: mass/length 3). As there is as yet no obvious way to assign a
definite value to the constant R 0 , we shall leave it unspecified for the present.
We select combinations of d, g, and R0 so that variables without asterisks,
defined as follows, are dimensionlesst:
x* t* p* v* p*
x = d, t = (d/g)112, p = Ro' v = (dg)112, p = Ro dg.
If we introduce these variables and also use (le) to modify (la), we obtain
Dv
Vv=O, p Dt = -Vp - pk, (2a, b)
Dp =0 w = 0 for z = 0, 1. (2c, d)
Dt '
Here w is the vertical component of velocity, and the operators V and DJDt
are to be interpreted using the new (unstarred) variables. The equations of (2)
are essentially the same as those of (I), except that in the nondimensional
variables of (2) the distance between the bounding planes and the gravita-
tional accelerationboth havethe valueunity.
Does a motionless layer correspond to a possible solution of (2)? With
v = 0, (2a) and (2d) are identically satisfied, while the remaining equations
become
P:x:= 0, P1 = 0, P::= - P, Pi = 0, (3a, b, c, d)
t Distances are compared to the width d of the layer. A factor of 2 1' 2 is omitted because it
clutters the equations, but otherwise times are compared to the time (2d/g)112 that it takes a freely
falling particle to traverse the layer, starting from rest. Similarly, velocities are essentially referred
to the terminal speed (2dg)1 ' 2 attained by a particle falling a distance d, starting from rest.
where the subscripts denote partial differentiation. Equations (3a) and (3b)
require that p be independent of x and y, while (3d) requires that p be indepen-
dent oft. With this information, (3c) becomes
pz(z, t) = - p(x, y, z), (4)
where we have explicitly indicated the remaining possible independent
variables in p and Pz. If for any times t1 and t 2 we have
pz(z, t 1) -::!-p.,(z, t2),
then there is a contradiction with (4). Consequently, P: cannot vary as t

varies. Similarly, p cannot vary as x and y vary. Thus Pz and p can depend
only on z.
Let
p = R(z), (5)
where R(z) is an arbitrary function of z. Then from (4)
p(z) = - fR(z) dz. (6)
The constant of integration (which can depend on time) that is implicit in the
indefinite integral of (6) can be used to make p(z) = 0 at some height z = z0
This merely means that p(z) is measured with respect to a zero value at the
reference level z0 Note that changing the reference level means changing the
definition of p by adding a constant. This does not have an effecton the motion,
since p occurs only in the form Vp. Note also that there now appears a
number of sensible ways to define the reference density R0 , e.g., R 0 = RH)
(the density at the center of the layer) or R0 = JAR(z) dz (the average density).
We have shown that
V::0, p = R(z), p= - fRdz, (7)
is an exact solution of the governing equations (2) for any (integrable)

function R(z). We shall now examine the stability of this solution to small
perturbations in a manner analogous to our study of the stability of the
equilibrium states of the pendulum.
LINEARIZED EQUATIONS FOR THE PERTURBATIONS
We first define perturbations v', p', and p' to the exact solution (7):
v = 0 + v', p = R + p', p= - fR dz + p'. (8)
On substitution of (8) into the governing equations (2), we obtain a set of

nonlinear partial differential equations for the perturbation quantities. For
example, the incompressibility condition (2c) becomes
0 = (R + p'), + u'(R + p')x + v'(R + p'), + w'(R + p')z
(9)
Sec. 15.2) Stability of a Stratified Fluid
519
As with the pendulum, we should get significant information even if we

limit ourselves to perturbations so small that an expression containing the
product of two perturbation terms can be neglected as insignificant compared
with expressions containing as a factor only one perturbation term. It should
be clearly understood that we restrict ourselves to small perturbations not
because only these perturbations are physically significant, but because only
with this restriction can we obtain a relatively simple linear problem to solve.
Application of the linearizationprocedure(deletion of nonlinear terms)
to (9) yields
, dR ,
p,+ dz w =0, (10)
while application to the momentum balance condition (2b) gives the three
scalarequations
Ru;= -p;, Rw; = - p; - p'. (lla, b, c)

The continuity equation (2a), without approximation, implies the linear
equation
u;+ v;+ w;= 0, (12)

and boundary condition (2d) requires
w'=O for z = 0, l. (13)
Equations {10}-(13) comprise a linear problem for the perturbations. There

are five partial differential equations for the five unknown functions u', v', w',
p', and p'.
In the set of linear partial differential equations (10), (11), and (12), the
independent variables x, y, and t appear only in the form of derivatives. To
put it another way, as far as these variables are concerned, the partial
differential equations have constant coefficients. Extending the procedure
Those familiar with scaling (Section 6,3) can better appreciate the linearization. In (9), for
example, terms like u'p~ and r/ p; can be neglected while a tenn like p', is retained if all dimension-
less perturbation terms are small compared to unity. But the nondimensionalization process we
used was based on scales comparable to those which describe free fall of a particle. Thus, for
example, the dimensionless vertical velocity component v' will be small compared to unity if the
actual velocity in the perturbed flow is small compared to the terminal velocity in free fall. This
and similar arguments give an idea of what is meant by a "small perturbation."
If we wished to utilize the scaling technique in full, we should (for example) have chosen
(dg) 112 a for a velocity scale, 0 < e ~ 1. This would explicitly indicate that we are considering a
situation in which the maximum expected speed is small compared to the terminal speed of free
fall. Similar introduction of a small parameter e into the scales for the other dependent variables
would lead to the appearance of e before all nonlinear terms in the governing equations. The
stage would then be set for a perturbation analysis, the first equations of which would form the
linearized problem that we derived less formally in the main body of the text.
Inviscid Fluid Flow [Ch. 15
520
used for linear ordinary differential equations with constant coefficients, we

can find solutions in which the dependence on each of the variables x, y, and t
is exponential-again, on substitution each term will have an exponential
factor that can ultimately be canceled.
The previous paragraph partially motivates our search for solutions of the
form
u'(x, y, z, t)l [u(z)l

v'(x, y, z, t) a(z)
w'(x, y, z, t) = lV(Z) ei(k1x+k21)+crt. (14)
[ p'(x, y, Z, t) {J(z)
p'(x, y, z, t) ft(z)
At this point it may seem that we are only considering a very small class of
perturbations. Even if true, however, for these perturbations our equations
reduce to ordinary differential equations, which makes their solution relatively
easy.* A more telling point, which we shall discuss below, is that our
special perturbations can be "added" to give a rather general perturbation.
We remark that as is always the case for linear problems, we can satisfy
the requirement that velocity, pressure, and density be real by taking as our
final solutions either the real or the imaginary part of the complex solution
(14).
In (14), <1may be complex, but we take k 1 and k 2 to be real. If either k 1 or k 2
has a nonzero imaginary part, then the perturbations will grow without
bound as x--. oo or as y--. oo. For example, if
then
if k~) < 0,
Iexp ik. 1x I =exp(-ki' O x)-+oo {as x --. + oo
asx-+-oo if k~) > 0.
The restriction to real values of k 1 and k 2 limits the set of perturbations
under discussion, but this set seems to contain enough elements to synthesize
all perturbations of interest. It might be helpful to think of the contrasting
case of disturbances generated at the entrance region of a long pipe. Here
'' interesting perturbations " are bounded in time but grow in space.t
Having established the form of perturbation to be considered we can
proceed with the calculations. Upon substituting (14) into (10), (11), and (12),
Investing a great deal of work to obtain a special solution might be unwise, but if only a
relatively small amount of work is necessary, then consideration of a special case is probably
worthwhile. This is a general rule of procedure in applied mathematics.
t In the last analysis, only complete specification of an initial boundary value problem can
determine the class of perturbations that should be considered-but a more informal approach
is appropriate in preliminary investigations such as the present one.
and canceling the common exponential factor, we obtain the linear ordinary
differential equations
up+ dR
A
dz -~ O
w = , (15)
RulV = - : - {}, (16a, b, c)
ik1u+ik2e+
':: =O, (17)
and the boundary conditions

(0) = lV{l) = 0. (18)
CHARACTERIZATION OF THE GROWTH RATE a AS AN

EIGENVALUE
Since the boundary conditions apply to w(z), we try to eliminate the other
variables. This can be done as follows.* Solving (16a) and (16b) for ik1u and
ik 2 f) in terms of faand then substituting into (17), one obtains
dlti
P = -Ruk- 2
dz (19)
where
(20)
One can then use (15) and (19) to write (16c) entirely in terms of w.To
simplify our notation, we write W = w.We finally obtaint
! (Rd:)+ k (a-!- R)w= o,

2 2 (21a)
W(O)= W(l) = 0. (21b)
We note at once from (21) that k1 and k 2 appear only in the combination
k 2
= kf + kf Now the perturbations that we are considering are pro-
portional to
exp [i(k 1x + k 2y)] = exp (ik r)
In a given problem, considerable trial and error might precede the discovery of an efficient
elimination scheme.
t The equations of (21) form a Stunn-Liouville problem. We shall repeat some of the general
discussion of this problem in Section S.2. The repetition serves the dual purposes of illustrating
general ideas in a particular context, and of making this section self-contained.
522
But [Exercise l(a)]
exp [ik (r+ 2; ~)] = exp (ik r).

Thus the perturbations are constant along the lines k r = constant (lines
that are perpendicular to the wave number vector k). They have a spatial
period of 2n/k, where k = (kf + k~)1' 2 = Ik 1- The fact that only k enters (21)
means that the stability behavior depends only on the spatial period of the
perturbations; it does not depend on their orientation. This is to be expected,
because there is no preferred direction in the problem-but keep in mind
that nonlinear interactions, not considered here, may act preferentially
among modes with certain combinations of wave-number vectors.
N o T E. The terms overall wave number or horizontal wave number are
sometimes used for k, the magnitude of k.
There seems little possibility that we can find an explicit formula for the
solution of (21) in terms of the function R(z). We do know, however, that
the linear second order ordinary differential equation (21) has a general
solution of the form
(22)
where w<l) and w<i are linearly independent solutions of (21). (We have
2
explicitly indicated the dependence of the solution on the parameters k 2 and

0'2.)
Application of the boundary conditions (18) gives a pair of homogeneous
linear algebraic equations for the constants C1 and C2 :
C1 w<n(O; k 2 , u 2 )+ C2 W(2)(0; k 2 , u 2 ) = 0,
(23)
c1w<1>c1;k 2 , u 2 ) + c2w<2 io; k 2 , u 2 ) = o.
These equations have the trivial solution C1 = C2 = 0 corresponding to
= =
W 0. This must be the case, for W 0 implies that u = v p p 0 and ===
an identically zero perturbation to an exact solution is certainly permissible.
In order that there be nontrivial solutions to (23), we must have
w<tl(O; k2, u2) w12ico;k2, u2) 1-
I w<oo; k2,0'2) w<2>0;k2, u2) - 0, (24)
or, defining the function F,

F(k 2 , u 2 ) = 0. (25)
The {homogeneous)equation and boundary conditions (21) thus constitute an
eigenvalueproblem that possesses nontrivial solutions if and only if there is a
relation between the parameters k 2 and u2 If k is regarded as given, (25)
implicitly determines u2 . Generally, there are many permitted values of u that
correspond to each k.
523
The consequences for stability are these. Given k 1 and k 2 , the x and y wave
numbers of the perturbation, we determine k by (20). For a given value of k,
values of <12 can in principle be obtained from (25). Suppose that for some
value of k there exists a corresponding u with positive real part:
a<>> 0.
Since
Iexp ut I = exp (a(r)t ), (26)
the perturbation velocity, pressure, and density will increase exponentially

with time. If we can find at least one such growing perturbation to the
motionless layer, then we term the motionless state unstable. If all possible
values of a<>are negative for a certain class of perturbations, we term the
motionless state stable to that class of perturbations. If, for a certain class of
perturbations, all values of a<>satisfy a<>~ 0 and at least one value satisfies
a<>= 0, then we term the motionless state neutrally stable to that class of
perturbations.*
Let us return to the particular problem under study. This will enable us to
illustrate the general discussion of the last four paragraphs and will prepare
the ground for a somewhat more comprehensive discussion at the end of
this section.
QUALITATIVE GENERAL DEDUCTIONS
Given the eigenvalue problem (21), there are two ways of proceeding. The
first is to try to make some general deductions about the behavior of the
eigenvalue a, for classes of initial density distributions R(z), without actually
solving the equation. The second is to choose a particular function R, solve
the problem in detail, and thereby to obtain a complete analysis of the
behavior of u for a particular problem. Let us start with the first approach.
Equation (21a) may have complex solutions. We are going to take as final
answers only the real and imaginary parts of (14) so that there is no physical
reason to require that W itself be real. With the possibility of complex
solutions in mind, then, we multiply (21a) by its complex conjugate W and
integratefrom zero to l :
fw !(Rd:)dz+k ::dz-k{RIWl dz=O.

a- f1w1 2
(27)
2 2 2 2
Unstable perturbations will eventually grow so large that linear theory becomes inconsistent.
Even when disturbance growth is predicted, however, it presumably is always possible to select
for examination initial perturbations which are so small that linearization is valid for a certain
time interval. Thus an instability result derived from linear theory should be meaningful in its
prediction of instability for the motionless state, in spite of the fact that the ultimate condition
of the flow cannot be ascertained by linear theory in this case.
524 Invisci"dFluid Flow [Ch. 15
This is not an obvious thing to do, but, as we shall see, it yields some interesting
information. The manipulations that we are in the process of relating are but
the simplest example of a number of ways which have been developed over
the years for obtaining qualitative information about the solutions to eigen-
value problems such as (21) without actually solving them explicitly.
If we integrate the first term in (27) by parts, we obtain
WR-dW1.z=1- f1-dW(R-dW)
dz.
dz .z=o o dz dz
The boundary conditions (2lb) imply that W(O)= W(l) = 0, so we can
deduce from (27) the equation
(f-
2
s::: Iw 1 dz= s:RI w1 dz+ k-
2 2 2
s:R rdz.
ldd: (28)
Deduction1. If dR/dz > 0 for O::;;z;::;;1 (density continually increasing

with height), then every value of u 2 satisfies q 2 = P for some positive quantity
P. Hence
u= JP. (29)
This follows at once from (28), since k is real, once we make explicit the
physical requirement that density be a positive quantity:
R(z) > 0, 0 ::;;z ::;;l. (30)
From (29) we see that eigenvalues of (21), if they exist, occur in pairs, one
positive and one negative. In treatises on ordinary differential equations it is
shown that an infinite number of eigenvalues do, in fact, exist provided that R
satisfies certain conditions. It is sufficient,for example, that R satisfy (30) and
also possess two continuous derivatives for O ;::;;z ::;;1. Equilibrium states
corresponding to such functions R are thus definitely unstable, owing to the
presence of the positive root in (29). The reason for the negative root in (29)
is entirely analogous to the corresponding reason in the case of the inverted
pendulum. Whatever the initial density perturbation, the initial velocity
perturbation can be chosen in exactly the right way so that the fluid finally
ends up in its (unstable) equilibrium state wherein heavy fluid strata are at
the top of the layer.
Deduction 2. If dR/dz < 0 for O;::;;z;::;;l, then we obtain u = iJP

for some positive quantity P. The motionless state is neutrally stable to the
class of perturbations corresponding to eigenfunctions of (21). [Again, this
follows at once from (28).J When the density continually decreases with
height, velocity, density, and pressure oscillate in time, in analogy with the
ordinary pendulum's periodic oscillation about equilibrium. Wave motion
must thus be expected when a stably stratified fluid is disturbed. Such "inter-
525
nal waves" are frequent and important in various atmospheric and oceano-
graphic phenomena.
We conclude our illustration of the qualitative deductions that can be
made from (21) by stating the following noteworthy result (Ince, 1927,
p. 237). If
dR
dz > 0 for z0 < z < z1, where O :::;;z0 < z1 :::;; 1,
then there are an infinite number of positive eigenvalues a 2 Thus a motionless

layer of stratified inviscid fluid is unstable if it contains any sublayer whatever,
no matter how thin, wherein the density increases with height.
DETAILED RESULTS FOR A PARTICULAR STRATIFICATION
We shall now consider (21) for a particular function R. If our interest in
stratified fluids were directly motivated by a particular application, then the
function R would be selected to represent some experimental measurements.
Such a function would probably be rather complicated. Sophisticated analysis
might yield a solution of the resulting version of (21), but it is quite possible
that a numerical approach would be a more appropriate way to obtain the
details that are often desired when one attempts to deal with a practical
problem. Our interests here are more general. As a first step we therefore
seek a particular function R which makes (21) as simple as possible. The
reader might like to cover up the next paragraph and spend a few moments
trying to select such a function.
As Lord Rayleigh was the first to observe, it turns out that if for some
constant /3
R(z) = exp (Pz),
then (2la) reduces to the following equation with constant coefficients:
d2 W
dz 2 + p dz
dW
+ k 2 (/3a- 2 - )
1W=
0
. (31a)
We still have the same boundary conditions

W(O)= W(l) = 0. (3lb)
The constant coefficient linear equation (3la) has solutions of the form
exp (rz), provided that the constant r is a root of the quadratic equation
r 2 +/Jr+ k 2 (/3a- 2 - 1) = 0. (32)
Thus
r= -tP tJQ, (33)
where
Q = f32 - 4k 2 (/3a-2 - 1). (34)
It is convenient to consider separately the cases where the roots r 1 and r 2

of (32) are (1) real and unequal, (2) real and equal, and (3) complex conjugates.
CASE (1): Q > 0. The general solution to (3la) is
W = C1 exp (r 1z) + C2 (exp r 2 z),

where
In order that there be a nontrivial solution satisfying W(O)= W(l) = 0,

direct calculation or application of (24) yields the requirement exp (r 1) =
exp (r 2 ), which is not satisfied for positive Q.
CASE (2): Q = 0. The general solution to (31) is
W = C1 exp ( -t/3z) + C2 z exp ( -t/3z).

The boundary conditions of (3lb) imply that C1 = C2 = 0. As in case (1),
there is no nontrivial solution.
C As E (3): Q < 0. We write Q = - 2 so that the general solution to
(31a) is
W = exp ( -t/3z)[C 1 cos (tz) + C2 sin (!-z)]. (35)
The requirements W(O)= W(l) = 0 imply that
C1 =0, C2 sin (!-) = 0,
so for a nontrivial solution,
f=mr:, n= 1, 2, 3, .... (36)
Since Q = - 2
, from (36) and (34) we deduce that if u 2
= u;, where
u;= 4k2/3(4k2+ 132+ 4n21t2)-1, (37)
then (31) has the nontrivial solutions (eigenfunctions)
W = exp ( -t/3z) sin (n1ez), n = l, 2, 3, .... (38)
Since sin (-n1tz) = -sin (mr:z), no new eigenfunctions are obtained by
allowing the possibilities n = -1, - 2, - 3 in (38).
N o TE. Equation (37) is an example of the general relation (25), which
we asserted must connect u 2 and k 2 if the eigenvalue problem is to have
nontrivial solutions.
From (37), if /3< 0 (density decreases exponentially with height), then u is
a purely imaginary number. The equilibrium state is neutrally stable with
respect to the class of perturbations for which Wis given by (38); and p, u,
and v are found from (19) and (16). (Remember, W lV.) =
In considering (37) for /3 > 0 (density increases exponentially with height)
we note first that u 2 is positive (instability) in accordance with deduction 1.
Next, let us explore the effect of varying n. Remember that for fixed horizontal
527
disturbance period 2n/k there is a disturbance mode of vertical period2/n for

every positive integer n. Each such disturbance is multiplied by a different time
factor exp (unt), where Unis given by (37). The growth rate Unis largest when
n = 1, and it approaches zero monotonically as n -+ oo.
The explanation for the variation of growth rate with n lies in the different
streamline patterns for different values of n. Figure 15.7 shows these lines,
@
n= I n=2
(a) (b)
n=3
(c)
FIGURE 15.7. Schematic streamline patterns for perturbation modes

of various vertical periodicities. The perturbations are superimposed on
a motionless layer of fluid whose density variation with height is given by
exp (ftz), where f3is small and positive. The overturning or vortical motion
is viewed along the axes of the vortices.
along which the flow circulates with ever-increasing speed. (In this figure the
density stratification exponent f3has been taken to be small, to simplify the
calculations required to obtain the streamlines. These calculations are re-
quested in Exercise 2.) Evidentally, the motion accelerates faster when the
heavy fluid falls in a single overturning vortex which fills the whole layer
[Figure 15.7(a)] than it does when there is a double vortex [Figure 15.7(b)],etc.
For fixed n and f3,u 2 increases monotonically from zero to p as k increases
from zero to infinity (Fig. 15.8). It is as expected that the maximum value of
u2 increases as p increases. This means that, other things being equal. the
a2
FIGURE 15.8. Variation of the perturbation growth rate as a function

of the horizontal perturbation wave number k.
motion is faster the greater is the contrast in density between the heavy fluid
above and the light fluid below. Less obvious is the reason why motion is
faster for larger k's and hence thinner vortices. The explanation lies in the
fact that the rate of increase in perturbation kinetic energy is equal to the
rate of decrease of potential energy due to the falling of the relatively heavy
top fluid (Exercise 3). Since gravity acts vertically, horizontal motions do
not alter potential energy. Motion in narrow vortices is faster than motion
in wide vortices because the latter contain a larger amount of horizontal
motion, which is unproductive in the release of potential energy.
Similar observations flow from (37) when P < 0. These are left to the
reader [Exercise l(b)]. Even without further thought, however, the reader must
agree that a surprisingly large amount of physical understanding has emerged
from a careful examination of (37). One natural question remains: how much
of this understanding applies only to the particular exponential function R
for which (37) holds and how much applies to all functions R, or a large class
of them? No convincing answer is possible without further investigation,
but the general character of the explanations in the previous paragraph makes
one fairly confident of their wide applicability.
Investigations of special problems are usually more productive than in-
experienced persons anticipate. As in the mechanism for the release of
potential energy, focusing one's attention on an unusual aspect of a particular
problem frequently leads to insights having general applicability. More
prosaic aspects of a particular problem also often turn out to be characteristic
of large classes of problems. The fact that a class of problems shares a common
aspect may only be suggested after several particular problems are solved and
their solutions compared. Usually, a general argument can then be formulated,
which proves that the suggestion is valid.
SUPERPOSITION OF NORMAL MODES
Our entire discussion thus far has been based on the behavior of certain
particular solutions. For example, in the case R(z) = exp (Pz) the only vertical
variation considered was that given by one of the solutions exp ( -tPz) sin mrz
529
of (31). But because the problem is linear, superposition of these solutions
is permitted. How general a function can be represented by such a super-
position? That is, for what class of functions f is it true that for appropriate
constants dn,
a,
f(z) = L d" cPn(z)
n=l
for O :s:z :s:1, (39)
where
= exp ( -!Pz) sin mtz?
cl>n(z) (40)
If we multiply both sides of (39) by exp HPz), the problem becomes one of
selecting constants dn to guarantee that
a,
F(z) = L d" sin mr:z,

n= 1
0 :S:z :S:I,
where
F(z) =f(z) exp (!Pz).
This is a classical problem in Fourier series. We have discussed such problems

