Texts in Applied Mathematics 28

Editors
].E. Mar5den
L. Sirovich
M. Golubit5ky
WJăger

Advisor
G. 10055

Springer-Science+Business Media, LLC

Texts in Applied Mathematics

I. Sirovich: Introduction to Applied Mathematics.

2. Wiggins: Introduction to Applied Nonlinear Dynamical Systems and Chaos.
3. Hale/KO!;ak: Dynamics and Bifurcations.
4. Chorin/Marsden: A Mathematical Introduction to Fluid Mechanics, 3rd ed.
5. Hubbard/West: Differential Equations: A Dynamical Systems Approach:
Ordinary Differential Equations.
6. Sontag: Mathematical Control Theory: Deterministic Finite Dimensional
Systems, 2nd ed.
7. Perko: Differential Equations and Dynamical Systems, 2nd ed.
8. Seaborn: Hypergeometric Functions and Their Applications.
9. Pipkin: A Course on Integral Equations.
10. Hoppensteadt/Peskin: Mathematics in Medicine and the Life Sciences.
II. Braun: Differential Equations and Their Applications, 4th ed.
12. Stoer/Bulirsch: Introduction to Numerical Analysis, 2nd ed.
13. Renardy/Rogers: A First Graduate Course in Partial Differential Equations.
14. Banks: Growth and Diffusion Phenomena: Mathematical Frameworks and
Applications.
15. Brenner/Scott: The Mathematical Theory of Finite Element Methods.
16. Van de Velde: Concurrent Scientific Computing.
17. Marsden/Ratiu: Introduction to Mechanics and Symmetry.
18. Hubbard/West: Differential Equations: A Dynamical Systems Approach:
Higher-Dimensional Systems.
19. Kaplan/Glass: Understanding Nonlinear Dynamics.
20. Holmes: Introduction to Perturbation Methods.
21. Curtain/Zwart: An Introduction to Infinite-Dimensional Linear Systems
Theory.
22. Thomas: Numerical Partial Differential Equations: Finite Difference
Methods.
23. Taylor: Partial Differential Equations: Basic Theory.
24. Merkin: Introduction to the Theory of Stability of Motion.
25. Naber: Topology, Geometry, and Gauge Fields: Foundations.
26. Polderman/Willems: Introduction to Mathematical Systems Theory:
A Behavioral Approach.
27. Reddy: Introductory Functional Analysis with Applications to Boundary-
Value Problems and Finite Elements.
28. Gustafson/Wilcox: Analytical and Computational Methods of Advanced
Engineering Mathematics.
29. Tveito/Winther: Introduction to Partial Differential Equations: A
Computational Approach.
30. Gasquet/Witomski: Fourier Analysis and Applications: Filtering, Numerical
Computation, Wavelets.
31. BrtEmaud: Markov Chains: Gibbs Fields, Monte Carlo Simulation, and Queues.
32. Durran: Numerical Methods for Wave Equations in Geophysical Fluid Dynamics.
Grant B. Gustafson Calvin H. Wilcox

Analytical and Computational Methods of

Advanced Engineering Mathematics

With 192 Illustrations

" Springer
Grant B. Gustafson
Calvin H. Wilcox
Department of Mathematics
University of Utah
Salt Lake City; UT 84112
USA

Series Editors
].E. Marsden L Sirovich
Control and Dynamical Sytems, 107-81 Division of Applied Mathematics
California Institute of Technology Brown University
Pasadena, CA 91125 Providence, RI 02912
USA USA

M. Golubitsky W]ăger
Department of Mathematics Department of Applied Mathematics
University of Houston Universităt Heidelberg
Houston, IX 77204-3476 69120 Heidelberg
USA Germany
Cover photo © Lightscape.

Mathematics Subject Classification (1991): 00-01, 00A06, 35-01

Ubrary of Congress Cataloging-in-Publication Daia
Gustafson, Grant B.
Analytical and computational methods of advanced engineering
mathematics. Grant B. Gustafson, Calvin H. Wilcox.
p. cm.-{Texts in applied mathematics ; 28)
Includes index.
ISBN 978-1-4612-6847-5 ISBN 978-1-4612-0633-0 (eBook)
DOI 10.1007/978-1-4612-0633-0
1. Engineering mathematics. 1. Wilcox, Calvin H. (Calvin Hayden)
II. Title. II. Series.
TA330.G87 1998
620'.OOl'Sl-dc21 97-49494

Printed on acid-free paper.

© 1998 Springer Science+ Business Media New York

Originally published by Springer-Verlag Berlin Heidelberg New York in 1998
Softcover reprint of the hardcover Ist edition 1998

All rights reserved. This work may not be translated or copied in whole or in part without the written permission of
the publisher (Springer-Science+ Business Media, LLC, except for brief excerpts in connection with reviews or
scholarly anaIysis. Use in connection with any form of information storage and retrieval, electronic adaptation,
computer software, or by similar or dissimilar methodology now known or hereafter developed is forbidden.
The use of general descriptive names, trade names, trademarks, etc., in this publication, even if the former are not
especially identified, is not to be taken as a sign that such names, as understood by the Trade Marks and
Merchandise Marks Act, may accordingly be used freely by anyone.

Production managed by Karina Mikhli; manufacturing supervised by Joe Quatela.

Typeset using authors' ID'pc files by The Bartlett Press, Marietta, GA.

98 76 S4 3 2 1
SPIN 10557758
 Preface

This book is a textbook of mathematical methods for undergraduate engineering and

science students. Such texts have been published under the title Advanced Engineering
Mathematics (AEM for short) ever since the appearance of C. R. Wylie's text of that name
in 1951. This preface begins with some remarks concerning the choices and constraints
that must be faced by the authors of an AEM text. It then describes the authors' new text
and how it differs from existing texts.
Role of AEM Texts in Engineering and Science Curricula. The prerequisite for most
AEM courses is one year of calculus (sometimes three semesters or four quarters).
Thus the AEM courses are intended for sophomore and junior engineering and sci-
ence students, and they are often the final mathematics courses for these students. By
de-emphasizing rigor and proofs they attempt to teach topics that are learned by math-
ematics students in their junior and senior years, after they have finished an advanced
calculus course.
Content of AEM Courses. During the last three decades there has been a strong con-
sensus as to the appropriate topics for an AEM course. These include ordinary differential
equations (first-order and linear second- and higher-order), Laplace transforms with ap-
plications to differential equations, linear algebra, vector calculus, Fourier analysis and
Sturm-Liouville theory and an introduction to partial differential equations and boundary
value problems of mathematical physics.
Use of Computer Technology. The development of flexible computer algebra systems
such as Mathematica, Maple, Matlab, and many others has greatly increased the size
vi PREFACE

and complexity of problems that students, and their teachers, can solve. The capability of
checking calculations and answers with electronic assistance is appreciated by instructors
and students alike.
Textbook vs. Handbook. Traditionally, engineers have liked to get double duty from
their AEM texts, using them first as textbooks during their student years, and later as
handbooks of mathematical methods during their professional lives. The second use has
led to very large books with many chapters and topics. Unfortunately, these two uses are
in conflict. A textbook should be written in a discursive, conversational style that allows
the student to read linearly and learn the material without constant help. By contrast, a
handbook should be terse and tightly organized to make it a convenient reference.
Time Allotted to AEM in Engineering Curricula. Most engineering curricula are very
full, crowding essentially five years of courses into four years. As a consequence, engineer-
ing faculties allow strictly limited time for mathematics and other service courses. In par-
ticular, this leads to crowded syllabi for the AEM courses and the omission of many topics.
Description of the New AEM Text. The Table of Contents provides a detailed topical
outline. This will be used to describe and comment on the manuscript.
Chapter 1. Numerical Analysis. The placing of this material at the beginning of the
book is new. Traditionally, a differential equation is considered solved if an implicit
equation for the solution is found. But a modern engineer or applied scientist does not
consider a problem to be solved until he or she can program a computer to generate
the values of the function on demand and to generate tables and graphs of the function.
Chapter 1 provides elementary methods to do this. The topics treated in Chapter 1
are direct applications of first-year calculus. They provide a seamless transition from
elementary calculus to ordinary differential equations.
Chapter 2. Ordinary Differential Equations of the First Order. Most AEM texts begin
with this topic because of its widespread use in applications. Here it is emphasized that a
first-order ordinary differential equation plus an initial condition is one of the best ways
to define and generate a function. The numerical algorithms presented in Chapter 2 are
natural extensions of the numerical integration methods of Chapter 1.
Chapter 3. Ordinary Differential Equations of Higher Order and Systems. Here
the emphasis is on second-order linear ordinary equations, a topic that pervades
mathematical physics (Chapters 4, 8, and 9). The structure theorem for solutions of
such equations provides concrete examples of the abstract vector spaces that will be
encountered in Chapter 5.
Chapter 4. Laplace Transform. This standard topic is still required by electrical
engineers. All AEM texts present it. It provides a flexible method to solve systems of
 PREFACE vii

linear ordinary differential equations with constant coefficients. Computer technology

has caused this text to have shorter Laplace tables and less emphasis on details like partial
fractions and simultaneous linear equations, compared to texts written in the 1960s.
Chapter 5. Linear Algebra. Linear algebra is developed in the context of solving systems
oflinear algebraic equations. General vector spaces are defined only after several concrete
examples have been studied. Matrix operations are defined only when they appear
naturally in solving linear systems. Computer algebra systems (CAS) provide a reliable
means for checking textbook computations. Largely, these systems extend the usefulness
of the basic text methods by increasing reliability and speed.
Chapter 6. Vector Analysis. Most first-year calculus textbooks include an introduction
to vector analysis. Moreover, science and engineering students are also introduced to
concepts of vector analysis in their introductory physics and engineering courses. For
this reason the new text reviews the basics topics only briefly before moving on to the
main theorems and applications.
Chapters 7, 8, and 9. These three chapters present an introduction to the boundary
value problems of mathematical physics and their solution by Fourier's method of
separation of variables.
Chapter 7. Partial Differential Equations of Mathematical Physics. This chapter
presents careful derivations of the best-known partial differential equations of mathemati-
cal physics. A novel feature is the use of uniqueness theorems to discover which boundary
and/or initial values are sufficient to determine the solution of a physical problem.
Chapter 8. Fourier Analysis and Sturm-Liouville Theory. The boundary value
problems of Chapter 7 and the search for product solutions are used to motivate the
study of Sturm-Liouville problems. This is done first in the special cases of Dirichlet
boundary conditions (Fourier sine series) and Neumann boundary conditions (Fourier
cosine series). Periodic boundary conditions then lead to the complete Fourier series.
The boundary value problem of heat diffusion in an infinite rod motivates the Fourier
integral theorem. The remainder of the chapter develops an elementary approach to
Sturm-Liouville theory and associated eigenfunction expansions.
Chapter 9. Boundary Value Problems of Mathematical Physics. This final chapter
uses Fourier's separation of variables method to solve a broad variety of boundary
value problems of mathematical physics. Problems solved include heat diffusion in one
dimension (8 problems), vibrating strings (4 problems), steady-state diffusion of heat
in plates (10 problems), transient diffusion of heat in plates (2 problems), vibrations of
drums (2 problems) and steady-state diffusion of heat in solids (2 problems). A final
section presents an alternative method based on the Laplace transform.
vi ii PREFACE

Additional Remarks on Differences Between the New Text and Existing Texts. The
focus of the new book is narrow. It does not try to be a treatise or handbook on engineering
mathematics.
This book has evolved over fifteen years into a text that is independent of any partic-
ular computing environment. This plan emerged in response to changes in computing
equipment and software systems, which tend to be subject to rapid flux, quickly dating
any text that is based upon specifics of a system or language. Instead, we adopted the
viewpoint that the mathematics is the fundamental element to be learned, and the soft-
ware system is certain to change, and soon! Accordingly, no computer-specific elements
appear in the main text or in the exercise statements.
Linear algebra is a strong point of the book. Unlike other texts that follow an approach
used in pure algebra texts, we use a practical approach based upon solving equations. All
ideas are introduced from this basic viewpoint, which allows engineers in their second
year to understand concepts that would otherwise be impossibly abstract.
Partial differential equations are introduced in the context of physical problems,
rather than in an abstract setting (classification into hyperbolic, parabolic and elliptic,
etc.), which is so common in other texts. Emphasis is on the solution of a variety of
models that arise in physics and engineering. Competing books have few applications
that are treated with so much completeness and depth.
Numerical analysis is treated as an extension of the calculus, and it is available
from the first days of the course. The text uses the numerical material as a bridge from
the mathematics to the applications. We understand that engineering involves graphs,
numbers, and experiments. The numerical methods provide a platform for interpretation
and evaluation of abstract results.
The chapters are as independent as possible. To achieve this it is sometimes necessary
to repeat or summarize material that was developed earlier. An example occurs in the
Sturm-Liouville theory of Chapter 8, which uses ideas and methods from the theory of
second-order linear ordinary differential equations. Here the text reviews the necessary
theory from Chapter 3.
Short One-Year Course. The shortest course over major topics is organized as shown
below. A quarter system one-year course does a short course in three quarters, Chapters
1-4, Chapters 5-6, Chapters 7-9. A semester system one-year course might do a short
course in two semesters, Chapters 1-5, Chapters 6-9. The pace in a short course is slow,
allOwing for diverse backgrounds and only a calculus prerequisite.

Chapter 1. Treat linear interpolation only. Skip most of 1.2 and all of 1.3 through 1. 7.
Introduce topics from Chapter 1 when demand dictates. Start Chapter 2 as
quickly as possible.
 PREFACE ix

Chapter 2. Skip exact equations 2.4 and numerical methods 2.7. Skip most applications
in 2.5.
Chapter 3. Skip power series solutions 3.4, numerical methods 3.5, and higher-order
linear equations 3.6.
Chapter 4. Skip applications to differential equations 4.4 and 4.5.
Chapter 5. Skip most of the fundamental theorem of linear algebra 5.5, treat just
dimensions less than 5 for determinants 5.6, and skip quadratic forms in 5.7.
Chapter 6. Already this chapter assumes a calculus course over the same topics. The
topics in 6.5 are used in Chapter 7. Skip unrelated topics in 6.1-6.4, gUided
by student background. Generally, three semesters of calculus will cover the
majority of topics in 6.1-6.4. Skip 6.6 for a short course.
Chapter 7. The chapter can be skipped and inserted into future lectures on topics from
Chapter 8 and Chapter 9. Chapter 7 is designed to be reference material for
reading rather than lecturing. The design originated with short courses where
there is not adequate class time for partial differential equations.
Chapter 8. Short courses do 8.1-8.5, 8.11-8.13 only. This leaves out the more difficult
topics and especially the Fourier integral.
Chapter 9. Short courses skip all topics that use Legendre and Bessel functions, multiple
integrals, Laplace and Fourier transforms. Only selected topics from 9.3 are
done in the lecture.
9.1: Problems 9.1.1, 9.1.2, 9.1.3, 9.1.4.
9.2: Problems 9.2.1, 9.2.2, 9.2.3.
9.3: Problems 9.3.1, 9.3.2, 9.3.7, 9.3.8, 9.3.9.

Normal One-Year Course. This course is defined as the short course above plus topics
selected to fit the curriculum. The pace through the text is faster, and much of the
workload is passed on to the student. It is suitable for well-prepared calculus students who
can dedicate time to independent learning. In designing a course, we suggest beginning
with the selection of topics from Chapter 9. For example, selection of a problem that
uses the Fourier transform requires deeper coverage of Chapter 8 and possibly Chapter
6. One critical issue for course design is whether or not to study power series methods
in the first term. This topic generally consumes five or more lectures. Some instructors
prefer to do the topics in the same term as the Bessel and Legendre applications.
Intensive One-Year Course. This course covers all the short course topics for the term
in 70% of the allotted time, e.g., six weeks for the quarter system and nine weeks for the
semester system and nine weeks for the semester system. Satisfactory completion of the
short course topics is scored as a passing grade in the course. To earn an honor grade,
the remaining 30% of the course is devoted to new topics from the text.
X PREFACE

Solutions Manual. The manual is a separate publication, which is available with

textbook orders. It consists of two parts published as one volume. Part I contains the
answers to all exercises. Part II contains complete solutions to all exercises, including
computer programs in Maple.
Additional Exercises with Maple Notes. Published electronically at
www . math. utah. edu are class-tested J6Tp<. sources with answers and solution notes.
Although the Maple notes are extensive, there is no Maple requirement in the exercises
themselves. The sources can be compiled and printed with or without the notes and
answers. The author's email address is gustaf son@math. utah. edu.
Errata. Errors found in the published text will be available electronically at
www . math. utah. edu and via the Springer website www. springer-ny . com. Please
make contributions to the errata list directly to the authors.
Acknowledgments. The authors wish to thank the Electronic Publishing and Book
Production departments at Springer-Verlag New York. They are espeCially indebted to
Achi Dosanjh, Mathematics Editor, and Karina Mikhli, Production Editor.

Salt Lake City, Utah Grant B. Gustafson

Calvin H. Wilcox
 Contents

Preface v

Introduction xv

1 Numerical Analysis 1
1.1 The Nature of Numerical Analysis. 1
1.2 Polynomial Interpolation 2
1.3 Numerical Integration and Differentiation . 22
1.4 Solution of Equations 41
1.5 Inverse Functions . . .. 55
1.6 Implicit Functions . 64
1.7 Numerical Summation of Infinite Series. 70

2 Ordinary Differential Equations of First Order 76

2.1 The Nature of Differential Equations 77
2.2 Separable Equations . 83
2.3 Linear First-Order Equations 92
2.4 Exact Equations 98
2.5 Applications to Some Second-Order Equations 106
2.6 The Initial Value Problem 123
2.7 Numerical Methods for the Initial Value Problem 127
xii CONTENTS

3 Ordinary Differential Equations of Higher Order 138

3.1 Examples from Engineering and Physics . . . 139
3.2 Linear Second-Order Equations - Structure of Solutions 155
3.3 Linear Second-Order Equations with Constant Coefficients 164
3.4 Linear Second-Order Equations with Analytic Coefficients 182
3.5 Numerical Methods for Second-Order Equations 204
3.6 Linear Equations of Order n > 2 . . . . . . . . 211

4 The Laplace Transform 218

4.1 The Nature of the Laplace Transform 220
4.2 The Laplace Transforms of Some Elementary Functions 224
4.3 Operational Rules for the Laplace Transform .. . 231
4.4 Applications to Differential Equations . . . . . . . 249
4.5 Applications to Systems of Differential Equations . 263

5 Linear Algebra 276

5.1 Systems of Linear Equations . 277
5.2 The Gauss Elimination Method 284
5.3 Vector Spaces . . . . . . . . . 294
5.4 Matrices and Matrix Algebra. . 305
5.5 The Fundamental Theorem of Linear Algebra. 314
5.6 Determinants and Cramer's Rule 326
5.7 Eigenvalues and Eigenvectors 345

6 Vector Analysis 364

6.1 Vector Algebra 366
6.2 Vector Calculus of Curves in Space 381
6.3 Vector Calculus of Surfaces in Space 392
6.4 Calculus of Scalar and Vector Fields. 403
6.5 Integral Theorems of Vector Calculus 417
6.6 X-Ray Diffraction and Crystal Structure 431

7 Partial Differential Equations of Mathematical Physics 440

7.1 Vibrating Strings: D'Alemberts Wave Equation 442
7.2 Heat Diffusion in Rods: Fourier's Heat Equation .. 451
7.3 Heat Diffusion in Plates . . . . . . . . . . . . . . . 456
7.4 Steady-State Heat Diffusion in Plates: The Laplace Equation 461
7.5 Vibrations of Drums . . . . . . . . . . . . . . . . . . . . . 463
 CONTENTS xiii

7.6 Heat Diffusion in Solids . . . . . . . 468

7.7 Steady-State Heat Diffusion in Solids 470

8 Fourier Analysis and Sturm-Liouville Theory 472

Part I Fourier Series . . . . . . . . . . . . . . . . . . . 474
8.1 Dirichlet Boundary Conditions and Fourier Sine Series . 474
8.2 Orthogonality and Fourier Coefficients . . . . . . . . . 479
8.3 Convergence of Fourier Sine Series . . . . . . . . . . . 480
8.4 Neumann Boundary Conditions and Fourier Cosine Series . 486
8.5 Periodic Boundary Conditions and the Complete Fourier Series 492
8.6 Proofs of the Convergence Theorems (Optional) 501
Part II Fourier Integrals . . . . . 511
8.7 Heat Diffusion in an Infinite Rod 512
8.8 Orthogonality Calculation. . . . 514
8.9 The Fourier Integral . . . . . . . 516
8.10 Fourier Sine and Cosine Integrals 529
Part III Sturm-Liouville Theory. 533
8.11 Heat Diffusion in Nonhomogeneous Rods 533
8.12 Sturm-Liouville Problems: Basic Theory . 534
8.13 Construction of Eigenvalues and Eigenfunctions 539
8.14 Singular Sturm-Liouville Problems . . . . . . . 548

9 Boundary Value Problems of Mathematical Physics 556

9.1 Heat Diffusion in One Dimension. . . . 557
9.2 Vibration of Strings and Traveling Waves 579
9.3 Steady-State Diffusion of Heat in Plates 598
9.4 Transient Diffusion of Heat in Plates. . 630
9.5 Vibrations of Drums . . . . . . . . . . 638
9.6 Steady-State Diffusion of Heat in Solids 643
9.7 The Laplace Transform Method . . . . 653

Appendix: Answers and Hints to Selected Exercises 662

References 709

Index 711
 Introduction

This book presents an introduction to the methods by which mathematical analysis and
the computing power of modern digital computers are combined to solve problems of
engineering science.
Most physical phenomena obey mathematical laws that can be formulated as
differential equations. This astonishing fact was first discovered by Isaac Newton in the
years 1665-1666 when he formulated the laws of mechanics as differential equations
and used them to predict the orbits of the planets. During the three centuries that
have followed Newton's discoveries many differential equations that govern physical
phenomena have been found. They include Euler's equations for the dynamics of rigid
bodies, the Navier-Stokes equations for the motion of viscous fluids, the Cauchy-Green
equations for the static and dynamic behavior of elastic bodies, Fourier's equations for the
diffusion of heat in solids, Kirchhoff's equations for electrical circuits, Maxwell's equations
for the electromagnetic field and Schrbdinger's equation for the dynamics of atoms and
molecules, to mention only the best-known examples. An examination of the recent
engineering and scientific literature reveals an almost unlimited number of differential
equations that have been introduced to model physical, chemical and biological systems
and processes. The goal of this book is to provide an introduction to the differential
equations that arise in engineering science and to the analytical and computational
methods that are used to solve them.
Mathematical Analysis of a Physical Problem. The mathematical analysis of a
physical problem involves the following four distinct steps.
xvi INTRODUCTION

1. Mathematical Formulation of the Physical Problem. This step is some-

times called "constructing a mathematical model." It involves quantifying the
variables in the physical problem, deciding what physical hypotheses and sim-
plifying assumptions will be made, and then translating these into a well-defined
mathematical problem.
2. Analytical Solution of the Mathematical Problem. In this step the methods of
mathematical analysis are used to construct the function, or functions, that solve
the mathematical problem of step 1. Such methods include algebraic calculations,
integration, differentiation, the summation of infinite series, the use of integral
transforms, among many others.
3. Numerical Computation of the Solution of the Mathematical Problem.
The results of step 2 typically provide the solution of the mathematical problem
as a definite integral, or the sum of an infinite series, or the solution of a set of
linear or nonlinear equations, or in one of a number implicit forms. The purpose
of step 3 is to provide a definite numerical procedure, or algOrithm, that permits
the effective numerical computation of the solution. A computer can then be
used to compute the solution and to display it in tabular or graphical form.
4. Physical Interpretation of the Solution. The final step is to use the results
of steps 2 and 3 to understand the original physical problem and its solution,
including the phenomena that may occur, how they depend on the various
parameters of the problem, and how the results may be used to answer practical
questions. Frequently this will lead to refinements of the original mathematical
model, additional analysis and computation.

At this point some readers will ask why we don't simply "put the problem on the
computer" at the outset, i.e., eliminate steps 1, 2, and 4. This question reflects a lack of
appreciation of the complementary roles of theory and computation in applied science.
In order to compute anything there must be equations to be solved. If the solutions of
these equations describe the states of a physical system then the equations themselves
are a mathematical model of the system. They can come only from considerations of
the kind indicated in steps 1 and 2. Moreover, the results of a computation can be
related to the states of a physical system only by means of some theory or model. Of
course, in some cases step 3 may be emphasized over the other steps. Nevertheless, the
successful application of mathematical analysis and computation to any physical problem
necessarily involves all four steps.
The method outlined above will now be illustrated by means of a simple example
for which all four steps can be carried out easily The example will also show the origin
and method of solution of a simple differential equation.
 INTRODUCTION xvii

A Flask Cooling Problem. A flask of water (see Figure 1) is heated to the boiling
point (100 degrees Celsius). It is then removed from the heat and stirred constantly while
it cools in a room with air temperature 20°e. After 10 minutes the water temperature is
60°e. Ihe problem is to predict when the water temperature will be 40°C and when it
will have cooled to room temperature (20°C).
Step 1. Mathematical Formulation. Ihe wording of the problem suggests the
follOwing hypotheses.
When removed from the heat the water is uniformly at 100°e.
Because of stirring, the water temperature at any instant is uniform throughout
the flask.
Ihese hypotheses suggest that the water temperature I, expressed in degrees Celsius,
is a definite function of the time. If

t = elapsed time, in minutes, after removal

of the flask from the heat

then the hypotheses imply that

(1) T = u(t), t ~ 0,

where u(t) is an as yet unknown function of t. It is given that

(2) u(O) = 100, u(lO) = 60.

Moreover, experience tells us that u(t) is a steadily decreasing function of t that satisfies
20 ~ u(t) ~ 100. Ihe problem is to determine the unknown function u(t).
A physical hypotheSiS is needed to determine the rate at which the flask of water
cools. A good approximation is Newton's law of cooling. In the present context it states
that the rate of decrease of I (i.e., -dI/dt) is proportional to the difference between
the instantaneous flask temperature and the ambient air temperature (i.e., T - 20). Ihis
leads to the differential equation
dT
(3) - = -h(T - 20)
dt '

FIGURE 1. Cooling of a Flask

xviii INTRODUCTION

where h is a positive constant, as yet unknown, that presumably depends on the size and
shape of the flask. The minus sign in (3) is needed because dTldt < 0 and T - 20 > O.
Step 2. Solution of the Mathematical Problem. Equation (3) is understood to
mean that the unknown function u(t) satisfies
duet)
(4) - - = -h{u(t) - 20} for each t ~ O.
dt
Now, the chain rule of calculus implies that

d 1 duCt)
dt log{u(t) - 20} = {u(t) - 20} dt .

Hence, equation (4) is equivalent to

d
(5) -log{u(t) - 20} = -h for each t ~ O.
dt

The fundamental theorem of calculus, applied to (5), gives the relation

(6) log{u(t) - 20} = -ht +C for all t ~ 0,

where C is some constant. This may be rewritten as u(t) - 20 = e- ht+C = KC ht , or

(7) u(t) = 20 + Ke- ht for each t::: o.

Finally, hypotheses (2) determine the values of the two constants K and h. First, (2) and
(7) give u(O) = 100 = 20 + K, or K = 80, so

(8)

Equations (2) and (8) give u(lO) = 60 = 20 + 80e- 10h , whence C 10h 112, or
h = (0.1) log 2, and

(9)

Hence, combining equations (7), (8) and (9), the solution of the mathematical problem is

(10) T = u(t) = 20 + 80 . rt/IO.

It is now easy to verify directly that (10) defines a function u(t) that satisfies Newton's
law (3) and the conditions in (2).
Checking the Answer. To verify that equation (10) satisfies the initial conditions,
compute u(O) = 20 + 80 = 100 and uclO) = 20 + 80(112) = 60. The differential
 INTRODUCTION xix

T
100
80
60
40
20 t
0 20 40 60 80 100 FIGURE 2. Temperature versus time

equation (3) is satisfied, because by (0), U'(t) = 0 + 80( -h)e- ht and -h(u(t) - 20) =
( -h)(80e- ht ).
Step 3. Numerical Computation. This step is easy in the present case. The solution
may be written in the form

u(t) = 20 + 80e-(o.J)t log 2.

In this form values of u(t) can be computed easily on a pocket calculator. The graph of
u(t) appears in Figure 2.
Step 4. Physical Interpretation. It is clear from equation (10) that the temperature
of the water is a strictly decreasing function of t such that 20 < T = u(t) S 100 for all
t ~ 0 and

(ll) lim u(t) = 20.

t---++oo

From equation (0) it is clear that u(t) = 40 when t = 20 minutes. More generally, the
temperature difference u(t) - u(oo) = u(t) - 20 is cut in half every 10 minutes. The
answer to the question "when does the water reach room temperature" is "never,"
because u(t) > 20 for all t ~ O. If temperature differences ofless than 0.1 °C cannot be
measured, then the water is effectively at room temperature when u(t) = 20.1°C, or

20.1 = 20 + (80)r tllO ,

or

2t/IO = 800.

Taking logarithms gives

t = 10 log 800/log 2 = 96.4 minutes.

A Refinement with Stirring Effects. When a liquid is stirred, the mechanical

energy of the stirrer is not lost but appears as heat in the liquid. The effect can be included
XX INTRODUCTION

in the cooling model. Assume that the stirrer adds energy to the water at the constant rate
calories
A--.
minute
Let the specific heat of the water be
calories
c
(gram) (oC)

and let the flask contain a mass of water of

M grams.

Then, in the absence of cooling, the stirrer would raise the temperature of the water at
the rate
cal 1 DC A °C
A-x--=-=Q-
min eM cal eM min
When this effect is added to the cooling by Newton's law, one gets the model equation
dT
dt = -h(T - 20) + Q.
It will be convenient to write this as
dT
(12) dt = -h(T - 20 - To),

where the temperature To is defined by

Q
(13) To = h'
Equation (12) has the same form as (3) but with 20 replaced by 20 + To. Hence, as
in the first case, T = u(t) must satisfy

u(t) = 20 + To + Ke- ht .
Conditions (2) give first K = 80 - To and then

40 - To -lOh
(14) ---=e
80 - To
On combining this with (13) it is found that the value of h is determined implicitly by
the equation

40h - Q -lOh
(15) ----e
80h - Q - .
 INTRODUCTION xxi

If, for example, A, c and M have the values

calories
A=lO---
minutes'
calories
c = 1 ____________---,--
(gram)(oC) ,
M = 1000 gram,

then Q = AI(cM) = 0.01 and h must satisfy

40h - 0.01 -lOh

(16) -----e
80h - 0.01 - .
It will be shown in Chapter 1 that equation (16) has exactly one positive solution
h = 6.9495 X 10-2 and To = O.Ollh = 0.1439°C. The temperature function may then
be written

(17) u(t) = 20 + To + (80 - To)e- ht

= 100 + (0.~1 _ 80) (1- e- ht ).

Note that by (14)

h= ~ 10 (80 - To)
10 g 40 - To '
and hence the first equation of (18) may be written

40 - To)U10
(18) u(t) = 20 + To + (80 - To) (
80 - To
From this form it is clear that if To is small, as in the numerical example, then u(O is
nearly the same as the solution (10) of the first model problem, and the effect of the
stirring can be ignored. It is also clear that the time t at which T = u(t) = 40° will be
greater than for the first model. This t can be found from (18) by putting u(t) = 40 and
solving for t. Numerical methods for doing this are presented in Chapter 1. Note that
the temperature function u(t) defined by (18) is again a strictly decreasing function of t.
However, in the present case,

u(t) > 20 + To for every t ::: 0

and

lim u(t) = 20 + To > 20.

t--++oo
xxii INTRODUCTION

Hence in this more refined model the flask never appears to reach room temperature
unless To is less than the smallest measurable temperature difference.
Each of the analytical and computational methods presented in this book is illustrated
by application to one or more physical problems. In many cases all four steps of the
analysis, as outlined above, will be presented. In other cases the main emphasis will be
on steps 2 and 3, which usually contain most of the mathematical work. In such cases
the construction of the model (step 1) and the interpretation of the results (step 4) may
be discussed briefly, or references to specialized textbooks may be given.
Analytical and Computational Methods of
Advanced Engineering Mathematics
CHAPTER 1
 Numerical Analysis

1.1 The Nature of Numerical Analysis

A digital computer can be programmed to solve a mathematical problem only if a
numerical algorithm for the problem is available. Here the term numerical algorithm
means a definite sequence of machine-executable arithmetic and logical operations that
when applied to the data of the problem will produce, in a finite number of steps, the
solution of the problem, correct to a prescribed degree of accuracy. of course, the number
of steps required to produce the solution may depend on the degree of accuracy required.
The principal theme of this book is the application of computer-based methods to the
solution of problems of engineering and applied science. Hence, numerical algorithms
playa central role throughout the book. New ones will be found in nearly every chapter.
This first chapter presents some of the simplest and most basic algorithms, together
with some applications to physical problems. These algorithms have been placed at the
beginning because they are used repeatedly in the remainder of the book.
The theory underlying the construction and evaluation of numerical algorithms is
called numerical analysis. This theory is based directly on differential and integral
calculus, and especially Taylor'S theorem with remainder.
The history of numerical analysis extends back to the beginnings of the calculus and
the work of Newton. In fact, two of Newton's methods are presented in this chapter.
More recently, the invention and development of electronic digital computers, beginning
2 NUMERICAL ANALYSIS

in the 1940s, stimulated vigorous research in numerical analysis. The search for new,
more effective algorithms continues at a high level and may be expected to do so in the
foreseeable future.
The enormous power of modem computers has led many laymen and some scientists
to believe that any problem can be solved by "putting it on the computer." This is certainly
not true. First of all, there are problems for which no satisfactory mathematical model has
been found. It should be clear from the discussion in the Introduction that computers
cannot be applied to such problems. Secondly, it has been found in recent years that
some problems possess chaotic solutions that are essentially unpredictable. This class
probably includes some problems of turbulent fluid flow and weather prediction. Finally,
some applications lead to mathematical problems whose computer solution would take
so long as to be beyond the capabilities of any existing or foreseeable computers.
In recent years many prepackaged computer programs that implement some of
the algorithms presented in this book have become available. For this reason students
of engineering and science sometimes ask why they should take the trouble to learn
the principles underlying such programs. The Simplest reply to this question is that
all algorithms, and all programs, have their limitations. Users must understand these
limitations in order to apply the programs successfully Moreover, engineering science
leads to an ever-increasing number of mathematical models, most of which do not fit
preexisting programs. Hence it is necessary for engineers to understand the principles
that underlie the analysis of engineering problems. In fact, many engineers become
involved in designing large programs to simulate complex physical systems. For the
successful completion of such projects it is essential that they understand the principles
of numerical analysis.

1.2 Polynomial Interpolation

The computational analysis of many engineering problems requires that one or more
functions be stored in the computer so that their values can be used at various steps
in the computation. The functions may be defined either by experimental data or by
mathematical algorithms. In either case only a finite number of function values can be
stored in the computer memory A numerical algOrithm is then needed to fill in, or
interpolate, the remaining values. A common method is polynomial interpolation. It
will be introduced by means of two problems.

Problem 1. Filling of a Reservoir. The flow of water into a reservoir is measured during
a 24-hour period by means of a flow meter whose output is recorded on magnetic
tape. The tape then contains a complete analog record of a function q = q(t), where
 1.2 POLYNOMIAL INTERPOLATION 3

t is the elapsed time in hours from the beginning of the measurements and q(t) is the
instantaneous rate of flow at time t, in cubic feet per hour (abbreviated Cfh).l

Several problems connected with such flow records are solved in this chapter and
the next chapter. The first is to compute the total volume V of water in the reservoir as
a function of t. If Vo is the volume present at t = 0 then V is given by

(1) V = Vo + it q(r) dr, 0::: t ::: 24 hr.

This function is computed numerically in Section 1.3 below.

The function q(t) will be stored in the computer by sampling its values on the tape at
a set of t-values to, tl, t2,"" tn by means of an analog-to-digital converter. The sample
values qo = q(to), ql = q(tl), q2 = q(t2), ... , qn = q(tn) will then be stored in computer
memory, and values of q(t) for t-values distinct from to, tl, t2, "', tn will be produced
by a numerical algOrithm that uses the sample values qo, ql, q2, ... , qn' As a specific
example, the t-values to = 0 hr, tJ = 1 hr, t2 = 2 hr, ... , t24 = 24 hr will be used.
The corresponding data are given below in Table 1. Two interpolation methods will be
applied to these data. First an example of a mathematically defined function is given.

Problem 2. The Bessel Function]o(x). The function]o(x), defined by the equation

(2) ]o(x) = -1
7(
l
0
rr
cos(x sin t) dt,

is called the Bessel function of order zero. 2 It occurs in the solution of many problems of
engineering and science, and for this reason it has been studied extensively Two standard
reference works are [McL] and [A-SJ.
The properties]o(O) = 1, lJo(x) I ::: 1 andio( -x) =.]o(x) follow immediately from
equation (2). Table 2 below is a short table of ]0 (x) for 0 ::: x ::: 6.0 and is accurate
to four decimal places. It is shown below how polynomial interpolation can be used to
compute values of]o(x) to table accuracy for any x on the interval 0.0 ::: x ::: 5.9.
The Interpolation Problem. In problems 1 and 2 the values of a function j(x)
are known at a finite set of x-values, say Xo < Xl < X2 < ... < x n. The notation
jo = j(XO),fl = j(xl),h = j(X2) , .. . ,fn = j(xn) will be used for these values. The
numbers XO,Xl,X2, ... ,Xn and jO,fl ,h, ... ,fn are called the nodes and nodal values of
j(x), respectively They are to be stored in computer memory A numerical algorithm is
then to be used to compute j(x) for other values of x.

'The conventional unit is cubic feet per second (abbreviated cfs). Here it will be more convenient to use cfh.
2The function is named for Friedrich Wilhelm Bessel (1784-1846). German astronomer and mathematiCian, who
applied it to the solution of an astronomical problem in a paper published in 1824.
4 NUMERICAL ANALYSIS

TABLE l. Sample flow meter data

t, (hr) qk (cfh) tk (hr) qk (dh) tk (hI) qk (cfh)

0.0 1.49 x 10' 9.0 2.26 X 10' 18.0 1.98 x 10'

1.0 1.34 X 10' 10.0 1.90 X 10' 19.0 1.63 x 10'
2.0 1.41 X 10' 11.0 2.15 x 10' 20.0 1.52 X 10'
3.0 1.64 X 10' 12.0 2.16 X 10' 21.0 1.52 x 105
4.0 1.49 X 10' 13.0 2.15 X 10' 22.0 1.34 X 10'
5.0 1.71 X 105 14.0 2.32 x 10' 23.0 1.31 X 10'
6.0 1.80 X 10' 15.0 2.22 x 10' 24.0 1.42 x 105
7.0 1.79 X 10' 16.0 1.98 x 10'
8.0 2.18 X 10' 17.0 2.08 X 10'

Linear Interpolation. The simplest polynomial interpolation method is linear inter-

polation. In this case J(x) is approximated by a linear function of x between each pair
of consecutive points Xk, Xk+l. Graphically, this means that the graph of J(x) is approxi-
mated by a polygonal line joining the consecutive points (xo,fo), (XI,fI), ... , (xn,fn) in
the plane.
Linear interpolation produces a function PI (x), defined for Xo ::: x ::: x n , such
that PI(X) is a linear function of x on each subinterval (Xk,Xk+l) and PI (Xk) = Jk for

TABLE 2. Jo(x) for 0.0 ::5 x ::5 5.9

x ]0 (x) x ]0 (x) x ]0 (x) x ]0 (x)

0.0 1.0000 1.5 0.5118 3.0 -0.2601 4.5 -0.3205

0.1 0.9975 1.6 0.4554 3.1 -0.2921 4.6 -0.2961
0.2 0.9900 1.7 0.3980 3.2 -0.3202 4.7 -0.2963
0.3 0.9776 1.8 0.3400 3.3 -0.3443 4.8 -0.2404
0.4 0.9604 1.9 0.2818 3.4 -0.3643 4.9 -0.2097
0.5 0.9385 2.0 0.2239 3.5 -0.3801 5.0 -0.1776
0.6 0.9120 2.1 0.1666 3.6 -0.3918 5.1 -0.1443
0.7 0.8812 2.2 0.1104 3.7 -0.3992 5.2 -0.1103
0.8 0.8463 2.3 0.0555 3.8 -0.4026 5.3 -0.0758
0.9 0.8075 2.4 0.0025 3.9 -0.4018 5.4 -0.0412
1.0 0.7652 2.5 -0.0484 4.0 -0.3971 5.5 -0.0068
1.1 0.7196 2.6 -0.0968 4.1 -0.3887 5.6 0.0270
1.2 0.6711 2.7 -0.1424 4.2 -0.3766 5.7 0.0599
1.3 0.6201 2.8 -0.1850 4.3 -0.3610 5.8 0.0917
1.4 0.5669 2.9 -0.2243 4.4 -0.3423 5.9 0.1220
 1.2 POLYNOMIAL INTERPOLATION 5

k = 0,1, . .. , n. To find an algorithm for computing PI (x), note that we may write

(3)

Here ao and al are to be computed from the equations

PI (Xk) = fk = ao,
PI (Xk+l) = fk+1 = ao + al (Xk+1 - Xk),

which have the unique solution

(4) ao = fk,
a] = (jk+l - aO)!(xk+l - Xk).

Hence PI (x) may be computed for any given number x such that Xo :s x :s Xn by the
following alogrithm.

Linear Interpolation Algorithm.

(5) If x = Xo then PI (x) = fo.

If Xo < x :s x n, compute k such that Xk < x :s Xk+1
(k is the largest integer such that x - Xk > 0).
Then compute ao, al by equations (4).
Then compute PI ex) by (3).

For the data of Table 1 the graph of the interpolant PI ex) is shown in Figure 1 below.
Uniformly Spaced Nodes. If

Xk+l - Xk = h = constant for k = 0,1, ... , n - 1,

PI

232000.0

131000.0
o 12 24

FIGURE 1. Linear interpolation of the data ofTable 1

6 NUMERICAL ANALYSIS

so that all pairs of consecutive nodes have the same separation h, then the last two steps
of the algorithm for Pi (x) are usually replaced by

(6) Compute r = (x - xk)!h

Compute t:.Jk = Jk+i - Jk
Compute Pi (x) = Jk + rt:.Jk.
The agreement of this with the preceding algorithm may be seen by combining (3), (4)
and Xk+i - Xk = h to get

Note that r is a fraction of the distance from Xk to Xk+i, so that 0 .:::: r .:::: 1.
The Error in Linear Interpolation. The function value Pi (x) will differ from j(x) for
most x-values other than the nodes. The error is given by the error function

(7) Ei (x) = J(x) - Pi (x)

(see Figure 2). The error may be expected to be small when Xk+i-Xk is small. Quantitative
estimates of Ei (x) may be obtained from a theorem (see [Ch-K]) that states that iff" (x) , the
second derivative of j(X) , is continuous, then for each number x such that Xk < x < Xk+l
there is a second number ~ such that Xk < ~ < Xk+i and
1 N
(8) Ei (x) = l(x - Xk)(X - Xk+l)j (~).

Equation (8) cannot be used to compute Ei (x) exactly because the value of ~ is
not known. However, useful estimates may be found by maximizing the factors on the
right-hand side of (8). It is a simple maximum problem to show that

(9)

FIGURE 2. Linear interpolation

 1.2 POLYNOMIAL INTERPOLATION 1

Hence, if M2 is any number such that

(10)

then it follows from (S) that

1 2
(11) IEl(X)1 s Slxk+l-xkl M2 for Xo S x S x n·

In particular, for uniformly spaced nodes we have the estimate

1 2
(12) IE] (x) I S Sh M2 for Xo S x S x n ·

As a simple application of (12) consider the function J(x) = cosx. In this case rex) =
- cos x and we may take M2 = 1. Thus the error in the linear interpolation of the cosine
function never exceeds h2/S. For example, to have lEI (x) I S 10-4 it is enough to require
that h2/8 S 10- 4 , i.e., h2 S 8 X 10- 4 or h S (8 X 10- 4)1/2 = 0.0283.
Quadratic Interpolation. Higher accuracy can be obtained if J(x) is approximated by
a quadratic polynomial in x on consecutive pairs of intervals, i.e., on

where it is assumed that n = 2m is even. Graphically, this means that on each interval
(X2k,X2k+2) the graph ofJ(x) is approximated by the unique parabola that passes through
the points (X2k,f2k), (X2k+l,f2k+l) and (X2k+2,f2k+2); see Figure 3.
This process produces a piecewise quadratic function P2 (x) whose graph is a parabola
on each interval (X2k, X2k+2) and that satisfies P2 (Xk) = Jk for k = 0, 1, ... ,2m = n. It
will be convenient to write P2 (x) in the form

(13)

___ f

t
12k P2(X) hk+l /2k+2

~ i ~ ~ ;0 X
X2k X X2k+l X2k+2 FIGURE 3. Quadratic interpolation
8 NUMERICAL ANALYSIS

where the constants ao, al and a2 are to be computed. Equation (13) clearly defines
a quadratic polynomial on (X2k,X2k+2). Moreover, the interpolation conditions at the
consecutive nodes X2h, X2k+l, X2k+2 give the equations

(14) P2(X2h) = hk = ao,

P2(X2k+l) = hk+l = ao + al (X2k+l - X2k),

P2(X2k+2) = hk+2 = ao + al (X2k+2 - X2k)

+ a2(x2k+2 - X2k) (X2k+2 - X2k+l) ,

which have the unique solution

(15) ao = hk'
hk+l - ao
al = ,
X2k+l - X2k
hk+2 - ao - al (X2k+2 - X2k)
a2 = .
(X2k+2 - X2k) (X2k+2 - X2k+l)

Hence P2(X) may be computed for any given number such that Xo .::s x .::s X2m by the
following algorithm.
Quadratic Interpolation Algorithm.
(16) If x = Xc, then P2(X) = fa.
If Xo < x .::s X2m, find k such that X2k < x .::s X2k+2
(k is the largest integer such that x - X2k > 0).
Compute ao, al, a2 by equations (15).
Compute P2(X) by (13).

For the flow meter data of Table 1, the quadratic interpolant P2 (x) is illustrated by
Figure 4 below.

P2
232000.0

131000.0
o 24

FIGURE 4. Quadratic interpolation of the data of Table 1

 1.2 POLYNOMIAL INTERPOLATION 9

Uniformly Spaced Nodes. If the node separations are all equal, Xk+l - Xk = h for
k = 0,1, ... , 2m - 1, then P2(X) is often written as
1 2
(17) P2(X) = hk + rt..hk + -r(r - l)t.. ilk for X2k:::: x:::: X2k+2,
2
where

(18) t..hk = j2k+l - j2k, t..hk+l = hk+2 - hk+l,

(19) t.. 2j2k = t..hk+l - t..hk,

and
X -X2k
(20) r = ---h
for X2k:::: x :::: X2k+2.

The last two steps of the algorithm for P2 (x) may then be replaced by

(21) Compute r from (20).

Compute t..hk, t..j2k+l from (18).
Compute t.. 2hk from (19).
Compute P2(X) from (17).

Note that °: :
r :::: 2 when X2k :::: x :::: X2k+2. Moreover, r = 0 when x = X2k, r = 1
when x = X2k+ 1, and r = 2 when x = X2k+2. It follows from these facts that (17) defines
the same quadratic polynomial as (13).
The Error in Quadratic Interpolation. This error is given by

(22)

It is shown in [Ch-KJ that if the third derivativej"'(x) is continuous, then for each x
such that X2k < x < X2k+2 there is a number ~ such that X2k < ~ < X2k+2 and

(23) E2(X) 1 -
= 6(x X2k) (x - X2k+l) (x - X2h+2)j '" (~).

If the nodes are uniformly spaced with Xk+l - Xk = h, then it is a simple maximum
problem to show that for X2. :::: x :::: X2k+2

(24)

Combining (23) and (24) gives the error estimate

h3M3
(25) IE2(X)1 :::: - - for Xo:::: X:::: X2m,
9v3
10 NUMERICAL ANALYSIS

where M3 is any number such that

(26) M3 :::: max If"'(~)1 for Xo S ~ S X2m·

In the example of the function f(x) = cos x we have 1'"(x) = sin x, and hence M3 = 1
is a suitable value. Thus the error in quadratic interpolation of the cosine function never
exceeds h3 /(9J3). In particular, to have IE21 S 10-4 it is sufficient to require that
h3 S 9J3 X 10-4 . On solving the equation

h3 = 9J3 X 10- 4 = 1.5588 X 10-3

by means of logarithms, we get h = 0.1159. Note that this is more than four times as
large as the maximum permissible h for the same accuracy with linear interpolation.
Error Control for Experimental Data. The error estimate (12) for linear interpolation
implies that for functions f(x) with continuous second derivative 1"(x) , the error can
be made less than any prescribed tolerance by chOOSing the sampling step size h
small enough. The estimate (25) implies that a similar statement holds for quadratic
interpolation whenever 1'" (x) is continuous. For experimental data, such as the flow
meter data of Problem 1, information about the derivatives off(x) is usually not available.
Hence no a priori determination of h can be based on the error estimates (12) and (25).
In such cases a suitable value of h must be found by trial and error. To do this, a first value
h = hI is selected, and PI (x) (or P2(X)) is computed for a representative set of x-values,
such as the midpoints of the intervals (Xk, Xk+I). These values are then compared with the
experimental values. If the errors are too great, then h is reduced, typically to h = ~ hI, and
the computation is repeated. This process is continued until a suitable value of h is found.
If at the first trial the errors are acceptable then a larger value such as h = 2hI may be tried.
Error Control for Mathematical Data. The trial-and-error method may also be applied
to mathematical functions f(x). In fact, this may be the only feasible method in cases
where 1" (x) and 1'" (x), or their maxima M2 and M3 , are difficult to calculate. However, if
good estimates for M2 and M3 are available, then the error estimates (12) and (25) can be
used to determine the optimum step sizes h that will yield a desired degree of accuracy:
This will be illustrated for the Bessel function of Problem 2.
The Bessel function Jo(x) was defined by the integral formula (2) above. The
derivatives of Jo(x) can be calculated by differentiation inside the integral sign. Thus,
one has

(27) J~(x) = - -1171: sin(x sin sin dt,

t) t
rr 0

J~ (x) = --1171: cos(x sin t) sin t dt,

2
rr 0
 1.2 POLYNOMIAL INTERPOLATION 11

]~' (x) = -1
7r
l0
rr
sin (x sin t) sin3 t dt.

The validity of this procedure is demonstrated in advanced calculus texts; see, for
example, [T-MJ. Notice that the integrands of these integrals are never larger than 1 in
magnitude. It follows that lJ~n)(x)1 ::: 1 for all x and for n = 0,1,2, .... Thus we may
take M2 = 1 and M3 = 1 in the error estimates for linear and quadratic interpolation.
In particular, (12) implies that If 1 (X) I ::: h2/8 ::: 10-4, provided that h ::: 0.0238. This
suggests, correctly, that linear interpolation in Table 2 for ]0 (x) does not maintain the
accuracy of the table. However, (25) implies that If2(x) I ::: h3/(9.J)) ::: 10- 4 provided
that h ::: 0.11592.
Higher-Order Polynomial Interpolation. Linear or quadratic interpolation is
adequate for most engineering computations. However, occasionally it is desirable to
have the greater accuracy provided by cubic or higher-order polynomial interpolation.
This will be the case if the nodal values Jk are difficult or expensive to obtain, or if
the computer memory space allocated to them is limited. In such cases, higher-order
interpolation can be used to reduce the number of nodal values needed to obtain an
accurate representation of J(x).
The general polynomial interpolation problem requires us to compute a polynomial
Pl(X) of degree.e that interpolates the nodal values at.e + 1 consecutive nodes Xo < Xl <
X2 < ... < Xl; i.e.,

(28)

It will be convenient to write Pi (X) in the form

(29) Pi(X) = ao + al (x - xo) + a2(x - Xo)(x - Xl) + ...

+ ae(X - Xo)(X - Xl) ... (X - Xi-I),

where ao, aI, ... , al are to be computed. Note that the number of these coefficients is
.e + 1, the same as the number of conditions (28).
It is important to note that there is exactly one polynomial Pi (X) that interpolates the
data at.e+ 1 distinct points. To see this, assume that qe(x) is a second polynomial of degree
.e that interpolates the data at Xo, Xl, ... , Xi such that qe (Xk) = Jk for k = 0, 1, ... , .e.
Then the difference ri(x) = Pi (X) - qi(X) is a polynomial of degree.e that is zero at the
.e + 1 points xo, Xl, ... , Xi. But a nonzero polynomial of degree.e can have at mostl real
zeros. It follows that re(x) = 0 for all x, so that Pe(X) = qi(X). This shows that there is
at most one interpolating polynomial. The existence of Pi (X) will be shown by deriving
an algOrithm for its construction.
12 NUMERICAL ANALYSIS

The assumed form (29) of Pe(x) and the interpolation conditions (28) imply that the
following equations hold.

(30) fo = ao,
fl = ao + al (Xl - Xo),
f2 = ao + al (X2 - Xo) + a2(X2 - XO)(X2 - Xl),

fe = ao + al (Xe - Xo) + a2(Xe - Xo)(Xe - Xl) + ...

+ ae(Xe - Xo)(Xe - Xl) ... (Xe - Xe-l).

This system of equations can be solved for ao, aI, ... , ae in succession because the
differences Xh - Xe are not equal to zero when k 'I- e. It is important to note that each
coefficient ah depends only on the nodes Xo, Xl, ... , Xh and nodal values fo, fl, ... , fh.
This property, which is evident from (30), implies that Ph (X) can be computed as soon
as ao, aI, ... , ak have been found. This property and equation (29) imply that

(31)

where Po(x) = fo = constant and

(32) k= 1,2, ... ,£.

These properties can be used to formulate an algorithm for computing pe(x). Note that
(30) can be rewritten as

(33) fo = ao,
fl = PO(XI) + alql(xl),
h= PI (X2) + a2q2 (X2),

The coefficients are uniquely determined by these equations as

(34) ao = fo,
aj = (h - po (Xj»)!ql (Xl),

a2 = (h - PI (x2»/q2 (X2),
 1.2 POLYNOMIAL INTERPOLATION 13

When ao, al, ... , ac have been computed then pc(x) is computed from (29). The
algorithm is easy to program. First ao is set equal to fo. Then PO(Xl) = fo, q(Xl) and al
are computed. With ao and al determined, Pl(X2), q2(X2), and a2 are computed, etc.
Uniformly Spaced Nodes. If Xk+l - Xk = h = constant for k = 0,1, ... ,.e - 1, then
another algorithm for pc (x) , based on the forward differences t:.fk, t:. 2fk, t:. 3fk, etc., is
often used. The forward differences are defined inductively by

(35) t:.fk = fk+l - fk'

t:. 2fk = t:.Jk+l - t:.A,
t:. 3fk = t:. 2fk+l - t:. 2A,

etc. The corresponding form of pc (x) is

1
(36) pc(X) = fo + rt:.fo + -r(r -
2
1)t:. fo + ...
2

+ .e!1 r(r - 1)··· (r -.e + 1)t:.f fo,

where

(37) r = (x - xo)Jh.

The correctness of (36) can be verified by noticing that r = 0 when x = Xo, r = 1 when
x = Xl, r = 2 when x = X2, ... , r = .e when x = Xf. It follows that the right -hand side
of equation (36) is a polynomial inx of degree.e that satisfies the interpolation conditions
(28). Hence it coincides with Pe(x) by the uniqueness of the interpolating polynomial.
The algorithm (36) is called the Gregory-Newton forward difference formula 3 It
may be given a compact form by introducing the binomial coefficients

(38)
\
G) = 1,
r) = r(r - l)(r - 2) ... (r - k + 1)
(k k! (k > 0).

Then (36) may be written

(39)

For the programming of (39), the computation may be broken into the following steps.

(40) Compute r = (x - xo)/h

Compute Po(x) = fa.

3The fonnula was first stated by James Gregory 0638-1675) in 1670. The first proof was given by Isaac Newton.
14 NUMERICAL ANALYSIS

Compute !lJo and PI (x) = Po(x) + r!lJo.

Compute !l2Jo and P2(X) = PI (X) + (;)!l2JO.
Compute !l3Jo and P3(X) = P2(X) + G)!l3JO'
etc.

The Error in Higher-Order Interpolation. The error is given by the error function

(41) Ee(X) = J(x) - Pe(X).

It is shown in [Ch-KJ that for each x such that Xo < x < Xe there is a point ~ such that
Xo < ~ < Xe and

(42)

Quantitative error estimates may be obtained from this equation. In particular, for
uniformly spaced nodes with Xh+1 - Xh = hand

(43)

we have (see [Ch-K])

(44)

This result shows the increased accuracy provided by higher-order interpolation. Note
that for .e = 2 the estimate (44) is not quite as good as the estimate (25).
Spline Interpolation. Linear interpolation produces functions that are continuous
and piecewise linear. Their derivatives are piecewise constant functions that in general
are discontinuous at the nodes. Similarly, quadratic interpolation produces continuous
functions whose derivatives are discontinuous at the alternate nodes where two quadratic
polynomials are matched. Corresponding remarks apply to higher-order polynomial
interpolation.
For some applications it is highly desirable to have an interpolating function whose
first derivative and (possibly) higher-order derivatives are continuous throughout their
domains of definition. The process of constructing such functions is called spline
interpolation. The name is derived from the early draftsmans spline, which is a flexible
elastic rod used to draw a smooth curve through a prescribed set of points.
The simplest spline functions are piecewise quadratic functions q(x) that interpolate
the nodal points (Xh,fh) and whose derivative q'(x) is required to be continuous. An
algorithm for computing such quadratic splines is given below, followed by an example
 1.2 POLYNOMIAL INTERPOLATION 15

of their application. The most widely used splines are the cubic splines, which interpolate
the data and have continuous first and second derivatives.
Quadratic Splines. The quadratic spline function q(x) for the nodes Xo < Xl < ... <

I
Xn and nodal values fo,fl, ... ,fn is a function of the form

qO(X) for XO:SX:SXI,

ql (x) for XI:S X :s X2,
(45) q(x) = .

qn~l (x) for Xn-l:S X :s Xn ,

where each function qo(X) , ql (x), ... , qn-l (x) is a quadratic polynomial. The inter-
polation conditions can be written

Finally, q' (x) is required to be continuous for Xo :s x :s X n . This is automatically true at

points other than the nodes because polynomials are always differentiable. The values of
the derivative at the nodes will be denoted by Zk, so

(47) q'(Xk)=Zk for k=O,l, ... ,n.

Note that the numbers Zk are as yet unknown and must be computed. The continuity of
q' (x) for Xo :s x :s Xn will hold if the derivatives q~ (x) match at the nodes. This means that

(48) q~(Xk) = Zk and q~(Xk+l) = Zk+l for k = 0,1, ... , n - 1.

Each quadratic polynomial qk(X) has three coefficients, as yet unknown, for a total of 3n
coefficients. The interpolation equations (46) provide 2n equations, while the matching
of q~ (Xk+ I) with q~+ I (Xk+ I) at the internal nodes Xl, X2, ... ,Xn-l gives an additional n - 1
equations, for a total of 3n - 1, i.e., one less than the number of unknown coefficients.
It will be shown that Zo can be prescribed arbitrarily and that the 3n coefficients of the
qk(X), as well as the derivative values Zl, Z2, ... , Zn, are then uniquely determined.
To derive an algorithm for q(x) it will be convenient to write qk(X) and q~(x) in the
form
1 2
(49) qk(X) = "2ak(X - Xk) + Zk(X - Xk) +fk'
q~(x) = ak(X - Xk) + Zk,
for k = 0,1, ... , n - 1, where the constants ao, aI, ... , an-l are to be computed. Note
that both the first of equations (46) and the first of equations (48) are satisfied by (49).
16 NUMERICAL ANALYSIS

The second equation of (48) gives

(50)

or

(51)

Finally, the second equation of (46) gives

1 2
(52) -ak(Xk+l - Xh) + Zh(Xk+l - Xh) + Ih = Ik+l.
2
Solving this for ak and equating the result to the quotient in (51) gives the equation

2(fh+l - Ik)/(xk+l - Xk) - 2Zk Zk+l - Zk

(53) ~= =.
(Xh+l - Xh) Xk+l - Xh

Finally, writing

(54) k = 0,1, ... , n - 1,

and solving (53) for Zk+l gives

(55) Zk+l = 2mk - Zk, for k = 0, 1, ... , n - 1.

If a value for Zo is selected then (55) and the data (Xk,Jk) can be used to compute Zl,
Z2, ... , Zn. Then (51) gives the values of ao, al, ... , an-l, and (49) can be used to evaluate
q(x). The algorithm may be outlined as follows. The input data are xo, Xl, ... , x n, 10,
11, ... ,In, and Zo, as well as the value of x.
(56) Compute mo, ml, ... , mn-l by means of (54).
Compute Zl, Z2, ... , Zn by means of (55).
Compute ao, a], ... , an-l by means of (51).
If X = xo, put q(x) = 10.
If Xo < X ::: X n , compute the integer k
such that Xk < x ::: Xk+l.
Compute q(x) = qk (x) by means of (49).
If desired, compute q'(x) = q~(x) by means of (49).

This procedure is illustrated in the following problem.

Problem3. A Nonlinear Resistor. The characteristics of a resistor R may be determined

by means of the circuit of Figure 5.
 1.2 POLYNOMIAL INTERPOLATION 17

ammeter
voltmeter

~
./?'

battery R

FIGURE 5. Circuit for testing a nonlinear

resistor R

It is assumed that the internal resistance of the ammeter and voltmeter have been
chosen to be close to zero and infinity, respectively, so that they do not affect the current in
the circuit. By adjusting the variable resistance Rv, the current i in the loop, as measured
by the ammeter, can be made to have any value from i = 0 to a maximum value i = imax
amperes. As i is varied, the voltage v across R, as measured by the voltmeter, also changes
in a definite reproducible way. Thus to each value of i on the interval 0 ::: i ::: i max there
corresponds a unique voltage v; i.e., v = F(i), where F is a function defined on the interval
o ::: i ::: i max . It will be assumed that F(i) has been recorded in analog form, for example
on magnetic tape, so that the experimental values of F(i) can be reproduced on demand.
It will be convenient to work with the function f(i) = F(i)/i rather than F(i). Note
that f(i) is the resistance of the resistor R, in ohms, at the current level i. In particular, if
f(i) = Ro = constant for 0 ::: i ::: i max then the resistor is a conventional linear one, and
vii = Ro is Ohm's law.
If f(i) is not constant then the resistor R is said to be nonlinear. In this case its
performance in any circuit is determined by the function f(i). An example is given in
Section 1.3. To make a computational analysis of such a circuit, the function f(i) must
be represented in the computer, as discussed at the beginning of this section. It will be
assumed that the values of f(i) are sampled at uniformly spaced values of i, say

(57) io = 0, i 1 = h, i2 = 2h, ... , in = nh = imax ,

1<0 = f(O), Rl = f(h), R2 = f(2h), .. . , Rn = f(nh).
A specific numerical example is provided by Table 3, where i max = 5 amps, h = 0.5,
amps and n = 10.
The table value Ro = 5.00 represents the measured values of f(i) for very small
positive values of i. Note that f(i) is nearly constant, and hence R, is nearly constant, for
o ::: i ::: 3 amps.
After the sampled values (57) have been stored in computer memory, intermediate
values of f(O may be computed by any of the interpolation methods discussed above.
18 NUMERICAL ANALYSIS

TABLE 3. Sample data for a nonlinear resistor

k i. v. R. k i. v. R.

0 0.0 0.00 (5.00) 5 2.5 12.43 4.97

1 0.5 2.51 5.02 6 3.0 14.88 4.96
2 1.0 4.98 4.98 7 3.5 18.97 5.42
3 1.5 7.50 5.00 8 4.0 24.12 6.03
4 2.0 10.10 5.05 9 4.5 29.93 6.65
5 2.5 12.43 4.97 10 5.0 37.40 7.48

Many methods of analysis require a representation offCO that has a continuous derivative.
Quadratic spline interpolation provides a suitable method. Application of the quadratic
spline algorithm to the data ofTable 3 gives Table 4. Note that the value Zo = 0 has been
selected. This is appropriate because f(O is nearly constant near i = 0 and hence has
slope near zero there.
A graph of the quadratic spline interpolation is shown in Figure 6. For comparison,
the graphs of the linear and quadratic polynomial interpolations are also shown.

TABLE 4. Computation of quadratic spline coefficients forf(i)

k i. R. mk Zk ak

0 0.0 5.00 0.04 0.00 0.16

0.5 5.02 -0.08 0.08 -0.64
2 1.0 4.98 0.04 -0.24 1.12
3 1.5 5.00 0.10 0.32 -0.88
4 2.0 5.05 -0.16 -0.12 -0.16
5 2.5 4.97 -0.02 -0.20 0.72
6 3.0 4.96 0.92 0.16 3.04
7 3.5 5.42 1.22 1.68 -1.84
8 4.0 6.03 1.24 0.76 1.92
9 4.5 6.65 1.66 1.72 -0.24
10 5.0 7.48 1.60
 1.2 POLYNOMIAL INTERPOLATION 19

Exercises 1.2

Linear Interpolation. Compute the linear interpolation segment equations and plot the interpolant.

1. 3.
x 0.5 0.9 1.2 x 0.5 1.0 1.5

y 0.469 0.677 0.755 Y 0.4 1.5 -0.75

2. 4.
x 0.5 0.8 1.2 x 0.5 1.0 1.5

y -1.0 0.5 0.75 Y -0.6 1.2 -0.7

Quadratic Interpolation. Compute the quadratic interpolant equation in the form ao + al (x - xo) +
a2 (x - xo)(x - Xl) and plot.

5. 7.
x 0.5 0.9 1.2 x 0.5 1.0 1.5

y 0.469 0.677 0.755 Y 0.4 1.5 -0.75

6. 8.
x 0.5 0.8 1.2 x 0.5 1.0 1.5

y -1.0 0.5 0.75 Y -0.6 1.2 -0.7

9. (Flow Meter) Compute the first three segment equations for the flow meter problem, using linear
interpolation.
10. (Flow Meter) Compute the first three parabolic equations for the flow meter problem, using quadratic
interpolation.
11. (Flow Meter) The first three parabolic equations for the flow meter problem, using natural quadratic
spline interpolation, Zo = 0, are

qo(x) = 149000 - 15000x - 15000x(x - 1),

ql (x) = 134000 + 7000(x - 1) + 37000(x - l)(x - 2),
q2(X) = 141000 + 23000(x - 2) - 21000(x - 2)(x - 3).
Verify this by shOwing that these equations interpolate the data items and that the piecewise quadratic
on 0 ::: x ::: 3 is continuously differentiable.
12. (Linear Interpolation) An engineer involved with embedded systems wants to replace the library
function sinh(x) by linear interpolation PI (x) on a table of values 0 ::: x ::: 0.5.
(a) Find a table of hyperbolic sine values with fewer than 49 rows that for ,,(x) == IPI (x) - sinh(x) I
satisfies ,,(x) < 10- 5 Hint: M2 = sinh(0.5).
(b) Report the best and worst accuracy ,,(x) at increments of 0.01 across the interval 0 ::: x ::: 0.5.

13. (Cubic Interpolation) Forfourequidistant nodesxk = xo+hh, h = 0, 1,2,3, verify that I(x-xo)(x-
Xl)(X - X2)(X - x3)1 ::: h4 for Xo ::: x::: X3.

14. (Cubic Interpolation) Use the previous exercise and equation (42) with I = 3 to derive the error
estimate
20 NUMERICAL ANALYSIS

for xo ::: x ::: X2.

15. (Linear Interpolation) Write a computer program to implement the linear interpolation algorithm
with data items XO, Xl, ... , Xn and Yo, Yl, ... , Yn . The program should accept input x in the domain
Xo ::: x ::: Xn and produce output Y = PI (x), the linear interpolant of the data points at x.

16. (Quadratic Interpolation) Write a computer program to implement the quadratic interpolation
algorithm with data items Xo, Xl, ... , X2m and Yo, YI, ... , Y2m. Testthe program on Table 1, producing
a data file of 49 pairs, suitable for use in plotting Figure 4.

17. (Cubic Interpolation) Repeat the previous exercise, using cubic interpolation instead of quadratic
interpolation, replacing 2m by 3m.

18. (Quadratic Spline Interpolation) Write a computer program to implement the quadratic spline
interpolation algorithm. Test the program on the nonlinear resistor data and produce a figure for f(O
onO::: i::: 5.

19. (Quadratic Spline Interpolation) Apply quadratic spline interpolation to the flow meter data of Table
1 in order to produce a plot very different from the ones in Figures 1 and 4.

20. (Bessel Function}o(x» Write a computer program that accepts a table of values oflo(x) and a number
x in 0 ::: x ::: 6.0. The output is a quadratic interpolation value oflo(x), correct to 4 significant figures
in the sense that IE2 (x) I < 10-+ Verify the accuracy theoretically using error estimates. Part of the
problem is to determine the table size.

21. (Arctangent Identities) The principal value of the Arctangent function is given by the integral
tan -1 (x) = (X ~. Prove these identities:
10 1 +u
(a) limHoo tan-I(t) =~. (b) tan-I (-x) = -tan-I(x). (c) tan-I (x) + tan-I (l/x) =~.

22. (Arctangent Table) A table of arctangent values is to be computed such that linear interpolation with
uniform spacing gives lEI (x) I < 10- 4 on the interval 0 ::: x ::: 1. Use estimate (12) and report the
smallest number of table entries that will suffice.

23. (Arctangent Evaluation) Use the results of the two preceding problems in order to develop a computer
program that evaluates tan -I (x), accurate to 4 significant digits, for 0 ::: x ::: 1, in the sense that
lEI (x) I < 10-+

24. (Bessel Function of Order One) The formula

1I (x) = -1 17f cos (x sin(t) - t)dt

7T 0

is used to produce an evenly spaced table with n + 1 entries on 0 ::: x ::: 1. Find the least n that produces
h (x) I < 10- 4 with quadratic interpolation.
25. (Cubic Interpolation) Develop the theory of cubic interpolation for uniformly spaced nodes. To be
givenarexo and h and the values Yo, Yl, ... , Y n , wheren = 3m. Inparticuiar,ifyo = Y3r,YI = Y3r+],
Y2 = Y3r+2, Y3 = Y3r+3, then derive these formulas:

ao = YO,
al = (YI - ao)/h,
a2 = (Y2 - ao - 2alh)/(2h2),
a3 = (Y3 - ao - 3alh - 6a2h)/(6h3).
 1.2 POLYNOMIAL INTERPOLATION 21

R
7.5

5.0 FIGURE 6. Quadratic spline interpolation of the data of

o 5 Table 3

26. (Runge's Example) I,et j(x) = 11(1 + x 2 ) on -5 < x < 5. For any integer n, there is a uniformly
spaced table of x and y values with n + 1 rows that is reproduced by f. Plot pn (x) for this table, n = 8
and n = 12. Runge showed that max IPn(x) - j(x) I --* 00 as n --* 00.

27. (Nonlinear Resistor) Plot the interpolation polynomial of degree 10 for the nonlinear resistor data. Is
the plot reasonable?

28. (Bessel Function JO(x)) Prove the following properties:

(a) Jo(O) = 1.
(b) lJo (x) I ~ 1.
=
(c) Jo( -x) Jo(x)

29. (Quadratic Spline) Derive the explicit formula

for the slopes Zk in the quadratic spline algorithm.

30. (Linear Interpolation) Compute by linear interpolation j(1.0) from the data items j(0.5) = 0.469,
f(0.9) = 0.677,fO.2) = 0.755.

31. (Quadratic Interpolation) Compute by quadratic interpolationj(1.0) from the data items f(0.7) =
0.566,f(0.9) = 0.612,f(1.5) = 0.857.

32. (Linear and Quadratic Interpolation) Compute v'55.25 and v'54.75 with a 10-digit calculator and
also by linear interpolation and quadratic interpolation, using data values .J53, ../54, y'55, .J56.
Compare your answers in a brief table.

R
7.5

5.0
o 5 FIGURE 7. Linear interpolation of the data of Table 3.
22 NUMERICAL ANALYSIS

R
7.5

5.0
o 5 FIGURE 8. Ouadratic interpolation of the data of Table 3.

33. (Linear Interpolation) Suppose that cos(0.15728) is to be evaluated by linear interpolation, using
adjacent table values of cos(0.15) and cos(0.16). How many digits of the linear interpolation answer
agree with the exact answer?

34. (Comparison of Linear, Quadratic and Quadratic Spline Interpolation) Apply the linear interpo-
lation, quadratic interpolation and quadratic spline interpolation programs to the data of Table 1 to
produce data files suitable for producing the graphs of Figures 6, 7 and 8.

35. (Linear Interpolation) Complete the derivation of the error bound (12) by verifying the inequality (9).
Hint: Introduce the new variable ~ = x - ~ (Xk + Xk+ 1)'

36. (Quadratic Interpolation) Complete the derivation of error bound (25) by verifying inequality (24).

1.3 Numerical Integration and Differentiation

Integration and differentiation are among the commonest mathematical operations that
are used in solving engineering problems. If the function to be integrated or differentiated
is mathematically defined then an exact formula or algorithm for computing its integral or
derivative may be available. However, in the case of integration, even among the simplest
elementary functions are those that have integrals that are not themselves elementary
functions. In such cases numerical integration must be used. A simple example is
provided by the error function, which is defined by

2 t
(1) erf(x) = J1i 10
2
e- t dt.

It occurs frequently in heat transfer problems and in probability theory. If the function to
be integrated or differentiated is known only numerically, either from experimental data
or as a result of computation, then only numerical methods are available to compute
its integral and derivative. Examples involving both kinds of functions are given below.
Numerical integration is treated first.
 1.3 NUMERICAL INTEGRATION ANO DIFFERENTIATION 23

Numerical Integration. The mathematical problem to be solved is to compute, to a

prescribed degree of accuracy, the numerical value of an integral

(2) I = lb f(x) dx.

The computation is to be based on a set of nodes Xo, Xl, ... , Xn such that a = Xo < Xl <
... < Xn = b and the corresponding nodal values fh = f(Xh), as in the interpolation
problem of Section 1.2. Only the case of uniformly spaced nodes will be treated, so that

(3) Xh = a + kh, k = 0,1, ... , n,

where h = (b - a)/n.