at some length in Chapter 4 and shall not go further into them here.
From a larger point of view, as mentioned above we are considering one
of the basic problems of the Sturm-Liouville theory discussed in Chapter 5.
This theory deals with the properties of a class of eigenvalue problems of
which (21) is a representative. Suppose, then, that we consider the expansion
(39), where the </Jn(z)are eigenfunctions of the Sturm-Liouville problem (21).
We assume only that R(z) possesses two continuous derivatives for O :s:z :s:I.
It can then be shown, as in Ince (1927) or Coddington and Levinson (1955),
that if
dn = f f(z)<j)n(z)dz,
0
t
(41)
then the right-hand side of (39) converges uniformly to f(z) for O :S:z :S:1,
provided that f(z) satisfies the same boundary conditions (31b) as the
eigenfunctions and also possesses two continuous derivatives for O :S:z :s:I.
Onecan go furtherand attemptto dealwitha generalhorizontalvariation
by superposing (adding together) solutions of the form (14) for various
values of k 1 and k 2 An example of this procedure is given in the discussion
of water waves found in Chapter 8 of II. For the present, suffice it to say
that a very large class of perturbations can be handled by appropriate
superposition of the eigenfunctions (14) which satisfy the homogeneous
linear partial differential equations (10)-(12), the boundary condition (13),
and the requirement that solutions remain bounded as x 2 + y 2 -+ oo. (The
last requirement was used in discarding complex values of k 1 and k 2 .)
Looking for eigenfunctions like those of (14) and hoping to be able to

handle a general initial value problem by superposition is called the normal
mode approach to stability. The eigenfunctions found by this approach are
the normal modes. We observe that a rigorous normal mode treatment of the
linearized problem requires not only a proof that a given class of initial
conditions can be synthesized from the normal modes, but also that the
corresponding infinite series satisfies the equations and boundary conditions.
R E M A R K . When the normal mode approach fails, as it sometimes does,
the use of Laplace transforms usually saves the day. For an example, see G.
F. Carrier and C. T. Chang, "On an Initial Value Problem Concerning
Taylor Instability of Incompressible Fluids," Q. Appl. Math. 16, 436-39
( 1959). [Some might find it helpful to view the situation in the following terms.
Normal modes correspond to isolated eigenvalues (point spectrum). In many
instances, a continuous spectrum is also present and the normal modes do not
form a complete set. Integral superpositions are required to solve the initial
value problem.]
NONLINEAR EFFECTS
We have treated a set of perturbations which are so small that nonlinear
effects can be ignored. But is this set nonempty? That is, can one show that
the behavior of perturbations of sufficiently small amplitude is given by the
linearized theory? In many cases, recent work allows an affirmative answer
to be given to this question. (We speak here and below of stability theory in
general.)
If linear theory reveals even one perturbation that grows, we feel quite
confident that the equilibrium state is unstable. Nonlinear terms might act
to restrict the growth of a disturbance to very small amplitudes. Nevertheless,
the resulting flow will differ from the equilibrium state, so the latter can
properly be said to be unstable.
By contrast, if all disturbances die out according to linear theory, this is
only evidence that sufficiently small disturbances ultimately will leave no
effect upon the equilibrium state. This is a useful result, for it means that if
one is sufficiently careful in isolating the equilibrium flow from disturbances
then the flow should persist. But how careful does one have to be? Might it
not be the case that the equilibrium flow will be destabilized by disturbances
which are physically "moderate" but which are sufficiently large so that linear
theory is not applicable? Such questions form the matter studied in nonlinear
stability theory whose development "took off" in the 1960s. In a sentence
one can report that some flows are extremely sensitive to large-amplitude
perturbations, while for others increasing the amplitude of the perturbation
does not hasten the onset of instability.
The computer is indispensable in studying large perturbations. As an
example of the results achieved, consider a numerical study by B. J. Daly
[(Phys. Fluids JO, 297 (1967)) of the stability of a heavy layer of fluid
placed above a lighter layer. Daly reports that" at a density ratio of 10:1,
the heavy fluid falls through the lighter fluid in the form of a sharp narrow
spike whose amplitude accelerates with time. At density ratios of 2: 1 or less,
however, the tip of the spike is broadened ... so that its late time velocity ...
is constant." A later calculation by the same author [Phys. Fluids 12,
1340 (1969)] examines the effect of surface tension upon the breaking up of
the spike into separated drops.
Calculations of the initial development of the flow, when perturbations
are small, agree with linear stability theory. Throughout, the calculations
are in agreement with relevant experiments.
WORKED EXAMPLE: A MODEL OF VISCOUS FLOW
INSTABILITY
A feature of linear stability theory has not yet been illustrated, namely,
the prediction that at the onset of instability growing disturbances will have
a particular finite wavelength. We have seen that for a stratified inviscid fluid,
disturbances of all wavelengths are unstable when the fluid is top-heavy.
Our study of the particular case R(z) = exp (f3z)showed that growth speed
decreased with wavelength, because thinner vortices are more efficient at
using up the available potential energy. When viscosity is considered, how-
ever, very small wavelengths are inefficient because the resulting large
velocity gradients dissipate a great deal of energy. A "compromise" leads
to a "most dangerous disturbance" of intermediate wavelength.
Thus, when viscous fluid layers become sufficiently top-heavy, they begin
a circulatory motion with a well-definedwavelength. Sufficient top-heaviness
is required because the stabilizing nature of viscosity will otherwise prevent
motion.
It is beyond our scope here to pursue the investigation of stability in top-
heavy fluids. [For further information see S. Chandrasekhar's book Hydro-
dynamic and Hydromagnetic Instability (New York: Oxford U.P., 1961).]
We do wish to point out, however, that there are many instances where in-
stability sets in at a specific wavelength, when the destabilizing influence
is sufficientlylarge. One such instance is illustrated in Exercise IO, a study of
chemical reaction in the presence of diffusion. Another will now be outlined.
Simple viscous flows become unstable if they are sufficiently speeded up.
This fact can be appreciatedwithout studyingviscousfluid motion, by an
analysis of a model equation. This equation is contrived so that the analysis
preserves the essential features of the full calculations but does not share
their complication.
Example. Considerviscousfl.ow(viscosity= v) betweenparallelplatesa distance
d apart. Let the bottom plate be stationary and the top plate move with speed U.
Pretend that the dimensionlesstangential speed u(x, y, t) satisfies the following
equation and boundary conditions:
u(x, 0, t) = 0, u(x, 1, t) = 1. (42)
The true equations have the same general structure as (42). In particular, the
dimensionless speed parameter or "Reynolds number" R Ud/v appears in a very =
similar fashion.
It is easy to verify that (42) hasthe exact solution u(x, y, t) = y. t Consider small
perturbations to this solution of the form/(y) cos kx exp (at). Show that the (R, k)
plane is divided into regions of stability and instability as shown in Figure 15.9.
Find Re and ke. If the speed of the flow is slowly increased from a small value,
instability will set in when R = Re at a wave number of ke . Why?
I
I_______
RC,_ _
Stable
I
1
FIGURE 15.9. The neutral curve in the plane of Reynolds number R

versus wave number k divides the stability region from the instability
region.
Sol11tion. If u(x, y, t) = y + u'(x, y, t), the lineari7.Cdequation for u' is

+ u;,J- u;.,..,..,.
R-. [u;.,.
1 - u~- u;:r= 0.
If u'(x, y, t) = /(y) cos (kx) exp (at), then/satisfies
f" + 'Af= 0, /(0) = /(1) =0, where 'A= Rp. - Ra - R.2 - ..
The only nontrivial solutions of the eigenvalue problem for f occur when 'A= n 2r,
n = 1, 2, .. [The solutions are constant multiples of sin (mrz).] Thus
Rp.- Ra-Rp. 2
- . = n2 1F2
and a < 0 (stability) if and only if
11-+nz,,.2
R<R(.,n) where R(., n) = .- .2
t The dimensional speed and vertical coordinate y* uc related to II and y by u*JV= 11,
11
y* Jd= y. Thus the solution 11= y corresponds
to u* = V.,./d. 1bis is a ftow whose speedincreuel
linearly from zero at the motionless plate y* ==O to V at tbe movinaplate y* .. d
Sec. 15.2} Stability of a Stratified Fluid
533
Instability will set in as R increases when the above inequality is reversed, at the
smallest value R. of R(., n). By inspection the smallest value occurs when n = 1.
Minimization with respect to . then gives
.. + 17'2
R.=---2,
..- ..
where..= k:= -17'
2
+ 'l'l''V'1'1'2 + 1.
For R < R., u < 0 no matter what the wave number k of the perturbation. When
R is just barely bigger than R., perturbations can grow, namely, those with wave
numbers at or near k .
EXERCISES
1. (a) Remarks on the geometric significanceof the wave number vector k
were made under (21). Verify these remarks.
(b) Discuss in physical terms the variation of the oscillation frequencies
that are given by (37), when p < 0.
2. This problem requests verification of the qualitative streamline picture
of Figure 15.7. With no loss of generality the y axis can be aligned
along the axis of the perturbation overturnings, so v' = 0, k 2= 0. For
simplicity restrict your consideration to the case of small p.It is virtually
sufficient to determine the lines x = constant and z = constant along
which the fl.owis vertical or horizontal.
3. From (11) deduce that if the flow pattern consists of the periodic
repetition of flow in a domain D (e.g., a rectangle or a hexagon), then
~ fff
ut D
tR(v'v') dt = - Jff
D
w'p' dt ,- ( , ,
[V=U,V,W. ')]
(In the linear approximation the left side of the above equation is the
rate of change of the perturbation kinetic energy; the right side is the
rate at which work is done by the buoyancy force.)
4. An alternative to assuming (14) and taking the real part of the final
answer is to assume that
u' = [u1 cos (k 1x + k 2 y) + u2 sin (k1x + k2y)]e.,.t,
with similar assumptions for the other dependent variables. Convince
yourself that this is a bad alternative by attempting to carry through the
required calculations. (If you complete the calculations, you should
obtainthe sameresultas that obtainedusingthe complexnotation,but
with much more effort.)
is. In biology, the number of individuals p in a given population at time t
is often approximately governed by the equation dp/dt = kp, where
k = k(t) is the reproduction rate. Consider the possibility
dp 2
k = ex- pp (ex,Ppositive constants) - PP.
so dt = r:1.p
534 Inviscid Fluid Flow [Ch.15
(The {Jpterm in k represents the decline in population growth rate due

to overcrowding and the concomitant decrease in the food supply,
increase in level of toxins, etc.) Clearly, there are two exact steady
solutions to the governing equation:
a:
p = Po where Po = 0 or Po = -
. fJ
Investigate the stability of these solutions to small disturbances.
6. This problem concerns inviscid fluid of unit uniform density (p I) =
contained between two infinite rigid parallel plates y = 0 and y = 1.
Gravity will be ignored and only two-dimensional motions will be
considered (w = o/oz= 0). The governing equations and boundary
conditions
Dv
Dt = -Vp, Vv = O; v=O on y=O, 1,
have the solution

u = U(y), v=O, P=Po,
where U is an arbitrary differentiable function and p 0 is a constant.
(a) Determine a set of linear partial differential equations that govern
small perturbations to the above solution.
(b) Assume solutions proportional to exp [ia:(x- ct)] with exreal. In
particular, take v = V(y) exp [ia:(x- ct)]. Eliminate pressure and
horizontal velocity. Thereby show that the problem can be reduced
to
(U - c)(V" - a:2 V) - U"V = O; V(O)= V(l) = 0. (43)
Why does the existence of nontrivial solutions to (43) imply a
functional relation between c and a:2 ?
(c) Let c = c, + ic; and assume that c; #: 0. Divide (43) by U - c.
Multiply by V, the complex conjugate of V, and integrate between
zero and 1. Perform an integration by parts. Take the imaginary
part of the resulting equation. Obtain
C; t U"I v1 1
1
2
U- ci-2 dy = 0.
Deduce that U must have at least one inflection point if c; .,,,,.
0 and
comment on the significance of the result.
7. As a model of fluid motion, J. M. Burgers considered the following
system of equations for u(t) and U(t):
dU 2 du
-=P-u -vU dt = Uu -vu.
dt '
Here P and v are positive constants.
535
(a) Show that there is a solution U = P/v, u = 0, which is stable if and

only if P < v2
(b) Show that when P > v2 , there is another pair of stable solutions
with constant values for U and u. [It is reminiscent of real fluids
that a relatively simple flow (with u = 0) becomes unstable at a
critical value P = v2 of the driving "pressure gradient" P and that
a new type of flow emerges which is stable but more complicated.]
8. In this problem we consider the stability of an inviscid incompressible
fluid layer of uniform density R1 , moving horizontally with uniform
speed Vi over a layer of uniform density R 2 and uniform horizontal
speed U2 Gravity acts vertically downward. We restrict ourselves
throughout this problem to two-dimensional motions (v = o/oy 0.) =
It is reasonable to conjecture that our results will have something to do
with the generation of water waves by wind.
(a) Find Pi(z) and pz(z) so that if U1 , U2 , R1 , and R2 are constants,
then the following provides a solution to the inviscid incompressible
flow equations (la-c).
For O < z < oo: u* = U 1 , w* = 0, p* = R 1,p* =Pi(z).

For - oo < z < 0: u* = U2 , w* = 0, p* = R 2 , p* = pz(z).
Also insure the continuity of pressure at the interface z = 0

between the two immiscible fluids [Exercise 14.4.8(b) with sn = 0,
t = -pn].
Account must be taken of the kinematic boundary condition
w* = C,+ uC,,on z = C(x,t), where z = C(x,t) is the equation of
the interface. (This condition is derived in Section 7.1 of II.) For
the present simple solution, z = 0 is the equation of the interface
and the kinematic boundary condition is clearly satisfied.
(b) In the upper layer, write
u* = U 1 + u;,, w* = w;,,p* = R1 + Pi, p* = P1(z) + P1 (44)

Derive a linearized set of partial differential equations for the
perturbations [the primed quantities in (44)]. Do the same for the
lower layer perturbations. (Use the subscript 2 for these.)
(c) Write
w;(x, z, t) = w;(z)exp [ici(x - ct)] (cireal, c complex),
plus similar expressions for u;,

p\, p;; i = 1, 2. Reduce each set of
partial equations in (b) to an ordinary differential equation, one
for w1 and one for w2 .
(d) Let the perturbed interface between the two layers be written
z = C(x, t) = { exp [ia(x - ct)], ea constant.

From the kinematic boundary condition derive

'IV1 'IV2
--=--
U - c U2 - c
at z = ,.
1
(e) We are neglecting surface tension, so the pressure (not the pertur-
bation pressure) must be continuous at z = (. Deduce
at z = C
(f) It is consistent with our small perturbation assumption to apply

the boundary conditions of (d) and (e) atz = 0. (Why?) It simplifies
the algebra a little to consider the case U1 = - U2 = V. (Why
does this involve no real loss of generality?) Use these two remarks
to solve for 'IV1 and 'IV2 (impose sensible conditions as izl -+ oo)
and hence to deduce
c = c'" (c4 + c.,)112
Here c'"is a weighted average velocity of the two layers
R 1 U1 + R 2 U2
C,n = R1 + R2 '
and c., and c., result from the density and velocity discontinuities,
respectively:
g R1 - R2 R1RiCU1- U2 ) 2
C11=-aR + R 2 ' c., =
(R1 + R 2) 2
1
(g) Show that there is always instability when heavy fluid is above
lighter (Taylor instability).
(h) Compare the result for V = 0 with the text's results for exponential
variation of density with height. As predicted here, an oil-water
interface oscillates much more slowly than an air-water interface.
This was noticed by Benjamin Franklin (Lamb, 1932, p. 371).
(i) Demonstrate the existence of instability when R1 = R 2 (Kelvin-
Helmholtz instability). Show that sufficiently short wavelengths
are still unstable when R1 < R 2 so that, contrary to experience,
the slightest zephyr should cause the appearance of short waves
in water. [When the approach of this exercise is modified by the
inclusion of surface tension effects, theory predicts that a wind
speed well in excess of 10 miles per hour is required before water
waves are generated. This speed is much greater than the observed
value so that a more subtle mechanism of wave generation must

be sought. Much research has been devoted to this problem.
See, e.g., A. Gupta, M. Landahl, and E. Mollo-Christensen, J.
Fluid Mech. 33, 673-91 (1968).]
t9. In the case R = constant, the term involving the growth rate u dis-
appears from (21a). Reconsider this case. Since the assumption of ex-
ponential time dependence leads to difficulty, instead of (14) start with
the assumption
u' = u(z, t) exp [i(k1x + k 2 y)]; etc.
(a) Show that for some functions oc, 0 , v0 , and w0 ,
u= ik 1k- 2 oc'(z)t+ u0 (z), D= ik 2 k- 2 a.'(z)t + vo(z),
~ = oc(z)t+ w0 (z), p = R[k- 2 oc"(z)- cx(z)], p = -Rk- 2
cx'(z).
(b) Interpret the results (i) when the problem is regarded as examining
the stirring up of a layer of uniform density (so p = O); (ii) when
the problem is regarded as examining the fate of a layer whose
initial density is nearly uniform (sop is some given function of z).
10. Publication of a paper* by the mathematician A. Turing (better known
for his fundamental work in logic) has sparked interest in the possibility
that chemical instabilities play important roles in biology. A contribution
to this matter was made by I. Prigogine and R. Lefever,t who studied
the (artificial) reaction scheme
k1 k2 k3
A--X, 2X+ Y-- 3X, B+X-- Y+D, X~E.
(45)
To obtain governing equations, diffusion terms are added to terms that
are required by the law of mass action (Section 10.1). For simplicity, we
consider only a single cartesian spatial coordinate r. The reader may
wish to verify that one can obtain the following equations for X and Y.
ax a2 x
-=k 1A+k X 2 Y-k BX-k X+Dx-;;-z, (46)
ot 2 3 4
ur
ay a2 y
2 X Y + k 3 BX+ Dr
2
at= -k , (47)
02
Since the same number of molecules of X and Y are represented on
the left of the expressions in (45) as on the right, (45) gives the over-all
reaction scheme A + B-+ E + D. We shall regard the initial constituents
as having essentially fixed uniform concentrations A and B, so A and B
Proc. Roy. Soc. (London) B237, 37 (1952).

t J. Chem. Phys. 48, 1695 (1968).
will be taken as constants in all that follows. The purpose of this

problem is to show that there is a critical value of concentration at
which an instability sets in.
(a) To make the calculations slightly easier, we shall assume that Dx
and Dy are equal and shall take Dx = Dy= D. What is the physical
meaning of this assumption? When might it be expected to be
valid, or nearly so?
(b) Show that (46) and (47) have a unique uniform solution X = X 0 ,
Y = Y0 , where the constants X0 and Y0 are given by
(48)
(c) Show that upon introduction of the change of variables

x y
x=-, y=-, (49)
Xo Yo
(46) and (47) become
ox= ,c + {3x2y -
- {3x- ,cx + -
o2x, (50)
ot os2
oy z a2y
Ot = -yx y + '}'X + os2 ' (51)
with
If you have studied the material on dimensional analysis and

scaling, describe fully why this change of variables is a good idea.
In doing so, mention what conclusions you can draw from (50)
and (51) without further work, and state what simplifications could
now be employed (if necessary) in a further study of these equations.
(d) Show that by writing x = I + x, y = 1 + ji, with the further
assumption
x = ~ cos s e'", ji = Pcos s e'",
the stability of the uniform solution to small perturbations can be
ascertained from a study of the equations
o = ~(-<1-A) + PJ, 0 = -y~ - (.B+ <1)J,
where
A = " + 2 - p,
Sec. 15.3] Compression Waves in Gases
539
2
(e) Show that u must satisfy an equation of the form u +bu+ c = O.
Demonstrate that disturbances die out if and only if b > O and
c > 0. In this part and the next we consider the possibility that
instability arises due to a change of c from positive to negative.
Under these circumstances demonstrate that a necessary and
sufficient condition for stability to a perturbation of wave number
is AB + {3y > 0, or
{3< {3c(m) where {3c(m)= (m + ,c)(ym- 1 + 1), m =2
t(f) Continuing part (e), find the value of {3 at which instability will
first set in as {3is slowly increased, and find the expected wavelength
of this instability.
t(g) Now consider the possibility of instability due to a change in the
sign of b. Find the value of {3 at which this type of instability will
first set in as {3is slowly increased. Show that this value is less than
the corresponding value of (f) and thereby deduce that instability
will actually commence in the oscillatory manner described here
in part (g). [If the assumption Dx = Dr is dropped, then there are
parameter ranges such that instability begins in the fashion des-
cribed in parts (e) and (f).]
15.3 CompressionWaves in Gases