The Trapezoidal Rule. A simple approximation to the integral (2) is obtained by

replacing the integrand f(x) by the piecewise linear interpolation function PI (x) of Section
1.2. This gives

(4) lb f(x) dx ~ lb PI (x) dx,

where the error should be small when h is small. Recall that

Moreover,

(6) l Xh +1

Xk
dx = h, lXk
XH1
rdx = -h
2 '
1

whence

(7)

and

(8) lb f(x) dx ~ lb PI (x) dx

1 n-l
= "2h l)fh + fh+l)
h=O
1
= "2h(jo + 2fl + 2f2 + ... + 2fn-1 + fn).
Notice that the integral (7) is just the area of the trapezoid in Figure 2, Section 1.2, with
vertices (Xh, 0), (Xh+l, 0), (Xk+l,fk+l), (Xk,fh). For this reason the approximation (8) is
called the trapezoidal rule.
24 NUMERICAL ANALYSIS

Error Estimate. The error in (8) is given by

(9) El (h) = lb f(x) dx - lb PI (x) dx = lb El (x) dx,

where El (x) = f(x) - PI (x). Moreover, by (8), Section 1.2,

(10)

where MI :::: max If''(x)l. Integrating (10) gives the estimate

(11) 11 ~
xk 1
+ El(X)dx I :S-M2h3.
I
12
Finally, (9) and (11) imply that
1 3 1 2
(ll) IE1(h)1 < -nM2h = -(b - a)M 2h ,
-ll II
if M2 :::: max If''(x) I on a :s x :s b, because n = (b - a)lh. The error estimate (12)
gives us control of the error in the trapezoidal rule (8) whenever M2 can be computed.
This will be illustrated by a computation of the function value erf(O.!) defined by (1).
Equation (8) will be applied with a = 0, b = 0.1 and f(x) = (21 ",fii)C X2 , so that

(13) j'(x) = .)nC-2x)e-X2 ,

j"(x) = 5rr(4x 2 - 2)e-x2 •

It is not difficult to show that the maximum of If'' (x) I for 0 :s x :s 0.1 occurs when
x = 0, so that we may take

M2 = max If'') 4 = 2.2568.

(x I = ",fii

Hence the estimate (12) gives

(14)

Inparticular,IE1(h)l:s 5xlO- 5 H(1.8807xl0- 2 )h 2 :s 5xlO- 5 ,orh 2 :s 2.6586xlO- 3 ,

or h :s 0.05156. A suitable value of h is h = 0.05, which gives n = (b - a)lh =
0.11(0.05) = 2 and by (8),
1
(15) erf(O.I) ~ 2(0.05)([0 + 2fl + h)
= 0.05 {eo + 2e-CO.05)2 + e-cO.l)2}
",fii
 1.3 NUMERICAL INTEGRATION AND DIFFERENTIATION 25

= 0.11242

with error less than ±O.OOOOS.

Simpson's Rule. A more accurate approximation to the integral (2) may be obtained
by replacingf(x) by the piecewise quadratic interpolating function P2 (x) of Section 1.2.
This gives

(16)

Recall that in quadratic interpolation on the interval a ::: x ::: b one takes n = 2m =
(b - a)lh and

(17)

where r = (x - X2k)lh for X2k ::: x ::: X2k+2. It is easy to verify that
X2k x2k 2
l
+2
dx = lx2k+2 r dx = 3 l +r(r - 1) dx = 2h.
X2k X2h X2k

Hence, integrating (17) gives, after some algebraic simplification,

x2k+
2 1
(18) l P2(X) dx = '3 h(f2k + 4hk+l + hk+2)
X2k
and

(19) lb f(x) dx ~ lb P2(X) dx

1 m-l
= '3 h L(f2k + 4hk+l + hk+2).
k=O

This approximation is known as Simpson's rule. 4

Error Estimate. The error in (19) is given by

(20) E2(h) = lb -lb

f(x) dx P2(X) dx = lb f2(X) dx,

where f2(X) = f(x) - P2(X). Moreover, by (19), Section l.2,

(21)

4 Thomas Simpson (1710-1761), self-taught English mathematician. Authored several popular texts. HIS rule
appears in computations of Gregory in 1668 and Newton in 1676.
26 NUMERICAL ANALYSIS

for X2k~ X ~ X2k+2, where M3 2: max If'''(x) I· If this estimate is integrated over
X2k ~ x ~ X2k+2 and then summed over k = 0,1, ... , m -1, as in the derivation of (12),
the estimate

is obtained. However, it is a remarkable fact that a much better error estimate is available.
It is shown in [Ch-KJ that in fact,

(22) 11 X2

X2k
h+2 I
f2(X) dx ~ -M4 h5 ,
1
90
where M4 2: max If(4)(x)l. Combining (20) and (22) gives the estimate
1 5 1 4
(23) IE 2 (h)1 < -mM4h = - ( b - a)M4h ,
- 90 180
where M4 2: max If(4) (x) 1 for a ~ x ~ b, because m = (b - a)/(2h).
It is clear from (21) that both quadratic interpolation and Simpson's rule are exact
when f(x) is a quadratic polynomial. However, (22) implies that Simpson's rule is also
exact when f(x) is a cubic polynomial, since f(4) (x) = 0 in this case. This property is
easy to verify directly from (18).
Other Methods. Numerical integration algorithms are available that are based on
higher-order interpolation or unequally spaced nodes (see [Ch-KJ and [He]). The latter
include the very accurate Gauss quadrature formulas (see [Ch-K]). However, the simple
methods presented above are satisfactory for most engineering computations.
Error Control. The errors produced by the numerical integration of experimental data
must be assessed by the method of trial and error, just as in the interpolation problem.
The theoretical estimates (12) and (23) tell us how these errors should behave when
h is varied. For the integration of mathematical functions, the error estimates provide
rigorous control of the errors whenever realistic values for M2 or M4 can be found. This
is illustrated in the next example.

EXAMPLE 1. Computation of erf(x). The function erf(x) is defined for all x by the
equation (1). It is an odd function, erf( -x) = - erf(x) , and it is a monotone increasing
function of x with derivative

(24)

Moreover, it is known that

(25) lim erf(x) =

x-+oo
2r:;:;
v rr
1 0
00
e~t 2 dt = l.
 1.3 NUMERICAL INTEGRATION AND DIFFERENTIATION 27

It is shown below that

(26) I erf(x) - 11 :::: 5 x lO-5 for all x::: 3.

This means that erf(x) = 1.0000, with four-place accuracy, when x ::: 3. Hence, to this
accuracy only the values of erf(x) with 0 :::: x :::: 3 need to be computed. This will be
done by applying Simpson's rule to evaluate the integral in (1). The error estimate (23)
will be used to determine a suitable step size h.

The integrand in (1) and its first two derivatives are given by (13). Two additional
differentiations give

(27) rex) = In(-4X)(2X 2 - 3)e- X2 ,

f(4l(X) = 5n(4X4 - 12x2 + 3)e- x2

•

It is not difficult to show that the maximum of If(4l(x) I occurs at x = 0, so

24
(28) M4 = If(4) (0)1 = y'1i = 13.5406.

Hence the error in using Simpson's rule to evaluate erf(x) with 0 :::: x < 3 may be
estimated by (23) with a = 0, b = 3 and M4 given by (28), so that
3
(29) IE 2 (h)1 < -(l3.5406)h4 = (0.2257)h 4.
- 180
Thus four-place accuracy will be achieved if (0.2257)h 4 :::: 5 x lO-5, or h4 :::: 2.2153 x
lO-4. On taking the square root twice, it is found that the condition h :::: 0.1220 will
guarantee four-place accuracy. Table 5 was computed with h = 0.1. Note that

(30) erf(x2k+2) = erf(x2k) + 2r;;

",n
l
X2.
x2k 2
+ e-[2 dt

h
;::;;; erf(x2k) + 3(hk + 4hk+l +hk+2),
where fk = f(Xk) = J"e-(Xh)2. The table was computed by means of (30) with Xh = kh =
(O.l)k and k = 0,1, ... ,14.
It is easy to verify that neither linear nor quadratic interpolation in Table 5 will
maintain table accuracy. Hence, if values of erf(x) are needed with x =1= (0.2)k, it is best
to compute them directly using Simpson's rule with a suitable value of h :::: (0.1).

EXAMPLE 2. Water Stored in a Reservoir from Depth Data. The reservoir problem in
Problem 1 provides an opportunity to apply numerical integration to experimental data.
28 NUMERICAL ANALYSIS

TABLE 5. erf(x) for 0 :::: x :::: 3 step 0.2

x erf(x) x erf(x)

00 0.0000 1.6 0.9763

0.2 0.2227 1.8 0.9891
0.4 0.4284 2.0 0.9953
0.6 0.6039 2.2 0.9981
0.8 0.7421 2.4 0.9993
1.0 0.8427 2.6 0.9998
1.2 0.9103 2.8 0.9999
1.4 0.9523 3.0 1.0000

The volume in cubic feet of the water in the reservoir at time t is the function defined by
(1), Section 1.2; i.e.,

(31) v = Vo + let), l(t) = 1t q(r) dr.

This function is computed below for the data of Table 1, Section 1.2, and an assumed
value of Vo = 1.13 X 10 5 cubic foot (cD.

A more interesting problem arises if during the filling of the reservoir the depth x
of the water at the deepest point in the reservoir is recorded as a function of t. Assume
that an analog record of this function x = X(t) is available. Clearly, X(t) is a monotone
increasing function and has an inverse t = X-I (x). The function

(32) v = Vo + I(X- 1 (x)) == F(x)

then gives the volume of water stored in the reservoir as a function of the depth-gauge
reading x. This function is computed below for a set of sample data.
The trapezoidal rule and Simpson's rule will be applied to estimating the values of
the integral in (31) for the data ofTable 1, Section 1.2. Thus h = 1.0, tk = (1.0)k, and
Vk = Vo + l(tk) for k = 0, 1, ... ,24.
Application of the Trapezoidal Rule. Equation (31) implies that

(33)

Approximation of the integral by the trapezoidal rule with h = 1.0 gives

(34)
 1.3 NUMERICAL INTEGRATION AND DIFFERENTIATION 29

This relation, and the assumed initial value

(35) Vo = 1.13 X 105 cf,

determine VI, V2 , ... , V 24 . The results for the data of Table 1, Section 1.2, are shown in
Table 6.
Application of Simpson's Rule. Equation (31) also implies the relation

(36)

Approximation of the integral by Simpson's rule with h = 1.0 gives

1
(37) Vk+l = Vk-I + 3(qk-1 + 4qk + qk+l), k = 1,2, ... ,23.

If Va is given by (35) and VI is defined by the trapezoidal rule

1
(38) VI = Va + 2:(qO + ql),
then (37) determines V2, V3, ... , V24. The results, for the data of Table 1, Section 1. 2,
are shown in Table 6.
Notice that the results obtained by the two integration methods agree quite well in
view of the large step size h = 1.0. Of course, if more accurate results were needed then

TABLE 6. Water stored as a function of t

Trapezoidal Simpson Trapezoidal Simpson
t. V. x 10- 6 V. X 10- 6 tk Vk x 10-6 Vk X 10- 6

0.0 0.113 0.113 13.0 2.48 2.47

1.0 0.255 0.255 14.0 2.70 2.71
2.0 0.392 0.388 15.0 2.93 2.92
3.0 0.545 0.542 16.0 3.14 3.15
4.0 0.701 0.704 17.0 3.34 3.33
5.0 0.861 0.852 18.0 3.54 3.56
6.0 1.G4 1.04 19.0 3.73 3.72
7.0 1.22 1.21 20.0 3.88 3.89
8.0 1.41 141 21.0 403 402
9.0 1.64 1.63 22.0 4.18 4.19
10.0 1.84 1.85 23.0 4.31 4.30
11.0 2.05 203 24.0 4.45 4.46
12.0 2.26 2.27
30 NUMERICAL ANALYSIS

the original flow meter data could be sampled at a shorter interval such as h = 0.5 or
h = 0.25, and the computation of Vk could then be repeated.
Next assume that the depth record x = X(t) for the reservoir, when sampled at the
same one-hour intervals as in Table 1, Section 1.2, produces the data shown in Table
7. The corresponding quantities of water stored, obtained from Table 1, Section 1.2, by
Simpson's rule, are also shown.
Our goal is to find an algorithm for computing values of the function V = F(x) ,
°
giving volume versus depth, for arbitrary values of x on the interval ::s x ::s 133.3 feet.
The most straightforward approach is to recognize that in Table 7,

(39) Vk = F(Xk), for k = 0,1, ... ,24.

This suggests using quadratic interpolation in the table to obtain intermediate values.
The algorithm described in Section 1.2 (following equation (11), Section 1.2) will be
applied with n = 2m = 24, and Xk as in Table 7 for k = 0,1, ... ,24. The results are
shown in Table 8, where for brevity, only the values for x = 100, 105, ... , 130 ft. are
shown. For other x-values, the method gives the values of V = F(x) with equal ease.
Numerical Differentiation. The solution of many engineering problems requires
an estimate of the derivative of a function that is known only numerically or from
experimental data. A first estimate may be based on the definition of the derivative:

(40) l' (a) = lim j(a + h) - j(a) .

h~O h

TABLE 7. Sample depth gauge and volume data

th hrs Xh ft Vh x 10-6 cf th hrs Xh ft Vh x 10-6 cf

0.0 8.1 0.113 13.0 94.2 2.47

1.0 21.6 0.255 14.0 99.4 2.71
2.0 30.4 0.388 15.0 104.0 2.92
3.0 38.1 0.542 16.0 108.5 3.15
4.0 45.1 0.704 17.0 112.3 3.33
5.0 50.7 0.852 18.0 116.6 3.56
6.0 57.1 1.04 19.0 119.6 3.72
7.0 62.5 1.21 20.0 123.0 3.89
8.0 68.3 1.41 21.0 125.4 4.02
9.0 74.3 1.63 22.0 128.4 4.19
10.0 79.9 1.85 23.0 130.4 4.30
11.0 84.3 2.03 24.0 133.3 4.46
12.0 89.9 2.27
 1.3 NUMERICAL INTEGRATION AND DIFFERENTIATION 31

TABLE 8. Water stored as a function of depth

x (ft) V = F(x) (cD

100 2.74 x 106

105 2.97 x 106
110 3.22 X 10 6

115 3.47 x 106

120 3.74 x 106
125 4.00 x 106
130 4.28 X 10 6

This suggests the approximation

(41) l' (a) ~ of(a, h) = f(a + h~ - f(a) ,

where h is a small positive number. The number of(a, h) is called the forward difference
approximation to l' (a).

EXAMPLE 3. j(x) = X1l2, a = 1. In this simple case1'(x) = ix-1/2 and1'(l) = 0.5,

while of0 , h) = [(1 + h)1/2 -l]/h. Computation to 6-place accuracy gives the following
values.

8f(1, h) 0.488088 0.498756 0.499875 0.499990 0.500000

Loss of Accuracy. These numbers look satisfactory. However, let us try some smaller
values of h. The follOwing values were obtained with a calculator that maintains lO-place
accuracy

h x 10" 10 3 2 1.5 1. 499999999

8fO, h) 0500000 0.333333 0.'500000 0.666667 0.000000

The reason for these inaccurate results is not difficult to discover. For these values ofh,
fO +h) andf(l) agree to 10 digits and hence the calculator delivers fO +h) -j(l) = 0, to
this accuracy In fact, it is clear that for any continuous function, and any digital computer,
f(a + h) and f(a) will be equal, to machine accuracy; when h is sufficiently small. This
means that the approximation (41) will fail on any computer if h is taken too small.
An Error Estimate. The example shows that the approximation (41) must be used with
care. If h is chosen too large then the approximation is poor. If h is chosen too small then
32 NUMERICAL ANALYSIS

cancellation of significant digits causes loss of accuracy. Hence an accurate estimate of

the error

(42) fl (a, h) = l' (a) - of(a, h)

is needed in order to choose an optimal value of h.

An error estimate may be obtain\d from Taylor's theorem. The theorem implies that
if 1" is continuous then

(43) f(a + h) = f(a) + h1'(a) + lh21"(~),

where ~ is a number between a and a + h. Rearranging (43) gives
-1
1'(a) - of(a, h) = -h1"(~),
2
which implies the error estimate

(44)

where M2 is any constant that satisfies

(45) M2::::maxlf"(~)1 for a::::~::::a+h.

In Example 3 above,1"(x) = _~X-312. Thus if a = 1 and h > 0 then we may take

M2 = ~ and the error does not exceed hiS. In particular, the error does not exceed
5 x 10-7 if h = 4 X 10-6 In fact, [(1 + h) i-1]!h = 0.500000 forthis value of h. Notice
that if (44) with M2 = ~ is used to guarantee an error IflO, h)1 :::: 5 x 10- 10 then we
must take h :::: 4 x 10-9 . These values of h lead to loss of accuracy due to cancellation of
significant digits. We conclude that1"(l) cannot be computed to this accuracy by means
of (41) with a computer that maintains only 1O-place accuracy.
The Central Difference Approximation. LH6pital's rule for calculating limits implies
that

(46) f l( a) -- l'1m f(a + h) - f(a - h)

h .
h..... O 2
This suggests the alternative approximation

(47) l' (a) ~ Df(a, h) = f(a + h) ;I(a - h) ,

where h is a small positive number. The number Dj(a, h) is called the central difference
approximation to l' (a). We shall find that for a given value of h, the approximation (47)
 1.3 NUMERICAL INTEGRATION AND DIFFERENTIATION 33

is usually more accurate than (41). To see this, note that Taylor's theorem implies that if
1"'(x) is continuous then

(48)

where a < ;+ < a + h. Similarly,

(49)

where a - h < ;_ < a. Subtracting (49) from (48) and rearranging terms gives

(50) rea) = f(a + h) - f(a - h) _ ~h2 [f"'(;+) + f"'(;-)] .

2h 6 2
1
Now, [f'1/(;+) +1"'(;-) I lies between the maximum and minimum of1"'(;) for a- h ~
;~ a + h, and hence equals flll(;) for some; on the interval. Thus (50) can be written

(51) rea) = DfCa, h) - ~h2f"'W, I; - al ~ h.

If the error in the approximation (47) is defined by

(52) 1'2 (a, h) = rea) - Df(a, h),

then equation (51) implies that

(53)

where M3 ?: max [flll(;) I for I; - al ~ h.

In Example 3, the error estimate (44) for the forward difference approximation gave
11'1 (1, h)1 ~ 5 x 10- 7 if h ~ 4 X 10- 6 Estimate (53) guarantees the same accuracy if

~M3h2 ~ 5 X 10- 7 . Now, forf(x) = X1l2 we have 1"'(;) = ~;-5/2 and the maximum of
rw for I; - 11 ~ h occurs when; = 1 - h. If we restrict h by 0 < h ~ 10- 2 then
M3 = ~(l - 102 )-5/2 = 0.3845 is a suitable value. Thus

(54)

ifh ~ 10- 2 and ~(0.3845)h2 ~ 3 X 10- 7 , or h2 ~ 7.8023 X 10-6 , or h ~ 2.79 X 10-3 <
3 X 10- 3 . This is more than 750 times as large as the h provided by the estimate (44).
Richardson Extrapolation. If both Df(a, h) and Df(a, hJ2) have been computed then a
r
still more accurate estimate of (a) can be obtained. To see how this may be done, note
that by Taylor'S theorem,

(55) f(a+ h) = L fen)

00 ( )
_,_a hn,
n=O n.
34 NUMERICAL ANALYSIS

j(a - h) = L f(n)(
00
__a) (_h)."
"=0 n!
When these equations are subtracted, alternate terms cancel, and division by 2h gives

(56) ') a, h) = f(a + h) 2h

D.r( a + a2 h2 + ~ h4 + ... ,
- f(a - h) = f'()

where a2 = 1'"(a)/3!, a4 = j<5)(a)/5!, etc. Replacing h by hl2 gives

1 2 1 4
Df(a, hl2) = f (a) + 4a2h + 16 ~h + ....
I
(57)

It follows that

Df(a, h) - 4Df(a, hl2) = - 3f' (a) + ~~h 4 + ....

4
Hence, if we define
If 4 1
(58) D}a,hl2) = -Df(a,hl2) - -Df(a,h),
3 3
then

(59)
I I
f (a) = Df(a, hl2) + -~h
1 4
+ ....
4
This equation implies that an error estimate of the form

(60) [f'ea) - D'f(a, h)1 S Ch4

is valid, where C depends on f. The h4 term in (60) implies that the error in the
approximation l' Ca) ~ D'fCa, h) decreases very rapidly with h. The process of computing
D'fCa, hl2) from DfCa, h) and DfCa, hl2) is called Richardson extrapolation. The process
can be repeated one or more times to get still greater accuracy

EXAMPLE 4. CExample 3 continued). The accuracies of the forward and central difference
formulas, and of Richardson extrapolation are illustrated in Table 9 for the case of
fCx) =Xl/2 and a = 1.

Notice that for h = 2-13 the forward difference method gives only 4-place accuracy,
while for h = 2- 9 the central difference method gives 6-place accuracy Richardson
extrapolation gives 5-place accuracy with h = 2- 4 = 0.0625 and 6-place accuracy with
h = 2- 5 = 0.03125.

EXAMPLE 5. f(x) = Jo(x), a = l. The Bessel function]oCx) was defined by (2), Section
1.2, and was tabulated in Table 2, Section 1.2 to 4-place accuracy for 0.0 to 6.0 in steps
of 0.1. Application of the central difference formula and Richardson extrapolation to
 1.3 NUMERICAL INTEGRATION AND DIFFERENTIATION 35

TABLE 9. Comparison of three numerical differentiation methods

h 8fO, h) Df(l, h) D'fC1, h)

0.485281 0.500983
0.492422 0.500245 0.499998
0.496154 0.500061 0.500000
0.498062 0.500015
0.499027 0.500004
0.499513 0.500001
0.499756 0500000
0.499878
0.499939
0.499970
0.499982

the functionjo(x) and a = I gives Table 10. Details of the computation are left for the
Exercises.
As a check, note thatjb(x) = -j1(X) is tabulated to IS-place accuracy in [A-5].
These tables, when rounded to four places, givejbO) = -0.440l.

EXAMPLE 6. Surface Area of a Reservoir from Depth Data. The rate of evaporation of
water in a reservoir depends on the area of its exposed surface. Thus in order to compute
estimates of the effect of evaporation, it is essential to know the surface area A(x) as a
function of the water depthx. We shall compute A (x) from the volume function V = F(x)
by numerical differentiation. The influence of A(x) on reservoir dynamics is analyzed in
the next chapter.

The volume V = F(x) of water in the reservoir of Figure 9, when filled to depth x,

f:
is related to A(x) by the equation

(61) V = F(x) = Va + A(~) d4.

TABLE 10. J~(l) estimated by numerical differentiation

h DJoO, h) D'JoO, h)

0.40 -0.43l4
0.20 -0.4379 -0.4400
0.10 -0.4395 -0.4400
0.05 -0.4399 -0.4401
36 NUMERICAL ANALYSIS

FIGURE 9. Surface area ofa reservoir as afunction of depth.

From the data of Table 7 we may take Xo = 8.1 ft and Vo = 0.113 X 106 cf, then
A(x) can be estimated from the (Xk, Vk) data of Table 7 by interpolation and numerical
differentiation. However, it will be more convenient to use the depth record x = X(t) of
the table. We have

(62) v = Vo + Lx(t) A(~) dg = Vo + it qCr) dr,

where q(r) is the flow meter output tabulated in Table 1, Section 1.2. Differentiation of
(62) with respect to t gives the relation

(63) A(X(t)) X'(t) = q(t).

Thus if t = tk = k hours and Xk = X(tk),X;' = X'(tk), qk = q(tk), and Ak = A(Xk), then
one has AkX;' = qk, or
qk
(64) Ak= "x
k

The values of qk are given in Table 1, Section 1.2. The values of x;' may be obtained
from Table 7 by numerical differentiation. The results are shown in Table 11 with the
corresponding values of Ak.
The values of x~ and x~4 were obtained by the forward difference formula and the
remaining values by the central difference method.
Table 11 gives Ak = A(Xk) at the unequally spaced depths Xk of the table. Values of
A(x) for other x-values may be obtained by quadratic interpolation of the tabular points
(xk,Ak).
The availability of A(x) offers an opportunity to check the consistency of our
computations by recomputing V = F(x) through numerical evaluation of the integral in
(61). The results of such a check are shown in Table 12. The computation was restricted
to 35 S x S 135 to avoid the comparatively inaccurate value Ao = A(8.1) produced
by the forward difference computation of x~. This term affects the interpolated values
 1.3 NUMERICAL INTEGRATION AND DIFFERENTIATION 37

TABLE 11. Flow rate, depth, depth rate of change, and surface area vs. t
tk qk x 10- 5 Xk X'
k
Ak x 1O-4(sq ft)

0.0 1.49 8.1 13.50 1.10

1.0 1.34 21.6 IUS 1.20
2.0 1.41 30.4 8.25 1.71
3.0 1.64 38.1 7.35 2.23
4.0 1.49 45.1 6.30 2.37
5.0 1.71 50.7 6.00 2.85
6.0 180 57.1 5.90 3.05
7.0 1.79 62.5 5.60 3.20
8.0 2.1S 6S.3 5.90 3.69
9.0 2.26 74.3 5.80 3.90
10.0 190 79.9 5.00 3.80
11.0 2.15 84.3 5.00 4.30
12.0 2.16 89.9 4.95 4.36
130 2.15 94.2 4.75 4.53
14.0 2.32 99.4 4.90 4.73
15.0 2.22 104.0 4.55 4.88
16.0 198 10S.5 4.15 4.77
17.0 2.08 112.3 4.05 5.14
18.0 198 116.6 3.65 542
19.0 1.63 119.6 3.20 5.09
20.0 1.52 123.0 2.90 5.24
210 152 125.4 270 5.63
220 1.34 128.4 2.50 5.36
23.0 1.31 130.4 2.45 5.35
24.0 1.42 133.3 2.90 4.90

of A(x) for Xo = 8.1 S x S 30.4 = X2. The table interval for x is taken to be 5 ft, so
that Xk = 35.0 + (5.0)k, k = 0,1,2, ... ,20. Column 3 gives the values of Ak = A(Xk)
obtained from Table 11 by quadratic interpolation. Column 4 gives the approximate
values of Vk = F(Xh) obtained from (61) by applying Simpson's rule

Column 5 gives the corresponding approximate values of Vk as obtained from Table

7 by quadratic interpolation. In Table 12 the entry Vo in column 4 is taken from column
5, while Vj in column 4 is obtained from Vo and (61) by the trapezoidal rule; i.e.,
38 NUMERICAL ANALYSIS

TABLE 12. Consistency of reservoir volume data

Xk Ak x 10- 4 Vk X 10-6 Vi X 10- 6
(from A k ) (from qk)

0 35.0 2.07 0.477 0.477

40.0 2.30 0.586 0.584
2 45.0 2.37 0.704 0.702
3 50.0 2.81 0.829 0.833
4 55.0 3.03 0.982 0.976
5 60.0 3.09 1.13 1.13
6 65.0 3.37 1.29 1.29
7 70.0 3.78 1.47 1.48
8 75.0 3.90 1.67 1.67
9 80.0 3.82 1.86 1.85
10 85.0 4.34 2.06 2.07
11 90.0 4.36 2.28 2.27
12 95.0 4.56 2.50 2.51
13 100.0 4.76 2.74 2.74
14 105.0 4.88 2.97 2.97
15 110.0 4.93 3.22 3.22
16 115.0 5.33 3.47 3.47
17 120.0 5.08 3.75 3.74
18 125.0 5.60 3.99 4.00
19 130.0 5.37 4.29 4.28
20 135.0 4.90 4.44 4.46

VI = Vo + ~(Ao + AI). Note that the agreement of columns 4 and 5 is quite good in
view of the large step size for x of h = 5 ft.

Exercises 1.3

Trapezoidal Rule. Compute f~ j(x)dx using the trapezoidal rule with 10 divisions and report the error
compared to the exact value.

1. j(x) = x 5 3. j(x) = logO + x).

2. j(x) = sinrrx. 4. j(x) = (x - 1)2(x + 1)3

Simpson's Rule. Compute f~ j(x)dx using Simpson's rule with 10 divisions and report the error compared
to the exact value.
 1.3 NUMERICAL INTEGRATION AND DIFFERENTIATION 39

5. j(x) = x 5 7. j(x) = 10g(2 + x)

6. j(x) = sin 7lxJ2. s. j(x) = (x - l)3(x + 1)2

Jo
9. (Trapezoidal rule) Verify numerically l (sin(x)/x) dx = 0.9460830704 using the trapezoidal rule with
enough subdivisions n to insure 4-place accuracy. Determine from error estimates a minimum number n
that will suffice.
10. (Simpson's Rule) Evaluate numerically the Fresnel Integral Jt
sin(x 2)dx for X = 20. The integral is
known to approach ../lJ8 as X -> 00. Use Simpson's rule with at least 200 subdivisions.
11. (Trapezoidal Rule) Write a computer program to implement the trapezoidal rule. The program accepts
as input the endpoints a and b and the number n :::: 1 of subdivisions. Internally, the program generates
the node values Xk = a + kh, step size h = (b - a)ln, and the function values j(Xk). The output of
J:
the program is the approximate value of j(x)dx. Test the program onj(x) = sin(x), a = 0, b = n,
n = 100. The function values j(x) may be generated by a function subroutine.
12. (Trapezoidal Rule) Write a computer program for the trapezoidal rule that accepts the interval
endpoints a, b, the number n :::: 1 of subdivisions and a set of data points (Xi,Yi), i = 1, ... ,N instead of
a function f. Accomplish this by coding the linear interpolant of the data points in a function subroutine
j that hides all details of interpolation. Test J~ j(x)dx = 27.16 for n = 100 using the nonlinear resistor
data
(0.0,500), (0.5,502), (1.0,4.98), (1.5,5.00),
(2.0,5.05), (2.5,4.97), (3.0,4.96), (3.5,5.42),
(4.0,6.03), (4.5,6.65), (5.0,7.48).

13. (Trapezoidal Rule) Apply the trapezoidal rule to compute a table for the error function erf(x) on
x = 0.1 to x = 3 in steps of 0.1. How many subdivisions are needed in the trapezoidal rule to insure
error IEl (h)1 < 10- 5 ,
14. (Simpson's Rule) Write a computer program to implement Simpson's rule. The program accepts as
inputs the values of a, band m = nl2. It generates internally the values h = (b - a)ln and nodes
Xk = a+kh. The function values j(x) may be generated by a function subroutine. Test it on J(;12 sin(x)dx
with 20 divisions.
15. (Error Function) Write a computer program based upon Simpson's rule that will accept any floating
point number x in 0:::: x:::: 1.0 and output erf(x) , accurate to four places.
16. (Bessel Function Jo(x)) Use the integral representation

]0 (x) = 2. [" cos(x sin t) dt

n io
together with Simpson's rule to develop a computer program that evaluates]o (x), with Simpson error
IE2(h)1 < 10- 4 , in 0:::: x:::: 15.0.

17. (Reservoir Problem: Depth Versus Volume) Use Table 7 and quadratic interpolation to develop a
computer program that evaluates V as a function of x for 8.1 :::: x :::: 133.3. Apply the program to verify
that V = 2.8 x 10 6 when x = 102 (approximately).
IS. (Reservoir Problem: Consistency of Surface Area and Flow Data) Compare numerically the two
integrals in equation (62) by using Simpson's rule on the integral and quadratic interpolation on the
integrand. Produce a table of values for the two integrals for t = 0 to t = 24 in steps of 2.0 and compare
the relative errors.
40 NUMERICAL ANALYSIS

19. (Error Function) Letf(x) = Jrre- x2 Compute for 0::: x::: 1 the derivative!"(x) and show that
max If''(x) I = If''(O) I = Jrr = 2.2568.

Apply the trapezoidal error estimate to verify that erf(0.2) = 0.22270 with absolute error less than
0.00001.

20. (Error Function) Find the first four derivatives of f(x) = Jrr exp( _x2). Verify that
max If(4) (x) I = If(4\0)I = 2!. = 13.5406.
0::;x::;3 v'rr
21. (Reservoir Surface Area) Verify Table 11 using a computer program that includes internal tables for q
and x. The program should compute x' and A, then print the columns of the table.

22. (Numerical Differentiation) Let f(x) = y'x. Compute the forward, central difference and Richardson
extrapolation approximations at x = 1 for step size h = 2- 9

23. (Numerical Differentiation ofJo(x)) Design a calculator program that reproduces Table 13 below. Use
the values

10(0.60) = 0.9120048635, 10(0.80) = 0.8462873528,

Jo(0.90) = 0.8075237981, Jo(0.95) = 0.7867870686,
Jo(1.05) = 0.7427955564, Jo0.10) = 0.7196220185,
JO(1.2O) = 0.6711327443, Jo0.40) = 0.5668551204.

The answers can be checked using the identity 1b(x) = -11 (x) and Bessel function tables, such as in
[A-Sl, which show JbO) = -0.4401 to four places.

24. (Central Difference Approximation tof") Use Taylor's theorem with fourth-order remainder to show
that iff(4) (x) is continuous, then

f" (a) = f(a + h) - 2f(a) + f(a - h) _ ~h2f(4)(e)

h2 12

for some e, Ie - al ::: h.

25. (Richardson Extrapolation for f") Define

Ji( h) = f(a + h) - 2j(a) + f(a - h)

'V a, - h2

TABLE 13. J~(l) Estimated by numerical differentiation

h Central Richardson
Difference Extrapolant

040 -04314
0.20 -04379 -0.4400
0.10 -04395 -04400
0.05 -04399 -04401
 1.4 SOLUTION OF EQUATIONS 41

Then Taylor's theorem implies that VfCa,h) = fl/Ca) + b2h2 + b4 h4 + .... Show by Richardson's
extrapolation method that

, 4 1
V fCa, h) == - VfCa, h) - - VfCa, 2h)
3 3

satisfies lfl/Ca) - V'fCa, h)1 :s Ch 4 for some constant C.

26. CNumerical Estimation of J~) Compute the approximations VJoO, h) and V'JoO, h) to J~O) for
h = 0.40,0.20,0.10 and 0.05. Check the answers by using theformula V'Jo(l, h) +D'JoO) + JoO) ""
J~Cl) + JbO) + Jo(l) == 0

1.4 Solution of Equations

In engineering analysis it is frequently necessary to solve an equation

(1) j(x) = 0,

where j(x) is a known function that may represent either experimental data or a
mathematical function. Examples of both types are given below. The problem may also
arise in different forms. For example, we may wish to solve the equation g(x) = c where
the constant c and the function g(x) are known. This problem takes the form (1) if we
define j(x) = g(x) - c. Similarly, we may wish to solve the equation g(x) = hex) where
the functions g(x) and hex) are known. This problem also takes the form 0) if we define
f(x) = g(x) - hex).
The solution set of an equation (1) is the set of all numbers x that satisfy the equation.
A second equation g(x) = 0 is said to be equivalent to 0) if it has the same solution
set. For example, the equations {f(X)}2 = 0 and exp {f(x)} - 1 = 0 are both equivalent
to 0). This freedom to change from 0) to an eqUivalent equation is frequently useful in
solving equations.
The solution set of (1) is the set of points where the graph of y = j(x) crosses the x-
axis. Hence, the number of solutions of (1) may be determined by graphing the equation
y = j(x) and determining how many times it crosses the axis. Even a rough graph may
suffice for this purpose.
The ideas mentioned above are illustrated and clarified in this section by means of
two examples. Many additional examples occur in the remainder of the text.

EXAMPLE 1. A Heat Transfer Problem. In the Introduction the cooling of a stirred flask
of water subject to Newtons law of cooling was analyzed. When the heat due to the
stirring was included it was found that the flask temperature was given by equation (18)
42 NUMERICAL ANALYSIS

of the Introduction, i.e., T = u(t), where

u(t) = 100 + (0.~1 - 80) (1 _ e- ht ) ,

and h is a positive solution of the equation

40h - 0.01 -lOh
(2) ----=e
80h - 0.01
Thus to complete the solution of the problem it is necessary to solve this equation. If the
new variable x = lOh is introduced, the equation may be written as
4x - 0.01 x
(3) - - - - - e - =0
8x-0.01 '
which has the form (1). This equation can be written in many equivalent forms. Here
the form
a
(4) C=2+--, a=2.5x10- 3 ,
x-a
will be used. Note that x = 0 is a solution of (4). However, only positive solutions are
relevant to the heat transfer problem. Moreover, r > 1 and 2 + al(x - a) < 1 for
o < x < a. Hence, any positive solution of (4) must satisfy x> a. To see that (4) has
exactly one solution x > a, note that (4) characterizes x as the ordinate of the point
of intersection of the curves defined by y = rand y = 2 + a/ex - a). Figure 10 is a
computer-generated plot of the two curves and indicates that they cross when x :::::: 0.7.
This solution of (4) is unique. A proof is outlined in the Exercises below.
The Method of Bisection. This is a simple method of solvingf(x) = 0 that is based
on the intermediate value theorem of the calculus [T-MJ. The theorem implies that if f(x)
is continuous and iff(a)f(b) < 0, so thatf(a) andf(b) have opposite signs, then there
is at least one solution of f(x) = 0 such that a < x < b. When such an estimate for x has
been found, it can be improved by computingf(c), where c = ~(a + b) is the midpoint
of the interval (a, b). If f(c) = 0 then a solution has been found. If f(a) f(c) < 0 then

a
y=2+--

--
x-a

FIGURE 10. Graphical solution of

a 0.25 0.5 0.75 equation (4)
 1.4 SOLUTION OF EQUATIONS 43

there is a solution in the interval a < x < c, while ifjCc)jCb) < 0 there is a solution
in the interval C < x < b. This process of bisection can be repeated until the desired
accuracy is obtained.
The method of bisection, as outlined above, may be formulated as an algorithm
that is easy to program for a computer. To begin, an interval Cao, bo) is found such that
jCao) jCbo) < O. Then a sequence of intervals Cal, bl ), (a2, b2), ... is computed as follows:

At step n(= 0,1,2, ... ) compute Cn = ~(an + bn) andjCc n).

If jCc n) = 0, then the solution is x = Cn.
IfjCan)j(c n) < 0, set an+l = an, bn+l = Cn
IfjCan)jCc n) > 0, set an+l = Cn, bn+ l = bn.
Continue until bn - an < E where E is the desired degree of accuracy.
The solution is x = Cn to accuracy E.

Application to Example 1. The method of bisection will be applied to the example by

defining
a
(5) j(x)=c-2---, a=2.5xlO- 3 ,
x-a
so thatj(x) = 0 is equivalent to (4). Note thatjCO.5) = -0.3563 andjC1.0) = 0.7158.
Hence ao = 0.5 and bo = 1.0 are suitable starting values. To estimate the number of steps
required for 5-place accuracy, we note that the error En after step n satisfies En ::::: bn - an
and EO ::::: bo - ao = 112. Moreover, bn+l - an+l = ~(bn - an) for n = 0, 1,2, . .. , whence
bo - aO
(6) En -< bn - an = - -2n
- for n = 0, 1,2, ....

Thus for 5-place accuracy it will suffice to terminate when n satisfies Cbo - ao)l2 n =
_ 2 X 10- 5 ,or
2 - n - l <!

log 105 5
n > --- = = 16.6.
- log 2 0.30103
Table 14 shows the result of the computation, terminating with n = 17. The solution,
correct to 5 decimal places, is x = 0.69495.
Advantages and Disadvantages of the Method. The primary disadvantage of the
method of bisection is the slowness with which it converges. In fact, the error estimate
(6) implies that more than three steps are needed to gain one additional decimal place
of accuracy. Thus in Example 1 fifteen steps were required for 5-place accuracy. An
advantage of the method is its generality. It works for any continuous function whose
values can be computed. Also, each step is simple, requiring only one function evaluation
at x = Cn, since Jean) is available from the preceding step.
44 NUMERICAL ANALYSIS

TABLE 14. The method of bisection applied to Example 1

n an bn Cn JCan) JCcn)

0 0.500000 1.000000 0.750000 -0.356304 0.113656

0.500000 0.750000 0.625000 -0.356304 -0.135770
2 0.625000 0.750000 0.687500 -0.135770 -0.014912
3 0.687500 0.750000 0.718750 -0.014912 0.048376
4 0.687500 0.718750 0.703125 -0.014912 0.016487
5 0.687500 0.703125 0.695312 -0.014912 0.000727
6 0.687500 0.695312 0.691406 -0.014912 -0.007108
7 0.691406 0.695312 0.693359 -0.007108 -0.003194
8 0.693359 0.695312 0.694336 -0003194 -0001235
9 0.694337 0.695312 0.694824 -0.001235 -0.000254
10 0.694824 0.695312 0.695068 -0.000254 0.000236
11 0.694824 0.695068 0.694946 -0.000254 -0.000009
12 0.694946 0.695068 0.695007 -0.000009 0.000114
13 0.694946 0.695007 0.694977 -0.000009 0.000052
14 0.694946 0.694977 0.694962 -0.000009 0.000022
15 0.694946 0.694962 0.694954 -0.000009 0.000006
16 0.694946 0.694954 0.694950 -0000009 -0.000001
17 0.694950 0.694954 0.694952 -0.000001 0.000003

The Method of Fixed Points. A fixed point of a function y = g(x) is a value of x

that is carried into itself by the function, so that

(7) x = g(x).
Graphically, the fixed points of y = g(x) are the ordinates of the points in which the
graph of y = g(x) intersects the straight line y = x. Thus in Figure 11 the points Xl, X2,
X3 are fixed points of g(x).

y
y=x

y = g(x)

~----~------~----~--?>x

FIGURE 11. Fixed points of y = g(x)

 1.4 SOLUTION OF EQUATIONS 45

Frequently, a fixed point C of g(x) can be computed by the method of successive

approximations. The first step in the method is the choice of a point Xo as a first
approximation to C. The point Xl = g(xo) is then computed. Under certain conditions,
described below, Xl is a better approximation to Cthan Xo. The operation is then repeated
to getx2 = g(Xl),X3 = g(X2), etc., so that

(8) Xn+l = g(xn ), n = 0,1,2, ....

This process is called iteration of the function g(x). If the sequence {xo, Xl, X2, ... }
converges to a limit C and if g(x) is continuous, then

C = lim Xn+l = n--+oo

lim g(xn ) = g(£),
n~oo

so that C is a fixed point of g(x).

EXAMPLE 2. (Example 1 continued). The method will be illustrated by application to

the heat transfer problem of the Introduction. This problem led to equation (4), which
is equivalent to the fixed point equation (7) with

(9) g(x) = log (2 + _a_) , a= 2.5

x-a
x> x lO-3.

Table 15 shows the results obtained by iteration of the function (9) with various choices
of the initial point Xo.
The table reveals the surprising fact that for any Xo > a the fourth iterate X4 =
0.69495 is accurate to five decimal places. Examination of Figure 12, the graph of
the function g(x), reveals the reason for this behavior. If Xo > 0.69495 then log 2 =
0.6932 < Xl < 0.69495, while if a < Xo < 0.69495 then Xl > 0.69495 and the
remaining steps are as in the first case.
Simple conditions that guarantee the convergence of the method of fixed points can
be given. Note that to be sure that all the iterates (8) are defined, we must choose Xo from
an interval I of the real axis that is carried into itself by g, so that g(x) is in I whenever X
is in I. (The interval I defined by a < X < 00 is suitable in the case of Figure 12.) If this

TABLE 15. Solution of Example 1 by the fixed point method

Xo 1.00000 100.000 1 X 109 0.00250000001

Xl 0.69440 0.69316 0.69315 19.33697

X, 0.69495 0.69496 0.69496 0.69321
X3 0.69495 0.69495 0.69495 0.69496
X4 0.69495 0.69495 0.69495 0.69495
46 NUMERICAL ANALYSIS

0.69495
log2
~--------------------~x FIGURE 12. Graph of g(x) =
0.69495 log(2 + x~,,). ct = 2.5 X 10- 3

is the case then the conditions

(10) Ig(x)-g(Y)1 :sLlx-yl for all x andy in I, and

(ll) o :s L < 1,

are sufficient to guarantee that I contains a unique fixed point £ = g(£) and that for every
choice of Xo in I the sequence {XO,XI,X2, ... } defined by (8) converges to £. Condition
(10) is called a Lipschitz condition, and L is called the Lipschitz constant 5
To see that the Lipschitz condition with 0 :s L < 1 ensures that £ is unique, suppose
that £1 and £2 are two fixed points in I, so that £1 = geed and £2 = g(£2). Then by (10)
applied to x = £1 and y = £2 we have

This is possible with 0 :s L < 1 only if 1£1 - £21 = 0; i.e., £1 = £2, and the two points
are the same. The existence of a fixed point, which also follows from (10) and (11), will
not be proved here.
To verify the convergence of xo, Xl, X2, ... to £, for any choice of Xo in I, note that by
(8) and (10), we have

(12) IXn - £1 = Ig(xn-l) - g(£)1 :s Llxn-l - £1

:s L2 1xn_2 - £1 :s ... :s Lnlxo - £1·

The right-hand side of this inequality tends to zero when n -+ 00 if 0 :s L < 1, and
hence limxn = £.
It is not always easy to verify the Lipschitz condition (10). A simpler condition is
available if g(x) has a derivative g'(x) such that

(13) max Ig'(x) I = L < 1,

I

'Named after Rudolph Lipschitz 0832-1903), German mathematician, who applied the condition to the theory of
differential equations; see Chapter 2.
 1.4 SOLUTION OF EQUATIONS 47

i.e., the maximum of Ig'(x)I is less than 1. In this case the mean value theorem,

(14) g(x) - g(y) = g/(~)(X - y), ~ between x andy,

together with (13) gives (10). Notice that conditions (10) and (11) are sufficient to
ensure convergence but are not necessary The function of Figure 12 does not satisfy
these conditions on the interval a < x < 00, but the iterates xn+l = g(xn ) converge to
1= 0.69495 for every choice of Xo > a.
EXAMPLE 3. Currents in a Nonlinear Circuit. As a second example of the method of
fixed points we shall analyze the currents in the nonlinear circuit of Figure 5, Section 1.2.
The voltage across the nonlinear resistor R will be assumed to be given by VI = F(i) =
if(i), where f(i) is defined by the data of Table 3, Section 1.2, extended by quadratic
spline interpolation as in Table 4, Section 1.2. The second resistor is assumed to operate
in the linear range and to provide an adjustable resistance of Rv ohms, so that the voltage
across it is V2 = Rvi. The battery delivers a constant voltage E. As before, the ammeter and
voltmeter are assumed to have internal resistances near zero and infinity, respectively, and
thus do not affect the current i. Kirchhoff's voltage law then implies that E = VI + V2, or

(15) E = Rvi + if(i).

This is a nonlinear equation that for given values of E and Rv is to be solved for i. To explore
the solvability of (15) it is convenient to draw a current-voltage plane and to regard the
solution i as the ordinate of the point of intersection of the straight line V = E - Rvi and
the voltage graph of the resistance, given by V = Fei) = if(i). For definiteness the value
E = 40 volts will be used; see Figure 13.
It is evident from the figure that there is a unique positive value of Rv for each current i
that satisfies 0 < i :::: 5. Moreover, i increases as Rv decreases. The smallest value of Rv that

v
40
. (5,37.4)
v = E - iRv
30

20

10

o
2 3 4 5
FIGURE 13. Graphical solution of
-10 E = 40, Rv = 10 E = iRv + if(i)
48 NUMERICAL ANALYSIS

gives a current i ~ 5 satisfies 40 = 5Rv + 37.4 (see Table 3, Section 1.2), so Rv = (40-
37.4)15 = 0.52 ohms. Thus our problem is to solve (15) for i when E = 40 volts and Rv
has a prescribed value such that 0.52 ~ Rv < 00. Note that a carefully drawn diagram will
give a first numerical approximation to the current i corresponding to a given value of Rv.
Equation (15) with E = 40 is

40 = Rvi + if(O = i (Rv +f(O).

An equivalent equation is
40
(16) i= -----:-c-:-
Rv + f(i)'
which has the fixed point form with g(O = 40/(Rv + f(O). We shall solve (16) for several
values of Rv.
Case 1. Rv = 10. Equation (16) with Rv = 10 is equivalent to

(17) C(O == 40 - Wi - if(O = O.

Now, Table 3, Section 1.2, gives

C(2.5) = 40 - 10(2.5) - (2.5)(4.97) = 2.58,

C(3.o) = 40 - 10(3.0) - (3.0)(4.96) = -4.88.

Hence the solution of (17) satisfies 2.5 < i < 3.0 by the intermediate value theorem.
On this interval, by Table 4, Section 1.2, and (49), Section 1.2, we have
1
f(O = qs(O = "2as(i - 2.5)2 + Zs(i - 2.5) + fs,
which can be written

(18) f(O = (0.36)l + (-0.20)y + 4.97, y = i - 2.5.

The fixed point method applied to

40
(19) g(O = 10 + f(O ' 2.5 ~ i ~ 3.0,

with f(O defined by (18), and three different starting values io, gives Table 16.
Clearly, the iteration converges rapidly in this case.
Case 2. Rv = 5. In this case, (16) with Rv = 5 is equivalent to
(20) C(O == 40 - 5i - if(i) = O.
 1.4 SOLUTION OF EOUATIONS 49

Table 3, Section 1.2, gives

G(3.5) = 40 - 5(3.5) - (3.5)(5.42) = 3.53,

G(4.0) = 40 - 5(4.0) - (4.0)(6.03) = -4.12.

Hence the solution satisfies 3.5 < i < 4.0. On this interval, by Table 4, Section 1.2, we
have

(21) f(i) = q7(i) = (-0.92)l + (1.68)y + 5.42, y =i - 3.5.

The fixed point method applied to

40
(22) g(i) = 5 + f(i) , 3.5:S i :s 4.0,

with f(i) defined by (21) gives Table 17.

In this case the iteration also converges, but this time quite slowly. The reason for
this behavior will be determined below.
Case 3. Rv = 1. In this case, (16) with Rv = 1 is equivalent to
(23) G(i) == 40 - i - if(i) = O.

Table 3, Section 1.2, gives

G(4.5) = 40 - 4.5 - (4.5)(6.65) = 5.58,

G(5.0) = 40 - 5.0 - (5.0)(7.48) = -2.40.

Hence the solution satisfies 4.5 < i < 5.0. On this interval, by Table 4, Section 1.2, we
have

(24) f(i) = qg(i) = (-o.12)l + (l.72)y + 6.65, y =i - 4.5.

Note that the function
40
(25) g(i) = 1 + f(n ' 4.5:s i :s 5.0,

TABLE 16. The method of fixed points applied to (19)

2.5000 2.7500 3.0000
2.6720 2.6769 2.6738
2.6763 2.6763 2.6763
50 NUMERICAL ANALYSIS

TABLE 17. The method of fixed points applied to (22)

io 3.5000 3.7500 4.0000
it 3.8388 3.7097 3.6265
i2 3.6753 3.7272 3.7673
i3 3.7431 3.7195 3.7026
i4 3.7126 3.7229 3.7304
is 3.7259 3.7214 3.7181
i6 3.7200 3.7220 3.7235
i7 3.7226 3.7217 3.7211
is 3.7215 3.7219 3.7222
i9 3.7220 3.7218 3.7217
ilO 3.7218 3.7219
ill 3.7218 3.7218

with 1(0 defined by (24), does not satisfy (13). In fact,

g'(.)
1 =
-401'(0
,
(l +1(0)2
and by Table 4, Section 1.2, we have

'() 401'(4.5) 40(1.72)

(26) Ig 4.5 1= = = 1.1756.
(1 +1(4.5))2 (1 + 6.65)2
Moreover, g(4.5) = 40/(1 + 6.65) = 5.23, and g(i) does not carry the interval 4.5 :s
i :s 5.0 into itself. Hence, the iteration of the function defined by (25) can be expected
to converge slowly or not at all. In fact, if we take io = 4.5, the iteration is found to
diverge. The solution of (23) is shown below to be i = 4.85304. However, even if we
choose io = 4.850, close to the correct value, we find that 46 iterations are required
to produce the approximation i 46 = 4.853, correct to only 3 decimal places. Thus the
method of fixed points fails in this case.
Error Estimates. The basic error estimate for the method of fixed points, given by (12),
may be used to understand the convergence results found in the preceding examples.
When Rv = 10, the function g(i) defined by (19) has derivative

, . -401'(0
g (1) = (10 + 10))2' 2.5 :s i :s 3.0.
The maximum value of If'(Ol for 2.5 :s i :s 3.0 is IZsl = 0.2 (see Table 4, Section 1.2)
and the minimum value of1(0 is easily found to be 1(0.28) = 4.94. Hence (13), Section
 1.4 SOLUTION OF EQUATIONS 51

1.3, gives the estimate

40(0.02) -2
(27) L :::: (10 + 4.94)2 = 3.58 x 10 .

Applying the error estimate (12) with n = 2 to the first column of Table 16, with
f = 2.6763, and using (27) gives

Hence h approximates f to at least 5 decimal places. yvhen Rv = 5, the function g(i) is

given by (22) and has derivative

I . -401'(0
g (I) = (5 + j(0)2' 3.5:::: i :::: 4.0.

In this case
40(1.68)
L = max Ig (01 = Ig (3.5)1 = (5 + 5.42 )2 = 0.619.
I I
(28)

Applying the error estimate (12) with n = 9 to the second column of Table 17, with
f = 3.7218, and using (28) gives the estimate

This guarantees convergence to at least 3 decimal places after 9 iterations. The estimate
is slightly pessimistic because i g is actually correct to 4 places.
Advantages and Disadvantages of the Fixed-Point Method. The principal advantage
of the method is that it is easy to program and requires only one function evaluation
per step. Moreover, if an interval I containing the solution can be found for which
the Lipschitz constant L is small compared with 1, then the method converges rapidly
Equation (13) shows that a key parameter is the value of Ig' (f) I. If it is small compared
with 1, then an interval containing f can be found in which L is small. The main
disadvantages of the method are that it may converge slowly and that it may be difficult
to put a given equationj(x) = 0 into the fixed point form x = g(x) in such a way that
g'(f) is small.
Newton's Method. One of the oldest iterative methods for solving j(x) = 0 was
invented by Newton. The geometric idea behind his method is shown in Figure 14.
Beginning with a first approximation Xo, obtained graphically or by another method,
the tangent line to y = j(x) at the point (xo,f(xo» is computed. The point Xl where the
tangent line crosses the x-axis is taken as a second approximation. The process is then
repeated until a sufficiently good approximation is found.
52 NUMERICAL ANALYSIS

y
tangent lines
I I
I (xo,f(xo))

'---t-----:==-""-"'----'<-----~--;> x FIGURE 14. Newton's method for

Xo solvingj(x) =0

The tangent line to y = j(x) at (xo,J(xo)) has the equationy = j(xo) +1'(xo)(x-xo).
Hence Xl is defined by the equationj(xo)+1'(XO)(XI-XO) = 0, orXI = xo-(f(xo)/1'(xo)).
Repetition leads to the sequence XO,XI,X2, ... defined by

(29) n = 0,1,2, ....

This is just the iterative method for finding a fixed point of g(x) = x - (f(x)I1' (x)). It
is clear that if lim Xn = 1'- exists and l' (£) =j:. 0 then 1'- satisfies j(£) = O. Of course the
procedure will fail if one of the numbers 1'(xo) , 1'(XI), 1'(X2) , ... is zero. Hence, the
method is usually applied within an interval in which1'(x) =j:. 0.

EXAMPLE 4. (Solved by Newton's Method). The method of fixed points was found to
fail when applied to the nonlinear circuit equation (16) with Rv = 1. The equivalent
equation (23) will now be solved by Newton's method. Note that

(30) G(i) = 40 - i - ij(i),

G'(i) = - 1 - j(i) - i1'(i) ,

wherej(i) and1'(i) are defined for 4.5 ::; i ::; 5.0 by

(31) = (-0.12)/ + 1.72y + 6.65,

j(i)
1'(i) = (0.24)y + 1.72,
y = i - 4.5.

Newton's method applied to G(i) gives the sequence io, iI, i2, ... defined by

. . GCin)
In+! = In - G'Cin) , n = 0,1,2, ....

Applying this with three different starting values io gives Table 18.
 1.4 SOLUTION OF EQUATIONS 53

TABLE 18. Newton's method applied to Example 2

io 4.5000 4.7500 5.0000
i, 48622 4.8538 4.8544
h 4.8530 4.8530 4.8530
i3 4.8530 4.8530 4.8530
i4

Error Estimates. Newton's method for solvingJ(x) = 0 was seen to coincide with the
method of fixed points for the function

J(X)
(32) g(x) = x - 1'(x)'

Moreover,
, J(x)j" (x)
(33) g (x) = [1'(x)]2'

and J(I) = O. Thus ifl' (I) t 0 then g' (I) = 0 and there will be intervals I containing I in
whichL = maXI Ig'(x) I < 1. Hence, the error estimate (12) will be applicable withL < 1.
An improved error estimate that explains the rapid convergence of Newton's method
can be obtained by applying Taylor's theorem with second derivative remainder to
g(x). If xo, Xl, X2, ... are defined by Newton's method (29), i.e., x n+! = g(xn ) and if
lim n --+ oo Xn = £ = ge£) then Taylor's theorem permits us to write

(34) 1 " (~))(xn - £) 2 ,

x n+! - £ = g(xn) - g(£) = 2:g

where ~ is between Xn and £. In particular,

1 2
(35) IXn+! - £1 ~ 2:M2lxn - £1 for n = 0,1,2, ... ,

where M2 = maxI Ig"(x) I and I is any interval that contains £ and Xo, Xl. X2, .... Note
that if M2 ~ 2 then

(36) IXn+l - £1 ~ IXn - £1 2 for n = 0, 1,2, ....

This means that each iteration doubles the accuracy of the preceding one; i.e., if IXn - £1 ~
lO-h then IXn+! - £1 ~ 1O-2h This behavior may be seen in Table 18.
Advantages and Disadvantages of Newton's Method. The principal advantages of
Newton's method are the ease with which it is programmed, which it shares with all
iterative methods, and its rapid convergence, typified by (35). A disadvantage is the
54 NUMERICAL ANALYSIS

necessity of computing l' (x) at each iteration. If l' (x) is difficult or costly to compute
then it may be preferable to use a slower method.

Graphical Solution. Find the given solution by graphical methods. Use a domain small enough to ensure
that the solution given is accurate to one decimal place.

1. 3x = X2 + 1, x = 2.62. 3. x 3 = 1 - x, x = 0.68.

2. 2 sin x = x, x = 1.895. 4. e- x = tanx, x = 0.53.

Bisection Method. Find the solution to 4 digits by the bisection method. Select the initial interval [a, b]
by graphical methods and record the number of iterations required for 4-digit accuracy

5. x 2 - 3x + 1 = 0, x = 0.3819660113. 7. x 3 + x + 1 = 0, x = -0.6823278038.

6. 4sinx - x = 0, x = 2.474576787. 8. e- 2x = tanx,x = 3.143453158.

Fixed-Point Method. Find the solution to 4 digits by the method of fixed points, given the initial
approximation Xo. Choose the interval [a, b] graphically. Record the number ofiterations required for 4-digit
accuracy.

9. x = (x 2 + 1)/4, Xo = 0.26 11. x = (1- x 3 )/5, Xo = 0.2

10. x = cosx, xo = 0.74 12. x = -0.210g(tanx), Xo = 0.26

Newton's Method. Find the given solution to 4 digits by Newton's method, given the initial approximation
xo. Choose the interval [a, b] graphically Record the number of iterations required for 4-digit accuracy

13. x 3 + 5x - 1 = 0, xo = 0.2 15. x 2 + x-I = 0, Xo = 0.62

14. x+sinx -1 = O,xo = 0.51 16. e- 3x - tanx = 0, xo = 0.34

Interpolation Methods. Let g(x) be the quadratic spline interpolant made from a data table of 11 elements,
x = 0 to 1 in steps of 0.1, using the functiony = 1 - sinx and gl(O) = -1.
17. Solve the equation x = g(x) graphically to obtain x = 0.51.
18. Find the quadratic polynomial q(x) = ao + al (x - xo) + a2 (x - xo)(x - Xl) that satisfies q(x) = g(x)
near x = 0.51.
19. Solve x = g(x) near x = 0.51 by the bisection method for the 4-place answer x = 0.51097.
20. Solve near x = 0.51 the quadratic equation x = g(x) by Newton's method for the 5-place answer
x = 0.510966.
21. (Heat Transfer Problem) Show that the fixed point equation

x = log (2 + _ a )
x-a
a
is equivalent to ~ = 2 + - - on the interval x> a.
x-a
 1.5 INVERSE FUNCTIONS 55

22. (Increasing Functions) A function j(x) is defined to be increasing on an interval a :::: x :::: b if j(XI) <
j(X2) for a :::: Xl < x2 :::: b. Show that iff' (x) > a on a :::: x :::: b, then j(x) is increasing on a :::: x :::: b.
23. (Decreasing Functions) A function j(x) is defined to be decreasing on an interval a :::: x :::: b if
j(XI) > j(X2) for a:::: Xl < X2 :::: b. Show thatj(x) is decreasing on a:::: x :::: b if and only if -j(x) is
increasing on a:::: x :::: b.

24. (Monotone Functions) A function j(x) is said to be monotone on an interval I if it is either increasing
or decreasing on I. Show that if j(x) is monotone on I, then the equation j(x) = c can have at most one
solution x in I.

25. (Flask Cooling) Show thatj(x) = ex - 2 - -a- is increasing for x> a = 0.0025. Explain why this
x-a
shows the uniqueness of the solution x = 0.69495 to the equation j(x) = O.

26. (Graphical Solution) Write a computer program that generates a data file suitable for producing the
graphs of y = eX and y = 2 + a/(x - a) in the region a < x :::: 1,0 :::: y :::: 7, for parameter value
a = 0.0025.
27. (Method of Bisection) Write a computer program to implement the method of bisection and apply it
to duplicate the results of Table 14.

28. (Method of Fixed Points) Write a computer program to implement the method of fixed points. The
computation should continue for N steps or until IX n - xn-ll :::: IXn IE. Program inputs include the
function g(x), the endpoints a and b of the interval where a fixed point is sought, the initial guess Xo
for a root, the value of E, and the maximum number N of steps.

29. (Method of Fixed Points) Make a table [or the first four iterates xi (i = 1,2,3,4), accurate to five
decimal places, in the method of fixed points for the function g(x) = log(2 + a/(x - a)), a = 0.0025,
using the interval [0,10 10 ] and initial guesses Xo = 1, xo = 100, Xo = 109 , Xo = a + 10- 11 . This
duplicates Table 15.

30. (Newton's Method) Write a computer program to implement Newton's method. The program should
fail if anyone of the follOWing conditions holds:

a. The derivative vanishes at an iteration step.

b. An iterate is outside the interval of definition off.
c. The iterates do not converge, i.e., Ixn+ I - Xn I fails to be small in the given number N of iterations.

It should succeed if accuracy E is achieved. The inputs are the functionj(x), the derivative j'(x) ,
the interval la, b] in which j is defined, the initial approximation xc, the accuracy E and the maximum
number N of steps.

31. (Existence of Fixed Points) This exercise assumes a strong advanced calculus background. Let a function
g(x) map an interval [a, b] into itself and satisfy a Lipschitz condition with constant l. < 1. Show that
for any starting point xo in [a, b], the sequence of points xc, Xl = g(xo), X2 = g(Xj), ... is a Cauchy
sequence and hence converges in [a, b] to a fixed point of g.

1.5 Inverse Functions

In Problem 3 of Section 1.2, the voltage v across the terminals of the nonlinear resistor
R is a function v = F(i) of the current i flowing through it. It is observed experimentally
that F(i) is an increasing function of i. Hence we may also regard i as a function of v.
56 NUMERICAL ANALYSIS

This function is the inverse of F, usually denoted by i = F- I (v). This last relation is of
direct physical interest because it tells us the current i that will flow through R when a
voltage v is placed across the terminals.
The need to construct a computational algorithm for the inverse of a known function
Y = j(x) arises frequently in engineering analysis. The function j(x) may represent
either experimental or mathematical data. Three examples are presented in this section.
Additional examples occur in subsequent sections and chapters.
A continuous function y = j(x) , defined on an interval 1: a :s x :s b, will have a
single-valued inverse x = j-I (Y) if and only if j(x) is either an increasing function of x
or a decreasing function of x. Such functions are called monotone functions (see the
exercises in Section 1.4). The function j will map I onto the interval]: j(a) :s y :s j(b)
if j is increasing and onto the interval]: j(b) :s y :s j(a) if j is decreasing. A sufficient
condition forj(x) to be monotone increasing (resp., decreasing) is thatJ'(x) > 0 (resp.,
rex) < 0) for every x in 1.
Use of Inverse Interpolation. If a table of values (xo,Yo), (XI,YI), ... , (xn,Yn), with
Yk = j(Xk) , is known for a monotone function Y = j(x) then a simple method to construct
x = j-I(y) is to interchangexk = j-l(Yk) andYk = j(Xk) in the table and then interpolate
the data (YO,xo), (YI,XI), ... , (Yn,xn). This procedure is called inverse interpolation.

EXAMPLE 1. A Nonlinear Resistor. Inverse interpolation will be illustrated by applica-

tion to the nonlinear resistor of Problem 3, Section 1.2. The numerical example given
there is defined by Table 3, Section 1.2. A computational algOrithm for the function
i = F-I(v) may be constructed by interpolating the table data with the ik and Vk columns

TABLE 19. Computation of quadratic spline coefficients for F- 1 (v)

k Vk ik mk Zk ak

0 0.00 0.0 0.199203 0.200000 -0.000635

2.51 0.5 0.202429 0.198406 0.003257
2 4.98 1.0 0.198413 0.206452 -0.006380
3 7.50 1.5 0.192308 0.190373 0.001488
4 lD.lD 2.0 0.214592 0.194242 0.017468
5 12.43 2.5 0.204082 0.234943 -0.025193
6 14.88 3.0 0.122249 0.173221 -0.024925
7 18.97 3.5 0.097087 0.071278 O.OlD023
8 24.12 4.0 0.086059 0.122897 -0.012681
9 29.93 4.5 0.066934 0.049220 0.004743
10 37.40 5.0 0.000000 0.084648 0.000000
 1.5 INVERSE FUNCTIONS 57

interchanged. Of course, any of the interpolation methods of Section 1.2 may be applied.
Here the quadratic spline algorithm will be used. Application of the algorithm to the
pairs (Vk, ik), k = 0,1, ... ,10, gives Table 19. A graph of i = F- 1 (v) computed by this
algorithm is given in Figure 15. Note that the value Zo = 0.20 has been selected. This
was motivated by the observation that F(i) = if(i) , where f(i) is the nonlinear resistance
function of Table 3, Section 1.2. It follows that F/(i) = f(i) + i1'(i), and in particular,
F/(O) = f(O) + 01'(0) = 5.00. Hence Zo = (F- 1 ),(0) = [F /(0)]-1 = 0.20.
The Method of Solution of Equations. If a value of y is given then the computation of
the value x = f- 1 (y) is eqUivalent to solving for x the equation y = j(x) or the equivalent
equation f(x) - y = O. This last equation, with y fixed, may be solved for x by one of
the methods of Section 1.4. Two examples will be given.

EXAMPLE 2. Computation of cos- 1 y. The number x = cos- 1 y will be regarded as the

solution of the equation

(1) f(x) == y - cosx = 0, y fixed.

Of course, cos- 1 y is a multiple-valued function. However, cosx is a monotone decreasing

function on the interval I : 0 S x S n and maps it onto the interval]: -1 S Y s 1.
Hence equation (1) has exactly one solution x in I for each choice of y in]. This value
x = cos- 1 Y is called the principal value of cos- 1 y. A graph of the principal value of
x = cos- 1 Y is shown in Figure 16.
Equation 0) may be solved for x by Newton's method. Sincej'(x) = sinx, Newton's
formula (29), Section 1.4, becomes
y - cosXn
(2) Xn+1 = Xn - n = 0,1,2, ....
sinxn
As an example, (2) was used to compute C05- 1 0 = nl2 to five-place accuracy using
three different values for Xo. The results are shown in Table 20.

5.0

2.5

0.0 " ' - - - - - - - - - - - - - . . . . . . ; ; . v FIGURE 15. Graph of the quadratic spline

a IS.7 37.4
interpolation of i = F- 1(v)
58 NUMERICAL ANALYSIS

:rr12

o +------~---~y

-1 o 1 FIGURE 16. Graph of x = cos- 1 y (principal value)

EXAMPLE 3. Kepler's Equation. The equation

(3) E - e sin E = M

is called Kepler's equation. 6 Its solutions determine the positions of the planets in their
orbits around the sun, as well as the positions of other satellites such as the planetary
moons, comets, and artificial Earth satellites. Solving Kepler's equation for E as a function
of M, e being fixed, is equivalent to computing the inverse of the function fCE) = E -
e sin E. Algorithms for doing this are given below for values of e that are of astronomical
interest. First, the role of Keplers equation in orbital calculations will be explained.

Kepler's Laws of Planetary Motion. In 1609, after years of calculation based on

the excellent observational data of the astronomer Tycho Brahe, Kepler published the
following two laws governing the motion of the planets.

The orbit of each planet is an ellipse with the sun at one focus.
II The line joining the sun to a planet sweeps out equal areas in equal times.

We shall show how these two principles make it possible to compute the position of
a planet, or other satellite, by solving Kepler's equation.

TABLE 20. Computation of 1r/2 = cos- 1 0 by Newton's method

Xo 0.50000 1.00000 1.50000 2.00000
Xl 2.33049 1.64209 1.57091 l.S4234
X2 1.38062 1.57068 1.57080 1.57080
X3 1.57312 1.57080
X4 1.57080
Xs

6 After Johannes Kepler (1571-1630), German mathematician and astronomer who discovered the laws of planetary
motion. Kepler's first two laws were published in 1609. His third law was published in 1618.
 1.5 INVERSE FUNCTIONS 59

a
b
c
I----~--------~--------~--~A
o

p FIGURE 17. An ellipse and its parameters

Mathematical Description of Ellipses. A typical ellipse is shown in Figure 17.

Its size and shape may be characterized by any two of the parameters a, b, c, e, p, where

(4) a = OA = BF, b = OB,

c = OF = Ja 2 - b2 ,
e = cia, p = b2/a = a(l - e2 ).
The ellipse may be defined by the property that for every point P on it,

(5) PF + PF' = 2a = constant.

Equations for the ellipse may be expressed in terms of the coordinates of Figure 18.
The equation

(6)

is well known to follow from (5). Moreover, from the figure and (6) we have

FR=rcosv, x=OR=acosE,
y = PR = r sin v = b sin E,

y
... Q(x,y')

o R F

FIGURE 18. Coordinates used to describe an ellipse

60 NUMERICAL ANALYSIS

so that

(7) x = r cos v + c = a cos E, y = r sin v = b sin E.

On eliminating E from (7) we get

(8) r = P
1 + fCOS v'
while eliminating v gives

(9) r = a(l- fCOS E).

Finally, eliminating r between (8) and (9) gives the following relations between v and E:
cos E - f cos v + f
(10) cos V = , cosE= - - - -
1- fCOS E 1+ fCOS V

Mathematical Formulation of Kepler's Second law. In Figure 19, the area A of the
sector FPoP of the ellipse is given by the equation

(11)

where r is the function of v defined by (8).

Now let us assume that P = PCt) is the position of a planet moving in as elliptic orbit,
so that v = vCt) is an as yet unknown function of time t. Then A = ACt) and Kepler's
second law is interpreted to mean that ACt) is a linear function of t, so that

(12) ACt) = -1
2
I
Vo
V
(t)
r2 dv
1
= -hCt
2
- to),

where h and to are constants and Vo = vCto). In particular, if T is the period of the planet,
i.e., the time needed to traverse the ellipse once, then ACT + to) = nab = !hT, so that

2nab
(13) T = -h- .

y
Po

P /v
········1.

F Fl GURE 19. Area of a sector of an eII ipse

 1.5 INVERSE FUNCTIONS 61

It can be shown that h is just the angular momentum of the planet in its motion around
the sun.
Derivation of Kepler's Equation. The function vet) that describes the planetary motion
is defined implicitly by (12). However, to evaluate the integral it is more convenient
to change the integration variable to the angle E of Figure 18, defined by (10). On
differentiating (10) and using (8) and (9), it is found after a short calculation that with r
defined by (9),
dv b
(14)
dE r
Thus the change of variable gives

(15) 1 V (t)
r2 dv = b
lE(t)
rdE

l
Vo Eo
E(t)
= ab (l - e cos E) dE
Eo

= ab(E - e sin E) I~~t)

= ab (E(t) - e sin E(t) - Eo + e sin Eo) ,
where E(t) is related to vet) by (10) and Eo = E(to). Now let us choose to = 0 and
Eo = Vo = O. This means that t is measured from the perihelion, or orbital point nearest
the sun, where E = v = O. Then combining Kepler's law (12) and (15) gives
ht 2m
(16) E(t) - esinE(t) = - = - .
ab T
This is just Kepler's equation (3) with
2m
(17) M=-.
T
Thus if E = j-l (M) is the solution of Kepler's equation then the position of the planet
at time t is given by

The notation v, E, M used in this example is due to Gauss. 7 Astronomers call v the
true anomaly, E the eccentric anomaly and M the mean anomaly of the planet.
Solvability of Kepler's Equation. For satellites moving in elliptic orbits, the eccentricity
e satisfies 0 :s e < 1. The left-hand side of Kepler's equation (3) is the function

7Karl Friedrich Gauss (1777-1855). famous German mathematician. His book Theory of Motion of the Heavenly
Bodies (1809) gave the first complete account of methods for determining the orbits of celestial bodies.
62 NUMERICAL ANALYSIS

TABLE 2l. Eccentricities of the planetary orbits

Planet Mercury Venus Earth Mars Jupiter Saturn Uranus Neptune Pluto
0.2056 0.0068 0.0167 0.0934 0.0484 0.0557 0.0472 0.0086 0.2486

feE) == E - e sin E with derivative l' (E) = 1 - e cos E. Clearly, l' (E) ~ 1 - e > 0 and
hence feE) is a monotone increasing function for 0 ::: E ::: 2n. It maps this interval onto
itself, sincef(O) = 0 andf(2n) = 2n.
Solution by the Method of Fixed Points. Kepler's equation can be rewritten as

(18) E = M +esin E,

which has the fixed-point form E = geE) with geE) = M +esin E. Here g'eE) = ecos E
and Ig'(E)1 ::: e. Hence, the method of fixed-points can be expected to work well if e is
small compared with 1. Now, the known planets have the eccentricities shown in Table
21. Note that they lie in the range 0 ::: e ::: 0.25. It follows that the method of fixed points
will work well for them. In fact, it is the method that was used by Gauss. The method
was used to compute Table 22, which is applicable to the planet Mercury (e = 0.2056).
Solution by Newton's Method. Most of the known comets move in elliptic orbits
that have eccentricities greater than 0.9. For example, Halley's comet has eccentricity
e = 0.967. For such orbits the convergence of the method of fixed points for Kepler's
equation is very slow. For example, ifthe method is applied with e = 0.967 and M = 3.0,
then 50 iterations, beginning with Eo = 3.0000, produces Eso = 3.0583, while the
correct value to four decimal places is E = 3.0696. On the other hand, Newtons method
works well for this value of e. To apply the method, define

(19) F(E) == E - e sin E - M,

so that Keplers equation is equivalent to F(E) = O. Then F'(E) 1 - ecos E and

Newton's method gives the iteration En+1 = En - F(En)/F'(En) , or

En - e sin En - M
(20) En+1 = En - 1- ecos En
, n = 0,1,2, ....

TABLE 22. Solution of Kepler's equation for Mercury (e = 0.2056)

M 0 rr/4 rr/2 3rr/4 rr 5rr/4 3rr/2 7rr/4 2rr
E 0.0000 0.9530 1.7722 2.4822 3.1416 3.8010 4.5109 5.3302 6.2832
 1.5 INVERSE FUNCTIONS 63

x
6.615

4.410

2.205
x = 0.967 sin(x) +y y
0.000 FIGURE 20. Graph of the solution of
0.0 2.1 4.2 6.3 Kepler's equation for Halley's comet (e = 0.967)

TABLE 23. Solution of Kepler's equation for Halley's comet

y a n/4 rr/2 3rr/4 5rr/4 3rr/2 7rr/4 2rr
x 0.0000 1.7388 2.2951 2.7369 3.1416 3.5463 3.9881 4.5444 6.2832

Table 23 shows the solution of Kepler's equation for the case of Halley's comet (e =
0.967). It was computed by means of (20) with Eo = M. The maximum number of
iterations required was 5. Figure 20 shows a graph of the solution.

Inverse Table. A table of x and y values has an inverse table, formed by swapping the x and y columns
and sorting the first column to nondecreasing order. Produce a table from y = f(x) with II rows, x = 0 to
1 in steps of 0.1. Then determine the inverse table.

l.y=1+x 2 4. Y = sin(nx).
2. Y =x3 5. Y =..;x.
1
3. y= - - . 6. Y = cos(.nxl2).
l+x

7. (Inverse Functions) Produce a table fromy = x 2 - x with 11 rows, x = 0 to 2 in steps of 0.2. Does the
inverse table represent a functional relationship x = g(y)?

8. (Computation of sin- 1 (y» Compute an evenly spaced inverse table for the function y = sinx, using
Newton's method, y = -1 to y = 1 in steps of 0.25. Check the inverse table with the inverse function
equation x = sin -1 (y).

9. (Nonlinear Resistor) Given the nonlinear resistor data (Vk, ik) (h = 0, ... , lO)in the firstthree columns
of Table 19, apply quadratic spline interpolation with Zo = 0.2 to verify the columns for mh, Zh, ah.
10. (Computation of cos- 1 (y» Apply Newton's method toy = cos x at initial guessesxo = 0.5, Xo = 1.0,
xo = 1.5, Xo = 2.0 in order to produce a table of approximations to n12. In the process, verify Table 20.
11. (Kepler's Equation for Mercury) Apply the method of fixed points to Kepler's equation for Mercury,
x = y + 0.2056 sinx, for y = 0 to y = 2n in increments of n/4, in order to produce an inverse table
accurate to five digits. The table of nine y and x values duplicates Table 22.
64 NUMERICAL ANALYSIS

12. (Kepler's Equation for Halley's Comet) Apply Newtons Method to Kepler's equation x =
0.967 sin (x) + y for y = 0 to y = 2:n: in increments of :n:/4 to produce an inverse table. The table of
nine y and x values duplicates Table 23 (see also Figure 20).
13. (Kepler's Equation for Pluto) Apply the method of fixed points to x = y + 0.2486 sinx, Keplers
equation for the planet Pluto, to create an evenly spaced inverse table with y = 0 to y = 2:n: in
increments of :n:/4. The result is similar to Table 2.2..
14. (Ellipse Parameters) Verify p = b2 /a in equation (4), using the focal distance given by equation (5)
and Figure 17.
15. (Elliptical Coordinates) Verify equations (8), (9) and (0).
dv b
16. (Kepler's Equation Derivation) Verify -
dE
= -;
r
see equation (4).

l.6 Implicit Functions

In the heat transfer problem of the Introduction, when the heat input Q of the stirrer was
taken into account it was found that the equilibrium temperature To of the system was
related to Q by the equations

40h - Q -lOh Q
(1) -,--------'. = e
SOh - Q ' To = h'
where h was the coefficient in Newton's law of cooling. On eliminating h between these
equations we find that
40 -
- - -
To =e-lOQIT0
(2)
80 - To
Clearly, To is a function of Q that is defined implicitly by equation (2). An algorithm for
computing To = To(Q), is given below.
Equation (2) may be reformulated as the equation F(Q, To) = 0, where

(3) F(Q T ) = 40 - To _ e-lOQITo.

, 0 80 - To

The problem of finding the solution set of an equation F(x,y) = 0 occurs frequently
in engineering analysis. For example, if T = f(x, y) is the temperature at a point in a
body with coordinates (x, y) then the equation f(x, y) = c defines a curve of constant
temperature, or isotherm. Note that this equation has the form F(x,y) = 0 with
F(x,y) = fex,y) - c. Similarly, if p = gex,y) describes a pressure distribution then the
equation g(x,y) = c defines a curve of constant pressure, or isobar. A third example
is provided by the stream function 1jI(x,y) of fluid dynamics. In this case the curves
defined by 1jI(x,y) = c are the stream lines of the fluid. Physically, these are the curves
 1.6 IMPLICIT FUNCTIONS 65

that would be followed by a small test particle that drifts with the fluid. Mathematically,
they are curves that at each point have the direction of the fluid velocity Many additional
examples of physical problems that lead to an equation F(x,y) = 0 will be found in
Chapter 2 and subsequent chapters.
The Implicit Function Theorem. In the examples given above it is intuitively clear that
the equation

(4) F(x,y) = 0

should define y as a function y = f(x) that satisfies F(x,f(x)) = 0 for every value of x.
However, this is not always the case. For example, the function F(x,y) = x 2 +y2 + 1 is
never zero, and hence in this case equation (4) has no solutions. Another case for which
(4) cannot be solved for y occurs when F(x,y) is independent of y. These examples
show the need to find properties of the function F(x, y) that guarantee the solvability of
equation (4). The following three properties are known to be sufficient.

(A) F(x,y), of(x,y)lox, and of(x,y)lOy are defined and continuous in a region D
in the (x, y)-plane.
(B) There is a point (xo,yo) interior to D such that F(xo,yo) = O. (A point is said
to be interior to D if it is the center of a disk that is contained in D.)
(C) of(xo,yo)!oy i= O.
If conditions (A), (B) and (C) hold then there is a rectangle al < x < a2, bl < Y < b2
contained in D with the property that for each x in the interval al < x < a2 the equation
(4) has a unique solution y = f(x) such that bl < Y < h The function f(x) satisfies the
equations Yo = f(xo) and F(x,f(x)) = 0 for al < x < a2· Moreover,f(x) is continuous
and has a continuous derivative.
The preceding results are one form of the implicit function theorem of advanced
calculus [T-MJ. The situation described by the theorem is illustrated in Figure 21.

y
y = y(x)
domainD

®~~'k
YO +b

Yo
YO - b

'--------------:~ x FIGURE 21. The implicit function

xo - a xo Xo +a theorem
66 NUMERICAL ANALYSIS

A simple illustration of the theorem is provided by the function F(x,y) = X2+y2_1

with derivatives aFlax = 2x and aFlay = 2y. For this function condition (A) is satisfied
in the entire (x,y)-plane. If we choose (xo,Yo) = (0,1) then (B) and (C) are also
satisfied and the solution of F(x, y) = 0 guaranteed by the implicit function theorem
is the function y = j(x) = .Jl=X2. It exists for -1 = al < x < a2 = 1. Note that
rex) = -xl";1 - x 2 is not defined at x = ±1 and hence the interval in which the
solution exists cannot be enlarged.
If in the preceding example we make the choice (xo,Yo) = (0, -1) then the
conditions of the theorem again hold, and the solution is y = j(x) = -.Jl=X2. Again
we may take -1 = al < x < a2 = 1. Note that the theorem is not applicable when
(xo,Yo) = (1,0) because aF(1, O)!ay = 0 and so condition (C) is not satisfied. However,
aF(1,O)!ax = 2, and hence the theorem may be applied with x and y interchanged.
Thus F(x,y) = 0 may be solved for x in terms of y near (1,0). The solution is clearly
x= J1- y 2

Note that the uniqueness of the solution y = j(x) of F(x, y) = 0 may fail if condition
(C) is not satisfied. A simple example is provided by the function F(x,y) = x 2 - i at
the point (0,0). For this case aF(O, O)!ay = 0, so that (C) fails and F(x,y) = 0 has the
two distinct solutions y = x and y = -x that pass through the point (0,0).
To compute a value Yl = j(Xl) of the implicit function defined by F(x,y) = 0 we
need only substitute x = Xl in the equation and then use one of the numerical methods
of Section 1.4 to solve the equation F(Xl,y) = 0 for y = Yl. This process is illustrated
here by two examples. Many additional examples will be found in subsequent chapters.

EXAMPLE 1. A Heat Transfer Problem. Recall that in the heat transfer problem of the
Introduction, To = To(Q) was the equilibrium temperature of the stirred flask, in degrees
Celsius, when the stirring produced a warming of Q degrees Celsius per minute. To(Q)
is defined implicitly by equation (2). However, it will be more convenient here to solve
the first of equations (1) for h = h(Q) and then obtain To(Q) from the second equation.

Proceeding as in Example 1, Section 1.4, we can show that the first equation of (1)
is equivalent to the equation

(5) x = log (2 + _Q_),

4x- Q
where x = lOh.

This equation may be solved easily by the method of fixed points. The case Q = 0.01
was treated as Example 1, Section 1.4. Application of the same method to (5) gives x =
x(Q), h(Q) = (O.l)x(Q) and To(Q) = Qlh(Q) = lOQlx(Q). This process gives Table 24.
 1.6 IMPLICIT FUNCTIONS 67

TABLE 24. Equilibrium temperature To as a function of heat input Q

Q 0.0 0.1 0.2 0.3 04 0.5 0.6
x(Q) 0.693 0.711 0.729 0.747 0.766 0.784 0.802
To(Q) 0.00 141 2.74 4.02 5.22 6.38 748

EXAMPLE 2. A Fluid Flow Problem. The flow of an incompressible nonviscous fluid

near a circular cylinder can be described by a stream function 1/1. The function 1/1 will be
defined by means of rectangular coordinates x, y, z with the z-axis along the axis of the
cylinder. If at large distances from the cylinder the fluid has a uniform velocity V in the
direction of the x-axis then the stream function is given by

(6) 1/I(x,y) = V (r - ar 2
) sin (),

where rand () are polar coordinates (so that x = r cos (), y = r sin ()) and a is the radius
of the cylinder.

The flow near the cylinder may be visualized by means of the stream lines 1/I(x, y) =
c. Figure 22 shows the stream lines for the parameter values V = 1, a = 1 and
c = 0,0.2,0.4, ... ,2.0. Note that the dividing streamline 1/I(x,y) = 0 consists of the
portions of the x-axis where Ixl > 1 (so that () = 0 and () = n, r > 1) plus the
circumference ofthe cylinder x 2 + y2 = 1 (so that r = 1,0 ::: () ::: 2n). Streamlines with
positive (respectively, negative) values of c pass above (respectively, below) the cylinder
in Figure 22.

2.0
1.6

1.2

0.8

0.4
0.2

t
values
of c

FIGURE 22. Stream lines for fluid flow

past a circular cylinder
68 NUMERICAL ANALYSIS

The stream function (6), with V = 1 and a = 1, may be written

(7) 1/r(x, y) = (1 - r12) r sin e= (1 - x2 ~ y2 ) y.

Hence, an equation for the streamlines is

(8) ( 1-~)Y=C.
x +y

On subtracting e from each side and multiplying by the denominator, we find the
equivalent equation

(9)

This has the form (4) with

(10)

For fixed values of e and x, the equation F(x,y) = 0 is a cubic equation for y that is
difficult to solve algebraically However, it is easy to solve by Newton's method

F(x,Yn)
(11)
Yn+! = Yn - 8F(x,Yn)18y'

Note that

(12) 8F(x,y)18y = 3l- 2ey + (x 2 - 1).

On substituting this into (11) one obtains, after simplification, the iteration

2y~ - ey~ + ex 2
(13)
Yn+l = 3y~ - 2eYn + (x 2 - 1)

To construct a specific streamline, the numerical value of e is substituted in (13) and

the iteration is carried out for a succession of x-values. The streamlines in Figure 22 were
computed by this method. The coordinates of several points on the streamline e = 0.2
are given in Table 25.

TABLE 25. Points on the streamline 'IjJ(x,y) = 0.2

x 0.00 0.25 0.50 0.75 l.00 1.50 2.00 3.00 4.00 5.00
y 1.10 l.08 l.00 0.86 0.66 0.35 0.27 0.22 0.21 0.21
 1.6 IMPLICIT FUNCTIONS 69

Classification. Classify as explicit or implicit in y.

1. y = tanx. 2. ytanx+ 1 = l/x.

4. (Multiplicity) Find an example of an implicit equation F(x,y) = 0 that is satisfied by three explicit
equations of the form y = j(x), all of which pass through the point (l, 1).

5. (Implicit Function Theorem) Let F be continuously differentiable. Suppose F(x, y) = 0 is satisfied by

the three explicit equations y = 1 + x, y = 1 + 2x and y = 1 + 3x. Does the implicit function theorem
apply to F at x = 0, y = 17

6. (Heat Transfer) Use the method of fixed points on the equation x = log(2 + yl(4x - y» for y = 0 to
Y = 0.6 in increments of 0.1 to produce an inverse table of y and x values. Apply the inverse table to the
heat transfer problem with To = 10ylx to verify the entries in Table 24.

7. (Streamlines) Use Newton's iteration method Yn+ 1 = Yn - F(x,Yn)/Fy(x,Yn) on the streamline equation
F(x,y) = y3 - c/ + (xl - l)y - cx l to generate graph data for the streamline at c = 0.2. Use x = I
to x = 0 in steps of -0.25, solving for the unknown y in each case. This verifies one of the curves in
Figure 22 and also portions of Table (25).

8. (Implicit Function Theorem) Explain why the equation yl + 2y + 2 + x 2 = 0 cannot be solved for y
in terms of x.

9. (Fluid Flow) Convert the polar streamline equation (r - r- 1) sine = c into rectangular coordinates to
obtain (l - (x 2 + /)-1)y = c.

10. (Streamlines) The streamline 0- (xl + /)-1)y = c gives rise to a differential equation for the curve

Verify this equation and explain why all streamlines have the same differential equation, regardless of
the value of c.

n. (Implicit Function Theorem) Suppose curve y = y(x) satisfies the equation F(x,y) = 0, a :::: x :::: b,
and assume that F and yare continuously differentiable. Find a differential equation for y in terms of
the partial derivatives JFlJx and JFlfJy.
Implicit Function Theorem. Verify that the equation F(x,y) = 0 satisfies F(O,O) = 0 and compute
FyeO,O)

12. F(x,y) = (y-x)(y-x l ). 14. F(x,y)=/-y+sin(x).

13. F(x,y) = y - xl. 15. F(x,y) = y3 + 2y + sin(x).

16. (Implicit Function Theorem) Apply the implicit function theorem to the equation xl + yl = 1 to
solve for y in terms of x at any point ex,y) on the curve. Which points (x,y) give trouble with the
hypotheses of the theorem 7 In what sense must a given solution agree with one of the two solutions
obtained by college algebra (namely, -~ and )1 - x 2 )? In your answer, pay close attention to
the domain of definition of a [unction.
70 NUMERICAL ANALYSIS

17. (Heat Transfer) Solve 4y - x = c-Y for y by clearing fractions, obtaining

Sy-x

y = C(x,y) = (x -xe-Y )/(4 - Se-Y ).

Explain, by means of the size of ICy (x,y) I, why the method of fixed points fails.
18. (Implicit Function Theorem) Let F be continuously differentiable. Assume that two explicit solutions
Y1 (x) and Y2 (x) satisfy F(x,y) = 0 and pass through (XO,YO)· Given that Yl and Y2 differ in every
interval about x = Xo, show that Fy(xO,Yo) = O.
19. (Implicit Function Theorem) Read a proof of the implicit function theorem and then explain it in your
own words. If you can construct a better proof, then outline the ideas of the proof and defend your
position.

40h-- -
20. (Heat Transfer Problem) Write - Q= e- lOh
in the fixed point form x = log ( 2 + -Q)
- ,
SOh - Q 4x - Q
where x = lOh.
lOQ
21. (Heat Transfer Problem) Derive the temperature formula To = - from To = QIh and x = lOh.
x

1.7 Numerical Summation of Infinite Series

The solution of many problems of applied mathematics requires the computation of the
sum 5 of an infinite series of known terms ak:

L
00

(1) 5 = ak = al + a2 + a3 + ....
k=l

Many such problems will be encountered in Chapter 3, Chapter 8 and Chapter 9. Let us
introduce the nth partial sum of the series,
n
(2) 5n = L ak = a1 + a2 + a3 + ... + an,
k=l

and the corresponding nth remainder

L
00

(3) Rn = ak = an+l + a n+2 + ... ,

k=n+1

so that

(4) 5=5 n +Rn , n= 1,2, ....

Then the convergence of an infinite series in (1) means that to every number E > 0 (no
matter how small) there corresponds a positive integer N = N(E) such that

(5) IRnl = IS - 5nl :::: E for every n :::: N(E).

 1.7 NUMERICAL SUMMATION OF INFINITE SERIES 71

Note that if (5) holds for a particular integer N(E) then it will also hold for larger integers.
Forthe theory, the exact value of N(E) is unimportant. For practical computations we shall
wish to choose N(E) as small as possible. This point will be illustrated in several examples.
The Geometric Series. This series is defined by (1) with ah = X k- I , where x is a
real parameter:

(6) 5 = Sex) = 1 + x + x 2 + x 3 + ....

Thus

(7)

Clearly (6) implies that 5(1) does not exist. For x ::f 1 a well-known algebraic identity
implies that
1
(8) --=l+x+x + .. ·+x
l-x
2 n-I
+--,
Xn
l-x
so that
xn
(9) Rn =- -
l-x'
n = 1,2,3, ....
It follows that Rn (x) --+ 0 as n --+ 00 if and only if -1 < x < 1 and hence

(10) 5= -1- = 1 + x + x 2 +... for - 1 < x < l.

l-x
We shall calculate the optimum N(x, E) for E = 0.1 and several values of x.

EXAMPLE 1. Geometric Series with x = 0.9, € = 0.1. Here (9) gives

(0.9)n n
Rn(0.9) = - - = 10(0.9) < 0.1,
1- 0.9
or

(ll) (0.9)" < 10- 1 .

We shall find the allowable integers n by solving this inequality. Taking the logarithm
of (11) and recalling that logx is a monotone increasing function gives the equivalent
inequality

n 10g(0.9) < -2 or n(-0.0458) < -2,

or
2
(12) n > -- = 43.67.
0.0458
72 NUMERICAL ANALYSIS

Since N(x, E) is a positive integer, (12) gives

(13) N(0.9, 0.1) = 44.

Thus 44 terms are needed to reduce the error to less than 0.1.

EXAMPLE 2. Geometric Series with x = 0.99, € = 0.1. Here we get

(0.99)n n
Rn ( 0.99) = = lOO(0.99) < 0.1,
1 - 0.99
or

(0.99t < lO-3.

In the same way one can show that

(14) N(0.99, 0.1) = 688.

Clearly, N(x, E) grows rapidly as x approaches 1.

Series whose Terms are Functions. Consider a series whose terms are functions of one
or more variables:
00

(15) Sex) = Lik(x)

= fl (x) +hex) +hex) + . . . for a::: x ::: b.

The corresponding partial sums and remainder will be written

n
(16) Sn(x) = Lfk(X) = fleX) +hex) + f3(X) + ... + fn(x),
k=l

00

(17) Rn(x) = L fk(X) = fn+l(x) + fn+2(X) + ....

k=n+l
An example is the geometric series (10). Note that for each value x of the interval a :::
x ::: b the series (15) is a different numerical series. Thus the convergence statement (15)
means that for each x on a ::: x ::: b and for each E > 0 there is an integer N = N(x, E)
such that

(18) IRn(x) I = IS(x) - Sn(x) I ::: E for every n 2: N(X,E).

We have seen with the geometric series that N(x, E) can vary greatly with x.
Uniform Convergence. If we can find an integer N = N(E), independent of x, such that

(19) IRn(x)1 = IS(x) - Sn (x) I ::: E

 1.7 NUMERICAL SUMMATION OF INFINITE SERIES 73

for all n :::: N(x, E) and all a ::::: x ::::: b then we say that the series (15) converges uniformly
for a ::::: x ::::: b. For numerical summation of a series (15), uniform convergence on an
interval a ::::: x ::::: b is one of the most useful properties a series can have. Uniform
convergence is illustrated by the geometric series (lO). Using the remainder Rn(x) =
xnl(l - x), it can be shown that it converges uniformly on any interval a ::::: x ::::: b such
that -1 < a < b < 1.

EXAMPLE 3. A Trigonometric Series. Consider the series

( ) _ ~ cos(2k + 1)x _ cosx cos3x ...

(20) Sx - ~ (2k + 1)2 - 12 + 32 + , -Jr ::::: x S n.
h=O

It is known that the sum Sex) equals ~ (~ - Ix l) for -n ::::: x ::::: n; see Section 1.4,
Example 2. However, this fact will not be used here. Instead, we shall show how to
evaluate Sex) numerically to within a prescribed error E.

To estimate Rn(x), let ¢(x) = 1/(2x + 1)2 and note that because ¢(x) is monotone
decreasing for x :::: 0 we have
[h+l
(21) ¢(k + 1) S Jh ¢(x) dx;

see Figure 23.

For the remainder Rn(x) we have

R (x) = ~ cos(2k + l)x

n ~
h=n
(2k + 1)2 '

and so by the triangle inequality and (21),

:xl Icos(2k + l)x I 00

IRn(x) I S {; (2k + 1)2 S ( ; ¢(k)

y
~</J(X)

</J(k) </J(k + 1)

k k+1 FIGURE 23. The integral test for L </J(k)

74 NUMERICAL ANALYSIS

::::
1n-1
00

</lex) dx = 2(2n - 1)
1

for all x. Thus IRn (x) I :::: E if 0.5/(2n - 1) :::: E. For example,

1
IRn(x) I :::: 2(2n _ 1) < 0.01 if n::: 25.5.

Thus a suitable value of n is N(O.01) = 26. We conclude that

cosx cos3x cos51x
Sex) = -F- + -3-
2
+ ... + -SF- + R26 (X) '
where

IR26 (X) I :::: 0.01 for - n :::: x :::: n.

Other trigonometric series that can be analyzed similarly are

j(x) = sinx _ sin3x + sin5x _... for 0:::: x:::: n

12 32 52

(see Section 8.3) and

jie) sin x sin 3x sin 5x

x =--+--+--+
P 33 53
... for 0:::: x:::: n.

Many others occur in Chapter 9.

Geometric Series Estimates. Estimate the minimum number of terms N(E) for 2-digit accuracy of the
truncated geometric series Sn (x), using geometric series formulas.

1
1. The geometric series whose sum is - - . Use x = -0.9 and E = 0.01.
l+x
2. The series whose sum is InO + x), obtained by termwise integration of the series for 1/0 + x). Use
X = -0.9 and E = 0.01.

1
3. The series whose sum is --2' Use x = 0.5 and E = 0.01.
l+x

4. The series whose sum is tan -1 (x), obtained by termwise integration of the series for 1/0 + x 2 ). Use
x = 0.5 and E = 0.01.
 1.7 NUMERICAL SUMMATION OF INFINITE SERIES 75

Trigonometric Series Estimates. Estimate the minimum number of terms N(E) for 2-digit accuracy of
the truncated series Sn (x), using the integral test method of the text. The estimate should be unifonn in x
with E = 0.01.
~(_ )k sin(2h + l)x
5. ~ 1 2 .
k=O (2h + 1)

~ sin(2h + l)x
6. ~ 3 .
k=O (2h + 1)
C HAP T E R 2
 Ordinary Differential
Equations of First Order

This chapter takes up a major theme of this text: differential equations and their
applications in science and engineering. Differential equations were discovered by Isaac
Newton some time before 1666. Newton was so convinced of the importance of his
discovery that he revealed it only in an anagram, one of the scientific puzzles that scholars
of his time set for one another. The anagram, when resolved into Latin and then translated
into modern English, states, "It is useful to solve differential equations." The soundness of
Newton's judgment is confirmed by the many textbooks and journals of modern science
and engineering that are filled with examples of differential equations and their solutions.

2.1 The Nature of Differential Equations

In the reservoir-filling problem of Section 1.2 the volume V of water in the reservoir at
time t satisfies
dV
(1) at = q(t),

where q(t) is the rate of flow into the reservoir, as recorded by a flow meter. Equation (1)
is one of the simplest examples of a differential equation. The volume V = Vet) in the
reservoir is to be determined from its rate of change q(t) and its initial value Vo = Veto).
The fundamental theorem of calculus implies that the equation

(2) V = Vo + t q( r) dr
llO
78 2 ORDINARY DIFFERENTIAL EQUATIONS OF FIRST ORDER

defines the unique solution of (1) that passes through the point (to, Vo) in the (t, V)-plane.
The heat transfer problem of the introduction led to the differential equation
dT
(3) dt = -h(T - 20).

The solution T = T(t) that passes through a prescribed initial point (to, To) may be
found by the method of the Introduction to be

(4) T = 20 + (To - 20)e- h(t-t o).

It is easy to verify directly that this function satisfies (3) and T(to) = To. Again there is
a unique solution through each point (to, To) of the (t, T)-plane.
As a third example, consider the simple L, R electrical circuit of Figure 1. Here Land
R are prescribed positive constants, the circuit inductance and resistance, respectively,
and E(t) is the prescribed output of a generator or other voltage source.
Kirchhoffs circuit laws imply that the electric current i(t) satisfies the differential
equation

(5) ti' (t) + Ri(t) = E(t).

For simplicity, only the special case L = 1, R = 1 and E(t) = sin t (alternating current)
will be analyzed here, so that

(6) i' (t) + i(t) = sin t.

It is not immediately obvious how to solve this equation. However, if the differentiation
formula

(7) (e t 0' =8 (t) + et i(t) = l (i' (t) + i(t))

is noted, then (6) is seen to be eqUivalent to the equation

(8)

R
FIGURE 1. L. R electrical circuit with voltage
source E(t)
 2.1 THE NATURE OF DIFFERENTIAL EQUATIONS 79

The fundamental theorem of calculus applied to (8) gives

(9) ct i(t) = f ct sin t dt = ~ct(sin t - cos t) + c.

The constant C is determined by the initial condition i(to) = i o, whence

(10) C = Cto (.
10
1 1.)
+ "2 cos to - "2 SIn to .

A short calculation then gives the solution

(ll)
2
1
i(t) = - (sin t - cos t) + c-(t-to) (1
+-
io
2
1)
cos to - - sin to
2
,

as may be verified directly. There is exactly one solution curve i = i(t) through each
point in the (t, i)-plane.
Each of the three examples above involves a differential equation of the form

dx
(12) dt = F(t,x),

where x is an unknown function of the independent variable t and F(t, x) is a prescribed

function of its two variables. Thus in the first example, x = V and F(t, x) = q(t), which
is independent of x. In the second example, x = T and F(t,x) = -hex - 20), which is
independent of t. In the third example, x = i and F(t, x) = -x + sin t, which contains
both variables.
Equations of the form (12) provide models for a great variety of physical, chemical
and biological processes. Examples presented in this and subsequent chapters include
problems from mechanics, hydraulics, electrical circuits, heat transfer and chemical
kinetics. The goal of this chapter is to develop analytical methods and numerical
algorithms to solve any differential equation of this type.
What is a Solution of a Differential Equation? The notation used in equation
(12) implies that the unknown x is a function x = ¢(t) of the independent variable t. In
each application the variable t will range over some definite interval Q of the real axis.
This suggests the following.
Definition. A function ¢(t) is said to be a solution of the differential equation (12) on
the interval Q if

1. ¢(t) is defined for every t in Q.

2. The first derivative ¢'(t) exists for each t in Q.
3. ¢'(t) = F(t, ¢(t)) for each t in Q.
80 2 ORDINARY DIFFERENTIAL EQUATIONS OF FIRST ORDER

Geometric Representation of Solutions of a Differential Equation. If x = </J(t) is

a solution of (12) then the point (t,</J(t)) traces a curve C in the (t,x)-plane as t runs
through the points of Q. The existence of </J' (t) implies that C is a continuous curve. The
differential equation (12) has a simple geometric interpretation. It means that the slope
of the solution curve C at each of its points is a function of the coordinates of that point.
The Direction Field of a Differential Equation. The preceding geometric representa-
tion of solutions of (12) offers a way to visualize the geometric meaning of the differential
equation. Imagine drawing through each point of the (t,x)-plane a small line segment
with slope F(t, x). This diagram, which is called the direction field of the differential
equation, gives a good idea of the form of the integral curves of the differential equation;
see Figure 2.
The integration of the differential equation may now be interpreted as threading a
curve C through the direction field in such a way that C passes through each of its points
(t, x) with the slope F(t, x).
Equation (12) is called an ordinary differential equation because the unknown
x is a function of a single independent variable t. Equations that involve the partial
derivatives of an unknown function of two or more independent variables are called
partial differential equations. Such equations will be treated in Chapter 7, Chapter 8
and Chapter 9. Equation (12) is said to be of the first order because only the first
derivative of x appears in it. Equations that contain derivatives of the second or higher
orders are studied in Section 2.5 below and in Chapter 3 and subsequent chapters.
Sections 2.2 through 2.5 treat several speCial classes of equations that can be
solved by direct integration. The numerical integration methods of Section 2.3 provide
numerical algorithms for the solution of these equations. Sections 2.6 and 2.7 develop
new algorithms that are applicable to any first-order equation.

,7////,7/,7// //
jjjjjjjj,//jjj
Illlllll 1111
~~~~~~~!~~~~j
1llllll'!lllll
I11111 I11111
///// ///////

~~~~~~~~~~~~~ FIGURE 2. Direction field of the equation x' = x(l - x)

 2.1 THE NATURE OF DIFFERENTIAL EOUATIONS 81

IExercises 2.1 I
f ;,'"' ~~'=~""~_~'''''''' ___ '' ~ _,oj

f:
1. (Reservoir Volume) In each of the cases below, given the flow rate q(t), the initial volume Vo and the
time interval [a, bl, verify the given volume V using the formula V = Va + q(t)dt.

[a,b] q(t) Vo V

[0,24] 1 + 2t 601
[0,3] 1 + 2t + t 2 22
[0,1] 2e- t 1.13 3.13 - 21e
[0, rr] sin 2 (t) 1 1 + rr/2
[0,2]
[0, rr/4]
sinh(t)
tan(t) ° cosh(2) - 1
1 + log(2)/2

2. (Reservoir Volume) Let the flow rate q(t) be given as the linear interpolant of the table below. Using the
volume formula V = 1.5 + f5 4 q(t)dt, determine V. The units for q are 2.3 x 105 acre-feet per hour and
t is in hours.

3. (LR Circuit) Solve the equation i'(t) + i(t) = sint with initial values iCO) = ° = 5. Plot both
°
and i(O)
solutions together, using ~ t ~ 6rr and -1 ~ i ~ 5.

4. (LR Circuit) Consider the LR circuit of Figure 1 with parameter values L = 1, R = 1 and voltage source
E(t) = sin t. Specialize the solution i(t) obtained in (11) to the initial condition i(O) = iO to obtain

i(t) = (iO + ~) e- t + ~ (sin t - cos t).

TABLE 1. Flow meter data

q q q

0.0 1.49 9.0 2.26 18.0 1.98

1.0 1.34 10.0 1.90 19.0 1.63

2.0 1.41 11.0 2.15 20.0 1.52

3.0 1.64 12.0 2.16 21.0 1.52

4.0 1.49 13.0 2.15 22.0 1.34

5.0 1.71 14.0 2.32 23.0 1.31

6.0 1.80 15.0 2.22 24.0 1.42

7.0 1.79 16.0 1.98

8.0 2.18 17.0 2.08

82 2 ORDINARY DIFFERENTIAL EQUATIONS OF FIRST ORDER

5. (Flask Cooling) Apply the cooling equation u' (t) = -h(u(t) - 20) and initial data u(O) = 100 to verify
the second and third columns of the table below.

Assumption Value of h Cooling Time to 21 0 C

log SO
u(lO) = 60 h = 0.1In(2) t = 1 0 - - = 63.22
log2
10gSO
uclO) = 70 h = 0.lln(S/5) t = 1 0 - - - = 93.23
log(S/5)
10gSO
uclO) = so h = 0.lln(4/3) t = 10--- = 152.32
log(413)
log SO
u(20) = 50 h = (lI20) In(S/3) t = 2 0 - - - = S9.35
log(S/3)
log SO
u(30) = 35 h = (1130) In(16/3) t = 30 = 7S.53
logCl6/3)

6. (LR Circuit) Solve the initial value problem

i'(t) + O.4i(t) = e- t , i(O) = 120,

to obtain the exact answer i(t) = (-5e- t + 365e- 2US )/3.

7. (Solution of a Differential Equation) Verify that y = e- 3x + x 2 is the solution of the first-order

differential equation y' + 3y = 3x2 + 2x with initial condition yeO) = 1.

8. (Direction Field) Determine a direction field for the equation x' (t) = sine4t)x(t) on -1 ~ t ~ 1,
-1 ~ x ~ 1, using 25 line segments.

9. (Direction Field) Plot the direction field for the equation x' (I) = x(t)(2 - xU)) in order to determine
all possible values oflimHoo x(l).
Direction Fields. In the folloWing exercises, sketch a direction field and plot several approximate solution
curves.

10. x' = 3x1t. 13. XXi +t = O.

11.x' =t+x. 14. x' = -tx.

12. x' = Ux. 15. x' =x3

Method of Isoclines. An isocline is a curve of constant inclination. In a direction field, line segments
with the same slope are called isoclines. A direction field can be created by plotting the curves F(t, x) = c
for certain values of c, then adding line segments to these curves. This is called the method of isoclines.

16. (Identifying Isoclines) In Figure 2 above, identify four sets of isoclines. Suggestion: Trace the figure,
and then color the isoclines (red, green, blue, yellow).

17. (Plotting Isoclines) Plot some isoclines for the equation x' (I) = (l - x(t)) (2 - x(t)), using a square
of side 10 and center (0,0). Add line segments to make a direction field plot.
 2.2 SEPARABLE EQUATIONS 83

=
18. (Isoclines and Exact Solutions) The equation Xl (I) (l-x(t»(2 -x(l) has constant solutions 1 and
2. The exact solution for x(O) =f. 1 and x(O) =f. 2 is log Ix(t) -II-log Ix(t) - 21 + t = C. Superimpose
the solution curves for x(O) = 1, x(O) = 2 and x(O) = 3 onto the direction field.

2.2 Separable Equations

The first-order equation (12) of Section 2.1 is said to be separable if it has the form
dx
(1) dt = F(t)G(x).

We shall assume that the functions F(t) and G(x) are defined in intervals al < t < a2
and b1 < x < b2 , respectively. The rectangle indicated in Figure 3 will be called the
domain of definition of equation (1). The graph of each solution x = ¢(t) of equation
(1) must lie in this rectangle, as illustrated in the figure, since F(t) or G(x) , or both, are
undefined at points (t,x) outside the rectangle.
We shall see that exactly one solution curve of (1) passes through each point of the
rectangle.
If G(b) = 0, where b1 < b < b2 , then the horizontal line x = b = constant is a
db
solution of (1), since dt = F(t)G(b) = 0. The line x = b then divides the rectangle of
Figure 3 into two rectangles, whose integral curves are distinct and which may be studied
separately. For this reason we shall restrict our attention to rectangles in which G(x) #- 0.
Let x = ¢(t) define the solution of (1) that passes through a point (to, xo), as
indicated in Figure 3. Then rjJCt) will be defined in some interval tl < t < t 2 , where
al :s tl < to < t2 :s a2. To calculate the value ¢(t) at any point of this interval, note that
(2) ¢(to) = Xo

and
(3) ¢/(r) = F(r)G(¢(r))

..... ~
x(t)

Xo
•
. .

to FIGURE 3. Domain of definition for equation 1

84 2 ORDINARY DIFFERENTIAL EQUATIONS OF FIRST ORDER

for every r between to and t. The last equation may be rewritten as

</J1(r)
(4) G(</J(r)) = F(r),

since it is assumed that G(</J( r)) 1= O. Now, integration of (4) over the interval to :::: r :::: t
(or t :::: r :::: to if t < to) gives

(5) 1t </J1(r)
( ( )) dr =
1t F(r)dr.
to G</Jr to

In the integral on the left, the variable of integration may be changed from r to ~ = </J(r).
The rules for changing the variable of integration in a definite integral give

t-,:-.....,--,-dr
</J1(r) 14>(t) -d~-
(6) 1to G(</J( r))
-
- Xo G(~)'

If x = </J(t) then (5) can be written as

(7) l -~() =
x

xoG~
1t to
F(r)dr.

This equation defines the solution x = </J(t). In fact, the implicit function theorem of
Section 1.6 implies that (7) can be solved for x = 4>(t). Moreover, it may be verified by
differentiating (7) that the function </J(t) so defined is a solution of (1).
The solution x = </J(t) of (7) may be constructed explicitly by means of the functions

(8) let) = 1t F(r)dr, al < t < a2,

to

and

(9)
J(x) =
l Xo
X
d~
G(~)' b1 < X < b2 .

Equation (7) can be written as

(10) J(x) = let).

Notice thatl'(x) = I/G(x) 1= 0 and hence y = J(x) is a monotonic function and has an
inverse x = r 1 (y). Applyingr 1 to the two sides of (10) gives the explicit solution

(ll)

If the integrals let) andJ(x) are elementary functions then (11) gives </J(t) as an elementary
function. In the more typical case that I and/or J are not elementary functions, they may
 2.2 SEPARABLE EQUATIONS 85

r
be computed by numerical integration, and I may be computed by the methods of
Section 1.5. Both possibilities are illustrated in the examples and exercises that follow.
The Equation dxldt = F(t). This simple differential equation, obtained by taking
G(x) = 1 in (1), is integrated in most elementary calculus texts. The solution

(12) x = Xo + t F(r)dr,
1o
t

which follows from the fundamental theorem of calculus, can also be obtained from (7)
by taking G(~) = 1. Two applications of this equation will be presented.

EXAMPLE 1. Filling a Reservoir. The volume V of the water contained in a reservoir at

time t satisfies the equation

dV
(13) dt = q(t), t::: 0,
where q(t) is the fill rate. Thus

(14) V = Vo + 1t q(r) dr,

where V = Vo when t = 0. The key feature of this example is that q(t) is known only at
a table of values (see Table 1 in Section l.2). In numerical work q(t) is represented by
an interpolation function.

The function V was computed by numerical integration in Example 2 of Section 1.3,

using the numerical data for q(t) of Table 1 in Section l.2. The results are given in Table
6 of Section 1.3.

EXAMPLE 2. Compression of an Elastic Column. Consider a vertical column made of

an elastic material, such as steel, and subject to a load F uniformly distributed over the
top, as indicated in Figure 4. The amount that the column will yield under the load F
will be determined.
The column need not have an axis of revolution. The important characteristic is A(x) ,
the cross-sectional area of the column in a horizontal plane at the distance x from the
top. The total length of the column when unloaded will be denoted by £. The length
when loaded will be written as .c - u, where u, the total compression, is to be calculated.
To begin, imagine that a load F is applied to a cube with side 1 made from the
column material. In this case Hooke's law predicts that the compression u will satisfy
F = Eu, where E is a constant, called Young's modulus, whose value is characteristic of
the material. Values of E for some typical metals, expressed in English units, are below.
86 2 ORDINARY DIFFERENTIAL EQUATIONS OF FIRST ORDER

o
x
x+dx

x FIGURE 4. Elastic column with load F

Material Young's modulus in lb/sq in

steel 30 x 106

aluminum 10.2 x 106

magnesium 6.5 x 106

If a load F is applied uniformly to a cylindrical column of cross-sectional area A and

height 1 then each unit of cross-sectional area experiences a load FIA and hence the
compression u satisfies FIA = Eu, or u = F/(EA). Finally, if the load F is applied
uniformly to a cylindrical column of cross-sectional area A and height e then the
displacement is u = F.et(EA) because each column segment of unit height is compressed
F/(EA) units. If the column has a variable cross-section A(x), the above rule may be
applied to a very thin slice of the column between horizontal planes at heights x and
x + ax.
Each such slice will experience a compression du = Fdx/(EA). Hence the total
compression is obtained by integrating the differential equation

du F
(15)
ax EA(x) ,
0:::: x:::: e,
which gives

F t ax
(16)
u= E10 A(x)'

As a numerical example consider a conical steel column with radii fO = 1 and fl = 2 at

the ends and length .e = 72, all units inches. Then the radius at distance x from the end
is f = fO + e-l(fl - fO)x = 1 + (xl72), and hence A(x) = nf 2 = n(l + (xl72))2. The
 2.2 SEPARABLE EQUATIONS 87

integral in (16) is easily evaluated in this case to give

36
(17)

and the total compression is

(1S) u= ( 36
30n x 10
6) F = (3.S2 x 1O-7)F inches,

where F is the load in pounds. In particular, a load of F = 3000 lb will produce a

compression of u = l.15 X 10-3 in.
The Equation dxldt = G(x). If the factor F(t) = 1 in equation (1) then (S) gives
l(t) = t - to and the solution formula (11) becomes

(19)

In this case only the functionr 1 needs to be computed. Two applications ofthis type of
equation will be given.

EXAMPLE 3. Heat Transfer. The heat transfer problem of the Introduction led to the
differential equation dT/dt = -h(T - 20). This has the form (1) with F(t) = 1 and
G(T) = -h(T - 20). The initial values were to = 0 and To = 100. Application of (19)
gives T = r I (t), or t = J(T), where by (9),

(20) J(T) = iT ~ = -liT

100 GCT') h 100
dT'
T' - 20
= ~log(~).
h T - 20
Solving the equation t = JCT) for T gives T = r l (t) = 20 + SOe- ht , as before.
EXAMPLE 4. Draining a Reservoir. The time needed to drain a reservoir from an initial
depth Xo to a final depth x is estimated in this example. The water is assumed to drain
through a valve located at the deepest point in the reservoir. If the depth of the water is
measured upward from this point then the water will exit with the speed

(21) v= /fiX,
where g is the acceleration of gravity. This principle is known as Torricelli's law. 1 If the
valve channels the water into a pipe of cross-sectional area a, then the rate at which the
water drains is avo This is the rate of decrease of the volume V of water in the reservoir,
so that
dV
(22) at = -avo

1 Published in 164] by Evangelista Torricelli (1608-1647), Italian physicist and mathematician.

88 2 ORDINARY DIFFERENTIAL EQUATIONS OF FIRST ORDER

dV
Now, dx = A(x) is the surface area of the reservoir when filled to depth x, and hence

(23) dV = dV dx = A(x) dx .
dt dx dt dt
Combining (21), (22) and (23) gives a differential equation for x that may be written as

dx _CX ll2
(24)
A(x) ,
c= aJ2g.
dt
dx
This has the form -:it = G(x) and hence the solution that satisfies x = Xo when t = 0 is
given by (10) with let) = t. Thus,

(25) t = -1
c Jxo
r Am~ = ~ l
~lI2 C x
xo A(~)d~ .
~1/2

Cylindrical Reservoir. Consider a cylindrical reservoir of radius R feet equipped with

a circular valve at the bottom of radius r ft. Then A = nR2 = constant and a = nr2. If
the value g = 32 is used, this gives c = 8nr 2 and Nc = (l/8) (Rlr) 2. With this notation
the draining time, in seconds, from depth Xo ft to depth x ft is

(26) t = ~ (~)21XO ~-1I2~ = ~ (~)2 (Xll2 _ X ll2 ).

8 r x 4 r 0

For example, if R = 10 ft, r = 0.25 ft, and Xo = 20 ft, then the time required to drain
the reservoir completely (x = 0) is
1
t = _(40)2(20)112 = 1789 sec = 29.8 min.
4
Natural Reservoir. In Example 6 of Section 2.3, the area function A(x) for a natural
reservoir was computed numerically from measured volume data. In this case the draining
time must be computed by numerical integration. As an example, the data of Table 12
in Section 1.3 will be used. A circular valve of radius r = 0.5 ft is assumed, so that
a = nr2 = 0.25n, and t is assumed to be measured in hours, so that

g = 32 ftlsec 2 = (3600)2 x 32 ftlhr2.

These values give

(27) c = (0.25)nJ0600)2 x 64 = 7200n = 2.2619 x 104 .

 2.2 SEPARABLE EQUATIONS 89

The initial depth, denoted by Xo above, will be taken to be X20 = l35 ft, so that

(28)
_ ~ f135 A(~)~
t(x) - 1/2
C x ~

is the time required to drain [rom a depth of l35 ft down to x ft (d. (25»). The notations
tk = t(Xk) and

(29)

will be used, so that

(30)

and

(31)

Note that (30) gives t20 = 0, since X20 = l35. Moreover, the trapezoidal rule (8) of
Section 1.3 applied to (31) with k = 20 gives the approximation

(32)

where h = X20 - XI9 = 5 feet and Fk = F(Xk)· For k .::: 18, equation (31) gives

and application of Simpson's rule (19) in Section 1.3 gives the approximation

(33)

where again h = Xk+1 - Xk = 5 ft. Table 2 shows the results obtained from (32), (33),
where Fk = Akl(C,jXh) and Ak is obtained from the data of Table 12 in Section 1.3.
90 2 ORDINARY DIFFERENTIAL EQUATIONS OF FIRST ORDER

TABLE 2. Draining times of a reservoir

k Xh tk hr k x, t, hr

20 135 0.00 10 85 10.42

19 130 0.99 9 80 1140
18 125 2.07 8 75 12.36
17 120 3.12 7 70 13.36
16 llS 4.17 6 65 14.33
15 110 5.23 5 60 15.23
14 105 6.27 4 55 16.12
13 100 7.33 3 50 17.02
12 95 8.37 2 45 17.85
11 90 9.39 40 18.65
0 35 1943

Separable Equations. Each of the exercises below is an initial value problem for a separable differential
equation. Solve each problem by the method of the text to obtain the solution in the implicit Jorm](x) = r(t).
Solve analytically for the explicit form x(t) = ]-1 (l(t)) when possible.

I. x' = t/x 2 , x(O) = l. 6. x' = _et cosx, xeO) = o.

2. x' = x 2 (t + 3)3, x(-3) = l. 7. x' = -1 + e-X, x(O) = l.
x+3 I t+1
x4 + l' x(O) = l.
I
3. x = --,x(O) = O. 8. x =
t+5
I (x 3 + 2x)t I (l + 2et )x ()
4. x = (x 2 + 1)(t2 + 1) ,x(l) = l. 9. x = t + 2et x 0 = l.

5. x, = x 2 - 6x + 13, x(O) = 3. 10. x' = eX 2 , x(O) = O.

Initial Value Problem. Each of the exercises below is an initial value problem for a separable differential
equation of theforrnx' (t) = J(t). Graph the solution on [0, IJ. In some cases a numericalintegration method
is required to produce the graph.

II. x' = et,x(O) = l. 16. x, = e21 , x(O) = O.

12. x' = sin nt, x(O) = O. 17. x' = 10g(1 + t), x(O) = l.

13. x' = I/(t + 5), x(O) = O. 18. x' = .JlTt, x(O) = l.

14. x' = e- t cos nt, x(O) = l. 19. x' = Jl+t3, x(O) = O.

15. x' = tan t, x(O) = 3. 20. x' = et2 , x(O) = O.

 2.2 SEPARABLE EQUATIONS 91

21. (Reservoir Volume) Using the flow-rate data for q in Table 2, Section 2.1, produce a graph of Von

a :<S t :<S 0.6. Use the trapezoidal rule applied to theformula V = Vo + fat q(v)dv with Vo = 1.13 x 105
cubic feet.
22. (Conical Column) Consider a conical aluminum column with radii rO = 1.5 and q = 2.8 at the ends
and length 1: = 140, all units inches. Find the total compression for a uniform top load of 2000 lb.
23. (Cylindrical Column) Consider a cylindrical steel column of radius r = 1.2 and length 1: = 60, all
units inches. Find the total compression for a uniform top load of 2000 lb.
24. (Flask Cooling) A flask of water is heated to 95 degrees Celsius and then removed from the heat and
stirred constantly while it cools in a room with air temperature 22°e. Find the differential equation for
the temperature u(t) at time t and predict the cooling time to 23°C, given that the flask cools to 90°C
in 5 minutes.
25. (Thermometer) An outdoorthermometer brought indoors at 7:00 in the morning reads 12°F. Afterthree
minutes, it reads 17°F. Assuming that the room is at 70°F, at what time will the thermometer read 69°F?
26. (Torricelli's Law) Derive from Newton's laws the formula v = y'2gh for the velocity v of a body
undergoing free fall from rest, through height h, neglecting air resistance. Assuming that the falling
body is a small cell of fluid, give an argument for the plausibility of Torricelli's law.
27. (Torricelli's Law) Let T1 be the time it takes to empty half of a cylindrical tank. Let T2 be the time it
takes to completely empty the tank. Explain phYSically why T2 = 2 T1 is false. Determine both T1 and
T2 for a cylindrical tank of radius 10 feet, filled to a depth of 20 feet, that drains from a circular valve
at the bottom of radius 0.25 ft.
28. (Natural Reservoir Surface Area) Let A(x) be the surface area at depth measurement x(t) and time t,
q(t) the flow rate and Vet) the reservoir volume. Show by the chain rule that q(t) = V' (t) = A(x(1) )x' (I)
and derive the formula A(x(t» = q(t)/x' (t).
29. (Draining a Natural Reservoir) Verify the drain time of 19.43 hours in Table 2, Section 2.1, using
Simpsons rule (12) in Section 1.3 and the formula

1 [135 A(u)du
t = 7200n 135 --;J72'

30. (Draining a Spheroidal Tank) A water storage tank has the shape of an oblate spheroid with major
axis a feet and minor axis b feet (so a < b < a). An equation of the tank surface is

r2 (x - b)2
a2 + -b-2- = 1,

where r and x are cylindrical coordinates with origin at the bottom of the tank; see Figure 5. Establish
these results:
(a) The radius of the circular cross-section at height x is

r=a/l-(x-b)2/b2,0:<sx:<s2b.

(b) The area of the cross-section is A(x) = :lIT 2 = na2 (l - (x - b)2/b 2 ).

A
(c) Let g = 32 ftlsec 2 , c/ = n sq ft, a = 50 ft and b = 25 ft. The time required for a full tank to
drain is 39.28 minutes.
31. (Reservoir Loss by Evaporation) Open reservoirs lose water by evaporation at a rate proportional to
their surface area. Establish these results:
92 2 ORDINARY DIFFERENTIAL EQUATIONS OF FIRST ORDER

(0,2b)
(r,x)

(0, b)
a

(0,0) FIGURE S. Spheroidal tank

(a) Loss due to evaporation alone satisfies dV/dt == -JLA(x), where JL > 0 is a constant.
(b) Combine the differential equation of part (a) with the equation dV/dt == A(x)dxldt to solve for
x(t) == x(O) - JLt.
(c) Determine the evaporation time for a cylindrical reservoir of radius 10 feet, filled to a depth of
20 feet, assuming JL == 0/300,000) ftlsee.

32. (Reservoir with Draining and Evaporation) Consider a reservoir in which both draining and
evaporation occur. Establish these results:
dV
(a) The rate equation for V is dt == -av - JLA(x), for some constant JL > o.
dx cx l/2
(b) The water depth x satisfies - == - - - ) - JL, where c == afig.
dt A(x
(c) Determine the time to empty a cylindrical reservoir of radius 10 feet, filled to a depth of 20 feet,
that drains from a circular valve at the bottom of radius 0.25 ft, assuming an evaporation constant of
JL == (1/30,000) ftlsee.

33. (Special Relativity) Let c be the speed of light and rna the rest mass of a particle traveling at velocity
v in a field of constant gravitation g. Einstein's special theory of relativity predicts a variable mass rn
according to the formula
ma
m == -v't=l==-==(=v/==c)=:;'2

Use this formula to solve the initial value problem

d(rnv)
----;It == mg, v(O) == 0,

and show that v == c tanh(gtlc).

2.3 Linear First-Order Equations

The first-order equation (1) of Section 2.1 is said to be linear if F(t, x) is a linear function
of x: F(t, x) = A(t)x + B(t). Otherwise, it is said to be nonlinear. This section presents
a method for solving the general linear first-order equation
dx
(1) - = A(t)x + B(t)
dt
 2.3 LINEAR FIRST-ORDER EQUATIONS 93

by direct integration. The method is illustrated by an application to an electrical circuit.

Additional applications are developed in the Exercises.
If the coefficients A(t) and B(t) in equation (1) are both constant then the equation is
separable and can be integrated by the method of Section 2.2. Hence we shall assume that
at least one of the coefficients is not constant. Functions A(t) and B(t) will be assumed
to have a common domain of definition at < t < a2. It follows that the graph of any
solution x = </J(t) of (1) must lie in the vertical strip illustrated in Figure 6. We shall see
that exactly one solution curve of equation (1) passes through each point of the strip.
Equation (1) with B(t) = 0 reduces to

dx
(2) - = A(t)x.
dt
It is called the homogeneous linear equation associated with equation (1). It is separable
and the method of Section 2.2 provides the solution (11) of Section 2.2. Moreover,
comparison of (2) and (1) in Section 2.2 gives F(t) = A(t), G(x) = x and (9) of Section
2.2 becomes

(3) j(x) = lXoX~

- = logx -
~
logxo = log -
Xo
(X) .

Note that y = j(x) if and only if cY = x/xo, and by inversion, x = rt

(y) = xocY . Hence
the solution curve of (2) that passes through the point (to,xo), as defined by (11) in
Section 2.2, is defined by

(4) let) = /t
to
A( r)dr.

The general linear equation (1), with B(t) "I 0, is called a nonhomogeneous linear
equation. We shall now show how a particular solution of (1) can always be found by
Lagrange's method of variation of parameters. The idea of the method is to seek a

x(1)
Xo

FIGURE 6. Domain of definition for equation (1)

94 2 ORDINARY DIFFERENTIAL EQUATIONS OF FIRST ORDER

particular solution of the form

(5) x = ¢p(t) = c(t)cl(t) ,

where let) is defined as in (4) and the function c(t) is to be determined. On differentiating
(5) and noting that dI/dt = A, we find that

dx dc
(6) - = _cI(t) + c(t)AcI(t)
dt dt '
while

(7) Ax + B = B + cAd(t) .

Thus (5) is a solution of (l) if and only if c(t) satisfies

dc = c-l(t)B(t).
-
(8)
dt
Hence a particular solution of (1) is defined by

(9) x = ¢p(t) = cl(t) t c-I(r) B( r)dr.

1to
Note that ¢pCto) = O.
Finally, let x = ¢(t) denote any solution of (1) and consider the difference X =
¢(t) - ¢p(t). A simple calculation shows that dXldt = A(t)X, and hence X must have the
form (4): X = ¢H(t) = X(tO)eI(t) Thus every solution of (1) can be written in the form
(10) x = ¢H(t) + ¢p(t) = cI(t) (xo + i t c-I(r)B(r)dr) .

Clearly, (10) defines a solution of (1) that passes through the point (to,xo). The preceding
derivation shows that equation (1) has a unique solution that passes through a given
point (to,xo) of its domain of definition. Moreover, from (8) we see that the solution is
defined on the entire interval al < t < a2 in which A(t) and B(t) are defined.

EXAMPLE 1. An LR Circuit. Consider the simple loop circuit of Figure 7. The inductance
L and the resistance R are assumed to be positive constants, while the applied voltage
E(t) may be any function of t.
If i = ¢(t) is the electric current in the loop at time t, then Kirchhoff's laws give the
linear first order equation

di
(11) L dt + Ri = E(t).
 2.3 LINEAR FIRST-ORDER EQUATIONS 95

0v
E(t)

R
FIGURE 7. An LR circuit

Note that this may be written in the form (1) with x replaced by i, A(t) = -R/L =
constant and B(t) = E(t)!L. Hence if we choose to = 0 then (4) and (10) give

(12) l(t)
r (-R)
= 10 L dr =L -Rt
and

(13)

where io is the current at time t = O.

AC Applied Voltage. Let the applied voltage be the alternating current defined by
(14) E(t) = Eo sin wt,

where the amplitude Eo and the frequency ware positive constants. Then a simple

lit
integration gives

E ~tfL
(15) - 2 0 22(Rsinwt-wLcoswt)
eRT1LE(r)dr=
L o R +w L
EowL
+ R2 + w 2 L2
Substituting this expression into (13) gives the solution

(16) ._
1-
-RtfL (.
e 10
+ 2
EowL )
R +w 2L2
Eo
+ 2 2 2 (Rsinwt - WLC05Wt).
R +w L
Let ¢ denote the unique angle, measured in radians, such that 0 < ¢ < nl2 and

(17)
96 2 ORDINARY DIFFERENTIAL EQUATIONS OF FIRST ORDER

Then (16) and the identity sin(a- b) = sin a cos b - cosa sin b imply that the response
of the L, R circuit to the AC voltage (14) is given by

(18) i = e- RUL ( io + 2EOWL)

2 2 + Eo sin(wt - ¢).
R +w L y'R2+W 2[2
The first term on the right in (18) is called a transient current because it is negligibly
small for all sufficiently large values of t. The second term is called the steady-state
current for the circuit with applied voltage (14):
E
(19) isteady-state = 0 sin(wt - ¢).
y'R2 + w2 L2
It is an alternating current with amplitude EoIJR2 + w2L2, frequency wand phase lag
¢ relative to the applied voltage.

Initial Value Problems and Separation of Variables. In the exercises below, solve the given initial value
problems using separation of variables.

1. x' + Aox = 0, x(O) = Xo. 6. x' + (2t)x = 0, x(O) = 10.

2. tx ' + Aox = 0, x(l) = Xo. 7, x' - (3t 2 )x = 0, x(O) = 5.

3. x' + (sint)x = 0, x(O) = 1. 8. x' - (5/t 2 )x = 0, x(l) = l.

4. x' + (l +( 2 )-lx = O,x(O) = 1. 9. x' = (2U(t 2 + 2))x, x(O) = 2.
5. x' + e-tx = 0, x(O) = 1. 10. x' + (2/(t 2 + t))x = 0, x(l) = 1.

In the exercises below, solve the given initial value problem by using basic methods or variation of
parameters.

11. x' = (-4It)x + t4 , x(l) = O.

12. x' = -x + 4 sin t, x(O) = o.

13. x' = (-l/t)x + cost, x(l) = O.
14. x' = (lIt)x + t2 + 3t - 2, x(l) = O.

15. x' = (2tant)x + 1- 2t tan t - 2sect,x(0) = O.

Growth and Decay. The model x' = Ax holds for decomposition of a radioactive substance, population
growth and continuous interest. The constantA is positive for growth problems, negative for decay problems.
16. (Population) A population of size xU) obeys the law x' = Ax, according to Malthus's law. Given a
population of 100 million in 1910 and 250 million in 1990, predict the population size in the year 2001.

17. (Atmospheric Pressure) Assume the model pi = Ap for the pressure p(x) at elevation x feet above sea
=
level. Given p(l7500) p(0)/2, find a formula for p(x).
 2.3 LINEAR FIRST-ORDER EQUATIONS 97

18. (Yeast Culture) The population xCt) in a yeast culture satisfies Xl = Ax and doubles every 16 hours.
In how many hours will the population be 10 times its original size?

19. (Half-Life) A radioactive substance obeys the decay law Xl = Ax with A = -0.0004332. Find the time
t in years when 2x(t) = x(O), called the half-life of the radioactive substance.

20. (Radiocarbon Dating) The idea of radiocarbon dating originates with the work of Nobel laureate
W Libby (Chemistry, 1960). It is based upon absorption of carbon-14, 6C 14, which stops when an animal
dies. The amount x of carbon-14 obeys the decay law x' = Ax and it is known that 2x(5730) = x(O)
(half-life is 5730 years). What percentage of carbon-14 is present in a fossil that is supposed to be 2500
years old?

21. (Loss of Moisture) The drying of a beach towel satisfies the decay law x' = Ax, where x(t) is the amount
of moisture present. If it is half dry in 30 minutes, then when will it be within 1 %of being completely dry?

22. (RC Series Circuit) Use Kirchhoff's laws to derive the linear first -order equation

dq q
(20) R- + - = vet)
dt C
for the RC series circuit of Figure 8, where q is the charge on the capacitor C and vet) is an applied voltage.

23. (RC Series Circuit) Solve the initial value problem for (20) with R = 1, C = 11l0, vet) = 0 and
q(O) = 5.

24. (RC Series Circuit) Solve the initial value problem for (20) with R = 1, C = 11l0, vet) = 20 and
q(O) = o.

25. (RC Parallel Circuit) Solve the initial value problem for (21) with C = 1, R = 11l0, iCt) = 20 sin t
and v(O) = o.

26. (RC Parallel Circuit) Use Kirch hoff's laws to derive the linear first -order equation

dv v
(21) C- +- = i(t)
dt R
for the RC parallel circuit of Figure 9, where v is the voltage drop from junction P to junction Q and
i(t) is a prescribed current generated by a current source.

27. (Continuous Interest) The investment x(t) after t years compounded n times per year at interest rate r
per annum is given by x(t) = x(O)(l +rln)nt. Continuous interest satisfies the growth equation x' = Ax,
where A = r. If x(O) is one hundred dollars and the rate is 8 percent, then for some n both methods of
interest result in identical amounts, to the nearest cent, for the first 50 years. Is n < 365 or is n > 365?

vet)
~C
R

FIGURE 8. An RC series circuit

98 2 ORDINARY DIFFERENTIAL EQUATIONS OF FIRST ORDER

i(t)

FIGURE 9. An RC parallel circuit

2.4 Exact Equations

Families of curves that are defined by an equation

(1) U(x,y) = c = constant

occur in many applications (isobars, isotherms, streamlines, etc.; see Section 1.6). In this
section the implicit function theorem of Section 1.6 is used to construct a differential
equation
dy
(2) dx = F(x,y)

for which the solutions are precisely the curves (1). Then the construction is reversed to
construct the solutions of (2) as a family (1).
Recall that if u, U x and uy are continuous and if uy 1= 0, then the implicit function
theorem implies that for each fixed constant c (in the range of u), equation (1) has a
solution

(3) y = j(x),

which is defined and has a continuous derivativej'(x) on a suitable interval a < x < b.
Of course, j(x) also depends on c, but this will be suppressed in the notation. The
definition of j(x) implies that

(4) u(x,J(x)) = c for a < x < b.

Consequently, one has

d
(5) dx u(x,J(x)) = 0 for a < x < b.
 2.4 EXACT EQUATIONS 99

The chain rule of calculus implies that this can be rewritten as

(6) uxCx,f(x» + uyCx,f(x»f'(x) = 0 for a < x < b.

But this equation means that y = jex) satisfies the differential equation
dy
(7) uxex,y) + uyex,y) dx = 0 for a < x < b.

This is equivalent to equation (2) with

(8) Fex,y) = -uxex,y)/uyCx,y).

EXAMPLE 1. If u(x, y) = y/x then (1) defines the family of straight lines through the origin

(9) ~- c
x
In this case Ux = _ylx 2 , uy = l/x and hence the family (9) satisfies

-y 1 dy
(10) -+--=0
x2 X dx '
or
dy y
(ll)
dx x
The procedure described above can often be reversed to obtain the solutions of a
differential equation (2) as a family (1) For if a function u(x,y) can be found such that
(8) holds, then (2) can be rewritten as (7), which clearly implies (1).

To discover how a function u(x,y) that satisfies (8) may be found when F(x,y) is
given, it is convenient to rewrite (2) as
dy
(12) - F(x,y) + dx = O.

Notice that an equivalent differential equation

dy
(13) M(x,y) + N(x,y) dx = 0

is obtained by multiplying (12) by a nonzero function Nex,y) and writing M(x,y) =

-N(x,y) F(x,y).
Definition. Equation (13) is said to be an exact equation if and only if there exists a
function uex,y) that satisfies

(14) Mex,y) = uxex,y) and Nex,y) = uyCx,y).

100 2 ORDINARY DIFFERENTIAL EQUATIONS OF FIRST ORDER

Note that an exact equation can be solved immediately: Its solutions are defined implicitly
by u(x,y) = c. Numerical values of the solutions can be constructed by the methods of
Chapter 1, as illustrated in Section 1.6.

EXAMPLE 2. Equation (10) is exact with u(x,y) = y/x, while the equivalent equation
dy
(15) -y+x- =0
dx
is not exact. Still another equivalent equation is

(16) - -1 + --
1 dy
= O.
x ydx
It is exact and has the integral u(x,y) = log(y/x).

A Criterion for Exactness. Clearly, it is desirable to have a criterion for when an

equation of the form (13) is exact and a procedure for calculating the corresponding
integral u(x, y). A clue to the right condition is the calculus theorem that states that if u,
Ux , u y , Uxy and u yx are continuous then

(17) Uxy = u yx .

Combining this with (14) gives

(18)

More precisely, we have the following theorem.

Exactness Theorem. Let M(x,y) and N(x,y) have continuous first partial derivatives.
Then if (13) is exact then (18) holds. Conversely, if (18) holds in a simply connected
region D then (13) is exact in D.

Intuitively, a planar region is simply connected if it has no "holes" in it. A precise

definition and a proof of the exactness theorem can be found in [D-S].

EXAMPLE 3. The equation

(19)

has

(20) M(x,y) = 2xy, N(x,y) = x 2 + 3l.

Thus,

(21)
 2.4 EXACT EQUATIONS 101

and (19) is exact (for allx,y).

Solution Procedure for Exact Equations. Example 3 will be used to illustrate the
procedure. We must construct u(x,y) such that

(22) ux(x,y) = M(x,y) = 2xy,

(23) uy(x,y) = N(x,y) = x 2 + 3/.

Integration of (22) with respect to x, treating y as a constant, gives

(24)

where h(y) is an arbitrary function of y. This result must be reconciled with (23). Equation
(24) implies that

(25) uyCx,y) = x 2 + h'(y).

This coincides with (23) if and only if

(26) h'Cy) = 3/ or hCy) = l.

Thus, a solution of (2), (23) is

(27)

and the solutions of the differential equation (19) are defined implicitly by the equation

(28)

where c is an arbitrary constant.

The procedure illustrated above will work for any exact equation for which the
integration can be carried out. One integrates Ux = M "partially" with respect to x, treating
y as a constant. The result is then substituted in u y = N to determine the integration
"constant" hCy). The consistency of the process is guaranteed by the exactness condition
My = N x . Alternatively, one may also integrate uy = N with respect to y and then
substitute into the equation Ux = M.
Integrating Factors. If p, (x, y) is any nonzero function then equation (13) is equivalent
to the equation
dy
(29) p,(x,y)M(x,y) + p,(x,y)N(x,y) dx = O.

If (13) is not exact then it may be possible to choose p,(x, y) such that (29) is exact. Such
a function p,(x,y) is called an integrating factor.
102 2 ORDINARY DIFFERENTIAL EQUATIONS OF FIRST ORDER

EXAMPLE 4. Equation (15) is not exact, but multiplication by /-L(x,y) = (xy)-l yields
the exact equation (16). Note that another integrating factor for (15) is /-L (x, y) = l/x2,
which yields the exact equation (10).

The exactness criterion for (29) is (/-LM)y = (/-LN\, or

(30)

or

(31) /-LxN _ /-LYM = My - N x .

/-L /-L
In particular, if (13) is exact then /-L(x,y) = 1 satisfies (31). Of course, (31) is a partial
differential equation. In general, it is at least as difficult to solve as (13). However, there
are certain special cases of (31) that are solvable.
Integrating Factors /-L(x). If (13) has an integrating factor /-L = /-L(x) that is independent
of y then (31) becomes

(32) -/-Lx = -1 (My - N x) == hex)

/-L N
and (13) has an integrating factor

(33)

Clearly, the criterion for this case is that (My - Nx)/N should be a function of x alone.

EXAMPLE 5. Equation (15) has M = -y, N = x. Since My - N x = -2, it is not exact.

However, (My - Nx)/N = -21x = hex) is a function of x alone and (33) gives the
integrating factor

(34)

The corresponding exact equation is (10).

Integrating Factors /-L(y). A similar argument shows that there is an integrating factor
/-L = /-L (y) if and only if

(35)

In this case

(36) /-L(y) = ef h(y) dy

is an integrating factor.
 2.4 EXACT EQUATIONS 103

EXAMPLE 6. Consider the equation

(37)

Here M(x,y) = xy and N(x,y) = x 2 + l + y, so

1 1
(38) - -(My - Nx ) = - == key).
M y
Equation (36) gives M(y) = Y and

(39)

is exact. The solution procedure for exact equations then yields the implicit solution

(40)

EXAMPLE 7. Streamlines in Fluid Flow. The steady planar flow of an incompressible

nonviscous fluid is described by a velocity field

(41)

where <p(x, y) is a scalar field called the velocity potential. Moreover, the incompress-
ibility condition implies that

(42)

The streamlines of a velocity field (41) are the curves in the (x,y)-plane that would
be followed by a particle that moved with the flow velocity V. Thus they satisfy
dx dy
(43) dt = <Px(x,y), dt = <p/x,y),

where t denotes time. Elimination of t gives the differential equation dy/dx = <Pyi<Px, or
dy
(44) - <Py + <Px dx = O.

The integral curves of this equation are precisely the streamlines of the flow (41). Note
that this has the standard form (13) with

(45) M = -<Py, N = <Px.

Moreover, My-Nx = -<Pyy -<Pxx = 0, by (42). Thus the streamlines of an incompressible
flow (41) are the solution curves 1/J(x,y) = c of the exact equation (44).
104 2 ORDINARY DIFFERENTIAL EQUATIONS OF FIRST ORDER

Flow Past a Circular Cylinder. The planar flow past a circular cylinder has velocity
potential

(46) ¢(x,y) = V (r + :2) cose,

e
where r, are polar coordinates, a is the cylinder radius and V is the fluid speed at large
distances from the cylinder. Here we derive the stream function 1fr(x,y) of Example 2,
Section 1.6.
The potential (46) may be written

(47) ¢(x,y) = V x ( + -2--2

X
a2x) .
+y
Thus

(48)

and

which has the standard form Mdx + Ndy = 0 with

-2a 2xy
(51) M(x,y)=V(2 2)2'
x +y
a2(y2 _X2»)
N(x,y) = V (1 + (x 2 + y2)2 .

Moreover, (50) is exact. Hence the streamlines are the level curves

(52) 1fr(x,y) = c,

where the stream function 1fr(x,y) satisfies 1frx = M,1fry = N. Integration of the first
equation gives

(53) 1fr(x,y) = V J (x 2
2a2xy
+ y2)2 dx + hey)
a2y
= - V-2- -2 + hey).
x +y
 2.4 EXACT EQUATIONS 105

This also satisfies the second equation if

a 2 (y2 _ X 2 ) ,
(54) 1/ryCX,y) = V (2
x +y
2)2 + h (y)
a2(y2 _ X2)
= V (2
x +y
2)2 + V.
Thus h'(y) = V and we may take hey) = Vy. Hence the stream function is given by

(55) 1/r(x,y) = V (y - a2y ) .

-2--2
X +y
Note that this agrees with equation (6) of Section 1.6.

Level Set Equations. Verify that the equation is exact and find the level set equation u(x, y) = c.

1. y+xy' = 0. 6. eY + (xeY + 2y)y' = 0.

2. 2xy + (l +x2)y' = O. 7. (3x 2y + Sxy2) + (x 3 + Sx 2y + 12/)y' = 0.

xy + 1 x - 2y ,
3. (y + eX) + xy' = 0. 8. - - = - - 2 - Y .
Y y
4. cos(4/) = Sxysin(4y 2)y'. 9. siny + (xcosy)y' = 0.

5 --=--y
y x, 4/ - 2x2 S/ - x 2 ,
. x2 + y2 x2 + y2 . 10. 2
4xy-x
3 + 4y-xy
3 2 Y = O.

Integrating Factors. Find an integrating factor of the indicated form, determine the new exact equation
and find the level set equation u(x, y) = c.

11. x - y/l - x 2y' = 0, fJ.. = fJ..(x).

12. y + (x 2y - x)y' = 0, fJ.. = fJ..(x)
13. y + (3 + 3x - y)y' = 0, fJ.. = fJ..(y).

14. 2xy + (/_x 2 )y' = 0, fJ.. = fJ..(y).

15. ( / +xy + 1) + (x 2 +xy+ 1)y' = 0, fJ.. = fJ..(xy).

Initial Value Problems. Use the theory of exact equations to solve the given initial value problem.
16. y+xy' = O,yO) = 2.
17. 2xy + (l +x2 )y' = O,y(O) = 1.

18. (x + siny) + (x cosy - 2y)y' = 0, y(2) = Jr.

19. (ysinx+xycosx) + (xsinx+ 1)y' = O,Y(JrI2) = 3.
20. eY + (xeY - 1)y' = 0, y(5) = °
21. (Fluid Flow) Show that equation (50) is exact, and then derive equation (54) from the exactness
criterion and differential equation (50).
106 2 ORDINARY DIFFERENTIAL EQUATIONS OF FIRST ORDER

22. (Separable Equations) Show that x'/g(x) = J(t), a completely separated equation, is also exact.
23. (Linear Equations) Show that a linear differential equation x' = a(t)x + bet) can be multiplied by a
suitable function JL(t) to make it exact.
24. (Integrating Factors) Assume that JLMdx + JLNdy = 0 is exact and assume that an integrating factor
JL(x,y) is a product of powers xnym Find conditions on nand m involving the functions M and N.