Compression waves in gases abound in nature. In everyday life we ex-
perience acoustic waves of human communication, noise and music. Ex-
plosions produce "shock waves, " which are compression waves of high
intensity. During the late stages of stellar evolution, a star explodes and
sends out shock waves that eject much of its mass. In the solar atmosphere,
acoustic waves are believed to be responsible for keeping the sun's corona
region at a temperature of 1 or 2 million degrees, whereas the surface of the
sun remains relatively cool at a temperature of about 6000K.
In the present section we shall briefly describe the elementary theory of
compression waves in ideal gases, beginning with the basic equations and
the theory at small amplitudes.
A general discussion of compression waves leads to the development of the
theory of partial differential equations of the "hyperbolic" type. We cannot
go at all deeplyinto the generaltheoryhere,but we shallreproducesomeof
the early work of Riemann for one-dimensional compression waves of finite
amplitude. We shall thereby afford one of the few glimpses at nonlinear
effects which are offered in the present volume.* At present, fuller treatment
of nonlinearity belongs in advanced courses. Moreover, new approaches to
The perturbation methods of Part B are used on some nonlinear problems, but the non-
linearity does not play such a large role there.
540
nonlinearity are a central concern of contemporary research. Older books

sometimes inculcated the attitude that the appearance of nonlinearity was a
matter to be viewed with near despair, but we would like to leave the correct
modem impression. Nonlinearity today offers a challenge to be viewed with
enthusiasm. Perhaps existing numerical and analytical methods will work,
at least after some adaptation; or perhaps a reputation can be made by invent-
ing a new method.
INVISCID ISENTROPIC FLOW OF A PERFECT GAS
We shall employ the following basic equations of inviscid hydrodynamics:

the equation of mass conservation (14.1.7),
op
ot
+ I
1=1 ox1
(pv;) = O;
.!.__ (I)
the Euler equations of momentum balance (1.10), in the absence of body force,
ov1+
ot j=l
f
v. ov1= _..!_op ;
1 oxj pox-
(2)
and the thermodynamic equation (14.4.14),
+ f vi:s)
o(!s =0. (3)
vt i=l vXi
Here pis density, the v 1are velocity components, pis pressure, f:Jis temperature,
and s is entropy per unit mass.
In (3) we have omitted the term Dq/Dt, i.e., we have assumed that the heat
flow into a given piece of gas is negligible during any time that this piece
undergoes significant distortion. Thus the distortion process can be regarded
as adiabatic. Moreover, we neglect the dissipation of mechanical energy into
heat, an assumption that is consonant with the neglect of viscosity. (Compare
Exercise 3.1.8 of II.) According to (3), the flow is isentropic.
The equations above are to be solved together with an equation of state.
From Section 14.4 or from previous knowledge of elementary thermo-
dynamics, we recall that we can use either
p = Rp8 or p = K0 e'1cvp1, K0 a constant, (4a, b)
for either one of these equations can be derived if the other is regarded as
known (and specffic heats can be regarded as constant). Here Cy, the specific
heat at constant volume, satisfies
Cy = R/{1 - I), (5)
where "I= C,/Cy is the ratio of specific heats and Risa constant for a given
gas.
Sec. 1.5.3] Compression Waves in Gases
541
Equation (3) states that the entropy of a fluid particle is constant. We shall
only consider cases (such as flow started from rest) where the entropy is
uniform throughout the flow. In such cases s = constant and we can take as
an equation of state (4b) in the form
P =Kp', K a constant. (6)
This is frequently called the polytropic state equation. It is sometimes replaced
by the more general barotropic relation p = p(p), since the theory is not
greatly complicated thereby.
If we regard the pressure as a known function of density, then the phenom-
ena are governed by (I) and (2), the equivalent of four scalar equations for the
three velocity components and the density. The initial distribution of density
and velocity must be prescribed, and the normal component of velocity must
vanish on any fixed solid boundaries. On moving solid boundaries, the normal
velocity of the gas must be the same as that of the boundary.
WAVES OF SMALL AMPLITUDES
We wish to consider small disturbances to a homogeneous motionless
atmosphere. Let the undisturbed atmosphere have uniform density p 0 To
consider a small disturbance, we write
!....= 1 + u. (7)
Po
The quantity u, sometimes called the condensation, measures the departure of
the dimensionless density ratio pf p0 from its unperturbed value of unity.
We look for a solution where the condensation CT is small compared to
unity and where the magnitudes of the velocity components are also small.
The product of two small factors should thus be negligible, so our strategy
will be to delete all nonlinear terms.
Substituting (7) into (1), we find that
OU 3 OV 3 0
- + L -' + L - (uv;) = 0. (8)
ot i=l ox; i=l ox;
We neglect the third term compared with the second to obtain the linear
equation
OU+
ar
t
i=l
OV;= o.
8x;
(9)
The linearization is valid if not only the condensation u but also its spatial
derivatives are small compared to unity. If we also linearize (2), we obtain
[Exercise I(a)]
OV; 2 OU
-= -Co- (10)
or ox;
542
We have written
2( ) p"(po)(p - Po>2
P(p) = P(po) + Co P - Po + 2
! + , (11)
where
c~ =p'(po) (12)
What has been implied in the linearization that resulted in (10)? To see,
let U and I be typical magnitudes of the velocity components and con-
densation. Let spatial derivatives multiply the magnitude of a quantity by
about L - i and temporal derivatives by about -r- 1 Then (9) and (10) imply
that
(13)
On elimination of I the above equations yield
(14)
(This is a dimensionally correct result, for c0 does have the dimensions of a
speed.) Thus the convective acceleration v Vv is, as assumed, small compared
to the local acceleration ov/otif
i.e., if U ~ c0
The ratio U/c0 is called a Mach number. Our linearization appears to be
valid if this ratio of fluid speed to sound speed is sufficiently small.
From (9) and (IO) we derive the wave equation for the condensation u:
or (15)
In general the velocity components v; will not satisfy the wave equation
but both ov.fotand "f.f=iovk/oxk
do [Exercise l(f)).
THE SPEED OF SOUND
If u = a(x 1, t), at a given time the density is constant on the planes x 1 =
constant. The development of such plane waves is governed by
ou
2
ou
2
(16)
ar2=~ oxf
-Solutions in unbounded space of this one-dimensional wave equation have
been discussed in Section 12.2. There the equations governed small displace-
ments in an elastic medium. The formal nature of the solutions is of course
independent of their physical interpretation (and the discussion of the wave
equation in Section 12.2 is independent of the rest of Chapter 12.)
Plane wave solutions of (16) propagate with unchanged shape at speed c0
Thus c 0 is the speed of soundat the density p 0 This speed is not unequivocally
Sec. 15.3] Compression Waves in Gases 543
specified by our analysis, for the assumed barotropic relation p = p(p) can
be obtained in more than one way (as we now discuss).
In his study of acoustic waves, Newton made the seemingly plausible
assumption that the gas temperature does not change as it expands and con-
tracts. Thus he regarded the equation of state as having the form p = p(p, lJ)
with 8 = constant. For a perfect gas this would amount to using the equation
(4a) with lJ = constant, i.e., pp- 1 =constant.With this,
c2= (op) E.
op o P
=
(It was actually Euler who derived the above result by appeal to the wave
equation, thereby improving on Newton's less formal argument.)
Observation shows that the Newton-Euler isothermal sound speed is not
correct. It was Laplace who pointed out that the expansion and compression
in sound waves should be regarded as an adiabatic and hence isentropic
process. [See the discussion of (3).] The state equation p = Kp 1 is thus the
correct one for perfect gases, and
c2 = (op) yp.
op P
= (17)
SPHERICAL WAVES
If the wave has spherical symmetry, the Laplace operator becomes
2 o2 -1 o
v = or2 + 2r or'
where r = (xf + x~ + x~)112 is the distance of (x1, x 2 , x 3 ) from the origin.
The wave equation takes the form
Co2 -
o2 u = cl2u
-+2
oa
2r-l - (18)
ot2 or or
To solve this equation, we introduce the transformation of variable
p
U=-. (19)
r
Then Psatisfies the equation of plane waves
-2 o2p a2p
Co Ot2= or2 ' (20)
whose solutions we have studied. Thus the general solution for spherical
waves has the form
u = r- 1/(r - ct)+ ,- 1g(r + ct). (21)
The two terms represent outgoing and incoming waves, respectively.
The amplitude of the spherical waves decreases with distance according to

the factor 1/r. This could have been anticipated [and hence the change of
variable (19) could have been motivated] by the following considerations.
The energy in the wave is proportional to the square of its amplitude. No
mechanisms for energy dissipation have been included in our model, so
energy on a spherical shell will propagate outward on a surface whose area
increases like r 2 Conservation of energy requires that the energy be pro-
portional to r - 2 , so the disturbance amplitude itself must be proportional to
-1
The solution (21) generally has a singularity at r = 0. This singularity

usually does not appear in the solution to a physical problem however. For
example, if we consider a pulsating sphere that sends out spherical waves,
we exclude small values of r.
A solution of the form (21) that has special interest is found by taking
f(r - ct) = sin k(r - ct), g(r + ct) = sin k(r + ct), (22)
where k is a constant. With (22), (21) becomes
1
u= 2,- sin kr coskct. (23)
This standing wave solution is finite throughout infinite space. Many other
interesting solutions can be obtained by superimposing such standing spherical
waves.
NONLINEAR WAVES IN ONE DIMENSION
In the linear theory of wave propagation, the wave shape is preserved.
When the full nonlinear equations of hydrodynamics are used, as is ap-
propriate when the amplitudes of the waves are no longer small, we do not
expect the general nature of the propagation to be changed. We do expect,
however, that the wave form may be distorted. We also expect that the
waves propagating to the right and to the left may interact with each other.
To demonstrate these effects, we consider the nonlinear problem following
the method of Riemann.
We begin our investigation by looking for solutions that propagate in
only one direction. In the linearized case, such waves are a function of a
combination of x and t, either x + ct or x - ct. Let us see if there are solutions
in the nonlinear case that depend on some such combination. We seek simple
wave solutions in which the density and velocity are functions of a single
quantity, cx(x,t).
For flows that depend on a single spatial coordinate x and have velocity
u(x, t), the continuity and momentum equations (I) and (2) can be written
op op ou
ot+ u ox+ P ox= o, <24>
OU OU c2 op
!'It
v
+ u -ox+ -p -ox= 0, (25)
545
2
where c = dp/dp.
Upon introducing
u = u(oc), p = p(oc), (26)
we obtain
'( >(ort.
p oc
orx.)
ot+ u ox + pu '(') ox
orx.
= o, (27a)
2
u '(rt.
)(ooc -ort.)
- + u ox + -c p'( rt.ooc
) -
at P ox= 0. (27b)
These equations are compatible if and only if
c2p _ 2 (dp)
_
doc
2 = (du)
_
da.
2
(28)
or
u =I
c(p)p-1 dp. (29)
The equation for exbecomes
ooc acx
ot+ (u c) ox= o. (30)
Equation (30) is quasi-linear;i.e., it is linear in the derivatives. (It is not

linear in the general sense, since the " coefficients " depend on the dependent
variables u and c.) Such an equation can be treated by the "method of
characteristics," which we shall not discuss in detail here. Instead we shall
write down the solution of (30) and verify it by direct substitution.
We assert that a general solution of (30) is
oc= F(), =x - (u c)t, (31)
which is obtained from (30) as if u c were a constant. To verify that (31) is
indeed a solution, we note that
aoc= F'()
ox [1- t(u' c') ox
0
'],
0
'= F'()[-(u
at -+ c)- t(u'-+ c') orx.]
ot
From this it follows that ooc/ox
and orx./ot
are in the ratio -(u c), and (30)
is satisfied.
The interpretation of (31) can be accomplished in a manner similar to that
for linear waves. Givenvalues of oc,and hence given values of u and c, displace
to the right at the speed u c. That is, the particularstate correspondingto a
given ocpropagatesat a speed u + c for one simple waveand u - c for another.
In other words, the wave propagation relative to the moving fluid element has
a speed c. To this extent, the result is similar to that in the linear theory.
However, in contrast to the results oflinear theory, the propagation involves
a distortion of the wave shape. This is a result of the fact that the propagation
speeds u c vary with a: and hence with the physical quantities u and c, the
latter being uniquely related to the density of the gas.
A relatively clear picture of the distortion can be obtained in the polytropic
case. It is convenient to write the state equation p = Kp 1 in the form
!!_ =
Po
(i..)',
Po
(32)
where p 0 and p 0 are the values of pressure and density at some reference state.
Using this, the definition c2 = dp/dp, and (29) in the more explicit form
U = rc(r)r-
Po
1
dr, (33)
it is not difficult to show [Exercise l(e)] that

2 y- I
u= --(c-c 0) SO C = - - U + Co (34a, b)
}' - 1 2
From this it follows that
u c = ,t(y + I )u c0 , (35)
so from (26), (31), and (35),
u = F{x - [t(y + l)u c0 ]t}, (36)
where the positive (negative) sign is taken for disturbances that propagate
rightward (leftward) with respect to the moving medium.
If the term proportional to t(y + I) were not present in (36), then the
graph of u would propagate either to the left or to the right with speed c0 .
The presence of ,t(y + l)u distorts the graph by adding to the velocity of
FIGURE 15.10. Velocity profile of a shear flow. The arrows represent

velocity vectors; speed increases linearly with height. The graph of (36)
changes just as if it were a material line placed in such a flow.
547
propagation a term proportional to the ordinate. The graph distorts just as if

it were composed of dye particles placed in a stream whose speed increased
linearly with height, as in Figure 15.10.
It can be shown that as long as the speed of sound increases in some fashion
with density, the same qualitative features hold as in the polytropic case.
Although we have found two simple waves, it is not possible to solve the
nonlinear initial value problem by superposition of these solutions. There is a
well-developed theory for this problem, but we shall not pursue it further.
Example. (Details are requested of the reader in Exercise4.) Consider a piston

gradually started from rest and pushed into an infinitely long tube at a speed
substantially lower than the speed of sound. Let the tube be filled with an ideal gas
at density po .
We shall consider the early stages of the solution, during which the leading part
of the disturbance should be of very small magnitude. Thus the leading edge of the
disturbance should travel forward with the sound speed, so we expect
for x ~ cot, p =Po and u = Uo
Let the position g of the piston at time -r be described by g = /( -r). Then it can be
shown that the appropriate simplewave solution can be parametrically described by
u =f'(-r), x - /(-r) = (u + c'X.t- -r); f(-r) :S; x :S; Cot. {37)
Here c is related to u by (34b), taking the positive sign.
SHOCK WAVES
As the distortion of shape progresses, a point will be reached when the
solution becomes three-valued (Figure 15.11). To have three values of density
at a point is unphysical, and some spectacular phenomenon must occur that
can no longer be described without refinement of the theory. Indeed, a shock
wave is formed whose description by inviscid theory requires acceptance of
discontinuous solutions.
The formation of a shock must involve some physical processes hitherto
ignored. These are the diffusive processes of viscosity and heat conduction.
However small the coefficients of viscosity and heat conduction are, just
before the solution develops an infinite gradient [Figure 15.1l(b) ], derivatives
are so large that diffusive processes are no longer negligible. If one assumes
that the distortion process continues outside this diffusive region, it appears
that a discontinuity would tend to form; this is the shock wave.
One can visualize the formation of the shock in another manner. Consider
the piston problem mentioned in the example above. The disturbances have
a front progressing at speed c0 The gas behind the front, being compressed,
has a higher local sound speed. Thus the later disturbances tend to catch up
with the earlier ones and produce stronger compressions.
I
I
I
I
I
I
I
I
I
I
I
L--------------------------x
(a)

_J
I
I
I
I
I
I
I
I
I
I
I
I
L---------------------------x
(b)
;I
I
I
I
5
~--------------------------x
(c)
FIGURE 15.11. Three stages in the spatial variation of the density p

according to simple wave theory.
Analysis of the detailed propagation of compression waves, including

shocks in general, requires a numerical procedure. But we shall confine
ourselves to a brief examination of conditions at the discontinuity.
To see how variables "jump" across shock discontinuities, consider an
observer moving with the shock. This observer sees a shock with a fixed
location. Let flow conditions ahead of the shock be designated with a
subscript a and those behind by a subscript b. Thus the flow into the shock
has a speed u0 , and the flow leaving the shock has a speed ub. It is easy to see
[Exercise S(a)] that conservation of mass requires that
(38)
Other exercises show that the momentum relation requires that

-pa+ PaU;= -pb + Pbu;, (39)
549
while the energy equation requires that
PaUa+ m(tu.; + ea)= p,, u,,+ m(tu; + e,,), (40)
where e is the internal energy per unit mass.
It is a remarkable fact that the detailed events within the thin shock
region can be ignored, and that various compressible flows can be well
described by treating shocks as discontinuities in the solutions. Values of the
dependent variables on either side of the discontinuity are related by "jump
conditions" like (38), (39), and (40).
The successful use of discontinuous solutions in compressible flow theory
emphasizes the fact that theoreticians always deal with models of natural
phenomena. One might even grant the assertion " nature really changes
continuously" but this would not bar the use of discontinuous solutions,
for (as in compressible flow) such solutions might provide excellent models
when regions of rapid change are anticipated.
The use of discontinuous solutions in one-dimensional elasticity has been
treated in Section 12.3. Exercises on deducing general jump conditions
across discontinuities have been given in Section 14.4. Specialization of the
general results to shocks is requested in Exercises 8 and 9. The most important
result is this. As a consequence of the Clausius-Duhem integral inequality
(14.4.29), when an inviscidfluid passes through a shock its entropy cannot
decrease*[Exercise 9(d)].
EXERCISES
1. Verify the following equations:
(a) (10).
(b) (20).
(c) (27).
(d) (28), (29), and (30).
(e) (34) and (35).
(f) Verify the statement after (15).
2. Formulate and solve a problem concerned with the radiation of sound
waves from a pulsating sphere.
3. If you have studied the material on nondimensionalization and scaling,
apply these ideas to the derivation of the wave equation for the conden-
sation. Your answer should take the form of a brief essay.
4. (a) Derive (37).
(b) Discuss the solution given by (37) in the special case /('r) = ar2,
where a is a constant.
According to the energy equation (3), a particle's entropy remains constant. But the essence
of the treatment of shocks as discontinuity surfaces is to regard all dissipative processes [which
were neglected in (3)] as present but only in narrow shock regions. In particular, (3) cannot be
used to determine jumps across the shock.
t(c) Comment on the statement: "This problem is without physical

signiftcance because there is no such thing as an infinite tube."
S. (a) Derive (38) by equating the rate at which mass flows into and out
of a unit area of shock surface.
(b) Derive (39) by equating force to rate of increase of momentum.
(c) Derive (40) by equating the rate of working of pressure to a rate of
energy change.
6. Suppose that a gas in an unbounded medium is initially at rest but hasa
perturbation in condensation with a nonuniform distribution cr= F(x).
Verify that the solution is
cr= t[F(x - ct) + F(x + ct)].
Sketch when F(x) is given by a triangular "bump" at the origin.
7. Consider gas in an infinite tube with one end closed. If the boundary at
the closed end moves back and forth according to the law = G(t), show e
that the induced motion in the gas is given according to linearized theory
by
U1 = G'(t- ~)
Find the corresponding density distribution.

8. Consider an inviscid fluid with no heat flux (h = r = 0). This exercise and
the next one apply the results of Exercise 14.4.8 to shocks, thereby
obtaining a number of results concerning the jumps of various quantities.
Since fluid crosses a shock, in the notation of Exercise 14.4.7, #- 0, s.;
s;;#- 0. Let v1 denote the vector obtained by projecting the velocity
vector v onto the shock surface. Manipulate the formulas of Exercise
14.4.S(b) to obtain
[psn] = 0, [ps; + p] = 0, [v1] = 0, Hs; + /] = 0, (4la, b, c, d)
where I= e + p-r is the specific enthalpy and -r = p- 1, the specific vo!ume.
(The letter v, normally used to denote volume, is already associated with
velocity.) Also
+ -
p -p + -
+
p -p
- = Sn Sn (42)
To do this, write v = vnn + v1 T, where Tis in the tangent plane to the

shock surface. Take the T and n components of the momentum jump
condition, remembering to use the constitutive equation for an inviscid
fluid. Use the mass jump condition to show that v1 must be continuous;
etc.
9. t(a) Show that for any shock s.; and s;; must have the same sign.
(b) Show that the common sign of s.;and s;;must be negative if fluid
flows across the shock from - to + . (See definitions and figure for
Exercise 14.4.7.) The region R- is then called the frontof the shock.
Sec. 15.4] Uniform Flow Past a Circular Cylinder
551
(c) Extend Exercise 14.4.8 to show that for an inviscid fluid with no
heat flux [psns] ~ 0, wheres.is specific entropy.
(d) Deduce [s] ~ 0.
i(e) The quantity i-- -i-+ is called the shock strength. It can be shown
that this quantity is positive and that
s+ - s- = K(i-- - i-+)3 + ,
where K is a constant and indicates omitted higher order terms.
Show that the speed at which an infinitely weak shock moves
relative to the fluid is the sound speed (iJp/op)1' 2. Start with the fact
that there is an equation of state p = p(p, s). Use (42).
10. Use Lagrange's method of characteristics (see Section 1.2) to derive (31).
15.4 Uniform Flow Past a Circular Cylinder

Determination of the force on an obstacle in a moving fluid has been a
leading problem in the study of fluid motion for centuries. Although much
progress has been made, major aspects of the problem remain unexplained.
In this section we start a discussion of flow past obstacles by examining a two-
dimensional inviscid flow past an infinite circular cylinder.
FORMULATION
We are trying to start with the simplest particular problem that might bear
on the general phenomenon in question. Let us then consider a circular
cylinder
x2+y2=a2, -00 < Z < oo,
immersed in a steady two-dimensional flow of a uniform-density fluid.
Suppose that the fluid moves with uniform velocity Ui far from the cylinder,
so that the velocity vector v satisfies
v(x, y)-+ Ui as x 2 + y 2 -+ oo. (1)
Making an assumption that will be justified toward the end of the section,
we shall write the velocityas the gradient of a potential function </>.The
assumption of uniform density requires that cpbe harmonic if mass is to be
conserved.For p = constant and Dp/Dt + pV v = 0 implies that V v =
v2 4,= o.
We shall employ the method of separation of variables to determine a
harmonic function cpthat satisfies appropriate boundary conditions. Once
this velocity potential cJ,is determined, we can find the pressure p-for the
inviscid version of momentum conservation links velocity and pressure. The
integrated effect of pressure gives the total force on the body.
552
Since we are considering a circular cylinder, we should use a velocity
which depends on polar coordinates r and O:
potential 
2 1
r = (x + y 2 )112, 0 = tan- (~)
(2)
q,(rcos 0, r sin 0) = <l>(r,0).
We shall need an expression for the gradient in polar coordinates. From
Equation 3 of Appendix 15.2 or Exercise 5, this is
(3)
Here, the subscripts on denote partial derivatives. Also e<)= e<')(r,8)
[e<11> = e<11>(r,0)] is a unit vector at (r, 0) which points in the direction of
increasing r (increasing 0). (See Figure 15.12.)
--
-
u
u
-u
FIGURE 15.12. Uniform/low at speed U past a right circular cylinder

of radius a. Shown is a typical pair of unit vectors e<>(r, {!)and e<>(r, {!)
that form the basis of polar coordinates.
We are ready for the final formulation of the problem. According to

Exercise 2 of Appendix 13.1 or other sources, Laplace's equation V2 q,= 0 in
polar coordinates is
r>a. j (4)
From (3) and (1), we find that the requirement of uniform flow at infinity
becomes
I,.....u cos o. ,-
1
<1>
11-+ - U sin O as r-+ oo. j (Sa, b)
Since no fluid can penetrate the cylinder, the radial component of V
must
vanish at r = a, so
I<1>,(a,O)= o. j (6)
553
One more condition must be imposed, because the point with polar coordi-
nates (r 0 , 80 ) also has the coordinates (r 0 , 80 + 2mr), n = 1, 2, .... We
must certainly require that the velocity be the same whether we consider it
at (r, 8) or at (r, 8 + 21r).Thus VCI> must have period 21r in (), or
8 + 21r)= Cl>,(r,
Cl>r(r, 8), 8 + 21r)= Cl>
Cl>e(r, 11(r, 8). (7a, b)
The boxed equations provide a definite mathematical problem. In solving

it, we shall see that another condition must be prescribed if a unique solution
is desired.
SOLUTION BY SEPARATION OF VARIABLES

Although a solution to (4)-(7) can almost be guessed, we shall solve these
equations by the method of separation of variables. Those who have not had
previous experience with this procedure should note that the six steps (A)-(F}
that we apply to this particular separation of variables problem apply with
little or no modification to any such problems. (Compare Section 4.1; see
also Exercise 7.)
STEP A. Assume a product solution:
8) = R(r)0(8).
Cl>(r,
STEP B. Substitute into the governingdifferential equation: From (4),
R"0 + r- 1R'0 + r- 2 R0" = 0 (