25. (Streamlines) Let u(x,y) = V (y - 2a

x +y
2y
2)' a = 1 and V = l. Verify that u(2, 1) = 0.8, uy(2, 1) =

22/25, ux (2, 1) = 4/25. Explain what this has to do with solving for x in terms of y and y in terms of
x near the point (2, 1), using the implicit function theorem.
26. (Implicit Functions) Assume that u(x,y) is continuously differentiable and u(XO,yo) = c. Give two
conditions for solving for one variable in terms of the other, based upon the implicit function theorem.

2.S Applications to Some Second-Order Equations

Second-order ordinary differential equations are equations of the form

(1) d2x (
dt 2 = F t,x, dt
dx)
where the unknown x is a function x = </J(t) of the independent variable t. Such
equations, together with the first-order equations (1) of Section 2.2, occur as models for
a large number of physical, chemical and biological phenomena. In fact, the theory of
such equations is a main theme of this book and many applications are presented here
and in subsequent chapters.
This section treats several special types of second-order equations that can be solved
by integrating two first -order equations. Such equations can be solved by the methods
of this chapter. More general second-order equations, not solvable by these methods, are
treated in Chapter 3.
Second-Order Equations Independent of x. If the variable x does not appear
explicitly in equation (1), so that

(2)

then it is natural to introduce the new unknown

(3)
dx
v = dt'
dv d2x
Clearly, - = - 2 ' and hence v is a solution of the first -order equation
dt dt
dv
(4) dt = Fet, v).
 2.5 ApPLICATIONS TO SOME SECOND-ORDER EQUATIONS 107

Let us assume that a solution v = 1jI(t) of (4) has been constructed by one of the methods
dx
of this chapter. Then (3) can be written as at = 1jI(t) and a second integration produces
a solution of (2) of the form

(5) x = ¢(O = Xo + t 1jI( r)dr.

lto
Two examples of this type will be given.

EXAMPLE 1. Free Fall Without Air Resistance - Plane Earth Model. Consider the
free fall of an object from a height h above the surface of the earth. If h is a small fraction
of the earth's radius then the gravitational field may be treated as a constant of magnitude
g = 32 ftlsec 2 (or 980 cm/sec 2 ), see Figure 10, where the x-axis is directed vertically
downward, with the origin at the earth's surface.
The gravitational force on an object of mass m is F = mg. Hence, if the resistance of
the air is neglected then Newton's laws imply that the position of the object at time t is
a function x = ¢(O that satisfies md2 x1dt 2 = mg, or

d2x
(6) dt 2 = g = constant.

This has the form (2) and v = dxldt is the speed of the object. It satisfies
dv
(7) dt = g,

and hence

(8) v = Vo + gt,
1 2
X = Xo + vot + 2. gt ,

where Vo and Xo are the initial speed and position, respectively. Equations (8) are Galileo's
law of free fall in a vacuum.

x = Xo =-h

F=mg

----1---- x = 0
earth
x FIGURE 10. Free fall of an object- plane Earth model
108 2 ORDINARY DIFFERENTIAL EQUATIONS OF FIRST ORDER

EXAMPLE 2. Free Fall with Air Resistance - Plane Earth Model. An object that falls
freely in a gas will experience a force R that opposes the motion, producing a net force
F = mg -.R. We expect that R = R(v) will be a function of the speed of fall v. Thus
Newtons laws give md 2x1dP = mg - R(dxldt), or

d2x
(9) dt 2 = g - <I> (dxldt) ,

where <I>(v) = R(v)/m. The precise form of the function <I>(v) will depend on the shape
of the falling object and on the gas through which it falls. It may be expected to be a
continuous function of v with the follOwing additional properties.

(10) <I>(v) :::: 0 and <1>(0) = o.

(11)

(12) <I>(VM) = g for some VM > O.

Notice that (9) has the special form (2), and equation (4) for the speed is
dv
(13) dt = g - <I>(v).

Moreover, (12) implies that one solution of (13) is

(14) v = VM = constant, for all t.

For this reason VM is called the terminal speed. If the object has this speed at time t = to
then it will continue in uniform motion

(15)

Equation (3) is separable. Application of equations (8)-00), Section 2.2, with F(t) = 1,
G(v) = g - <I>(v), and initial values x = Xo, v = Vo = 0 when t = 0 gives

(16) J(v) = 1° v dry

g - <I>(ry)
=t

and so

(17)

and

(18)
 2.5 ApPLICATIONS TO SOME SECOND-ORDER EQUATIONS 109

Notice that g - <1>( v) > 0 for 0 :s v < VM by (1) and (12), and hence r 1 exists, since
j' > O. These results will be applied to three special forms of the resistance function <I> (v).

Case 1. (<I>(v) = kv, k > 0). This function clearly has properties (0)-(12), and the
defining equation for the terminal velocity is kVM = g, or VM = glk. Moreover,

(19)
f~ g - kTJ
= -~f~
k TJ - VM
= -~logITJ-VMI.
k
Hence (16) becomes, for 0 :s v < VM,

(20) j(V) =
10r ~
g - kv
= -~ log
k
(1 - ~) VM
= t.

Solving the equation for v gives 1 - V/VM = e- kt , or

(21)

Note that v ---+ VM when t ---+ 00, as expected.

Case 2. (<I>(v) = kv 2, k > 0). This clearly satisfies (10)-02) with kv~ g, or
VM = /ilk. Moreover, a table of integrals gives
(22) f dTJ = _1_ log I VM + TJ I
g - kTJ2 2y'gk VM - TJ '

and hence (6) becomes

(23) j(V) =
10
r g -dTJkTJ2 =_l_IOg(vM+v)=t
2y'gk VM - V '

for 0 :s v < VM. On solving this equation for v, we get (VM + v)I( VM - v) = e2t .Jiik and
hence, after some algebraiC manipulation,

v e2t .Jiik - 1
e2t.Jiik + 1
or

(24) v = VM tanh(t y'gk), for all t::: O.

Again, one has v ---+ VM as t ---+ 00.
Case 3. (<I> ( v) from Experimental Data). More realistic estimates of the function <1>( v)
may be obtained from wind tunnel measurements. The operator of the tunnel can control
the air speed v. The corresponding resistance R( v) can be measured by an analog device,
whose output is recorded. Values of R(v) and <I>(v) = R(v)/m may then be reproduced at
110 2 ORDINARY DIFFERENTIAL EQUATIONS OF FIRST ORDER

will. Table 3 shows values of Vk and ct>k = ct>(Vk) that might have been obtained in this way.
The units are ftlsec for v and ftlsec 2
for ct>(v). Note that VM =
150 ftlsec, approximately.
When ct>(v) is obtained from experimental data then the integralj(v) of (16) must
be computed by numerical integration. The final column in Table 3 gives the values of
j(Vk) = tk that were obtained by computingj(v) by means of Simpson's rule with step
size h = 5. The value t30 = 16.40 is probably quite inaccurate because the integrand
(32 - ct>(V))-l varies rapidly in the interval 145 ::: v ::: 150. More accurate values of
j(Vk) may be obtained by reducing the step size h. Of course, this will necessitate the use
of the experimental values ct>(v) at additional values of v.
r
The velocity v is given as a function of t by equation (17): v = 1 (t). This function
can be computed from Table 3 by inverse interpolation.
Quadratic interpolation, using the algorithm of Section 1.2, was used to compute
the values of Table 4 below. For brevity, t is restricted to the values 0.0 to 3.0 in steps of
0.1. The positions Xk = X(tk) can be obtained from these data by numerical integration.
Second-Order Autonomous Equations. A physical system is said to be au-
tonomous if its governing differential equation does not contain the independent variable

TABLE 3. Sample air resistance data

k v. <l>. t. k Vh <l>. tk

0 0.0 0.00 0.0000 16 80.0 6.19 2.6832

5.0 0.23 0.1568 17 85.0 7.02 2.8825
2 10.0 0.40 0.3147 18 90.0 7.27 3.0821
3 15.0 0.64 0.4734 19 95.0 9.32 3.2923
4 20.0 0.94 0.6337 20 100.0 10.27 3.5201
5 25.0 l.02 0.7950 21 105.0 10.54 3.7502
6 30.0 1.39 0.9570 22 110.0 11.28 3.9879
7 35.0 l.58 1.1213 23 115.0 13.22 4.2384
8 40.0 l.68 1.2855 24 120.0 13.46 4.5132
9 45.0 2.20 l.4519 25 125.0 14.82 4.7837
10 50.0 2.35 l.6204 26 130.0 17.84 5.1088
11 55.0 2.48 1.7892 27 135.0 2l.47 5.5098
12 60.0 3.14 l.9602 28 140.0 23.98 6.0675
13 65.0 3.68 2.1355 29 145.0 27.08 6.8381
14 70.0 4.30 2.3135 30 150.0 3l.81
15 75.0 4.78 2.4962
 2.5 ApPLICATIONS TO SOME SECOND-ORDER EQUATIONS 111

TABLE 4. v = J- 1 (t) computed from Table 3

k tn Vk k t, v,

0 0.0 0.0 16 1.6 49A

0.1 3.2 17 1.7 52A
2 0.2 6A 18 1.8 55.3
3 0.3 9.5 19 1.9 58.2
4 OA 12.7 20 2.0 61.1
0.5 15.8 21 2.1 64.0
6 0.6 19.0 22 2.2 66.8
7 0.7 22.1 23 2.3 69.6
8 0.8 25.2 24 2A 72A
9 0.9 28.2 25 2.5 75.1
10 1.0 31.3 26 2.6 77.8
11 1.1 34A 27 2.7 80A
12 1.2 37A 28 2.8 82.9
13 1.3 40A 29 2.9 85A
14 1.4 43.4 30 3.0 87.9
15 1.5 46A

t explicitly In particular, autonomous second-order equations have the form

(25) ~:~ = F (x, ~) .

If x = ¢(t) is a solution of this equation and if v = dxldt = ¢'(t) for a < t < b then
t = ¢-l (x) exists and v = ¢' (¢-l (x» = 1/I(x) defines v as a function of x. The same is
true if v = ¢' (t) < 0 for a < t < b. Differentiation of v = 1/I(x) with respect to t, using
the chain rule, then gives

dv I dx dv
(26) dt = 1/1 (x) dt = v dx·

On combining (25) and (26) and using d2 x1dt 2 = dvldt, we find that

dv
(27) v dx = F(x, v).
Thus v may be determined as a function of x by integrating the first -order equation (27).
Moreover, when v = 1/I(x) is known then the equation v = dxldt becomes

dx
(28) dt = 1/I(x).
112 2 ORDINARY DIFFERENTIAL EQUATIONS OF FIRST ORDER

This is a separable first-order equation whose solution gives x as a function of t. Several

applications of this method are given in the examples and exercises below.

EXAMPLE 3. Launch of a Space Probe. A space probe is to be launched by a rocket

to explore the neighborhood of Jupiter. We wish to predict the necessary launch speed
and estimate the time needed to reach Jupiter. For simplicity we shall assume that at the
time of rocket burnout, the probe is at a distance Xo from the center of the earth and is
traveling directly away from the earth with speed Vo > 0; see Figure 1l.
Physical intuition suggests that there is a speed VE, called the escape velocity, such
that if Vo < VE, then the probe will fall back to the earth. If Vo > VE, we expect that the
probe will continue to move away from the earth with a speed v that will approach a
limiting value Voo at large distances from the earth. We shall use Newton's laws to verify
this behavior and to compute the speeds VE and Voo and the time needed for the probe
to reach the vicinity of Jupiter.
Let x denote the distance of the probe from the center ofthe earth (Figure 11). Then
the gravitational force on the probe is

(29)

by Newton's inverse square law, where m is the mass of the probe, g is the gravitational
acceleration at the earth's surface and R is the radius of the earth. 2 Approximate numerical
values are g = 32 frlsec 2 and R = 4000 mi. The minus sign is needed because F is directed
toward the center of the earth. Equation (29) and Newton's laws of motion imply that the
position x = ¢(t) of the probe satisfies the differential equation md 2x1dt 2 = -mk/x 2 , or

k
(30)
X2 '

earth
vo
, - - - - + - + -....- X

Xo
FIGURE II. Initial conditions for launch of a space probe

2A more realistic model would include the influence of the sun and Jupiter, but these effects will be ignored here.
 2.5 ApPLICATIONS TO SOME SECOND-ORDER EQUATIONS 113

for all x > R. Equation (30) has the form (25), and the associated first-order equation
(29) is vdvldx = -klx 2 , or

dv
(31)
dx
This equation is separable and the integration technique of Section 2.2 is applicable with
F(x) = l/x 2 and G(v) = -klv. We shall measure time t from the moment of rocket
burnout and assume that the initial values Xo and Vo satisfy

(32) Xo > R, Vo > O.

Then application of equations (8)-01), Section 2.2, gives

(33) lex) = IX ~ = ~ - ~,
I1
Xo ~ Xo x

= -- 1 (2
vo-v 2) .
V
J(v) I1 dl1=-
k Vo 2k
The solution v = 1j;(x) of (31) that satisfies 1j;(xo) = Vo is defined implicitly by the
equationJ(v) = l(x) , which may be written as

(34)

This equation gives the square of the probe's speed at each distance x from the earth. It
is clear that the escape velOCity VE is the speCial value of Vo for which v -+ 0 as x -+ 00.
From (34) we see that v~ = 2k1xo, or

(35)

Normally, Xo will differ from R only by one or two percent. Taking Xo = R in (35) gives
VE = J2gR as the escape velocity from the earth's surface. To obtain a numerical estimate,
take R = 4000 miles and g = 32 ftlsec 2 = 32 x (3600)2/5280 milhr2 These values
give VE = 25067 milhr.
Equation (34) can be written as

(36)

Clearly, if 0 < Vo < VE then v~ - V6 > 0 and v = 0 when 2klx = v~ - V6 > O. In this
case X max = 2k1( v~ - V6) is the maximum distance reached by the probe. On the other
114 2 ORDINARY DIFFERENTIAL EQUATIONS OF FIRST ORDER

hand, if Vo > VE then v2 > V6 - v~ > 0 for all x and v2 ---+ v~ - V6 as x ---+ 00. Thus

(37)

is the limiting value of v for large x. Note that by (36) and (37),

(38) v --~
V~ + v~ > V<Xl> for all x> R.

Time Required to Reach Jupiter. By (38) we have

dx
(39) -dt = v > voo , for all t -> o.
If x] is the distance to Jupiter and t] is the arrival time at distance x], then integration of
(39) over 0 :s t :s t] gives

(40)

Thus
x] -Xo
(41) tmax = - - -
Voo

is an upper bound to the travel time to Jupiter. The minimum distance from the earth
to Jupiter is about 4.2 A.U 3 If we require that tmax = 5 yr = 43800 hr, then we must
have Voo = 4.2 x (93 x lO6)143800 = 8917 milhr = JV6 -
v~. Thus V6 = v~ + v~ =
(25067)2 + (8917)2 or Vo = 26606 milhr.
Solution of the Equation of Motion (30). The solution x = ¢(t) of (30) may be
obtained by integrating the separable first-order equation

(42) dx =
dt
.j2kx + v 2
00

with initial condition x = Xo when t = O. Proceeding as in Section 2.2; we find that

(43)

where

(44)

30ne A.U. (= Astronomical Unit) equals the mean radius of the eanhs orbit around the sun. This distance is
approximately 93 x 106 miles.
 2.5 ApPLICATIONS TO SOME SECOND-ORDER EOUATIONS 115

A table of integrals gives

(45) f .jf+a2
~1/2d~

~+ a2
= J a2~ + e - a2log(J~ + a2 + .j€),
whence

,Ja2x + x2 - fa 2x 0 +x02 + a2log

V
(~+Fo)
.Ja2+x+.jX
(46) t= -------------'----~

The solution x = ¢J(t) may be obtained by solving (46) for x by one of the methods of
Chapter 1. It is interesting that the time tj obtained from (46) with Voo = 8917 milbr,
x = Xj = 4.2 x (93 x 10 6 ) = 3.906 X 108 mi, and Xo = 4000 mi is tj = 4.998 yr, in

excellent agreement with the estimate made above.

Second-Order Linear Equations and Riccati Equations. Equation (1) is said to
be linear if it has the form
d2 x dx
(47) dt 2 = F(t) dt + G(t)x.
Such equations are studied systematically in Chapter 3. Here it will be shown that (47)
can always be solved by solving two first-order equations, one of which is separable.
Let x = ¢J(t) > 0 be a solution of (47) and write
d Idx
(48) u = - log x = - -.
dt x dt

On substituting dxldt = xu and d2x1dt 2 = xduJdt + udxldt = x(duJdt + u 2 ) into (47), it

is found that u satisfies the first-order equation
du
(49) - = _u 2 + F(t)u + G(t).
dt
Note that if u = 1/r(t) is a solution of (49) then (48) with u = 1/r(t) is a separable equation
for x with solution

(50) x = Xo eICtl , let) = t 1/r( r) dr.

l to
Equation (49) is known as a Riccati equation. It is nonlinear and, in general, not
separable. However, we shall show that if a particular solution Uj = ¢Jl (t) has been found
then the remaining solutions can be constructed by integration. Indeed, subtracting

du]
dt = -u1 + F(t)Uj + G(t)
2
(51)
116 2 ORDINARY DIFFERENTIAL EQUATIONS OF FIRST ORDER

from (49) gives

(52)

Thus, if v = U - UI, so that U + UI = V + 2UI = V + 2¢1 (t), then one has

dv 2
(53) dt = -v + (F(t) - 2¢1 (t»v.

Finally, if v = l/w then dvldt = -(1/w 2 )dwldt and

dw
(54) dt = (2¢1 (t)) - F(t))w + 1.
To summarize, if UI = ¢l (t) is a particular solution of (49) and U is any other solution
then we may write

(55)

where w is a solution of the linear equation (54).

Further simplification can be made if two distinct solutions of (49) are knowrl. Note
that (52) can be written as
1 d(u - UI)
(56) -- = -u - UI + F(t).
U - UI dt
If U2 = ¢2 (t) is a third solution of (49) then similarly
1 d(u - U2)
(57) -- = -u - U2 +F(t).
U - U2 dt
Subtracting (57) from (56) gives
1 d(u - UI) 1 d(u - U2)
(58) - -- =U2- UI,
U - UI dt U - U2 dt
or

(59)

Integrating this equation gives

(60) : =:~~~; = c exp ( { [¢2(r) - ¢l (r)] dr) ,

where c is an arbitrary constant. This equation gives the general solution of the Riccati
equation in terms of two particular solutions.
 2.5 ApPLICATIONS TO SOME SECOND-ORDER EQUATIONS 117

EXAMPLE 4. L, R, C Electrical Circuits. The Riccati method will be used to solve

Kirchhoff's equations for the electrical circuit of Figure 12.
Here L, R, and C are the circuit inductance, resistance, and capacitance, respectively,
while

(61) Q = 1/f(t) = charge on capacitor,

i = ¢(t) = current in the circuit,
. dQ
1= di'
Kirchhoff's laws imply that
di
(62) VL = L- = voltage drop across L,
dt
VR = Ri = voltage drop across R,

Vc = cQ = voltage drop across C,

V L + VR + Vc = 0.

It follows that Q and i satisfy the same differential equation:

d2 Q dQ Q
(63) Ldt2 +Rdi+c=O,
d2 i di i
L - +R-
dt 2 dt
+-C =0.
Note that the equation is linear, homogeneous and second order, with constant
coefficients.
Impedance and Admittance. The impedance Z of a circuit element is defined to
be the voltage across the element divided by the current flowing through it. Thus, the
impedance of the inductor L is given by
VL L di
(64) Z= - = --.
i i dt

FIGURE 12. An L, R, C circuit

118 2 ORDINARY DIFFERENTIAL EQUATIONS OF FIRST ORDER

Similarly, the admittance Y of a circuit element is the reciprocal of its impedance:

Y = Z-l. Hence, the admittance of the capacitor C is given by

i C dQ
(65) Y=-=--.
Vc Q dt
The Riccati Method. We will solve the linear constant-coefficient differential equa-
tions of L, R, C circuits by the Riccati method. Essential use is made of the impedance
and admittance equations.
Clearly, Z and Y satisfy Riccati equations. A short calculation, using (62), gives

dZ 1 2 R 1
(66) = - -Z - -Z - -
dt L L c'
dY 1 2 R 1
--Y - - Y - -
dt C L L
These Riccati equations have constant coefficients and may be integrated by separation of
variables. Moreover, after they are solved for Z and Y, one can solve (64) and (65) to get

(67) i = ( exp (~fot Z(T) dT) ,

Q = c exp (~it YCr)dr).
General Solution. We will find the general solution of equation (66). Separating the
variables in the equation for Z gives

(68) -dt= LdZ

Z2 +RZ +UC
The integral of the right-hand side may be found in integral tables. Three cases occur
depending upon whether the discriminant of the denominator,

2 4L
(69) D=R - - ,
C
is positive, zero, or negative.
Case 1 (D > 0). Write 2fJ = y'D/L. Then

(70) f Z2 + LdZ
RZ + UC = -731 tanh -1 ( Z
f3L
R)
+ 2fJL = -t + (1,

where (1 is a constant. Solving this for Z gives

R
(71) Z = -- + f3L log (tanh f3(t - (1)) .
2
 2.5 ApPLICATIONS TO SOME SECOND-ORDER EQUATIONS 119

A second integration gives

(72) f Z(t) dt = -~t + Llog (cosh ,B(t - (1))

and hence, by (67),

(73)

where C2 is a second constant of integration. Note that the identity

(74) cosh(x - y) = cosh x cosh y - sinh x sinh y

implies that

(75) i(t) = e- RtI(2L)(A cosh,Bt + B sinh ,Bt),

,B = JR2 - (4UC)/(2L),

where

(76)

It can be shown that (75) is a solution of (63) for any values of A and B.
Case 2 (D = 0). In this case one has

(77) f L~
Z2+RZ+UC =L
f ~ ~
(Z+Rl2)2 = Z+Rl2 = -t+C1,
whence
R L
(78) Z= --+--.
2 t - C1

A second integration gives

(79) f Z(t) dt = -~t + L log(t - (1)

and hence, by (67),

(80)

where C1 and C2 are arbitrary constants. This can be written as

(81) i(t) = e- RtI(2L)(A + Bt),

where A = -C1C2 and B = C2 are arbitrary. This is a solution of (63) for any values of A
andB.
120 2 ORDINARY DIFFERENTIAL EQUATIONS OF FIRST ORDER

Case 3 CD < 0). Write 2y = yCD/L. Then

(82)

whence
R
(83) Z = -- - YL tan yet - (1)'
2
Integrating again gives

(84) f Z(t) dt = - ~t + L log (cos yet - (1)) ,

(85)

where (1 and (2 are arbitrary. The formula for cos(x - y) implies that

(86) i(t) = e- RU(2L)(A cosyt+B sinyt),

y = J(4UC) - R2)/(2L).

Again, this is a solution of (63) for all values of A and B.

Physical Interpretation. The three cases distinguished above correspond to three
qualitatively different behaviors of the L, R, C circuit of Figure 12, which may be described
as follows.
Case 1. (D > 0). The current i(t) of (75) may be written
1 1
(87) i(t) = -(A + B)e-(R1(2LJ-fi)t + -(A - B)e-(R1(2L)+fJJt.
2 2
Moreover,

(88) ~ - f3 = ~ (R - JR2 - (4UC)) > O.

2L 2L
It follows that i(t) tends steadily to zero as t -+ 00. The circuit is said to be overdamped
in this case. The function i(t) has at most one zero for t > O. It can be shown that i(t)
has a zero t > 0 if and only if either A > 0 and A + B < 0 or else A < 0 and A + B > O.
Case 2. (D = 0). In this case i(t) is given by (81) and i(t) -+ 0 as t -+ 00 (since
RlL > 0). Moreover, i(t) will have at most one maximum or minimum and at most one
zero. It can be shown that a zero occurs if and only if AB < O. The circuit is said to be
critically damped in this case.
Case 3. CD < 0). In this case i(t), given by (86), is oscillatory, with infinitely many zeros
at those times t for which
A
(89) tanyt = --.
B
 2.5 ApPLICATIONS TO SOME SECONO-OROER EQUATIONS 121

y
0.2

-0.2
y = e- O.2x _ 2e- 0 .4x
-0.6
x
-1.0
o 10 20 30 FIGURE 13. Overdamped oscillations

0.8 Y

0.0

Y = (x - l)e- O.4x
x
-1.0
0 10 20 FIGURE 14. Critically damped oscillations

y
0.8
Y = e- O.4x sin(2x)

0.0
x
-0.4
0 10 FIGURE 15. Underdamped oscillations

If R > 0 then the oscillations are damped, with amplitude that tends to zero like e- RU(2L).
The circuit is said to be underdamped in this case. Figures 13, 14 and 15 illustrate the
three cases.

Equations Independent of x. Convert the differential equations below into two equations of the form
v = x' and Vi = F(t, x') and solve for the general solution, containing two arbitrary constants.
122 2 ORDINARY DIFFERENTIAL EQUATIONS OF FIRST ORDER

1. x" = o. 6. x" -x' = 1.

2. x" = 1. 7. x"-x'=t.
3. x" = 1 + t. 8. x" + 2x' = et.
4. x" +x' = O. 9. x" + (l/t)x' = O.

5. x" + 2x' = O. 10. x" - (l/t)x' = t 2 .

Equations Independent of t. Convert the differential equations below into an equation of the form
v(dvldy) = F(y, v), usingy = x(t) and v = x'(t). Solve for the general solution, containing two arbitrary
constants, if possible.

11. x" = O. 16. x" -x = 1.

12. x" = 1. 17. x" + 2x' = O.

13. x" -x = o. 18. x" + 2x = O.

14. x" +x' = O. 19. 2XX" = 1 + (x,)2

15. x" +4x = O. 20. x" = (l + (x,)2 )3/2

21. (Free Fall) Solve Galileo's problem of free fall from rest with air resistance if ¢(v) = ev 2 + kv with
e > 0 and k > O.
22. Oupiter Rendezvous) Verify the implicit solution of the Jupiter problem given by equation (46).

23. (Rockets) The equation of motion for the velocity vet) of a rocket in a constant gravitational field g is

(M - at)v'(t) - ab = -geM - at).

In this equation, M is the initial rocket mass in grams and the rocket loses gas at a grams per second at
constant velocity b centimeters per second relative to the rocket (all constants are positive). Solve the
equation for the position x and velocity v. Hint: yeO) = 0 and x(O) = Re (radius of the earth). The
answers involve the reciprocal and logarithm of M - at.

24. (Cantilever) A uniform cantilever beam of length L and of constant weight w pounds per foot has a
shape curve y(x) known to satisfy the initial value problem

"() w(L - x)2

Ely x = 2 yeo) = /(0) = 0
(See Fig. 16.) The constant product El is the flexural rigidity ofthe beam. The values E and I are Young's
modulus and the moment of inertia of the beam, assumed constant. Solve for the elastic curve y(x).

25. (Riccati Equation) The Riccati equation y' = xy2 - Y + 1 - x has a particular solutionYl == 1. Use the
transformationy = Yl + l/u to develop a formula forthe solution in terms ofthe integral ofx exp(x 2 -x) .

..---x . . '" l-x -

FIGURE 16. Uniform cantilever beam

 2.6 THE INITIAL VALUE PROBLEM 123

26. (Riccati Method) The differential equation x" (t) - x' (t) = 0 can be solved by transformation to a
Riccati equation. Show that the Riccati equation is u' (t) = _u 2 (t) + u(t) with solution u/(l - u) = het .
Use the equation u(t) = x' (t)lx(t) to find the general solution x(t) = q + [2 et .
27. Verify the tk column of Table 3 by applying Simpson's rule to evaluate the integral](v) in (16), using
the data in the Vk and <Pk columns. Take g = 32 and use the trapezoidal rule for tl.
2S. Verify Table 4 by applying inverse quadratic interpolation to the Vk and tk columns of Table 3.

2.6 The Initial Value Problem

The preceding sections have provided many examples of first-order ordinary differential
equations
dx
(1) - = F(t,x).
dt
In each example, the function F(t, x) is a prescribed function of t and x, defined in some
region D in the (t,x)-plane. The initial value problem for such an equation asks for
a solution curve x = x(t), lying in D, that passes through a prescribed "initial point"
(to, xo) in D such that

(2)

It has been shown above that if equation 0) is separable, or linear, or exact then the initial
value problem has a unique solution which can be constructed by direct integration.
Moreover, the integral curves fill the region D and there is exactly one of them through
each point of D. An example was shown in Figure 2 of Section 2.l.
It is natural to ask whether this unique solvability of the initial value problem also
holds for more general equations (1) that are not solvable by the special methods of
Sections 2.2 to 2.4. The purpose of this section is to present simple and general conditions
on F(t, x) that guarantee the unique solvability of the initial value problem. They are
described below.

Fundamental Theorem. Let F(t,x) and its partial derivative aF(t,x)lax be defined and
continuous in the region D. Then the initial value problem (1), (2) has a unique solution
for each initial point (to,xo) in D.

Notice that there are two parts to the fundamental theorem. First, it states that for
each point (to, xo) in D there is a solution of 0) that passes through that point. Second,
it states that that solution curve is unique.
Meaning of the Existence Statement of the Fundamental Theorem. The
existence part of the theorem means that there is a function x = xU), defined on an
124 2 ORDINARY DIFFERENTIAL EQUATIONS OF FIRST ORDER

interval to - a < t < to + b, such that (t, x(t)) is in D and satisfies (1) for each point of
the interval. If lim x(t) = x(to + b) exists and (to + b, x(to + b)) is in D then one can
t-Ho+b
apply the fundamental theorem with (to, xo) replaced by (to + b, x(to + b)) to extend the
interval of existence to a larger one. Thus we may assume that a and b are chosen to be as
large as possible. These values will in general depend on to and Xo, which will be indicated
by the notation a = aCto, xo), b = b(to, xo). The intervalto - aCto, xo) < t < to + b(to, xo)
will be called the maximal interval of existence of the solution through (to,xo).

EXAMPLE 1. The concept of the maximal interval of existence of a solution may be

illustrated by the simple separable equation
dx
(3) - = 1 +x 2 .
dt
The region D for this equation is the whole (t,x)-plane. The solution through (to,xo)
obtained by the method of separable equations is

(4) tan- 1 t = t - to + tan- 1 Xo,

where tan- 1 x denotes the principal value of the inverse tangent function, defined by
-rr/2 < tan- 1 x < rr/2. Clearly, the solution of the initial value problem for (3) is

(5) x = tan(t - to + tan- 1 xo).

Moreover, since the function tan(t) has singularities at t = ±rr/2, it is clear that for this
example,

(6)

Notice that if Xo is very large and positive then b(to, xo) will be very close to 0, while if Xo
is very large and negative then aCto, xo) will be very close to O. This example demonstrates
that the size of the maximal interval of existence will not be obvious from the differential
equation alone.

Meaning of the Uniqueness Statement of the Fundamental Theorem. The

uniqueness statement of the fundamental theorem means that there is exactly one solution
curve of (1) through each point of D. In particular, solution curves cannot branch. This
may no longer be true if the hypotheses of the fundamental theorem are not satisfied.
This is illustrated by the following example.

EXAMPLE 2. Consider the equation

dx
_ =
(7) 3X2/3
dt '
 2.6 THE INITIAL VALUE PROBLEM 125

whose region D is the whole (t,x)-plane. Through the point (to,xo) = (0,0) there are
the distinct solutions x(t) = 0 and also x(t) = t 3. Thus solutions of the initial value
problem are not unique for this equation. It is easy to see that the fundamental theorem
does not apply to equation (7) in any region D that includes points of the t-axis, because
F(t, x) = 3x2/3 and FxCt, x) = 2x- 1I3 is not continuous for x = O. In particular, solutions
can branch from (to,xo) when aFlax fails to be continuous in D.

Geometrical Meaning of the Fundamental Theorem. The preceding discussion

shows that if F(t,x) and Fx(t,x) are continuous in D, then the integral curves of (1)
provide a simple covering of D. This means that there is one and only one integral curve
of (1) through each point of D.
The proof of the fundamental theorem will not be given here. It requires techniques
of advanced calculus. proofs may be found in most advanced textbooks on differential
equations; see, for example, [B-RJ.

EXAMPLE 3. Notice that the fundamental theorem gives no indication of how the solutions
of 0) may be computed. Consider, for example, the equation

(8) dx = /1 _ t2 _ x2
dt Y .

Here the function

(9) F(t,x) = J1 - t 2 - x2

is defined in the unit disk D : t 2 + x2 < 1 and one has

(10)
aF(t, x) -x
ax

Equation (8) is neither separable nor linear nor exact and it cannot be solved by
elementary methods. Nevertheless, F and aFlax are continuous in D and hence the
fundamental theorem guarantees that there is a unique solution of (8) for each (to, xo)
in D. Numerical methods for computing the values of that solution are given in Section
2.7 below.

The fundamental theorem suggests a new point of view. We may define a function
x = x(t) by means of a differential equation (1) and an initial condition (2). In Section
2.7, accurate numerical algorithms to compute the values of such functions are presented.
The algorithms work directly from the differential equation and the initial condition.
Such methods frequently offer the most efficient way to generate and compute a function.
Differentiability of Solutions. It is shown in elementary calculus that every differen-
tiable function is continuous. Hence, solutions of a differential equation (1) are necessarily
126 2 ORDINARY DIFFERENTIAL EQUATIONS OF FIRST ORDER

continuous. Moreover, if x = x(t) is a solution of (1) for a < t < b then

(ll) x'(t) = F(t,x(t)) for a < t < b.

In particular, if F(t, x) is a continuous function of t and x then (11) implies that x' (t) is
also continuous for a < t < b. Continuous functions x(t) such that x'(t) is continuous
are said to be continuously differentiable. Solutions may have higher-order derivatives.
In particular, if F(t, x) and its first partial derivatives Ft(t, x) and FxCt, x) are continuous
then (] 1) implies that x"(t) exists and satisfies

(12) x"(t) = Ft(t,x(t)) + Fx(t,x(t»x'(t) for a< t < b.

It follows that x(t), x' (t) and x" (t) are continuous. We say that x(t) is twice continuously
differentiable. Proceeding in this way, we may prove the following.

Differentiability Theorem. Let F(t, x) have continuous partial derivatives of orders less
than or equal to k. Then every solution x = xCi) of (1) has continuous derivatives of orders
less than and equal to k + 1. In particular, if F(t, x) has continuous partial derivatives of
all orders then every solution of (1) has continuous derivatives of all orders.

In Section 2. 7 these results will be used to derive algorithms for the numerical
solution of first-order differential equations.

1. (Uniqueness) Consider the initial value problem

X' =~, x(O) =0.

Determine the natural domain D of the fundamental theorem. Show that this problem has a unique
solution, defined for -00 < t < 00, which is given by

t :s -n/2,
1
-1 for
x(t) = ~in t for -n/2 :s t :s n/2,
for t ~ n/2.

2. (Nonuniqueness) Show that the initial value problem

X' =~, x(O) = -1,

has infinitely many solutions given by the one-parameter family

:s a,
xU) = 1 -1
~in(t - a - nl2)
for
for
for
-00 < t
a:st:sa+n,
t ~ a+n,

with parameter a ~ O. Explain why this example does not violate the existence-uniqueness theory
 2.7 NUMERICAL METHODS FOR THE INITIAL VALUE PROBLEM 127

3. (Nonuniqueness) Show that the initial value problem

x' = 3x 2/ 3 , xCO) = 0,

I°
has an infinite two-parameter family of solutions, defined by

u- a)3 for t ~ a,
x(t) = for a ~ t ~ b,
(t - b)3 for t ?:. b.

where -00 ~ a ~ 0 ~ b ~ +00

4. (Maximal Interval of Existence) Consider the initial value problem

x' = x2, xCO) = xo > O.

Determine the separation of variables solution and use it to show that the maximal interval of existence
is -00 < t < l/xo.

5. (Differentiability of Solutions) Show that if F(t,x), FtCt,x) and Fx(t, x) are continuous in a region D of
the (t, xl-plane and if x(t) is a solution of x' = F(t,x) in D, then xU) has a continuous second derivative
that is given by

x" = Ft(t,x) + Fx(t,x)F(t,x).

6. (Differentiability of Solutions) Show that if F(t,x) and its partial derivatives of orders one and two are
continuous in a region D and if xCt) is a solution of x' = F(t, x) in D, then xCI) has a continuous third
derivative that is given by

2.7 Numerical Methods for the Initial Value Problem

The fundamental theorem of Section 2.6 guarantees the unique solvability of the initial
value problem
dx
(1) dt = F(t,x), x(to) = xo,
but gives no information about how the solution X = x(t) may be computed. The purpose
of this section is to present several numerical algorithms that can generate accurate values
of x(t) directly from the differential equation and the initial value.
Direction Fields. The solution curve x = xCi) of (1) will pass through any of its
points (tl ,Xl) with the slope F(tl, Xl). Thus a short line segment centered on (tl, Xl) and
having slope FCtl, Xl) will closely approximate the integral curve through (tl, Xl)' Such
a segment will be called a line element of (1). A graph that depicts the line elements
through the points of a grid in the (t, x)-plane is called a direction field for (1). An
example is shown in Figure 17.
128 2 ORDINARY DIFFERENTIAL EQUATIONS OF FIRST ORDER

////////// //
jjjjjjjjjljjj
Illlllllllll
jjjj~jj!jj~~~
1llllll!lllll
I111I1 I11111
///// ///////

~~~~~~~~~~~~~ FIGURE 17. Direction field for the equation x' = xCl - x)

Note that a carefully prepared direction field gives an excellent qualitative description
of the solution curves of a differential equation. Computer programs that generate
direction fields are widely available.
The line elements for (1) were defined geometrically above, and it is easy to write
an equation for them. The line element through (to,xo) is a straight line segment with
slope F(to, xo). Thus it is given by

(2) x = Xo + F(to,xo)(t - to), It - tol :s h,

where h is a small parameter that controls the length of the line element.
The Initial Value Problem on a Prescribed Interval. The numerical algOrithms
of this section are designed to generate the values of the solution x = x(t) of (1), to a
prescribed accuracy, on a prescribed interval to :s t :s to + T. Of course, this interval must
lie within the maximal interval of existence of problem (1). If this condition is violated,
the algorithms may appear to work, but they will yield meaningless numbers.
Finite Difference Methods. The methods described below are finite difference
methods. A discrete mesh of t-values

(3)

is introduced and approximate values Xj ~ x(tj ) are computed at these points only: Values
x(t) at other t-values may then be computed by the interpolation methods of Chapter 1.
It will be shown that x(t) can be computed on the interval [to, to + TJ to prescribed
accuracy by finite difference methods based on a suitably fine mesh.
Euler's Method. The simplest way to generate an approximate solution of (1) is to
follow the line element (2) from (to,xo) to (tl,XI), where Xl = Xo + F(tO,XO)(t1 - to).
The process may then be repeated Lo get (t2,X2), etc. The process is implemented by the
following algorithm:

(4)
 2.7 NUMERICAL METHODS FOR THE INITIAL VALUE PROBLEM 129

TABLE 5. Euler's method for x' = x(l - x), x(O) = 0.5

N 4 64 256 4096
0.73684 0.73143 073115 0.73106
-000578 -0.00037 -0.00009 0.00000

The process is recursive: Each step provides the data for the next. The approximate
solution is a polygonal curve with vertices (to,xo), (tl,Xl), ... , (tN,XN) and sides that
are line elements. Algorithm (4) was proposed by L. Euler in 1768 and is usually called
Euler's method.
The case of a uniform mesh (3) is defined by tl - to = t2 - tl = ... = tN - tN-l = h,
which gives tk = to + kh, k = 1,2, ... , N. In particular, tN = to + Nh = to + T, so that

(5) Nh= T,

and the algorithm becomes

(6) to, Xo, T, N given, h = TIN,

then for k = 0, 1, ... , N - 1,
tk+l = tk + h, Xk+l = Xk + hF(tk,Xk)·
EXAMPLE 1. Euler's method will be illustrated by its application to the problem
dx
(7) - =x(l-x) X(O) = 0.5.
dt '
The differential equation is called the logistic equation. The exact solution to this
problem can be found by the method of separation of variables. It is defined for all values
of t and is given by

(8)
Application of Eulers method on the interval 0 .:s t .:s 1 with various values of N
gives the results in Table 5, where only the value XN ~ x(I) is recorded. 4 Note that
xCI) = 0.73106, to five decimal places. Hence, Euler's method gives the result to the
same accuracy if the step size h = liN is suffiCiently small.

4The third row of the table gives the error EN = x(l) - XN.
130 2 ORDINARY DIFFERENTIAL EQUATIONS OF FIRST ORDER

Error Analysis for Euler's Method. The differences

(9)

measure the error produced by Eulers method at the mesh points tk. How small must
h be to guarantee that IEk(h)1 is less than a prescribed tolerance? This question will be
answered by means of a careful estimate of IEk(h)l.
Taylors theorem with remainder implies that

(10)

where tk ~ ~k ~ tk+l. Moreover, since x(t) is a solution of (1), we have X'(tk)

F(tk,X(tk)) and hence

(11)

On subtracting the Eulers method equation

(12)

from (11) and using the definition (9), we get the relation

(13) Ek+l(h) = Ek(h) + h(F(tk,X(tk») - F(tk,Xk» + ~h2X"(~k).

This gives the error at step number k + 1 as the error at step number k plus two additional
contributions. These may be estimated in terms of bounds for F, F t and Fx. Thus if

(14) IFx(t, x) I ~ L for to ~ t ~ to +T

for all relevant values of x then the mean value theorem gives

and hence

(15)

Moreover, (1) implies that

(16) xl/(t) = Ft(t,x(t))+ Fx(t,x(t))x'(t)

=Ft(t,x(t)) + FxCt,x(t))F(t,x(t)).
Thus if

(17) IF(t,x)1 ~ M and IFt(t,x)1 ~ K for ~ t ~ to +T

 2.7 NUMERICAL METHOOS FOR THE INITIAL VALUE PROBLEM 131

for all relevant values of x then

(18)

Combining the estimates (13), (15) and (18) gives the estimate
1
(19) IEk+l(h)1 ::; (1 + hL)IEk(h)1 + "2h2(K + LM),
for k = 0,1,2, ... , N - 1. Moreover, E(h) = x(to) - Xo = 0. On applying (19) for
k = 0, 1,2, ... ,N - lone gets the following error estimate.
An Error Estimate for Euler's Method. If F(t,x) satisfies (14) and (17), then

(20) IEk(h)I::;Ch for k=0,1,2, ... ,N,

where

(21) C= ~ (M + ~) (e LT - 1).
The details of the proof of (20) will not be given here. They can be found in most
numerical analysis texts; see, e.g., [HeJ or [Ch-KJ. Note that (20) implies that if Ch ::; E,
so that IEk(h)1 ::; E, then

(22)

Thus, at worst, halving the step size h will reduce the error to one-half its former value.
This explains the rather slow convergence shown in Table 5. Note that the four h values
in the table are hI = 1/4, h2 = 1/64, h3 = 1/256 and h4 = 1/4096 with the successive
ratios hk+I/h k = 1/16. Hence, each EN should be about 1/16 of the preceding one. The
table entries are consistent with this estimate.
The remainder of this section presents some more accurate finite difference methods
for solving the initial value problem (1). In each case an error estimate of the form

(23)

holds, where m is a positive integer. A method for which (23) holds is said to be an mth-
order method. Thus, Euler's method is a first-order method. Methods of orders 2 and 4
are described below. Note that (23) implies that

(24)

Hence methods of order m 2: 2 converge much more rapidly than Euler's method.
Taylor Series Methods. In the error analysis of Euler's method, Taylor'S theorem
with second-order remainder, equation (10), led to Euler's equations xk+l = Xk + hx~,
132 2 ORDINARY DIFFERENTIAL EQUATIONS OF FIRST ORDER

TABLE 6. The Taylor2 method applied to Example 1

N 8 16 32 64
0.73127 0.73111 0.73107 0.73106
-000021 -0.00005 -0.00001 0.00000

x~ = F(tk,Xk) when the second-order error term h2X"(~k)/2 was dropped. It is natural
to construct higher-order numerical methods by retaining additional terms in the Taylor
series. The simplest case, after Euler's method, is suggested by Taylor's theorem with
third-order remainder:

(25)

where th :::; ~k :::; tHI. Again, the differential equation (1) gives X'(tk) = F(tk,X(tk».
Moreover, differentiaLing (1) gives
d
(26) x"(t) = d/(t,x(t») = Ft(t,x(t» + FxCt,x(t»x'(t)
and hence

(27)

This suggests the algorithm

(28) Let to, Xo, T, N be given; set h = TIN.

Fork=O,l, ... ,N-l,
x~ = F(th,Xh),
x~ = Ft(tk,Xk) + FxCtk,Xk)x~'
tk+l = tk + h,
Xk+l = Xk + hx'k + 2:Ih 2x h. /I

An error analysis similar to that for Euler's method above confirms that this algorithm is
second order, as expected. We shall refer to (28) as the Taylor2 algorithm.

EXAMPLE 2. For equation (7) we take to = 0, Xo = 0.5, and T = 1. Moreover, F(t,x) =

x(l - x), Ft(t,x) = 0, and Fx(t,x) = (1 - 2x), so that
, 2
(29) Xk = Xk - xk ,
X~ = (1 - 2Xk)x~'

The Taylor2 algorithm with these values gives the results in Table 6.
 2.7 NUMERICAL METHODS FOR THE INITIAL VALUE PROBLEM 133

Note that the Taylor2 method yields x(l.O) to 5-place accuracy in only 64 steps, many
fewer than the 4096 steps needed with Euler's method. The error results in the table are
consistent with the 2nd order error estimate lEN I :::: Ch 2 , since the h values for the table
are hI = 118, h2 = 1/16, h3 = 1132 and h4 = 1164 with common ratio hk+llhk = 112.
The Taylor4 Method. As a final example of the Taylor series method we shall
consider the Taylor4 method. This is defined by keeping the terms through order 4 in
the Taylor series. Thus

(30) x = x + hx(l) + ~h2X(2) + ~h3X(3) + ~h4X(4)

h+l k k 2 k 6 k 24 k'

where the coefficients x~£) = dixkldt£ for f = 1,2,3,4 are generated from F(t,x) and its
partial derivatives.

EXAMPLE 3. Application of (30) to problem (7) gives

(31) X~l) = XkO - Xk),

X~2) = 0 - 2Xk)X~l) ,
X~3) = (1 - 2Xk)X~2) - 2[x~l)12,
X~4) = (1- 2xk)X~3) - 6X~l)X~2).

Computation with these formulas gives the results in Table 7.

The Taylor4 algorithm requires only 8 steps [or 5-place accuracy. The method can be
shown to be fourth order, with an error estimate of the form IEN(h) I :::: Ch4. In particular,
halving h = liN reduces EN by a factor of about 1116, as is borne out by the table.
Heun's Method. The Taylor2 method requires us to calculate the partial derivatives
Ft(t, x) and Fx(t, x) analytically and then, at each step of the method to evaluate the three
functions F(t,x), Ft(t,x) and FAt, x) at the point (tk,Xk). For some functions F(t,x) it
may be prohibitively difficult or costly to calculate and evaluate these partial derivatives.
A way to avoid the use of Ft and Fx and still have a second-order algorithm was discovered
by K. Heun in 1900. Heun's method can be motivated as follows. The exact solutionx(y)

TABLE 7. The Taylor4 method applied to Example 1

N 2 4 8
0.72917 0.73098 0.73105 0.73106
0.00189 0.00008 0.00001 0.00000
134 2 ORDINARY DIFFERENTIAL EQUATIONS OF FIRST ORDER

of the initial value problem (1) satisfies the integral equation

(32) x(to + h) = x(to) + I

to
to +h
F(r,x(r)) dr,

obtained by integrating x' (r) = F( r, x( r)) over the interval to ::: r ::: to + h. If h is small,
the integral in (32) may be estimated accurately by one step of the trapezoidal method.
Thus
h
(33) x(to + h) ~ x(to) + "2 [F(to, x(to)) + F(to + h, x(to + h))].
Of course, the value x(to + h) on the right is not yet known. However, a good
approximation to it is given by Euler's method:

(34) x(to + h) ~ x(to) + hF(to,x(to)).

Let us set x(to) = Xo and define Xl, a new approximation to x(to + h), by
(35) Yl = Xo + hF(to,xo),
h
Xl = Xo + "2 [F(to,xo) + F(to + h,Yl)].
This is Heun's Method. We can iterate the process to get the general algOrithm:

(36) Let to, xo, T, N be given; set h = TIN.

Fork=O,l, ... ,N-l,
Yk+l = Xk + hF(tk,Xk), tk+l = tk + h,
Xk+l = Xk + ~ [F(tk,Xk) + F(tk+l,Yk+l)].
Error Analysis. Combining the two equations (35) gives Xl as a function of the step
size h:
h
(37) Xl (h) = Xo + "2 [F(to, xo) + F(to + h, Xo + F(to,xo))]·
If Taylor's theorem is used to expand Xl (h) in a power series then it is found, after a brief
calculation, that

(38)

where IR(h)1 S Mh 3 . Thus Heun's approximation Xl (h) differs from the Taylor2
approximation by a term of order 3. It can be shown, by a calculation similar to that for
the error analysis of Euler's method, that the overall error EN(h) for Heuns algorithm (36)
satisfies IEN(h)1 ::: Ch 2 Therefore, the method has order 2. In contrast with the Taylor2
 2.7 NUMERICAL METHOOS FOR THE INITIAL VALUE PROBLEM 135

method, second-order accuracy is achieved with only two evaluations of F(t, x) per step,
and the partial derivatives Ft(t,x) and Fx(t,x) are never used.

EXAMPLE 4. Application of (36) to problem (7) gives the results in Table 8.

We note that the error decreases by a factor of 4 when h is halved, as it should for a
second-order method.
Runge-Kutta Methods. Heun's method is one of a family of higher-order methods
that require several evaluations of F(t,x) at each step but involve no partial derivatives
of F(t,x). These methods were introduced by C. Runge in 1895, refined by K. Heun
and W Kutta in 1901, and again by C. Runge in 1905. 5 The best known of these is
a fourth-order method, often called the Runge-Kutta method. Here we shall call it the
Runge-Kutta4 method. The algorithm takes the following form.

(39) Let to, xo, T, N be given; set h = TIN.

Fork=0,1, ... ,N-1,
Fl = F(tk,Xk),
F2 = F(tk + ~h,Xh + ~hFl)'
F3 = F(tk + ~h,Xk + ~hF2)'
F4 = F(tk + h,Xk + hF3),
Xk+l = Xk + ~(Fl + 2F2 + 2F3 + F4).
If Xl = Xl (h) is expanded in a power series in h, then it is found, after a long calculation,
that

X (h) = X + x(l)h + ~x(2)h2 + ~x(3)h3 + ~x(4)h4 + R(h)

1 ° ° 2 ° ° 24 ° '
where IR(h) I S Mh5. It follows that the Runge-Kutta4 method is a fourth-order method,
like the Taylor4 method. The great advantage of the Runge-Kutta4 method over the
Taylor4 method is that it requires only four evaluations of F(t, x) per step. The Taylor4

TABLE 8. Heun's method applied to Example 1

N 8 16 32 64
0.73090 0.73102 0.73105 0.73106
0.00016 0.00004 0.00001 0.00000

'The original references, (in German), are C. Runge, Math. Ann. 46, 167-178 (1895); K. Heun, Z. Math. Phys. 45,
23-38 (1900); W Kutta, Z. Math. Phys. 46, 435-453 (1901); and C. Runge, Nachr. K. Ges. Wiss. G6ttingen, 252-257
(1905)
136 2 ORDINARY DIFFERENTIAL EQUATIONS OF FIRST ORDER

method requires us to calculate and evaluate F(t, x) and its partial derivatives through
order three at each step.

EXAMPLE 5. Application of (39) to problem (7) gives the results in Table 9.

The decreases in the error dearly show the fourth order convergence. Moreover, only
4 steps are needed for 5-place accuracy, while the Taylor4 method required 8 steps.
Remarks on Adaptive Methods. In practical applications of finite difference methods
to differential equations it is desirable to estimate the error at each step and to change
the step size h accordingly Algorithms that automate this process are called adaptive
methods. Many such algorithms have been developed. One effective adaptive method
is the Runge-Kutta-Fehlberg method. Details may be found in [Ch-KJ.

IExercises 2.7 ~
"~,ok~,~~.%,;n,mijJ , $W11>

Direction Fields. Apply the method of direction fields to produce a graphical solution to the given initial
value problem. Use 25 grid points.

1. x' = x, x(O) = 1. 3. x' = x(l - x), x(O) = 0.5

2. x' = (t + x)2, x(O) = O. 4. x' = x(1 - x), xeo) = 1.

Euler's Method. Apply Euler's method to the following initial value problems using 4 steps. Compare with
the exact solution.

5. x' =x,x(O) = 1, h = 0.1. 7. x' = x(l- x), xeo) = 0.5, h = 0.1.

6. x' = 1 +x 2 ,x(0) = 0, h = 0.05 8. x' = tx, x(O) = 1, h = 0.05.
Taylor2 Method. Apply the Taylor2 method to the following initial value problems using 4 steps. Compare
with the exact solution.

9. x, =x+t,x(O) = 1, h = 0.1. 11. x' = x(l- x), x(O) = 0.5, h = 0.1.

10. x' = t2 +x, x(O) = 0, h = 0.1. 12. x' = t2x, xeO) = 1, h = 0.1.
Taylor4 Method. Apply the Taylor4 method to the follOwing initial value problems using 2 steps. Compare
with the exact solution.

TABLE 9. The Runge-Kutta4 method applied to Example 1

N 2 4
0.73091 0.73105 0.73106
0.00015 0.00001 0.00000
 2.7 NUMERICAL METHODS FOR THE INITIAL VALUE PROBLEM 137

13. x' = t +x, xeO) = I, h = 0.2. 15. x' = x(l - x), x(O) = 0.5, h = 0.2.

14. x' = t2 +x,x(O) = 0, h = 0.2. 16. x' = t 2x, x(O) = I, h = 0.2.

Heun's Method. Apply Heun's method (Improved Euler) to the following initial value problems using 2
steps. Compare with the exact solution.

17. x, = x, x(O) = 1, h = 0.2. 19. x' = x(l - x), x(O) = 0.5, h = 0.2.

18. x, = 1 + x 2 , x(O) = 0, h = 0.1. 20. x' = tx,x(O) = I, h = 0.1.

Runge-Kutta Method. Apply the Runge-Kutta4 method to the following initial value problems using I
step. Compare with the exact solution.

21. x' = x, x(O) = 1, h = 0.4. 23. x' = x(I - x), x(O) = 0.5, h = 0.4.

22. x, = 1+ x 2 , x(O) = 0, h = 0.2. 24. x' = tx,x(O) = 1, h = 0.2.

25. (Euler) Write a computer program to implement Euler's method. Test it on the logistic equation and
verify Table 5.
26. (Taylor2) Write a computer program to implement the Taylor2 method. Test it on the logistic equation
and verify Table 6.
27. (Taylor4) Write a computer program to implement the Taylor4 method. Test it on the logistic equation
and verify Table 7.
28. (Heun) Write a computer program to implement the Heun method. Test it on the logistic equation and
verify Table 8.
29. (Runge-Kutta4) Write a computer program to implement the Runge-Kutta4 method. Test it on the
logistic equation and verify Table 9.
30. (Comparisons) Use the Euler, Heun, and Runge-Kutta methods to compute x(l) for the problem
x'(t) = x 2 (t) + 1, x(O) = 0, with step size h = 0.1. Compute the exact answer and compare the four
answers in a table.
31. (Quadrature) The integral f;/2 cos(t)dt can be evaluated by Simpson's rule. It can also be evaluated
by solving the initial value problem X'(t) = cos(t), x(O) = 0, for the value x(rr/2). Choose step size
h = O.lrr12. Apply the Euler, Heun, and Runge-Kutta4 methods to evaluate x(rr/2) and compare the
answers with the corresponding answer obtained by Simpson's rule using 10 divisions.
C HAP T E R 3
 Ordinary Differential
Equations of Higher Order

This chapter continues the theme of Chapter 2: differential equations and their
applications in science and engineering. Chapter 2 emphasized first-order equations

dx/dt = F(t,x).

Here equations that contain second, and higher-order derivatives are analyzed. Section
3.1 shows how such equations arise naturally in applications. Sections 3.2 through "3.5
develop the theory and applications of linear second-order equations. Section 3.6 treats
linear equations of the third and higher orders.

3.1 Examples from Engineering and Physics

Differential equations that model several physical systems are developed in this section.
Consider first the LRC series circuit.
LRC Series Circuit. Figure 1 depicts an LRC series circuit with constant parameters
L, R, and C. The function E(t) is a prescribed function of time t that represents the output
of a generator or other voltage source.
Thus L, Rand C are the circuit inductance, resistance and capacitance. If q(t) is the
charge on the capacitor C and i(t) is the electric current in the circuit then the voltage
drop formulas on these circuit components are known to be
di
(1) VL = L- = voltage drop across L,
dt
140 3 ORDINARY DIFFERENTIAL EQUATIONS OF HIGHER ORDER

IV
E(t)
c

R
FIGURE 1. An LRC circuit with voltage
source E(t)

VR = Ri = voltage drop across R,

Vc = ~ = voltage drop across C.

In addition, the algebraic sum of the voltage drops equals zero, by Kirchhoff's law:

VL + VR + Vc - E = O.

The charge q(t) on the capacitor C and electric current i(t) in the circuit are related by
. dq
(2)
1= dt'
whence
d'
(3) Ldl + Ri + 3.. = E(t).
t C
Elimination of i(t) gives a single second-order equation for q:

(4) Ld
2
q + R dq + 3.. = E(t).
dt 2 dt C
In Section 2.5 it was shown that if E(t) = 0 then (4) can be solved by means of a Riccati
equation. This method fails if E(t) =f=. O. More general methods that are applicable in all
cases are developed in Sections 3.2 and 3.3 below, and in Chapter 4.
Mechanical Oscillators. Consider the mechanical oscillator of Figure 2.
A mass m is suspended from a spring and attached to a damper such as a
pneumatic door stopper or an automotive shock absorber. The displacement of m from
its equilibrium position under gravity is denoted by x. It will be assumed that three forces
act on m:

-kx = restoring force of the spring,

dx
-c dt = resisting force of the damper,
F(t) = a prescribed external force,
 3.1 EXAMPLES FROM ENGINEERING AND PHYSICS 141

Tx
m

c
FIGURE 2. A mechanical oscillator with damper

where k and c are positive constants that characterize the spring (Hooke's law) and the
damper, respectively. The acceleration of the mass m is d2x1dt 2 . Thus Newton's second
law of motion implies that x satisfies the differential equation

d2x dx
(5) m dt 2 + c dt + kx = F(t).
Notice the similarity between this equation and equation (4). The equations become
identical if one makes the identifications L = m, R = c, lie = k and E(t) = F(t). Thus
if (4) has been solved then (5) may be solved by just renaming the letters. Such equations
are said to be isomorphic, meaning that they have the same form.
A Torsional Pendulum. Yet another oscillator that is modeled by an equation
isomorphic to (4) is depicted in Figure 3. It is assumed that the structure is cylindrical, is
made of an elastic material such as steel, and is fastened rigidly at the top. Such torsional
pendulums are used in some clocks.
If the pendulum is rotated about its axis, the displacement is described by the twist
angle e. If I denotes the moment of inertia of the pendulum then Newton's law for
rotating bodies states that

d2e
(6) 1 - = net torque on the pendulum.
dt 2

....... ~
.. ~.

FIGURE 3. A torsional pendulum

142 3 ORDINARY DIFFERENTIAL EQUATIONS OF HIGHER ORDER

It will be assumed that three torques act:

(7) - ke = elastic restoring force (Hooke's law),

de
- ( - = linear damping torque,
dt
TCt) = a prescribed external torque,

where k and ( are positive constants characteristic of the particular pendulum. Combining
(6) and (7) gives the differential equation

(8)

which dearly is isomorphic to (4). It will be seen below that this kind of isomorphism
extends to many mechanical and electrical systems. Further examples may be found in
[DenH].

Equations with Variable Coefficients. In the LRC circuit described above the
coefficients L, R, and C are constants. However, in some applications, L, R, or C may be
caused to vary with t in a prescribed way, either by mechanical or by electronic means.
This leads to equations of the form

(9)

where the notation means that one or more of the coefficients L, Rand C may vary with t.
Isomorphic equations arise from mechanical oscillators whose parameters m, ( and k vary
with t. A variety of examples from the theory of mechanical vibrations may be found in
[DenH], Chapter 8. Here two of the problems from [DenH] will be described. Solutions
are presented in Section 3.5 below. Both problems involve an equation of the form

(10)

where bet) is positive and periodic with period p > 0, so that for all t,

(11) bet + p) = bet).

An example is (see [DenH], Chapter 8)

(12) bet) = k + ~k sin(wt),

 3.1 EXAMPLES FROM ENGINEERING AND PHYSICS 143

where k and I'.k are constants and w > 0, so that p = (2rr)/w. Equation (10) with
periodic bet) is called a Hill's equation. 1 The special case (12) is usually called Mathieu's
equation. 2
Two examples of physical problems that lead to Hill's equation will be given. The
examples are from page 340 in [DenHJ.
Pendulum of Variable Length. The small amplitude oscillations of a pendulum are
governed by the equation

(13)

see Figure 4 and page 340 in [DenHJ. If f. is varied periodically with period p (and hence
with frequency w = (2rr)/p) by pulling and releasing the string through the point of
support, then (13) has the form (10), (1l).
This example suggests how solutions of (10), (11), and (12) may be unstable due to
resonance. The frequency of the unperturbed pendulum (I'.k = 0) is Wo = .Jk radians
per sec. If w = 2wo and if the initial conditions are chosen properly, then f. is shortened
when x = 0 and the tension in the string is a maximum and f. is lengthened when
x = ±xmax , and the tension is a minimum. Indeed, this is essentially what a child does
to "pump" a schoolyard swing. The result is that work is done on the system, twice for
each oscillation, producing an ever increasing amplitude of the motion. It will be shown
below how mathematics can be used to predict and analyze such instabilities. A second
physical example is the following:
Oscillation of a Mass on a String with Variable Tension. Consider a mass m,
vibrating at the midpoint of a taut string with length f. and tension T; see Figure 5 and
page 340 in [DenHJ.

FIGURE 4. Pendulum of variable length

lAfter George William Hill (1838-1914). American mathematician and astronomer.

2After Emile-Leonard Mathieu (1835-1900), French mathematician.
144 3 ORDINARY DIFFERENTIAL EQUATIONS OF HIGHER ORDER

m
T
J
FIGURE 5. Mass on string with variable tension

Newton's laws imply that for small vertical displacements x of the mass m, one has

(14) ~:~ + (:~)x = O.

If T = TCt) varies periodically then (14) is equivalent to (10), (11). Clearly, if TCt) is a
maximum when m moves toward the center, and is a minimum when it moves away,
then work is done on the system and the amplitude of the oscillation will increase. This
effect may occur in electrical transmission lines, where an alternating current produces
periodic heating and corresponding periodic variations of the tension. Resonance may
occur if the natural frequency of the line is twice the current frequency.
Examples from Partial Differential Equations. The partial differential equations
of mathematical physics are a source of a great variety of linear second-order equations
with variable coefficients. A number of these partial differential equations are derived
and studied in Chapters 8 and 9. Here some of these equations are quoted and used to
derive examples of linear second-order ordinary differential equations.
Temperature Distributions in a Nonuniform Rod. Consider a thin heat-con-
ducting, nonuniform rod of length f, as in Figure 6.
If u(x, t) is the temperature in the rod at section x and time t then it is shown in
Chapter 8 that u(x, t) is a solution of a partial differential equation of the form

(15) w(x)- au = -a ( r(x)-

au) - q(x)u + F(x, t).
at ax ax
Here w(x), r(x), and q(x) are positive functions that are combinations of the physical and
geometric parameters of the rod such as mass density, specific heat, thermal conductivity
and cross-sectional area. (Details are in Section 7.2 and Section 8.11). The function F(x, t)

o x FIGURE 6. NonUniform Rod and Coordinate Axis

 3.1 EXAMPLES FROM ENGINEERING AND PHYSICS 145

is determined by the external and internal sources of heat. For static (time-independent)
sources F(x) and temperatures u(x) one gets the equation

(16) dx dU) -
d ( rex) dx q(x)u = -F(x),

or
d u (1 dr(X)) du q(x)
- - - - - - U = - -F(x)
2
(17) -+ -.
dx2 rex) dx dx rex) rex)

This is a linear second-order equation with variable coefficients.

Vibrating Circular Membrane. Consider a circular elastic membrane that is
stretched over a circular frame of radius a, as in Figure 7. The motion of the membrane
will be described by a function z = u(r, e, t) that gives the displacement out of the
(x,y)-plane.
It will be shown in Chapter 9 that for small displacements, u is a solution of the
partial differential equation

1 a2u a2u 1 au 1 a2u

(18) -c2 -=-+-
at 2 ar2
-+--.
r ar r2 al]2

Here c2 = Tip, T is the tension in the membrane and p is its mass density. Next consider
the synchronous vibrations of the membrane, defined by the product

(19) u(r, e, t) = vCr, e) cos(wct).

Substituting this into (18) gives for vCr, e) the partial differential equation

(20)
a2v 1 av 1 a-
-ar2 + -r -ar + -
2v
+w2 v = o.
r2 ae 2

FIGURE 7. Circular membrane and coordinates

146 3 ORDINARY DIFFERENTIAL EQUATIONS OF HIGHER ORDER

In analogy with (19), consider product solutions of (20) of the form

(21) v(r,e) = R(r)G(e).

Substituting this into (20) gives, after some rearrangement of terms, the equation
rlRI/(r) + rR'(r) GI/((})
(22) __ ----,--,----__ + (j}r l = _ --.
R(r) G(e)
This must hold for all points of the membrane; i.e., 0 < r < a and 0 :::: e::
2n. Hence
the two sides of (22) must equal one and the same constant, which we shall assume
positive and write as JL 2 . Thus

(23)

and

(24) rlRI/(r) + rR'(r) + (j}r2R(r) - JL 2R(r) = 0 for 0 < r < a.

Equation (23) has constant coefficients and can be solved by the method of Section 2.5.
The general solution is

(25) G(e) = A cos JLe + B sin JLe.

This function is single-valued at the points of the disk only if JL = n is an integer. Thus

(26) JL = n = integer (::: 0),

G(e) = Gn(e) = An cos ne + Bn sin ne.
Equation (24) is more complicated. With JL = n it reads

(27)

The parameter w may be eliminated by defining

(28) x =wr,
y(x) = y(wr) = R(r).

Making these substitutions in (27) gives

(29)

or, equivalently,

-d y2 + -1 -dy + ( 1 -
2 n2)
(30) - Y = O.
dx X dx x2
 3.1 EXAMPLES FROM ENGINEERING ANO PHYSICS 147

This is the well-known Bessel equation of order n, which has many applications in
physics and engineering. 3 Properties of its solutions are studied below, in Section 3.4 and
in Chapters 8 and 9. The analysis of the vibrating membrane problem will be completed
in Chapter 9.
Steady-State Temperature in a Disk. Consider the disk of Figure 7 and let T =
uCr,8) denote the steady-state Ctime-independent) temperature there. It will be shown
in Chapter 9 that u(r, 8) satisfies the Laplace equation in polar coordinates r, 8, namely,

a2 u 1 au 1 a2 u
(31) -+--+--=0
ar2 r ar r2 a8 2 .

Note that this is identical to equation (20) with w = O. Hence the product solutions of
(31),

(32) u(r,8) = R(r)8(8),

satisfy (23) again, and (24) will be w = 0, or

d2 R 1 dR n 2
(33) -+----R=O.
dr2 r dr r2
This equation will be solved below. The solution of the steady temperature problem for
the disk will be completed in Chapter 9.
Steady-State Temperature in a Sphere. The steady temperature in a solid homo-
geneous sphere is a function T = u(x,y, z) of the three rectangular coordinates x, y and
z that satisfies the partial differential equation of Laplace:
a2 u a2 u a2u
(34) -+ -+-= 0
2
ax 2 ay2 az .

This result is derived in Chapter 9.

To study the temperature in the ball of radius a, defined by x 2 + y2 + Z2 ::: a2, it is
natural to use the spherical coordinates r, 8, cp as shown in Figure 8, defined by

(35) x = r cos 8 sin cp,

y = rsin8sincp,
Z = rcoscp.
Clearly, r ~ 0, 0 ::: 8 ::: 2rr, and 0 ::: cp ::: n. Here, for Simplicity, only steady
temperatures of the form T = vCr, cp) will be discussed. Thus T is independent of the
longitude coordinate 8 and hence is symmetric about the z-axis. For this case it will be

3The word "order" is overworked in mathematics. Here it denotes the parameter n rather than the second derivative.
148 3 ORDINARY DIFFERENTIAL EQUATIONS OF HIGHER ORDER

()
x FIGURE 8. Spherical coordinate system

shown in Chapter 9 that the Laplace equation (34) is equivalent to the partial differential
equation

(36) -J(2JV)
Jr
r -
Jr
+ -1- -J(.sm ¢ -
sin¢ J¢
av) = O.
J¢
A family of product solutions of this equation,

(37) v(r,¢) = R(r)<I>(¢),

will be sought, as in the preceding cases. Substitution of (37) into (36) gives, after
separation of variables,

(38) 1 d ( 2dR) -1 d ( . d<l»

R dr r dr = <I>(¢)sin¢ d¢ sm¢ d¢ .

This relation must hold for 0 < r < a and 0 < ¢ < 1T, and hence, as in the disk problem,
the two sides of (38) must equal the same constant f,A,. Thus

(39) ~ (r2 dR) _ f,A,R = 0 for 0 < r < a,

dr dr
and

(40) d~(sin¢~:)+f,A,Sin¢<I>(¢)=O for 0<¢<1T.

The equation for R can also be written as

d2R 2 dR f,A,
(41) -+----R=O
dr2 r dr r2 '
which is similar to, but not identical with, equation (33). It will be solved below. Equation
(40) for <I> is much more complicated. It can be simplified somewhat by the change of
variable

(42) x = cos¢, y(x) = <I> (¢).

 3.1 EXAMPLES FROM ENGINEERING AND PHYSICS 149

Then the chain rule of calculus gives

d<l> dy dx . dy
- = - - = -(sm¢)-
d¢ dx d¢ dx
and hence

(43) :£ (sin ¢ d<l»

d¢ d¢

Combining this with (40) gives the Legendre equation 4

d ( (1- x 2) dy
(44) dx dx ) + J-LY = 0,

or, equivalently,

2 d2y dy
(45) (1 - x ) - - 2x- + J-LY = O.
dx 2 dx
Solutions of this equation are derived below, in Section 3.4, and are applied in Section
3.9 to construct steady-state temperature distributions in spheres.
Examples of Systems from the Theory of Oscillations. The remaining examples
illustrate the occurrence of systems of ordinary differential equations in the theory of
electrical and mechanical systems with several degrees of freedom. It will be shown that
such systems lead by elimination to higher-order linear differential equations; that is,
equations that contain derivatives of order two or more.
Multiloop Electric Circuits. Kirchhoff's laws, when applied to multiloop circuits,
lead to systems of ordinary differential equations with two or more unknown functions.
This will be illustrated here with the simple two-loop circuit of Figure 9.
The variables ql and q2 denote the charges on the capacitors CI and C2 , while iI,
i2, and i3 are the currents in the three branches, as indicated in the figure. Kirchhoff's
current laws applied to the branch junctions give

4 After A.M. Legendre 0752-1833), French mathematician, who introduced it in the theory ofthe Laplace equation.
150 3 ORDINARY DIFFERENTIAL EQUATIONS OF HIGHER ORDER

q]

t ~i2

C] i3 q2
i]
C2
~

FIGURE 9. Two-loop LRC circuit

(46)

and
. dq]. dq2
(47) I] = dt' 12 = dt'
Moreover, Kirchhoffs voltage laws applied to the two loops give (since b =h - i2 )

(48)

and

(49) ~ - R(h - i2) = O.

C2
Eliminating i] and i2 by means of (47) gives

(50)

and

(51) Rdq2 _ Rdq ] +~ = O.

dt dt C2
This is a system of two differential equations for the unknown functions q] and q2.
In Chapter 4 it will be shown how the Laplace transform can be used to solve the
system (50), (51) directly for ql and q2. Here an alternative method, based on elimination,
is described. The idea is to eliminate one of the unknowns to get a single equation for
the other. To carry this out, note first that adding (50) and (51) gives

(52)
 3.1 EXAMPLES FROM ENGINEERING AND PHYSICS 151

If this is solved for q2 and the result is substituted for q2 in (50), then the following
equation for ql alone is found:

(53) (RC L) d3ql + Ld2ql + (R + RC2) dql +~ = O.

2 dt3 dt2 C1 dt C1
For example, the special parameter values R = lO, L = 6, C2 = 1/60, Cl = 1/6 give
d3 d2 d
(54) ~3
dt
+ 6~2 + 11 ~ + 6 1 = O.
dt dt q
This is a linear differential equation of order three for q1. In Section 3.6 it will be shown
that the general solution of (54) is

(55)

where AI, A 2 , and A3 are arbitrary constants. Moreover, substituting this in (52) gives
1 -t 1 -2t -3t
(56) q2 = --Ale - -A2e - A3e .
5 2
It can be shown that (55), (56) give the general solution of the system (50), (51) for the
special parameter values of R, L, Cl and C2 chosen above.
Mechanical Oscillators with Two Degrees of Freedom. The electric circuit of
Figure 9 is said to have two degrees of freedom because two functions, ql (t) and q2 (t),
are needed to specify the state of the system. Similarly, mechanical oscillators are said
to have two degrees of freedom if their states are specified by two displacements. An
example is shown in Figure lO.
The masses ml and m2 are attached to springs with spring constants hI and h2 ,
respectively The displacements of ml and m2 are both measured downward from their
eqUilibrium positions, which are separated by the fixed distance Xo. The masses are
coupled by a damper, which opposes the relative motion of m1 and m2 with a force

FIGURE 10. Mechanical oscillator with two degrees of freedom

152 3 ORDINARY DIFFERENTIAL EQUATIONS OF HIGHER ORDER

±c(d/3ldt), where /3 is their instantaneous separation. From the figure, Xl + /3 = Xo + X2,

and so d/3ldt = dx 21dt - dxl/dt. The damper opposes the motion, so Newton's second
law gives the system

(57) ml d;~l = _ hlXI + c ( ~2 _ ~l ) ,

(58) m2 d2x2 __ h2X2 _ C (dx 2 _ dx l ) .

dt 2 - dt dt

This is a system of two second-order differential equations for the unknown displace-
ments Xl (t) and X2 (t). In Chapter 4 the system is solved directly by means of the Laplace
transform. Here it will be shown how it may be solved by eliminating X2 to obtain a
single fourth-order equation for Xl.
It will be convenient to write
d
(59) D=-
dt
and rewrite (57), (58) as

(60) (m1D 2 + cD + hl)xl = cDx2,

(61) (m2D2 + cD + h2)X2 = CDXI.
If the operator (m2D2 + cD + h2) is applied to both sides of (60) then by (61) we have
(62) (m2D2 + cD + h2) (m1D2 + cD + hI) Xl
= cD (m2D2 + cD + h2) X2
= cD(cDxI) = c2D2XI'
On multiplying out the left-hand side of (62) and canceling the like terms c2D2XI, one
obtains a single fourth-order equation for Xl, namely

d4xI d3XI
(63) (mlm2) dt 4 + (ml + m2)c dt 3
d2XI dx l
+ (ml k2 + m2 hl) dt 2 + (hI + h2 )CTt + (k l h2 )x1 = O.

For example, for the special parameter values ml = m2 = 1, c = 1, hI = h2 = 1, this

equation becomes

(64)
 3.1 EXAMPLES FROM ENGINEERING AND PHYSICS 153

In Section 3.6 it will be shown that the general solution of (64) is

(65)

where A I ,A2 ,A3, and A4 are arbitrary constants. Moreover, with Xl known, the function
DX2 = dx 21dt can be calculated from (60) and X2 from (58). This procedure gives

(66) X2 = (-Al)e-t + (-A 2 )te- t + A3 cos t + A4 sin t.

It can be shown that (65), (66) is the general solution of the system (57), (58) for the
special parameter values chosen above.
One special case of the solution is worth mention. If ml = m2 and hI = h2, as in the
numerical example, then there is a special solution with Xl = X2. In this case 8 = constant
and the damper does not affect the motion. This case corresponds to Al = A2 = 0 in
(65), (66), and the motion is periodic with no damping.

LRC Circuits. Determine the differential equation and initial conditions for the given circuit.

1. R = 100, L = 1, E = 5 and no capacitor. Initial charge and current are zero.

2. C = 1110, L = 1, E = 5 and no resistor. Initial charge and current are zero.

3. R = 100, C = 1110, L = 1, E = 5. Initial charge and current are zero.

4. R = 100, L = 1, E = 5 and no capaCItor. Initial charge is zero and the initial current is 5.
5. C = 1110, L = 1, E = 5 and no resistor. Initially the capacitor is charged to 0.05 farads and the current
is initially zero.

6. R = 100, C = 1110, L = 1, E = 5. The capacitor is charged to 0.001 farads and an ammeter reads 3
amperes at time t = O.
Mechanical Oscillators. Determine the differential equation and initial conditions for the given
mechanical system. All units are cgs.

7. No damping is present, the mass is 10 and the Hooke's spring constant is 2. No external force. At time
t = 0 the position and velocity are zero and one, respectively
8. No damping is present, the mass is 10 and the Hooke's spring constant is 2. The externalforce is 100 sin wt.
At time t = 0 the position and velocity are both zero.
Torsional Pendulum. Determine the differential equation and initial conditions for the given system. All
units are cgs.

9. The moment of inertia is 1, no damping is present and the elastic restoring force constant is 10- 2 No
external torque is present. The twist angle and velocity are initially one.

10. The moment of inertia is 5, the damping constant is 2 and the elastic restoring force constant is 1. No
external torque is present. The twist angle and velOCity are initially one.
154 3 ORDINARY DIFFERENTIAL EQUATIONS OF HIGHER ORDER

11. The moment of inertia is 1, the damping constant is 3 and the elastic restoring force constant is 2. The
external torque is 10sinwt. The twist angle and velocity are initially zero.

12. The moment of inertia is 1, no damping is present and the elastic restoring force constant is 2. No
external torque is present. The twist angle is initially 90 degrees and the velocity is initially zero.
Pendulum of Fixed Length. Determine the differential equation and initial conditions for the given system.

13. A string oflength 10 creates a pendulum with attached mass 0.01, all in fps units. The small oscillations
of the pendulum begin by pulling the pendulum to angular position 7 degrees and releasing the mass.

14. A string of length 5 creates a pendulum with attached mass 10, all in cgs units. The small oscillations of
the pendulum begin by pulling the pendulum to angular position rrll2 radians and releaSing the mass
with velOCity 1 em/sec.
Temperature Distribution in a Nonuniform Rod. Determine the linear second-order ordinary differential
equation for the time-independent temperature u(x).

15. The partial differential equation is

au = -a ( yix-
- au) - x 2 sin(x)u.
at ax ax

16. The partial differential equation is

17. (Bessel Equation) In r 2R"(r) + rR'(r) + (<<}r2 - n 2 )R(r) = 0, make the change of variables X = wr,
y(x) = R(r) to obtain the Bessel differential equation x 2y" + xy' + (x 2 - n 2 )y = O.

18. (Steady-State Temperature in a Disk) Derive the differential equation r2R" (r) + rR'(r) - n2R(r) = 0
by substitution of the product solution u = R(r)EJ(iI) into the partial differential equation r 2 u rr + fUr +
ulJIJ = O. Assume that the separation constant A has the form A = n2 (see Chapter 8).

19. (Legendre Differential Equation) Explain why (0 - x 2 )y')' + f.LY = 0 is equivalent to the equation
(l - x 2 )y" - 2xy' + f.LY =
O.
20. (Euler's Differential Equation) Define variables r = e:", y(x) = R(r). Find relations for r 2d2R1dr2
and rdRldr in terms of y and x. Use them to show that the Euler equation r2 ~ + r¥r - f.LR = 0 is
equivalent to the constant coefficient equation y" - f.L y = 0

21. (Multiloop Circuit) Verify directly that q = Aj c- t + A2c-2t +A3c-3t is a solution of the loop equation
q'" + 6q" + 11q' + 6q = O.
22. (Mechanical Oscillator) Verify directly that x = Aj c- t + A2te-t + A3 cos t + A4 sin t is a solution of
the oscillator equation X'III + 2X'" + 2x" + 2x' + x = O.

23. (Parallel LRC Circuit) Consider the parallel LRC circuit of Figure 11 with current source i(t). Let ij,
i2, i3 and i be the currents in the four branches, as shown, and let u(t) = Li3(t). Use Kirchhoff's laws
to show that

(u " 1, + -u
+ -u 1 = i(t).
R L
24. (Mass-String System) For the mass-string system of Figure 5, show that for small displacements x
the vertical force on the mass m is -2Tsin8 "" -2Ttan8 and tanil = 2x1e. Then use Newton's law to
verify the differential equation (14).
 3.2 LINEAR SECOND-ORDER EQUATIONS - STRUCTURE OF SOLUTIONS 155

«-- i
FIGURE 11. Parallel LRC circuit with current source

FIGURE 12. Two-loop electrical circuit

25. (Two-Loop Circuit) Consider the two-loop electrical circuit of Figure 12. Let i), i2 be the electrical
currents in the two branches, as shown, and let q), q2 be the corresponding electrical charges on the
plates of the two capacitors.
Use Kirchhoff's laws to derive a system of two differential equations for q), q2. Then show that the
system is isomorphic to the system (57), (58) for the mechanical oscillator of Figure 10.

3.2 Linear Second-Order Equations - Structure of

Solutions
Second-order ordinary differential equations x" = F(t, x, x') were introduced in
Section 2.5. Such an equation is said to be linear if it can be put into the form

(1) x" + a(t)x' + b(t)x = jet),

where a(t), bet) and jet) are prescribed functions of t. The corresponding homogeneous
equation is

(2) x" + a(t)x' + b(t)x = o.

If jet) #- 0 then equation (1) is said to be nonhomogeneous. It is said to have constant
coefficients if aCt) = ao and bCt) = bo are constants. Otherwise it is said to have variable
coefficients.
The purpose of this section is to develop some properties of the solutions of linear
second-order equations that will be used repeatedly in the remainder of this chapter and
in later chapters.
156 3 ORDINARY DIFFERENTIAL EQUATIONS OF HIGHER ORDER

For applications to electric circuits,f(t) may represent an input voltage that has simple
jump discontinuities such as a square wave or sawtooth wave. Similarly, in applications
to mechanics, J(t) may represent a discontinuous applied force, and aCt) and bet) may
be system parameters that vary discontinuously. Hence, in what follows aCt), bet) and
J(t) are allowed to be sectionally continuous. The solution concept for (1) may then be
formulated as follows.

Definition. A function x(t) (defined on a certain interval to ~ t ~ tl) is said to be a

solution of (1) if (i) x(t) and x' (t) are continuous, (ii) x" (t) is sectionally continuous with
discontinuities only at discontinuities of aCt), bet) andJ(t) (if any), and (iii) equation (1)
holds at each point of continuity of aCt), bet) andJ(t).

The appropriateness of this definition is confirmed by the following theorem.

Fundamental Theorem. For each pair of numbers Co and CI, equation (1) has exactly
one solution x(O such that
(3)

Clearly, this theorem guarantees that (1) has a two-parameter family of solutions on
any interval to ~ t ~ tl, where aCt), bet) andJ(t) are defined and sectionally continuous.
This fact underlies all the work that follows on second-order linear equations. The proof
of the fundamental theorem, like that for the first-order case in Section 2.6, requires
techniques of advanced calculus, see, e.g., [B-RJ.
The Homogeneous Equation. Note that the identically zero function x(t) == 0 is
always a solution of (2). It will be called the trivial solution. Next, let Xl (t) and X2{t) be
two solutions of (2) and let Al and A2 be constants. Then the function

(4)
is said to be a linear combination of XI(t) and X2(t), with coefficients Al and A 2. A
simple but fundamental property of the linear homogeneous equation (2) is that linear
combinations oj solutions are again solutions. The reader is advised to check this property.
Linear Dependence. A linear combination (4) is said to be trivial if Al =
0 and
A2 = O. Otherwise, it is said to be nontrivial. Two functions Xl (t) and X2 CO are said to
be linearly dependent if a nontrivial linear combination of them equals zero:

(5)

This may be expressed by stating that Xl (0 and X2 (t) are linearly dependent if and only
if one of them is a constant multiple of the other. For example, if (5) holds with Al =I- 0
thenxI(t) = -(A2/A I )X2(t).
 3.2 LINEAR SECOND-ORDER EQUATIONS - STRUCTURE OF SOLUTIONS 157

Linear Independence. Two functions Xl (0 and X2 (t) that are not linearly dependent
are said to be linearly independent. Thus, if Xl (t) and X2 (0 are linearly independent and
AIXI (t) + A2X2 (t) = 0 then Al = 0 and A2 = O. This may be expressed by stating that
Xl (0 and X2 CO are linearly independent if and only if neither is a multiple of the other.
The Wronskian. A simple test for linear independence uses the Wronskian
functions

(6)

With this notation one has the

Test For Linear Dependence. Two solutions of (2), Xl(t) and X2(t), are linearly
dependent on an interval to ::: t ::: tl if and only if W(Xl, X2, t) = 0 on to ::: t ::: tl.
To verify this, note that if Xl (t) and X2 (t) are linearly dependent then Xl (0 = e X2 (0
or X2(t) = eXl (t). In either case, substitution in (6) gives W(Xl,X2, t) = O. Conversely, if
W(XI, X2, t) = 0 and Xl (0 = 0 or X2 (t) = 0, then Xl, X2 are linearly dependent. Finally,
if W(XI, X2, t) = 0 and neither Xl (t) nor X2 (t) is zero, then (6) implies that

(7) ~ (In)X2(t))) = ~ dx 2 _ ~ dx l = W(Xj,X2,t) =0

dt Xl (t) X2 dt Xl dt XIX2
and so X2(t)/XI (t) = e, whence Xl and X2 are linearly dependent.
Abel's Formula. 6 For any two solutions of (2), Xl (t) and X2(t), on an interval to :::
t ::: tl, one has

(8) W(XI,X2, t) = W(XI,X2, to) exp {- it a(r)dr} .

It follows that W(XI, X2, t) = 0 for to ::: t ::: tl if and only if W(XI, X2, to) = O. Thus
linear dependence can be checked by calculating W(XI, X2, t) at a single point.
To verify Abel's formula (8), differentiate (6) and use the assumption that Xl (t) and
X2 (t) are solutions of (2) to give

(9) d W(XI,X2, t) = XIX2/I - X2XI/I

dt
- aCt) (XIX; - x2xD
- a(t)W(xI,X2, t).

Then (8) follows by integration of the first-order differential equation (9).

Existence of Linearly Independent Solutions. Abel's formula and the fundamen-
tal theorem imply that (2) has many pairs oflinearly independent solutions. For example,

5 After j.M. Hone (Count Wronski) (1778-1853), Polish mathematician.

6After N.H. Abel (1802-1829), NorwegIan mathematiCIan.
158 3 ORDINARY DIFFERENTIAL EQUATIONS OF HIGHER ORDER

the fundamental theorem implies that (2) has a pair of solutions, Xl CO and X2 (t), with
initial values

(10) Xl (to) = 1, x~ (to) = 0,

X2(tO) = 0, x;(to) = 1.

This pair is linearly independent because W(Xl,X2, to) = 1. Of course, Xl and X2 can be
defined in many other ways. It is only necessary to choose their initial values such that
W(Xl,X2, to) 1= o.
Structure Theorem for Equation (2). It was noted above that if Xl (t) and X2(t) are
solutions of (2), then so are all of their linear combinations (4). It will now be shown
that if Xl (t) and X2(t) are any linearly independent pair of solutions of (2) and if x(t) is
any third solution, then there exist unique constants Al and A2 such that (4) holds. To
verify this, let x(t) be any solution of (2) and define Co = x(to), Cl = x'(to). Choose AI,
A2 to satisfy the equations

(11) AlXl(tO) +A2x2(to) = Co,

Alx~(to) +A 2x;(to) = (1·

This is a pair of linear equations for AI, A 2, and it has a unique solution for AI, A 2.
This is true because the determinant of the system is W(Xl, X2, to), and W(Xl, X2, to) 1= 0
because Xl (t) and X2 (t) are linearly independent. Now let

(12)

where AI, A2 satisfy (11). Then yet) is a solution of (2) and yet) satisfies y(to) = co,
y' (to) = (1, by (11). But this means that the solutions x(t) and yet) have the same initial
values. Hence x(t) and yet) are identical by the fundamental theorem. Thus if Xl and X2
are linearly independent solutions of (2) then every solution of (2) has the form (4). For
this reason a pair of linearly independent solutions of (2) is called a solution basis for (2).

The Inhomogeneous Equation. Next let x(t) and xp(t) be any two solutions of (1)
with the same right-hand sidej(t):

(13) X" + a(t)x' + b(t)x = jet),

x; + a(t)x~ + b(t)xp = jet).

If these equations are subtracted, then the jet) terms cancel, and yet) = x(t) - xp(t) is
seen to be a solution of the homogeneous equation. Thus yet) has the form (12):

(14)
 3.2 LINEAR SECOND-ORDER EQUATIONS - STRUCTURE OF SOLUTIONS 159

This proves the following theorem.

Structure Theorem for Linear Second-Order Equations. Let xp(t) be any particular
solution of (1) and let Xl (t), X2 (t) be any pair of linearly independent solutions of (2).
Then for each solution x(O of (1) there exist unique constants AI, A2 such that

(15)

The constants A I, A2 play the role of the two constants of integration that are expected
in the general solution of a second-order equation. Thus the structure theorem reduces
the calculation of the general solution of (2) to two steps: CO find a pair of linearly
independent solutions of the homogeneous equation (2), and (ii) find a particular
solution of the inhomogeneous equation (1). Of course, the steps can be carried out in
either order.
The most difficult step in finding the general solution (5) of 0) is to find a single
nontrivial solution Xl (t) of the homogeneous equation. In fact, the surprising result
will be shown that if a nontrivial solution Xl (t) of (2) can be found then a second
independent solution of (2) and a particular solution xp(t) of 0) can be calculated by
simple integrations.
The Method of Reduction of Order. This method starts with a solution Xl (0 =1= 0
of (2) and seeks a second solution X2 (t) of the form

(16)

Note that if vet) is not constant then the pair Xl (0, X2 (0 will be linearly independent.
Differentiation of (16) gives

(17) X; = vx~ + V'Xl,

x~ = VX" + 2v'x~ + V"XI.
If (6), (7) are substituted in (2) and the assumption that Xl is a solution of (2) is used,
it is found that v must satisfy

(18)

Clearly, vet) = ( = constant is a solution, but this gives X2 (t) = (Xl (t), and Xl, X2 are
linearly independent. What is needed is any nonconstant solution of (18). Now, (18) is
a linear separable first -order equation for

(19) z(t) = v' (t).

160 3 ORDINARY DIFFERENTIAL EQUATIONS OF HIGHER ORDER

In separated form the equation for z is

(20) -dz = - 2dx-l - a(t)dt,

z Xl
and integration gives

(21) log Izl + 2 log IXII = log Ixizl = - f a(t)dt,

and, by exponentiation,

(22) xi(t)z(t) = xi(t)v'(t) = e- ja(t)dt.

Another integration gives

(23)

Note that constants of integration can be chosen in any convenient way, since only one
solution is needed. It is clear from (22) that v' is not zero and hence Xl and X2 are linearly
independent.

EXAMPLE 1. Reduction of Order. Consider the equation with constant coefficients

(24) x" - 6x' + 5x = O.

By inspection, Xl = et is a nontrivial solution. To find a second solution X2 = VXl one

has, by (23),

and so

Since multiples of solutions are also solutions, it is seen that Xl = el and X2 = eSt are
linearly independent solutions of (24). Hence, the general solution of (24) is

by the structure theorem, where Al and A2 are arbitrary constants.

The Method of Variation of Parameters. To complete the solution of the non-

homogeneous equation (1), a method is needed to find a particular solution x/t). The
 3.2 LINEAR SECOND-ORDER EQUATIONS - STRUCTURE OF SOLUTIONS 161

method of variation of parameters, usually attributed to Lagrange,7 will produce a solu-

tion xp(t) by simple integrations whenever two linearly independent solutions Xl (t) and
X2(t) of the homogeneous equation (2) are known. The method constructs an xp(t) of
the form

(25)

where VI (t) and V2(t) are to be calculated. Of course, if VI and V2 are constants then
(25) defines a solution of the homogeneous equation. It will be shown that nonconstant
functions VI (t), V2 (t) can be constructed such that (25) defines a particular solution of
the nonhomogeneous equation.
Differentiation of (25) gives

(26)

Before differentiating again, note that two conditions can be imposed on the functions
VI and V2. One condition is the differential equation (1). As a second condition it is

traditional to take

(27)

Other choices are possible, but this one will prove to be convenient. Combining (26)
and (27) gives

(28)

and a second differentiation gives

(29)

On substituting (25), (28) and (29) into (1), one finds that

(30)

The terms containing VI and V2 (undifferentiated) cancel because Xl and X2 are solutions
of the homogeneous equation (check this!). Equations (27) and (30) are a pair of
simultaneous linear equations for v~ and v;. Moreover, the determinant of the system
is the Wronskian W(XI, X2, t) and is not equal to zero because Xl and X2 are linearly
independent. Hence, (27) and (30) have a unique solution for v~ and v;, which is readily
found to be

(31)

7Joseph Louis Lagrange 0736-1813), famous European mathematician. He was born in Turin and worked there
and in Berlin and Paris.
162 3 ORDINARY DIFFERENTIAL EQUATIONS OF HIGHER ORDER

Functions VIet) and V2et) are obtained by integration of (31). Note that constants of
integration can be ignored, since they change xp only by a solution of the homogeneous
equation. If the equations of (31) are integrated and substituted into (25) then the explicit

()-1t
formula
[X2 (t)XI (r) - Xl (t)X2 (r) If( r) d
(32) Xp t - r
to W(XI,X2, r)

is obtained, where to is any convenient value.

EXAMPLE 2. Variation of Parameters Method. Consider the inhomogeneous equation

(33) X" - 6x' + 5x = et .

It was seen above that one may take Xl = et , Xl = e5t , whence W(XI, X2, t) = 4e6t . Thus,
by (31)
, 1
V ---
1- 4'

Note that since et is a solution of the homogeneous equation, a Simpler particular integral
is xp(t) = -tet/4. Thus, by the structure theorem, the general solution of (33) is
1
x(t) = Alet + A2 e5t - 4"tet ,

where Al and A2 are arbitrary constants.

IExercises 3.21
~''C'",,>",'_ 'wVv,"")Llh",-_-,,,_v~"_Cvv __ -'_"

Wronskians and Linearly Independent Solutions. In each of the following exercises, (a) compute the
Wronskian determinant of the given solutions, (b) verify that the solutions satisfy the differential equation,
(c) apply the Wronskian test to establish linear independence of the solutions on the indicated interval.
1. XI(t) = et ,x2(t) = e- t , -00 < t < OO,X" -x = O.
2. Xl (I) = cosh(t), X2 (I) = sinh(l), -00 < t < 00, x" - X = O.

3. Xl (I) = cosh(t - to), X2(t) = sinh(t - to), -00 < t < 00 (to fixed), x" - x = O.
4. XI(t) = t,X2(t) = tlogt, 0 < t < 00, [2X" - tx' +x = O.
Abel's Formula. In each of the following exercises use Abel's formula to evaluate the Wronskian of the
given solutions Xl and X2.
 3.2 LINEAR SECONO-OROER EQUATIONS - STRUCTURE OF SOLUTIONS 163

5. 2x" + 3x' + tx = 0, Xl (0) = 0, x~ (0) = 1, X2(0) = 2, x;(O) = 4.

6. x" + 0 + t)x' + X = 0, Xl (0) = 0, x~ (0) = 1, X2(0) = 1, x;(O) = o.

Method of Reduction of Order. In each of the following exercises a linear homogeneous differential
equation and one nontrivial solution Xl (t) are given. Use the method of reduction of order to construct a
second solution X2 (t) such that XI (t) and X2 (t) are a solution basis on the interval.

9. x" +X = O,XI(t) = cost, -00 < t< 00.

10. x" + 2x' +X = 0, Xl (t) = e- t , -00 < t< 00.

11. t 2x" - txt + X = 0, XI (t) = t, 0 < t < 00.

12. (t + l)2x" - 2(t + l)x' + 2x = 0, Xl (t) = t + 1, -1 < t < 00.

13. (t _l)x" - txt +x = O,XI (1) = et , -00 < t < 1.

14. t 2x" + tx' + (t 2 - *)x = 0, XI (1) = t- 1/2 cost, 0 < t < 00.

15. 0- t 2 )x" - 2tx' + 2x = O,XI(t) = t, -1 < t < 1.

16. t 2x" - 2tx' + 2x = 0, Xl (t) = t, -00 < t < 00.

Method of Variation of Parameters. In each of the following exercises a linear nonhomogeneous equation
is given, together with a solution basis for the homogeneous equation. Use the method of variation of
parameters to find a solution of the inhomogeneous equation.

18. tlx" - 2tx' + 2x = t-l,X] = t,x2 = tl, t > O.

(Divide the equation by t2 to find a(t), bet) and f(t).)

19. t 2x" - 2tx' + 2x = t 2 log t, Xl = t, X2 = t 2 , t > O.

20. x" + X = cos t, Xl = sin t, x2 = cos t, 0 < t < n,

21. x" +x = sint,xl = sint,x2 = cost, t > O.

22. (Linearity) Let Xl (1) and Xl (t) be solutions of the linear homogeneous second-order equation (2) on
an interval to ~ t ~ tl and let Al and Al be constants. Show that the linear combination x(t) =
A IX] (t) + A2Xl (t) is also a solution of (2) on the same interval.

23. (Rockets) A skyrocket is fired vertically upward from the ground. Find the time at which it returns
to the ground if the rocket has mass m = 100 gm, thrust T = 2 x 10 5 dynes, and if it burns for
164 3 ORDINARY DIFFERENTIAL EQUATIONS OF HIGHER ORDER

to = 3 sec. Ignore air resistance. Suggestion: Let the rocket's height at time t sec be x(t) cm. Show that
if g = 980 cmlsec 2 then

x" (t) = {2000 for O:s t :s to,

-980 for t> to.

Hencex(t) is the unique solution ofthis equation that satisfies the initial conditionsx(O) = 0, x' (0) = O.

3.3 Linear Second-Order Equations with Constant

Coefficients
The class of linear second-order equations with constant coefficients can be solved
completely with at most two integrations. In many important cases the solutions may be
found by algebraic methods, without integration. These facts are derived in this section
and then applied to physical problems.
The Homogeneous Equation. The general linear second-order homogeneous
equation with constant coefficients can be written in the form

(1) x" + ax' + bx = 0,

where a and b are constants. Note that in Section 3.1, the LRC circuit equation (4),
with source term E(l) = 0, can be put in the form (1) by dividing by L. The analogous
statements are true for the mechanical oscillation equation (5) and the torsional oscillator
equation (8) of Section 3.1.
Pairs of solutions of (1) with
R
(2) a= -
L'
b=~
LC
were derived in Section 2.5 by the Riccati equation method. Three cases were found,
depending upon whether the parameter
a2 R2 1 R2 - 4UC
(3) - - b=- - - = ---::--
4 4U LC 4U
was positive, zero, or negative. The solutions may be written as follows.
a2
Case l. 4- b > 0.

(4) Xl (t) = e( -al2+fiJt, X2 (t) = e( -al2-fi Jt, f3 j - b.

= ~
a2
Case 2. - - b = 0.
4
(5)
 3.3 LINEAR SECOND-ORDER EQUATIONS WITH CONSTANT COEFFICIENTS 165

a2
Case 3. - - b < 0.
4

(6) Xl CO = e- aU2 cos yt, X2(t) = e- aU2 sin yt, y = Jb - ~.

These solutions are valid, and they give linearly independent solution pairs for all
values of the coefficients a and b.
Euler's Method. The solutions above can be found by a simple method due to Euler.
Euler looked for solutions of (1) of the form

(7)

where the parameter A was to be determined. It is easy to check that (7) defines a
nontrivial solution of (1) if and only if A is a root of the characteristic equation

(8) A2 +aA+b=0.

The three cases defined above correspond to whether the discriminant of (8), a2 - 4b,
is positive, zero, or negative.
In Case lone has a 2 - 4b > 0, and (8) has the two real distinct roots:

a a2
(9) A2 = -- -f3; f32 = - - b.
2 4

°
Equation (7) with these values of A gives the two solutions (4). In Case 2 one has
a2- 4b = and (8) has the double real root
a
(10) Al = A2 = --.
2
Equation (7) with this value of A gives the solution Xl (t) of (5). The second solution of
(5), X2(t) = t e- at12 , can then be found by the method of reduction of order. In Case 3
one has a2 - 4b < 0, so that the roots of (8) are complex and (7) does not immediately
give the solutions (6). However, in this case one can look for solutions of (1) of the form

(ll)

Then a short calculation shows that v must satisfy

(12)

By inspection, a pair of linearly independent solutions of this equation are

(l3) VI (t) = cos yt, V2 (t) = sin yt.

166 3 ORDINARY DIFFERENTIAL EQUATIONS OF HIGHER ORDER

Combining this with (11) gives the solution pair (6) of Case 3. Note that formally, Case
3 is the case where (8) has the complex roots

a
(14) A2 = -2 - iy,

Thus, in every case, the roots of Euler's characteristic equation (8) permit one to calculate
a solution basis for (1).

The Nonhomogeneous Equation. The nonhomogeneous equation corresponding

to (1) is

(15) x" + ax' + bx = f(t),

where f(t) is a prescribed function of t. The structure theorem of Section 3.2 guarantees
that the solutions of (15) have the form

(16)

where Xl, X2 is a solution basis for (1), constructed as above, and xp(t) is any particular
solution of (15). The variation of parameters method of Section 3.2 provides the integral
formula (32), Section 3.2, for xp(t). It will be shown that it takes a particularly simple form
when the homogeneous equation has constant coefficients. The three cases described
above will be considered separately
Case 1. In this case Xl(t) = d- 1t , X2(t) = d- 2t , and W(Xl,X2, t) = 0"2 - AI) eO.j +'"2)t,
where AI, A2 are defined by (9). Thus equation (32), Section 3.2, with these choices and
to = 0 gives

(17)

or

(18)

where the function k(t) is given by

(19)
 3.3 LINEAR SECOND-ORDER EQUATIONS WITH CONSTANT COEFFICIENTS 167

Substituting Al and A2 from (9) gives, after simplifying,

sinh f3t
(20) h(t) = e- atl2 _ _ (Case 1).
f3
Case 2. Substituting the solution basis (5) into (32), Section 3.2, again gives a particular
solution xp(t) of the form (18). The function h(t) is most easily calculated from (20) by
fixing a and making b -+ a2/4 so that f3 -+ O. This gives

(21) h(t) = te- atl2 (Case 2).

Case 3. Equation (32), Section 3.2, again gives a particular solution of the form (18). The
simplest way to find h(t) forthis case is to use the relation f3 = iy (see (4), (6)). This gives
sin yt
(22) h(t) = e- atl2 _ _ (Case 3).
y

Alternatively, the use of complex quantities may be avoided by using the solution basis
(6). This gives W(XI ,X2, t) = ye at and hence the trigonometric identity

sin yet - r) = sin yt cos yr - cos yt sin yr

gives

X2(t)XI (r) - Xl (t)X2(r) e- 1Ct+r ) .

(23) - - - - - - - = - - sm yet - r) = h(t - r),
W(Xl,X2, r) ye- ar
where k(t) is defined by (22). Note that in all cases, k(t) is the unique solution of (1) that
satisfies k(O) = 0 and hl(O) = 1. The integral on the right-hand side of (18) is called the
convolution of h(t) and f(t). The notation

(24) (h * j)(t) = it h(t - r)f(r)dr

is often used. One of its useful properties is symmetry; that is,

(25)

or

(26) it h(t - r)f( r)dr = it f(t - s)h(s)ds.

This is easily verified by making the change of variable r = t - s (t fixed) in the integral
of (24). The convolution will play an important part in the applications of the Laplace
transform in Chapter 4.
168 3 ORDINARY DIFFERENTIAL EQUATIONS OF HIGHER ORDER

The Method of Undetermined Coefficients. The convolution integral (18)

provides a solution of the inhomogeneous equation (15) for any integrable right-hand
side J(t). Thus, even ifJ(t) is very complicated, or is derived from experimental data, the
convolution can be evaluated by numerical integration. Nevertheless, it is of considerable
interest that for a restricted but rather large class of functions J(t), particular integrals
xp(t) can be found algebraically, without integration. The method for doing this is called
the method of undetermined coefficients. First the method will be illustrated by several
simple examples. Then the general rules for its use will be formulated.

EXAMPLE 1. Consider first the equation

(27) x" + 3x' + 2x = 3.

It is natural to try xp(t) = A = constant. Since then the derivative terms are zero, one
has a solution if lA = 3, or x~(t) = 3/2. This is clearly much Simpler than using the
convolution integral (18).

EXAMPLE 2. Next consider the equation

(28) x" + 3x' + 2x = t.

Since the derivative of a linear function is a constant, it is natural to try Xp(t) = At + B.

Substituting into (28) gives the equation 3A + 2(At + B) = (lA)t + (3A + 2B) = t,
which implies the two equations lA = 1, 3A + 2B = 0 with unique solution A = 112,
B = -3/4, and so xp(t) = At + B = (1/2)t - (3/4).

These examples suggest that the equation

(29) x" + 3x' + 2x = Pn(t),

where Pn(t) = potn + P1tn- 1 + ... + Pn is a polynomial of degree n, has a particular

solution xp(t) = Aotn + A1tn- 1 + ... + An. The (initially) undetermined coefficients are
found by substitutingxp into (29) and equating coefficients of like powers of t. Additional
examples are given in the exercises at the end of the section.

EXAMPLE 3. Now consider the equation

(30) x" + 3x' + 2x = e21 •

Since repeated differentiation of e21 gives multiples of that function, it is natural to try
xp(t) = Ae2t , where A is to be determined. Substitution gives 4Ae2t +6Ae2t +lAe2t = e2 t ,
whence llA = 1, and so xp(t) = (1I12)e2t .
 3.3 liNEAR SECOND-ORDER EQUATIONS WITH CONSTANT COEFFICIENTS 169

EXAMPLE 4. The method of Example 3 fails for the similar equation

(31) x" + 3x' + 2x = e- t .

Indeed, xp(t) = A e- t is a solution of the homogeneous equation for (31), and so it cannot
produce the right-hand side of (31). The method of variation of parameters applied to
(31) yields a solution of the form xp (t) = Ate- t . Indeed, substituting this into (31) yields
A = 1, so xp(t) = te- t

EXAMPLE 5. A more difficult case is

(32) x" + 2x' +x = e- t .

Here A = -1 is a double root of the characteristic equation, and hence both e- t and
t e- t are solutions of the homogeneous equation. Hence the trial solution xp(t) = At e- t
will not succeed. The variation of parameters method yields a solution of the form
xp(t) = At 2 e- t If this is substituted into (32), it is readily found that A = 112, so
xp(t) = (l/2)t 2 e- t .

EXAMPLE 6. A third type of function for which the method of undetermined coefficients
is effective is illustrated by the equation

(33) x" + 2x' + x = sin t.

Since differentiation of sin t gives multiples of sin t or cos t, it is natural to try xp(t) =
A sin t + B cos t with undetermined coefficients A and B. Thus x~ = A cos t - B sin t,
x; = -A sin t- B cos t = -xp, and substituting into (33) gives 2A cos t-2B sin t = sin t
or 2A cos t - (2B + 1) sin t = O. Since cos t and sin t are linearly independent, the last
equation implies that A = 0 and 2B + 1 = 0, or B = -112. Thus xp(t) = -(1/2) cos t.

Scope of the Method of Undetermined Coefficients. The examples suggest

that the method yields a particular solution of (15) when J(t) is a polynomial in t, an
exponential, or a sine or cosine. Actually, the method also is effective for functions J(t)
that are sums and products of such functions. Thus J(t) may be a sum of terms of the form

(34)

where n > 0 is an integer and p and q are real. These functions have the important
property that all their derivatives are sums of functions of the same type. Thus if a function

(35) Xp(t) = (Aot n +A1tn- 1 + ... + An)ept cosqt

+ (Botn + B1t n- 1 + ... + Bn)ept sin qt
170 3 ORDINARY DIFFERENTIAL EQUATIONS OF HIGHER ORDER

is substituted into the left side of (15) then a right-hand side of the form

(36)
J(t) = (pot + Plt n- l + ... + Pn) ept cosqt
+ (Qotn + Qltn- l + ... + Qn) ept sin qt
is generated. It is a remarkable fact that the converse of this statement is essentially true.
Exceptions occur only when some of the terms in (35) are solutions of the homogeneous
equation. This happens precisely when the real or complex number p + iq is a simple
or double root of the characteristic equation (8). Complete rules for finding xp(t) in all
cases may be formulated as follows.
Rules for the Undetermined Coefficients Method. Consider the nonhomoge-
neous equation (15) with right-hand side J(t) given by (36). Then (15) has a particular
solution xp(t) that is given by one of the following three cases.
Case 1. p + iq not a root oj (8). Then there is an xp(t) of the form (35).
Case 2. p+iq a simple root oj (8). Then the correct form of xp(t) is found by multiplying
the right-hand side of (35) by t.
Case 3. p+iq a double root oj (8). Then the correct form of xp(t) is found by multiplying
the right-hand side of (35) by t 2 .
Once the correct form of xp(t) is determined, the "undetermined coefficients" AD, A l ,
... , An, Bo, Bl , ... , Bn are calculated by substitutingxp(t) and (36) into the differential
equation and then equating coefficients of like terms. This leads to a system of 2n linear
equations for the unknowns AD, Al, ... ,An, Bo, Bl, ... , Bn.
An argument that the procedure described above always works can be based on the
Laplace transform.
Note that Case 3 above can occur only if q = O. This is because complex roots of
(8) occur only in complex conjugate pairs Al = P + iq, A2 = P - iq. The method may
also be applied when J(t) is a sum of several terms of the form (36), each with different
values of n, p, anellor q. This follows from the follOwing theorem.

The Addition Theorem. Let Xl (t) and X2 (t) satisfy

(37) X~ CO + ax~ (t) + bXl (t) = h (t),

x~ (t) + ax; (t) + bX2 (t) = h (t).

Then the sum x(t) = Xl (t) + X2 (t) satisfies (15) with J(t) = Jl (t) +h (t).
This is verified by adding the two equations of (37).
The remainder of this section illustrates the methods developed above by applications
to physical problems. Additional examples and applications are given in the exercises at
the end of the section.
 3.3 LINEAR SECOND-ORDER EQUATIONS WITH CONSTANT COEFFICIENTS 171

Applications to LRe Series Circuits. A simple LRC series circuit with voltage
source E(t) is shown in Figure 13. The charge q(t) on the capacitor C and the current
i(t) in the circuit satisfy the differential equations (2) and (4) in Section 3.1. In particular,
if E(t) is a given function of t ~ 0 and if q(O) = qo and i(O) = dq(O)!dt = io are given,
then q(t) is a solution of the initial value problem for (4), Section 3.1, namely

1
(38) Lq" + Rq' + - q = E(t) for t ~ 0,
C
q(O) = qo, q' (0) = i o·
The fundamental theorem of Section 3.2 guarantees that this problem has one, and only
one, solution. Moreover, the solution can be constructed explicitly by the methods of
this section. Here several physically distinct cases are discussed below that correspond
to different voltage sources E(t) and parameter values L, Rand C.
The general form of the solution of (38) may be obtained from (15)-(18) above with
a = RlL, b = l/LC andj(t) = E(t)IL. Thus

(39) q(t) = A1ql (t) + A2q2(t) + qp(t),

(40) qp(t) = -1
L 0
1t
k(t - r)E( r)dr,

where ql (t) and q2 (t) are any linearly independent solutions of the homogeneous
equation and k(t) is the particular solution of the homogeneous equation that satisfies
k(O) = 0 and h'(O) = 1. Note that

(41) qp(O) = 0 and q~(O) = o.

The first equation is obvious from (40). The second is proved by calculating

(42) q~(t) I1t

= -
L 0
k'(t - r)E(r) dr.

E(t)
c

FIGURE 13. An LRC circuit with voltage source E(t)

172 3 ORDINARY DIFFERENTIAL EQUATIONS OF HIGHER ORDER

The constants AI, A2 in (39) are determined by the initial conditions of (38). In fact,

(43) q(O) = A1ql (0) + A 2q2(0) = qo,

q'(O) = Alq~ (0) + A 2 q;(0) = i o.

This is a pair of simultaneous linear equations for Al andA 2 . They have a unique solution
because the determinant of the system is

(44)

and W(ql, q2, 0) f. 0 because ql and q2 are assumed to be linearly independent.

Natural Oscillations, E(t) == o. If there is no voltage source in the circuit then
qp(t) == 0 and hence
(45)

The precise forms of ql and q2 depend on the values of L, Rand C. This was discussed
in Section 2.5. Three cases were distinguished depending on whether the discriminant
D = R2 - (4UC) was positive, zero, or negative. If R > 0 and D ~ 0 then q(t) was
seen to tend exponentially to zero with at most one change of sign, while for R > 0
and D < 0 the charge q(t) was oscillatory with exponential damping. Such solutions are
called natural oscillations of the circuit. The initial state of the circuit, defined by qo
and i o, gradually decays to zero.
The Case R = O. In an idealized circuit with R = 0, called a superconducting circuit,
energy is conserved and the solution is purely oscillatory with no decay The exact form
is given by the solution basis (see (87), Section 2.5)

(46) ql (t) = cos yt, q2 (t) = sin yt, y = 1I.JLC.

DC Applied Voltage. If E(t) == Eo = constant, then
1
(47) Lq" + Rq' + - q = Eo
C
and qp can be found by the method of undetermined coefficients. In this simple case
qp = constant, and substituting in (47) gives

(48)

Thus

(49) q(t) = Alq] (t) + A 2q2(t) + CEo,

i(t) = Alii (t) + A 2i 2 (t).
 3.3 liNEAR SECOND-ORDER EQUATIONS WITH CONSTANT COEFFICIENTS 173

This differs from the case of no applied voltage (Eo = 0) only by the shifted value of q(t).
Thus q(t) oscillates about the constant charge CEo rather than about zero. The current
i(t) is unchanged.
AC Applied Voltage. If E(t) = Eo sin wt with w > ° then
,1
+ Rq + Cq =
If
(SO) Lq Eo sin wt.

This can be solved by the method of undetermined coefficients.

Case R = 0. The superconducting case will be treated first. The characteristic equation
in this case can be written as LA 2 + ~ = 0, or A. 2 + y2 = 0, where y2 = I/(LC). Thus
the rules of the method give two cases, depending on whether or not p + iq = iw is a
root of this equation.
Case 1. w =1= y = 11 JLC. Here the rules give

(SO qp(t) = Ao cos wt + Bo sin wt

and (50) may be written as

(52)
If
q + Y2 q = -Eo.SIn wt.
L
Substituting (51) into (52) gives
E
(53) (y2 _ ( 2) (Ao cos wt + Bo sin wt) = ~ sin wt.
L
Equating coefficients of cos wt and sin wt gives, since y2 - w2 =1= 0,
Eo 1
(54) Ao = 0, Bo = -L 2 2'
Y -w
and
Eo sinwt
(55) qp(t) = L(y2 _ (2) .

The general solution of (52) is

E
(56) q(t) =A1cosyt+A2 sinyt+ (2 0 2 sinwt.
L y -w)

This is a superposition of two oscillations with distinct frequencies (y =1= w).

Case 2. w = Y = II JLC. Here A. = iw is a simple root of the characteristic equation
and the rules give

(57) qp(t) = Aotcoswt + Bot sin wt.

174 3 ORDINARY DIFFERENTIAL EQUATIONS OF HIGHER ORDER

Substituting this into (52) with w = y gives, after simplification,

(58)

Thus (57) is a solution of (52) with y = w if (and only iO

Eo
(59) 2Bow = 0, - 2Aow = -
L'
whence Bo = 0, Ao = -Eo/2wL, and
Eo
(60) q(t) = Al cos wt + A2 sinwt - - - t cos wt.
2wL
Resonance when R = O. When R = 0 and w =j:. y, the solution (56) is the sum of an
oscillation with the natural frequency y = 1I.J[C of the undriven oscillator and a second
oscillation (the last term) with the frequency w of the driving voltage. The term y2 - w 2
in the denominator shows that if w is close to y then the amplitude of the output may
be much larger than that of the input. This phenomenon is called resonance. If R = 0
and w = y, so that the driving frequency coincides exactly with the natural frequency,
a more extreme form of resonance occurs. Equation (60) shows that q(t) oscillates with
ever increasing amplitude because of the factor t multiplying cos wt. Of course, this