R' = dR 0 , = d0) (8)
dr' -de
STEP c . Separate variables.We write (8) in the form
R" + r- 1
R' 0"
r- 2R = - 0. (9)
The variables are now separated so that the left side is a function only of the
independent variable rand the right side is a function only of the independent
variable 8. Reasoning just as under (2.4), we deduce that both sides of (9)
must equal a separation constant k so that we obtain the separated equations
R" + r - R' = kr - 2 R, 0"= -k0. (IOa, b)
STEP D . Determine
permissible
valuesof theseparation
constant
from an
eigenvalueproblem consistingof a separatedequationand suitable homogeneous
boundary conditions. The equations of (10) are both second order, so we
seek a pair of homogeneous boundary conditions. Trial and error indicates
that of the various possibilities, only the conditions stemming from (7) are
appropriate. From (7a)
R'(r)0(8 + 21r)= R'(r)0(8),
so either
R'=O (11)
or
0(0 + 2n) = 0(0). (12)
From (7b)
0'(0 + 2n) = 0'(9). (13)
Since R = 0 gives the trivial solution, there is no alternative to (13). As an
eigenvalue problem to determine k, we therefore take (lOb), (12), and (13).
[We shall consider the possibility (11) separately. J
If k > 0, we write k = 2 , #- 0. The general solution to (lOb) can then be
written
0 = C1 cos O + C2 sin O. (14)
From (12) and (13)

C1 [cos (O+ 2n) - cos OJ+ C2 [sin ((} + 2n) - sin OJ= 0, (15)
C1 [- sin (8 + 2n) +sin 8J + C2 [ cos (8 + 2n) - cos 8] = 0.
(16)
Equations (15) and (16) have only the trivial solution C1 = C2 = 0 unless we
require that the determinant of the coefficients vanishes. By the use of simple
trigonometric identities, this requirement gives
[cos (O+ 2n) - cos 8]2 + [sin (8 + 2n) - sin 0]2 = 0,
or
2 - 2[cos (8 + 2n) cos 8 + sin (8 + 2n) sin 8] = 0,
or
cos 2n = 1 so= 1, 2, 3, .... (17)
When is an integer, (15) and (16) are indeed satisfied for arbitrary constants
C1 and C2 .
If k = 0, the only nontrivial solution of (}Ob),(12), (13) is 0 = 80 , where 80
is a nonzero constant. If k < 0, there are no nontrivial solutions of (1 Ob), (12),
and (13) [Exercise l(a)].
STEP E . Solve the remaining separated equation when the separation
constant takes on the values determined in step (D). When k = n2 the equidimen-
sional or Euler equation (lOa) has as a general solution a linear combination
of r" and r-n, n > 0. When k = 0 the general solution is a linear combination
of 1 and ln r (Boyce and DiPrima, 1963, Sec. 4.4).
Sec. 15.4] Uniform Flow Past a Circular Cylinder SSS

STEP F. Superpose all possible product solutions and attempt to satisfy the
remaining boundary conditions. By "superpose" we mean "form a linear
combination of." To include all possible product solutions in this combi-
nation, we must consider the possibility (II). Equations (II) and (8) imply that
0" = 0 and lead to the conclusion that an arbitrary multiple of(} is a product
solution of (8) and (7), which must be added to the list of such solutions al-
ready obtained. From this conclusion and the results of steps (D) and (E) we
are led to assume for (I>the infinite series
(l>(r,(}) = A(} + B0 + B1 ln r
co
+ L [(Cnr" + Cnr-") cos (n(J) + (Dnr" + dnr-") sin (n9)]. (18)
=1
The condition at infinity (S) is satisfied if
C1 = U; D 1 = O; Cn = Dn = O; n =2,3,4, ....
Boundary condition (6) then requires that
n = 2, 3, 4, ....
A solution to all equations and boundary conditions is thus
(I>= A(}+ U(r + a2 r- 1) cos(}, (19)
as can easily be checked by direct substitution.
INTERPRETATION OF SOLUTION
The presence of the term A(} in (19) is noteworthy. This term is independent
of Ubut depends on the arbitrary constant A, so there are an infinite number
of uniform flows past the cylinder. The velocity vector corresponding to A(} is,
using the polar coordinate gradient formula (3),
V(A9) = Ar- 1
e<8 >. (20)
The corresponding flow is purely circumferential. Its speed approaches zero
as r-+ oo, so the uniform flow at infinity is unaffected by this circulatory
motion.
As a measure of over-all rotation in a flow, Kelvin in 1869introduced the
concept of the circulation K(C) around a simple closed curve C, which he
defined in terms of a line integral as follows:
K(C) = fcv dr. (21)
When v = V(I>one can show without difficulty [as in Exercise l(e)] that
circulation around C change of (I>when C is traversed once
in a given direction in the given direction.
556 Inviscid Fluid Flow [Ch. JS
When <ll is given by (19), K( C) = 2nA for any curve C circling the cylinder
once in the counterclockwise direction [and K(C) = 0 for any curve C that
does not enclose the cylinder].
Our separation of variables approach would thus have given a unique
answer if we bad required as r -+ oo not only that the velocity approach a
constant U but also that the circulation approach a constant 21rA.
Let us return to the question that motivated our investigation: What is the
force on the cylinder? To find the force we must compute the pressure from
the momentum equation for steady motion. From (1.10) with f = O* we
obtain
or (Vq, V)Vq,= -v(~) (22a, b)

The left side of (22) is known, so p can be computed by direct calculation.
Even using polar coordinates, however, this is somewhat unpleasant. A
simpler approach stems from the observation that (22) can be written
so
114> + ~
V 1
2
= constant. (23)
By considering (23) in the limit r -+ oo, we can determine the constant. Our
final result is
-21 Ivi2 +-=-U

p
p 2
1 2
+-Pa:,
p
(p = limp). (24)
(This is a special case of Bernoulli's equation, proved in Exercise 1.7.) The

force F on the cylinder of radius a is given in terms of the exterior normal
n and arc length s by
F =ft ds = - I pn ds = - f2 ..
0
(cos 8i + sin 8j) pa dB. (25)
By using (24) and (19), a little calculation shows [Exercise l(c)] that
F= -pUKj (K = 21tA= counterclockwise circulation). (26)
With flow at infinity in the direction of increasing x, lift and drag are
defined as the force components in the j and -i directions. We have found
a lift proportional to the circulation, and no drag. To put it dramatically,
If the effect of gravity were taken into account, there would merely be an additional lifting
force due to buoyancy (Exercise 9).
557
the lack of drag can be interpreted as predicting that there would be no down-
ward force exerted on your arm if you held it under Niagara Falls. Something
is wrong somewhere.
Suspicion first falls on the assumption v = Vcj,.But it is a central result of
inviscid flow theory that V /\ v = 0, and hence that this assumption is
inevitable, for flow starting from rest-even if gravity is introduced and even
if a varying density is considered, provided that the equation of state is of the
form p = p(p) (Exercise 1.6).
Perhaps there is another uniform flow past a circle that is not revealed
by separation of variables. Perhaps the no-drag paradox holds only for the
completely symmetric circular shape, and more sensible results emerge for
more general shapes. But it can be shown that for two-dimensional uniform
inviscid incompressible flow past any rigid bounded object,
as r-oo, (27)
and (26) continues to give the force (Kutta-Joukowsky theorem).

The three-dimensional counterpart of (27) is
v = Ui + O(r- 3
)
and F = 0, in some ways an even more striking result.

In Section 15.1 we stated that it appears reasonable to assume that shear
stresses are always zero in water and air. This inviscid fluid assumption led
to good results when used in our discussions of the stability of a stratified
layer of an incompressible liquid, and of compression waves in gases. But we
cannot accept the result that drag is absent in the flow past two- and three-
dimensional obstacles. There are two valuable lessons here: (i) that a" reason-
able" assumption may or may not turn out to be a good one; and (ii) that a
given mathematical model may be perfectly adequate in some contexts and
yet inadequate in others.
The predicted absence of drag is called D' Alembert's paradox.* One's first
reaction on hearing such a counterintuitive result is that it was derived as a
consequence of an error in reasoning. Only a mathematical proof can fully
convince one that it is the basic assumptions which are at fault. A proof for
the three-dimensional case is outlined in Appendix 15.1.
Resolution of D'Alembert's paradox requires the study of viscous fluids,
wherein the possibility of shear stresses is considered. This is done in Chapter
3 of II. It is found that however small the viscosity, shear stresses are always
important in thin layers located near solid boundaries and perhaps elsewhere.
The thickness of these layers approaches zero as the viscosity tends to zero,
but for "blunt" bodies there is an effect on drag that does not tend to zero.
A proof for the two-dimensional case was given by Euler in 1745. D' Alembert "rediscovered
or appropriated" it in 1752. Later (1768) "he reasserted it in sensational terms" (Truesdell and
Toupin, 1960, p. 541.)
For streamlined bodies such as airfoils, on the other hand, the thin layers
have little effect on the flow; the drag approaches zero as the viscosity goes
to zero. For streamlined bodies, then, the D'Alembert "paradox" is virtually
a correct result, for the drag on such bodies is very small. Furthermore, the
two-dimensional force formula (26) is the basis for an explanation of the lift
provided by an airfoil.
In summary, the inviscid model provides a good first approximation to the
study of flow past streamlined bodies, but the flow past blunt bodies is grossly
different from the simple predictions of classical inviscid flow theory. The
most important elements in the theory of flow past streamlined bodies are
discussed in Chapter 3 of II. The reader will search in vain, however, for a
definitive treatment of flow past blunt bodies; the main questions here
remain open.
EXERCISES
1. (a) Show that there are no nontrivial solutions of (IOb), (12), and (13)
if k is negative.
(b) Verify by direct substitution that (19) satisfies all required equations
and boundary conditions.
(c) Verify (26).
(d) Use the relation between pressure and speed given by the Bernoulli
equation (24) to show that (26) is plausible.
(e) Verify the relation between circulation and the change of</>that is
stated below (21).
2. (a) We have found that the velocity potential, W(r, 8), describing the
uniform flow around a circular cylinder, is given by (19). Determine
the corresponding stagnationpoints, i.e., the points at which v = 0.
Be sure to distinguish the cases 2aU > A, 2aU = A, and 2aU < A.
Indicate the positions of the stagnation points relative to the
cylinder in each of these instances.
(b) Assume that v(x, t) is differentiable and show that streamlines have
a well-defined tangent except at stagnation points. Conclude that
streamlines can cross only at such points.
3. Noting that the flow at infinity is uniform, that the circle must be part
of a streamline, and that a streamline can branch at a stagnation point,
one can make a surprisingly accurate sketch of the streamlines for this
problem. Do so.
4. Suppose that a velocity potential cf, exists. Show that cf,(P1) - cf,(P0)
equals the line integral of the velocity along any curve joining the points
P 0 and P 1 . A meaning is thus ascribed to the potential function.
5. (a) Suppose that a velocity potential cf,(x,y) exists for the velocity field
v(x, y). Consider the transformation into polar coordinates. Show
that the relation between the radial velocity component V,(r, 0)
Sec. 15.4) Uniform Flow Past a Circular Cylinder 559
and the velocity potential in polar coordinates ~(r, 8) is V, = o~/

or.Do this by direct calculation in the following way. Since e<r>=
cos 8i + sin 8j is a unit vector making angle 8 with positive x axis
and v, = v. e<r>, then
a<t, a<t,.
ax
v, = cos 9 + aysm 9.
Change o<f,/oxand o<J,/oy
into polar coordinates via the chain rule.
(b) Show that the azimuthal velocity component Vi,can be expressed
as Vi,=r- 1(0'1>/09)
by justifying and completing the following
steps:
d<f, i}(f) d()
V9 = V<f,e= ds = iJ() ds = ....
Here d<J,Jdsis the directional derivative in the direction of increas-
ing 9.
i6. Derive (26) by using the momentum transfer equation. (See Exercise
14.2.5.)
t7. Consider potential fl.ow (v = V'1>),past a sphere of radius a, of an
inviscid fluid of uniform density. Orient axes so that far from the
sphere p = a there is uniform fl.owof velocity - Uk, U a constant. Use
spherical coordinates. (See Figure 15.13.) Look for axisymmetric flow
+ u+ u+ u+ uJ uJ
u
(p, ~. 8)
I
I
I
.._ I
' ........
I
FIGURE 15.13. Diagram of the situation discussed in Exercise 1.
(0/09 = O), and show that the problem reduces to the following set of
equations for '1>(p,
<f,):
. 0(P op00>)+ o<J,
op
sID 
2 o(. sID oq,
oct>)
= o,
o'1>
'1>__.-Up cos <J, asp__. oo, - = 0 at p = a.
op
Find the most general solution that emerges from a separation of

variables approach. You will need the following facts: the Legendre
equation for y( q,) is
I
sin
d (. dy)
q,dq, sin q,dq, + ky = O.
Solutions to this equation remain finite for all q,only if k = n(n + 1),
n = 0, I, 2, .... The corresponding solutions are multiples of P.(cos q,),
where the Pn are (Legendre) polynomials of degree n, for example,
P0 (x) = 1, Pix)= t(3x 2 - 1).
8. Find the net force exerted on the sphere by the flow of Exercise 7. Do
this by integrating the local effects of pressure.
9. (a) Show hovr (23) would be altered if a body force term f =-gj were
added to the right side of (22a).
(b) Show that the effect considered in (a) would add a buoyancy term
to the formula for F in (26).
Appendix 15.1 A Proof of D' Alembert's Paradox in the

Three-dimensional Case
Consider three-dimensional irrotational inviscid fluid flow generated by the
motion of a bounded object and suppose that the fluid is motionless at
infinity. Let v = Vet,,where q,is the velocity potential. If the fluid is of uniform
density, V v = 0, <f,is harmonic, and (16.1.18) can be applied. This equation
represents a harmonic function at a point P in terms of an integral over the
boundary oRof a region R that surrounds P. If P is the origin, the equation
reads
q,(O)= - ;1tI[[q,:n (~) - ~ :: ] du (o/on= exterior normal derivative).
With a slight reinterpretation we can write
q,(x) = -
41tff
..!..
BR
[q,!.._(~) - !. oq,]du,
on r r' on (1)
where r' is the distance from x to the variable of integration that runs over oR.
Let R be a region that contains x and is bounded within by S, a surface
containing the obstacle mentioned above; and is bounded without by I,
a large sphere of radius p centered on x (Figure 15.14).
Appendix 15.J] A Proof of D'Alembert'sParadox 561
--..-----
,AX
s
......
---- p
FIGURE 15.14. Region R used in representing the velocity potential

at x for flow past obstacle 0. The vectors n are typical unit exterior
normals to R.
It is not difficult to show (Exercise 1) that (1) implies that
- ff[V<f, o] du+.!_ ff [
ff<f,du+ 4,rp !._(!_)
on r' - !_
1 1
<f,(x)= - 2
4,rp :t :t 4,r s
ocp]
r' on
du.
(2)
In (2) and below, the time t is not explicitly designated, since it is regarded
as fixed throughout our discussion.
The first integral in (2) is the mean value of </>over a sphere of radius p
centered at x. For the moment, let us take for granted the plausible assertion
that this value is independent of x in the limit p --+ oo. We shall sketch a proof
of this assertion below.
That the second integral in (2) vanishes in the limit can be deduced from the
fact that Vcf,--+ 0 at large distances from the origin, because of the assumption
that the fluid is motionless at infinity.
Suppose now that the obstacle moves with a uniform velocity - U. Take
new coordinate axes that move with velocity - U relative to the old ones.
{In the new coordinate system, of course, the obstacle is stationary and fluid
far from the obstacle moves with velocity U.) By using the deductions made
so far, we see that the velocity potential satisfies
<J,(x)= U x ff
..!... [<J,i (;) - ~ o<J,]du.
+ constant + 4n on r r on
5
(3)
Exercise 2 shows that as r = Ixi -+ oo,we can deduce from (3) that
<J,(x)-+ U x + constant - -
F + O(r - 2). (4)
4nr
Here
F=ff~<f>
sun
du= Jfv<t>ndu= ffvndu
s s
is the net outflow of fluid through S. As there are no sources, we must have
F=O. (5)
To prove this, note that the divergence theorem shows that
F- ff V<J,ndu=O.
obstacle
(6)
But the second integral in (6) is zero, since v n = 0 on the stationary obstacle.
From (4), (5), and Bernoulli's equation (4.24) we can conclude [Exercise
2(b)] that
as r-+ oo, (7)
Since we know that the speed and pressure approach their values at infinity
in this fashion, it follows from Exercise 14.2.6 that there is no force on the
obstacle.
No TE. For a two-dimensional body, the estimate of behavior at infinity
cannot be obtained by the argument used above: the velocity potential may
be multiple-valued. One can utilize the Poisson integral representation of the
solution,* or one can proceed from analytic function theory.
It remains to prove the lemma that the mean value of <J,on a sphere
S.,,of radius p is independent of the position x of its center. Denote this
mean value by Mp(x), and let u denote a unit vector. Then
Mp(x + au) - Mp(x) = : fJ<J,du - 4: P ff<J,du

4 P2 Sx+au
2
Sx
= 4 1 ff<<J,*
2 - <t>)du, (8)
1tp Sx
The Poisson integral representation is the two-dimensional version of (16.2.13), with a

logarithmic singularity.
Appendix 15.2] Polar and Cylindrical Coordinates
where </J* is </J evaluated according to the change of variables necessary to

shift the domain of integration in the first integral from S,.+aato S,.. By the
integral mean value theorem, we can deduce from (8) that
Mp(x + au) - Mp(x) = </J*

- <P= V</J
(vector of magnitude a), (9)
where </J*
- </Jand V</Jare evaluated at some point on I -. 0 as
S. Since IV</>
p-. oo, we conclude that
=0
+ au) - M CIO(x)
M CIO(x
for arbitrary a and u. Hence M CIO(x)
is a constant.
EXERCISES
1. Verify equations (2), (3), and (6).

2. (a) Verify (4).
(b) Verify (7).
3. Explicitly carry out the change of variables required in (8) and thereby
obtain an expression for the "vector of magnitude a" mentioned in (9).
Appendix 15.2 Polar and CylindricalCoordinates

Let e<>(r0 , 90 ) be a unit tangent vector at the point (r0 , 90 ) to the "r-
varying" coordinate curve 9 = 90 . Let e<9>(r0 , 90 ) be a unit tangent vector at
the same point to the "8-varying" coordinate curve r = r 0 e<>and e<0>are
to point in the directions of increasing r and 9, respectively (Figure 15.15).
In polar coordinates the components of a vector v(r, 9) are defined to be v<>
and v<9 >,where
(1)
r =r0
FIGURE 15.15. Fixed cartesian unit vectors i and j, and a typical pair
of varying polar unit vectors e<>and e<>.
and e<11
Since e<r> >are orthogonal unit vectors,
v<r)= v. e<>, (2)
In particular, when v = VWthe relations
v<r)=aw' aw
= ,-1 _,
v<B) (3)
or ae
follow almost immediately from the following formula for the directional
derivative of win the direction of the unit vector e:
The following formulas relate polar coordinates with cartesian (unit

vectors i and j, velocity components u and v). If
i = C1 e<r>+ C2 e<11
>,
then
C1 = cos O;
Thus
i = cos (} e<>- sin (} e<11
>. (4)
Similarly,
j = sin (}e<>+ cos (} e<'>, (5)
and
e<rJ= cos (}i + sin 8 j, (6)
e<11
> = - sin 8 i + cos 8 j. (7)
From (1), (6), and (7), since v = ui + vj,
= u cos 8 + v sin 8,
v<> (8)
v<11> = -u sin(}+ v cos 8. (9)
In like manner, using (4) and (9),

u = v<r>
cos (}- v<')sin (}, (10)
v = v<> 11
sin 9 + v<>cos 0. (11)
Appendix 15.2] Polar and Cylindrical Coordinates 565
In cylindrical coordinates, polar coordinates are used in the (x, y) plane,

while the z-cornponent is the same as in cartesian coordinates. For example,
0 , 80 , z 0 ) is a unit tangent vector at the point (r 0 , 80 , z 0 ) to the curve
e<>(r
r = r 0 , z = z0 The components of the velocity vector v(r0 , 80 , z 0 ) are
8
0 , 80 , z 0 ), v<>(r0 , 80 , z0 ), w(r0 , 80 , z 0 )]
[v<>(r
so
+ v<>e<
v = v<>e<> 9>+ wk.
As usual, k denotes the unit vector along the z axis.

CHAPTER 16
Potential Theory
THE THEORY of gravitational potential, and its extension to the theory

of electrostatics and to electrom~gnetic and sound waves, provides fine
examples of the power of mathematical analysis. It also gives rise to a
number of very subtle and challenging problems. In this chapter we shall
discuss a few of the highlights of this subject to give a reader a feeling of the
breadth of its implications and ramifications.
In Section 16.1, using the gravitational potential for motivation, we derive
various results for the equations of Laplace and Poisson. Section 16.2 deals
primarily with the construction and use of Green's functions; electrostatics
provides physical background. The final section treats an inviscid flow prob-
lem that reduces to Helmholtz's equation: the problem of the diffraction
of planar sound waves impinging on a hole in a screen.
16.1 Equationsof Laplace and Poisson

The first major formula in this section, equation (7), gives the gravitational
potential V(x) due to mass that is distributed in various regions throughout
space. We show that if xis outside the mass distributions, then V(x) satisfies
Laplace's equation (3), while if x is within a mass distribution, then V(x)
satisfies Poisson's equation (8). We then turn to a brief study of functions
that satisfy these equations within bounded regions R, with the object of
representing the functions by formulas which involve integrals over the boun-
dary oR of R. For Laplace's equation, the representation leads to the mean
value theorem and to the maximum principle for harmonicfunctions. We also
conclude that (7) must be satisfied by a function which satisfies Poisson's
equation provided that its behavior at infinity is specified in a certain way.
This section closes with some remarks on uniqueness. A theorem-proof
style seems appropriate.
GRAVITATIONAL POTENTIAL OF DISCRETE MASS

DISTRIBUTIONS
As is discussed in introductory physics courses, the gravitational potential

V at a field point x, due to a point mass m located at the source point ;, is
given by
m
V(x)= -G--. (1)
lx-;I
566
Sec. 16.1] Equations of Laplace and Poisson
(Here G is the universal gravitational constant 6.67 x 10-s dyne cm 2 /g 2 .)

By definition, the potential is a scalar quantity whose negative gradient gives
the gravitational force g. Thus from (1) we obtain the inverse square force due
to a point mass m:
g.(x)
av
= - - = -Gm
x.-e.
' ' . (2)
' ox; Ix - 1;13
It is easy to verify [Exercise l(b)] that, except when x = I;, the gravitational
potential satisfies the Laplace equation
3 a2v
v2 v= I - 2 =o. (3)
i=l OX;
In complicated problems, e.g., in the presence of boundaries of all but the

simplest shapes, it is not possible to guess an exact expression for the potential.
It is usually best, in such cases, to work from the differential equation that
the potential satisfies. Preparing ourselves for these cases, we observe that
we can obtain (1) by seeking a solution of the partial differential equation (3)
with a singularity at the point I;. Such a specification is really quite vague,
since there are other solutions [e.g., the partial derivatives of (l)] that satisfy
the same general requirement. One must be more precise in specifying the
nature of the singularity. In fact, to recover (1), one should impose the re-
striction that
limrV= -Gm. (4)
r-+O
The gravitational potential due to a number of mass points m 1 , m 2 , , m.

located at 1;1 , 1;2 , , I;. is
m
V(x) = -G
k=l
L Ix - \ k I . (5)
This function V also satisfies the differential equation (3) (when x ':/:l;k).
To recover (5) from (3), we have to stipulate
lim Ix - l;kI V = -Gmk, k = 1, 2, ... , s . (6)
....
~
As we shall see later, there is also a condition at infinity to be specified.
GRAVITATIONAL POTENTIAL OF CONTINUOUS MASS

DISTRIBUTIONS
If there is a mass distribution of density p(I;), the gravitational potential
is given by
p(I;) dt
V(x)=-G ffflx-1;1. (7)
568 Potential Theory [Ch. 16
Here de is the element of volume; integration is with respect to ;. It is

convenient to take infinite space as the region of integration, with the under-
standing that p =0 in regions devoid of mass. For x within such regions,
the Laplace equation is again satisfied. What about a point that lies within
a mass-filled region?
First, there is a question of the convergence of the integral, but this can
be proved without difficulty. Second, if we wish to calculate the force vector
by formally calculating the gradient of (7), the problem becomes more serious,
of (7) can stillbejustified.However,whenweat-
but formaldifferentiation
tempt to calculate the second derivatives of (7), the formal procedure is
no longer justifiable. Indeed, we find that the gravitational potential (7)
does not satisfy the Laplace equation (3), but satisfies the Poisson equation:
v2 v= 41rGp. (8)
Although the formal derivation of this relationship is quite easy-as we shall

see immediately-the proof of the existence of the second derivatives (for
sufficiently smooth density distributions) is more complicated. [See Jeffreys
and Jeffreys (1962), Sec. 6.04.] We begin the discussion with a famous formula
that relates gravitational flux to mass.
Theorem 1 (Gauss). The flux F of the gravitational force field g = -VV

is defined by
(9a)
Here n is a unit exterior normal to oR, a surface (assumed to be smooth

enough to permit various manipulations) that encloses a region R. The flux F
is related to the total mass M enclosed within oR by the relationship
F= -41tGM. (9b)
A mass point on the surface is considered half inside and half outside.
Proof. Consider first a single point mass. If the surface oR does not en-
close it, then by the divergence theorem we have
(10)
If there is a mass point inside R, say at point P, the above theorem cannot be
applied. But let us draw a spherical surface S of small radius e, centered at
the mass point P (Figure 16.1). *
The use of a small sphere to exclude a singularity is a very useful device. We shall use it
again subsequently.
Sec. 16.1] Equations of Laplace and Poisson 569
oR
FIGURE 16.1. UseofasphereSofsmallradius e to exclude a singular-
ity point P. The vectors n are unit exterior normals to the region enclosed
bys and aR.
Since (10) holds if it is applied to the region between Sand oRwe have
ffg
oR
n du + ffg n du = 0.
S
In this equation, the unit normals n to the surfaces oRand Sare both directed
outward from the region between these surfaces; in particular, the normal
on S points toward the mass point. On the surface S, the law of inverse
squares enables us to calculate the flux [Exercise 3(a)], and we immediately
obtain the desired result (9b) for a single mass point. By the principle of
superposition, it also holds in general. O
Theorem2. If x is in a domain that contains continuously distributed

mass of density p, then the potential V(x) of (7) satisfies Poisson's equation
(8).
Proof. We write (9) as
ffg n du=
oR
-4nG IIIp dt.
R
If the divergence theorem may be applied to the left-hand side of this equation,
we have
Jff (V g + 4nGp) dt = 0,
R
(11)
for an arbitrary region R. Thus, for any point with a neighborhood in which
the integrand is continuous, we arrive at the desired result,
-v g = v2 v = 4nGp, (12)
by using the Dubois-Reymond lemma (Section 4.1). 0
THEOREMS CONCERNING HARMONIC FUNCTIONS

Functions satisfying the Laplace equation in a domain PJare said to be
harmonic in PJ.For example, as stated in (3), the function
r -1= Ix-~ ~,-1 (13)
is harmonic in the variable x, with ; fixed, in any domain not containing

the point !;. This function is also harmonic in the variable !;, with x fixed,
in any domainnot containingx.
We shall now prove three theorems about harmonic functions. The second
and perhaps most important of these asserts that the value of a harmonic
function <f:,
at x is the mean valueof the samefunction overany sphericalsurface
centeredaroundx (and within the domain of the function). If </>is interpreted
as steady state temperature (Section 4.1) this theorem is physically reasonable
-for if the temperature at x were not equal to its average, one would expect
heat to flow, in order to "smooth out" the heat distribution. If</>is given the
probabilistic interpretation of Section 3.3, then the theorem states the highly
plausible result that for a particle starting at x the probability of leaving a
specified portion of the boundary is equal to the average of the correspond-
ing probabilities for particles starting on a sphere around x (and within the
domain of the function).
Theorem 3. A harmonic function <f:,(x) is determined within a region R

by its values and those of its normal derivative on the boundary aR of R.
The formula is given in (18). (We assume sufficient smoothness to carry out
the steps of the proof.)
Proof. We start with the symmetric Green's formula
III(uV v - vV u) d-r: ~ (uanav -

R
2 2
=
OR
V aau)du,
n
(14)
where a;an denotes a derivative in the direction of the exterior normal. For-
mula (14) is an immediate consequence of the divergence theorem. (See
Exercise 3.4.8.) In (14) we set
u=<f:,, v = r-1, (15)
and regard I; as the variable of integration. There is a singularity of v =

Ix - I; I - 1 at the point P with coordinates x.To exclude it, we again surround
P with a sphere S of small radius e (Figure 16.1). We then apply (14) to the
region between Rand S. Since the left-hand side of (14) vanishes identically,
fJ
ilR
!._
[<1>
an r
(!)-!r a<f:,]
an
da + fJ
s
[<1>!...(!)-!r a<f:,]
an r an
du= o. (16)
Sec. 16.1] Equations of Laplace and Poisson 571

On S, the normal derivative satisfies a;an= -afar, and hence the second
integral becomes
ffs [2"'+ -1 -a<J,]

du f f
2
= " " ["'
26 e2 sin ~ d~ dr,,
+ -1 -a<J,] (17)
6 ar
6 o o 6 or
where ~ and r, are the polar and azimuthal angles of spherical coordinates.
If we now let the radius e of the sphere S approach zero, we obtain, from
(16),
ff[<J,~On (!)
iJR r
_! 0
"'] du+ 4rr<J,(x)= 0,
r On
or
<J>(x)
= _ _!_ff ~ (!)_! 0"'] du.
[<t> D (18)
41tiJR On r r On
The representation (18) is quite useful-indeed, we have already used it
in the proof of D'Alembert's paradox (Appendix 15.1).
Theorem4 (Mean value theoremfor harmonicfunctions). Let x be the

center of a sphere R of radius p. Suppose that <Pis harmonic in a domain
containing R. Then
</)(x)= 4 1 2
1tp
</>du.
iJR
Proof. Applying (18) and using the fact that o/iJn

= o/oron oR,we obtain
1 ,ir 1 ,ir o<J>
<J,(x)= - 2 lJ' </>du
41tp iJR
+ -4 lJ' -;--du.
1tpilR un
Employing the divergence theorem, we see at once that the integral in the
second term vanishes, for it equals
The mean value theorem will now be used to show that a nontrivial har-
monic function cannot have an interior maximum or minimum. (Compare
the remarks in Section 4.1 concerning the maximum principle for the heat
equation.)
Theorem5 (Maximumprinciplefor harmonicfunctions).Let q,be a function

that is harmonic within a region R. Then either </)is identically constant,
or the maximum and minimum values of</>occur on the boundary of R.
Proof. Suppose that </>has an interior local maximum M at a point ;,

so that Icp(x) I :::;;;
M for Ix - !; I :::;;;
8, 8 sufficiently small. Apply Theorem 4,
taking !; as the center of a sphere of radius 81 , 81 :::;;;8. This gives
cp(;)=~ffdu
41t8
:::;;;~ffMdu
41t8
= M.
1 1
But </>(;)= M, so the equality sign must be used in the above equation. Thus
[Exercise 4(a)]
</>(X)=M, Ix-;! 58. (19)
Now take new circles, centered on Ix - !; I = 8, and

repeat the argument.
The desired result follows. The proof ruling out internal minima is essentially
the same. 0
INTEGRAL REPRESENTATION FOR THE SOLUTION TO

POISSON'S EQUATION
Again let us employ Green's identity (14). As before, we set v = ,- 1 ,
but now we put u = V, a function satisfying Poisson's equation (8). We
obtain the following integral representation for a function V that satisfies
Poisson's equation within a bounded region R [Exercise 4(b)]:
v= - _!:_
4,r
ff [v~On (!)-! anav]du - G fffe_d-r.
cR r r R Y
(20)
The last term in (20) is the volume integral (7), giving the gravitational
potential due to the density distribution p. The other term represents con-
tributions from the surface S, which can also be given physical interpreta-
tions. Such interpretations are quite real when we consider problems in
electrostatics but are rather artificial for the gravitational problem. In the
latter case the surface integral should really be regarded as giving contribu-
tions from mass distributions outside of R. We shall therefore consider the
case where the surface oR-+ oo. In the limit we are dealing with mass distri-
buted in an unbounded region, so it is expected that (7) would follow.
If (7) holds, we can make formal appraisals of the magnitudes of V and
VVat infinity, and obtain
v- ,- 1, vv- ,- 2 , as r-+ oo. (21)
The conditions (21) should be regarded as stipulations on the solutions we
seek. Otherwise, there is no assurance of uniqueness; in fact, harmonic
functions such as r" cos n9 could be added to V. With conditions (21) imposed,
it is a matter of simple calculation to show that indeed the surface integral
in (20) approaches zero as the surface oRrecedes to infinity [Exercise 5(b)].
We thus obtain the following result.
Sec. 16.JJ Equations of Laplace and Poisson
573
Theorem 6 (A partial converse of Theorem 2). Let V satisfy Poisson's

equation within a region R. If R is bounded and has a sufficiently smooth
boundary, then the representation (20) applies. If R is unbounded and V
has behavior (21) for large r, then (7) holds.
UNIQUENESS
Formula (18) suggests that a harmonic function is completely specified

in a domain once the value of the function and its normal derivative are given
on the enclosing surface. This is certainly true, but the function is actually
uniquely determined by either the value of the function or its normal derivative.
(In the latter case, there is still the freedom of an arbitrary additive constant.)
If a harmonic function is sought with prescribed boundary values, it is
customary to speak of the Dirichlet problem or the first boundary value
problem for harmonic functions. If the normal derivative of the function is
prescribed on the boundary, the literature refers to the Neumann problem
or the second boundary value problem.
For both Dirichlet and Neumann problems (or a combination of them)
uniqueness can be proved by following an argument similar to that used to
prove uniqueness for heat conduction (Exercise 6). Alternatively, the maxi-
mum principle can be employed (Exercise 7). Extension of the proofs to the
Poisson equation is very simple (Exercise 8).
EXERCISES
1. (a) Verify (2).

(b) Verify that the function V defined in (1) satisfies (3) except when
x=;.
2. Supply sufficient conditions to permit two differentiations to be taken
under the integral sign in (7) and thereby prove that V(x) satisfies La-
place's equation when x is in a domain that is devoid of mass.
3. (a) Complete formal verification of Theorem 1.
(b) Fill in the details that are required to complete the proof.
4. (a) Verify (19) by an argument similar to that used under (4.1.15).
(b) Verify (20).
5. (a) Provideformalcalculationswhichindicatethat (7)implies(21).
(b) Show that if the conditions of (21) are satisfied, then the surface
integral in (20) approaches zero as the surface approaches infinity.
6. (a) Prove a uniqueness theorem for the Dirichlet problem by following
the general approach that leads to (4.1.17). (The proof in the
present case is actually easier.)
(b) For the Neumann problem, show that two solutions can differ
at most by a constant. Give a physical argument which shows that
this is a reasonable result.
(c) Prove a uniqueness theorem for the problem wherein a harmonic

function is sought with prescribed values on part of the boundary
and a prescribed normal derivative on the rest of the boundary.
7. (a) Use the maximum principle to prove that the Dirichlet problem
has at most one solution.
(b) Use the maximum principle to prove that a solution to the Dirichlet
problem depends continuously on the boundary data.
8. Extend the various proofs so that they apply to the Poisson equation.
16.2 Green's Functions

In this section we construct a variety of Green's functions. Roughly
speaking, these give the effect at a field point x of a unit source (of heat,
electrostatic potential, gravitational potential, etc.) that is concentrated at l;.
Explicit formulas for Green's functions are given for free space and for
situations in which a planar or spherical boundary is present. We show
how the corresponding general boundary value problems can be solved,
once these functions are determined.
GREEN'S FUNCTION FOR THE DIRICHLET PROBLEM
To deepen our understanding of harmonic functions and to provide a
useful method for solving specific problems, we should introduce the concept
of Green's function. This is best done in the context of electrostatics.
Consider a cavity inside a perfect conductor, at potential zero(" grounded").
If there is a positive point charge e at an internal point I; (the source point),
there would be induced negative charges on the surface. These would be
distributed in such, a manner that the potential on the surface remains
zero. The potential at any other internal point x (a field point) is then a
we1I-defined quantity, and we shall denote it by eG(x, l;). G is a Green's
function. As is typical, it can be regarded as the " effect " (potential) at x due
to a unit" cause" (charge) located at I;. Near the point l; we would expect the
potential directly due to the charge e to dominate effects of surface charges,
and therefore we anticipate that
as r = Ix - l; I -+ 0. (1)
Moreover, the potential (up to a factor e)
G1(x, l;) = G(x, l;) - r- 1
(2)
must be due to the induced surface charges only. Thus G1(x, l;) can have no
contributions from interior singularities; this function must be harmonic
everywhere inside the cavity.
The Green's function G(x, !;)for a domain Dis thus mathematically defined
as a solution of the Laplace equationfor x '=I=l; which (i) vanishes on the boundary
of D and (ii) satisfies the condition (1). This function is also known as the
Sec. 16.2] Green's Functions
575
fundamental solution (of the Laplace equation) for the domain D and repre-
sents a generalization of the function 1/r, which is the fundamental solution
for infinite space.
The existence of the Green's function is not easy to prove. Clearly, its
existence proof is equivalent to that for the function G1 (x, ~), which is har-
monic in D and satisfies the boundary condition G1(x, ~) = -1/r. This
special boundary condition does not make matters any better, so the existence
proof for G is in general no easier than the existence proof for the Dirichlet
problem itself.
REPRESENTATION OF A HARMONIC FUNCTION USING
GREEN'S FUNCTION
We now show that the Green's function provides a representation for a
harmonic function which, in contrast to (1.18), does not use values of both
the function and its normal derivative on the boundary. We assert that the
value of a harmonic function at any interior point of a region R may be
expressed by
fj
<f>(x)= - _.!._ oGdu. (3)
41tiJR On
To prove this relationship, we simply follow the same steps as in the proof
of Theorem 1.3, with 1/r replaced by G. Since G vanishes on the boundary,
the second term on the right-hand side of (1.18) is now replaced by zero.
Equation (3) thus defines the harmonic function in terms of its values on
the boundary, in agreement with the uniqueness theorem.
SYMMETRY OF THE GREEN'S FUNCTION
In free space the potential at a point x due to a unit charge placed at the
point ~ is the same as the potential at ~ due to a unit charge at x, for the
function 1/r is symmetric in (x, ~). This is true also for potentials inside a
cavity in a conductor, so we expect that
(4)
To prove this reciprocity property, we again follow the same steps as in the
proof of Theorem 1.3. In Green's identity (1.14) we use the functions
U = G(X1,~). v = G(x 2 , ;),
where ; is the variable of integration. The detailed calculations are left to
the reader (Exercise l).
EXPLICIT FORMULAS FOR SIMPLE REGIONS
Green's function provides a key to the solution of the boundary value
problem for the harmonic function. Unfortunately, it is not possible to write
down the Green's function except for the simplest kind of regions. We shall
Potential Theory [Ch. 16
now consider a pair of related cases in which such a function is known:

(i) the semiinfinite space bounded by an infinite plane, and (ii) the region
inside or outside a sphere. The geometry in these two cases is related by an
inversion process.
Infinite plane. If there are two electrical charges of opposite sign situated
in the position of mirror images relative to a plane x 1 = 0, the potential on the
plane is zero. Thus the Green's function for x 1 ~ 0 is
1 1
(5)
G(x,I;) = Ix - I;I - Ix - I;' I '
where I;' is related to I; by
e2.e;)= <-e1,e2,e3).
ce~. (6)
(The points I; and I;' are mirror images with respect to the plane x 1 = 0.)
The solution of the Dirichlet problem for the right half-space, according to
(3), is then found to be (Exercise 2)
(7)
where the integration (with respect to I;) is taken over the infinite plane
e1=o.
Sphere. Consider a sphere with center O and radius a. In this case the
" image" I;' of a source point I; turns out to be defined by the relationship
11;
I 11;
I = a2, (8)
plus the requirement that point and image lie on the same radial line. If
the field point x is on the sphere, then the potential at x must be zero, by the
definition of the Green's function. In this case, the geometric situation is
depicted in Figure 16.2. The triangles OQ'P and OPQ are similar so that
lx-1;'1 a
Ix-;! =m (9)
To obtain the zero boundary condition, then, the charge placed at Q' must be
opposite in sign, but smaller by a factor a/II; I, compared with the charge
placed at Q. Thus, for a point x outside the sphere, the Green's function is
1 a/11;I
G(x,I;) = Ix - I; I - Ix - I;' I ' (10)
w bile for a point x inside the sphere it is

, J 1;1/a (11)
G(x,I;)= Ix -1;'1- Ix- !;I.
Sec. 16.2] Green's Functions
577
FIGURE 16.2. Source point Q and image point Q' for a conducting
sphere of radius a. The field point x is on the sphere, so the potential there
must be zero.
From these functions it can be shown after some calculation (Exercise 4)

that the solution for a point x outside the sphere is
cj,(x)= lxl2 -a2 ff cj,(f,)dt13 (12)

4,ra Ill =a IX - f, I
and (Exercise 5) for a point x inside the sphere is
cj,(x)= (a2 - Ix 12) ff cj,(f,')~t13. (13)
41ta ll'l=a Ix - f, I
Equations (12) and (13) are referred to as Poissonintegralformulas.
WIDESPREAD UTILITY OF SOURCE, IMAGE, AND
RECIPROCITY CONCEPTS
The general concepts that we have been discussing for the Dirichlet
problem are widely applicable. We already have used them, in Chapter 3,
for the diffusion equation. There, as here, we solved general problems with
integrals involving source solutions. We also used the method of images.
In particular, we mention that the reciprocity property (4) also holds for the
fundamental solution u0 (x - e.t) of the heat equation. [Recall from Section
3.4 that this solution is the temperature at point x, time t, which is present
due to a unit amount of heat that is released at time zero from point e.
As can be seen from the explicit formula (3.4.4), the value of u0 (x - t) e.
e
remains the same if x and are interchanged.]
In Chapter 3 and the present chapter, then, we have seen reciprocity
relationships that are inherent in heat flow, gravitational attraction, and
electrostatics. But this is far from a complete picture. To mention one other
example, the Nobel Prize in chemistry was awarded to L. Onsager in 1968
for his contributions to irreversible thermodynamics, especially for his
discovery of certain reciprocity relationships in the transfer coefficients.
We now extend our present discussion of Green's functions by taking into

account different boundary conditions for Laplace's equation and by dealing
with a new equation,
k>O. (14)
This Helmholtz equation, as we shall see, arises in wave propagation problems.

Thus we will have presented examples of the use of source solutions in the
three great areas of classical mathematical physics-wave propagation,
equilibrium (governed by Laplace's equation), and diffusion.
GREEN'S FUNCTION FOR THE NEUMANN PROBLEM
The boundary condition on the Green's function for the Dirichlet problem
enabled us to obtain the solution (3), which depends only on the boundary
values of <J,,not on its normal derivative. There is an alternative to the
choice of the boundary condition. We may choose to define a Green's
function of the second kind H(x, !;) satisfying the boundary condition
on the boundary. (15)
To be more precise, besides the condition (15), the function H(x, !;) with !;
fixed and x the running variableis a solution of the Laplace equationin an infinite
domain D, except at the point!;, in whoseneighborhoodwe have
r= Ix-!;!. (16)
If we follow through the reasoning of Sectio~ 16.1 (Exercise 6), we shall find,
instead of (3), that
ff
<J,(x)= - 1 H-;-a<J,
41tDR un
d<1. (17)
A minor observation should be made here. If His a Green's function of

the second kind, then H + constant is another. However, formula (17) is
not influenced by this change, since the extra term is zero (as in the proof of
Theorem 1.4). We thus have a formal solution of the boundary value problem
of the second kind. We see that the harmonic function <J, is completely speci-
fied by its normal derivative on the boundary. Bear in mind, however, that
two functions which differ by a constant have the same normal derivative.
The physical interpretation of the Green's function of the second kind
best emerges in the context of fluid dynamical examples. In Section 16.3
we shall discuss such an example in some detail.
The explicit construction of the Green's function of the second kind is also
possible in the presence of simple boundaries like a plane or a sphere. In the
Sec. 16.2] Green's Functions 579

case of a plane x 1 = 0, the method of images may again be applied (Exercise
7) to obtain the solution
1 1
H(x,;) =-lx---;I + Ix- ;'I' (18)
where; and!;' are image points related by (6). Here, unlike in the Dirichlet
case (5), one has two terms of the same sign. The reader should work out the
Green's function for the case of the sphere.
GREEN'S FUNCTION FOR THE HELMHOLTZ EQUATION
The concept of the Green's function can be extended to other partial differ-
ential equations. Consider the Helmholtz equation (14). In spherical co-
ordinates (p, <f,,0) it is
; 2 [:p(p:p)+ ~]u+
2
k
2
u= O, (19)
where
~=sin
1 a(. a) 1 a
cpa4>sm q,a4>+ sin q,ao 2
2
2 (20)
If the solution has spherical symmetry, then ~ =0 and we may rewrite

(19) as
d2
d2 (pu) + k 2 (pu) = 0. (21)
p -
This constant coefficient equation for pu yields the solutions
(22)
Either of the solutions (22) corresP.onds to the fundamental solution r - i

in the case of potential theory (k-+ 0). We may therefore construct two
Green's functions,
(23a)
and
-iklx-~I
G
c2> ) e
(x, !;, k "" -Ix---!;-1. (23b)
These are spherically symmetric solutions of the Helmholtz equation (14),

except at the point !;, and they satisfy the appropriate boundary condition
G -+ 0 as r -+ oo.
EXERCISES
1. (a) Prove (4).
2. (a) Verify that (5) has the required properties of a Green's function.
(b) Derive (7), by showing that an integral over an infinite hemisphere
is negligible.
3. (a) Verify (9).
(b) Show that (10) and (11) are the appropriate Green's functions.
4. Derive (12).
5. Derive (13).
6. Derive (17).
7. (a) Verify that (18) gives the Green's function for the Neumann prob-
lem in a half-space.
(b) Generalize to the spherical case.
16.3 Diffraction of Acoustic Waves by a Hole