behavior is a consequence of the assumption that R = O. In reality, superconductivity
will fail when q(t) and i(t) = dq(t)/dt become large enough.
The Case R > O. In this case p + iq = iw is not a solution of the characteristic
equation U. 2 + RA + l/C = O. For if A = iw were a solution then A2 = -w 2 and w
would satisfy Lw 2 - l/C = iwR or, on squaring both sides, (Lw 2 - l/C)2 = -w 2R2
Clearly, this is possible only if R = 0 and w 2 = 1I(LC). Thus when R > 0, the first case
of the undetermined coefficients method holds, and

(61) qp(t) = Ao cos wt + Bo sin wt.

Substituting this into (50) gives, after a short calculation,

(62) ( ( - Lw 2 + ~) Ao + RWBo) cos wt

+ (-RWAo + C-LW2 + ~)Bo) sinwt = Eosinwt.
Equating coefficients of cos wt and sin wt gives a pair of linear equations for Ao and Bo ,
namely,

(63) (-LW2 -~) Ao + (Rw)Bo = 0,

 3.3 LINEAR SECOND-ORDER EQUATIONS WITH CONSTANT COEFFICIENTS 115

(-Rw)Ao + (-LW2 +~) Bo = Eo.

The determinant of this system is

(64) Ll = (LW2 _ ~)2 +R2w 2 > 0,

and hence (63) has a unique solution. In fact, elementary algebra gives

-RwEo
(65) Ao = -Ll-'

and hence
-Rw cos wt + (_LW2 + lIC) sin wt
(66) q ( t) - E - - - - - - - - - - -
p - 0 (Lw2 _ l/C)2 + R2w2
Resonance when R > O. Notice that the point in the (x,y)-plane with coordinates

-Lw 2 + lIC
(67) y= ~

satisfies x 2 + y2 = 1, by (64). Hence there exists a unique angle ¢ such that

-Rw -Lw 2 + lIC
(68) cos¢= ~' sin¢= ~

and nl2 < ¢ < 3nl2 (since cos ¢ < 0). On substituting (68) into (66) and using the
identity

cos ¢ cos wt + sin ¢ sinwt = cos(wt - ¢),

one gets
(t) = Eo cos(wt - ¢)
(69)
qp .J(Lw2 - lIC)2 + R 2w 2
It is clear that resonance occurs at frequency w = W max such that x = w~ax maximizes
the amplitude

2 1
(70)
Y = LC

Calculus gives the derivative

176 3 ORDINARY DIFFERENTIAL EQUATIONS OF HIGHER ORDER

Two cases must be distinguished.

Case 1. y2 ~ R2/(2L)2. In this case

(71)

for all x> O. Hence A'(x) < 0 and A(x), is monotone decreasing for all x> O. There is
no resonance.
Case 2. y2 > R2/(2L)2. In this case A'(x) has exactly one positive zero at x = cv~x'
and resonance occurs at frequency

~
(72) CVmax = yy2 - W'

Figure 14 shows sample graphs of A(cv2 )/Eo versus cv for L = 1, C = 1, and several
values of R.
General Applied Voltages. Let E(t), defined for t ::: 0, be an arbitrary applied
voltage such as a signal from a transmission line. If the circuit is quiescent for t < 0, so
q(O) = 0, q'(O) = 0, then by (40) one has the circuit response

(73) q(t) = qp(t) = - lit

L 0
k(t - T)E(T)dT.

This defines the unique solution of the initial value problem for (38) with zero initial
values, for any function E(t) that is sectionally continuous for t ::: O.
Causality. Equation (73) clearly expresses the causality principle for circuits: The
state of the circuit at a time t > 0, as described by q(t) and i(t) = q'(t), is determined
by the values of the applied voltage E( T) for 0 :::: T ~ t only.
Numerical Integration. If E(t) is known only from experimental data then the
circuit response q(t) may be obtained directly from the convolution (73) by numerical
integration. Recall that k(t) is an elementary function, given by (20), (21), (22). For

o
o 0.5 1.0 1.5 2 FIGURE 14. Graphs of A(o/)IEo vs. w
 3.3 LINEAR SECOND-ORDER EQUATIONS WITH CONSTANT COEFFICIENTS 177

brevity, only Case 3 will be discussed. In this case

(74) Os R2 < 4UC,

and

(75)

Moreover, h(t) = q2 (t)/y, and hence, by the addition formula sin(a - b) = sin a cos b -
cos a sin b, one has

(76) 1 R, sin yr ) qi (t)

h(t - r) = ( -yell + (1
yell
R,
cos y ) q2{t).

Combining (73) and (76) gives the representation

(77)

where

(78) 1
VI (t) = - yL 10t( ell sin yr
R, )
E(r)dr,

1
V2 (t) = yL 10t( ell cos yr
R, )
E( r)dr.

These last two functions can be computed directly from the data E(r) by numerical
integration.
Applied Voltage of Finite Duration. If E(t) == 0 for all t 2: to > 0 then VI and V2
are constants Al and A2 for t 2: to and one has

(79)

From (77), (78) one has

(80)
1
Al = - yL
to
10 (e~ sin yr) E(r) dr,
A2
1 to
= yL 10 (e~ cos yr) E(r) dr.
Isomorphic Physical Problems. Each result obtained above for LRC series circuits
has an analogue for each of the isomorphic physical problems derived in Section 3.l.
These are the LRC parallel circuits and the linear and torsional mechanical oscillators.
To obtain the analogous result, one need only rename the variables and parameters
178 3 ORDINARY DIFFERENTIAL EQUATIONS OF HIGHER ORDER

TABLE l. Parameters of isomorphic physical systems

Series Parallel Linear Torsional
Circuit Circuit Oscillator Oscillator

q u x (J

i = dqldt v = duldt v = dxldt w = dlJldt

L C m

R lIR

C IlL 11k k

E(t) l(t) F(t) ret)

according to the following table. Some applications of these isomorphisms are given in
the Exercises below.

Euler's Method. Find solution bases for the following equations by Euler's method.

1. x" - 3x' + 2x = O. 5. 10x" + 6x' +x = O.

2. x" - 6x' + 9x = O. 6. x" + 3x' - 4x = O.

3. x" - 3x' + 25x = O. 7. 25x" + 20x' + 4x = O.

4. x" - 3x' = O. 8. x" + 2x' + 5x = O.

Method of Undetermined Coefficients. Find a particular solution of the following nonhomogeneous

equations by the method of undetermined coefficients.
9. x" + 2x' + x = cos 2 1. Use cos 2 t = i + i cos 2t.
10. 10x" - 6x' +x = 15sin2t.
11. x" +x' - 2x = e- t
12. x" - x' = et + 2e2t .
13. x" +x' - 6x = (t 2 - 4)l + e2t
Applications to Sourceless LRC Circuits. Consider the LRC circuit of Figure 13 with zero voltage source
E(t) = O. Find the general solution of the LRC circuit equation LQ" + RQ' + (l/C)Q = O.
14. Resistor of 20 ohms, inductor of 4 henries, capacitor of 0.008 farads.
15. Capacitor of 0.001 farads in series with an inductor of 0.4 henries. No resistor.
16. Inductor of 5 henries, resistor of 5 ohms, capacitor of 0.08 farads.
17. Inductor of 5 henries, no resistor, capacitor of 80000 microfarads.
 3.3 LINEAR SECOND-ORDER EQUATIONS WITH CONSTANT COEFFICIENTS 179

L
1 c

FIGURE 15. LRC circuit with switch

18. Inductor of 5 henries, no resistor, capacitor of 500 microfarads.

Applications to LRC Circuits.
Consider the LRC circuit of Figure 15. Assume that for t < 0 the switch is open, the capacitor is charged
to Ve volts and no current is flowing. At time t = 0 the switch is closed. In the exercises below, find the
transient current i(t) and plot the graph when L, R, C and Ve have the given prescribed values.
Hint: Kirchhoff's voltage law at t = 0 gives Li' (0) + Ri(O) + ve = 0 Use i(O) = 0 to get i' (0) = -veiL.

19. Compute iCt) and plot its graph if L = 0.10 H, R = 2 n, C = 20 /-LF, and Ve = 200 V.
20. Compute iCt) and plot its graph if L = 0.10 H, R = 2 n, C = 10 /-LF, and Ve = 100 V.
Applications to Mechanical Oscillators.
Consider the mechanical oscillator of Figure 16. The damper is attached to a fixed support, while the
spring is attached at its lower end to an oscillating platform. The displacements of the mass and the platform
are denoted by x(t) and y(t), as indicated in the figure. Newton's second law gives

(81) mx" = -ex' + F,

where F, the force due to the spring, is given by Hooke's law F = key - x + K), where K is a constant.
It will be assumed that the spring is unstretched when x = Xo and y = YO, so that Yo - Xo + K = 0, or

(82) F = key -YO - (x -xo)) .

.. yet)

J
platform FIGURE 16. Mechanical oscillator
180 3 ORDINARY DIFFERENTIAL EQUATIONS OF HIGHER ORDER

TABLE 2. Data set for E(t)

(0.0,0.0000), (0.1,0.3090), (0.2,0.5878), (0.3,0.8090),

(0.4,0.9510), (0.5,1.0000), (0.6,0.9510), (0.7,0.8090),
(0.8,0.5878), (0.9,0.3090), (1.0,0.0000).

If we write
(83) x(t) = xo +u
and define yet) by

(84) yet) = Yo + Ao sin wt

then combining (82)-(84) gives the differential equation

(85) mu" + cu' + ku = hAo sin wI.

21. Calculate the steady-state solution of (85),

ussCt) = Al (w) sinew!) + A2(W) cos(wt),

for arbitrary positive parameters m, c, k, Ao and w.
22. Find a formula for the steady-state amplitude

A(w) = JAl(W)2 +A2(W)2

and show that A' (w) = 0 when 2m2 w 2 = 2km - [2

23. Calculate Amax = maxA(w) on w ::": O. Give separate answers for the two cases 2km - c2 > 0 and
2km - [2 ::: O. Explain the connection with resonance.
Applications of the Convolution Integral to lRC Circuits. In the exercises below, refer to the lRC
circuit of Figure 13. The input voltage E(t) represents a signal, delivered to the circuit by a transmission
line. The goal is to compute the voltage across the capacitor,

Vet) = q(t)/C,

where q(t) is the charge on the capacitor It is assumed that q(O) = 0 and q' (0) = i(O) = 0, so that q(t) is
given by (73) and hence

Vet) = ~ t k(t - ,)E(,) dT.

LC 10
The form of the function k(t) depends on the parameters L, Rand C. The three possibilities are listed in the
convolution formulas.
24. Compute and graph the function Vet) on the interval 0 ::: t ::: n if L = 1, R = 2 and C = 0.5, while
E(t) = {1O for 0::: t ::: n/4,
o for t> n/4.

25. Graph the function Vet) on the interval 0 ::: t ::: 1.0, if L = 1, R = 2, and C = 0.5, while E(t) is the
linear interpolant of the 11 experimental data points given in Table 2. The data points are (x, sin(nx))
for x = 0 to x = 1 in steps of 0.1. Without computer assistance, use E(t) = sin(nt).
 3.3 LINEAR SECOND-ORDER EQUATIONS WITH CONSTANT COEFFICIENTS 181

Euler's Method Verification. The following exercises complete statements made in the text about the
validity of Euler's method and linear independence of the solutions.

26. Show by direct differentiation that the solutions Xl, x2 obtained for cases 1, 2, 3 indeed satisfy the
differential equation x" (t) + ax' (t) + bx(t) = 0

27. Calculate the Wronskian of the solution pair xl (t), x2 (t) given in cases 1,2,3. Verify in each case that
the Wronskian never vanishes. Apply the theory of linear independence in Section 3.2 to verify that the
solution pairs given in cases 1, 2, 3 are linearly independent.
Overdamped Spring. A spring with attached mass and fluid dashpot has its free vibration x(t) satisfying
the model

X" (t) + 3x' (t) + 2x(t) = O.

Complete the follOwing exercises.

28. Find the roots of the characteristic equation}.. 2 + 3}" + 2 = O. Use the analogy of the LRC circuit
equation to solve for x(t) = Ae- t + Be- 2 t.

29. Graph x(t) = Ae- t + Be- 2t for A = B = 1 on -2 ~ t ~ 2.

30. Graphx(t) = Ae- t + Be- 2t for A = 1, B = -Ion -2 ~ t ~ 2.
Critically Damped Spring. A spring with attached mass and fluid dashpot has model

X" (t) + 16x' (t) + 64x(t) = O.

The damping effects are in the constant 16, because if 16 is decreased by any amount at all, then oscillations
are produced. Therefore,

X" (t) + 15x' (I) + 64x(t) = 0

is underdamped and it exhibits oscillations. Complete the following exercises.

31. Show by analogy with the LRC circuit models that the solutions of these two equations are respectively
fl (0 = (A + BOe- 8t and
h(t) = (Acos(51U2) + Bsin(51U2))e- 7.5t.
32. Graph the solutions on a Single plot for A = B = 1 and 0 ~ t ~ 1. Where do the graphs cross?

33. Plot the general solutions of x"Ct) + 16.5x'Ct) + 64x(t) = 0 and x"(t) + 17x'(t) + 64x(t) = 0 for
A = B = 1 and 0 ~ t ~ 1. Where do they cross?
Underdamped Spring. A spring with mass and fluid dashpot usually displays sustained oscillations of
decreasing magnitude. A characteristic model is

X" (t) + 2x' (t) + 2x(t) = O.

Complete the follOwing exercises.

34. Solve the differential equation by analogy with the LRC circuit equation.

35. Consider the particular solution x(t) = e- t sin(t). Graph this solution for 0 ::s t ::s T, where T is the
solution of e- T = 11100 (then x(T) is 1% of the maximum amplitude 1). How many times does x(t)
cross the t-axis on 0 < t < T7
182 3 ORDINARY DIFFERENTIAL EQUATIONS OF HIGHER ORDER

3.4 Linear Second-Order Equations with Analytic

Coefficients
This section develops solution methods for linear second-order homogeneous equations

(1) x" + a(t)x' + b(t)x = °

whose variable coefficients aCt) and bet) have convergent Taylor series expansions. This
class, although quite restrictive, includes some of the most important equations of
mathematical physics. The principal method for solving these equations is the power
series method.

The Power Series Method. This method seeks to construct solutions x(t) of
equation (1) as the sums of convergent power series:

L [k(t -
00

(2) x(t) = to)k, It - tol < r.

k=O

Taylor's theorem implies that the series above must be the Taylor series of x(t), so that

(3) [k = X(k) (to)!k! for k = 0,1,2, ....

Functions x(t) with such a representation are said to be analytic at to. The basic fact
concerning the power series method is the follOwing:

The Power Series Theorem. Let the coefficients aCt) and bet) in (1) be analytic at to:

L ak(t - L bk(t -
00 00

(4) aCt) = to)k, bet) = to)h

k=O k=O

for It - tol < r. Then for each choice of the constants Co = x(to) and [1 = x/(to) the
equation (1) has an analytic solution (2). The solution may be constructed by substituting
the series (2) and (4) into (1) and calculating the coefficients [2, [3, ....

Proofs of this theorem may be found in advanced textbooks such as [B-RJ. No general
proof will be given here. However, the theorem will be verified in several important
special cases.

Rules for Power Series. The application of the power series method depends on
three simple rules for calculating with series. The proofs of these rules, which are based
on Taylor's theorem, will be left for the exercises.
 3.4 LINEAR SECOND-ORDER EQUATIONS WITH ANALYTIC COEFFICIENTS 183

The Differentiation Rule. The power series representation (2) can always be
differentiated term by term to get a power series representation for the derivative:
00

(5) x' (t) = L kCk(t - to)k-l.

k=O

Moreover, the derived series converges in the same interval It - tol < T as for x(t).
The Product Rule. Let aCt) and bet) be analytic at to with power series (4) convergent
for It - to I < T. Then a(t)b(t) is analytic at to and
00

(6) a(t)b(t) = LYk(t - to)k, It - tol < T,

k=O

with coefficients
k
(7) Yk = Laebk- e, k=0,1,2, ....
£=0

Note that the coefficients (7) are those that result from the formal multiplication of the
series (4) for aCt) and bet)
The Vanishing Rule. If the sum of the power series in (2) is zero for It - tol < T

°
then the coefficients Ck = for k = 0, 1,2, ....
The power series method is applied below to the Bessel and Legendre equations.
First, the method will be illustrated by application to two simpler examples.

EXAMPLE 1. Consider the equation with constant coefficients

(8) x" - 2x' + x = 0.

The method of Section 3.3 provides the solution basis Xl(t) = ct , X2(t) = tc t . These
solutions will be rederived here by the power series method.

For (8) the constant coefficients aCt) = -2, bet) = 1 are analytic at to for every to.
Taking to = °gives the trial solution
00

(9) x(t) = L Cktk.

h=O

Two applications of the differentiation rule give

00 00

(10) x'(t) =L kCktk-l = I)k + l)ck+l tk

k=l k=O
184 3 ORDINARY DIFFERENTIAL EQUATIONS OF HIGHER ORDER

and
00 00

(ll) x"(t) = L k(k + 1)Ck+lt k- 1 = L(k + 1)(k + 2)ck+2 tk .

k=1 k=O

The second summations in (10) and (11) are obtained from the first by increasing the
summation index from k to k + 1. Verify that the two summations represent exactly
the same power series. On substituting the series for x(t), x'(t), and x"(t) into (8) and
combining the three sums, one gets
00

(12) L [(k + l)(k + 2)ck+2 - 2(k + l)ck+l + ckl tk = o.

k=O

By the vanishing rule, this equation implies that the coefficient of each power tk must
equal 0, so that for k = 0, 1,2, ... ,

(13) (k + 1)(k + 2)Ck+2 - 2(k + 1)ck+l + Ck = O.

Solving for Ck+2 gives the recursion relation

(14) C = 2Ck+l _ Ck k=0 1 2

k+2 k+2 (k + 1)(k + 2) , , , , ....

Clearly, when Co and CI are given then (14) with k = 0 gives C2. Next, (l4) with k = 1
gives C3, etc. Thus all the Ck for k 2: 2 are generated from Co and CI by (l4).
To generate the solution XI (t) = et (so x~ (t) = et ), one must take Co = XI (0) = 1
and Cl = x~ (0) = l. Then (14) with successive values h = 0,1,2, . .. gives
2Cl 1 1
C2 = -
Co
- - =-
2 2 2!'
C3 = -2C2
3
- CI
--
2·3
= -
3!'
etc.

It is easy to guess that

1
(15) Ck = k! for k = 0, 1,2,3,4, ....

The correctness of this equation for all k 2: 0 may be verified by mathematical induction.
Thus if (15) holds for C£ and CHI then (14) gives
2 1 1 1 1
C£+2 = £ +2 (£ + 1)! - (£ + 1)(£ + 2) £! = (£ + 2)!'
This completes the induction. Finally, substituting (15) into (9) gives the familiar
exponential series
00 tk
(16) x(t) = " - = et =Xl(t).
~ h!
k=O
 3.4 LINEAR SECOND-ORDER EQUATIONS WITH ANALYTIC COEFFICIENTS 185

The second solution X2 (t) = t if (so x~ (t) = (t + I)e t ) can be obtained by the same
method from (14) and the initial values Co = X2(0) = and C1 = x~(O) = l. °
Example 1 was simple because (8) has constant coefficients. A more interesting case
is the following.

EXAMPLE 2. Consider the equation

(17) x" + 2tx' + 2x = 0.

This has the form (1) with variable coefficients aCt) = 2t and bet) = 2, which are analytic
at all points to. Taking to = ° gives
oc
(18) x(t) = L Ck tk ,
k=O
oc
x'(t) = L kCktk-1,
k=O
oc
tx'(t) = L kCk tk ,
k=O
oc oc

X"(t) = L k(k - I)ck tk - 2 = L(k + 2)(k + l)ck+2 tk .

k=O k=O
Substitution into (17) gives

L
00

(19) [(h + 2)(h + 1)Ch+2 + 2hch + 2chl tk = o.

k=O
Thus, by the vanishing rule, the coefficients must satisfy

(20) (k + 2)(k + I)ck+2 + 2(k + l)q = ° for k = 0,1,2, ... ,

or

(21) Ck+2=(k~\)(k for k=0,I,2, ....

Consider first the solution with Co = 1, C1 = 0. Then (21) with k = 1,3,5, ... gives
C1 = (3 = C5 = ... = 0; that is,

(22) C2k+! = ° for k = 0,1,2, ....

Next, (21) with k = 0,2,4, ... gives C2 = -1, C4 = -(212 = 112, C6 = -I/3!, etc. It is
easy to guess that

(23) C2k = (-l)klk! for k = 0,1,2, ....

186 3 ORDINARY DIFFERENTIAL EQUATIONS OF HIGHER ORDER

To verify this, note that (21) with k = 2l and (23) with k = £ give

(24) C2£+2 = ( -2)

+
2l 2 C2e =
( -1+ )
£ 1
(_1)f
-£!-
(_1)£+1
(£ + 1)!
Hence the induction is complete and (23) is valid. Finally, substituting (22) and (23)
into x(t) of equation (18) gives a first solution of (17), namely
(_t 2 )k
=L =L
00 00
(25) Xl (t) C2k t2k -k-'- = e- t2
,
k=O k=O'

where the series was summed by the familiar exponential series. It is easy to verify, by
direct differentiation, that Xl (t) = e- t2 is a solution of (17).

To find a solution basis for (17), a second solution X2 CO with initial values Co = 0,
Cl = 1 will be constructed. Note that Co = ° and (21) imply that

(26) C2k=0 for k=0,1,2, ....

Moreover, after calculating C3 = -2/3, C5 = 22/3.5, C7 = -2 3/3 . 5 . 7, etc., one can

guess, and then verify by induction, that for k = 0, 1,2, ... ,
C-2)k C-4)kk!
(27)
C2k+l = 1 .3 . 5 .... e2k + 1) = e2k + I)! .
This gives the solution
00 k! (_4t 2 )k
(28) X2(t) = t {; (2k + 1)! .

It is easy to show that X2 (t) is not an elementary function. In fact, the method of reduction
of order gives a second solution

(29)

where by (23), Section 3.2, with aCt) = 2t one has

(30) vet) = it Xl (i)-2 e-r 2dt = it e2r2 . e- r2 di = it er2 di,

whence

(31) X2 (t) = e- t2 1 t r2
e di.

This is the solution whose Taylor series is (28). In fact, from (31),

(32)
 3.4 LINEAR SECOND-ORDER EQUATIONS WITH ANALYTIC COEFFICIENTS 187

Hence, both (28) and (31) define solutions of (17) that have initial values X2(0) = 0,
x;(O) = 1. It follows that they are identical, by the uniqueness part of the fundamental
theorem.
The Legendre Equation. The Legendre equation (45) of Section 3.1 can be written

"lx, J.l 0
(33)
Y - I - x 2Y +1- x 2Y = .

Thus it has the form (1) with nonconstant coefficients

-2x J.l
(34) a(x) = --2' b(x) = --2'
I-x I-x
These functions are analytic at all points Xo except the two points Xo = 1 and Xo = -1.
Here the power series method will be used to construct a solution basis near Xo = O.
The solutions will have the form
00

(35) y(x) =L Ck xk ,
k=O
so
00

(36) y'(x) = LkckXk-l,

k=O
00

(37) xy' (x) = L kCkX h ,

k=O

2: k(k -
00 00

(38) y"(x) = 1)Ckxk-2 = 2:(k + 2)(k + l)Ck+2Xk,

k=O k=O
00

(39) x 2y" (x) = L k(k - 1)ckxk.

k=O

Thus, rewriting (33) in the more convenient form

(40) (l - x 2)y" - lxy' + J.lY = 0

and substituting the series (35)-(39) gives, after collecting terms,

(41) [(k + l)(k + 2)ck+2 - (k(k + 1) - J.l) ChI xk = 0,

The vanishing rule gives the recursion formula

k(k + 1) - J.l
(42) CH2 = (k + 1)(k + 2) Ck, k = 0, 1,2, .. , ,
188 3 ORDINARY DIFFERENTIAL EQUATIONS OF HIGHER ORDER

Clearly, once Co and Cl are chosen, the remaining coefficients Ck are uniquely determined.
Two cases will be discussed.
Case 1. (co = 1, Cl = 0). In this case, (42) with odd indices gives C2k+! = 0 for k = 0,
1, 2, ... and so one solution of (40) is

L
00

(43) Yl (x) = C2kX2k,

k=O

where
(2k)(2k + 1) - J.l
(44) C2k+2 = (2k + 1)(2k + 2) C2k for k = 0, 1,2, . ...
This relation implies that if no C2k = 0 then

(45) · -
11m -= 1
C2k+2
k-->oo C2k

and hence by the ratio test, the radius of convergence of (43) is 1; that is, Yl (x) is defined
by (43) for Ixl < 1. The successive coefficients C2k are determined by Co = 1 and (44).
Thus
(-J.l)
C ---
2 - 1.2'
2·3 - J.l (-J.l) (2 . 3 - J.l)
C4 = 3 . 4 C2 = 4!
It is not difficult to prove by induction that
(-J.l)(2·3 - J.l)(4· 5 - J.l) ... «(2k - 2)(2k -1) - J.l)
(46) C2k =
(2k)!
Case 2. (co = 0, Cl = 1). This second choice gives C2k = 0 for k = 0, 1, 2, ... and
hence a second solution of (40) is

(47)

where
(2k - 1)(2k) - J.l
(48) C2k+l = (2k)(2k 1) C2k-l+ for k = 1,2,3, . ...

Note that W(Yl ,Y2, 0) = Yl (O)Y~ (0) - Y2 (O)Y~ (0) = 1 and hence Yl ,Y2 are a solution
basis for the Legendre equation. If no C2k+! = 0 then

(49)
· -
11m -= 1
C2k+l
k-->oo C2k-l
 3.4 LINEAR SECOND-ORDER EQUATIONS WITH ANALYTIC COEFFICIENTS 189

and hence the radius of convergence of (47) is l. The successive coefficients C2h+1 are
determined from C1 = 1 and (48). Thus

1·2-11
2·3
3·4 - 11 (1 ·2 - 11)(3 . -11)
C5 = 4.5 C3 = 51

By induction one gets the formula, for k ::: 1,

(l . 2 - 11)(3 ·4- 11) ... «2k - 1)(2k) - 11)

(50) =
C2k+l
(2k + 1)1
Legendre Polynomials. There is one exception to the radius of convergence calcu-
lations for the series solutions Y1 (x) and Y2 (x). Note that by (42), if 11 = n(n + 1) is
the product of two successive nonnegative integers then Cn+2 = 0, CnH = 0, etc. In par-
ticular, if 11 = (2n) (2n + 1) then Y1 (x) = P2n (x) = polynomial of degree 2n, while if
11 = (2n + 1) (2n + 2) then Y2 (x) = P2n+ 1 (x) = polynomial of degree 2n + 1. It will
shown below that the Legendre equation has a solution that is finite at x = ± 1 if and
only if the parameter 11 = n(n + 1), where n = 0, 1,2, . ...

Equations with Singular Points. Equation (1) is said to have a singular point
at to if one or both of the coefficients aCt), bet) is not analytic there. Thus the Bessel
equation (30) of Section 3.1 has a singular point at Xo = 0 and the Legendre equation
(33) has Singular points at Xo = ±l. Moreover, for applications it is often important to
discover the behavior of the solution of the differential equation near such points. All of
the examples given ahove can be treated by the theory of linear second-order equations
with regular singular points. The theory is based on the following definition.
Definition. A linear second order equation (1) is said to have a regular singular point
at t = to if it can be put into the form

a(t) , bet)
+--x +
/I
(51) x x=o,
t - to (t - toF

where aCt) and bet) are analytic at t = to.

Clearly, the Bessel equation has a regular Singular point at Xo = 0, while the Legendre
equation has regular singular points at Xo = 1 and Xo = -1.

Euler Equations. Equation (51) is said to be an Euler equation if aCt) = ao and

bet) = bo are constants. It will be shown that such equations can always be solved
explicitly in terms of elementary functions. It will suffice to treat the case to = 0 and
190 3 ORDINARY DIFFERENTIAL EQUATIONS OF HIGHER ORDER

write the equation as

(52) x"+ -x
£10 '+ -x=
bo 0
t t2 '

or, equivalently,

(53)

Note that in the examples from physics of Section 3.1, equations (33) and (41) are Euler
equations. Euler's method for solving (53) is to look for solutions of the form x(t) = ItI",
where the index a is to be calculated. For simplicity it will be assumed here that t > 0,
so that

(54) x(t) = t = eOilogt,

Oi (t> 0).

Then

(55)

and similarly,

(56) x" (t) = a(a - 1)tOl - 2.

On substituting these into (53) one gets

(57) [a(a - 1) + aoa + bol eOilogt = o.

Since eOi log t =1= 0, it is clear that (54) is a solution of (53) if and only if a satisfies the
indicial equation

(58) a 2 + (£10 - l)a + bo = O.

Three cases arise, depending on whether the roots of this equation are real and distinct,
real and equal, or complex conjugates. The results may be formulated as follows.
Case 1. (al and a2 real and distinct). Then (53) has the solution basis

(59)

Clearly, the ratio of the two solutions is not constant if al =1= a2, and hence they are
linearly independent.
Case 2. (al = a2 = a real). Then (53) has a solution basis

(60)

It is clear that t Oi is one nontrivial solution. The second can be obtained from the first by
the method of reduction of order.
 3.4 liNEAR SECOND-ORDER EQUATIONS WITH ANALYTIC COEFFICIENTS 191

Case 3. (al = p + iq, a2 = p - iq, q > 0). Then (53) has a solution basis

(61) Xl (t) = tPcos(q log t), X2(t) = tP sin(q log t).

These are obtained from Euler's solution
X(t)= e(p+iq)logt = eplogt . eiqlogt
(62)
= tP (cos(q log t) + i sin(q log t))
by separating into real and imaginary parts. The solutions are linearly independent
because X2(t)!Xl (t) = tan(p log t) is not constant.

EXAMPLE 3. The search for product solutions u(r, ()) = R(r)8«()) of the Laplace equation
in polar coordinates, (31), Section 3.1, led to the Euler equation (33) (Section 3.1); i.e.,

(63)

Putting R(r) = r a gives the indicial equation

(64)

with roots al = n, a2 = -no Thus one has the general solution

R(r) = /Alr n +A 2 r- n (Case 1, n > 0),

(65)
Al + A 2 10g(r) (Case 2, n = 0).

Note that limHo R(r) is finite only if A2 = O.

The Method of Frobenius 8 • If (51) has a regular singular point at to but is not
an Euler equation (so aCt) and bet) are not both constants), then Euler's substitution
x(t) = It - tol a does not give a solution. For this more general case Frobenius proposed
the trial solution

L Ch(t -
00

(66) x(t) = (t - to)a to)k, Co f. 0,

k=O

where, for simplicity, it is assumed that t > to. This clearly generalizes the power series
method (where a = 0) and the Euler method (where Ck = 0 for k = 1,2, .. .). The
idea is to substitute (66) and its derivatives into (51) and to determine, in succession,
the values of the exponent a and the coefficients co, C1, C2, .... Here the main steps of
the calculation will be written down to display the equations by which a and the Ck are
determined. For brevity, only the case to = 0, t > 0 is shown.

BAfter EG. Frobenius, German mathematician 0849-1917).

192 3 ORDINARY DIFFERENTIAL EQUATIONS OF HIGHER ORDER

Equation (51) with to = 0 is equivalent to the equation

(67) t 2x" (t) + ta(t)x' (t) + b(t)x(t) = 0

and aCt) and bet) are assumed to be analytic at to = 0:

00 00

(68) aCt) = L aktk, bet) = L bktk.

k=O k=O
The Frobenius trial solution has the form (for t > 0)
00 00

(69) x(t) = t a L Ck tk = L Ck tk+a .

k=O k=O
Moreover, the differentiation rule gives
00

(70) x'(t) = L(k + a)ck tk+a-\

k=O
00

(71) x"(t) = L(k + a)(k + a - 1)cktk+a-2,

k=O
and hence
00 00

(72) tx' (t) =L (k + a)ck t k+a == L c~tk+a

k=O k=O
and
00 00

(73) t 2x"(t) = L(x + a)(k + a - l)cktk+a == L c~tk+a,

k=O k=O
where for brevity,

(74) c~ = (k+a)ck, = (k+a)(k+a -1)ck, k = 0, 1,2, ....

c~

Moreover, the product b(t)x(t) is analytic at to = 0 with power series obtained by the
product rule (6):

(75)

Similarly,

(76)
 3.4 LINEAR SECOND-ORDER EQUATIONS WITH ANALYTIC COEFFICIENTS 193

If the series in (69) is to define a solution of the differential equation (67) then the sum
of the series (73), (76), and (75) must be zero:

(77)

By the vanishing rule this implies that the coefficients are all zero:

=°
k k
(78) C~ + L ak-jC; + L bk-jCj for k = 0, 1,2, ....
j=O j=O

In particular, taking k = 0, c~ = aco, c~ = a(a - 1)c0 gives

(79) [a(a - 1) + aoa + bol Co = 0.

Since Co i= °
by assumption, the exponent a must satisfy the indicial equation

(80) Ind(a) =a(a - 1) + aoa + bo = 0.

Note that if (67) is an Euler equation, so that aCt) = ao, bCt) = bo, then (80) is the same
equation as in Euler's method.
Next note that for k > 0, equations (74) and (78) give

(81) [(a + k)(a + k - 1) + ao(a + k) + bol Ck

k-1
+ LCa + j)Cjak-j + cjb k- j = 0, k = 1,2, ....
j=O
If the indicial polynomial Indea) is defined by (80) then (81) can be written as the
recursion formula
k-1
(82) Ind(a + k)ck =- L [(a + j)ak-j + bk - j] Cj, k = 1,2, ....
j=O
This determines C1, C2, •.. successively as multiples of Co (i= 0), provided that Ind(a +
°
k) i= for k = 1,2, .... Now, a is one of the two roots of the indicial equation (80). In
°
particular, Ind(a + k) i= for k = 1,2, ... if a is the larger of two real roots or if a is
either of the two complex roots. This leads to the following result.

Main Frobenius Theorem. Let a be the larger of two real roots of the indicial equation
or either of two complex roots. Then for each Co i= 0, equation (67) has a nontrivial
solution (69) with coefficients generated by the recursion relation (82).
194 3 ORDINARY DIFFERENTIAL EQUATIONS OF HIGHER ORDER

To complete the proof of this theorem one need only prove that the series (69)
generated by the recursion (82) must converge. Such proofs require techniques of
advanced calculus. They may be found in advanced texts such as [B-RJ.
Of course, if the roots of the indicial equation do not differ by an integer or
are complex, then both roots generate solutions, and a solution basis is obtained.
However, the Main Frobenius Theorem suffices for applications because a second, linearly
independent, solution can always be constructed by the method of reduction of order.
These techniques are illustrated in the remainder of this section by application to the
Bessel and Legendre equations.
Bessel Functions. Only Bessel functions of integer order n ~ 0 will be discussed.
For the application of the Frobenius method it will be convenient to write the Bessel
equation (30) of Section 3.1 in the alternative form

(83)

Here the coefficients a(x) and b(x) of (51) are a(x) == 1 and b(x) = _n2 + x 2, so in (68)
ao = 1, bo = _n 2 , and the indicial equation (80) is
(84)

with roots al = nand a2 = -no Thus the main Frobenius theorem guarantees the
existence of a solution
00 00

(85) y(x) = xn L CkX k = L CkX k+ n .

k=O k=O
Application of the differentiation rule gives
00

(86) xy' (x) = L (k + n)ck xk+ n

k=O
and
00

(87) x 2y"(x) = L(k + n)(k + n - l) ckx k+ n .

k=O
Moreover, one can write
00 00

(88) x 2y(x) = L Ck xk + n+ 2 = L Ck_ 2Xk+ n ,

k=O k=O
if it is agreed that Ll = L2 = O. Finally,
00

(89) - n 2y(x) = L (-n2ck}xk+n.

k=O
 3.4 LINEAR SECOND-ORDER EQUATIONS WITH ANAlYTIC COEFFICIENTS 195

Now, the Bessel equation (83) implies that the sum of the last four displayed equations
is zero:

L [(k + n)(k + n -
00

(90) 1)ck + (k + n)ck - n2ck + Ck-2] x n+k = O.

k=O

Hence all the coefficients in the square brackets must vanish by the vanishing rule. These
conditions can be written as

(91) [(k + n)2 - n 2] Ck + Ck-2 = 0 for k = 0, 1,2, ... ,

where C2 = Cl = ° and Co 1= O. For k = 0 this is satisfied because C2 = O. For k = 1

one has

(92)

because Cl 0, and hence C1 = O. It then follows from (91) that C2k+l °for
k = 0, 1,2, ... , and (85) takes the form

L C2kX2k+
00

(93) y(x) = n.

k=O

Moreover, the recursion (91) with k replaced by 2k + 2 can be written as

-C2k
(94) C2k+2 = 22(k + l)(k + n + 1) for k = 0, 1,2, ....

This determines the coefficients C2, C4, ... uniquely when Co is given. A short calculation
gives
-Co
C2 = -22-=-(-1-!)-Cn-+-1-) ,
Co
C4 = -24-C-2-!)-(n-+-1-)C-n-+-2)'
-Co
C6 = -26-C-3-!)-Cn-+-1-)(-n-+-2)-(n-+-3-r
This suggests the general fonnula

( -1) kco
(95) C2k = 22kk!Cn + 1)(n + 2) ... (n + k)' k = 1,2, ... ,

which is easy to check by induction. The standard choice of Co is Co = 1I(2 nn!), which
gives the formula
(_l)k
(96) C2k= 2 2k +n k!(k+n)!' k=0,1,2, ....
196 3 ORDINARY DIFFERENTIAL EQUATIONS OF HIGHER ORDER

The corresponding solution is the standard Bessel function of order n, denoted by

Yl (x) = In(x) , where by definition,

00 (_l)k (x12)2k+n
(97) In(X)=~ k!(k+n)! ' n=0,1,2, ....

It is easy to verify by the ratio test that this series converges for all finite x. Note in
particular the important speCial case

00 (_1)k(x12)2k
(98) lo(x) =~ (k!)2

Other Solutions of the Bessel Equation. The method of reduction of order applied to
the Bessel equation and the particular solution Yl (x) = In(x) yields a second linearly
independent solution (if Xo is any positive constant):

(99)

See (16) and (23) in Section 3.2. Thus Yl (x) = In (x) and Y2 (x) are a solution basis, and
every solution of the Bessel equation can be written Y = AnYl + BnY2, or

(100)

where An and Bn are arbitrary constants. The theory of these functions is developed in
standard references, e.g., [A-SJ. Here only the following fact, needed for applications in
Chapter 8 and Chapter 9, will be presented.
Finiteness Property for In. The solution (l00) of the Bessel equation has a finite limit
as x ~ 0 if and only if Bn = O. In other words, the only solutions of the Bessel equation
that are finite at x = 0 are the scalar multiples of ln (x).
It is clear that if Bn = 0, then yeO) = AJn (0) is finite. To verify the converse it will
be enough to show that

(101) . l
hmln(x)
x...,. a x
Xo
~ = +00.
- 2-
~n (~)

Consider first the case n = O. By equation (98),

(x12) 2 (X/2)4 (x12) 6

lo(X) = 1- (1!)2 + (2!)2 - (3!)2 + ...
 3.4 LINEAR SECOND-ORDER EQUATIONS WITH ANALYTIC COEFFICIENTS 197

is a convergent alternating series. The Leibnitz theorem for alternating series implies that
for x small enough, 0 < ]o(x) :::: 1, and hence
1 1
-- > - for 0 < x :::: Xo.
x]~(x) - x

Integration gives

(102) [
xO d~
-2-::::
[xO - d~
=logxo-logx=log(-).
Xo
x ;Jom x ~ x
Since]o(O) = 1, this implies (101) with n = O. Proceeding similarly for n :::: 1 one has

1 (x)n [(xJ2)2 (xJ2)4 ]

(103) ]n(x) = n! 1- n+1 2: + 2!(n + l)(n + 2) - ... ,

and hence 0 < ]n(x) :::: (xJ2)n/n! for x small enough, whence
X2n+l
X]2 (x) < ---::--::-
n - (n!)222n

and
1 22n(n!)2 1
-->--->--
x]~(x) - X 2n + 1 x2n+l

for 0 < x :::: Xo == xo(n). Integrating over (x, xo) gives

(104) l x
Xo d~
;J~m
> 1 ( 1
- 2n x2n - x~n
1) for 0 < x :::: Xo.

For Xo small enough, Om) implies that

(l05) ]n(x) :::: :! Gr (~) for 0:::: x < Xo·

On multiplying the last two inequalities, one has

[Xo d~ ( 1 xn)
(106)
]n(x) lx ;J~(~):::: Yn xn - x~n for 0< x:::: xo,
where Yn = 1/[(n!)(2n+l)(2n)] > O. Making x --+ 0 in this inequality gives (101) for
n:::: l.
Behavior of ]n(x) for x --+ +00. Another approximation that gives inSight into the
nature of the function]n(x) is defined by the equation

(l07) ]n(x) = y(2 ( mr 7r:)

~ cos x - 2 - 4" + en(x).
198 3 ORDINARY DIFFERENTIAL EQUATIONS OF HIGHER ORDER

Of course, this may taken as a definition of the function cn(x). The equation is made
interesting by the statement that one has the estimate

(108)

where Kn is a suitable constant. The known proofs of (l08) and estimates of the constant
Kn are not elementary. They may be found in advanced texts such as [B-RJ.
The estimate (108) implies that for large x the function 1n(x) is approximated by
the elementary function on the right-hand side of (l07). Moreover, the approximation
improves with increasing x. Such approximations are called asymptotic estimates.
Estimate (l08) implies that1n(x) is an oscillatory function for large x, with infinitely
many zeros. Moreover, the large zeros of 1n ex) are spaced about rr units apart. In fact,
the zeros of10(x) are close to those of cos(x -rr/4), even for relatively small values of x.
This will be explored numerically in the exercises.
Recurrence Relations. The values of Bessel functions and their first derivatives are
closely related. The most basic relations are

(l09) d n ) n
dx [x 1n(x J = x )n-1 (x)

and

(1l0)

These equations are verified by substituting the Frobenius series for 1n (x) on the left-
hand sides and checking that the resulting series are those of the right-hand sides. From
(109), (110) one can easily derive the recurrence relations for n 2: 1:
2n
(lll) - 1n(x) = 1n-1 (x) + 1n+l (x)
x
and

(1l2)

The details are left for the exercises.

Integral Representation. An integral representation for 10 (x) was introduced in Chap-
ter 1; see (2) in Section 1.2. A corresponding representation for the Bessel function of
integer order n 2: 0 is

(1l3) 1n(x) = -1 171: cos(x sin t - nt)dt.

rr 0
 3.4 LINEAR SECOND-ORDER EQUATIONS WITH ANALYTIC COEFFICIENTS 199

It is not obvious that this integral defines the same function as the Frobenius series (97).
This can be verified by expanding the integral in a power series in x and comparing it
with (97), but the calculation is very long. A more efficient method can be based on
Fourier analysis.
Numerical Evaluation of ln (x). Bessel's functionln (x) is a complicated function of the
two variables x and n, and no single computational algorithm can be effective for all
values. For small x and moderate n, truncation of the alternating series (97) suffices. This
is explored below in the exercises. In Chapter 1 the integral representation was evaluated
accurately by numerical integration for moderate values of x. This also is effective for ln (x)
with moderate x and n values. For large x and moderate n the asymptotic approximation
and its refinements are accurate. The recursion relations are also useful for computing
ln (x) and 1~ (x); see [A-51 for a useful discussion. In recent years sophisticated software
for computing Bessel functions has been developed that makes these functions almost
as easy to compute as the trigonometric functions.

Legendre Functions. The power series method was used above to construct a
solution basis Y1, Y2 of the Legendre equation

(114) (1 - X 2 )y" - 2xy' + ILY = O.

These series solutions, defined by (43), (46) and (47), (50), respectively, converge on the
interval -1 < x < 1 only and have no limits when x ~ ± 1. The only exceptions occur
when f1. = nCn + 1), with n = 0,1,2, .... In this case one of the solutions reduces to
a polynomial. Here the Frobenius method will be used to show that for any value of f1.
there is a nontrivial solution of the Legendre equation that is finite at x = 1. However,
this function will have no limit when x ~ -1, except when IL = n(n + 1). Hence these
IL values are the only ones for which the Legendre equation has a solution that is finite
on the closed interval -1 :::: x :::: 1.
The Legendre equation can be written in the form

II a(x) I b(x)
(115) Y +x - I Y + (x _ I)2Y = 0,

where

2x
(116) a(x) = - - = 1
x+I
+ a1 (x - 1) + a2(x - 1)
2
+". '
b(x) = f1. (I-X)
I+x
-- = 0 + b1(x -1) + b2 (x - 1)2.
200 3 ORDINARY DIFFERENTIAL EQUATIONS OF HIGHER ORDER

Hence ao = 1, bo = 0, and the indicial equation at Xo = 1 is a(a - 1) + a == a 2 = 0

with double root a = O. Thus the main Frobenius theorem guarantees the existence of
a nontrivial solution y + (x) of the form
00

(117) y+(x) = L::>k(X - l)h.

h=O

Therefore,

Lk
00

(118) y~(x) = Ch(X - 1)k-l

h=O

and

L k(k -
00

(119) y:(x) = 1)ck(X - 1)k-2.

k=O

It will be convenient to substitute these series into the Legendre equation in the form
(114). Note that

(120) (1 - x 2 ) = - (x - 1)2 - 2(x - 1),

-2x = - 2(x - 1) - 2,
whence

L [k(k -
00

(121) (l - x 2 )y: (x) = - l)ck + 2k(k + l)ck+ll (x - l)k

k=O

and

L [kCk + (k + l)ck+d (x -
00

(122) - 2xy~(x) = -2 l)h.

k=O

Substituting these into (114) gives, after some simple algebra,

L [(JL -
00

(123) k(k + 1)) Ch - 2(k + 1)2ch+l] (x - l)k = O.

h=O
Thus, by the vanishing rule, the coefficients Ch are determined by Co = 1 and the recursion
relation
JL - k(k + 1)
(124) Ch+l = 2(k + 1)2 Ch, k = 0, 1,2, ....
Note that the series (117) terminates, and y+(x) = Pn (x) is a polynomial of degree n, if
and only if JL = n(n + 1) en
= 0, 1, 2, ... ). In all other cases (117) is an infinite series
 3.4 liNEAR SECOND-ORDER EQUATIONS WITH ANALYTIC COEFFICIENTS 201

and (24) implies that

(125) lim
k.... oo
ICk+ll
Ck
=~.
2

Thus (17) has radius of convergence 2 and converges for - 2 < x-I < 2, or -1 <
x < 3. Note that any solution of (114) that is linearly independent of y+ will tend to 00
at x = 1. In fact, the reduction of order method produces all such solutions and (6),
(23) in Section 3.2 with aCt) = -2U(1 - t2 ) gives the solution

(126)

The technique used to prove (01) above can be used to show that

(127) [y(x) [ ~ y log (_1_)

I-x
for xo::: x < 1,
where y > 0 and Xo is close to 1. Hence, all solutions of the Legendre equation, other
than constant multiples of y+(x), satisfy limHl Iy(x) I = +00. Finally, a careful analysis
of the series (117) for y+(x) yields the following estimate (difficult!):

1
(128) [y+(x) [ ~ y log - - for - 1 < x::: -Xo.
x+l

This implies that the Legendre equation has a nontrivial solution that is finite at both
x = 1 and x = -1 if and only if /-L = n(n + 1) with integer n ~ O. This important fact
will be needed in Chapter 8.

Exercises 3.4

Power Series Method. Compute the coefficients co, q, ... , cy in the trial power series solution

L cn tn
00

x(t) =
n=O

for the following initial value problems. Find a general formula for Cn and sum the trial series (if possible).
A computer algebra system may be helpful.
202 3 ORDINARY DIFFERENTIAL EQUATIONS OF HIGHER ORDER

1. x" = 0, 8. x" +tx = 0,

x(O) = 1, x'(O) = O. x(O) = 0, x'(O) = 1.

2. x" -x' = 0, 9. x" + tx' + x = 0,

x(O) = I,x'(O) = O. x(O) = 0, x' (0) = 1.

3. x" +x' = 0, 10. x" + 2tx' - x = 0,

=
x(O) 0, x' (0) = 1. x(O) = 0, x' (0) 1. =
4. x" + 2x' + x = 0, 11. x" + t2x = 0,
x(O) = I,x'(O) =-1. x(O) = 1, x'(O) = O.

5. x" -x = 0, 12. x" + t2 x = 0,

x(O) = I,x'(O) = o. x(O) = 0, x' (0) = 1.

6. x" +4x = 0, 13. x" - tx = 0,

x(O) = 0, x' (0) = 2. x(O) = I,x'(O) = O.

7. x" -x = 0, 14. x" - t2 x = 0,

=
x(O) 0, x' (0) = 1. x(O) = 1, x'(O) = O.

Standard Basis. Find series formulas for the standard basis Xl, X2.

15. x" = O. 18. x" + 2x' + x = O.

16. x" - x' = O. 19. x" + tx = O.

17. x" +x' = O. 20. x" + t2x = O.

Euler Differential Equation. Apply Euler's method to calculate a solution basis Xl, x2 of the given Euler
differential equation.

21. t 2 x" = o. 25. 4t 2 x" +x = O.

22. t2x" + tx' = O. 26. t2x" - tx' + 2x = O.

23. t 2x" - tx' = o. 27. 4t 2x" - 4tx' + 3x = o.

24. t 2x" - 2tx' + 2x = O. 28. t 2x" - 5tx' + 9x = O.

Frobenius Method. Use the Frobenius method to calculate a nontrivial solution Xl (t) of the given
differential equation corresponding to the larger root of the indicial equation. Find an independent solution
X2 (I) corresponding to the smaller root. Sum the series where possible.

29. t 2x" = o. 33. t 2x" + 6tx' + (6 - t 2 )x = O.

30. t2x" + tx' = o. 34. tx" + x = O.

31. t 2 x" - tx' = O. 35. tx" - x' + 4t3 x = o.
32. 4tx" + 2x' + x = o. 36. 4t 2x" - 4tx' + (3 - 4t 2 )x = o.
Integral Representation of]o(x). These exercises verify the integral representation

(129) ]0(1) =~ t' cos(t sin e) de.

n 10
 3.4 LINEAR SECOND-ORDER EQUATIONS WITH ANALYTIC COEFFICIENTS 203

It is assumed known that Jo(t) is the unique solution of Bessel's differential equation of order zero x" +
(l/t)x' + x = 0 satisfying limHo+ x(t) = x(O) = 1. The derivation defines

(130) x(t) = -1111: cosU sin B) dB.

Jr 0

Then limt-+o+ x(t) = x(O) = 1. To prove relation (129), it suffices to show that x" + (l/t)x' + x = O.
37. Derive the formula

(131) x' (t) = - - 1111: sin(t sin B) sin BdB.

Jr 0

Then integrate by parts (u(B) = sin(t since», v(O) = - cos B) to show that

(132)
x'(t)
-- = - -
1111:
cos(t sin 0) cos 2 B dB.
t Jr 0

38. Differentiate (31) again to show that

(133) x" (I) = -.!:..

Jr
r
10
cos(t sin B) sin2 B dB.

39. Add equations (l33), (l32) and (130) to prove x" + (l/t)x' + x = o.
Recurrence Relations for Jn (x). Verify the statements using the given suggestions.
40. Verify equation (l09). Show that

by writing the left-hand side and the right-hand side as series.

41. Verify equation (lID). Show that

by writing the left hand side and the right hand side as series.
42. Verify equation (lll). Show that
2n
- In(x) = In-l (x)
x
+In+l (x)
by expanding the derivative terms in (l09) and (110) and equating the resulting formulas for J~ (x).
43. Verify equation (l12). Show that

2J~(x) =In-l(X) - In+l(X)

by expanding the derivative terms in (l09) and (lID) and adding the resulting formulas for J~ (x).
Approximations ofJo(x). Define the nth partial sum of the power series for Jo(l) by the formula

Sn(t) = ~
n (_l)h
(kl)2
(t2' )2h

The limitJo(l) = limn-+oo Sn (t) is valid for every real t. Complete the follOwing exercises.
204 3 ORDINARY DIFFERENTIAL EQUATIONS OF HIGHER ORDER

°
44. Plot on the same axes the graphs of]o(t), 54(t), 56(t) and 5s(t) for 0.0 =:: t =:: 8.0. Hint:]o(t) ~ 515(t)
to 5 digits on =:: t =:: 8 by the Taylor remainder formula.

45. Plot on the same axes the graphs of]o(t) and -!2/(rrt) coset - rr/4» for 0.1 =:: t =:: 10.0.

46. Justify from graphs the statement

lim (Jo(t) - J2/(rrt) cosct - rr/4» = O.

t~oo

Legendre Polynomials. The Legendre equation

(134) (1-X2)y" - 2xy' + fLY = 0, -1 =:: x=:: 1,

is known to have a nontrivial solution that is finite when x = ±1 if and only if fL = n(n + 1) where
n = 0, 1,2, .... That solutiony(x) = Pn(x) is a polynomial of exact degree n and is odd (resp. even) if n is
odd (resp. even). It is customary to normalize Pn(x) by the condition

Pn(l) = 1, n=0,1,2, ....

With this convention the polynomials are uniquely defined. Power series methods can be used to show that
they are related by the Legendre polyomial recurrence relation

(135) (n + 1)Pn+l (x) = (2n + l)xPn(x) - nPn-l (x), n::: 1.

47. Apply power series methods to show that Po (x) = landPl(x) = x.Hint: Solve (134) with fL = n(n+l),
°
n = and y(l) = 1 by the Frobenius method. This proves Po(x) = y(x) = 1. Repeat with n = 1 to
=
show PI (x) x.
48. Use the Legendre polynomial recurrence relation (135) to calculate P2 (x) and P3(X). A computer algebra
system is useful to check the answers.
49. Deduce from the Legendre polynomial recurrence relation (135) and mathematical induction that
PnO) = 1 for all n = 0,1,2, ... .
50. Produce graphs of Po (x), PI (x), ... , P3(X) on -1 =:: x=:: 1.

3.5 Numerical Methods for Second-Order Equations

This section continues the study of second-order linear equations

(1) x" + a(t)x' + b(t)x = jet), to::; t ::; to + T,

wheie the functions aCt), bet) and jet) are sectionally continuous on the interval to <
t ::; to + T.
The Initial Value Problem. The fundamental theorem of Section 3.2 guarantees
that (1) has exactly one solution with prescribed initial values x(to) = co, x' (00 = Cl. In
particular, this theorem implies the following theorem.

Unique Continuation Theorem. Let to < tl < to + T and let Xl (t) be a solution of (1)
for to ::; t ::; tl. Then Xl (t) has a unique continuation to thelarger interval to ::; t ::; to + T.
 3.5 NUMERICAL METHODS FOR SECOND-ORDER EQUATIONS 205

The continuation may be constructed by solving the initial value problem

(2) x~ + a(t)x; + b(t)x2 = j(t), t1':::: t .:::: to + T,

X2(t1) = Xl (t1), X;C(1) = x~ (t1),

and then defining

(3) x(t) = IX1 (t) for to':::: t .:::: t1,

X2(t) for t).:::: t .:::: to + T.

For the proof one notes that the sectionally defined function (3) has all the properties
of a solution of the initial value problem on the interval to .:::: t .:::: to + T; see Section 3.2.
Hence it is the solution, by the uniqueness part of the fundamental theorem.
The unique continuation theorem is useful in allowing one to build solutions of (1),
section by section, when aCt), bet) and jet) are discontinuous. This is illustrated in a
vibration problem below.

Numerical Methods for the Initial Value Problem. The fundamental theorem
guarantees the unique solvability of the initial value problem for (1) for all sectionally
continuous coefficients. However, the solutions can be represented by elementary
functions and their integrals only in very special cases: constant coefficients (a(t) = ao,
bet) = bo), Euler equations (aCt) = aolt, bet) = bolt 2 ), and coefficients that are sectionally
of these types. For more general coefficients it is natural to adopt the point of view of
Section 2.7: The differential equation (1) and the initial conditionsx(to) = Co, x'(to) = c)
define a unique function x(t). The analysis then shifts to a search for effective numerical
algorithms to compute the values of x(t). This search leads quickly to an important
discovery, namely, that the numerical methods for first-order equations of Section 2.7
can also be applied to second-order equations. To do this one must write the latter as a
pair of first-order differential equations.

First-Order Systems. The decisive step is to regard the derivative yet) = x'(t) as a
second unknown function and consider the first-order system

(4) X' =y,

y' = - a(t)y - b(t)x + jet).
This system is fully equivalent to equation (1). Thus if x(t) satisfies (1) and yet) = x'(t)
then (4) holds. Conversely, if (4) holds then x(t) satisfies (1).
206 3 ORDINARY DIFFERENTIAL EQUATIONS OF HIGHER ORDER

Before describing numerical methods for solving (4), it will be helpful to write it in
a more symmetrical form. First, note that the general second-order equation

(5) x" = F(t,x,x')

is equivalent to the system

(6) x'=y,

y' = F(t,x,y).

The symmetrical form of this is the general first -order system of two differential equations

(7) x' = F1(t,x,y),

y' = F2(t,X,y),

where Fl and F2 are any two functions of the three variables t,x and y. In particular,
system (4) has the form (7) with

(8) Fl(t,X,y) = y,

F2(t,X,y) =J(t) - b(t)x - a(t)y.

Finite Difference Methods. The methods of Section 2.7 can be extended to system
(7) almost without change. Approximate solutions will be constructed to the initial
value problem for (7) on a prescribed interval [to, to + TJ with prescribed initial values
x(to) = Co, y(to) = Cl. Of course, as remarked in Section 2.7, the existence of the
solution on this interval must be known if meaningless computations are to be avoided.
Uniform Mesh. For Simplicity, only uniform meshes will be discussed. Thus

(9)
tk = to + kh, Nh = T.
The approximate values of x(t) and yet) at the mesh points will be denoted by Xk, Yk, so
that

(10)

Euler's Method. Equation (7) is similar to the Single first -order equation x' = F(t, x)
of Chapter 1. It differs only in having two unknown functions and two equations. When
written in vector notation, (7) is identical in form to the single equation. This suggests
that we may extend the methods of Section 1.7 directly to the systems (7). The simplest
numerical method is the following:
 3.5 NUMERICAL METHODS FOR SECOND-ORDER EQUATIONS 207

The Euler Method For Systems

(ll) Given to, Xo, Yo, T, N with h = TIN,

For h = 0, 1,2, ... , N - 1,
Compute tk+l = tk + h.
Then compute
Xk+l = Xk + hFI (tk,Xk,Yk), and
Yk+l = Yk + hF2(tk,Xk,Yk).

This has the same recursive structure as the single-equation case. Thus for h = 0
one starts with the given initial data Xo = x(to), Yo = y(to) and one computes Xl, YI.
Then, with these data one computes x2, Y2, etc. For h = N - lone gets the final values
XN ~ x(to + T), YN ~ y(to + T).
Error Estimates. The analysis of Section 2.7 extends to the system (7). Thus

(12)

for a suitable constant C, provided that FI(t,x,y), F2(t,X,y) and their first partial
derivatives are continuous. For the linear equation (4) this is valid if a(t), b(t), and J(t)
have continuous first derivatives.

EXAMPLE 1. Euler's method for systems will be illustrated by application to the problem

CD) xl! +X = 0, x(O) = 0, x'CO) = 1,

with exact solution

(l4) X(t) = sin t, x' (t) = cos t.

Application of Eulers method on the interval 0 :s t :s n with various values of N gives

the results in Table 3, where the values XN ~ x(n) and YN ~ x' (n) are recorded, together
with the errors

(15)

As in Section 1. 7, the slow convergence of a first -order method is observed.

The Runge-Kutta4 Method. This method will be described as an example of a
higher-order algOrithm. It parallels the single-equation case of Chapter 2 closely, but with
two unknowns and two equations at each step. The step from (Xk,Yk) to (Xk+l,Yk+I) is
as follows. Compute
(16) Xl = hFI (tk,Xk,Yk) ,
YI = hF2 (tk,Xk,Yk) .
208 3 ORDINARY DIFFERENTIAL EQUATIONS OF HIGHER ORDER

TABLE 3. Euler's method applied to equation (13)

N 5 10 100

XN 0.7590168183 0.1560636299 0.001085156

EN -0.7590168183 -0.1560636299 -0.001085156

YN -2.1685690322 -1.5933623253 -1.050559787

E'N 1.1685690322 0.5933623253 0.050559787

Then compute

(17)

Then compute

(18)

Then compute

(19) X 4 = hFI (tk + h,Xk + X 3 ,Yk + Y3 ) ,

Y4 = hF2 (tk + h,Xk + X 3,Yk + Y3)'

Finally, compute
1
(20) Xk+l = Xk + -6 (Xl + 2X2 + 2X3 + X 4) ,
1
Yk+l = Yk + '6 (Yl + 2Y2 + 2Y3 + Y4 ) .
EXAMPLE 2. Application of The Runge-Kutta4 method to problem (13) gives the results
in Table 4.
Note the rapid fourth-order convergence.
Other Methods. The Taylor2, Taylor4, and Heun algorithms of Section 2.7 have
versions for second -order equations. The key to such generalizations is the vector notation
of Chapter 5, which shows that each method of Section 2.7 is a special case of a method
that is valid for equations of any order.
 3.5 NUMERICAL METHODS FOR SECOND-ORDER EQUATIONS 209

TABLE 4. The Runge-Kutta4 method applied to equation (13)

N 10 100

XN 0.0035138083 0.0002460702 0.0000000255

EN -0.0035138083 -0.0002460702 -0.0000000255

YN -0.9979640590 -0.9999340320 -0.9999999993

E'N -0.0020359410 -0.0000659680 -0.0000000007

Exercises 3.5

Unique Continuation and Exact Solutions. Below, solve the given initial value problem to obtain the
sectionally defined solution x(t) on the given interval.
1. x" = jet), x(O) = x/CO) = O,j(t) = ton [0, 1],j(t) = 1 - t2 on [1, 3J.

2. x" = jet), x(O) = 1, x/CO) = O,j(t) = ton [0, 1],j(t) = 1 - t 2 on [1, 3J.
3. x" = j(t),x(O) =x/(O) = O,j(t) = sint on [O,rrJ,j(t) = 1- sint on [rr,2rrJ
4. x" =j(t),x(O) = O,x/(O) = 1,j(t) = sint on [O,rr],j(t) = I-sinton [rr,2rrJ.
Numerical Calculations and Unique Continuation. A controller x(t) for a process is given as two initial
value problems: x" = 1, x(O) = x/CO) = 0 on [0, IJ and x" = j(t) on [1, 2J. Findx(2) numerically, assuming
that j(t) is a table of values given as below.
5. j(1) = 1,j(l.2) = 1.5,j(1.4) = 1.3.1(1.6) = 1.0,j(1.8) = 1.2,j(2) = 1.4.
6. f(1) = 0.9,f(1.2) = 0.5,f(1.4) = 0.7.10.6) = 0.85,j(1.8) = 0.95,j(2) = 1.1.
7. J(t) is a sampling of sin(rrU2) on [1, 2J at steps of 0.2.

8. J(t) is a sampling of t2 on [1,2] al sleps of 0.2.

Euler's Method. Solve the following initial value problems using four Euler steps of step size h = 0.1.
Present the answer as a table of values for th, Xk,Yk, Ek and Ek where th is a point of the uniform mesh {tklt=o'
9. x" + 16x = 0, x(O) = 1, x/CO) = O.
10. x" = sin t, x(O) = x/ (0) = o.
11. x" + 2x' +x = e- t , x(O) = x/CO) = O.

12. x" = cos t, x(O) = x/CO) = O.

13. x" + 16x = 0, x(O) = 0, x/ (0) = 1.

14. x" = t, x(O) = x/ (0) = O.

15. x" = sinh(t), x(O) = x/CO) = O.

Runge-Kutta4 Method. Solve the follOWing initial value problems using two RK4 steps of step size
h = 0.2. Present the answer as a table of values for th, Xh, Yk, and Eh, where th is a point of the uniform
mesh {thl~=o and Eh = J(Eh)2 + (Ek)2 is the mean error at th'

16. x" + 16x = 0, x(O) = 1, x/CO) = O.

210 3 ORDINARY DIFFERENTIAL EQUATIONS OF HIGHER ORDER

17. x" = sin t, x(O) = x'(O) = O.

18. x" + lx' + x = e- t , x(O) = x'(O) = O.
19. x" = lOcost, x(O) = x'(O) = 0

20. x" + SOx = 0, x(O) = 0, x'(O) = 1.

21. x" = t2 , x(O) = x'(O) = O.

22. x" = cosh t, x(O) = x' (0) = O.

Euler's Method and Graphing. Solve the follOwing initial value problems using four Euler steps of step
size h = 0.1, producing a table of values for th, Xh, Yh, and Eh, where th is a point of the uniform mesh
(thl~=o' Graph the linear interpolant of the data setth, Eh, k = 0, ... ,4. This graph represents the difference
between the exact solution and the approximate solution on [0,0.41.

23. x" + 16x = 0, x(O) = 1, x'(O) = O.

24. x" = sin t, x(O) = x' (0) = O.

25. x" + 2x' +x = e- t , x(O) = x'(O) = O.

26. x" = cosh t, x(O) = x' (0) = o.

27. (lC Circuit) Consider the LC circuit of Figure 17 with parameters L = 1, C = 1 and a battery
that supplies Eo = 10 volts. At times t < 0 the switch is open and there is no charge or current:
q(t) = i(t) = 0 for t < O. At time t = 0 the switch is dosed to position B and the capacitor begins to
charge. At time t = 10 seconds the switch is set to position A, shorting out the battery. Calculate the
charge q(t) for all t ::: 0 using the unique continuation theorem.
28. (Paratrooper) A paratrooper falls from rest at 12000 feet for to seconds, then the parachute opens.
Estimate the time of the journey to the ground using the model below.

W
-x
II {WW 0:::. t :::. to,
g (t) = - ex '()
t t> to,
with the follOWing physical constants:
Combined weight W = 200 pounds (paratrooper and parachute),
Gravitational constant g = 32 feet per second,
Parachute open at to = 3 seconds,
Parachute drag force constant e = 1.
29. (Unsolvable Numerical Problem) Verify that the initial value problem x" = lx' 12/3 , x(O) = x' (0) =0
has infinitely many solutions. Hence no numerical method can make sense for this problem.

A
open

B Eo

L
~ C

FIGURE 17. LC circuit with battery and switch

 3.6 liNEAR EQUATIONS OF ORDER n> 2 211

30. (Euler's Method) Verify Table 3.

31. (The Runge-Kutta4 Method) Verify Table 4.

32. (Conversion to Systems) The differential equation ulll (t) + 4u" (t) - Su' (t) = sin(t) can be converted
to a system of first-order differential equations by the substitutions x = u(t), y = u'(t), Z = u"(t).
Show that the resulting system is

x' =y,
y' = Z,
z' = Sy - 4z + sin(t).
33. (Third-Order System) The equation u'" (t) + 4u" (t) - Su' (t) = sin(t) can be converted to a system
of first-order differential equations and solved numerically by a generalization of the Runge-Kutta4
method for systems of dimension 2. Develop the general equations. Apply them to this problem with
initial conditions u(O) = 0, u' (0) = 0, u" (0) = 1 and step size h = 0.1 to compute the values of u, u',
u" at t = 0.1. Compare with the exact solution

3 1. 2 1 t
u(t) = - cos(t) - - sm(t) - - + -e + -9e-5t
26 13 5 4 260
and its derivatives.

3.6 Linear Equations of Order n > 2

In this section the main results of Sections 3.3 and 3.4 on linear second-order equations
are extended to orders n > 2. The most general linear nth-order equation can be put
into the form

(1)

where the coefficients al(t), ... , an(t) and the right-hand termj(t) are prescribed func-
tions. The equation is said to be homogeneous ifj(t) == 0, otherwise inhomogeneous.
It is said to have constant coefficients if aI, ... , an are all constants. Examples from
physics with n = 3 and n = 4 were given atthe end of Section 3.1, and more complicated
electrical and mechanical oscillators lead to equations of all orders.
The solution concept and fundamental theorem of Section 3.2 extend completely to
arbitrary n > 2. They may be formulated as follows.
Definition. A function x(t) (defined on an interval to :::: t :::: tl) is said to be a solution
of (l) if (i) x(t) and the derivatives dkxldt k with k :::: n - 1 are continuous, (ii) dnx(t)ldt n
is a sectionally continuous function with discontinuities only at discontinuities of
al (t), ... , an (t) and jet) (if any), and (iii) equation (1) holds at each point of continuity
of al (t), ... , an (t) and j(t).
With this definition, one has the following result.
212 3 ORDINARY DIFFERENTIAL EQUATIONS OF HIGHER ORDER

Fundamental Theorem. For each set of n real numbers Co, Cl, ... , cn-l, equation (1)
has exactly one solution x(t) such that

(2) X(k)(tO)=Ck for k=O,l, ... ,n-l.

This theorem makes precise the sense in which the solutions of (1) form an
n-parameter family Proofs of the theorem may be found in advanced texts such as [B-RJ.
The Homogeneous Equation. If each of a set of functions Xl (t), X2 (t), ... , Xk (t) is a
solution of the homogeneous equation

(3)

and if AI, A2 , ... , Ak are any constants, then the function

(4)
is easily seen to be a solution of (3). The function x(t) defined by (4) is said to be
a linear combination of the functions Xl(t), X2(t), ... , Xk(t). The constants AI, A 2,
... , Ak are said to be the coefficients. A linear combination (4) is said to be trivial if
Al = A2 = ... = Ak = 0. Otherwise, it is said to be nontrivial.

In Section 3.2, functions Xl (t) and X2 (t) were said to be linearly dependent if one of
the functions was a constant multiple of the other. The extension of this important idea
to sets of k > 2 functions is contained in the following.
Definition. A set of functions Xl(t), X2(t), ... , Xk(t) (defined on a common interval
to :::s t :::s tl) is said to be linearly dependent if and only if there is a nontrivial linear
combination of them that equals zero:

(5)

If the set Xl (t), X2 (t), ... , Xk (t) is not linearly dependent then it is said to be linearly
independent.
If Xl, X2, ... ,Xk is a linearly dependent set, so that (5) holds with not all Ai = 0, then
some of the functions in the set are equal to linear combinations of the others. Thus, if
°
Al =j:. in (5) then Xl can be written as a linear combination of X2, X3, ... ,Xk, etc. Note
that when k = 2 the new definition of linear dependence coincides with the old. Also,
if Xl, X2, ... , Xk are linearly independent and (5) holds then it must be the case that
Al = A2 = ... Ak = 0. The basic facts concerning linearly independent solutions are
described by the following theorem.

Theorem. There are sets of n linearly independent solutions of (3). Every set of n + 1
solutions is linearly dependent.
 3.6 LINEAR EQUATIONS OF ORDER n> 2 213

These statements follow from the fundamental theorem. To prove the first statement,
let Xl, X2, ... ,Xn be solutions of (3) with the initial values

(
6) Xl ( = 1,
to ) XlI (to) = 0, = 0, ... ,Xl(n -1) (to) = 0 ,
Xl/I (to)

to = 0 ,
X2 () x2I (to) = 1, 2 to = , ... ,x2
x"() 0 (n-l) (to) = 0 ,

Then assume that these solutions satisfy

Taking t = to in (7) and using (6) gives Al = O. Next, differentiating (7) once and setting
t = to gives A2 = 0, etc. Hence, the coefficients AI, A 2 , ... , An must all be zero, and
therefore Xl, X2, ... , Xn are linearly independent.
At this point we have not developed enough linear algebra to verify the second part
of the theorem. This will be completed in Chapter 5. Note, however, that if x(t) is any
solution of 0), characterized by its initial values (2), and if Xl, X2, ... , Xn is the special
set of solutions defined by (6), then

(8)

This follows immediately from (2) and (6). Thus the set XI, X2, ... , Xn is linearly
dependent.
Solution Bases. Equation (6) defines one set of n linearly independent solutions of 0),
Section 3.3. There are (infinitely) many other such sets, each generated by initial values.
Each such set of n solutions will be called a solution basis for 0). The reason for this
terminology is the following theorem.

Basis Theorem. If Xl (t), X2 (t), ... ,Xn (t) is any solution basis for 0), then every solution
of that equation can be written in one and only one way as

(9)

Note that the existence and uniqueness statements for (9) follow directly from the
preceding theorem. Indeed, if X, Xl, X2, ... , Xn are linearly dependent, then there must
be constants B, BI , B2 , ... , Bn , not all zero, such that

(10)
214 3 ORDINARY DIFFERENTIAL EQUATIONS OF HIGHER ORDER

Moreover, B #- 0 because otherwise, Xl, X2, ... , Xn would be linearly dependent. Thus
we may divide by B in (10) and write Ak = -BkIB to get (9). The uniqueness of the
representation (9) follows similarly.
Now consider the nonhomogeneous equation (1). In complete analogy with the case
n = 2 one has the following.

Structure Theorem for Linear nth-Order Equations. let xp(t) be any particular
solution of (1) and let Xl (t), X2(t), ... ,Xn (t) be any solution basis for (3). Then for each
solution x(t) of (1) there exist unique constants AI, A2 , ... , An such that

(ll)

This result is immediate from the basis theorem, since the difference x(t) - xp(t) is
a solution of the homogeneous equation (3).
Solution of the Nonhomogeneous Equation. The method of variation of parameters
has a straightforward extension to nth-order nonhomogeneous equations.
Equations with Constant Coefficients. Consider the homogeneous equation

(12)

with constant coefficients al, a2, ... , an. Eulers method of exponential solutions

(13) X(t) =It

yields a solution basis, just as in the special case n = 2. Thus (13) defines a solution of
(12) if and only if A is a root of the characteristic equation

(14)

Thus, for the third-order equation (54),

(15) X'" + 6x" + l1x' + 6x = 0,

one has

(16) P3(A) = A3 + 6A 2 + 11A + 6 = (A + 1)(A + 2)(A + 3)

with roots -1, -2 and -3. Hence Xj(t) = e- t , X2(t) = e- 2t and X3(t) e- 3t are
solutions that are easily seen to be linearly independent, and it follows that

(17)

is the general solution of (15), as stated in Section 3.1.

 3.6 LINEAR EQUATIONS OF ORDER n > 2 215

Similarly, for the fourth-order equation (64),

(18) X(4) + 2X"' + 2X" + 2x' +X = 0,

one has

(19)

The double root A = -1 yields solutions Xl (t) = c t and X2 (0 = t e- t , while the

complex roots A = ±i yield X3 (t) = cos t and X4(t) = sin t. These may be shown to be
linearly independent and hence the general solution of (18) is

(20)

Rules for Calculating a Solution Basis. The preceding examples illustrate the fact
that if the roots of the characteristic equation (14) can be found, then a solution basis
of (12) can be written down. Two cases arise, corresponding to real and complex roots,
respectively.
Case 1. Real Roots. Let A = Al be a real root of (14), of multiplicity m ::: 1, so that
(A - Al)m is a factor of Pn(A). Then (12) has the m linearly independent solutions

(21)

Case 2. Complex Roots. Let A = P + iq be a complex root (q > 0) of multiplicity

m ::: 1, so that I = P- iq is also a root of multiplicity m and [(A - p)2 + q2t is a factor
of pn(A). Then (12) has the 2m linearly independent solutions

(22) Xm+l = ept sin qt,

X2 = t ept cos qt, xm+2 = t ept sin qt,

The final rule states that the set of n solutions obtained in this way is a solution basis
of (12). The algebraic proofs of these rules are rather long. They will not be written down
here. However, they are readily verified by the Laplace transform method of Chapter 4
and the theory of partial fractions.

IExercises 3.61
;Lk<i$~~"'1!(hb [,u"iG!k ".,C41\,,,,,=,

A computer algebra system will be helpful in solving some of these exercises.

216 3 ORDINARY DIFFERENTIAL EQUATIONS OF HIGHER ORDER

Solution Basis. Find a solution basis for the given differential equation. First try to factor the characteristic
polynomial into linear and quadratic factors. If this is possible then an exact basis should be given. If not
then compute and use approximate roots.
1. x''' + 6x" + 11x' + 6x = O.
2. x'1II + 2Xlll + 2x" + 2x' + x = O.
3. XIII + 3x" + 4x' + 2x = O.
4. 9x"" - 24xlll - 29x" + 120x' - 80x = O.
5. 2Xlll + 3x" - 12x' + 5x = O.
6. 4Xlll - 3x" - 18x' + 8x = O.
7. 100xlll - 334x" + 668x' - 536x = O.
S. 100xlll - 334x" + 768x' - 670x = O.
9. 9x"" - 42x1ll + 22x" + 126x' - 147x = O.
10. 2500xlll + 9600x" - 14325x' - 5500Sx = O.
Initial Value Problems. Calculate the solution of the initial value problem with the prescribed initial values.
11. 2Xlll + 3x" - 12x' + 5x = 0, xCO) = 1, x'CO) = 0, x"CO) = O.
12. 4x''' - 3x" - 18x' + 8x = 0, xCO) = 0, x'CO) = 1, x"CO) = o.
13. XIII + 6x" + llx' + 6x = 0, xCO) = 1, x'CO) = 1, x"CO) = 1.

14. x"" + 2x1ll + 2x" + 2x' + x = 0, xCO) = 1, x'CO) = 0, x"CO) = 0, XIII CO) = 1.
Cauchy's Solution. Let the homogeneous linear differential equation (3) have constant coefficients.
Cauchy's solution is defined by the initial conditions

KCO) = ... = K(n- 2l CO) = 0, K(n-ilCO) = 1.

In the exercises below calculate the Cauchy solution KCt) for the given differential equation.
15. XIII + 6x" + llx' + 6x = O.
16. x"" + 2x''' + 2x" + 2x' + x = O.
17. XIII + 3x" + 4x' + 2x = O.
IS. 9x"" - 24xlll - 29x" + 120x' - 80x = O.
19. 100xlll - 334x" + 668x' - 536x = O.
20. 100xlll - 334x" + 768x' - 670x = O.
21. 9x"" - 42xlll + 22x" + 126x' - 147x = O.
22. 2500xlll + 9600x" - 14325x' - 55008x = O.
Variation of Parameters. Solve for the particular solution xpCt) with zero initial conditions, as given by
the variation of parameters formula

xpCt) = t KCt - s)jCs)ds,

110
where KCt) is Cauchy's solution. As usual,jCt) represents the right-hand side of the differential equation and
to is the initial point. Choose to = 0 throughout.
 3.6 LINEAR EQUATIONS OF ORDER n> 2 217

23. x'" = 1. 28. x"" = t3

24. XIII = et. 29. /6) = t.

25. = sin t.
XIII 30. x"" - x = 1.
26. x"" = cost. 31. x"" -x = et.
27. 100x lll = t2 32. x"" - x = e-t.
General Solution. Find the general solution xCI) = xh (I) + xpCI) for the following equations.
33. XIII = I. 36. x"" = t3 .

34. x"" = cos t. 37. x"" - x = 1.

35. lOOxlll = t2 38. x"" - x = et.

CHAPTER 4
 The Laplace Transform

This chapter presents an efficient and powerful method for solving the differential
equations that govern electrical networks and many other problems of engineering and
applied science. The method, which originated in work of O. Heaviside 1 dating from
the 1890s, consists in transforming the differential equations, according to certain rules,
into corresponding algebraic equations. The latter are then solved algebraically and the
solutions to the original differential equations are obtained from them by Heaviside's rules.
Heaviside did not base his transform method on the Laplace transform, although
he was aware of the connection. However, most subsequent workers have founded their
treatments of Heaviside's method squarely on the Laplace transform. It is this approach
that is followed here.

t-Domain s-Domain
Original Problem Transformed Equation
=*
y" + 5y' + 6y = 0, s2y + 5sY + 6Y = 1
y(O) = 0, y'(0) =1
.IJ.
Solution of the Solution of the

Original Problem Transformed Equation

-1 1
y= e- 2t _ e- 3t Y=--+--
s+3 s+2

I Oliver Heaviside 0850-1925). British scientist famous for his many contributions to electromagnetic theory, based
on Maxwells equations, and for their applications to electrical engineering and technology
220 4 THE LAPLACE TRANSFORM

4.1 The Nature of the Laplace Transform

Many quantities of interest in engineering and applied science are functions jet) of a time
variable t that ranges over the interval 0 ~ t < 00. A number of examples from electrical
circuit theory, mechanics, heat transfer and other areas have been studied in the preceding
chapters. Typically, jet) is a solution of an initial value problem for a linear differential
equation. The Laplace transform method for solving such problems is to associate with
jet) a second function F(s), depending on an auxiliary variable s, in such a way that
the differential equation for jet) is replaced by an algebraic equation for F(s). The new
equation is solved for F(s) by simple algebra and the functionj(t) is found from F(s).
The official definition of the Laplace transform is given by the integral

(1)

If F(s) is related to j(t) by this equation then F(s) is said to be the Laplace transform of
jet). When it is necessary to express the dependence of F(s) onj(t), the notation

(2) F(s) = .c{f(t)}

will be used. The expression .c{j(t)} is a function of the variable s and does not depend
on t. The t-dependence has been integrated away. Here, .c is an example of an integral
operator. It assigns to eachj(t) from a certain class a second function F(s). The function
F(s) is referred to as the image ofj(t) under .c.
The integral in (1) is said to be improper because of the upper limit "00". The
definition of the improper integral is

(3) r>o ¢J(t) dt = lim t ¢Jet) dt.

10 T--..oo 10

The existence of this limit depends on the function ¢J(t). The improper integral is said to
exist if the limit exists. As an example, consider the function ¢J(t) = e;<H. Direct integration
gives

(4) io
T eat
eatdt=-
a
IT
0 a a
(a:j:. 0).

Hence, the improper integral exists, and

(5) io
oo 1
eat dt =--
a
 4.1 THE NATURE OFTHE LAPLACE TRANSFORM 221

if and only if a < O. A simple sufficient condition for the Laplace transform of J(t) to
exist is

(6) 1 00
e-sotlf(t)I dt < 00 for some so.

In fact, the following theorem is proved in advanced calculus courses:

Theorem. If condition (6) holds then the Laplace transform F(s) = .c{f(t)} exists for
all s ~ so.

A useful criterion for the existence of F(s) is given by the following corollary

Corollary. If J(t) is integrable on every finite interval 0 :::: t :::: T and if

(7) for all t ~ T,

where M and a are constants, then F(s) exists for all s > a.
The proof is based on showing that (7) implies that (6) holds whenever So > a.
Linearity of .c. IfJl (t) and h (t) have Laplace transforms and if Cl and C2 are constants,
then elementary properties of the integral imply that

(8)

Equivalently,

(9)

implies

(10)

for all numbers s for which Fl (s) and F2 (s) exist. Operators.c that satisfy (8) are said to
be linear operators. This simple property of the Laplace transform plays a fundamental
role in applications.
The Heaviside Function. The step function defined by

= (~
for -00 < t < 0,
(ll) H(t)
for 0:::: t < 00,

g
is known as the Heaviside function. A related function is

for -00 < t < a,

(12) H(t - a) =
for a:::: t < 00,
222 4 THE LAPLACE TRANSFORM

where a is a constant. If a ::: 0 then H(t - a) = 0 for t :s a, and its Laplace transform
can be calculated from the equations

(13) iT
o
e- 5t H(t - a) dt =
iT e-
a
5t dt = -=-
e- IT
5t

5 a 5 5

It follows that .c{H(t - a)} exists and

e- a5
(14) .c{H(t - a)} = - for all 5> 0 and a::: O.
5

This result plays an important role in many applications of the Laplace transform.
Uniqueness of the Transform. For applications it is often necessary to recover the
original function J(t) when its transform F(5) is known. Thus the question arises whether
eachF(5) comes from a uniqueJ(t), orwhethertwo or moreJ(t),s can have the same F(5).
If discontinuous J(t)'s, such as the Heaviside function, are allowed, then J(t) is certainly
not unique. Indeed, in the definition (1) J(t) can be changed at a finite number of points
without affecting the values of F(5). Nevertheless, the original function is "nearly unique"
in the sense of the following theorem.

Uniqueness Theorem. Let Jl (t) and !I (t) have Laplace transforms .c{fd = Fl (5) and
.c{f2} = F2 (5). Then

(15)

if and only if

(16) l tJ1 (T)dT= I t

!I(T)dT forall t:::O.

Moreover, ifJl(t) and!I(t) are continuous for all t ::: 0, then

(17)

The rather complicated proof of this theorem may be found in advanced texts on
Laplace transforms, e.g., [WiJ.
The remainder of Chapter 4 is organized as follows: In Section 4.2 it is shown that
the Laplace transform (14) of the Heaviside function and the linearity and continuity
properties of the operator .c allow one to calculate the Laplace transforms of many
elementary functions without carrying out the t-integration of the definition (1). Section
4.3 presents the operational rules for the Laplace transform that are needed for the
applications to differential equations. These rules also permit one to calculate the direct
and inverse transforms of large classes of functions. Section 4.4 develops applications
 4.1 THE NATURE OF THE LAPLACE TRANSFORM 223