Light waves can propagate through a hole in a wall only as a narrow
beam, whereas sound waves can propagate rather freely. The latter case is
said to involve the diffractionof waveswith wavelength much longer than the
size of the hole. We shall now consider this phenomenon with the aid of a
Green's function for a wave of arbitrary length. We shall see that the phenom-
enon of diffraction is most interesting when we are dealing with wavelengths
comparable to the size of the hole. For waves of such lengths we derive the
characteristic diffraction pattern, alternating rings of sound and silence.
FORMULATION
Consider acoustic waves governed by the following linearized equations

of inviscid hydrodynamics, (15.3.9) and (15.3.10):
00' +
at
t axj = 0,
i=1
OVj (la)
OVj 2 00'
-= -Co- (lb)
at axj
Here c0 is the speed of sound for the medium at rest, and v(x, t) is the velocity
of the medium at point x, time t. The condensation s(x, t) (dimensionless
density perturbation) is related to the density* u(x, t) by the formula
O' = O'o(l + s). (2)
We are reserving the letter p, usually used to denote density, for a spherical coordinate.
Sec. 16.3) Diffraction of Acoustic Waves by a Hole 581
where a 0 is the density of the unperturbed medium. If we eliminate vi from

(la) and (1b), we obtain the wave equation for the condensations:
iPs 3as 2
,:=c~
at
r-2
i=laxi
(3)
At a solid boundary the normal component of velocity vanishes. Hence,

using (lb}, we have
OV
0 = - n = -c 20 n Va = -c 2 aa
0 -
at an
That is,
as=0. (4)
an
Equations (3) and (4) may be used as the basis for the study of acoustic
waves.
Alternatively, if we introduce the velocity potential <J,
such that
a<J,
V=-, (5)
I axi
we may derive, using (1b) and absorbing a function of t into <J,,
<1= -co-2 at/>

a, (6)
Hence (la) leads to the wave equation for <J,:

a2</> 3 a2"'
a12 = c~ i~l axf. (7)
The condition
=o
a<J, at a solid boundary (8)
on
then follows easily from the definition of cp.We shall use (7) and (8) as the
basicequationsinsteadof (3)and (4).
To consider waves at a particular frequency <.o,we study solutions of (7)
that have the form
q,= e- io>'t/l(x). (9)
The spatial function t/1satisfies the Helmholtz equation

ai
k2=- (10)
c~
This is to be solved subject to the boundary condition
ol/1= o (11)
on
at a solid wall.
No TE. From (10), k is the wave number and 2rc/kis the spatial period
of the waves under consideration. [Compare the discussion of (12.2.9.))
Let us consider plane waves approaching a screen x 3 = 0 with a hole
(Figure 16.3). We assume that the waves below the screen will be modified
by a negligible amount. This idealization is equivalent to simulating the
incoming plane wave at the hole by a vibrating piston. That is, we assume
that the motion at the hole is given by
(12)
where vis a given function of x 1 and x 2 We are interested in obtaining a
solution for the half-space above the screen. At the bounding screen, then,
Screen
ttttt
Incoming waves
FIGURE 16.3. Diffraction of waves by a hole in a screen. (A quarter

of the screen is shown.) The effect at the field point x is caused by the
superposition of influencesfrom source points l; that are distributed over
the hole.
Sec. 16.3] Diffraction of Acoustic Waves by a Hole
the exterior normal points in the direction of the negative x 3 axis. Thus, over
the hole we have the boundary condition
o<J, o<J, -i<ot
(13)
- = -v 3 = -ve
on= - -
OX3 '
and hence, by (9),

ot/1
-=-v. (14)
on
One other physical consideration must be introduced. We are interested
in effects caused by the waves arriving at the screen from below and from no
other cause. Such effects must be a superposition of outward traveling waves
only. We thus impose the condition that the above italicized phrase holds.
(Such a radiation condition has already been encountered in the discussion
of longitudinal elastic waves in Section 12.2.)
Recapitulating, we present the final formulation of a mathematical prob-
lem that represents diffraction of plane acoustic waves by a hole in a screen.*
We shall designate by n the set of (x 1 , x 2 ) coordinates that correspond to
the hole.
(10)
ot/1
-=-v; (14)
on
(11)
Solution must be a synthesis of outward-moving waves.

k 2 and v(xi, x 2) are given.
SELECTION OF THE APPROPRIATE GREEN'S FUNCTION

In (2.23) we have recorded solutions to Helmholtz's equation (10) that
possess spherical symmetry. Using (9), we are led to the following basic
source solutions for the velocity potential <J,:
p- 1
exp(-irot + ikp) = p- 1
exp[ik(p - ct)], (15a)
p- 1
exp(-iwt- ikp) = p- 1
exp[-ik(p + ct)]. (15b)
Here p is the spherical coordinate, which measures distance from some
origin, and c = w/k.
Incorporation of the various subtleties connected with this problem remains a matter for
continuing study. A recent paper, with references to earlier work, is A. V. Chinnaswamy and
R. P. Kanwal's "Uniform Asymptotic Theory of Diffraction by a Plane Screen by the Method
of Boundary Layers," STAM J. Appl. Math. 23, 339-SS (1972).
The first of the above solutions corresponds to outgoing waves, the second
to incoming. Both sets of waves move with speed c, and in both cases the
effect can be regarded as due to a "cause" concentrated at the origin.
We wish to consider "causes" that are located at any point !; in the hole,
and we have stipulated that only outgoing waves should be considered. Thus
we make use of the Green's function having the singularity
eiklx-tl
Ix- !;I.
Since the problem involves given values of the normal derivatives, we choose
the Green's function of the second kind,
eiklx-fl eiklx-f'I
Jl<1l(x, 1;,k) = Ix- 1;1+ Ix- !;'I. (16)
Here!;' is the reflection of!; in the wall, so if xwis a point on the wall, xw - I; =
-(xw - !;'). By choosing the positive sign in (16), then, we have used the
method of images to construct a Green's function whose normal derivative
vanishes on the wall.
DERIVATION OF THE DIFFRACTION INTEGRAL
We shall now show that the solution is given by the "diffraction integral"
1
lj,(x) = Jf
47t n
Hc1 >(x,!;, k) ~t/1
un
du, (17a)
where the variable of integration is ; and integration is carried out over the
hole in the screen. Here olj,/on
is assumed known. On the wall, Hc1 >= 2eikr/r,
where r = Ix - I; I Thus the integral (l 7a) can also be written in the form
lj,(x) = -
1 ff
- -du.elkr olj, (17b)
21t n r on
The physical interpretation of this formula is not difficult to discern. The
amplitude of the wave at any point x is obtained by a superposition of waves
generated from a "smear" of sources at the hole, of strength per unit area
(21t)- 1 olj,/on.Each source point!; is associated with its own r = lx-1;1,
so each effect acquires an appropriate phase delay during its propagation
to the field point x. The fact that complicated wave phenomena can be re-
garded as resulting from the superposition of waves from point sources, with
appropriate phase differences and concomitant cancellation and reinforce-
ment, is known as Huygen's principle.
The form of the solution (17) may be anticipated from (2.17), the solution
of the boundary value problem of the second kind for the harmonic function.
Derivation of (17) requires a more subtle discussion, however, because of
conditions at infinity.
If we consider a domain bounded by the wall and a hemisphere S centered
Sec. 16.3] Diffraction of Acoustic Waves by a Hole 585
at some point on the hole in the screen [Exercise 3(b)], we may readily derive
the following representation of the solution:
l{,(x) = - 1
41t
ffH
n
0 J- oi/1du+
on
-1 ff(H<
41t s
1 oi/1- i/1--
>-
on
aH<))
on
1
du. (18)
When the radius of the hemisphere Sis made infinitely large, it is expected-
and it must be verified-that the integral over the hemisphere approaches
zero. Equation {l 7a) then follows.
To verify that the contribution from S vanishes in the limit, let us note
that the largest terms in H<1 J and i/1have the following behavior:
ikr eikr'
n<1),..,:_ i/1--.r' (19a, b)
r '
Here r is the distance from a point ~ on the hemisphere S to a point x at a
finite distance from the screen, and r' is the distance from x to the image
point ~, {see Figure 16.4). Thus, in the integrand n<1 l(oi/l/on) - ljl(offOljon)
that occurs in (18), the largest term arises from the differentiation of the ex-
ponential factor and is of the form
eikr eikr' . (or or') (20)
--lk ---
r r' on an '
3
other terms being O(r- ) as r-+ oo.* Some calculation (Exercise 4) shows
that
or- or'=o(!.), (21)
an on r
and hence the integrand over the spherical surface is O(r - 3 ). Since the area
of the hemisphere is O(r 2), the desired result holds.
Hemisphere S
I
I
I
I
I
I
\
\ //t'
',, 'ii,
' .,./
',, _,.,,,.,.,
------
Image hemisphere
FIGURE 16.4. Source point; and its image;'.
See Appendix 3.1 for the definition of the O symbol.

586
APPROXIMATE EVALUATION OF THE DIFFRACTION

INTEGRAL
The diffraction integral (17b) can be simplified for easier evaluation.

We shall employ spherical coordinates (p, <J>,8) for the field point x, with the
origin at a convenient point in the hole (Figure 16.3). In (17b) we are only
concerned with source points ; that lie in the plane of the screen; we take
the spherical coordinates of such points to be (p 1 , n/2, 81). We shall approxi-
mate the integral under the assumption that p ~ p 1 ; i.e., we consider only the
field points which are many hole "diameters" away from the hole.
Using the law of cosines, we have
r = (p
2
+ Pi- 112
2ppl cos V1) = p[l -: 1
cos VI+o(:rJ' 1
(22)
where VIis the angle between x and!;. Unit vectors in the direction ofx and!;
are
i cos 8 sin </>+ j sin 8 sin cp+ k cos cp and
respectively. We obtain cos VIby taking the scalar product of these vectors:
cos VI= sin (cos 8 cos 81 + sin(} sin 81) = sin</>cos (8 - 81 ). (23)
Furthermore, in most cases, the hole is sufficiently small so that it is justifiable

to regard OVl/onas a constant v in the integrand. With this assumption we
obtain, using (22) and (23),
Vl(P,cp,8) ~
exp(ikp) v
-
ff
exp[ - ikp 1 sm
. cpcos (8 1 - 6)] du. (24)
p 2n n
In the case of a circular hole of radius a, we know that du= p 1 dp 1 d81,

p 1 varies from Oto a, and 81 from Oto 2n. We can set 8 = 0 in the integrand
without affecting the answer. But (3.1.29) gives an expression for the resulting
integral in terms of J O , the Bessel function of order zero:
1 2x
2
n Ia exp( - ikP1 sin</>cos 8 1) d81 = Jo(kp 1 sin cp). (25)
Hence (Exercise 5) we obtain the final result
.,.( ,1.. ) ~ 2 J 1(ka sin </>)

eikp.
"' p, "'' 8 va k asm"'. ,i, p
(26)
Here J 1 is the Bessel function of order 1 and is related to J 0 by
(27)
Sec. 16.3) Diffraction of Acoustic Waves by a Hole
We see from (26) that the wave amplitude vanishes on those directions <p
for which J 1(ka sin)= 0. The direction <p= 0 is an exception, since
lim J 1 (ka sin)=! .
.....oka sin <p 2
Consequently, "silent rings" are defined by the positive zeros of J 1 :
ka sin <p= 3.832, 7.016, 10.173, 13.324, etc. (28)
It is readily shown that the intensity is a maximum at <p= 0, as one would
expect from physical considerations.
Note that for there to be a real angle <psatisfying any of the equations (28),
ka ~ 3.832. That is, the wavelength 21t/k must be smaller than 2na/3.832 ~
0.8(2a). There are no "silent rings" unless the wavelength is somewhat less
than the diameter of the hole. On the other hand, if the wavelength is too small,
the rings will be too close together for satisfactory resolution.
Many kinds of waves are governed by the wave equation (7) with the con-
dition of vanishing normal derivative (8), and the above calculations apply
with little or no modification. To demonstrate the wave nature of matter,
G. P. Thompson sent electron beams through a small hole and obtained
diffraction patterns of the kind just calculated.
EXERCISES
1. Show that introduction of a velocity potential [as in (5)) is justified in a
problem governed by (I).
2. Write a paragraph or two on the advantages and limitations of the
assumption (12). Can less restrictive assumptions be made without
entailing too much additional effort?
3. (a) Verify that the normal derivative of n<1 > vanishes on the wall.
(b) Derive (18).
4. Verify (19), (20), and (21).
5. (a) From the series definition of the Bessel functions in Exercise 3.1.21,
prove that (27) holds.
(b) Obtain (26).
6. Discussthe locationand magnitudesof the variousmaximumintensities
given by (26).
7. Show that the image point I;' is just one of the possible choices in (I 9b);
any point behind the screen may be used to accomplish our purpose.
Bibliography
GeneralAppliedMathematics
CoURANT,R., and D. HILBERT.(1953) Methods of Mathematical Physics. Vol. 1. New
York: lnterscience Publishers, a division of John Wiley & Sons, Inc. 561 pp.
"Mathematical methods originating in problems of physics are developed and
the attempt is made to shape results into unified mathematical theories." (Volume
2 does not have the classical style of Vol. 1 but is a compendium of important
results in the theory of partial differential equations.)
FRIEDMAN, B. (1956) Principles and Techniques of Applied Mathematics. New York:
John Wiley & Sons, Inc. 315 pp.
"This book was written in an attempt to show how the powerful methods
developed by ... abstract studies can be used to systematize the methods and
techniques for solving problems in applied mathematics."
GENIN,J., and J. S. MAYBEE.(1970) Introduction to Applied Mathematics, Vol. 1.
New York: Holt, Rinehart and Winston, Inc. 289 pp.
"We have tried to present the material in a unified way, motivated by more or
less realistic problems drawn from a variety of fields. Moreover, we have tried to
give a clear description of the nature and extent of the practical information that
can be obtained with the techniques presented." Junior-senior level.
JEFFREYS, H., and B. JEFFREYS. (1962) Methods of Mathematical Physics. New York:
Cambridge University Press. 716 pp.
"This book is intended to provide an account of those parts of pure mathe-
matics that are most frequently needed in physics .... Abundant applications to
special problems are given as illustrations."
KARMAN,T. von, and M. A. B10T. (1940) Mathematical Methods in Engineering.
New York: McGraw-Hill Book Company. 505 pp.
"There are two ways of teaching the art of applying mathematics to engineering
problems. One consists of a systematic course comprising selected branches of
mathematics including a choice of appropriate examples for applications. The
other chooses certain representative groups of engineering problems and demon-
strates the mathematical approach to their solution . . . . This book might be
considered an experiment in the second method."
POLLARD,H. (1972) Applied Mathematics: An Introduction. Reading, Mass.:
Addison-Wesley Publishing Co., Inc. 99 pp.
"I wish to convince the advanced undergraduate, or beginning graduate
student ... that applied mathematics is interesting . . . by presenting a small
selection of problems of major importance ... and calling the attention of the
student to failure as well as success."
SAGAN,H. (1961) Boundary and Eigenvalue Problems in Mathematical Physics.
New York: John Wiley & Sons, Inc. 381 pp.
A senior-graduate text that provides a good alternative reference for much of
the material in Part A.
Bibliography
590
AdvancedCalculus
FRANKLIN,P. (1940) A Treatise on Advanced Calculus. New York: John Wiley &
Sons, Inc., 1940. Reprinted by Dover Publications, Inc., New York, 1964. 595 pp.
HILDEBRAND, F. B. (1962) Advanced Calculus for Applications. Englewood Cliffs,
N.J.: Prentice-Hall, Inc. 646 pp.
KAPLAN,W. (1952.) Advanced Calculus. Reading, Mass.: Addison-Wesley Publish-
ing Co., Inc. 678 pp.
DifferentialEquations
BoYCE,W., and R. DIPRIMA.(1969) Elementary Differential Equations and Boundary
Value Problems, 2nd ed. New York: John Wiley & Sons, Inc. 583 pp.
"This book is written from the ... viewpoint of the applied mathematician,
whose interests in differential equations may at the same time be quite theoretical
as well as intensely practical."
CODDINGTON, E., and N. LEVINSON. (1955) Theory of Ordinary Differential Equations.
New York: McGraw-Hill Book Company. 429 pp.
Authoritative advanced work.
GARABEDIAN, P. (1964) Partial Differential Equations. New York: John Wiley &
Sons, Inc. 672 pp.
A graduate text "written for engineers and physicists as well as mathemati-
cians."
INCE,E. (1927) Ordinary Differential Equations. Essex, England: Longman Group
Ltd. Reprinted by Dover Publications, Inc., New York, in 1956. 558 pp.
An excellent source for the classical theory.
KELLOGG,0. (1929) Potential Theory. Reprinted by Dover Publications, Inc.,
New York, 1953. 384 pp.
"It is inherent in the nature of the subject that physical intuition and illus-
tration be appealed to freely, and this has been done. However, in order that the
book may present sound ideals to the student, and also serve the mathematician,
both for purposes of reference and as a basis for further developments, the
proofs have been given by rigorous methods."
MOSER,J. (1973) Stable and Random Motions in Dynamical Systems, with Special
Emphasis on Celestial Mechanics. Princeton, N.J.: Princeton University Press.
198 pp.
A modern and deep study.
TYCHONOV, A., and A. SAMARSKI. (1964) Partial Differential Equations of Mathe-
matical Physics, Vol. I. San Francisco: Holden-Day, Inc. 390 pp. Vol. II, 1967,
250 pp.
Representative of a large class of books, this is relatively brief and has good
illustrations of how the mathematics is used.
DimensionalAnalysis
BIRKHOFF,G. (1960) Hydrodynamics-A Study in Logic, Fact, and Similitude,
2nd ed. Princeton, N.J.: Princeton University Press. Also reprinted by Dover
Publications, Inc., New York, 1955. 186 pp.
Bibliography 591
Noteworthy for the use of group theory in unifying ideas of dimensional

analysis and similarity theory.
BRIDGMAN, P. W. (1931) Dimensional Analysis, rev. ed. New Haven, Conn.: Yale
University Press. (Paperback edition, 1963.) 113 pp.
Discusses fundamental aspects of the subject with clarity and profundity.
HUNTLEY, H. E. (1967) Dimensional Analysis. New York: Dover Publications, Inc.
158 pp.
A good, inexpensive general reference.
KLINE,S. J. (1965) Similitude and Approximation Theory. New York: McGraw-Hill
Book Company. 229 pp.
A useful discussion of dimensional analysis and scaling.
LIGHTHILL, M. J. (1963) Laminar Boundary Layers, Chap. 1, L. Rosenhead (ed.).
New York: Oxford University Press.
Part 3 of this chapter, in just 11 pages, contains an excellent discussion of
dimensional reasoning and its application to the simplification of problems in
fluid mechanics.
PerturbationMethods
COLE,J. (1968) Perturbation Methods in Applied Mathematics. Waltham, Mass.:
Ginn/Blaisdell. 260 pp.
"This text is written very much from the point of view of the applied mathe-
matician; much less attention is paid to mathematical rigor than to rooting out
the underlying ideas, using all means at our disposal."
LAGERSTROM, P. A., and R. G. CASTEN.(1972) "Basic Concepts Underlying
Singular Perturbation Techniques." Society for Industrial and Applied Mathematics
(SIAM) Review 14, 63-120.
A readable and up-to-date survey.
NAYFEH,A. H. (1973) Perturbation Methods. John Wiley & Sons, Inc. 425 pp.
The latest and most comprehensive book on its subject.
VAN DYKE, M. (1964) Perturbation Methods in Fluid Mechanics. New York:
Academic Press, Inc. 229 pp.
The first modem book on the subject of perturbation methods, it contains
many insights.
ProbabilityandIts Relationto DifferentialEquations

FELLER,W. (1968) An Introduction to Probability Theory and Its Applications, Vol. I.
New York: John Wiley & Sons, Inc. 509 pp.
"It is the purpose of this book to treat probability theory as a self-contained
mathematical subject rigorously, avoiding nonmathematical concepts. At the
same time, the book tries to describe the empirical background and to develop
a feeling for the great variety of practical applications.
WAX,N. (ed.) (1954) Selected Papers on Noise and Stochastic Processes. New York:
Dover Publications, Inc. 337 pp.
A collection of six important papers. The text refers to S. Chandrasekhar's
"Stochastic Problems in Physics and Astronomy" and M. Kac's "Random
Walk and the Theory of Brownian Motion."
Bibliography
592
Foundationsof ContinuumMechanics
AR.rs, R. (1962) Vectors, Tensors, and the Basic Equations of Fluid Mechanics.
Englewood Cliffs, N.J.: Prentice-Hall, Inc. 286 pp.
" Sets out to show that the calculus of tensors is the language most appropriate
to the rational examination of physical field theories." The argument is illustrated
by detailed consideration of fluid mechanical equations.
JAUNZEMIS, W. (1967) Continuum Mechanics. New York: Macmillan Publishing
Co., Inc. 604 pp.
A graduate text that takes a "modern" rigorous approach to the subject.
Emphasizes elasticity.
SoMMERFELD, A. (1950) " Mechanics of Deformable Bodies," in Lectures on Theoret-
ical Physics, Vol. II. New York: Academic Press, Inc. 396 pp.
" My aim is to give the reader a vivid picture of the vast and varied material
that comes within the scope of theory when a reasonably elevated vantage point
is chosen." Unlike the other references, this requires no knowledge of tensors.
TRUESDELL, C., and W. NOLL.(1965) "The Non-linear Field Theories of Mechanics,"
in Encyclopedia of Physics, S. Flugge (ed.), Vol. 111/3.New York: Springer-Verlag
New York, Inc. 602 pp.
"The maximum mathematical generality consistent with concrete, definite
physical interpretation is sought."
TRUESDELL, C., and R. TOUPIN.(1960) "The Classical Field Theories," in Encyclo-
pedia of Physics, S. Flugge (ed.), Vol. IIl/1, pp. 226-790. New York: Springer-
Verlag New York, Inc.
"We present the common foundation of the field viewpoint. We aim to
provide the reader with a full panoply of tools of research, whereby he himself,
put into possession not only of the latest discoveries but also of the profound but
all too often forgotten achievements of previous generations, may set to work
as a theorist. This treatise is intended for the specialist, not the beginner."
Elasticity
GREEN,A. E., and W. ZERNA.(1954) Theoretical Elasticity. New York: Oxford
University Press, Inc. 442 pp.
"This book is mainly concerned with three aspects of elasticity theory which
have attracted attention in recent years ... finite elastic deformations, complex
variable methods for two-dimensional problems ... and shell theory."
LOVE,A. E. H. (1944) A Treatise on the Mathematical Theory of Elasticity, 4th ed.
New York: Dover Publications, Inc. 643 pp.
"It is hoped ... to present a fair picture of the subject in its various aspects,
as a mathematical theory, having important relations to general physics, and
valuable applications to engineering." Although not easy to read, this is the
classical reference work.
PREscon, J. (1924) Applied Elasticity. Essex, England: Longman Group Ltd.
666 pp. Reissued by Dover Publications, Inc., New York, 1946.
" In writing this book I have tried to see the subject from the point of view of
the engineer rather [than] from that of the mathematician."
Bibliography
593
SoKOLNJKOFF, I. S. (1956) Mathematical Theory of Elasticity. New York: McGraw-
Hill Book Company. 476 pp.
Perhaps the best single reference. "This book represents an attempt to present
several aspects of the theory of elasticity from a unified point of view and to
indicate, along with the familiar methods of solution of the field equations of
elasticity, some newer general methods of solution of the two-dimensional
problems."
TIMOSHENKO, S. P., and J. M. GOODIER.(1970) Theory of Elasticity, 3rd ed. New
York: McGraw-Hill Book Company. 567 pp.
"The primary intention ... [is] to provide for engineers, in as simple a form
as the subject allows, the essential fundamental knowledge of the theory of
elasticity together with a compilation of solutions of special problems that are
important in engineering practice and design."
A leading journal is the Journal of Applied Mechanics.
FluidMechanics
BATCHELOR, G. K. (1967) An Introduction to Fluid Dynamics. New York:Cambridge
University Press. 615 pp.
"A textbook which can be used by students of applied mathematics and which
incorporates the physical understanding and information provided by past
research."
LAMB,H. (1932) Hydrodynamics, 6th ed. New York: Cambridge University Press.
738 pp.
Also available from Dover Publications. An encyclopedic collection of classical
results.
LANDAU,L. D., and E. H. LIFSCHITZ,(1959) Fluid Mechanics. Elmsford, N.Y.:
Pergamon Press, Inc. 536 pp.
"The nature of the book is largely determined by the fact that it describes
fluid mechanics as a branch of theoretical physics." Translated from the Russian
by J. B. Sykes and W. H. Reid.
PRANDTL, L., and 0. G. TIETJENS.(1934) Applied Hydro- and Aeromechanics.
New York: McGraw-Hill Book Company. 270 pp. Also available from Dover
Publications, Inc., New York, 1957, as Fundamentals of Hydro- and Aero
Mechanics.
Brief and inexpensive, with considerable physical insight.
SEEGER, R. J., and G. TEMPLE(eds.). (1965) Research Frontiers in Fluid Dynamics.
New York: Interscience Publishers, a division of John Wiley & Sons, Inc.
738 pp.
A collection of useful survey articles.
SERRIN,J. B. (1959) "Mathematical Principles of Classical Fluid Mechanics,"
in Encyclopedia of Physics, S. Fltigge (ed.), Vol. VIII/I, pp. 125-263. New York:
Springer-Verlag New York, Inc.
"Our intent ... is to present in a mathematically correct way, in concise form,
and with more than passing attention to the foundations, the principles of
classical fluid mechanics."
594 Bibliography
The two principal journals are Journal of Fluid Mechanics and Physics of Fluids.
Beginning in 1969, Annual Reviews of Fluid Mechanics, a collection of authori-
tative summaries of research progress, has been published by Annual Reviews,
Inc., Palo Alto, Calif.
NumericalAnalysis
RALSTON, A. (1965) A First Course in Numerical Analysis. New York: McGraw-Hill
Book Company. 578 pp.
A standard introductory text.
WILKINSON, J. (1963) Rounding Errors in Algebraic Processes. London: Her Majesty's
Stationery Office. 161 pp.
A thin book, with many ideas.
HandbooksandTables
ABRAMOWITZ, M., and I. STEGUN. (1964) "Handbook of Mathematical Functions,"
National Bureau of Standards Applied Mathematics Series 55, 1046 pp. Reprinted
by Dover Publications, Inc., New York, 1965.
An invaluable collection of formulas, graphs, and tables.
FREIBERGER, W. (ed.) (1960) The International Dictionary of Applied Mathematics.
New York: Van Nostrand Reinhold Company. 1173 pp.
Defines the terms and describes the methods in the applications of mathematics
to 31 fields of science and engineering. Includes French, German, Russian, and
Spanish equivalents of the terms defined in the book.
MURPHY,G. (1960) Ordinary Differential Equations and Their Solution. New York:
Van Nostrand Reinhold Company. 451 pp.
Contains a tabulation of 2315 equations and their solutions.
Supplementary Bibliography (1994)
Companion Volume
A "following volume", designated as II, was mentioned on p. 9. This is
SEGEL,L. (1977) Mathematics Applied to Continuum Mechanics (with material on
elasticity by G. H. Handelman). New York: Macmillan Publishing Company. Cor-
rected edition reprinted by Dover Publications, 1987.
General Applied Mathematics

In the last years, a number of books have been published whose purpose is to
teach the art of constructing mathematical models. Some representative publications
are listed below.
ARIS, R. (1978) Mathematical Modelling Techniques. London: Pitman Publishing
Ltd. 191 pp.
The original title (was) ... "Notes toward the definition of the craft of mathematical
modelling."
BENDER,E. A. (1978) An Introduction to Mathematical Modeling. New York: John
Wiley & Sons, Inc. 256 pp.
"This book is designed to teach students how to apply mathematics by formulating,
analyzing, and criticizing models."
BURGHES,D. N., HUNTLEY,I., ANDMCOONALD,J. (1982) Applying Mathematics: A
Course in Mathematical Modelling. Chichester: Ellis Horwood Ltd. 194 pp.
"We regard mastery of model formulation as a crucial step, and much of this
book is devoted to attempting to make you more proficient in this process."
CROSS, M., AND MOSCARDINI, A. 0. (1985) Leaming the Art of Mathematical Mod-
elling. Chichester: Ellis Horwood Ltd. 155 pp.
"The primary aim of this book is to provide a framework which allows the 'novice'
modeller to 'catch' the ability to build mathematical models to analyse real world
problems arising from an industrial or commercial context."
HABERMAN,R. ( 1977) Mathematical Models: Mechanical Vibrations, Population Dy-
namics, and Traffic Flow. Englewood Cliffs, N.J.: Prentice-Hall, Inc. 402 pp.
"By studying three diverse areas in which mathematics has been applied ...
mechanical vibrations, population dynamics, and traffic flow ... this text attempts
to introduce to the reader some of the fundamental concepts and techniques of
applied mathematics."
Mathematical Methods
BENDER,C. M., AND ORSZAG, S. A. (1978) Advanced Mathematical Methods for
Scientists and Engineers. New York: McGraw-Hill Book Company. 593 pp.
595
596 SupplementaryBibliography
"The main purpose ... is to present and explain mathematical methods for ob-
taining approximate analytical solutions to differential and difference equations ...
(by) asymptotic and perturbative analysis .... We stl'C care but not rigor."
COLE, J. D., AND KERVORKIAN, J. (1981) PerturbationMethods in Applied Mathe-
matics. New York: Springer-Verlag, Inc. 558 pp.
A major revision of Cole ( 1968).
STRANG,G. (1986) Introductionto AppliedMathematics.Wellesley, Mass.:Wellesley-
Cambridge Press.758 pp.
"The central topics are differential equations and matrix equations-the contin-
uous and discrete .... We intend to try-working always with specific examples--
to combine the algorithms with the theory."
Numerical Analysis
LAPIDUS, L., AND PINDER, G. F. (1982) Numerical Solution of Partial Differential
Equationsin Scienceand Engineering.New York: John Wiley & Sons,Inc. 677 pp.
" ... Incorporates the essential elements of all the numerical methods currently
usedextensively in the solution of partial differential equations encountered regularly
in science and engineering."
PREss, W. H., TEUKOLSKY, S. A., VETTERUNG,W. T., AND FLANNERY, B. P. (1992)
Numerical Recipes in C: The Art of Scientific Computing, 2nd ed. Cambridge:
Cambridge University Press. 994 pp.
"Our aim in writing the original edition of NumericalRecipes was to provide a
book that combined general discussion, analytical mathematics, algorithmics, and
actual working programs .... We wanted, then and now, to write a book that is
informal, fearlessly editorial, unesoteric, and above all useful. There is a danger
that, if we are not careful, we might produce a second edition that is weighty,
balanced, scholarly, and boring."
RALSTON, A., AND RABINOWITZ, P. ( 1978) A FirstCoursein NumericalAnalysis, 2nd
ed. New York: McGraw-Hill Book Company. 556 pp.
A major revision of Ralston ( 1965).
Differential Equations and Dynamical Systems

ARNOLD, V. I. (1973) OrdinaryDifferentialEquations.Cambridge, Mass.:MIT Press.
200pp.
"I havetried to emphasize the geometric and qualitativeaspect The applications
... to mechanics are considered in more than customary detail."
DRAZIN, P. G. ( 1993)Nonlinear Systems. Cambridge: Cambridge University Press.
317 pp.
"My primary aim is to introduce simply the mathematical properties of nonlinear
systems as an integrated theory .... A secondary aim is to give an impression of
the diverse applications of the theory .... The approach is to discuss topics in as
concrete a way as possible, using worked examples and problems to motivate and
illustrate general principles."
GUCKENHEIMER, J., AND HOLMES, P. (1983). NonlinearOscillations,DynamicalSys-

tems, and Bifurcationsof VectorFields. New York: Springer-Verlag, Inc. 453 pp.
"There is now widespread interest in the engineering and applied science com-
munities in strange attractors, chaos, and dynamical systems theory. We have written
this book primarily for members of this community."
HIRSCH, M. W., AND SMALE, s. (1974) DifferentialEquations, Dynamical Systems,
andLinear Algebra. New York: Academic Press. 358 pp.
"This book is about dynamical aspects of ordinary differential equations and the
relations between dynamical systems and certain fields outside pure mathematics
a
... The importance of the ordinary differential equations vis vis other areas of
science lies in its power to motivate, unify, and give force to those areas. Our four
chapters on "applications" have been written to do exactly this."
JOHN, F. (1978) Partial DifferentialEquations. New York: Springer-Verlag, Inc. 198
pp.
"Emphasis is put on discovering significant features of a differential equation,
and on exploring them as far as possible with a limited amount of machinery from
mathematical analysis."
SupplementaryBibliography
STROGATZ, S. H. ( 1994). NonlinearDynamicsand Chaos:WithApplicationsto Physics,

Biology, Chemistry, and Engineering.Reading, MA: Addison-Wesley. 496 pp.
"Written in a friendly style, the treatment stresses analytical methods, concrete
examples, and geometric intuition."
Updated Material
Much of the material in this book is classical, but some concerns newer fields of
application. Largely by means of references, we here update the newer material.
Section 1.2
In recent years, especially in the 1980's, there have been extensive studies of the
dynamics of galactic disks, aimed at determining the mechanisms for the formation
and maintenance of the spiral structure in galaxies and at following the evolution of
such spiral structures in the course of time. Naturally, the details of these processes
depend on the basic model adopted for the galaxy. A variety of dynamical scenarios
have been described for different basic models. In particular, there is now a successful
theory for the global structures of all Hubble morphological types in terms of linear
modes calculated for a family of physically plausible basic states. In the following
papers, the modal approach to the global spiral structure in galaxies is studied theo-
retically and compared in detail with the structures observed in one prominent nearby
galaxy, Messier 81. Good agreement is found. The modal characteristics are clearly
visible despite the interaction with a companion.
BERTIN,G., LIN, C. C., LowE, s. A., ANDTHURSTANS,R. P. (1989) "On the Modal
Approach to the Morphology of Spiral Galaxies" I, II, Astrophys. J., Vol. 338, p.
74 and p. 104. III, LoWE, s. A., ROBERTS,W. W., YANG, J., BERTIN,G., AND
LIN, C. C. (1994) Astrophys. J., to appear, May 20. BERTIN,G. (1993) Publications
of the Astronomical Society of the Pacific, Vol. 105, p. 640.
Section 1.3
Aggregation of slime mold amoebae is an example of biological pattern formation.
A discussion of advances in both the particular and the general area of application
can be found in
LEVIN, s. A., AND SEGEL,L A. (1985) "Pattern Generation in Space and Aspect,"
SIAM Review, Vol. 27, No. 1, pp. 45-67.
The motility-chemotaxis equation ( 1.3.10) is one tool in an ongoing theoretical
analysis of cellular motion and its consequences in biology and medicine. See for
example
l'RANQUILLO,R T. ( 1990) "Theories and Models of Gradient Perception," in Biology
of the Chemotactic Response, J.P. Armitage and J.M. Lackie, eds., pp. 35-75.
Cambridge: Cambridge University Press.
Chapter 6
An illuminating example of scaling, simplification, and a "wretched consistent
approximation" is given by
598
Updated Material
RILEY, N. (1976) "A Note on 'Simplification and Scaling,'" Proc. Edinburgh Math.
Soc., Vol. 20, pp. 63-67.
Chapter 8
Interest has continued both in standing gradient flows, and in the approach discussed
here to find approximate solutions. For example see
McLAUGHLIN,s.,ANDMATHIAS,R. T. (1985) "Electr<rOSmosis and the Reabsorption
of Fluid in Renal Proximal Tubules," J. Gen. Physiol., Vol. 85, pp. 699-728.
Chapter 10
This chapter is based on the "traditional" dimensionless variables (10.1. 7) for the
enzyme-complex equations (10.1.5). These variables lead to a singular perturbation
development in terms of a small parameter e = e/s that reflects the assumed small
ratio of the initial enzyme concentration e to the initial substrate concentration s.
Recent research has provided a good rationale for a different choice of dimensionless
variables. It turns out that the "correct" small parameter is e* = e/(s + Km) where
Km = ( k_ 1 + k2 ) / k 1 Note that if e is small then e* is small, and indeed all results
obtained for small e are valid. But even if e is not small, e* may be small and the
various approximations may still be valid. See the following references.
SEGEL, L. A., AND SLEMROD,M. (1989) "The Quasi-Steady Staie Assumption: A
Case Study in Perturbation," SIAM Review, Vol. 31, pp. 446-477.
SEGEL, L. A. (1988) "On the Validity of the Steady State Assumption of Enzyme
Kinetics," Bull. Math. Biology, Vol. 50, pp. 579-593.
The ideas have been extended to a new class of problems in
SEGEL,L. A., ANDGoLDBETER,A. ( 1994) "Scaling in Biochemical Kinetics: ~on
of a Relaxation Oscillator," J. Math. Biology, Vol. 32, pp. 147-160.
A recent comprehensive reference is
TuRANYI, T., TOMLIN,A. S., ANDPILLING,M. J. (1993), "On the Error of the Quasi-
Steady-State Approximation,'' J. Phys. Chem., Vol. 97, pp. 163-182.
Hints and Answers

(for exercisesmarked with t)
3.3
SECTION
7(b). Hint: Separate off the tenns q[w(m + 1, N - 1) - 2w(m, N - 1) + w(m - 1,
N-1)].
SECTION 3.4
9(a). P{r) = (In R - In r)l(ln R - In e). (b) P(p) = (R- 1 - p- 1)l(R- 1 - e- 1).
SECTION 4.2
2(a). Hint: Reduce the problem to one in alternating series.
5. Hint: Use integration by parts.
SECTION 5.1
6. IC= 0.()()4crn I~.
2
7. Given the present value of the gradient, the cooling process would have lasted
about 3 x 107 years. It is now believed that heat given off by radioactive substances
is the major source of the observed temperature gradient.
SECTION 6.2
1 2
3(c). T/(L/g) = /(AIL), where A is the initial linear displacement of the bob.
'
When AIL is small, /(AIL) can be replaced by /(0), so the result of this part reduces
to the result of part (b).
IO(b). Hint: You must solve an equation of the form [yi,8/(yi, ... ,y.)/8yi]/
/(yi, ... , y.) = constant.
SECTION 6.3
2. U = A exp (-ab), L= a- 1
9(b). From the zeroth order solution (7.1.2) with a= 0, the time scale should
again be wo 1 ; the scale for 0 should be Owo 1 (c) The angle scale should be a
(or Owo 1) if a is considerably larger (or smaller) than Owo 1 If both quantities are
about the same size, then either scale or their average or geometric mean can be
used.
SECTION 7.1
2(a). Hint: Use the definition of derivative.
4. Hint: Derive an equation expressing conservation of energy, and consider its
consequences.
SECTION 7.2
I. Hint: If .i = v, dvldt = (dvldx)(dxldt) = v(dvldx).
ll(a). ,W>= -(2n 2/11')f~ x sin 2 nx dx = -(2n 2 /,r)(i,r 2
+ in- 2
).
600
Hints and .Answers 601
SECTION 8.2
6(a). Partial answer. The steady state mass conservation requirement is F 1(x) - F.
(x+ Ax) =0 for arbitrary Ax so that F. is constant.
SECTION 9.1
2(b). As the text indicates, begin by introducing s = em into (1), wheres can be
safely assumed to have a Maclaurin expansion in powers of e.
3. Hint: In expanding the square root, it is helpful to write (4-4e) 1' 2 =
2(1 -e) 112 and then to use the binomial theorem.
4(f). Hint: Introduce a new variable y = xe 112
11(a). Divide (13) by eE 2 (b) Divide (13) by E and remember that mand hence E
does not approach a nonzero limit as e-+ 0.
SECTION 9.2
2(c). Yo(x) = exp (-fx ), Yrt.6= kJ~ exp (-!s 413 ) ds; k- 1 = J: exp (-!s4 13 ) ds.
213
4(c). Partial answer. By balancing the first two terms in the differential equation,
one sees that the time scale for t small is Jc-1 From the second initial condition
it then foUowsthat Ip.- is the correct scale for the displacement.
S. Hint: There are boundary layers of O(e112) thickness near x = 1 and x = -1.
8. The limit is 1 if ex= o(l), e-t if ex= k + o(l), and O if 1/x = o(e).
10.2
SECTION
3(c). Hint: Use the method of undetermined coefficients.
5(a). Hint: Using (2), show that the equation for s 1 has a relatively simple inte-
grating factor and can be written
dt
(K+ So
d- --s1
So
)
= [ -K
+ ,c) +
So(So
KA ]
+ ,c) io
So(So 2

Completion of the solution thus requires computing
f[ -,c ,cA ]
u(u + ,c) + u(u + ,c)z du,
which can be accomplished by employing partial fractions.

8(d). Hint: It is often helpful to multiply the numerator and denominator by the
same parameter.
SECTION 11.2
3. Hint.' A system quite similar to (20) appears. Again, the approach is to find an
equation for the analogue of A 2 + B 2 When this is done, one of the functions
involved will be found to be identically z.ero.
7(b). In establishing boundary conditions, ignore TST terms.
SECTION 12.1
6. Hint: First use Leibniz's rule to perform the differentiation. Then employ the
integral mean value theorem to transform the integral. Finally, consider the limit
N-+M.
12. Hint: Use the integral mean value theorem.
602 Hints and Answers
12.2 SECTION
14(b). Hint: First consider two special cases, where f = 0 and g = 0, respectively.
16(a). Q = (I - ,c2 + 2i,c)/(l + ,c2), B = 1, o= tan - , (2,c/(1 - ,c2)].
SECTION 13.2
1. v=i+x2i+(x, -t)k.
2(b). Hint: The particle paths are given by x = x(t), y = y(t), where dx/dt = y,
dy/dt = x. The solution of these equations can be obtained by eliminating either
x or y.
3(a). !x 2 + x = t(y- Yo)+ !x5 + .xo).
+ Yo
4(a). x = -exp (-y) cost+ constant. (b) x = ! exp (-t + to - BXsin t - cost)+
A - ! exp (-BXsin to- cos t0), y = t - t 0 + B.
SECTION 13.3
1. Partial answer. o(~, t) = (A, + t) + (A2+ t 2 )sin t.
SECTION 13 .4
4(a). J = 1. (b) J = !Ai"' 12
APPENDIX 13 .1
2. 8 T/8r
2 2 + ,- 1
&7/8r + r- 2 2
8 Tj8(J2.
3(a). r :P(p + (psin rp)- :r/>( sin rp~) +

2 2
:)
2 1
(p2 sinz rp)-1 :::
APPENDIX 13.3
2. Hints: (a) Use the triangle inequality. (b) Use the fact that zero is an interior
point of R<>.
SECTION 14.2
2. (I0)-11Z(2Tu + T 21 -2T,2 -T22)dA.
3(c). Zero.
6. Hint: Integrate over a sphere of radius R. Use the fact that the absolute value
of the integral is no greater than 41rR 2 times the absolute value of the integrand.
14.3 SECTION
3(a). Hint: Use the integral mean value theorem.
SECTION 14.4
4. Hint: Write v = Vrp+ w. Show that K* = K + a nonnegative term.
SECTION 15 .1
1. Hint: Transform po da to a volume integral and use the hydrostatic expression
forp.
4 and 5. Hint: Use the identity curl (grad g) = O.
5. p- 1 Vp= V f p- 1 dp.
8. Hint: Use Bernoulli's equation and a simple form of the continuity requirement.
10. Hint: Use a simple form of the continuity equation.
Hints and Answers 603
SECTION 15 .2
5. The first steady solution is unstable, the second is stable.
=
9(a). Hint: If Q a2 w/ot2,show that 02 Q/oz2 = k 2 Q. (b). In case (i) the pertur-
bation velocity maintains its initial value. There is neither growth nor decay, since
there is no mechanism for altering the initial perturbation. (The perturbation would
disappear in this situation if the slightest viscosity were present.) In case (ii) the
slight variation in pressure, and the resulting pressure gradient, drive a flow that
increases with time until linearization ceases to be valid.
lO(f). Instability sets in at the minimum value of {3c(m),namely, (Vy+ v;)2.
This minimum occurs when m = me= (yic)112, so the expected wavelength of the
instability is 2TT/(yic)114
lO(g). According to the possibility investigated here, instability would set in when
f3 first exceeds the value r + ic.
SECTION 15.3
4(c). The solution should be appropriate at a given point of an actual tube until
the disturbance reaches the end of the tube and some reflected disturbance travels
back to the point in question.
9(a). Hint: Use (41b) and introduce m, the common value of p+s;t and p-s;
whose existence is guaranteed by (4la). (e). Hint: Employ Taylor's formula to
approximate p+-p- = p(p+, s+)-p(p-, s-).
SECTION 15.4
6. Hint: Use as a control surface the cylinder x 2 + y 2 = R 2 , z = 1. Let R-+ oo.
7. g, = - Up[l - !(a/p) 3 ] cos 'P
Authors Cited
Citations from the bibliography are not indexed.
Acrivos, A., 305 Heineken, F. G., 305 Prager, W., 31