of the Laplace transform to linear ordinary differential equations of any order. Finally,
Section 4.5 develops applications of the transform to linear systems of differential
equations of the kind that arise in electric circuit theory and mechanics.
In Section 4.2 many transforms will be calculated by applying three simple properties
of the operator C. The indicated integration in equation 0) will be actually performed
for only a limited number of examples.

Exercises 4.1

Heaviside's Function. Graph the given combination of Heaviside functions on [0, (0).

1. f(t) = H(t - rr)

2. f(t) = m(t - 5) + 3H(t - 7).

3. f(t) = HU - 1) - H(t - 2)

4. f(t) = H(t - a) - H(t - b), given b > a> 0.

Calculation of .e-transforms by direct integration. Verify the given .e-transform equation by evaluating
the integral fle-stf(t)dt, using an integral table or a computer algebra system, and then letting T -+ 00.
Assume throughout that a :::: 0.

5 . .e{H(t - a)} = e-as/s for s > 0

6 . .e{ (t - a)H(t - a)} = e- as /s 2 for s > O.

7 . .e{et-aH(t - a)} = e-as/(s - 1) for s > l.

8 . .e{et-aU - a)H(t - a)} = e-as/(s - 1)2 for s > l.

Calculation off(t) from F(s) = .e{j(t)}. Apply the uniqueness result (17) to £indf(t).

9 . .e{j(t)} = .e{e- 7t }

10 . .e{j(t)} = .e{e- 5t } - U{l}.

11. .e{j(t)} = s1 + -e-52s- '

12 . .e{jU)} = (s_\)2'

-2s
13 . .e{j(t)} = ~-l

e-l. IS
14 . .e{jU)} = ~.
224 4 THE LAPLACE TRANSFORM

4.2 The Laplace Transforms of Some

Elementary Functions
The Laplace transform £ defined in Section 4.1 by (1) may be shown to have the following
three properties.

(1) £ is linear, £{Cdl + c2h} = cl£{fd + C2£{f2}.

(2) £{H(t - a)} = e- as /5 for all 5> 0 and a:::: O.
(3) £ is continuous.

The last property means that if jn(t) -+ jet) as n -+ 00 then £{fn(t)} -+ £{f(t)} as

n -+ 00, for all 5 > 50.

Here it will at first be assumed that £ is any operator that has the three properties
listed above. It will be shown that these properties alone make it possible to calculate the
£-transforms of many elementary functions. Later it will be shown that any operator £
with these three properties must be given by the Laplace integral 0) of Section 4.l.
To derive a second £-transform from (2) we shall replace a by b and integrate with
respect to b over a :s b < 00, where a :::: O. This gives

(4) 100 £{H(t -

a
b)} db =- 1100 e-
S a
bs db = -c 2 .
S
a

Next, the linearity and continuity of £ imply that the integration on the left in (4) can
be brought inside the operator £. Thus

(5)

The integral of H(t - b) is elementary and may be evaluated by the substitution

t- b = r, -db = dr. Thus
(6) 1 a
00
H(t - b) db = 10
rt-a
H(r) dr = (t - a)H(t - a),

for all a:::: O. Hence we have the new £-transform

e- as
(7) £{(t-a)H(t-a)}=- forall s>O and a::::O.
52

This process may be repeated to get additional £-transforms. In fact, for n = 0, 1, 2,

... the change of variable t - b = r gives

(8) 1 a
00

(t - b)nH(t - b) db =
ret-a)
10 rnH(r) dr
 4.2 THE LAPLACE TRANSFORMS OF SOME ELEMENTARY FUNCTIONS 225

(t - a)n+l
= H(t-a).
n+l
Thus (7) and (8) with n = 1 give
2e- as
(9) C{(t - a)2H(t - a)} = - 3 - for all 5> 0 and a 2: o.
5

Proceeding in this way gives the C-transforms

n!e- aS
(10) C{(t - a»>H(t - a)} = -sn+
-1' n = 0,1,2, ....

An important special case is obtained by setting a = 0:

n!
(ll) C{tnH(t)} = sn+l' n = 0, 1,2, . ...
C-Transform of an Exponential. Recall that for any real or complex constant a one
has the Taylor series

(12)

which is valid for all t. If the operator C is applied to the series on the right, then C and
the summation may be interchanged, by the linearity and continuity of C. If (11) is then
used, one gets

Cleat} = 00
~ _
an = _1 ~
00 (a)n
_ .
(13)
~ sn+l 5 ~ 5
n=O n=O

Now recall the geometric series

(14) Lxn - --
n=O
00 1

1- x
for Ixl < l.

This shows that for la/51 < 1 the right-hand side of (13) equals 11(5 - a). Thus

1
(15) Cleat} = - - for all 5 > la!-
s-a
This equation may now be differentiated with respect to a to get further C-transforms.
The differentiation may be brought inside the C operator by the linearity and continuity
properties. Thus (15) implies

I' at 1
(16) L{te } = (
s-a)2
226 4 THE LAPLACE TRANSFORM

and

(17)

In general, one has

(18) .t.-{t e } =
I' n at
(5 - n!a)n+ l' n = 0, 1,2, ....

£-Transforms of Trigonometric Functions. Equation (15) is valid for any complex

constant a. In particular, if w is real, then

(19) £{eiwt } = _1_ = 5 + iw .

5- iw 52 + w2
Combining this with Euler's formula

(20) eiwt = cos wt + i sin wt

and the linearity of £ gives

+ i£{sinwt} = + i -2--2'
5 W
(2l) £{coswt} -2--2
5 +w 5 +w
Finally, equating real and imaginary parts gives the .c-transforms of cos wt and sin wt:
5
(22) £{coswt} = -2--2'
5 +w
w
(23) .c{sin wt} = -2--2'
5 +w
Products of Powers of t and Trigonometric Functions. Equation (18) with
a = iw (w real) gives

. t n!
(24) £{tneIW } = - - - -
(5 - iW)n+l

In particular, one has

(25)
£{teiwt } = (5 + iw)2 (52 - ( 2) + i(2W5)
(52 + ( 2 )2 (52 + ( 2 )2
whence on equating real and imaginary parts one gets
52 _ w2
(26) £{tcoswt} = ( 2
5 +w2)2'
2W5
(27) £{t sin wt} = (52 + ( 2 )2
 4.2 THE LAPLACE TRANSFORMS OF SOME ELEMENTARY FUNCTIONS 227

Similar, but more complicated, formulas for CW cos wt} and C{ tn sin wt} follow from
(24).
Products of Powers of t, Real Exponentials, and Trigonometric Functions.
Equation (18) with a = a + iw gives

(28) C {t nat
e e
iwt}
=
n!(s-a+iw)n+l
, n = 0, 1,2, ....
[(s - a)2 + w2]"+!

In particular, taking n = 0 gives

(29) C{e"te iwt } = s - a + iw .

(s - a)2 + w2
Separating real and imaginary parts gives

C{eat cos wt} = ( S - a

(30) )2 2'
s-a +w
w
(31) Cleat sinwt} = ( )2 2
s-a +w
Similar formulas follow from (28) with n 2: 1.
Functions Whose C- Transforms Are Rational Functions. A rational function
of the variable s is a quotient

(32) F( ) = N(s)
s D(s) ,

where N(s) and D(s) are polynomials in s. It has been shown above that functions J(t)
of the form tne"t cos wt and tneat sin wt have C-transforms that are rational functions of
s with degree N < degree D. More generally, the linearity of C implies that any linear
combination of such functions has the same property. It is interesting that the converse
of this statement is true. This may be stated as follows.

Theorem. Let F(s) be any rational function with degree N < degree D. Then F(s) =
C{j(t)}, where J(t) is a finite linear combination of functions of the form tne"t cos wt and
tme"t sin wt.

The parameters n, m, a, and w in the theorem may have different values in the terms
of the linear combination.
The proof of this theorem depends on the theory of partial fractions. These are
special rational functions of the form
A
(33) Rl (s) = (
s-a)m'
228 4 THE LAPLACE TRANSFORM

where a, ex, fJ, A, B, C are real, fJ > 0, and m is a positive integer. The partial fraction
decomposition theorem states that every rational function F(s) = N(s)/D(s) with real
coefficients and degree N < degree D equals a finite sum

= L Fj(s),
k
(34) F(s)
j=l

where each Fj is a partial fraction of the form Rl or R2 . The parameters A, B, C, a, b, c,

and m may vary from term to term. A proof of this theorem may be found in advanced
algebra texts; see, for example, [B-MJ.
The theorem stated above, which characterizes the inverse .c-transforms of rational
functions F(s), follows from the partial fraction decomposition (34) because each partial
fraction (33) is the .c-transform of a function of the type defined in the theorem. For
example, by (18),

(35) ()
Rl S
A ) = .c{ ( A ) t m-lat}
= (s-a e .
m m-1!

A similar, but more complicated, result holds for R2(S). The case m = 1 is covered by
equations (30) and (31). For m = 2 it can be shown, using (26) and (27), that

(36) [(
(s - ex)
s - ex
)2
at. }
2]2 =.c -te smfJt ,
+ fJ
{I
2fJ

[( 1
)2 = .c { -13 eat(.
smfJt - fJtcosfJt) } .
s - ex + fJ 2]2 2fJ

The Method of Undetermined Coefficients. In Chapter 3 this method was used

to solve inhomogeneous differential equations of the form

(37) x" + ax' + bx = J(t),

where a and b are constants and J(t) is a linear combination of terms of the form
tne"t cos wt and tme"t sin wt. These are precisely the functions J(O whose .c-transforms
are rational functions F(s) = N(s)lD(s) with degree N < degreeD. In fact, the partial
fraction decomposition can be used to show that the method of undetermined coefficients
always works. See, for example, [B-RJ.
 4.2 THE LAPLACE TRANSFORMS OF SOME ELEMENTARY FUNCTIONS 229

Calculation of 'c-transforms by direct integration. Verify the given 'c-transform by using equation 0)
in Section 4.1. Justify improper integrals by integration over 0 :::: t :::: T and calculation ofthe limit T ~ 00.
Assume that a 2: 0, b 2: O.
1. 'c{sin(at») = aJ(5 2 + a2) for 5> O.
2. 'c{cos(at») = 5/(5 2 + a2 ) for 5> O.
3. 'c{tsin(at») = 2a5/(5 2 + a2)2 for 5> O.
4. ,c{ t cos (at) ) = (52 - a2)/(5 2 + a2 )2 for 5> 0

5. 'c{e"t sin(bt») = b/((s - a)2 + b2) for s> a.

6. 'c{e"t cos(bt») = (s - a)/((s - a)2 + b2) for 5> a.

7. 'c{sinh(at») = aJ(5 2 - a2 ) for s > a.

8. 'c{cosh(at») = s/(s2 - a2 ) for s > a.
Heaviside's Coverup Method for Partial Fractions.
The constants A and B in the partial fraction problem
s- 1 A B
-,----,-:--..,-
(s-2)(s-3)
= --
5-2
+ --
s-3
can be found without dearingJractions and without writing, as follows:

s- 1 I s- 1 I -B
l(s-2)I(s-3) 5-2=0 =A, (s-2)I(s-3)1 5-3=0 - .

Thefactors (s - 2) and (s - 3) are to be covered by your thumb during evaluation. This engineering technique
is call the coverup method or Heaviside's coverup method.

9. Justify the answer for A in the above example by multiplying the original problem by (s - 2) and
substituting s - 2 = O.
10. Find A and B by the coverup method:
s+1 A B
-,----,-:--..,- = - - + --.
(s+2)(s-3) s+2 s-3
11. Find A, B, and C by the coverup method:
s-1 ABC
..,....,..-.,---- = - + - - + --.
s(s-2)(s-3) s s-2 5-3
12. Argue from uniqueness (17) that

A B
'c{f(t)) = - +-
s+1 s-3

impliesJ(t) = Ae- 2t + Be3t

Partial Fractions. Use algebra or a computer algebra system to decompose the given rational function
into partial fractions. Then determine the function J(t).
13. 'c{f(t») = (s - 2)/( (52 + 4) (5 + 7))
230 4 THE LAPLACE TRANSFORM

14. £(f(t)} = (52 + 45 + 1)/((5 - 2)2(5 + 3)).

15. £(f(t)} = 53/((5 + 3)2(5 + 2)2)
16. £(f(t)} = (52 + 85 - 3)/((5 2 + 25 + 1)(5 2 + 1)).

17. £(f(t)} = (5 - 2)/(5 2 - 45 + 8)

18. £(f(t)} = (45 - 5)/(5 3 - 52 - 55 - 3)

19. £(f(t)} = (305 - 5)/(1005 2 + 105 - 56).

20. £(f(t)} = (1015 2 - 2365 + 138)/(10053 - 3345 2 + 6685 - 536).

21. £(f(t)} = (35 2 - 65 + 7)/(5 2 - 25 + 5)2.

22. £(f(t)) = (53 - 35 2 + 65 - 4)/(5 2 - 25 + 2)2.

23. £(f(t)} = 25/(5 4 +4).
Derivation of £{ eat} = s~a from the Three Properties. Establish the results below to derive the equation
from the three properties (1), (2) and (3).

24. The geometric series formula is

Lx
00
n- --
1
for Ixl < l.
n=O I-x

Derive this formula for complex x.

25. For any real constant a, the Taylor series

is valid for all t. Justify, based upon the three properties, the following steps by providing missing details.

by (11)

where x = a/5

by (14)
51 -x
1
for a real
5-a
26. Show that (15) is valid for complex a, la/51 < 1, by assembling the results of the above two problems
and verifying that everything works for a complex.
Reduction Formulas. Apply the theory of integration by parts f udv = uv - f vdu from calculus to
establish the given reduction formula.
 4.3 OPERATIONAL RULES FOR THE LAPLACE TRANSFORM 231

27. To the integral fooo e-st(t - a)nH(t - a)dt, for a> 0, apply integration by parts to show that

(38) 'c{(t - a)nH(t - a)) = ~'c{(t - a)n-1H(t - a))

s
forn=1,2, ....

28. To the integral fooo e-steattndt, for a > 0, apply integration by parts to show that

(39) 'c{eattn) = _n_'c{eatt n- 1)

s-a

forn = 1,2, ... ,s > a.

29. Prove by induction on n that equation (38) implies

n!e- as
'c{(t-a)nH(t-a))= sn+l' n=0,1,2, ... , a>O,

and that equation (39) implies

nl
'c{eattn) = . , n = 0, 1,2, ... , s > a.
(s - a)n+l

4.3 Operational Rules for the Laplace Transform

The purpose of this section is to derive a number of operational rules for the C-
transform that make it possible to find the transforms of functions j(t) that are
possibly discontinuous or derived from experimental data. Many of these rules are those
formulated by Heaviside.
C- Transforms of Power Series. Let Jet) be a function that has an C-transform, and
assume that its Taylor series converges for all t:

(1)

If the C-transform is applied to both sides of (1), then the linearity and continuity
properties of C imply that on the right C may be brought inside the summation. It then
follows from (11) in Section 4.2 that
00 j(n)(o)
(2) C{{(t)} = F(s) = L~'
n=O

provided that the last series converges for s large enough. Equation (2) with jet) = eat
was used in Section 4.2 to calculate C{eat } = lI(s - a). It is used below to find a number
of operational rules for the C-transform.
232 4 THE LAPLACE TRANSFORM

The Derivative Rule. If (1) holds thenf'(t) is also represented by its Taylor series
for all t:

(3) f'(t) = f j<n)(o) tn- 1 forall

n=1 (n - I)!
t.

Applying the operator £ to this equation, as above, gives

(4) £{f'(t)} = ~J(n)(o) = ~j<n)(o) _ J(O) ,

~ sn ~ sn
n=1 n=O
whence by (2),

(5) £{f'} = self} - J(O).

This derivative rule is fundamental for the applications of the £-transform to differential
equations. It makes it possible to transform linear differential equations for J(t) into
algebraic equations for F(s).
Of course, (5) was derived above only for those special functions J(t) with a Taylor
series representation (1). However, it holds for much more general functions. The
following theorem may be proved by using the Laplace integral (1) of Section 4.1.

Theorem. Assume that

(6) J(t) is continuous for 0 S t < 00,

(7) If(t)1 S Meat for all t ~ 0,

(8) f' (t) is integrable on every finite interval.

Then f' (t) has an £-transform, and the derivative rule (5) holds for all s > a.

The proof of this theorem is left for the exercises.

Theorem 1 can be applied withJ replaced by f'. In this way, one gets the following
corollary

Corollary. Let (6)-(8) hold forJ(t) andf'(t). ThenJ"(t) has an £-transform, and

(9) £{f"} = s2£{f} - sJ(O) - f'(0).

For the proof note that (5) is applicable with J replaced by f'. Thus

(10) £{f"} = s£{f'} - f'CO).

Combining this with (5) gives (9). A similar argument gives the follOwing corollary
 4.3 OPERATIONAL RULES FOR THE LAPLACE TRANSFORM 233

Corollary. Let (6)-(8) hold for Jet), f'(t), ... , and j(n-OCt). Then j(nlCt) has an £-
transform and

Several simple applications of the derivative rule will be given next.

EXAMPLE 1. Takej(t) = eat, sof'(t) = aeat andj(O) = 1. Then the derivative rule (5)
gives

(12)

This is equivalent to the result £{eat } = l/(s - a) that was obtained earlier by the power
series method.

EXAMPLE 2. Takej(t) = sinwt, s0f'(t) = wcoswt,jl/(t) = -w2 sin wt,j(O) = 0 and

f'(0) = w. Then (10) gives

(13) - w2£{sinwt} = s2£{sinwt} - w.

This is equivalent to the result £{sin wt} = wl(s2 + d) that was obtained earlier.
EXAMPLE 3. Consider the simple LR circuit problem of Chapter 2 with constant applied
voltage. Then
di
(14) L dt + Ri = E for t::: 0, i(O) = io,

where L, R, and E are constants, with Land R positive. If

(15) I(s) = £{i(t)}

then the £-transform of (14) gives

E
(16) L{sI(s) - io} + RI(s) = -.
s
On solving this equation for 1(5) one finds, after some algebra, that
io (ElL)
(17)
1(5) = s + (RlL) + s(s + (RlL)) .
The problem now is to find the function i(t) whose £-transform is the right-hand side
of (17). Note that

(18)
234 4 THE LAPLACE TRANSFORM

To deal with the second term, note the simple partial fraction expansion

(19)
s(s
1
+ so) = ~
1(1~ - +1)s So .

Combining this with (17) gives

(20)

Thus the uniqueness theorem for the £-transform gives the solution

(2l) .
l(t)=
(.10-R:E) e -RUL
+R:'
E
It is easy to verify that this is indeed the unique solution to (14) Note that i(t) is the sum
of the steady-state current EIR and a transient current ce- RUL , as explained in Chapter 2.

EXAMPLE 4. Consider the LC circuit of Chapter 3 with a constant applied voltage. Then
if q(t) and i(t) = q'(t) are the charge and current respectively, one has
1
(22) Lq" + - q = E for t 2: O.
C
If the circuit is initially inactive, so q(O) = 0 and i(O) = q' (0) = 0, then the £-transform
of (23), Section 4.2, gives

(23)

where Q(s) = £{q(t)}. Solving for Q(s) gives

ElL
(24) Q(s) = S(S2 + k2)'
The partial fraction expansion of (24) is found to be

(25)

whence

(26) £{q(t)} = £{EC(l- coskt)}.

 4.3 OPERATIONAL RULES FOR THE LAPLACE TRANSFORM 235

The uniqueness of C-transforms then gives the solution

(27) q(t) = EC(l - cos kt), k = 1/..JLC.

Notice the simple physical interpretation of the solution. At time t = 0 there is no current
(i(O) = 0) or charge on the capacitor (q(O) = 0). Then a constant electromotive force
E is switched into the circuit. Immediately, q(t) begins to increase, reaching a maximum
i max = 2EC when kt = :n:. Thereafter, it oscillates between q = 0 and q = 2Ee. The
mean value of q is EC, which is the steady state charge on C. There is no damping because
circuit resistance has been ignored (R = 0).
The Integral Rule. The derivative rule has a counterpart for integrals. To find it, let
f(t) be continuous for 0 :s t < 00 and satisfy

(28) [f(t) I :s Meat for O:s t < 00.

Then F(s) = .e{f} exists, and if get) is defined by

(29) get) = fotf(r) dr, t::: 0,

then dearly,

(30) g'(t) = f(t) and g(O) = o.

it it
Moreover,

(31) Ig(t)1 :s [f(r) I dr :s M M

ear dr = _(eat - 1) :s _eat,
M
a a a a
provided that a> 0 (if a :s 0 in (28) we may always replace a by a positive value). Thus
G(s) = .e{g} exists. Application of the derivative rule to g gives

(32) .e{g'} = sC{g},

since g(O) = o. By (29) and (30) this gives the integral rule:
(33) .e {l tf(r)dr} = ~C{f(t)}.
That is, the C-transform of the integral off is F(s)/s.

ExAMPLE 5. The Sine Integral. The sine integral function Si(t) is defined by

(34) Si(t) = i t
-dr.
a r
sin r
236 4 THE LAPLACE TRANSFORM

It is known that this integral cannot be expressed in terms of elementary functions. The
integrand has the Taylor series
sin t _1) nt 2n
L (2n+ 1)!'
00 (
(35)
-t- =
n=O

obtained from the Taylor series for sin t by division by t. Application of the power series
method to (35) gives

(36) .c {sin t} _ ~ (_1)n 1

-t- - ~ 2n + 1 s2n+l'

It follows from the Taylor series

L----
00 x (_1)n 2n+l
(37) tan-Ix =
2n + 1
n=O

that

(38) .c {Si: t } = tan -1 (l) .

On combining this with the definition (34) and the integral rule, one finds that

(39) .c{Si(t)} = ~tan-l (~).

Important and useful properties of the Laplace transform are expressed by the following
two rules:

First Shifting Rule. Let J(t) have a Laplace transform F(s) for s > So and let a be
any real number. Then e"tJ(t) has a Laplace transform for s > So + a that is given by

(40)

Second Shifting Rule. LetJ(t) have a Laplace transform F(s) for s > So and let a be
a nonnegative constant. Then H(t - a)J(t - a) has a Laplace transform for s > So that is
given by

(41) .c{H(t - a)J(t - a)} = e- as .c{j(t)}.

The two rules have a geometric interpretation. The first shifting rule states that multiply-
ingJ(t) by eat has the effect of shifting the graph of the transform by a units (to the right if
a ::: 0, otherwise to the left). The second shifting rule states that shifting the graph of the
original function J(t) to the right by a units has the effect of multiplying the transform by
e- as . The second shifting rule has an alternative form that is useful for some applications.
 4.3 OPERATIONAL RULES FOR THE LAPLACE TRANSFORM 237

Second Shifting Rule (Alternative Form). Let get) have a Laplace transform and
let a :::: 0 be constant. Then get + a) has a Laplace transform, and

(42) £{H(t - a)g(t)} = e- as £{g(t + a)}.

Verification of the First Shifting Rule. The result as stated follows readily from the
integral formula (1), Section 4.1, for the transform. Under the more restrictive conditions
of the power series method, equations (1), (2) give

(43)
atfC )
e t = ~j(n)(o)
~--e t,
at n
n=O n!

whence, since £{eattn} = n!l(s - a)n+l, one gets by (2)

(44)

Verification of the Second Shifting Rule. Again, the rule can be proved from either
the integral formula (1) of Section 4.1 or the power series method. For the latter, one has

(45) H(t - a)j(t - a) = L jCn)(o)

00

--H(t -
n=O n!
a)(t - at.

Moreover, (10) in Section 4.2 states that £{H(t - a)(t - a)n} = n;:;;'. Hence the linearity
and continuity properties of .[ give
00 j(n)(o)

(46) £{H(t - a)j(t - a)} = e- as " " - - I

~ sn+
= e-asF(s).
n=O

The alternative rule (42) follows from (41) if get) is defined by

(47) Jet - a) = get) for t:::: o.

Replacing t by t + a gives

(48) Jet) = get + a).

Substitution into (41) gives (42).
The Convolution. The convolution operation of Section 3.3 assigns to each pair of
functionsjU) and get) (that vanish for t < 0) a third function k(t), defined by

(49) k(t) = 1t j( r)g(t - r) dr.

238 4 THE LAPLACE TRANSFORM

The notation k = j * g was introduced, and the symmetry property j * g = g * j was

noted; that is,

(50) 1 tj(r)g(t - r) dr = 1t g(r)j(t - r) dr.

Note that the vanishing of j(t) and get) for t < 0 implies that k(t) = 0 for t < O.
Moreover, for fixed t :::: 0 the integrandj(r)g(t - r) of (49) vanishes for r outside the
interval 0 :::: r :::: t, and hence one may also write

(51) k(t) = 1'.)0 j(r)g(t - r) dr = l:j(r)g(t - r) dr.

It will now be shown that the convolution is very useful, both in calculating the inverse
Laplace transforms of complicated functions and in the Laplace transform method of
solving differential equations. These applications are based on the following.
The Convolution Rule. Let Jet) and get) be two functions of t :::: 0 whose Laplace
transforms F(s) = £If} and G(s) = £{g} exist for s > So. Then the Laplace transform
of j * g exists for s > So and

(52) £If * g} = F(s)G(s).

This rule can be proved by applying the integral formula (1) of Section 4.1 directly
to the definition (49). Here an alternative verification will be based on properties (1)-
(3), Section 4.2, and the second shifting rule. If K(s) = £{k} then (52) states that
K(s) = F(s)G(s). Application of £ to (51) gives

(53) K(s) = £ {l OO
j(r)g(t - r) dr} = 1 00
j(r)£{g(t - r)} dr,

by the continuity property (3) in Section 4.2. Next, since get) = 0 for t < 0, one has, by
the second shifting rule,

(54) £{g(t - r)} = £{H(t - r)g(t - r)} = e-TSG(s).

Substituting this into (53) gives

(55) K(s) = (1 00
e-STj(r) dr) G(s).

Now, the symmetry property (50) implies that j and g may be interchanged in (55)
without changing K. Hence

(56) K(s) = (1 00
e-STj(r) dr) G(s) = F(s) (1 00
e-srg(r) dr) .
 4.3 OPERATIONAL RULES FOR THE LAPLACE TRANSFORM 239

In particular, on taking get) = H(t) in (56), so that

G(s) = -1 =
s
1 0
00
e-stH(r) dr,

one gets

(57) F(s) = 10 00
e-STj(r) dr.

The three properties (1), (2), (3) of Section 4.2 have led full circle to the Laplace integral
(l) in Section 4.1. Finally, (56) and (57) imply that K(s) = F(s)G(s), as required.
Differentiation and Integration of Transforms. The operations of differentiating
and integrating the transform F(s) = £If} correspond to simple operations on jet). To
find them assume, at first, that jet) has the Taylor series representation (1) and transform
(2). Then formal differentiation of F(s) gives
dF(s) pn)(O)(n + 1)
I:
00
(58) ds = - sn+2 .
n=O
Recalling that 1/sn+l = £Wln!}, one finds by the continuity assumption and (1),

dFd(SS) = _ fpn) (O)(n + 1)£ {( tn+l )' }

!I : - -I
(59)
n=O n+1.
00 pnl(o)tn+l
-£
n=O n!
= £{ (-t)j(t)}.
A careful analYSiS, based on the Laplace integral (1) in Section 4.1, gives the follOwing
extended version of (59).
Derivative Rule for Transforms. Let jet) be piecewise continuous and exponen-
tially bounded (condition (7) in Section 4.1). Then
dF(s)
(60) ds = £{ (-t)j(t)} for all s > a.

Next consider the integral

(61) [00 F(a) da,

where F(s) is given by the series (2). If j(O) i- 0 then

(62)
240 4 THE LAPLACE TRANSFORM

and the improper integral (61) diverges. However,

(63) fs
oo dal
a n+
= _1_
nsn
for n = 1,2, ....
Hence, if J(O) = 0 then term-by-term integration of (2) gives

(64) f s
oo
F(a)da= L--
00

nsn
n=l
jCn)(o)

= jCn)(o).c{ tn-I}
L
00

IL'----
n=l
n (n-l)!

00 jCn)(0)t n- 1 )
=.c
n=] n!

=.c V~)}.
A careful analysis based on the Laplace integral (1) of Section 4.1 yields the following
extension of (64).
Integral Rule for Transforms. Let J(t) be piecewise continuous and exponentially
bounded (condition (7) in Section 4.1). Moreover, assume that
.
11m J(O
(65) -
[-+0+ t
exists. Then

(66) 1 00
F(a) dr = .c {J~) } for all s > a.

Sectionally Defined Functions. For applications to electric circuits and mechanical

systems it is often desirable to construct a function J(t), such as an applied voltage or
force, that is defined by different equations in different intervals. An example would be an
applied voltage vet) that is obtained from experimental data by quadratic interpolation. In
this case, vet) would be defined by different quadratic polynomials in different intervals.
Here it will be shown how the calculation of the .c-transforms of such functions is
facilitated by use of Heaviside functions and the second shifting rule. The ideas are
illustrated first by an example.

I
EXAMPLE 6. Consider the sectionally defined function

t for 0 S t S 1,
(67) J(t) = 1 for 1 S t S n12,
sin t for t ~ nl2.
 4.3 OPERATIONAL RULES FOR THE LAPLACE TRANSFORM 241

The graph of J is shown in Figure 1.

Clearly,f(t) is continuous for all t ::: O. Of course, one can calculate F(s) = £{f(t)}
directly from the Laplace integral (1) of Section 4.1. Thus

(68) F(s) = 10
1
e-stt dt + f l!:

2 e- st dt + i2
00

e- st sin t dt

and the work can be completed by calculating three integrals. However, F(s) can be
calculated without integration by use of the operational rules. To see how to do this,
recall that if a < b then

H(t - a) - HCt _ b) = jl for a < t < b,

o otherwise.

This property makes it possible to write a single formula for J(t) , namely

(69) J(t) = (t) [H(t) - H(t - 1)] + (1) [H(t - 1) - H (t - nil)]

+ sin(t)H (t - nl2) .

This may also be written as

(70) j(t) = tH(t) + (l - t)H(t - 1) + (sin t - 1)H(t - nI2).

Each of the three terms in this sum can be calculated by the second shifting rule. Thus
£{ tH(t)} = £{ t} = 1/s 2 For the next term let get) = 1 - t in the second shifting rule,
equation (42), with a = 1. Then get + 1) = 1 - (t + 1) = - t and

(71)

Finally, take get) = sin t - 1, a = n12, so that get + nl2) = sin(t + n12) - 1 = cos t - 1
in (42) and

(72) £ {(sin t - I)H (t - nl2)} = e-rrs/2 £{ cos t - I}

Y
1.0
y=l Y = sin(t)
y=t
0.0

-1.0 FIGURE 1. Graph of a sectionally defined

0.0 :rr12 3:rr12 function
242 4 THE LAPLACE TRANSFORM

Combining these results gives

(73) F(5) = -1 - -e-

s
52 52
+ e-,,512 (5+
- - - -1 ) .
52 1 5

Next consider a sectionally defined function J(t) consisting of M sections with transitions
at the points 0 = to < t 1 < ... < tM' If J(t) = In (t) for tn-l < t < tn, then generalizing
(69), one can write

(74) J(t) = Jl (t) [H(t - to) - H(t - tl)]

+ h(t) [H(t - tl) - H(t - t2)]
+ ...
+JM(t)H(t - tM-l).

Notice that if eachJn(t) is defined for all t ::: tn - l then (74) can be rearranged, as in (70),
to get

(75) I(t) = II (t)H(t - to)

+ [I2(t) - II (t)]H(t - tl)
+ ...
+ [JM(t) - JM-l(t)]H(t - tM-l).

This represents J(t) as a sum of terms whose 'c-transforms can be calculated by the
second shifting rule when the transforms of Jl (t),fl (t + tl),h(t + tl)' ... are known.
Infinite Sums. Sectionally defined functions J(t) with infinitely many transition
points 0 = to < tl < ... < tn < ... are also useful. Thus we assume that J(t) = In (t)
for tn-l < t < tn and n = 1,2, .... For simplicity it will be assumed that tn -+ +00 as
n -+ 00. Then the analogue of (75) is

=L
00

(76) J(t) [fn+l (t) - In CO] H(t - tn), t::: 0,

n=O

where we define Jo(t) == O. This infinite series is a finite sum. Indeed, for any fixed value
of t, there is an index N such that t < tN, Thus if t < tn then H(t - t n) = 0 for n = N,
N + 1, .... However, the C-transforms of the individual terms in the series will not be
zero, in general. This is illustrated below.
 4.3 OPERATIONAL RULES FOR THE LAPLACE TRANSFORM 243

Periodic Functions. A functionf(t), defined for t ~ 0, is said to have period p > °if

(77) f(t + p) = fCt) for all t ~ 0.

Graphically, this means that the graph of y = f(t) repeats itself in sections of length p;
see Figure 2.
It is natural to build periodic functions sectionally, in sections of length p, with
transition points at t = 0, p, 2p, 3p, .... To do this define on (-00, (0)

(78) fp(t) = f(t)[H(t) - H(t - p)].

The graph of y = fp(t) coincides with the curve y = f(t) for °: :

t ::: p and is zero
elsewhere. Moreover, since f(t - p) = f(t) for t ~ p, one has for t ~ p,

(79) fp(t - p) = f(t) [H(t - p) - H(t - 2p)] ,

and the graph of y = fp(t - p) coincides with y = f(t) for p ::: t ::: 2p and is zero
elsewhere. Proceeding in this way one gets the representation

(80) f(t) = fpCt) + fp(t - p) + fp(t - 2p) + ....

For any u,fpCu) = fp(u)H(u), therefore,

(81) f(t) = fp(t)H(t) + fp(t - p)H(t - p)

+ fp(t - 2p)H(t - 2p) + ....
Equation (81) is particularly well suited to the calculation of the £-transform of periodic
functions, and it gives the following result.
The Periodic Function Rule. Let f(t) have period p and define fp(t) by (78). Let
F(s) = £{f(t)} and Fp(s) = £{fp(t)}. Then

(82) Fp(s) = l P
e-stf(t) dt,

3p FIGURE 2. Graph of a periodic function

244 4 THE LAPLACE TRANSFORM

and for s > 0,

F (s)
(83) F(s) = Fp (s) (1 + e- Ps + e- 2ps + ... ) = p
1 - cps
.

To verify the first equality of (83) note that C{fp(t - np)H(t - np)} = e-nsFp(s) by the
second shifting theorem. The second equality follows by noting that e- ns = (e- S ? and
o < e- S < 1 for s > O. Thus the series in (83) is the geometric series L~ xn with ratio
x = c S and 0 < x < 1.
In the remainder of this section the periodic function rule will be used to calculate
the C-transforms of several simple periodic functions.

EXAMPLE 7. A Square Wave. The definition of the square wave function J(t) is indicated
in Figure 3.
The square wave J(t) is given by equation (80) with P = 2a and

(84) Jp(t) = H(t) - 2H(t - a) + H(t - 2a).

Integration in (82) followed by some simple algebra gives

= lo a e- st dt - 12a (1 e- )2 as
(85) Fp(s) e- st dt =- ---
o a S

Combining this and (83) with P = 2a gives

(1 - e- as )2 (1 - x)2 1 1- x
(86) F(s) = = ---
s(1 - e- 2as ) s(1 - x 2 ) S 1 + x'

where x = e- as , and hence tanh(u) = (eU- e-U)!(eU+ C U) implies

(87) F(s) = ~1 tanh (as)

2" ' s> o.
This is a compact form of the C-transform of the square wave. However, of more use for
applications is the form obtained from (86) by expanding in a Taylor series in powers of

o
a 2a 3a 4a Sa 6a
-1
FIGURE 3. Asquare wave y = J(t)
 4.3 OPERATIONAL RULES FOR THE LAPLACE TRANSFORM 245

x = e- as This gives
1
(88) F(s) = - [1 - 2e- as + 2e- 3as _ 2e- 4as + ... J.
s
Combining this with the second shifting rule gives an alternative representation the
square wave:

(89) J(t) = H(t) - 2H(t - a) + 2H(t - 2a) - 2H(t - 3a) + ....

The correctness of the equation is clear from Figure 3 and inspection.

EXAMPLE 8. A Triangular Wave. The definition of the triangular wave function get) is
indicated in Figure 4.
The C-transform G(s) can be calculated from the periodic function rule, as in the
preceding example. The triangular wave is precisely the integral of the square wave:

(90) get) = it J( r) dr for all t::: O.

It follows from the integral rule (33) and (87) that

(91) G(s) =
1
S2 tanh (as)
2" .

Circuit Problems with Continuous Inputs. Solve the following differential equations using the £-
transform method and check the solution.
1. (LR Circuit) 3i'(t) + li(t) = 10, i(O) = 1.
2. (Re Circuit) Sq'(t) + O.Sq(t) = sinOOt), q(O) = O.
Circuit Problems with Discontinuous Inputs. Solve the following differential equations by the £-
transform method. Check the solution.
3. (LR Circuit) Li' (t) + Ri(t) = vet), L = 10, R = 100, vet) = 100e- t on 3 ~ t < 00, vet) = 0 otherwise.
Find the current i(t) given i(O) = o.

4. (LR Circuit) Li'(t) + Ri(t) = vet), L = 10, R = 100, vet) = 400t on 0 ~ t < 1, vet) = 0 otherwise.
Find the current i(t) given i(O) = o.

o a 2a 3a 4a Sa 6a FIGURE 4. A triangular wave y = get)

246 4 THE LAPLACE TRANSFORM

5. (RC Circuit) Rq'(t) + (lIC)q(t) = vet), R = 100, C = 0.1, vet) = 1000 on 0 ::: t < 1.1, vet) = 0
otherwise. Find the current i(t) = q' (t) given q(O) = o.
6. (RC Circuit) Rq'(t) + (l/C)q(t) = vet), R = 100, C = 0.1, vet) = 100(t - 2) on 2::: t < 00, vet) = 0
otherwise. Find the current i(t) = q' (t) given q(O) = O.
Vibrating Systems with Continuous Inputs. Solve the following differential equations by the £-transform
method.
7. (Constant Input without Damping) /'(t) + 4y(t) = 100,y(0) = /(0) = 0 Findy(t).
B. (Sinusoidal Input without Damping) yl/ (t) + 4y(t) = 10 sin(20t), yeO) = / (0) = O. Find yet).
Vibrating Systems with Discontinuous Inputs. Solve the following differential equations by the £-
transform method.
9. (Single Square Wave without Damping)yl/(t)+2y(t) =j(t),y(O) =/(0) = O,f(t) = 1£orO::: t < 1,
j(t) = 0 otherwise. Find yCt) and the steady-state oscillation.

10. (Single Square Wave with Damping) yl/(t) + 3/(t) + 2y(t) = j(t), yeO) = /(0) = O,j(t) = 1 for
o ::: t < 1,J(t) = 0 otherwise. Find yet) and the transient state for t > l.
Applications of the First Shifting Rule. Find £([(t)}, given j(t).

11. test. 14. e2t (sin t - 3 cos t).

12. e- t sinwt. 15. et L~~o(-I)nt"

13. t 3 et - te- tl2 16. e2t L~=o ant"

Applications of the First Shifting Rule. Find j(t), given £([(t)}.

2 1
17. (s+2)2 .
22. (s-3)2'

1 s-1
lB. (s-0.5)4 .
23. (s-1)2+3 .

2 100
19. (s-1)2+4 .
24. (s+0.3)1000·

1
20. s2+2s+2 .
25. 5s 22s+5
+lOs+6 .
s-1
21. 2A+Bs
s2+4 .
26. s2+2s+2 .

Applications of the Second Shifting Rules. Find £([(t)), given j(t).

27. (t - 2)H(t - 2) 30. cos(rrt)(H(t - I) - H(t - 3»

2B. t(H(t) - H(t - 1» 31. cosh(t)H(t - 2).

29. (t - 1)2H(t - I). 32. (tet)H(t - rr).

Applications of the Second Shifting Rules. Find j(l), given £([(t)}.

e- 2s e- S
33. 36. s2+1 .
52
e- 2s e1- s
34. (s-I)2 . 37. (s-1)2+4 .

se2 - s
35. 2 cosh
s 1
s-e' 3B. s2+4s+5 .
 4.3 OPERATIONAL RULES FOR THE LAPLACE TRANSFORM 247

Applications of the t-Integral Rule. Given .c(f(t)}, findJ(t).

9
39. 52+5 . 42. S15-1
5+1'

2
40. 53+25 43. 1 5-1
. 525 2+1'

1 1 1
41. 54 -165 2' 44. 545 2+16'

Miscellaneous Applications of the Rules.

45. (Laguerre Polynomials) Apply the t-derivative rule and also the s-derivative rule to the Laguerre
differential equation

ty" (t) + (l - t)y' (t) + Ny(t) = 0

to show that yes) = .c{y(t)} satisfies the differential equation

y'(s) N +-I+N
+(- - -) yes) = o.
s 1- s

46. (Laguerre Polynomials) Solve the differential equation

y'(s) N +-I+N
+(- - -) yes) = 0, YO) = 0,
s 1- s
N
by separation of variables to get yes) = eN (5~!{
5
. Conclude that the Nth Laguerre polynomial is given
by

47. Solve forf(t), given 'c{f(t)} = In ~t~.

48. Solve for J(t), given .c(f(t)} = arctan(s/lO).
Applications of the Convolution Rule. Find.c(f * g), given J(t) and g(t).

50. sin wt, cos wt. 51. sin wt, H(t - n).

Applications of the Convolution Rule. Find h(t) by the convolution rule, given .c{h(t)}.
1 1 1 1
52. (5~l)2' 53. 52 5-4' 54, S 52+4'

Applications of the Convolution Rule to Oscillations. Find yet) by the convolution rule.
55, (Undamped Forced Oscillation) y" +y = sin 2t, yeO) = y' (0) = O.
56. (Undamped Forced Oscillation with Shifted Data) /' + y = sin t, yeO) = 1, /(0) = O.
57. (Undamped Oscillation with Discontinuous Input) y" + 16y = sin t[H(t) - H(t - n)l, yeO)
/(0) = O.

58. (Damped Oscillation with Discontinuous Input) y" + 3/ + 2y = H(t - 1) - H{t - 2), yeO) =
/(0) = O.

59. (Integral Equation) yet) = 2 + fci y{x)dx.

248 4 THE LAPLACE TRANSFORM

o rrlw 3rrlw FIGURE 5. Half-wave sin wt

60. (Integral Equation) yet) = t + e- t - J~(t - x)y(x)dx.

Extensions and Proofs. The following problems extend the theory using mathematical techniques
appearing in the proofs.
61. (Extended Derivative Rule) Assume the t-derivative rule hypotheses, except that Jet) has a single finite
jump discontinuity at t = ta > O. Show that the rule holds in the revised form

£(f'(t)} = s£{f(t)} - j(O) - [jUa + 0) - j(ta - O)]e-tos,

where j(ta+O) and jUa -0) are the right- and left-hand limits ofj(t) at ta. Why does the extra term arise?
62. (Convolution Example) Letj(t) = 1/0, get) = 1/.;rt=li. Show that J~j(x)g(t - x)dx fails to exist
at t = 1 and explain why this example is excluded by the hypotheses of the convolution theorem.
63. (Convolution Rule) The main step in the proof of the convolution rule is the identity

fXl LOO F(t,x)dtdx = 1000 f F(t,x)dxdt,

where F(t,x) = j(x)e-stg(t - x) (s is constant in the argument). Justify the interchange of dx and dt
and the new limits of integration by citing a reference from advanced calculus.
£-Transforms of Periodic Functions. The following problems assume a periodic input.
64. (£-transform of a Triangular Wave) Use direct integration to calculate the integral

fo2a e-stg(t) dt,

where get) is the triangular wave of Figure 4. Then use the periodic function rule to verify equation
(91) for the £-transform of g(t).
65. (Half-Wave Rectifier) Determine £(f(t)}, wherej(t) with period 2rrlw is the half-wave rectification of
sinwt in Figure 5.
66. (Full-Wave Rectifier) Determine £(f(t)}, where j(O with period 2rrlw is the full-wave rectification of
sin wt in Figure 6.
67. (Sawtooth Wave) Letj(t + p) =Jet) andj(t) = kUp for 0 :'S t :'S p. Compute £(f(t)} .

.... t
o rrlw 3rrlw FIGURE 6. Full-wave sin wt
 4.4 ApPLICATIONS TO DiffERENTIAL EQUATIONS 249

68. (Steady-State Current) Find the steady-state current in an LR-circuit with 2a-periodic square wave
input vet) = A for 0 :::: t < a, vet) = 0 for a :::: t < 2a.
Power Series Methods. The following problems use a Taylor series representation to derive formulas for
the C-transform.
69. (Derivative Rule) Assume that l' (t) is represented by its Taylor series

L -(-- tn-
00 j(n) (0)
1'(t) = 1 for all t.
n=1 n - 1)1

Apply the operator C to this equation to obtain

C{f'(t)} = ~ln)(o) = ~j(n)(o) -j(O).

L.. sn L.. sn
n=1 n=O

Conclude by (2) that C{f'} = sC{f} - j(O).

70. (Integral Rule) Assume that jet) is represented by its Taylor series (1). Then term-by-term integration
gives

(92) io
t
j(r)dr= L (n+l!)
00

n=O
j(n)(O)t n+1

Derive the t-integral rule by applying the C-transform to the two sides of (92).
71. (First Shifting Rule) Assume thatj(t) is represented by its Taylor series (1). Then

Use the rule Cleatt n } = nlf(s - a)n+l to obtain the first shifting rule

C{eatj(t)} = f: j(n)(O~1 = F(s -

n=O (5 - a)n
a).

4.4 Applications to Differential Equations

This section illustrates the utility of the .c-transform by applications to simple electrical
and mechanical oscillators. The examples are governed by linear first- and second-order
equations with constant coefficients. Such equations were already solved in Chapter 2
and Chapter 3. Here the goal is to show the usefulness of the .c-transform in solving
initial value problems and in analyzing the structure of their solutions. The section begins
with an analysis of simple LR and RC circuits governed by linear first-order equations.
Then electrical and mechanical oscillations governed by linear second-order equations
are treated.
The plan for using the .c-transform to solve initial value problems for linear
differential equations with constant coefficients is clear from Section 4.3, Examples 3 and
250 4 THE LAPLACE TRANSFORM

4. The .c-transform of the differential equation is used to obtain an algebraic equation for
the .c-transform Xes) = .c{x(t)}. The equation is solved algebraically for Xes). Finally,
x(t) is constructed from its .c-transform X(s). This last step is based on the operational
rules of Section 4.3, supplemented by transform tables and computer programs for
calculating inverse transforms.
LR and RC Circuits. The LR circuit equation of Section 2.3,
di
(1) L- + Ri = E(t),
dt
and the corresponding RC circuit equation

(2) Rdq +i = E(t)

dt C '
are clearly isomorphic 2 under the correspondence L -+ R, R -+ llC, i -+ q. They are
equivalent to the general first -order linear equation with constant coefficients

(3) X' + ax = f(t), t::: O.

No Input (j(t) == 0). The homogeneous equation x' + ax = 0 has solution x = ce-at,
where c = x(O) can have any value. To reproduce this formula via the .c-transform, we
note thatthe transform of x' +ax = 0 is sX -x(O)+a.X = 0, whenceX(s) = x(O)/(s+a).
The inverse transform, from Section 4.2 or a table, is the solution x(t) = x(O)e- at .
Constant Input (j(t) == fo). The equation

(4) x' + ax = fo = constant

was solved via the .c-transform in Section 4.3, Example 3. The solution formula (21) in
Section 4.3 with i = x, L = 1, R = a, E = fo gives

(5) x(t) = (X(O) -~) e-at +~.

If a > 0, then the first term on the right, of the form ce- at , is a transient term, and

(6) lim x(t) =

1-+00
h..
a
Clearly, xp(t) = fola = constant is the steady-state solution of (4) to which all solutions
converge as when t -+ +00.
Sinusoidal Input. Let w > 0 and consider the equation

(7) X' + ax = sinwt.

2This term was defined in Chapter 3. It means that the equations are the same by renaming the letters.
 4.4 ApPLICATIONS TO DIFFERENTIAL EQUATIONS 251

The solution withx(O) = 0 has C-transform Xes), which satisfies (s + a)X(s) = W/(S2 +
or
( 2 ),

(8) Xes) = w =~+ Bs + C .

(s+a)(s2+w 2) s+a S2+W 2
The last form of Xes) is guaranteed by the theory of partial fractions. A simple algebraic
calculation gives A = -B = w/(w 2 + a2), C = aw/(w 2 + a2) and hence

C
(9) x(t) = Ae-at + B cos wt + - sin wt
w
w ) -at asin(wt) wcos(wt)
= ( e + - .
w 2 + a2 w 2 + a2 w 2 + a2
Clearly, the first term on the right is a transient term, while
a w
(10) xss(t) = 2 2 sin(wt) - 2 2 cos(wt)
w +a w +a
is the steady-state solution. This function is periodic, with the period of the input
function, even though the homogeneous equation has no periodic solutions. Of course,
the method of undetermined coefficients can also be applied to (7) to yield the solution
(10).
Periodic Square Wave Input. Next consider (3) with a = 1 and periodic square wave
input jet) with period 2. Thus jet) is given by (81) in Section 4.3 with p = 2,

(ll) f(t) = fp(t)H(t) +fp(t - 2)H(t - 2)

+ fp(t - 4)H(t - 4) + ... ,
while fp(t) is defined by (84) of Section 4.3 with a = 1,

(12) jp(t) = H(t) - 2H(t - 1) + H(t - 2).

Search for a Periodic Solution. Before analyzing the solution of

(l3) X' + x = fp(t)H(t) + jp(t - 2)H(t - 2»

+jp(t - 4)H(t - 4) + ...
via the C-transform, it will be helpful to study whether it can have a periodic solution
with the same period p = 2 as the input function. Thus, a solution of (13) will be sought
that is defined and satisfies the periodicity condition

(14) x(t + 2) = xU) for all t::: O.

252 4 THE LAPLACE TRANSFORM

Of course, the general solution of (13), as given in Chapter 2, is

(15) x(t) = x(O)e- t + i t e-(t-r)j(T) dT,

but this is of no immediate help in finding a periodic solution. Note that any solution of
(13) must be continuous and satisfy

(16) x'+x=I, or x(t) = I+c1e-t, O~t~I,

x'+x= -1, or x(t)=-I+c2e-t, I~t~2.

The constants C1 and C2 must be determined so that x(t) is continuous and periodic.
First, x(t) must be continuous at t = 1, so C1 and C2 must satisfy

(17)

Next, to be a periodic solution, x(t) must satisfy (14) with t = 0; that is, x(O) = x(2), or
(I8)

Now, (17) and (18) are a pair of linear equations for C1 and C2. By Cramers rule or the
method of elimination these equations have the unique solution
-2e 2e2
(I9)
C1 = e + l' C2=--·
e+I
Thus, if (13) is to have a periodic solution x(t), then for 0 ~ t ~ 2 it must be given by

2e)
1- ( e + 1 e- t for 0 ~ t ~ I,
(20) {
xp(t) = -1+ (2e
--
)
e
-t
for 1 ~ t ~ 2.
e+I
Moreover, we have
I-e
(2l) x(O) = x(2) = --.
I+e
This value determines the unique continuation of x(t) to t ~ 2. Since x(2) = x(O), the
continuation should be simply xp(t), translated to the right 2 units; i.e., xp(t - 2). To
check this, notice that by (ll), (12) and (13),

(22) X' + x = jp(t - 2) for 2 ~ t ~ 4,

and (2l) holds. To identify the solution of this initial value problem, let

(23) t- 2 = r, x(t) = yeT),

 4.4 ApPLICATIONS TO DIFFERENTIAL EQUATIONS 253

and make the change of variables (t,x) ~ (r,y) in (21), (22) to obtain

dy
(24) - +y =!p(r) for 0::: r::: 2,
dr
l-e
yeO) = - .
l+e
But the unique solution of this was shown above to be y(r) = xp(r). Thus, by (23),

(25) x(t) = xp(t - 2) for 2::: t ::: 4,

as expected. Similarly, (13) on 4 ::: t ::: 6 withx( 4) = (l-e)/(l +e) givesx(t) = xp (t-4),
4 ::: t ::: 6, etc. This argument proves the following result.

Theorem. The LR circuit equation with periodic square wave input, equation (13), has
a unique periodic solution x(t) = xss(t) defined by

(26) xss(t) = xp(t)H(t) + xp(t - 2)H(t - 2)

+xp(t - 4)H(t - 4) + ... ,
where xp(t) is defined by (20) for 0 ::: t ::: 2 and is zero elsewhere. Solution xss(t) is the
unique solution generated by the initial value xss(O) = (l - e)/(l + e). A graph of xss(t)
on 0 ::: t ::: 6 is shown in Figure 7.
The function xss(t) defined by (26) is a particular solution of the inhomogeneous
equation (13). Therefore, one has the following corollary

Corollary. The general solution of (13) is

(27) x(t) = xss(t) + ce- t ,

where c is an arbitrary constant. Hence, xss(t) is the steady-state solution of (13) to which
all other solutions converge.

x
0.6

0.0

-0.6
FIGURE 7. Graph of the steady state
o 6 response of an LR circuit
254 4 THE LAPLACE TRANSFORM

'c-Transform Solution of the Square Wave Problem. Equation (27) will be

derived by the 'c-transform method. It will be convenient to rewrite (13) using (89) in
Section 4.3 as

(28) x' +x = H(t) - 2H(t -1) + 2H(t - 2) - 2H(t - 3) + ....

The 'c-transform of this is
1 e- s e- 2s e- 3s
(29) sX - x(O) + X= - - 2- + 2- - 2- + ... ,
s s s s
whence

(30) Xes) = x(O) + K(s) [1 _ 2e- s + 2e- 2s _ 2e- 3s + ... J

s+ 1 '
where
1 1 1
(31) K(s) = 'c{k(t)} = = - - -,
s(s + 1) s s+ 1
and hence

(32) k(t) = 1 - e- t .

Now the second shifting rule gives

(33) K(s)e-ns = 'c{H(t - n)k(t - n)}, n = 0, 1,2, ....
Thus (30) is the 'c-transform of the equation

(34) x(t) = x(O)e- t + k(t) - 2k(t - 1)H(t - 1)

+ 2k(t - 2)H(t - 2) - ....

Note that for any particular value of t, the series in (34) terminates. Thus, for 2n < t <
2n + 1,
(35) x(t) = x(O)e- t + (k(t) - 2k(t - 1) + 2k(t - 2) + ... + 2k(t - 2n»

Next, note that by equation (32),

(36)

Substituting this into (35) gives, for 2n < t < 2n + 1,

(37) x(t) = x(O)e- t + 1- e- t (1 - 2e + 2e2 - ... + 2e2n ) .

Let Sln = 1 - 2e + 2e 2 - ... + 2e 2n . By the geometric sum formula
2 n-l I-un
l+u+u +···+u =--
l-u
 4.4 ApPLICATIONS TO DIFFERENTIAL EQUATIONS 255

it follows that

(38) 52n = 1 - 2e + 2e2 - ... + 2e2n

- 1 + 2(1 - e + e2 - ... + e2n )

= -1 +2 (
1 + e2n +!)
I+e
1- e 2e2n+1
=--+--.
I+e I+e
Combining (37) and (38) gives for 2n < t < 2n + 1,

x(t) = ( x(O) - 1 - -e) e- t + 1 _ __

- 2e e-(t-2n).
(39)
I+e l+e
Comparing this with (20) gives for 2n < t < 2n + 1,

(40) x(t) = ce- t + xp(t - 2n)

= ce- t + xss(t),
where
I-e
(41) c=x(O)---.
I+e
This verifies (27) for 2n < t < 2n + 1. A similar calculation shows that

(42) x(t) = ce- t + xp(t - 2n - 1)

= ce- t + xss(t) for 2n + 1 < t < 2n + 2,

and thus (27) holds for all t 2: O. In particular, (40), (41), (42) imply that x(t) = xss(t)
if and only if x(O) = (1 - e)/(l + e).
Electrical and Mechanical Oscillators. The electrical and mechanical oscillators
that were analyzed in Section 2.5 and Section 3.3 are modeled by the initial value problem

(43) x" + ax' + bx = jet), t 2: 0,

x(O) = Xo, x/CO) = VA,
where a and b are non-negative constants and j(t) is a prescribed "input" function. Here
it will be shown how readily the results of Section 3.3 are derived by the L-transform
method. The section ends with an analysis of the response to a periodic square wave.
The Homogeneous Equation. If jet) == 0 then the L-transform of (43) gives

(44) (S2X - sXo - vo) + a(sX - xo) + bX = 0,

256 4 THE LAPLACE TRANSFORM

whence

(45) Xes) = sXo + (vo - axo).

S2 + as + b
Three cases arise corresponding to roots of the denominator in (45). Thus one has the
following three cases (see (4), (5), (6), Section 3.3).
Case 1. (a 2 - 4b > 0). Then S2 + as + b = (A - Al)(A - A2), where the roots Al and
A2 are real and distinct. The partial fraction expansion of Xes) in this case is

(46)

with corresponding solution

(47)

The constants AI, A2 can be calculated from the initial values Xo and Yo.
Case 2. (a 2 - 4b = 0). Then S2 + as + b = (s - A)2, where A = -al2 is a real double
root. Thus
Al A2
(48) Xes) = -s--A + -(S---A)-2 '

and

(49)

Case 3. (a 2 - 4b < 0). Then 52 + as + b = (s + I)2 + y2, y = .Jb - a2/4, and Xes)
can be written

(50)

Therefore,

(51)

The Nonhomogenous Equation. The £-transform of (43) with initial values Xo =

0, Vo = 0 leads to the equation

(52) Xes) = F(s)

52 + as + b
Moreover, the equation
1
(53) K(s) = £{k(t)} = 2 b
5 +as+
 4.4 ApPLICATIONS TO DIFFERENTIAL EQUATIONS 257

can be solved for k(t) by the partial fractions method. Thus

(54) Xes) = K(s)F(s)

and the convolution rule gives

(55) x(t) = (k *f) (t) = 1t k(t - r)J( r) dr.

This reproduces the particular solution of (43) that was found in Section 3.3 by the
method of variation of parameters.
The Transfer Function. Equation (54) represents Xes) = C{x(t)}, the C-transform
of the oscillator response, as a product of a factor K(s) that depends only on the
physical properties of the oscillator and a factor F(s) = C{f(t)} that depends only on the
input function. The function K(s) is called the transfer function of the oscillator. This
simple example illustrates a concept that is widely used by engineers in the analysis of
complicated electrical networks and other linear systems. Additional examples of transfer
functions are given in Section 4.5.
The exact form of k(t) depends on the three cases discussed above. In Case 1 it is
easy to check that

(56)

and hence
~lt_~2t
(57) k(t) = .
Al - A2
This agrees with (19) in Section 3.3. The other two cases follow similarly
Damped Oscillator with Sinusoidal Input. This problem was solved in Sec-
tion 3.3, equation (50) and following. Here the following example will be solved by the
C-transform method:

(58) x" + lx' + 5x = sint, t::: 0,

x(o) = 0, x'(O) = o.
Application of the C-transform gives

1 As + B C(s + 1) + 2D
(59)
Xes) = (S2 + 1)(S2 + 2s + 5) = S2 + 1 + (s + 1)2 + 22 '
258 4 THE LAPLACE TRANSFORM

where the second equality is the appropriate partial fractions expansion. Simple algebra
gives the coefficients
1 1 1 -1
(60) A=--
10'
B=-
5'
c=-
10'
D--
- 20'
Thus one has
1 s 1 1 1 s+1
(61)
Xes) =- 10 S2 + 1 + 5" S2 + 1 + 10 (s + 1)2 + 22
1 2
20 (s + 1)2 + 22 '
where the last line is convenient for the use of C-transform tables. The latter give
1 1 1 1
(62) x(t) = - - cos t +- sin t + _e- t cos(2t) - _e- t sin(2t).
10 5 10 20
This can be written

(63) x(t) = xtr(t) + xss (t) ,

where
1 1
(64) XtrCt) = _e- t cos(2t) - _e- t sin(2t)
10 20
is a transient term, a solution of the homogeneous equation, and
1 1
(65) xss(t) = - - cos t + - sin t
10 5
is the steady-state solution. Again, comparing with the methods of Chapter 3, xss(t) is
the particular solution of (58) provided by the method of undetennined coefficients.
Damped Oscillator with Periodic Square Wave Input. Consider a damped
oscillator with a periodic square wave input j(t) of period 2n:

(66) x" + 2x' + 5x = jCt) , t 2: 0,

x(O) = 0, x'(O) = 0,

where by (89), Section 4.3,

(67) jet) = H(t) - 2H(t - n) + 2H(t - 2n) - 2H(t - 3n) + ....

The operational rules of Section 4.3 applied to these equations give

(68) (S2 + 2s + 5)X(s) = F(s),

 4.4 ApPLICATIONS TO DIFFERENTIAL EQUATIONS 259

where
1
(69) F(s) = - [1 - 2e- rrs + 2e- 2rrs _ 2e-3rrs + ... J.
5

Combining these, one finds that

(70) Xes) = M(s) [1 - 2e- rrs + 2e- 2rrs _ 2e-3rrs + ... J'

where
1
(71) M(s) = [{met)} = [( )2 2]·
55+1 +2
The method of partial fractions gives, after some algebra,
1 1 5 +1 1 2
M(s) = 5s - 5(s
(72)
+ 1)2 + 22 10 (s + 1)2 + 22
and hence
1 1 1
(73) met) = - - _e- t cos(2t) - _e- t sin(2t).
5 5 10
It will also be convenient to define
1
(74) let) = -(2cos(2t) + sin(2t»e- t
10
so that
1
(75) met) = - - let).
5
Combining (70) and the second shifting rule, M(s)e- mrs = H(t -rrn)} [{ met -rrn), gives
the solution

(76) x(t) = m(t)H(t) - 2m(t - rr)H(t - rr)

+ 2m(t - 2rr)H(t - 2rr) - ....

Also, since cos(2t) and sin(2t) have period rr, one has from (74) and (75),

(77) met - nrr) = 51 - l(t

) enrr , n = 0,1,2,· ...

Combining this with (76) gives the following two cases.

Case 1 «2n - l)rr < t < 2nrr).

(78) x(t) = met) - 2m(t - rr) + 2m(t - 2rr)

- ... - 2m(t - (2n - l)rr)
1
= - - - l(t) (1 - 2err + 2e2rr _ ... _ 2eC2n-ll7r) .
5
260 4 THE LAPLACE TRANSFORM

The last sum can be evaluated by means of the geometric series

Indeed, if a = -f!', then

(80) 1 - 2eIr + 2e1Ir - ... - 2e(ln-l)Ir = 1 + 2a + 2a1 + ... + 2a 2n - 1

- 1 + 2(1 + a + a 2 + ... + a 1n- 1)

1- a 2n
-1+2·--
I-a
1+a 2 - a 2n
I-a I-a
1 - f!' e2nn
=---2--.
l+e Ir l+eIr
Combining this with (78) and using (74) gives

(81) x(t) = _ ~ -let) [1 - f!' _ 2l+e

e2nIr ]
5 l+e Ir Ir

( eIr+l
f!'-I)l_~
5
2 e-(t-2Irn)
+ - -Ir . (2cos(2t) + sin(2t»
l+e 10
1 2
= cl(t) - - + - - l e t - 2Jrn)
5 1 + eIr '

where c = (f!' - l)/(f!' + 1). A similar analysis gives the following.

Case 2 (2nJr < t < (2n + l)Jr). By methods similar to case 1 above,

1 2
(82) x(t) = d(t) +- - - - l (t - (2n + l)Jr).
5 1 +eIr
Note that the term d(t) is a transient term, tending to zero as t -+ 00. If this term is
dropped in (81) and (82) then the remaining terms define a periodic solution xss(t), of
period 2Jr. For the interval 0 .::5 t .::5 2Jr we have from (82) with n = 0, and also (81)
with n = 1, xss(t) = xp(t), where

1 2
- - - - l e t - Jr) 0.::5 t .::5 Jr,
(83) 5 l+eIr '
1 2
-- + - - lIr( t - 2Jr) Jr .::5 t .::5 2Jr.
5 l+e '
 4.4 ApPLICATIONS TO DIFFERENTIAL EQUATIONS 261

Define xp(t) = 0 for t > 2n. We shall show that

(84) xss(t) = xp(t)H(t) + xp(t - 2n)H(t - 2n)

+ xp(t - 4n)H(t - 4n) + ...
is the periodic (and steady-state) solution of period 2n. To verify this, note that for
2n ::::: t ::::: 4n one has from (82) with n = 1 and also (81) with n = 2,

~- _2_l(t - 3n) 2n ::::: t ::::: 3n,

5 1 +e"" '
(85) xss(t) = { 1 2
-- + - - l e t - 4n) 3n ::::: t ::::: 4n.
5 1 +e"" '
This agrees on the interval2n ::::: t ::::: 4n with (84). The general case follows from (81),
(82). Finally, we can write

(86) x(t) = xss(t) + Xtr(t) ,

where Xtr(t) = cl(t) is a transient term of the form
(87)

Exercises 4.4

Applications to Initial Value Problems. Use the C-transform and the initial values to calculate the C-
transform Xes) = C{x(t)}. Then find the partial fractions form of Xes). Finally, use tables or a computer
algebra system to determine x(t).
1. x' - 0.05x = 0, x(O) = 200.
2. x" - 3x' + 4x = 0, x(O) = 1, x'(O) = 2.
3. x" +x' + 1.25x = O,x(O) = 2,x'(0) =-1.
4. 9X" - 6,( +x = O,x(O) = 5,x'(0) = o.
5. x" + 20x' + 200x = 0, x(O) = 0, x' (0) = 20.
Applications to Finding Particular Solutions. Find a particular solution of the nonhomogeneous
equation by applying the C-transform method with initial values xCO) = 0, x'CO) = o.

6. x" + 2x' + x = cos2 t.

7. x" +x' - 2x = e- t .
8. 10x" -7x' +x = 30 sin 2t.
9. x"-x' =et +2e2t .
10. x" +x' - 6x = t2 - 4.
Applications to Higher-Order Equations. Solve the initial value problem by the C-transform method.
262 4 THE lAPLACE TRANSFORM

II. ,I" + 6x" + llx' + 6x = 0, x(O) = 0, x'(O) = 0, x"(O) = l.

12. x"" + 2,1" + 2x" + 2x' + x = 0, x(o) = 0, x'(O) = 0, x"(o) = 0, x",(O) = l.
13. ,I" + 3x" + 4x' + 2x = 0, x(O) = 0, ,1(0) = 0, x"(O) = l.
14. 9x"" - 24,1" - 29x" + 120x' - 80x = 0, x(O) = 0, x'(O) = 0, x"(O) = 0, xll/(O) = l.

15. 2x'" + 3x" - 12x' + 7x = 0, x(O) = 1, x'(O) = 0, x"(O) = 0.

16. 4x'" - Sx" - 2,1 + 3x = 0, x(O) = 0, x'(O) = 1, x"(O) = 0.

17. 100x'" - 334x" + 668x' - 536x = 0, x(O) = 0, x'(O) = 0, x"(o) = l.

18. 100x'" - 334x" + 768x' - 670x = 0, x(O) = 0, x'(O) = 0, x"(O) = l.
19. 9x"" - 42x'" + 22x" + 126x' - 147x = 0, x(O) = 0, x'(O) = 0, x" (0) = 0, xll/(O) = l.

20. 2500x'" + 9600x" - 14325x' - 55008x = 0, x(O) = 0, x'(O) = 0, x"(O) = l.

Applications to Electrical Circuit Problems.
21. Find the general solution of the LRC circuit equation with emf E = 0, capacitor of 0.001 farads in series
with an inductor of 0.4 henries. No resistor.
22. (LRC Circuit) Find the general solution of the LRC circuit equation with emf E = 0, resistor of 20
ohms, inductor of 4 henries, and capacitor of 0.008 farads.
23. (LRC Circuit) Compute the transient current and plot its graph if E = 0, L = 0.10 henries, R =2
°
ohms, C = 20J1,F, Ve = 200 volts (voltage on capacitor), i(O) = and i'(0) = -VeIL.
24. (iRC Circuit) Compute the steady-state current and plot its graph if E = 0, L = 0.10 henries, R =2
°
ohms, C = lOJ1,F, Ve = 100 volts (voltage on capacitor), i(O) = and i' (0) = -VeIL.
°
° °
25. (iRC Circuit Voltage) Compute the voltage function V(t) = q(t)IC on the interval ~ t ~ 3.0, given
L = 1 henry, R = 2 ohms, C = 0.5 farads, E(O = 10 for ~ t < 1, E(t) = elsewhere. Assume zero
initial charge and current.
°
°
26. (LRC Circuit Voltage) Compute the voltage function Vet) = q(t)/C on the interval ~ t ~ 2.0, given
L = 1 henry, R = 2 ohms, C = 0.5 farads and E(t) = 10 for ~ t < 1, E(t) = 2 for 1 ~ t < 2,
°
E(t) = elsewhere. Assume zero initial charge and current.
Damped Oscillations with Triangular Wave Input. Consider the damped oscillator problem

x" + 2x' + 5x = j(t) , t~ 0,

x(O) = 0, ,1(0) = 0.

The inputj(t) is the 271"-periodic triangular wave that on [0,271"1 agrees with the function

-
jet) = {t7I"-t °
71"
~ t < 71",
~ t < 271".

In the exercises below, the general solution is obtained, and the steady-state solution is computed.
27. Use the L:-transform and the convolution method to solve the damped oscillator problem in convolution
integral form.
28. Evaluate the convolution integral x(t) of the damped oscillator problem on the interval
Hint: On [71",271"1 the answer is
° ~ t ~ 271".

x(t) = (571"e"" + 6e"" - 3)e- t sin(2t)/50

+ (571"e"" - 4elT + 2)e- t cos(21)125
+ (2 + 571" - 50125.
 4.5 ApPLICATIONS TO SYSTEMS OF DIFFERENTIAL EQUATIONS 263

29. Plot the oscillator solution x(t) on the interval 0 :::: t :::: 2n.

30. Calculate the initial values of the steady-state solution of the damped oscillator problem and plot this
solution on [0,2n].

Hint: The steady-state solution x(t) = Xh(t) + xp(t) has period 2n and therefore it satisfies
x(O) = x(2n), x' (0) = x' (2n). In these equations, xh (t) = (q cos(2t) + c2 sin(2t))e- t and xp(t)
is the particular solution defined and plotted above, satisfyingxp(O) = x~(O) = O. Solve equations
xeO) = x(2n), x' (0) = x' (2n) for q, C2 and plot xU) on 0 :::: t :::: 2n.

4.5 Applications to Systems of Differential Equations

Many engineering problems involving electrical networks and mechanical systems
are governed by systems of two or more linear differential equations with constant
coefficients, involving two or more unknown functions. Each equation may contain
derivatives of the first or higher order of each of the unknown functions. Two such
problems were given in Section 3.1, and several more are analyzed in this section. The
purpose of this section is to show that the .c-transform is a natural and efficient tool to
solve such problems.
The .c-transform method is particularly effective in the solution of systems with
a large number n of equations and unknown functions. However, it is difficult and
cumbersome to analyze such large systems without introducing the concise notation of
linear algebra and matrix theory. Therefore, this section will be restricted to cases of
n = 2 equations. Linear algebra will be developed in Chapter 5.
Engineering literature and textbooks provide many examples of systems of linear
differential equations that can be solved by the .c-transform method. A particularly rich
source of examples is the textbook Mechanical Vibrations by Den Hartog (see [DenH]),
although it does not use the .c-transform. The classic textbook by Churchill (see [Ch-B])
also has many excellent examples. The examples discussed below are mostly variants of
problems from these sources. They were chosen to illustrate the principles without undue
algebraic complications.
A Two-Loop LRC Circuit. Consider the two-loop circuit of Figure 8. The basic
unknown functions are the charges ql (0 and q2 (t) on the two capacitors. Kirchhoffs
laws imply the two differential equations (50), (51) in Section 3.1; that is,

(1)

(2)
264 4 THE LAPLACE TRANSFORM

ql

~ ~i2

CI i3 q2
iI
C2
t

FIGURE 8. Two-loop LRC circuit

Here only the special case L = 6, R = 10, l/C I = 6, 1/e2 = 60 will be discussed. In
this case (1), (2) become, after removal of constant factors,

(3) 3q~ + 5q~ - 5q; + 3ql = 0,

(4) q; - q~ + 6q2 = o.
In Chapter 3 this system was solved by eliminating q2 to obtain a single third-order
equation for ql. Here it will be solved by applying the £-transform directly to the system
(3), (4). The system is second order in ql and first order in q2. Hence the appropriate
initial data are

(5)

It will be shown that the system has one and only one solution for each choice of ai, a2, bl .
On transforming (3), (4) and using the operational rules of Section 4.3, one gets for
the transforms Ql (s) = £{ql (O}, Q2(S) = £{q2(t)} the pair of equations

(6) (3s2 + 5s + 3)QI - 5SQ2 = (3s + 5)al - 5a2 + 3bl ,

(7) -SQI + (s + 6)Q2 = a2 - al.

The determinant of this system is the cubic polynomial

(8) D(s) = (3s 2 + 5s + 3)(s + 6) - 5s 2

= 3(S3 + 6s 2 + lIs + 6)
= 3(s + 1)(s + 2)(s + 3).

Solving the system (6), (7) gives

(9) QI(S) = PI(S) = ~ + ~ +~,

D(s) S+ 1 S+ 2 S+ 3

Q2(S) = P2(s) = ~ + ~ +~,

D(s) S+ 1 S+ 2 S+ 3
 4.5 ApPLICATIONS TO SYSTEMS OF DIFFERENTIAL EQUATIONS 265

where PI(S) and P2(s) are quadratic polynomials whose coefficients involve at, a2, bl ·
The right-hand sides of (9) are the partial fractions appropriate to the denominator D(s).
Clearly, (9) implies that

(0) ql (t) = Alc- t + A 2c- 2t + A 3c- 3t ,

q2(t) = BJ c- t + B2c- 2t + B3c- 3t ,
where the coefficients are to be determined. Here only the special case

(ll)

will be calculated. This gives

3s + 5-5s
(12) PI (s) =
-1 s+ 6
= (3s + 5)(s + 6) - 5s
= 3(S2 + 6s + 10),

whence, by a simple calculation, Al = 512, A2 = -2, A3 = 112. A similar calculation

gives BI = -3/2, B2 = 3, B3 = -312. Thus

5 1
(13) ql CO = 2:c-t - 2e- 2t + 2: c- 3t ,
1 _/
q2(t) = - -e + e-2t --e
1 -3t
2 2
It is easy to check that (13) satisfies the original system (3), (4) and the initial conditions
(11).

A Two-Body Undamped Mechanical Oscillator. Consider the system of two

masses and three springs as depicted in Figure 9.
The two bodies are assumed to have equal masses mI = m2 = 1, while the springs are
identical, with Hooke's constant k > O. The mass displacements Xl and X2 are measured
from their equilibrium positions. The bodies are assumed to slide without friction on a
horizontal surface, so there are no dissipative forces in the system.

FIGURE 9. Two-body undamped mechanical oscillator

266 4 THE LAPLACE TRANSFORM

Each body experiences two Hooke's law forces, from the springs on either side of
them. Thus Newton's law gives the pair of equations

(14) x~ == - kxI + k(X2 - Xl),

X~ =- k(X2 - Xl) - Rx2,

or, collecting all terms on the left,

(15) X~ + 2kxl - kx2 = 0,

X~ + 2kx2 - kxl = o.
Synchronous Oscillations. Physical intuition suggests that the solutions of (15) are
oscillatory, and it is natural to look for synchronous oscillations of the form

(16)

since the second derivative of sin wt is -w 2 sin wt. Substituting these into (15) gives,
after some rearrangement, the pair of equations

(17) (2k - ( 2 )c1 - kC2 = 0,

-kCI + (2k - (2)C2 = O.
What is wanted is a pair (CI, (2) =f:. (0,0) and a positive constant w that satisfy (17). By
elementary algebra, such a solution exists if and only if the determinant of the system is
zero:

(18)

This has two positive solutions: w = ,,;'k and w = .J3k. Thus (17) has the solutions
w = ,,;'k, CI = C2 = 1 and also w = .J3k, CI = -C2 = 1. This gives the solution pairs
(19) Xl = sin -Jkt, X2 = sin -Jkt,
Xl = sin J3kt,X2 = - sin J3kt.
Moreover, one can replace sin wt by cos wt in (16) to get the two additional solution pairs

(20) Xl = cos -Jkt, X2 = cos -Jkt,

Xl = cos Y3kt, X2 = - cos J3kt.

Thus four distinct (linearly independent!) solutions have been found. Moreover, if each
is multiplied by a constant and they are added together, new solutions are obtained,
because system (15) is linear and homogeneous. The L:-transform will be used to show
that all solutions arise in this way.
 4.5 ApPLICATIONS TO SYSTEMS OF DIFFERENTIAL EQUATIONS 267

Solution via the 'c-transform. System (15) is second order in both Xl and X2, and
hence appropriate initial values are

(21) XI(O) = al'x~(O) = bl ,

X2(0) = a2,x;(0) = h
It will be shown that system (15) has exactly one solution for each choice of data (2l),
and the solution is a linear combination of the four solutions (19), (20). To do this
we introduce the transforms XI(s) = 'c{XI(O}, X 2 (s) = 'c{X2(t)} and 'c-transform the
system (21). The result is

(22) (5 2X I - sal - bl ) + 2kX I - kX2 = 0,

(S2X2 - sa2 - b2) + 2kX2 - kX I = O.

This may be written as a system of linear equations for Xl, X 2 :

(23) (52 + 2k)X I - + bl ,

kX2 = sal

-kXI + (S2 + 2k)X2 = sa2 + h

The determinant of the system is

(24)

and is never zero for real s. This implies that Xl (s) and X2 (s) are uniquely determined
by (23). In fact, a simple calculation gives

(25)
X (s) = (S2 + 2k)(sal + bl ) + k(sa2 + h)
I (52 + k)(S2 + 3k) ,

and

(26)
X (s) = (S2 + 2k)(sa2 + b2) + k(sal + bl ) .
2 (S2 + k)(s2 + 3k)

It is clear from the partial fractions expansion form of these functions that Xl (0 and
X2(t) are linear combinations of the four special solutions defined by (19) and (20). In
particular, al = a2 = 0, b[ = h = J'k gives the first solution pair of (19), while
al = a2 = 0, bl = -b 2 = J3k gives the second; al = a2 = 1, bl = b2 = 0 gives the
first of (20), and al = -a2 = 1, bl = b2 = 0 gives the second. On multiplying the four
solutions by constants AI, A 2, A 3, A4 and adding, one gets the general solution

(27) Xl (0 = Al sin J'kt + A2 sin 5kt + A3 cos J'kt + A4 cos 5kt,

X2(t) = Al sin J'kt - A2 sin 5kt + A3 cos J'kt - A4 cos 5kt,
268 4 THE LAPLACE TRANSFORM

with derivatives

(28) X~ (t) = JkA l cos ../ht + 5kA2 cos J3kt

- JkA3 sin ../ht - 5kA4 sin J3kt,
x~{t) = JkA l cos../ht - J3kA2 cos J3kt
- JkA3 sin ../ht + 5kA4 sin J3kt.

To calculate AI, A 2, A3, ~ from the initial values (21), set t = 0 in these equations to
get the equations

(29) XI(O) = A3 +A4 = aI,

X2(0) = A3 - ~ = a2,

and

(30) X~(O) = JkA l + 5kA2 = bl ,

x~(O) = JkA l - 5kA 2 = b2 ,
with unique solutions

(31) Al = (b l + b2 )/2../h,
A2 = (b l - b2)/2J3k,
A3 = (al + a2)12,
A4 = (al - a2)/2.

Equations (27) have been shown to define the general solution of the oscillator problem
of Figure 9.
Note that if A2 = A4 = 0, then the solution is a synchronous oscillation of frequency
.fii;(21r), while if Al = A3 = 0, it has frequency J3kt(21r). These two special solutions
define the characteristic modes of the oscillator. It will be shown in Chapter 5 that
systems of n coupled oscillators have n characteristic modes.
A Two-Body Damped Mechanical Oscillator. Consider two simple mass-spring
oscillators, coupled together by a damper as suggested in Figure 10.

FIGURE 10. Atwo-body damped

mechanical oscillator
 4.5 ApPLICATIONS TO SYSTEMS OF DIFFERENTIAL EQUATIONS 269

The springs are assumed to obey Hooke's law, while the damper opposes relative
motion of the two masses by a force proportional to their relative speed. Newton's laws
give the equations

(32) mx~ + kxl - C(X; - X~) = 0,

mx~ + kx2 + C(X; - X~) = O.

For simplicity, the numerical values m = 1, k = 1, c = 1 will be used in the analysis.

Then (32) can be written

(33) x~ +x~ +XI -x; = 0,

x~ +x; +X2 -x~ = O.

The solutions of this system will be determined by the initial positions and speeds of the
two bodies. If the notation (21) of the preceding problem is used, then the 'c-transforms
of equations (33) can be written

(34) (S2 + s + 1)XI - sX2 = sal + bl + al - a2,

-SXI + (S2 + S + 1)X2 = sa2 + b2 + a2 - al.

The determinant of system (34) is

(35)

Hence the partial fraction expansion of Xl (s), obtained from (34), (35) is
AIS+A2 A3 A4
(36) X I ()
S = +- - + -:----:_::_
S2 + 1 S + 1 (s + 1)2 .
This gives

(37)

with a similar expression for X2(t). The dependence of AI, A 2, A 3, A4 on the initial data
al, a2, bl , b2 can be calculated from (34).
There is an easier way to find the general solution of (33). To discover it, note that
by adding the equations of (33) one finds that

(38)

Similarly, on subtracting the equations, one finds that

(39)
270 4 THE LAPLACE TRANSFORM

It follows readily by the method of Section 3.3 that

(40) YI (t) = CI cos t + C2 sin t,

Y2(t) = C3e-t + C4 te- t ,

and hence

(41) Xl (t) = CI cos t + C2 sin t + C3e-t + C4te-t,

X2(t) = CI cos t + C2 sin t - C3e-t - C4te-t.

It is not difficult to calculate the constants CI, C2, C3, C4 from the initial data ai, a2, bl ,
h Thus (41) defines the general solution of the system (33).
It is not difficult to see from the structure of the device of Figure 10 that it may have
an undamped oscillatory mode of vibration. Indeed, if the distance X2 - Xl = constant
then the damper is not activated, and undamped oscillations occur (C3 = C4 = 0 in
(41)). In general, the motion consists of a damped part, or transient, and an oscillatory
part, or steady-state oscillation. For large t the two bodies oscillate in phase with constant
distance between them.

An Electric Circuit Analog. The two-loop electric circuit of Figure 11 will be shown
to be isomorphic to the damped mechanical oscillator just discussed.
For if the electric currents ii, i 2, i3 are as in the figure and ql, q2 are the charges on
the capacitors, then Kirchhoff's current and voltage laws give

(42)

FIGURE 11. Adamped electric circuit oscillator

 4.5 ApPLICATIONS TO SYSTEMS OF DIFFERENTIAL EQUATIONS 271

If the first three equations are used to eliminate il, i 2 , i 3 , the resulting system becomes

(43) Lq~ + ~ql - R(q; - q~) = 0,

Lq~ + ~q2 + R(q; - q~) = O.

This system is clearly isomorphic to (32) under the correspondence L = m, ql = Xl,

q2 = X2, l/C = k, R = c. In particular, if L = 1, C = 1, R = 1 then the solution of (43)is
given by (41) with ql (t) for Xl (t) and q2 (t) for X2 (t). In particular, the steady-state solution
of the circuit, which all solutions approach as t -+ 00, is characterized by i3 (t) = 0 and

(44)

A Vibration Absorber. Consider the simple damped oscillator of Figure 12.

The governing differential equation is

(45) mx" + ex' + kx = f(t),

where f(t) is the applied force. The analogous LRC circuit was analyzed in Section 3.3
above, equations (61)-(66). It was found that if

(46) fet) = fo sin wt, w > 0,

then the oscillator has an undamped steady-state solution

(47) Xss (t) = A sin wt + B cos wt,

where
k-mw 2 -we
(48) A=--- B=--
D(w) , D(w)'

Moreover, the general solution of (45) differs from (47) by an exponentially damped term,
so that all solutions approach xss(t) for large t. In some applications, such as synchronous
electrical motors, xss(t) may represent an undesirable vibration. For such cases it was
discovered by Frahm in 1909 (see [DenH]) that xss(t) can be completely eliminated by

J(t)

~ x FIGURE 12. Adamped mechanical oscillator

272 4 THE LAPLACE TRANSFORM

FIGURE 13. A vibration absorber

coupling m to a small undamped oscillator. To see how this can be done, consider the
coupled oscillators of Figure 13.
Newton's laws yield the system of equations

(49) rnx" + ex' + kx - hi (XI - x) = fo sinwt,

mlx~ + hI (XI - x) = o.
For simplicity, consider the solution of this system with initial values x(O) = Xl (0) =
X' (0)= xi (0) = O. The £-transform of this solution satisfies the linear equations

(50) (ms 2 + es + h + hl)X - hlXI = fow/(s2 + ( 2),

-klX + (mls 2 + kl)X I = o.
Solving for X gives

(51)

where

(52)

It can be shown that all the roots of D(s) = 0 have negative real parts; see [eh-B] for a
proof. The partial fractions expansion of (51) can be written

(53) () As + Bw
Xs= +--
pes)
52 + w 2 D(s) ,
where pes) is a polynomial of degree 3 or less. The corresponding equation for x(t) is

(54) x(t) = Acoswt + Bsinwt +xtr(t),

where
pes)
(55) £{xtr(t)} = D(s) ,

and Xtr(t) tends to zero exponentially as t -+ 00. Thus, in general, the compound
oscillator of Figure 13 has the same type of undamped steady-state oscillation (47) as the
 4.5 ApPLICATIONS TO SYSTEMS OF DIFFERENTIAL EQUATIONS 273

simple oscillator. The only exception to this occurs when the numerator in (51) cancels
the factor S2 + ui in the denominator. This is the idea behind the vibration absorber.
Given the frequency w of the applied force, one chooses hi and ml such that
hi 2
(56) -=w.
ml
Then, from (51),

(57) X( ) = fOmlw
S D(s) ,

and it follows that x(t) = Xtr(t) is exponentially damped. Of course, one has from (50),
fowh l
(58) Xl (s) = (52 + ( 2 )D(s) ,
and hence ml will have a steady-state vibration Xl (0 of the form (47). However, this may
be more acceptable in an application than the vibration of the primary mass m. Vibration
absorbers of this and other types are discussed starting on page 87 in [DenHl.

Uncoupled Systems. Use the .c-transform method to calculate x(t) and y(t).
1. (LR Circuit) x' = 3x, y' = 4y + 5, x(O) = 1, yeO) = 2.
2. (LR Circuit) x' = -x + e- t , y' = 2y. x(O) = 0, yeO) = -2.
3. (Mass-Spring System) x" + 4x = 1, y" + y = 0, x(O) = 0, x'(O) = 0, yeO) = 0, y'(0) = 1.
4. (Mass-Spring System) x" + 2X' + 5x = 1, y" + y = I, x(O) = 0, x'(O) = 0, yeo) = 0, y'(0) = 1.
First-Order Coupled Systems. Use the .c-transform method to calculate Xes) = .c{x(t») and yes) =
.c(y(t»). Then use tables or a computer algebra system to find x(t) andy(t).

5. x' = 3x +y,y' = 4x+3y,x(0) = I,Y(O) = 0

6. x' = 4x - y, y' = x + 2y, x(O) = 0, yeO) = 1.
7. x' = 2x - y,y' = 9x+2y,x(0) = l,y(O) = 1.
8. 2X' + y' - X - Y = e- t , x' + y' + 2x + y = et , x(O) = 0, yeO) = O.
9. x' +y = t, y' + 4x = 0, x(O) = I, yeo) = -1.
Second-Order Coupled Systems. Use the .c-transform method to calculate Xes) = .c{x(t») and yes) =
.c{y(t»). Then use tables or a computer algebra system to findx(t) andy(t).

10. x" + y' - x + y = 1, y" + x' - x + y = 0, x(O) = 0, x' (0) = 0, yeO) = 0, y' (0) = o.
11. x" + y' + 3x = 1, y" - 4x' + 3y = 0, x(O) = 0, x'(O) = 0, yeo) = 0, y'(O) = O.
Applications to Electric Circuits. Below are four two-loop electrical circuits. In each case use Kirchhoff's
laws to write a system of differential equations for the unknowns. Then use the .c-transform method to find
formulas for the unknowns.
274 4 THE LAPLACE TRANSFORM

FIGURE 14. Atwo-loop RL network

FIGURE 15. Atwo· loop RL network

FIGURE 16. Atwo-loop LRC network

FIGURE 17. Mutual inductance

12. The parameters in the circuit of Figure 14 have the values v = 30 volts, R1 = 10 ohms, R2 = 20 ohms,
L1 = 0.02 henries and L2 = 0.04 henries. Calculate i1 and i2 if i1 (0) = 0 and i2(0) = o.

13. The parameters in the circuit of Figure 15 have the values v = 100 volts, R1 = 20 ohms, R2 = 40
ohms, L1 = 0.01 henries and L2 = 0.02 henries. The currents i1 and i2 are as indicated. Calculate i1
and i2 if i1 (0) = 0 and i2 (0) = o.