Adam, G., 109 Herman, P., 32 Prigogine, I., 537
Andronov, A., 344 Hooke, R., 34 Proust, J ., 466
Aris, R., 305 Hoppenstead, F. C., 316
House, C.R., 261 Reiner, M., 510
Barber, M., 74 Rubinow, S., 311
Batchelor, G. K., 200 James, M. L., 222 Ruffner, J. A., 487
Bell, E. T., S, 114 Johnstone, R. B., 203
Bok,B.J., 20 Joseph, D., 241 Schey, H. M., viii
Bonner, J. T., 22 Segel, L., 24, 112, 185, 246,
Bossert, W., 245 Kac, M., 278 277
Bowen, J. R., 305 Kanwal, R. P., 583 Selwyn, Ph., 491
Brockett, R. W., 32 Keller, E. F., 24, 112 Shaffer, D., 34
Keller, E.G., 204 Silver, D., 466
Carrier, G. F., 35, 298, 530 Keller, H. 8., 34 Smith, G. M., 222
Chandrasekhar, S., 15, 531 Khaikin, S., 344 Smith, J. M., 345
Chang, C. T., 530 Klamkin, M. S., 208 Smith, W., 106
Chinnaswamy, A. V., 583 Sobolev, S. L., 170
Cohen,H., 32 Lanczos, C., 141, 187 Stakgold, I., 296
Courant, R., 33, 121 Landahl, M., 537 Stoker, J., 344
Lebowitz, J., 311
Dainty, J., 261 Lefever, R., 537 Thring, M. W., 203
Daly, B. J., 530 Lefschetz, S., 339 Titchmarsh, E. C., 147
Darvey, I. G., 344 Lewy,H., 121 Tormey, J., 245
Delbruck, M., 109 Lighthill, M. J., 68 Truesdell, C., 22, 33, 412, SOI
Den Hartog, J. P., 323 Lin, C. C., 14, 20 Tseng, H., 466
Diamond, J., 245 Tsuchiya, H. M., 305
Doherty, R. E., 204 McBride, E. B., 76 Turing, A., 537
Drew, D., 221 Matlak, R. F., 344
Mollo-Christensen, E., S37 Van Ness, H. C., 491
Edson, L., 8 Montroll, E.W., 34 Vasil'eva, A. 8., 315
Einstein, A., 5 Morrison, P., 14 Vitt, A., 344
Engel, A., 33 Munson, Z. Z., 241 von Kannan, T., 8
von Neumann, J., 32
Ferguson, N., 194 Nemenyi, P. F., 502
Feynman, R., 496 Ninham, 8., 74 Watson, G. S., 106
Finlayson, 8., 194, 466 Nordtvedt, K., 41 Weinbaum, S., 275
Friedrichs, K., 121 Whittaker, E. T., 69
O'Malley, R. E., 300 Widder, D., 176
Gadd,G., 202 Oppenheim, A. K., 305 Wigner, E. P., 33
Gal-or, 8., 500 Oppenheim, I., 491 Wilson, C., 37
Gasparoux, H., 466 Wimp, J., 76
Goldgraben, J. R., 275 Pitts, R. F. 247 Wolford, J.C., 222
Grenander, U., 34 Pollack, H. 0., SS Woltjer, L., 16
Gupta, A., 537 Polya, G., 76
Subject Index
Absorbing barrier, 96 Complex representation, 379-80
Action and reaction, 361, 457-58 Compression waves, 539-49
Active transport, 245 Computation, 32
Adiabatic process, 494 Condensation, 541
AMP,23 Conservation laws; See Mass conservation,
Analyticity, 59 Linear momentum balance
Angular momentum, 466 Consistency, 188, 189
conservation of, 37 Constitutive equations, 363, 487; see also
Anharmonic periodic behavior, 178 entries for specific substances, such as
Anomaly, 45 Motionless fluid, Inviscid fluid
Aphelion, 45 heat fl.ux,117
Applied mathematics, 4-8, 31-33 Contact force, 456
Astronomical unit, 11 Continuity equation, 357, 442
Asymptotic approximations, 81-89, 230, 232 Continuousdependence,63
to eigenfunctions and eigenvalues, 163-65 Continuous medium, 414
Asymptotic series, 230, 232 Continuum, 3, 24, 417-18
Asymptotic stability, 324 Continuum mechanics, validity of, 490-91
Attenuated wave, 386 Convection, 353
Autocorrelation function, 178 Convergence, 52
Autonomous systems, 339 absolute, 62
Availability of energy, 501 of perturbation series, 232-33
Average value, 76 uniform, 62
Axial strain, 363 Convergent series, compared to asymptotic, 84
Cosine series; see Fourier series
Balance law, 24; see also Linear momentum Covariance, 487
balance, Energy balance Critical points, 338
Balancing procedure, 282 behavior of trajectories near, 339-41
Barotropic fluid, 512, 541 Crocodiles, kidney tubules of, 273
Basic simplification procedure, 186-94 Cross-section paths, 350
Bernoulli equation, 513-14, 556 Cylindrical coordinates, 565
Bessel equation, 164-65
Bessel function, 80 D'Alembert's paradox, 557, 560-63
Bessel inequality, 146 Deborah number, 510
Binomial coefficient, 75 Deformation rate, 4 79-80
Binomial expansion Degree of freedom, 344
use in perturbation theory, 234 Delta function, 416
Birds, salt glands of, 273 Density, 355; see also Phase space number
Dode's law, 37 density
Body force, 358, 454 Deterministic system, 42
Boltzmann's constant, 78 Difference equation, 91, 153
Bond number, 202 Differentiability, 60, 64-65
Boundary conditions, 119, 121 Differential equations; see also Laplace
elastic bar, 366-68, 372 equation, VVaveequation
homogeneous, 124, 126 nondimensionalization of, 195-98
Boundary layer, 288 residuals for, 191-93
Brownian motion, 73-74, 78-79 system of, 42
recurrence propertv in, 108-110 Diffraction
Buckingham pi theorem, 207 of acoustic waves, 580-87
Buoyancy, 556 integral, 584-86
Diffusion coefficient, 78
Cauchy's differential equations, 461 Diffusion equation, 26, 93, 100, 118, 120-29,
Center, 341 151, 156
Chain rule, 433-34 fundamental facts concerning, 129
Change of units rule, 205 numerical integration of, 153
Characteristic curves, 17 solution by Fourier transform, 170-71
Characteristic equations, 17, 160 Diffusivity, 26, 255
Characteristics, Lagrange's method of, 17 Dilatation, 431
Chemical engineering, scale models in, Dimensional analysis, 195-204
202-203 comparison of physical and mathematical
Chemotaxis, 26 approaches, 259
Circular cylinder, flow past, 551-58 of functional relationship, 249-52
Circulation, 555 practical examples of, 204-205
Clausius-Duhem inequality, 479 Dimensional homogeneity, 208, 257
Coagulation, 97 Dimensionless groups, 197
Coin tossing, 72, 74 Dimensionless parameters
Col, 340 estimating the size of, 264
Collisions, 491 physical interpretation of, 272-74
Completeness, 163 Dimensionless variables, 128, 196, 517
dimensional, 207 Dirac delta function, 176
Complex, 304 Direct notation, 460
606 Subject Index

Dirichlet problem, 573 Fresnel integrals, 90
Discontinuous solutions, 549 Frog skin, 260--61
for elastic bars, 381-84, 388-99 Froude number, 201
for general media, 482-85 Function space, 63
Discrete spectrum, 160 Functional relationship,
Displacement, 352 nondimensionalization of, 198-200
Dissipation function, 481 Fundamental solution, 103-106, 160
Dissociation, 247 of Laplace equation, 575
Distortion, measure of, 363 Fundamental term, 127
Distribution function, 15-16 Fundamental units, 196
DNA molecules, collision of, 106
Drag, 556 Galactic spirals, 19-20
Dubois-Reymond lemma, 121, 441 Galaxies, 10-14
Dynamics, 418 Gammafunction,86-89
Gauss' theorem, 568
Effective radius, 260 Generating function, 76-77, 80
Eiconal equation, 396 Gibbs phenomenon, 141-43
Eigenfunctions, 157 Gibbs relations, 477
asymptotic approximation, 163--65 Gompertz equation, 325
completeness, 163 Gram molecular weight, 247
linear independence, 161 Gram-Schmidt process, 161--62
orthogonality, 161 Gravitation, 39
properties, 405 Gravitational potential, 15, 566
Eigenvalue problem, 522 Green's formula, 108
perturbation theory for, 242 Green's functions, 574-79
Eigenvalues, 157 Ground, temperature in, 151-53
asymptotic approximations to, 163--65
multiplicity of, 162 Harmonic analysis, generalized, 172-80
numerical determination of, 165--66 Harmonic functions, 570- 72
properties of, 405 Harmonics, 127
Elasticity, 349 Heat conduction, 115-29
Emergent osmolarity, 248 geophysical, 158-59
Energy, 45, 480 in ground, 151
Energy balance, 476 in nonuniform medium, 156
Engineering, applied mathematics in, 8 three-dimensional, 120
Ensemble average, 78, 179 Heat conductivity, 116
Entropy, 478-81, 496-501 Heat content, 496
Enzymes, 304-306 Heat equation; see Diffusion equation
Equations, heat; see Heat equation Heat flux density, 116, 477
Equilibrium, 321, 515 Heaviside function, 175
stability of, 322-24 Helmholtz equation, 578-81
Equilibrium thermodynamics, 474 Heurism, 34
Error Higher harmonics, 53
formula, 68 Homogeneity, 159
least squared, 144 dimensional, 208
weighted, 145 spatial, 217
Euler expansion formula, 355, 432 Homogeneous boundary condition, 124,
Euler momentum equation, 510 126
Eulerian strain, 375 Homogeneous function, 43
Eulerian variables, 357, 420 Hooke's law, 364
Euler's contributions to mechanics, 413, Horizontal wave number, 522
503-504 Hourglass, 515
Euler's laws of mechanics, 413, 503-504 Huygen's principle, 584
Existence, 57, 61 Hydrostatic equation, 509
Exotic materials, 487
Expected value, 75 Ideal gas, 471
Ill-conditioned problem, 189
Fallin~ sphere, dimensional analysis Images, method of, 103, 576-77
of flow, 205 Impedance boundary conditions, 368
Fick's law, 101, 255 Impenetrability of matter, 419
Field equations, 365, 440, 485-86 Incompressibility, 357, 426
Field point, 574 Infinite plane, Green's function for, 576
Finite differences. 66--68 Inhomogeneous function, 43
First passaiie time, 97, 104 Initial conditions, 119
Flux, 24, 100, 104 Initial layer, 308
Focus, 341 Initial value problem, 42, 46, 57, 104
Formal calculation, 59 simrular perturbation approximation to,
Fourier identity, 169 308-16
Fourier series, 77, 131-39 Inner approximation, 290
Fourier theorems, 137 Innerequation,293
Fourier transform, 167-72 Inner solution, 308
Fourier's book on mathematical theory Instability, 28; see also Stability
of heat, 114-15 Kelvin-Helmholtz, 536
Free will, 43 of stratified fluid, 515-30
Frequency, 380 Taylor, 536
Subject Index
Integral Mathematical model (cont.)

of differential equations, 17, 40, 56 DNA collision, 106
Integral equation, 61 enzyme-substrate chemical reaction,
Integral mean value theorem, 435-36 302-307
Integrals, asymptotic approximation of, 84 one-dimensional elastic bar, 371-72
Intermediate expansion, 312-13 planetary motion, 37-40
region, 293-94 simple pendulum, 48
Internal energy, 476 standing gradient flow, 244--68
Interphase, 22 stellar system, 14
Intrinsic reference quantities, 197-98 Maximum principle, 123, 571-72
Inviscid fluid, constitutive equation for, 510 Mean displacement, 76
Irreversible processes, 493 Mean free path, 491
Irrotational motion, 482, 513 Mean square displacement, 76
lsentropic processes, 498 Mean value theorem, 571
Isochoric motion, 427 Michaelis constant, 319
Iteration, 50-51, 61 Michaelis-Menten approximation, 308
for perturbation calculations, 23~0 Michelson Fourier analyzer, 141
for polynomial roots, 280 Model, mathematical; see Mathematical
model
Jacobian, 355 Modelequation,278
material derivative of, 429-32 Models, scale, 200-203
Jump conditions, 391-95, 484, 548-49 Mole, 247
Molecular average, 414-15
Kelvin-Helmholtz instability, 536 Molecular model, 506
Kepler's laws, 37 Moment, 76
Kinematic boundary condition, 425 force,465
Kinematics, 38 momentum, 466
of deformable media, 418-25 Momentum balance
Kinetic energy angular, 465-69
in elastic bars, 401 linear, 360-62, 454--61
Kinetics, reaction, 305 Momentum flux tensor, 463
Kutta-Joukowsky theorem, 557 Momentum transfer equation, 462
Morphogenetic motion, 4
Lagrangian strain, 375 Motility, 26
Lagrangian variables, 420 Motionless fluid, constitutive equation
Langevin's equation, 78 for, 506-508
Laplace equation, 107, 552, 567 Multiple scale expansions, 325-31
Laplace method, 84
Large box method, 446-48 Naive approximations, 279
Large numbers, laws of, 71 Natural boundary conditions, 407
Least squared error; see Error Neumann problem, 573
Legendre equation, 560 Neutral stability, 324, 523
Length scale, 128, 215 Newton's contributions to mechanics, 412,
Liapunov stability, 324 502-503
Lift, 556 Newton's laws, 412
Limit cycles, 339 Newton-Euler sound speed, 543
Line integrals, in thermodynamics, 494-95 Newton-Fourier law of cooling, 101
Linear combination, 44, 59 Node, 340
Linearity, 43 Nonlinearity, 16, 27, 53, 539-40
Linearization,29,322,368-71,519 in stability theory, 530
Lineweaver-Burke plot, 319 in waves, 544-47
Liouville's normal form, 164 Nonuniqueness, 66
Lipschitz condition, 57 Norm, 63
Local determinism, 490 Normal modes, 158, 530
Local equilibrium assumption, 475 Normal stress, 506
Local moment equilibrium, 467 Normalization, 145
Local stress equilibrium, 456-57 Notation, theoretical vs. particular, 421
Logistic equation, 324 Numerical analysis, 41
Lorentz transformation, 488-90 compared with analytical approach,
155, 157
Mach number, 542 for eigenvalues, 165-66
Mass action, law of, 302-304 of stratified fluid layer, 531
Mass conservation, 16, 254-55, 356-58,
440-49 O and o symbols, 112-113, 227
Mass distribution function, 416-17, 451-52 Observer invariance, 476
Matching, 293-94, 297, 312-15 Orbits, planetary and cometary, 40
Material derivative, 353, 426-28 Order, 52
Material description, 420 Order of magnitude, 213-14
Material region, 429 Order symbols, 112-13
Material variables, 357 Orthodoxy, 218-21
Mathematical model Orthogonal functions, 145
Brownian motion, 73-74. 78-79 Orthogonality, 161
chemotactic amoebae, 24-27 Oscillation theorems, 163
conceptual versus descriptive, 278 Osmol, 247
diffraction of acoustic waves, 583 Osmosis, 247
608 Subject Index

Outer approximation, 289 Random processes, 71
solution, 308 Random walk, 73-77, 91-97, 100-101
Over-all wave number, 522 Rate constant, 303
Rate of deformation tensor, 479
Parameter dependence, continuity of, 58 Rayleigh quotient, 404-409
Parametric differentiation, 59, 236 Rayleigh-Ritz method, 407-409
Parsec, 11 Reciprocity property, 575
Parseval's theorem, 146 Recurrence property, 108-10
Partial differential equations Reflected wave, 381
classification of, 10 Reflecting barrier, 95-96
first order, 17 Regular perturbation theory, 240
Particle, 419 Relative error, 191
Particle path, 419, 423-24 Relativity, 41
Pendulum problem, 48, SS Residual, 190-94
dimensional analysis applied to, 204 Resonant term, 53
multiple scale method applied to, 326-30 Reversible processes, 493
phase plane behavior for, 335-38, 342-43 Reynolds number, 201
residuals for, 193 Reynolds transport theorem, 374, 443
series method for, 226-33 Riemann method, 544
stability of, 322-24 Riesz-Fischer theorem, 147
Perihelion, 45 Rigid body, 487
advance of, 41, 56 Rigor, 34
Period,380
Permeability, 248 Saddle point, 340
Perturbation, 29, 515 Scale models, 200-203, 253
Perturbation theory, 41, 45-SS Scale-up problems, 202
multiple scale methods, 325-31 Scaling, 209-23, 226, 284
Poincare's, 53-55 in chemical kinetics, 306-307
series method, 225-36 regular perturbation theory and, 236
two-parameter, 265-67 in standing gradient flow problem, 262-64
unsuccessful use of, 265 Scattering, 384
Phase, 381 Secondary quantities, 206
Phase plane, 334-44 Secular equation; see Characteristic equation
Phase space number density, 15 Secular term, 53
Pi theorem, 207 Semipermeable membrane, 247
Piecewise smoothness, 131 Sensitivity, 71, 189, 193
Pipe flow, 205 Separation constant, 125
Piston, 492-93 Separation of variables, 123-27, 553-55
Plane waves, 542 Separatrix, 337
Planets, 36 Shear stress, 506
Point-particle interchangeability, 420 Ship design, 201-202
Poisson equation, 18, 568 Shock strength, 551
Poisson integral formulas, 577 Shock waves, 547-49
Polar coordinates, 563-64 Silent rings, 587
Polar fluids, 466 Silly Putty, 506
Polynomial, zeros of, 188 Similar regions, 436-39
Polynomial equations, singular perturbations Simple waves, 544
applied to, 278-84 Sine integral function, 143
Polytropic state equation, 541 Sine series; see Fourier series
Potential energy, in elastic bars, 402-403 Singular perturbation theory, 277-301
Potential theory, gravitational, 566-69 applied to biochemical kinetics, 302-20
Power spectrum, 179-80 higher approximations in, 311-16
Pressure, 478, 506 Singularity, excluded by small sphere, 568-69
Primary quantities, 205 Skin effect, 153
Pritulo method, 332 Slime mold amebae, aggregation of, 22-31
Probability, 73 Small box method, 444-46
Probability density function, 93 Small terms, neglect of, 210
Projectile problem Smoothness, piecewise, 131
air resistance in, 224 Sound speed, 377, 542
basic simplification procedures for, 187, 193 Source point, 574
iteration method in, 239 Source solution, 103
nondimensionalization of, 195-98 Spatial description, 420
parametric differentiation in, 237-38 Spatial variables, 357, 420
perturbation theory in, 233-40 Specific heat, 481
scaling of, 211-13 Sphere
Propeller thrust, 204 cooling of, 129
Pure mathematics, relation to applied flow past, 559-60
mathematics, 68-69, 115 Green's function for, 576-77
Spherical waves, 543
Quasi-linear equation, 545 Stability, 16, 28, 58, 322-24
in numerical analysis, 154
Radiation, 147 in population dynamics, 533-34
Radiation condition, 384, 583 of reaction-diffusion equations, 537-39
Railroad track, strain in, 365 of two-layer inviscid flow, 535-36
Random motion, 25 of viscous flow model, 531, 534-35
Subject Index
Stagnation points, 558 Transient, 124
Standing gradient flow, 245 Translation invariance, 103
Standing wave, 381, 403, 544 Transmitted wave, 381
Star-shaped regions, 436 Trial and error, 222
State variables, 492 Triangle, diffusion from, 106-108
Stationary time sequence, 176-78 Trigonometric identities, 229
Steady flows, 421, 423 Trivial solution, 125, 157
Stefan's law, 149 Truncated sinusoid, Fourier analysis of,
Stieltjes integral, 176 173-76
Stirling's approximation, 81, 86-89, 102
Stokes' law, 78, 204-5 Uniform approximation, 294-95, 310
Strain, 363 Uniform validity, 289-90
Stratified fluid, 516 Uniqueness, 58, 60, 122-23
Stream function, 511 Unorthodoxy, 218-21
Streamlines, 422-24 Unstable equilibrium, 515
Stress, 358-60, 454 Unstirred layers, 261
component, 361
in motionless fluids, 505-10
in one-dimensional elastic bar, 364 Variance, 76-77
tensor, 458-61 Vector,488
Stress-strain relation, 364 product, 465
Sturm-Liouville theory, 159, 167, 404-405, Velocity, 352, 420
521, 529 Velocity potential, 512-13, 558
Substantial derivative; see Material derivative Velocity scale, 214
Substrate, 304 Verhulst-Pearl equation, 324
Successive approximations; see Iteration Virtual displacement, 407
Superposition,39, 103, 114
Surface traction, 358 Wave equation, 377-78, 542
Symmetry, in perturbation calculations, 232 Wavefunction,396
Systems of ordinary differential equations, Wave number, 380
theorems for, 57~5 Wave number vector, 522
Wavelength, 380
Taylor instability, 536 Wedge, propagation of discontinuities in,
Taylor series, 46, 47 399
Tensors, 449 Weighting function, 145
Term-by-term integration, 89 Well-posed problem, 123
Theoretical science, contrasted with applied Work,401-402,493
mathematics, 7 Work-energy principle, 403
Thermal diffusivity, 118 Wretched consistent approximations, 189, 231
Thermodynamics, 491-501
Thermostatics, 474 Young's modulus, 3~5
Time scale, 128
Torque, 465 Zermelo paradox, 109
Trajectories, 336 Zero of a function, 190-91

Lin Segel Applied Mathematics

Uploaded by

Copyright:

Available Formats

Lin Segel Applied Mathematics

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

Lin Segel Applied Mathematics

Uploaded by

Copyright:

Available Formats

What is the book about?

What is the book about?

What topics are covered in the book?

What topics are covered in the book?

Downloaded 10/13/15 to 128.84.127.209. Redistribution subject to SIAM license or copyright; see http://www.siam.org/journals/ojsa.

with material on elasticity by

Classics in Applied Mathematics

*First time in print.

Classics in Applied Mathematics (continued)

Copyright 1988 by the Society for Industrial and Applied Mathematics

Library of Congress Cataloging-in-Publication Data

THIS volume is a classic in the pedagogy of applied mathematics. Its three

Robert E. O'Malley, Jr.

THIS text, an introduction to applied mathematics, is concerned with the

STYLE AND CONTENT

effort is required to obtain a deep understanding of the problem. This effort

RELATIONSHIPS BETWEEN THIS TEXT AND VARIO US COURSES

University.) A second course, Introduction to Applied Mathematics, was

1.2 Introduction to the analysis of galactic structure 10

1.3 Aggregation of slime mold amebae 22

APPENDIX 1.l Some views on applied mathematics 31

2.2 Elements of perturbation theory, including Poincare's method for

2.3 A system of ordinary differential equations 57

3.2 Asymptotic series, Laplace's method, gamma function, Stirling's

3.3 A differenceequation and its limit 91

3.4 Further considerations pertinent to the relationship between

APPENDIX 3.1 0 and o symbols 112

4.1 Conduction of heat II s

4.2 Fourier's theorem 131

4.3 On the nature of Fourier series I 37

5.1 Other aspects of heat conduction 150

S.3 Brief introduction to Fourier transform 167

5.4 Generaliud harmonic analysis 171

6.3 Scaling 209

12.2 One-dimensionalelastic wave propagation 376

12.3 Discontinuous solutions 388

12.4 Work, energy, and vibrations 401

13.2 Kinematics of deformable media 418

13.3 The material derivative 426

15.2 Stability of a stratified fluid 515

15.3 Compression waves in gases 539

15.4 Uniform flow past a circular cylinder 551

THE PURPOSE of this book is to foster an appreciation of the nature

4 What Is Applied Mathematics? [Ch. 1

forces involved. The morphogenetic(form-producing) motion of the amebae

1.1 On the Nature of Applied Mathematics

An answer often given* to this philosophical question is that

6 What Is Applied Mathematics? [Ch. 1

APPLIED MATHEMATICS CONTRASTED WITH PURE

cannot be solved in terms of existing methods, or the nature of its solution

8 What Is Applied Mathematics? [Ch. 1

To give an illustration, the origin of hurricanes would not be a subject of

APPLIED MATHEMATICS IN ENGINEERING

Sec. J .J] On the Nature of Applied Mathematics 9

in question have a large intuitive element, but still there is something to be

IO What Is Applied Mathematics? [Ch. 1

CONCEPTS THAT UNIFY APPLIED MATHEMATICS

elliptic ( such as ::~ + ::~ = 0),

hyperbolic ( such as ::~ - ::~ = 0).