14. The parameters in the circuit of Figure 16 have the values v = 500sin(10t) volts, C = 0.01 farads,
L = 1 henries and R1 = R2 = 10 ohms. Calculate the charge q on the capacitor if q(O) =
0 and
i(O) = q'(O) = o.

15. The two loops of the circuit in Figure 17 are coupled by a mutual inductance M. Kirchhoff's laws give
the pair of equations

L1i~ +Mi; +R1i1 =v,

Mi~ + L2i; + R2i2 = O.
 4.5 ApPLICATIONS TO SYSTEMS OF DIFFERENTIAL EQUATIONS 275

Calculate the currentsij andi2 ifij(O) = O,i2(O) = oand the parameters have the values Lj = L2 = 1
henry, Rj = 1 ohm, R2 = 0.25 ohm, M = 0.5 henry and v = 120 volts.
C HAP T E R 5
 Linear Algebra

Systems of simultaneous linear equations occur frequently in all branches of science and
engineering. Linear algebra is the theory of such systems and the methods available
for their solution. The goal of this chapter is to present the elementary concepts and
algorithms that are needed for applications.

5.1 Systems of Linear Equations

This introduction begins with three examples that illustrate how systems of linear
equations arise in science and engineering.

EXAMPLE 1. The Wheatstone Bridge. The resistive network of Figure 1 is known

historically as a Wheatstone bridge. It was invented as a device to measure an unknown
resistance by comparing it to three known resistances. It can be shown that the current
through resistor R5 is zero if and only if RI~ = R2 R3 . Thus if R3 and ~ are known fixed
resistances and R2 is a known variable resistance then one can adjust R2 until the current
through R5 is zero. The resistance RI is then calculated as RI = R2R3/~.
A more general problem is to compute the six branch currents iI, i 2 , i 3 , i 4 , i 5 , i6
when the resistances RI , R2 , R3, R4 , R5 , R6 and the voltage V are known. The simplest
way to do this is to compute the three mesh currents Xl, X2, X3 indicated in the figure.
278 5 LINEAR ALGEBRA

+
v
FIGURE 1. Wheatstone bridge

Ohms law and Kirchhoffs voltage law imply that Xl, X2, X3 satisfy the linear equations

(RI + R3 + RS)XI - RSX2 - R3X3 = 0,

(I) -RSXI+(Rl+~+Rs)xl -~X3=0,
-R3XI - ~X2 + (R3 + ~ + R 6 )X3 = V.
When this system has been solved for Xl, Xl, X3, the branch currents are then given by
il =Xl, iz =
Xl, i3 =
X3 - Xl, i4 = Xl - X3, is = Xl - Xl, and i6 = X3. As a numerical
example let RI =
2, Rl =
4, R3 = 1, ~ = 2, Rs = 10, ~ = 10 (ohms), and V = 15
(volts). Then system (1) becomes

13XI - lOxl - X3 0, =
(2) -lOxl + 16xl - 2X3 = 0,
-Xl - 2X2 + 13x3 = 15.
It will be shown below that this system has the unique solution Xl = 5/12, Xl = 5/12,
X3 = 5/4. Hence, the Wheatstone bridge with the given parameters has branch currents
il = Xl = 5/12, il = Xl = 5/12, i3 = X3 - Xl = 5/6, i4 = Xl - X3 = -5/6,
is = Xl - X2 = 0 and i6 = X3 = 514 amperes.

ExAMPLE 2. Mechanical Equilibrium. Consider the mass-spring system depicted in

Figure 2. Five masses ml, ... , ms are suspended between springs as shown. The six
springs are assumed to obey Hooke's law with the same Hookes constant k. The five
displacements Xl, ... , Xs are measured downward from their rest points under no load.
The problem is to compute the static displacements, under gravity, when the masses and
spring constant are given.
The conditions required for equilibrium are that the total force on each mass be
zero. For mass ml, Hooke's law gives spring forces -kxl up and k(X2 - Xl) down, while
the gravitational force is mIg (g being the gravitational constant). Thus the equilibrium
condition for ml is -kxl + k(X2 - Xl) + mIg = 0, or lxl - X2 = (glk)ml' Similarly, for
 5.1 SYSTEMS OF LINEAR EQUATIONS 279

FIGURE 2. Spring-mass structure

mass m2 the condition is -Xl + 2X2 - X3 = (glk)m2, etc. Proceeding in this way gives a
system of five linear equations for Xl, ... , xs, which may be written

lxl - X2 =gm/k,
-Xl + 2x2 - X3 =gmik,
(3) X2 + 2x3 - X4 = gm3lk ,
X3 + 2x4 - Xs = gm4 /k ,
- X4 + 2xs = gmslh.
It will be shown below that for any values of glk and masses m l, ... , ms this system
has a unique solution. For example, if glk = 1 and ml = 1, m2 = 2, m3 = 3, m4 = 4,
ms = 5 then Xl = 3516, X2 = 32/3, X3 = 27/2, X4 = 40/3 and Xs = 55/6.

EXAMPLE 3. Balancing a Chemical Reaction. l The problem is to balance the rather

complicated chemical reaction

(4) Xl CU25 + X2H+ + X3N03 --+

X4CU2+ + xsNO + X6SS + X7H20
lThis problem appears in FirstLeaves: A Tutorial Introduction, B.W Char et aI., Springer 1992, page 197 (ISBN
0-387-97621-3)
280 5 LINEAR ALGEBRA

by calculating the positive integer coefficients Xl, ... , X7. The principles are that the
number of atoms of each element must be conserved as well as the total electric charge
Q. This gives six equations for the seven coefficients:
eu 2x1 = Xi,
S Xl = 8X6,
H X2 = 2x7 ,
N X3 = Xs,
0 3X3 = Xs +X7,
Q X2 -X3 =~.

This system of equations will be shown below to have a one-parameter family of

positive integer solutions:

Xl = 24n,
X2 = 128n,
X3 = 32n,
Xi = 48n,

Xs = 32n,
X6 = 3n,
X7 = 64n,

where n is a positive integer. The desired least positive solution is given by n = 1.

Linear m x n Systems. Examples 1, 2 and 3 illustrate the central problem of linear
algebra, which is to solve a system of m linear equations in n unknowns, where m and
n are positive integers. Such systems have the general form

(5) allXI + aUX2 + ... + alnXn = bl ,

a2lXI + a22X2 + ... + a2nXn = b2,

where all, au, ... , amn and bl , b2 , ... , bm are prescribed numbers and the equations are
to be solved for the unknowns Xl, ... , X n . Examples 1,2, and 3 have m x n = 3 x 3,
5 x 5, and 6 x 7, respectively.
Section 5.2 below develops a method for computing all the solutions (if any) of any
linear system. For small values of m and n (say m ::::: 3, n ::::: 3) those solutions can
 5.1 SYSTEMS OF LINEAR EQUATIONS 281

be found by hand computation. However, real applications often lead to much larger
values of m and n. For example, practical electric circuits may have 100 or more meshes,
leading to linear systems with m ::: 100, n ::: 100. Problems of nuclear reactor design
lead to systems with several thousand unknowns. Clearly, such problems require efficient
computer algorithms for their solution.
For large systems (m and/or n large) it may not be helpful, or even possible, to write
the equations down on paper. Instead, the numbers all, au, ... , a mn and bl , b2 , ... , bm
that define the system may be generated in the computer as needed. For the theoretical
study of such systems it is essential to have a concise notation that does not require each
equation to be written explicitly The notation of linear algebra was invented for this
purpose.
Vector Notation. A solution of the linear system (5) is an ordered set of real numbers
(Xl, ... ,xn )· The term ordered means that each unknown Xk has its assigned place in the
list. For theoretical calculations each solution set will be denoted by a single letter:

x == (Xl,'" ,Xn).

The set of all such n-tuples x is denoted by Rn. Similarly, b == (b l , ... , bm ) denotes
the right-hand side of system (5). Members of Rn are called vectors or, more explicitly
n-vectors.
Matrix Notation. The set of mn coefficients aij in system (5) will be denoted by the
single letter A. Thus

en a,. )
au
a2l all a2n
(6) A == .

amI a m2 a mn

Such an array of mn values will be called a matrix or, more explicitly, an m x n matrix.
The m x n system (5) can now be written as a single equation, Ax == b, where Ax denotes
the m-vector

allXI + a12X2 + ... + alnXn )

(
a2l X I + anX2 + ... + a2nXn
(7) Ax == . .

amlXI + a m2X2 + ... + amnXn

°
A special n-vector is the zero n-vector == CO, 0, ... ,0) Call terms zeros). The special
system Ax == 0 is called a homogeneous linear system. If b =j:. 0 then Ax == b is called
a non-homogeneous linear system.
282 5 LINEAR ALGEBRA

Vector Algebra. Let n-vectors y = (YI, ... , Yn) and z = (ZI, ... , Zn) be two solutions
of the homogeneous system Ax = 0, so that Ay = 0, Az = 0, or, explicitly,

(8) allYl + a12Y2 + ... + alnYn = 0,

allYl + allY2 + ... + a2nYn = 0,

and

(9) allZI + a12Z2 + ... + alnZn = 0,

a21Z1 + anZ2 + ... + a2nZn = 0,

On adding corresponding equations of (8) and (9) one finds that the n-vector w
(WI, W2, ... , wn ) defined by

(10) W= (YI+ZI,Y2+Z2, ... ,Yn+Zn)

satisfiesAw = O. Thus two solutions of the homogeneous system, y andz, have generated
a third solution w. It is natural, and convenient, to write w = y + z and call w the vector
sum of the n-vectors y and z.
Similarly, if a is any real number then on multiplying each equation of (8) by a one
finds that u = (OIYI, OIY2, . .. ,00Yn) satisfies Au = O. This suggests that we write u = OIy
and call u the scalar multiple of the n-vector y by the number 01. The definition

(ll) u = (OIYI,OIY2,'" ,00Yn)

can be expressed by saying that u = OIy if and only if Uk = OIYk for k = 1,2, ... , n.
The notation of the last two paragraphs implies that the matrix A obeys the rules

(12) A(x+y) =Ax+Ay,

A(OIx) = OIAx,

for all vectors x, y and all numbers 01. It follows easily that if {J is another number then

(l3) A(OIx + {3y) = OIAx + {JAy

 5.1 SYSTEMS OF LINEAR EQUATIONS 283

for vectors x, y and all numbers ex, f3. In Section 5.5 these properties will be used to find
a simple description of the solution sets of the homogeneous systems Ax = o.
Complex Numbers. The examples given above dealt with matrices and vectors with
components in the real number field R. However, exactly the same rules hold when the
components are in the field C of complex numbers. Both cases will be treated in the
remainder of this chapter.

Conversion to Matrix Form. Identify matrices A, x and b that convert the systems below to matrix form
Ax = b. Verify the definitions of A, x and b using matrix multiplication and equality of vectors.

1. 2x - Y = 3, 2. x+y-2z= 1,
x+4y = 5. 4x +z = 2,
-x+y = 5.

A Resistive Network. Consider the resistive network of Figure 3. Apply Kirchhoff's voltage law and
Kirchhoff's current law to complete the following exercises.

3. Derive a system of five linear equations for the mesh currents Xl, ... , x5 in Figure 3.

4. Represent the five mesh equations of Figure 3 in the form of a matrix system Ax = b. Fully identify the
matrix A, the vector x and the vector b.

5. Solve the system Ax = b of the previous problem for the mesh currents x, verifying the formula

1
x = 6535 (9702, 3033, 612, 54, 18).

Transition Probabilities. A vector p is called a probability vector if each entry is nonnegative and the sum
of the entries is 1. A matrix A is called a transition probability matrix if each column of A is a probability vector.

6. (Transition Probabilities) Let A be a square transition probability matrix. Show that for each probability
vector p, the vector Ap is also a probability vector. Conclude that iterates Anp of probability vectors are
again probability vectors.

FIGURE 3. Aresistive network

284 5 LINEAR ALGEBRA

7. (Weather) Let

2'I 4
3 4I)
A -- (1 g 1
4 I
4 .
I I I
4 g 2'

The entries are interpreted as weather transition probabilities for rain, sun or clouds. Precisely, if p is
the 3-vector of today's probabilities for rain, sun, and clouds, then Ap is tomorrow's probabilities. Given
that today's probabilities are 0, ! and!, predict the probability of rain 1, 3 and 5 days ahead.

5.2 The Gauss Elimination Method

Let Ax = band A'x = b ' be two systems of linear equations for the same unknowns
x = (Xl, X2, ... , x n ). The systems will be said to be equivalent if their solution sets (the
set of n-vectors x that satisfy the systems) are exactly the same. For equivalent systems,
the solutions of the first system can be found by solving the second. This is useful if the
second system has fewer equations or is otherwise simpler than the first.
The systematic use of equivalent systems to solve systems oflinear equations is usually
attributed to Gauss. The Gaussian method is to apply a sequence of algebraic operations
to the system Ax = b to produce an equivalent system A'x = b ' that is easy to solve. The
method will be illustrated by three examples before it is formulated as a general algorithm.

EXAMPLE 1. Consider the 3 x 3 system

(1) + 2x2 + 3X3

Xl = 1,
4XI + 5x2 + 6X3 = 7,
4XI - 3X2 +X3 = 5.

This has the form Ax = b with

(2)
-3
2 3
5 6
1
). b=(1,7,5).
To solve (1) it will suffice to omit writing the unknowns Xl, X2, X3 and write only the
augmented matrix

(3)
 5.2 THE GAUSS ELIMINATION METHOD 285

Here the left three columns define A and the right column defines h. Thus each row of
the augmented matrix defines the corresponding equation ofthe system (1).

Systems equivalent to 0) will be obtained by applying the following operations to

the rows of (3):

(4) • Multiply a row by a nonzero number.

• Add a row to another row.
• Interchange two rows.

These operations are called elementary row operations. Each of them is reversible.
Hence each produces an equivalent system. This will be illustrated by transforming (3)
to a solvable equivalent form. The goal of the operations is to introduce into A as many
zeros as possible. The double arrow notation (A I h) {::::::} (A'l hi) will be used to
denote the equivalence of systems Ax = hand A'x = hi.
Multiply row 1 of (3) by -4 and add it to row 2 and row 3. This combination of
elementary row operations produces the equivalent system

2 3
(5) -3 -6
-11 -11

Next multiply row 2 by -11/3 and add it to row 3. The result is the augmented matrix

~(
1 2 3
(6) (A' Ib') 0 -3 -6 3I ) ,
0 0 11 -10

which represents the linear system

(7) Xl + 2x2 + 3X3 = 1,

-3X2 - 6X3 = 3,
llx3 = -10.

This system has a unique solution, which can be obtained by solving the equations in
reverse order. Thus X3 = -lOllI, X2 = -1 - 2x3 = -1 + (20/11) = 9/11 and
Xl = -2X2 - 3X3 + 1 = -18/11 + 30111 + 11111 = 23/11. This shows that system
(7), and hence also system (1), has the unique solution x = (23/11,9/11, -10/11). The
solution may be verified by direct substitution into (1).
286 5 LINEAR ALGEBRA

EXAMPLE 2. Consider next the system Ax = b with augmented matrix

(8) (A Ib) = (
1
4
2
5
31)
6 7 .
6 9 12 9

It looks similar to Example 1 but has quite a different solution set. To solve it, add -4
times row 1 to row 2 and -6 times row 1 to row 3. This gives the equivalent system

(9)
( o~ -~ -! ~).
-3 -6 3

Next add -1 times row 2 to row 3 and also multiply row 2 by -1/3. The result is the

0
eqUivalent system

(10) (A'lb') =
2 3
1
o
2
0
-l)
This represents the system

(11) Xl + 2X2 + 3X3 = 1,

X2 + 2x3 = - 1,
0=0.

The third equation, OXI + OX2 + OX3 = 0, is automatically satisfied and may be dropped.
System (11) has infinitely many solutions, given by X2 = -1- 2x3 ,XI = 1- 2X2 -3X3 =
3 + X3. Thus X3 = t may be any real number and then X2 = -1 - 2t, Xl = 3 + t. In
vector notation the solutions can be written

(12)

This also defines the solution set of (8), since it is equivalent to (11).

The solution makes sense geometrically, because the equations

represent planes in space, which, according to the solution above, intersect along a line
in space, as in Figure 4.
 5.2 THE GAUSS ELIMINATION METHOD 287

intersection FIGURE 4. Intersecting planes

ExAMPLE 3. Finally, consider the system Ax = b with augmented matrix

(l3) (A Ib) =(
1 2
4 5
31)
6 7 .
6 9 12 3

It differs from Example 2 only in the third component of b. Applying exactly the same
row operations as in Example 2 gives the equivalent system

(14) (A'l b') = ( 123 1)

0 1 2 -1 .
000 2

The last row represents the equation OXI + OX2 + OX3 = 2, which is always false. Hence
(14) has no solution-its solution set is empty. The same is therefore true of system (l3).

The Gauss Algorithm. The Gauss method of simplification by elementary row

operations was illustrated in Examples 1,2,3. Now it will be described in the context of
a general m x n system. It will be assumed that each of the n unknowns Xl, X2, ... , Xn
appears with a nonzero constant in at least one of the m equations. Thus each column of
the matrix A contains at least one nonzero number. If this is not the case then the system
can be rewritten with a smaller number of unknowns.
The Gauss method proceeds by columns, working from left to right.
Column 1. At least one coefficient in column 1 is not zero. Bring it to the top by
interchange with row 1. Then divide the new row 1 by its first coefficient to produce a
1 in the (1, 1)-position. Finally, subtract multiples of row 1 from the remaining rows to
produce zeros in positions (2, 1), (3,1), ... , (m, 1).
The preceding row operations transform the system (A I b) into an equivalent system
(A'l hi) in which only the first equation contains the unknownxi. It may happen that the
second through the mth rows of A' contain only zeros. In this case the Gauss procedure
is complete. Otherwise one proceeds to column 2.
288 5 LINEAR ALGEBRA

Column 2. Apply the procedure described for column 1 to the (m -1) x (n -1) system
obtained from (A'l b') by deleting the first row and first column. This produces an
equivalent system (A" Ib") in which there is a 1 in position (2,2) and zeros in positions
(3,2), (4,2), ... , (m, 2). There is only one exception. If a;2 i= 0, then it may happen
that a~2 = a;2 = ... = a~2 = O. In this case one leaves column 2 as it is and proceeds
to the first column (if any) that has nonzero coefficients on or below row 2.
The Gauss method continues as above through the remaining columns, each time
deleting an additional row and one or more columns. It terminates when all columns of
A have been processed. The final result is the replacement of the original system (A I b)
by an equivalent system (B I c) with the following properties.

(15) • All rows (if any) that consist entirely of zeros are at the
bottom of the matrix.
• In any nonzero row, the first nonzero number (from the
left) is 1. It will be called a leading 1 .
• In any two consecutive nonzero rows the leading 1 in the
lower row is to the right of the leading 1 in the upper row.

Row Echelon Matrices. A matrix that has properties (15) is said to be a row echelon
matrix. Thus matrix (10) of Example 2 is a row echelon matrix. Matrix (6) of Example 1
is not. However, division of rows 2 and 3 by - 3 and 11, respectively, give an equivalent
row echelon matrix:
1 2 3
(16) ( 012 1
-1 )
001 -10/11

The procedure described above for reducing a linear system (A I b) to row echelon form
is called Gauss Elimination. The row echelon form readily yields all solutions (if any)
of the original system. Examples 1, 2 and 3 illustrate the three possibilities. An m x n
system (A I b) is inconsistent and has no solutions if (and only in its row echelon form
has a row consisting of n zeros followed by a nonzero constant, as illustrated in Example
3. System Ax = b has a unique solution if (and only in its row echelon form has n
leading 1's along the main diagonal (possible only if m ::::: n). If m > n then rows n + 1
through m must consist entirely of zeros. Finally, Ax = b has infinitely many solutions
if (and only in its row echelon form has fewer than n nonzero rows.
Solution by Back Substitution. All solutions of an m x n system Ax = b can be
calculated from an equivalent row echelon system Bx = c. If Ax = b has a unique
solution then equation n of the system Bx = cis Xn = en. Next, equation (n - 1) reads
 5.2 THE GAUSS ELIMINATION METHOD 289

Xn-l + bn-l,nXn = (n-l· Substituting Xn = (n gives Xn-l = (n-l - bn-l,n(n. Continuing

in this way gives, in succession, Xn-2, X n-3, ... , Xl. This procedure is called back
substitution. It was illustrated by Example 1 with equivalent row echelon form (16).
If Ax = b has infinitely many solutions then an equivalent row echelon system
Bx = c will have a number r of nonzero rows with r < n. In this case all solutions of
Ax = b may be calculated by the following procedure. First, divide the unknowns into
two types, the "dependent unknowns" and the "independent unknowns." An unknown
Xk is a dependent unknown if and only if column k of B contains a leading 1. The
remaining unknowns are the independent unknowns. All solutions of Ax = b are found
by (0 assigning arbitrary values to each of the independent unknowns and (i0 solving
the r nonzero equations of Bx = c for the dependent unknowns by back substitution.
There will be exactly r dependent unknowns and hence the solutions of Ax = b will be
linear functions of the n - r independent unknowns. This was illustrated by Example
2 above with row echelon form (0). In that case m = n = 3 and r = 2. Hence the
solutions depend on n - r = 3 - 2 = 1 parameter.

Reduced Row Echelon Matrices. Elementary row operations can be used to

simplify further the row echelon systems. This is done by using each leading 1 to eliminate
the nonzero entries that lie above it in its column. The resulting matrix is a row echelon
matrix with the additional property that each column that contains a leading 1 has zeros
in the rest of the column. Such matrices are said to have reduced row echelon form
(rref). They are important for two reasons. First, each augmented matrix (A I b) is
equivalent to a unique rref matrix. A proof of this theorem may be found in [No J. The
result is illustrated by Example 2, where

(~ ~ ~
-1
(17) (Alb) {=:::} 2
000 o

Recall that" {=:::} " denotes equivalence. Equation (17) exhibits two distinct row echelon
matrices that are equivalent to (A I b), but only the latter is in reduced row echelon form.
The second advantage of the reduced row echelon form is that it is immediately
solvable without back substitution. Each nonzero row contains one dependent unknown
followed by zero or more independent unknowns. For example, from the last matrix in
(7) it is seen that X3 = t is an independent unknown, while Xl and X2 are dependent.
Moreover, by inspection one gets Xl = t + 3, X2 = -1 - 2t, X3 = t, as in (2).
290 5 LINEAR ALGEBRA

The extended version of the Gaussian algorithm described above is usually called
the Gauss-Jordan algorithm. 2 It associates to each augmented matrix (A I b) a unique
rref matrix (B I c), equivalent to (A I b), from which the solutions of Ax = b (if any) can
be written down by inspection. For systems of three or four equations the rref matrix
(B I c) can be calculated by hand. For larger numbers of equations and/or unknowns the
Gauss-Jordan algorithm becomes algebra-intensive, requiring large numbers of algebraic
operations. Fortunately, the algorithm can be automated by using one of the computer
algebra systems (CAS for short) that have been developed since the early 1970s. The
use of computers greatly increases the size of the linear systems that can be solved in
practice. This section concludes with three systems that are difficult to solve by hand but
are readily solved by use of a CAS. Many additional examples are given in subsequent
examples and exercises.

EXAMPLE 4. Consider the 4 x 6 system Ax = b with

(18) A= CBS 97
49
-255
291
147
85
-97
-49
-55
50
63
-705
779
469
-37 )
79
57
45 135 -45 -8 299 -93

and b = (-435,642,523,-444). It would be tedious to apply the Gauss-Jordan

algorithm to (A I b) by hand. However, a computer algebra system gives the equivalent
augmented matrix (B I c) in seconds. The result is

3 -1 0 7 0

U D
0 0 1 2 0
(19)
0 0 0 0 1
0 0 0 0 0

By inspection, X2, X3 and Xs are independent unknowns and

(20) Xl = 1 - 3X2 + X3 - 7xs,

X4 = 3 - 2xs ,
X6 = 5.

All solutions of the system Ax = b are obtained by assigning arbitrary values to X2, X3
and Xs and then computing Xl, X4 and X6 from (20). These solutions can be written in

2After Wilhelm Jordan (1842-1899), German geodesist and mathematician who popularized the reduced row
echelon method through his Handbook of Geodesy.
 5.2 THE GAUSS ELIMINATION METHOD 291

n-vector form (n = 6) as

(21) x = (1,0,0,3,0,5)

+S(-3,1,0,0,0,0)

+ t( 1,0,1,0,0, 0)
+u(-7,0,0,-2,1,0),
where S = X2, t = X3 and u = X5 are arbitrary parameters.
EXAMPLE 5. Consider the 4 x 6 system Ax = b where A is the matrix of Example 4 with
a new vector

(22) b = (-470,698,464,-352).

In this case machine computation gives

(23) (B I c) =
(
1 3
-1 ° 7 ° 1)
° ° °° ° ° ° 1 2 3
°° 1 5 .
°° ° °°° 1
The last row shows that the equivalent systems (B I c) and (A I b) are inconsistent and
have no solutions. Thus a simple change of the vector b changes the system Ax = b from
one with infinitely many solutions to one with no solutions. From the original forms of
the two systems it is impossible to recognize this. A simple algebraiC test to determine
the nature of solutions of Ax = b is given in Section 5.5 below.

EXAMPLE 6. Consider the 5 x 4 system Ax = b where

-85 -55 -37 -35
50 79 56 49
(24) A= 57 -59 45 -8
92 43 -62 77
54 -5 99 -61

and b = (-470,616,97, -12,473). The system Ax = b appears to be overdetermined

with 5 equations for 4 unknowns. However, the Gauss-Jordan algorithm and computer
292 5 LINEAR ALGEBRA

algebra give the equivalent rref matrix

1 0 0 0 1
0 1 0 0 3
(25) (B I c) = 0 0 1 0 5
0 0 0 1 1
0 0 0 0 0

and the system is seen to have the unique solution x = (1,3,5,1).

Complex Numbers. The examples given above all have components in the real
number field. However, the Gauss elimination method works equally well with complex
coefficients. An example is given in the Exercises.

Exercises 5.2

Equivalence. Use elementary row operations to establish the following equivalences.

l.(~! ~)~(~ ~ 3~2)~(~ ~ 3~2)'

-1 )
1 .
Tridiagonal Systems. Find the solution x of the system Ax = b where b is the first column of the identity
matrix.

3. A = ( _~ -12 ) .

4. A= ( -~
-1
2
-r) .

n
-1
Echelon Form. Find the reduced row echelon form of A.
-1 0 0
( -12 2 -1 0
5. A= ~ -1 2 -1

n
0 -1 2

-1 0 0 0
-1 2 -1 0 0
( 2
6. A= ~ -1 2 -1 0
0 -1 2 -1
0 0 -1 2
Gauss-Jordan Algorithm. In the problems below, solve the system Ax = b.

7. A= (~ ~2 1
1 1
~ ~)'b=(l).
-1 -1 0
 5.2 THE GAUSS ELIMINATION METHOD 293

8. A = (! 2 ~). b =( -0.
9. A= ( ~ ~ ). b = ( - D.
Geometric Classification. Classify the following systems of equations geometrically into the classes below.
Class A. The planes intersect in a unique point.
Class B. The planes intersect in a line.
Class C. The planes meet pairwise in skew lines that never intersect.
Class D. All the planes are the same.
Class E. The planes are parallel and there are no intersections.
5x+ 5y+ Z =0,
10. lOx + 25y + 2z = 0,
5x + 15y + 3z = 1.
5x+ 5y+ Z= 0,
11.lOx+lOy+2z=0,
5x + 15y + 3z = 1.
Y =0,
12. z= 0,
y+z=1.
5x+ 5y+ z=O,
13. lax + lOy + 2z = 0,
15x + 15y + 3z = O.
5x+ 5y+ Z= 0,
14.lOx+l0y+2z=l,
15x + 15y + 3z = 2.
5x+5y+ z = 0,
15. 5x + 5y + 2z = 0,
5x + 5y + 3z = 2.
16. (Solution of a Complex System) Solve the complex system

by rref methods and show that the unique solution is

x = (1, i).
17. (Inconsistent Systems) Three planes meet in a unique point that is not on a fourth plane. Explain the
connection with inconsistent systems of equations.
18. (Inconsistent System) Prove that a 3 x 3 linear system Ax = b is inconsistent and has no solution if
and only if there is a row echelon form of the augmented matrix (A I b) having a row with 3 zeros and
4th entry nonzero.
19. (Uniqueness) Prove that a 5 x 3 linear system Ax = b has a unique solution x if and only if a row
echelon form of its augmented matrix (A I b) has 3 rows with a leading 1 and rows 4 and 5 are all zeros.
294 5 liNEAR ALGEBRA

20. (Infinitely Many Solutions) Prove that a consistent n x n linear system Ax == b has infinitely many
solutions x if and only if a row echelon form of its augmented matrix (A Ib) has at least one zero row.
21. (Refinery) Two refinery sites X and Y produce total amounts K of kerosene and G of gasoline. At site
X, they process x barrels of crude oil, and y barrels are processed at site Y. Production at these sites is
slightly different, resulting in the equations

Kerosene 10x+ 2Sy == K,

Gasoline Sx+ ISy == G.
Solve for the raw material counts x and y.
22. (Consistency) Let

A == ( 4 -1 2 6)
-1
3
5
4
-1 -3
1 3
.

For any column vector b, determine consistency conditions for solvability of Ax == b and report the
solution x (which depends on the components of b).
23. (Tridiagonal Systems) Let A == (aij) be an n x n matrix defined by the formulas akh == 2, 1 ::: k ::: n;
ak+l,k = ak,k+l == -1,1 ::: k ::: n - 1; aij == 0 otherwise. Let b be the first column of the identity
matrix. Solve the system Ax == b for n = 7.

5.3 Vector Spaces

The operations of vector addition and scalar multiplication of vectors x = (Xl, X2, ... , xn)
in Rn (n-vectors) were defined in Section 5.1 by

(1)

and

(2)

respectively. This section develops the basic concepts and notations of vector algebra,
based on these operations, that are needed in Section 5.5 to describe the structure of
solutions of linear systems Ax = h.
A vector space V in Rn is a subset of Rn that is closed under vector addition and
scalar multiplication. This means that if x and yare any vectors in V and a is any scalar
then the vectors x + y and ax are also in V.
Vector spaces occur frequently in linear algebra and many will appear in the
remainder of Chapter 5. Here the concept is illustrated with two examples that are
associated with a matrix A.
The Null Space of A. If A is an m x n matrix then the set N(A) of all n-vectors x
that satisfy Ax = 0 will be called the null space of A. Thus N(A) is the solution set
 5.3 VECTOR SPACES 295

of the homogeneous system Ax = O. Here 0 denotes the zero m-vector (0,0, ... ,0) (m
terms). In Section 5.1 it was shown that for all n-vectors x and y and all scalars a, one
has A(x + y) = Ax + Ay and A(ax) = aAx. In particular, if x and yare in N(A) then
so are x + y and ax. Thus N (A) is a vector space in Rn.

The Range of A. If A is an m x n matrix, then the set R(A) of all m-vectors b such that
Ax = b has a solution x in Rn is called the range of A. Thus R(A) is the set of all right-
hand sides b for which the system Ax = b is solvable. To prove that R(A) is a vector space,
recall that if b' and b" are in R(A), then there are n-vectors x' and x" such that Ax' = b'
and Ax" = b". But then, by (12) of Section 5.1, b' + b" = Ax' + Ax" = A(x' + x"),
and also ab' = aAx' = A(ax'). Hence b' + b" and ab' are also in R(A). Thus R(A) is
closed under addition and scalar multiplication, and so R(A) is a vector space.

Subspaces of a Vector Space. The set Rn is clearly a vector space in Rn, since it
is closed under n-vector addition and scalar multiplication. All other vector spaces in
Rn are proper subsets of Rn. Such vector spaces are called vector subspaces of Rn. The
smallest subspace of Rn is the subspace Vo = {OJ whose only member is the zero n-
vector 0 = (0,0, ... ,0) E Rn. The subspace Vo is closed under vector addition and
scalar multiplication and thus is a vector subspace of Rn.

Linear Combination of Vectors. If Vi, V2, ... , Vk are vectors in Rn and ti, t2, ... ,

tk are scalars then the n-vector

(3)

is said to be a linear combination of the n-vectors Vi, V2, ... , Vk. The set of all linear
combinations of Vi, V2, ... , Vk is a vector subspace V of Rn. The proof of this is left for
the Exercises. V will be called the span of Vi, V2, ... , Vk and the notation

(4)

will be used. A given subspace V may be spanned by many different sets of vectors. This
is illustrated by the following example.

EXAMPLE 1. In R4 consider the vectors

(5) Vi = (1,-2,3,0),
V2 = (2,0,5,0),

V3 = (5,-2,13,0),
296 5 LINEAR ALGEBRA

and define the subspace

(6)

Then it is also true that

(7)

To see this, recall that span {VI, V2} is the set of all vectors

(8)
while span {VI, V2, V3} is the set of all vectors

(9)

Every vector V in span {VI, V2} is in V; just take t1 = 51, t2 = 52 and t3 = a in (9).
To show that every vector in V has the form (8), and thus is in span {VI, V2}, the easily
verified equation

(10)

may be used. It follows that every w in V can be written

(ll) I w =
=
+ t2V2 + t3(Vl + 2V2)
t1 v 1
(t1 + t3)VI + (t2 + 2t3)V2,

by the rules of vector algebra. This completes the proof. Equation (10) can be used
Similarly to eliminate VI or V2, and one has

(12)

which shows both the redundancy in the definition (6) and the nonuniqueness of the
spanning sets.

In constructing a vector subspace V = span {VI, ... , Vk} in Rn it is desirable to have

the spanning set VI, V2, ... , Vk as small as possible. In Example 1, the equation (10)
implied that the spanning set VI, V2, V3 was too big. In the same way, a spanning set VI,
V2, ... , Vk will be unnecessarily big if one or more of the vectors can be written as a
linear combination of the others. This concept may be defined precisely as follows.
Definition. A set of vectors VI, V2, ... , Vk in Rn is said to be linearly dependent if there
are scalars t1, t2, ... , tk, not all zero, such that

(13)
 5.3 VECTOR SPACES 297

Thus if (13) holds with tl # °then VI = (-t2/tj)V2 + ... + (-tk/tl)Vk and Vj is a linear
combination of V2, ... , Vk. Similarly, if t2 # 0, then V2 is a linear combination of Vj, V3,
... , Vk, etc. If a set of vectors VI, V2, ... , Vh is not linearly dependent, we shall say that
it is linearly independent. This is equivalent to the following definition.
Definition. A set of vectors Vj, V2, ... , Vk in RH is said to be linearly independent if
and only if equation (3) holds only when tj = t2 = ... = tk = 0.
Equation 00) can be rewritten as

(14)

and it follows that the set Vj, V2, V3 of Example 1 is linearly dependent. On the other
hand, any two of these vectors are linearly independent. To verify this for the pair Vj, V2
it must be shown that the equation tjVj + t2V2 = 0 can hold only if tl = t2 = 0. The
rules of vector algebra give

(15) tlVj + t2V2 = (t j + 2t2, -2tj, 3tj + 5t2, 0)

= (0,0,0,0).
Equating corresponding components gives the equations tl + 2t2 = 0, -2tl = 0,
3tl +5t2 = 0, whose only solution is tj = t2 = 0. Thusvj and V2 are linearly independent.
Bases for a Vector Space. A set of vectors Vj, V2, ... , Vk in RH is said to be a basis
for a subspace V in RH if and only if

(16)

and

(17) Vj, V2, ... , Vk are linearly independent.

An advantage to describing a subspace V by a basis, rather than a more general spanning

set, is described by the folloWing theorem.

Theorem. If VI, V2, ... , Vk is a basis for a subspace V in RH then every vector V in V
has a unique representation as a linear combination

(18)

Proof. Only the uniqueness needs to be shown. Suppose that a vector V in V had two
representations:

(19) V = tjVj + t2V2 + ... + tkvh,

V = t~Vj + t;V2 + ... + t~Vk.
298 5 LINEAR ALGEBRA

We wish to show that these are really the same: t1 = t~, t2 = t;, ... , tk = t;,. But (19)
implies that

(20)

Moreover, V1, V2, ... , Vk are linearly independent, by assumption. Hence, (20) implies
that t~ - t1 = ... = t~ - tk = 0, or t1 = t;, t2 = t;, ... , tk = t~. This completes the proof.

A given vector space V has many different bases. Thus in Example 1, each of the
pairs {V1, V2}, {V1, V3} and {V2, V3} is a basis for V; see (12). More generally, an infinite
number of different bases for V are defined by

(21) W1 = cos 8V1 + sin 8V2,

W2 = - sin8v1 + COS8V2,
where 8 is a real number such that 0 ::: 8 < 2n. The proof is left to the Exercises.
Dimension of a Vector Space. A vector space V in Rn is said to have dimension
k if V has a basis V1, V2, ... , Vk consisting of k vectors. This definition is unambiguous
because of the following theorem.

Theorem. Let VI, V2, ... , Vk and W1, W2, ... , We be any two bases of a vector space V
in Rn. Then k = £.
Proof (Optional). We shall show that if k < £ then the set W1, W2, ... , We is linearly
dependent and hence not a basis. The opposite case k > £ then leads to the conclusion
that VI, V2, ... , Vk is linearly dependent.

Each of the vectors Wj is in V and V1, V2, ... , Vk is a basis for V. Hence there are
constants aij such that

(22)

for j = 1, 2, ... , £. The coefficients aij define a k x £ matrix A = (aij) (where 1 ::: i ::: k,
1 ::: j ::: e). Using this matrix it will be shown that there exist constants t1, t2, ... , te not
all zero such that

(23)

and hence the vectors Wj are not linearly independent and hence are not a basis. To prove
this, substitute (22) into (23) and collect terms to get

(all t1 + al2t2 + ... + alftg)v1

+ (a2l tl + ant2 + ... + a2e tg)V2
 5.3 VECTOR SPACES 299

+ ...

Since the vectors Vj, V2, ... , Vk are linearly independent, the coefficients of each Vm must
be zero. This gives k linear equations in the e unknowns tj, t2, ... , te, which can be
written in matrix form as At = 0. But A is a k x ematrix with k < e, so there are more
unknowns than equations. The Gauss elimination method of Section 5.2 implies that
such a system always has nontrivial solutions

This completes the proof.

A Basis for Rn. A simple basis for the vector space Rn is defined by

(24) ej = (1,0,0, ... ,0),

e2 = (0,1,0, ... ,0),

en = (0,0,0, ... ,1),

where ej has the entry 1 in place j and zero in all others. A linear combination of the
vectors has the form

(25)

and this is the n-vector (0,0, ... ,0) if and only if tj = t2 = ... = tn = 0. Thus the set
of vectors ej, e2, ... , en is linearly independent. Moreover, each x = (Xj, X2, ... ,xn ) in
Rn can be written

(26)

and hence the set (24) spans Rn It follows that (24) defines a basis for Rn and R n
has dimension n. As a corollary, we conclude that any n + 1 vectors in Rn are linearly
dependent.
Testing Sets of Vectors for Linear Dependence. Let Vj, V2, ... , Vk be a set of k
vectors in Rn. By definition, the set is linearly dependent if and only if the vector equation

(27)
300 5 liNEAR ALGEBRA

holds for a coefficient set t = (t I, t2, ... ,tk) i- (0, 0, ... , 0). To test this, introduce the
components of the vector Vj by

(28)

After substituting this into (27), that equation becomes

(29) (alltl + ... + aIktk, a2ltl + ... + a2ktk,' .. ) = (0,0, .. .)-

If the n x k matrix A = (aij) is introduced then (29) can be written as the linear
homogeneous system

(30) At = 0.

Thus the vectors VI, V2, ... , Vk are linearly dependent if and only if (30) has a nontrivial
solution. This can be decided by using Gauss elimination to find the reduced row echelon
form of A. The matrix A is readily written down, since its columns are just the vectors

Example 1 concluded. The vectors VI, V2, V3 defined by (5) are known to be linearly
dependent. To verify this by the test described above we have the row equivalence

A= (-~ ~ -~)
3513
{=::} (~ ~ ~).
000
000 000

Clearly, At = 0 has nontrivial solutions. In fact, t3 is arbitrary, t2 = -2t3 and tl = -t3,

so t = t3(-1, -2, 1). On taking t3 = -1 we get

VI + 2V2 - V3 = 0,

as before.

EXAMPLE 2. Consider the 4 x 6 matrix A defined by (18) in Section 5.2. Then A has the
same null space as the reduced row echelon matrix B defined by (19) of Section 5.2. It
follows that x is in N(A) if and only if

(31) x= s( -3,1,0,0,0,0) + t( 1,0,1,0,0, 0) + u( -7,0,0, -2,1,0)

where s, t and u are arbitrary Hence the three 6-vectors

(32) VI = (-3,1,0,0,0,0),
 5.3 VECTOR SPACES 301

V2 = (1,0,1,0,0,0),
V3 = (-7,0,0,-2,1,0)

are a basis forN(A), since theyspanN(A) by (31). Moreover,x2 = S,X3 = tandx5 = U,

and so x = 0 if and only if S = t = U = 0, so VI, V2, V3 are linearly independent. We shall
show later that Gauss elimination always produces a basis for N(A) whose dimension
in the number of independent unknowns.

ExAMPLE 3. Balancing a Chemical Reaction This example was introduced in Section

5.1, Example 3. The coefficients Xl, ... , X7 that are needed to balance the chemical
reaction (4) of Section 5.1 must satisfy Ax = 0, where A is the 6 x 7 matrix

° ° ° ° °
2 -1

° ° °
1
° 0 -8

A=
° ° ° ° °
1 -2

°° ° ° ° 1 -1

°° ° ° 3 -1 -1

° ° ° °
1 -1 -2

Gauss elimination gives the equivalent matrix

2
°° ° -1
° °
° ° ° °
2
° -4

B=
°° ° 2 -2
° 0
0
°° 1 0 -16
°
°° 0 0 2
° -1
0
° ° N°
Thus X7 is the only independent unknown, and
0 -64 3

(A) has dimension 1. A brief calculation

gives a spanning vector

x=s (3
- 21- 3- -
1 -31 )
8' '2'4'2'64' ,

where s is an arbitrary constant. To solve the chemistry problem, the coefficients Xk must
be positive integers. This is achieved by taking s = 64n, where n is a positive integer.
The solution with the smallest coefficients is given by n = 1, so

x = (24,128,32,48,32,3,64),
302 5 LINEAR ALGEBRA

and the balanced chemical reaction is

24CU25 + 128H+ + 32N03 ----+

48Cu2+ + 32NO + 358 + 64H 20.
Other Vector Spaces. Vector spaces were defined above to be subsets V of Rn that
are closed under the operations of vector addition and scalar multiplication, as defined
by (1) and (2). These vector spaces have the following ten simple properties.

(I) x + y is in V whenever x and yare in V.

(II) x + y = y + x for all x and y in V.
(III) (x + y) + z = x + (y + z) for all x, y and z in V.
(IV) There is an element 0 in V such that x + 0 = x for all x in V.
(V) For each x in V there is an element -x in V such that x + (-x) = o.
(VI) ax is in V whenever a is in R and x is in V.
(VII) a(x + y) = ax + ay for all a in R and all x and y in V.
(VIII) (a + f3)x = ax + f3x for all a and f3 in R and all x in V.
(IX) a(f3x) = (af3)x for all a and f3 in R and all x in V.
(X) Ix = x for all x in V.
The properties are readily verified by using the definitions 0), (2) and elementary
properties of real numbers. An important inSight of nineteenth-century mathematics was
the discovery that there are many systems, besides Rn and its subspaces, that share these
ten properties. Moreover, such systems have found many uses in applied mathematics
and engineering. This suggests the following extended definition.
Extended Definition. A set V together with operations (x, y) ----+ x + y and a, x
----+ ax is said to be a vector space whenever properties (I)-(X) hold.
A review of our discussion of the vector space Rn will show that the concepts and
theorems stated for Rn and its subspaces are based entirely on properties (I)-(X). Thus the
concepts of subspace, linear combination, span, linear dependence and independence,
basis, and dimension extend immediately to all vector spaces. This section closes with
several examples of vector spaces. Many others will be met in other chapters.

EXAMPLE 4. Let V be the set of all solutions of a linear homogeneous second-order

differential equation

y" + a(x)y' + b(x)y = 0,

defined on a fixed interval a ::: x ::: b. Properties (I)-(X) are easily seen to hold with the
usual definitions of sums and scalar multiples. Thus V is a vector space. For this example
 5.3 VECTOR SPACES 303

the notions of linear independence, basis and dimension were already encountered in
Chapter 3, where V was shown to have dimension 2. More generally, the solution set of
a linear homogeneous nth-order equation,

is a vector space of dimension n.

EXAMPLE 5. Fix a positive integer n and let Vn be the set of all polynomials of degree less
than or equal to n,

where ao, ... , an are arbitrary scalars. Then Vn is a vector space. It is not difficult to
show that the monomials 1, x, x 2 , ... , xn are linearly independent and hence Vn has
dimension n + 1.

EXAMPLE 6. Let V denote the set of all polynomials eno restriction on n). Then V is a
vector space, and it contains Vn as a subspace for each integer n :::: O. Clearly, V cannot
have finite dimension. Such spaces are said to be infinite-dimensional.

EXAMPLE 7. Let C [a, b J denote the set of all continuous functions on a fixed interval
[a,bJ. Then C[a,bJ satisfies properties (I)-eX) and hence is a vector space. It contains
the set of all polynomials as a subspace and hence is infinite-dimensional.

EXAMPLE 8. Let V be the set of all functions jex) that are defined for -00 < x< 00 and
satisfy

Then V is a vector space. In this case it is not obvious that V is closed under addition.
This will be shown later to follow from the Schwarz inequality:

Complex Vector Spaces. Properties (I)-(X) are stated for scalars from the real
number field R. These properties characterize real vector spaces. If R is replaced by the
complex field C throughout properties (I)-(X), the resulting vector space is said to be a
complex vector space. For eigenvalue problems and some other applications it will be
necessary to use complex vector spaces.
304 5 LINEAR ALGEBRA

Exercises 5.3

Null Space and Range. Determine the null space and range of the matrix A.

1. A == (~ ~). 2. A == (~ ~).
Basis for the Null Space of A. Determine a set of basis vectors that span the null space of A.
1 0
1100)
3. A == ( 0 0 1 0 . ( o 1
000 0 4. A == 1 1
2 0
Independent Sets. Determine a smallest subset that has the same span as VI, V2, v3.

6. VI == G),V2 == G),V3 == G)·

Test Independence of Vectors. Determine whether the columns of the matrix A are dependent or
independent.

7. A == (~ ~). 8. A == ( 232102)
1 1 0 .

9. (Trivial Subspace) Prove that the origin is a proper vector subspace of Rn.
10. (Subspace) Assume n > 1 and let x be a fixed vector in Rn. Prove that the set V == {ax : a real} is a
proper vector subspace of Rn.
I l. (Subspace) Let XI, ... , Xk be certain vectors in Rn. Let V be the set of all linear combinations

k
LCiXi,
i=1

where q to Ck are arbitrary real numbers. Prove that V is a vector subspace of R" (it may be all of R").
12. (Span) Show that if VI and V2 are n-vectors then

span {VI, V2} == span {VI + V2, vI - V2}·

13. (Span) For the 4-vectors VI, V2 and V3 of Example 1, defined by (5), show that

span {VI, V2} == span {V2, V3}·

14. (Basis) Prove that vectors WI and W2 in (21) define a basis for the subspace V in Example 1.
15. (Vector Subspace of Rn) Let V be a vector subspace of R n of dimension n. Prove that V == R"-
16. (Vector Space of Tuples) Prove that the space Rn with the usual definitions of vector addition and
scalar multiplication satisfies the properties (I)-(X) for a vector space.
17. (Vector Space of Functions) Prove that the set V of functions on a set] with the usual definitions of
function addition and scalar multiplication satisfies properties (I)-(X) for a vector space.
 5.4 MATRICES AND MATRIX ALGEBRA 305

5.4 Matrices and Matrix Algebra

Matrices were introduced in Section 5.1 as a concise notation for systems of linear
equations. This section develops the algebraic operations on matrices that can be used
to solve such systems.
The notation Mm,n will be used to denote the set of all m x n matrices (m rows, n
columns) where m and n are positive integers.
Sums and Scalar Multiples of Matrices. If A and B are in Mm,n and if an n-vector
x is a solution of both Ax = bi and Bx = b 2 then x is also a solution of the system
(1)

This is a system of the form ex = b i + b 2 , where e = (Cjn) is defined by

(2)

We shall write e = A + B and call e the matrix sum of A and B. Similarly, if x is a

solution of Ax = b and a is a scalar, then x is also a solution of

(3) a (Ax) = abo

This is a system of the form C'x = ab, where C' = (cJ k ) is defined by

(4)
We shall write C' = aA and call C' the scalar multiple of A by a. The set Mm,n, with
operations of addition defined by (2) and scalar multiplication defined by (4), is a vector
space. The proof is a verification of properties (I)-(X) of Section 5.3. The space Mm,n is
spanned by the matrices Eij (1 :::: i :::: m, 1 :::: j :::: n), where Eij is obtained from the zero
matrix by replacing the element in row i and column j by 1. Moreover, these matrices
are readily seen to be linearly independent and hence form a basis of Mm,n. In particular,
Mm,n has dimension mn.
Some of the vector spaces Mm,n are familiar from earlier work. The space Ml,l is just
the scalar field R (or C) interpreted as a vector space of dimension 1. More generally,
MI,n is the vector space Rn (or Cn) of n-tuples of scalars, or row vectors. Similarly, Mm,l
is the space of n x 1 matrices, or column vectors, and is isomorphic to R n (or en).
Matrices as Linear Operators. Every matrix A in Mm,n determines a linear operator
from R n to Rm. Explicitly, A carries each x in R n to the vector b = Ax in Rm, where
n
(5) hi = L:>ijXj, I:::: i :::: m.
j=l
306 5 LINEAR ALGEBRA

The linearity of the mapping means that (see (13) in Section 5.1)

(6) A(ax + f3y) = aA(x) + f3A(y)

for all vectors x, y in Rn and all scalars a, f3. Conversely, any mapping A of Rn to Rmthat
satisfies (6) has the form (5), where aij is component i of Aej and el, e2, ... , en is the
basis for Rn of (24) in Section 5.3. The simple proof is left to the Exercises.
Composition of Linear Operators. Consider matrices A in Mm,n and B in Mn,p' If B
maps x in RP to y = Bx in Rn and A maps y in Rn to z = Ay in Rmthen the composition
map

(7) z = A(Bx)
maps x in RP into z in Rm. Moreover, it is clearly linear and hence has the form

z= ex,
where e is in Mm,p' The composite map e is called the matrix product of A and B and is
written e = AB. To calculate the components of e, we use the component representations
of A and B. Thus
m P
Zj = LajkYk, Yk = LbktXt.
k=l £=1

Combining these two maps gives

p m
(8) Zj = L L ajkbk£Xt.
f=l k=l

Thus C = (Cjt) is given by

m

(9) Cjt = L ajkbkt, 1::: j ::: m, 1 ::: £ ::: p.

k=l

Matrix Products. The order of the factors A and B in the matrix product e= AB
must be preserved. As a linear operator the product AB means first apply B then apply
A. For this to be meaningful it is necessary that the number of columns of A equal the
number of rows of B. This restriction implies that the product in the opposite order BA
may not be defined. In fact, if A is in Mm,n and B is in Mn,p then BA is defined only if p = m.
The Scalar Product Rule for Computing Matrix Products. If x = (Xl, X2, ... ,xn )
and y = (Yl,Y2,'" ,Yn) are in Rn then their scalar product is the number

L XkYk =
n
(10) (x, y) = XIYl + X2Y2 + ... + xnYn'
k=l
 5.4 MATRICES AND MATRIX ALGEBRA 307

With this notation the rule (9) for forming the matrix product C = AB is

(ll) Cjf = (row j of A, column £ of B).

This scalar product rule makes it easy to remember how to compute matrix products.

Associative and Distributive Rules. Matrix multiplication satisfies the rules

(12) (AB) C = A (BC) ,

(A + B) C = AC + BC,
A(B+C) =AB+AC,

whenever the indicated products are defined. The first rule implies that multiple matrix
products AIA2 ... An are meaningful without parentheses. Proofs of the rules are left to
the Exercises.

Square Matrices. Matrices in the set Mn,n are called square matrices of size n.
Products of matrices in Mn,n are always defined and Mn,n is closed under matrix
multiplication. Almost any example will show that matrix multiplication in Mn,n is not
commutative. For example, in M2,2 if

A=(12)
3 4'
B=(56)
7 8

then

AB = (1943 5022) ' BA = (2331 34).

46

On the other hand, it may happen that AB = BA. An example is

A= ( -21 0
2)' B= (-3 2)
-2 -4 '

for which

AB = ( -~ -6 ) = BA.
-4

Diagonal Matrices. A square matrix D = (dij) is said to be diagonal if dij = 0 for

i i= j. Thus, using Kronecker's delta (oij =
0 for i i= j, oij = =
1 for i j), writing djj dj , =
308 5 LINEAR ALGEBRA

The notation D = diag(d l , d2 , ... , dn ) will also be used. The multiplication rule gives
the products

cnd, and2 alndn

(13) AD=
a2l dl

anldl
.
and2

an2d2
a2n dn

anndn
)
and

c,an
)
dIan dlal n
d2 a2l d2an d2a2n
(14) DA= .

dnanl dnan2 dnann

In particular, if D = In where
(15) In = diag(l, 1, ... , 1)
then AIn = InA = A. The matrix In is called the identity matrix in Mn,n. Equations (13)
and (14) imply that if D is diagonal and AD = DA for all A in Mn,n then D = dIn, where
d is a scalar and AD = DA = dA.
Singular and Nonsingular Square Matrices. A square n x n matrix A defines a
linear operator x --+ y = Ax in Rn. We know that RCA), the range of A, is a vector
subspace of Rn. The matrix A will be said to be singular if and only if RCA) i= Rn.
Otherwise A is said to be nonsingular. Thus A is nonsingular if and only if RCA) = Rn.
Now, the system Ax = y can be written in the column vector form Csee the Exercises)

(16) Xl
(
a2l
all)
.
..
+ X2 (an)
.
.
an
+ ... + Xn
.
a2n
(aln)
.
. .
=
..
Y2
. (1) .

anI an2 ann n

 5.4 MATRICES AND MATRIX ALGEBRA 309

This says that vectors y in R(A) are in the span of the column vectors of A. In particular,
A is nonsingular if and only if every y in Rn can be written in this way This implies the
follOwing theorem.

Theorem. An n x n matrix A is nonsingular if and only if its column vectors are linearly
independent and thus form a basis for Rn.

Proof. If the column vectors of A are linearly independent, then they must span Rn ,
and hence A is nonsingular. Conversely, if A is nonsingular, then R(A) = Rn , and hence
(16) implies that the column vectors of A span Rn. To do this they must be linearly
independent.

Corollary. An n x n matrix A is nonsingular if and only if the null space N(A) = {OJ.
This means that the only solution of Ax = 0 is the vector x = O.

Proof. If A is nonsingular, then the column vectors of A are linearly independent. Hence
(16) with y = 0 implies that x = O. Conversely, if N(A) = {OJ then (16) with y = 0
implies that A has linearly independent columns and hence A is nonsingular.

Invertible Matrices. An n x n matrix A is said to be invertible if and only if there

is a second matrix B such that

(17) AB = BA = In.

Note that if A is invertible then both A and B must be nonsingular, because (17) holds
if and only if

A(Bx) = B(Ax) = x for all x in Rn.

The converse is also true. We state this as a theorem.

Theorem. An n x n matrix A is invertible if and only if it is nonsingular.

Proof. We have shown that if A is invertible then it is nonsingular. We must show that if
A is nonsingular then it is invertible. To prove this, we note that if A is nonsingular then
the map x ----+ y = Ax has range R(A) = Rn. Moreover, the map is one-to-one. For if
y = AXl = AX2 with Xl =f. X2 then the linearity of A implies that x = Xl - X2 satisfies
Ax = 0, so x is in N(A). But this implies that x = 0 by the Corollary above and so
Xl = X2, contrary to assumption. Thus A maps Rn one-to-one onto R(A) = Rn and hence
310 5 liNEAR ALGEBRA

has a unique inverse x = By. The linearity of A implies that B is also linear. Moreover,
x = BAx for all x in Rn, and y = ABy for all yin Rn, so AB = BA = In, and A is invertible.

Uniqueness of the Inverse. The inverse matrix B of (17) is unique when it exists. For
if (17) holds and also AC = CA = In then AB = In implies CAB = C and hence CA = In
implies that B = C.

Notation. The unique inverse of a nonsingular matrix A is denoted by A-[ .

Test for Invertibility of a Square Matrix. We have shown that a square matrix A
is invertible if and only if N(A) = {O}. Now, solution of Ax = 0 by Gauss elimination
generates all solutions. The matrix A is invertible if and only if x = 0 is the only solution.
This can happen only if the reduced row echelon form of A is the identity In. For if any
diagonal entry of the reduced matrix were zero then Ax = owould have a solution x =I O.

Computation of A -[ . The Gauss elimination procedure can be modified to give an

algorithm for computing A-I. This may be done by representing each elementary row
operation as left multiplication by a nonsingular matrix. There are three types of such
elementary matrices. These matrices will be called elementary row matrices.
Multiplying a Row by a Nonzero Number. This is equivalent to left multiplication by

E = diag(l, 1, ... ,c, . .. , 1), c =I O.

This follows readily from (14). The matrix E- 1 is found by replacing c by lie.

Adding a Row to Another Row. The corresponding matrix is obtained by applying the
same operation to In. Thus to add row q to row p we have

1 o o o

o 1 1 o
E=
o o 1 o

o o o 1

where only the first, pth, qth, and last rows and columns are shown. The remaining rows
and columns are those of In.
 5.4 MATRICES AND MATRIX ALGEBRA 311

Interchanging Two Rows. The corresponding matrix is obtained by applying the same
operation to In. Thus to interchange rows p and q we have

1 0 0 0

0 0 1 0
E=
0 1 0 0

0 0 0 1

where only the first, pth, qth and last rows and columns are shown. The remaining rows
and columns are those of In.
Algorithm for A-I. It was shown above that A is invertible if and only if the reduced row
echelon form is In. Moreover, from Section 5.2, the reduced form can be computed by a
sequence of elementary matrices EI , E2 , ... ,EN of the three types defined above such that

(18)

Multiplying on the right by A-I gives the construction

(19)

The algorithm for generating A -1 is contained in these equations. It states that if the
sequence of elementary row operations reduces A to In, then the same operations produce
A -1 when applied to In. The work may be organized by writing the augmented matrix
(A I In) and applying E I , E2 , ... , EN to it to get (In I A-I).

EXAMPLE 1. Simple row operations give the row equivalence

(A I13 )
(
2
2 -3
3 -6
-2 1 1 0
2 0 1
4 0 0 n
( )
1 0 0 0 2 -1
{=::} 0 1 0 -2 5 -2
0 0 1 -3 6 -2
(I3IA-l).
312 5 LINEAR ALGEBRA

U !)(
As a check we have
o
~ )~h
-2 2 -1
AA-' -3 -2 5 -2
-6 -3 6 -2

Vector Space Mm,n. Let EU be the m x n matrix obtained from the zero matrix by replacing the entry in
row i and columnj by 1. Complete the following.
1. Determine all matrices EU for m = n = 2.
2. Do the matrices Eij span Mm,"?
3. Are the matrices Eij linearly independent?
4. (Linear Operators) Define the mapping nx) = y by Yl = 3Xl - 2X2, Y2 = Xl + 2X2. Compute
anu) + bnv) and nau + by). Is T a linear operator?
Linear Operators on Polynomials. Let T be the mapping from the set V of all polynomials of the form
x = e3t2 + C2t + Cj into R3 defined by nx) = (el, e2, e3).

5. Calculate nax+by)for all polynomials x = al +a2t+a3t2, y = bl +b2t+b3t2 and all constantsaandb.

6. Calculate anx) + bny). Does nax + by) = anx) + bnyp

7. (Conformable Products) Let A = (2, -1, 3) and B = (~). Does AB make sense?

8. (Commuting Matrices) Find two 3 x 3 matrices A and B that commute, both of which have some
nonzero off-diagonal elements.

9. (Powers) Let A = (t ~). Compute A2 - A.

10. (Powers) LetA = diag(l,2, -1, -2) FindA 5

Diagonal Matrices. Complete the following exercises on matrices of diagonal type.
11. FindAB and BA, given A = diag(l,2,-I,-2) andB = diag(-I,-2,2, 1)

12. Find a 3 x 3 complex diagonal matrix D such that D2 +I = O.

13. Let A be the 3 x 3 matrix defined by aij = I/Ci +j), i,j = 1,2,3. Assume that D is a diagonal matrix
such that AD = DA. Determine all possible D.
14. Find a 4 x 4 diagonal matrix D such that D2 = I but D is not the identity I.
15. (Equivalent Systems) Write explicitly the column vector form of the system Ax = y in terms of the
columns of A, given that A is the 3 x 3 matrix

A=( ; -;
-I 5 4
i).
 5.4 MATRICES AND MATRIX ALGEBRA 313

16. (Column Vector Form of a System) Prove that an n x n linear system Ax = y can be written in the
column vector form

Xl (all)
~l
: +X2 : + ... +Xn (a :) = tl)
(al2)
an 1n n
~n
:.

anI an2 ann n

17. (Singular Matrices) Let A = ( ~ ~

-1
- !). Is A Singular or nonsingular?

18. (Singular Matrix) Find necessary and sufficient conditions for which a diagonal matrix of order three
is singular.

19. (Singular Matrix) Find necessary and sufficient conditions for which a diagonal matrix of order n is
Singular.

20. (Elementary Matrix) Are elementary row matrices Singular or nonsingular?

21. (Invertible Matrix) Is A = (

-1
~ ~ - !) invertible? Answer without finding the inverse matrix.

22. (Invertible Matrix) Suppose that a square matrix A is the product of invertible matrices. Explain why
A is invertible.

23. (Invertible Matrix) A given n x n matrix A has reduced row echelon form equal to the n x n identity
matrix. Explain why A is invertible.

24. (Inverse Matrix) Let A =( 32 3)

2 and B = (-2/5
3/5 3/5) . Apply the reduced row echelon
-2/5
inversion method to verify that B = A-I.

25. (Inverse Matrix) Let A = (

-1
~ -~
5 4
i) Apply the reduced row echelon inversion method to

verify that

1/16 19/48 -7/48 )

A- I = ( -3/16 7/48 5/48 .
1/4 -1/12 1/12

26. (Elementary Matrices) Write out all 3 x 3 elementary row matrices E that correspond to multiplying
a row by a nonzero number.

27. (Elementary Matrices) Write out all 3 x 3 elementary row matrices E that correspond to adding a row
to another row.

28. (Permutation Matrices) All possible products of 3 x 3 elementary row matrices E that correspond to
swapping rows in the identity matrix 13 are called the permutation matrices of order 3. There are six
distinct permutation matrices of order 3. Find them.
Calculation of the Inverse Matrix. Find explicitly the inverse of the given matrix.
314 5 LINEAR ALGEBRA

-3
29. A = ( 21 -3
-1
-1 2)
1 .
5 4
34. A= ( ~
-3
6
5
-1 )
4 .
2

-1 _1)
A~ 1
-1 )
30. A= (12 21 -4 , -2 1 5
35. ( o
3 0 -1 0 1 '
-3 0 o 2
A= ( 21

(l -1 )
31. 11 31) ,
0
-1 -3 2

-2 4)
36. A~ 2 -1
1 '
32. -4 -2 , -1 1

A~ (1
A=(l

-: )
-5 3 -2 2
3 2
37. 1 '
33. A= (43 21 1)
1 . 1
231 0 1

38. (Linear Operators) Show that if x --+ A(x) is any linear map from Rn to Rm (so that (6) holds), then
Ax = b is equivalent to the linear system

n
I>ijXj = hi, 1::: i::: m,
j=l

where A(ej) = I:::l aijei (the standard basis el. " ' . em was defined in (24) of Section 5.3)
Associative and Distributive Laws. Prove the following rules, assuming that all products are defined,

39. A(BC) = (AB)e.

40. A(B + C) = AB + Ae.

41. (A + B)C = AC + Be.

42. (Scalar Product Rule) Find the entry in row 3 and column 2 of the product matrix AB, given

5.5 The Fundamental Theorem of Linear Algebra

We have seen that an m x n linear system Ax = b has solutions if and only if b is in
the column space of A; see (16) in Section 5.4. The solutions, when they exist, can be
generated by applying Gauss elimination to the augmented matrix (A [b), This section
develops an alternative test for solvability based on the concepts of scalar product and
orthogonality in Rn ,
 5.5 THE FUNDAMENTAL THEOREM OF LINEAR ALGEBRA 315

Scalar Product in Rn. If x and yare two vectors in Rn then the scalar product, or
inner product, of x and y, written (x , y), was defined in Section 5.4 by the relation
n
(1) (x, y) = L XhYh = XlYl + X2Y2 + ... + XnYn·
h=l

It is easily seen to have the following three properties.

Linearity: (x, (lYl + (2Y2) = (1 (x, n) + (2 (x, Y2).
Symmetry: (x, y) = (y, x).
Positivity: (x, x) ~ 0, and (x, x) = 0 if and only if x = O.
The nonnegative function defined by

(2) IIxll = (x, x)1/2 ~ 0

is called the norm of x. A basic property of the scalar product is the Schwarz inequality.
Schwarz Inequality. For all vectors x and y in Rn,

(3) I(x, y)1 ::: IIxllllyll·

Moreover, equality holds if and only if x and yare linearly dependent.

Proof. Let x and y be fixed vectors and let C{ be any scalar. Then the three properties
of the inner product imply the relations

(4)
Taking C{ = -(x, y)/llyl12 (the minimizing value) and simplifying gives (3). Moreover,
equality holds if and only if x + ay = 0 and hence x and yare linearly dependent.
Geometrical Meaning of (x, y). If x and yare nonzero vectors, then the Schwarz
inequality can be written as

(x, y) I < 1
I IIxllllyll - .
It follows that there is a unique angle esuch that 0 ::: e ::: nand
(5) (x, y) = IIxlillyll cos e.

Thus e can be defined to be the angle between the nonzero vectors x and y. If e = n/2
then cos e = 0 and

(6) (x, y) = o.

Orthogonality. Two vectors x and yare said to be orthogonal if and only if (6) holds.
Note that the zero vector 0 is orthogonal to all vectors and is the only vector with this
316 5 LINEAR ALGEBRA

property (see the Exercises). Moreover, two nonzero vectors are orthogonal if and only
if the angle between them is nl2 radians.
Orthogonal Sets and Bases. A set of vectors in Rn

is said to form an orthogonal set if each Xj i: 0 and

Such a set is necessarily linearly independent. For if

then taking the scalar product of this equation with any vector Xk and using the
orthogonality gives

Since IIxkll i: 0, this gives CI = C2 = ... = cp = O.

As a corollary, we see that an orthogonal set (7) in Rn must satisfy p ::::: n = dim(Rn ).
Moreover, any set of n orthogonal vectors in Rn must be a basis (see the Exercises). Such
bases are said to be orthogonal bases.
Orthonormal Sets and Bases. If (7) is an orthogonal set in Rn and if

then the set YI, n, ... , YP consists of orthogonal unit vectors and has the same span as
the set (7). An orthogonal set of unit vectors is called an orthonormal set. In particular,
an orthogonal basis of unit vectors is called an orthonormal basis.
Orthonormal bases are particularly simple to work with. Thus if Yl, Y2, ... , Yn is an
orthonormal basis in Rn and x is any vector in Rn then

(8)

where

Ck=(X,Yk), k=1,2, ... ,n.

This is easily verified by taking the scalar product of equations (8) and Yk.
The Gram-Schmidt Construction. This method starts with a basis
 5.5 THE FUNDAMENTAL THEOREM OF liNEAR ALGEBRA 317

in Rn and builds an orthogonal basis

by a step-by-step procedure as follows.

Step 1.
Step 2.
Step 3.

Step k.

At Step k one never gets Yk = °because if this occurred then Xk would be a linear

°
combination of x I, X2, ... , Xk-l and the original set {Xl, X2, ... ,xn } would not be a basis.
Moreover, at Step k the construction gives (Yj , Yk) = for j < k. Hence the new set
{YI, Y2, ... ,Yn} is orthogonal.

EXAMPLE 1. In R2 take

so (Xl, X2) = 1 =j:. 0. Then Yl = Xl = (1,0) and

(X2, YI)
Y2 = X2 - IIYl1l 2 YI

= (1,1) - (1,0) = (0,1).

Orthogonal Subspaces and Orthogonal Sums. We shall write X .1 Y to mean

that X and yare orthogonal. More generally, let VI and V 2 be two subspaces of the same
space Rn. Then we shall say that VI and V 2 are orthogonal subspaces of Rn, written
VI .1 V 2 , if and only if every VI in VI is orthogonal to every V2 in V 2 . If this is the case

then we shall denote by VI EB V 2 the set of all vectors V = VI + V2 with VI in VI and V2

in V 2 . It is easy to verify that VI EB V 2 is a subspace of Rn Moreover, each vector V in
VI EB V 2 has a unique representation V = VI + V2 with VI in VI and V2 in V 2 . For if

where z is in VI and V2 . Since VI .1 V 2 , this implies that z .1 z, or (z, z) = 0. Hence

z = 0, which implies that VI = WI and V2 = W2. This proves the uniqueness.
318 5 LINEAR ALGEBRA

Orthogonal Complements. If Y is any subspace of Rn , then we shall denote by

y-L the set of all vectors in Rn that are orthogonal to each vector of y. y-L is called the
orthogonal complement of y. It is easy to see that y-L is a subspace and y-L 1. y. The
name is justified by the follOwing theorem.

Theorem. For each subspace Y of Rn we have

(9)

proof. Y EB y-L is a subspace of Rn. To prove (9) it is enough to show that any vector x
in Rn can be written as x = Xl + X2 with Xl in Y and X2 in y-L. To do this, let

be an orthonormal basis in Y (possible by the Gram-Schmidt procedure) and define

k
Xl = L(X' Vj)Vj.
j=l

Clearly, Xl is in y. Next, define X2 by

k
X2 =X-XI = x - L(x, Vj)Vj.
j=l

Then

(X2, vel = (x, vel - (x, Ve) = 0 for l = 1,2, ... ,k

by the orthonormal property of {VI, V2, ... , Vk}. Thus

and so X2 is in y-L as required. Since x = Xl + X2, the proof is complete.

Corollary. For each subspace Yin Rn one has

(0)

Proof. Apply the preceding theorem to the subspaces Y and y-L. This gives

Rn = y EB y-L = y-L EB (y-L)-L .

Thus every vector x in Rn can be written as

 5.5 THE FUNDAMENTAL THEOREM OF liNEAR ALGEBRA 319

where VI is in Y, V2, and WI are in y-L and W2 is in (y-L)-L. Since V2 - WI is in y-L and
W2 = Vj + (V2 - WI), we have

0= (W2, V2 - WI)

= (VI + (V2 - WI), V2 - WI)

= IIV2 - w1112.
Thus V2 = WI, and therefore also VI = W2. Hence W2 is in Y and (10) follows.
Matrix Transpose and Adjoint. With each m x n matrix A we may associate the
n x m matrix AT obtained from A by interchanging rows and columns. Thus the row
vectors (column vectors) of AT are the column vectors (row vectors) of A. The matrix AT
is called the transpose of A. If A = (aij) and AT = (a&) , then the definition means that

(ll) aJ = aji for 1 sis n, 1 S j S m.

Clearly, (AT) T = A.
A second n x m matrix A * may be associated with A by the equality

(12) (Ax, y) = (x, A*y) for all x in Rn and y in Rm.

The matrix A* is called the adjoint of A. We shall show that the matrices AT and A * are
the same. More precisely, we shall prove the following theorem.

Theorem. For each m x n matrix A there is exactly one n x m matrix A* such that (12)
holds and it is

(13)

Proof. For any vectors x in Rn and y in Rm we can write

n
(Ax)i = L aijXj
j=1

and
n
(Ax, y) = L (AX)iYi
i=1

n m
= LXjLaijYi.
j=1 i=1
320 5 LINEAR ALGEBRA

The last sum has the form (x, Ny), where

m m
(A*Y}j = LaijYi = LajiYi'
i=l i=l
so

This completes the proof.

After these preliminaries, we can now formulate the central result of this section as
follows.

Theorem. For all m x n matrices A one has

(14) R(A)EBN(A*) = Rm ,
(15) R(A*) EBN(A) = Rn.

This result is known as the fundamental theorem oflinear algebra because, as will
be shown below, it contains complete information on the solvability of the linear system
Ax=b.

Proof. Equation (15) follows from (14) on replacing A by A* and A* by (A*)* = A (so
m is replaced by n). Thus only (14) need be proved. To prove (14) we recall that for each
subspace Y of Rm one has Y EB y.l = Rm. Applying this to Y = R(A) gives

R(A) EB R(A).l = Rm.

Hence the proof may be completed by showing that

(16) R(A).l = N(A*).

To show this, we note that if y is in R(A).l, then for any x in Rn one has

0= (Ax, y) = (x, A*y),

whence Ny = 0 (take x = Ny). Thus y is inN(A*) and we have proved that R(A).l is
a subspace of N(N). For the converse, let y be in N(A*). Then for all x in Rn one has

o = (x, Ny) = (Ax, y)

and hence y is in R(A).l. ThusN(A*) is a subspace ofR(A).l. This completes the proof
of (16).

Fredholm Alternative Theorem. This result, which gives complete information about
the solvability of linear systems Ax = b, may be formulated as follows.
 5.5 THE FUNDAMENTAL THEOREM OF LINEAR ALGEBRA 321

Theorem. Let A be an m x n matrix. Then either

Ax = b has a solution for each b in Rm, or

A*y = 0 has nontrivial solutions.

In the second case, Ax = b has solutions if and only ifb J.. NCA*).

Proof. The first case occurs when dim RCA) = m and hence, by (14), NCA*) = {O}.
The second case occurs when dim RCA) < m and hence dimNCA*) > O. Moreover, in
the second case RCA) = NCA*)~ and hence vectors bin Rm are in RCA) if and only if
b J.. NCA*).

Rank and Nullity of a Matrix. The rank and nullity of a matrix A are the
nonnegative integers

(17) rankCA) = dim RCa),

nullityCA) = dim NCA),

respectively The fundamental theorem (14), (15) implies that if A is m x n then

(18) rankCA) + nullityCA *) = m,

rankCA*) + nullityCA) = n.

We have shown that the column space of A is the subspace RCA) in Rm. Similarly,
the row space of A is the span of the row vectors of A in Rn. This coincides with the
subspace R(AT) = RCA*) in Rn. It is a remarkable fact that R(A) and RCA*) always
have the same dimension, even when m =f. n. The result may be formulated as follows.

Theorem. For every m x n matrix A one has

(19) rankCA*) = rank(A).

Proof. We shall show that for every matrix A one has

(20) rank(A*) :s rank(A).

Equation (19) then follows on replacing A by A*, which gives

(21) rank(A) :s rank(A*),

since (A*)* = A. Inequalities (20) and (21) imply (19).
322 5 LINEAR ALGEBRA

To prove (20), let aI, a2, ... , an denote the column vectors of A = (ely), so

all)
a21 (an)
a22 (aln)
a2n
al = ( . ,
..
a2 = .,
..
an = . .
..
amI am 2 amn
If we write r = rank(A) then the definition of rank(A) implies that there is a basis for
R(A) with r vectors, sayvl, V2, ... , Yr. Hence there exists an n x r matrix (Cij) such that

(22)

These are vector equations in Rn. Let

Then taking kth components in (22) gives

akl CllVlk + CnV2k + ... + ClrVrk,

=
ak2 = e21 Vlk + C22V2k + ... + C2rVrk,

These can be written as a vector equation for the kth column of A*:

ak2
akl) C21
(ell) e22
(en) (Clr)
C2r
( : = Vlk : + V2k : + ... + Vrk : .
akn Cn I Cn2 enr
This says that R(A*) is spanned by r vectors. It follows that dim R(A*) :::: r == rank(A),
which is equivalent to (20). This completes the proof.

Corollary. For any m x n matrix A one has

(23) rank(A) :::: min(m, n).

 5.5 THE FUNDAMENTAL THEOREM OF LINEAR ALGEBRA 323

This follows from (18) on replacing rank(A*) by rank(A).

Corollary. For any m x n matrix A one has

(24) rank(A) + nullity(A) = n.

EXAMPLE 2. Considerthe 4 x 6 matrix A of (18) in Section 5.2 with reduced row echelon
form B of (19), Section 5.2. Here

rank(A) = number of dependent variables of B

= number of nonzero rows of B
=3,
nullity(A) = number of independent variables of B
=3.

So rank(A) + nullity(A) = 6, which is the number of unknowns of the system Ax = b.

Structure of Solutions of Ax = b. Linear systems Ax = b have been shown to have
solutions if and only if b ..1 N (A *). Moreover, when solutions exist, they have the form

(25)

where xp is any particular solution of Ax = b,

is a solution basis for Ax = 0, and tl, t2, ... ,tk are arbitrary scalars. In particular, by (24),

k = nullity(A) =n- rank(A).

The rank of A, k, and the vectors X p , Xl, X2, ... , Xk are all readily generated by Gauss
elimination.
Complex Field. The results of this section extend to linear systems with complex
coefficients, provided that the scalar product is redefined by
n
(x, y) = I>kYk,
k=l

where Z denotes the complex conjugate of z. The adjoint of a complex matrix A is then
-T
A* = A ; that is,
324 5 LINEAR ALGEBRA

1. (Orthogonal Subspace) Let V be a plane through the origin in R3 Describe geometrically how to
construct a basis for V-1.
2. (Orthogonal Complement) Let K be any subset of R", not necessarily a subspace, and let K-1 denote the
set of all vectors orthogonal to all of the vectors in K. Prove that K-1 is a subspace of R" and determine
(K-1 )-1

3. (Gram-Schmidt) Apply the Gram-Schmidt process to the basis

Xl = (1,0,0), X2 = (1,1,0), x3 = (1,1,1)

to obtain the orthonormal set

YI = (1,0,0), Y2 = (0,1,0), Y3 = (0,0,1).

4. (Orthonormal Sets) Show that if Yl, Y2, ... , Yn is any orthonormal set in Rn then it is also a basis in Rn.
5. (Gram-Schmidt) Apply the Gram-Schmidt process to the basis

Xl = (1,1), X2 = (1,0)
to obtain the orthogonal set

YI = (1,1), Y2 = (1/2,-1/2).

6. (Gram-Schmidt) Prove that at Step k in the Gram-Schmidt construction, Yh is orthogonal to Xl, ... ,
Xk-l and also to Yl, ... , Yk-I·

7. (Fredholm Alternative) Give an example of a 3 x 3 matrix A and a vector b such that Ax = b has a
solution and also A*y = 0 has nontrivial solutions.
8. (Orthogonality) Show that the zero vector is the only vector that is orthogonal to all vectors.
9. (Rank) Let A be a 6 x 8 matrix with rank 4. How many vectors are in a basis for the column space of A?

10. (Adjoint Relation) Prove the adjoint relation A* = :;f for the complex case.
11. (Rank) Let A be a 6 x 8 matrix with rank (A T) = 5. How many vectors are in a basis for the row space of A?
12. (Nullity) Compute the nullity of a 6 x 8 matrix A that satisfies rank(A*) = 3.
Fundamental Theorem of Linear Algebra. Compute bases for the four subspaces R(A) , R(A*), N(A),
N(A*) of the fundamental theorem.

13. A = (~ ~ ~).

=( ~ ~)

(l ::n
14. A

15 A=

16. (Inverse Matrix) Let a square matrix A have orthogonal columns. Show that A-I is formed from A by
replacing each column c of A by dllcll 2 followed by taking the transpose of the resulting matrix.
 5.5 THE FUNDAMENTAL THEOREM OF LINEAR ALGEBRA 325

17. (Inverse Matrix) Find the inverse of the matrix

-2 )
1 .
2

18. (Orthogonal System) Let a square matrix A have orthogonal columns aI, ... ,an. Show that the solution
of the linear system Ax = b is the solution of the vector equation

whose solution x is given by the scalar projection formulas

19. (Solution of an Orthogonal System) Find the solution X of the linear system Ax = b, given the matrix
A and vector b below.

Regression. A regression line is a straight line of the form Y = mx + b that best fits data points (Xl, Yl),
... , (Xn, Yn) in the sense that the sum-of-squares error

n
E = L(mxk +b- Yk)2
k=l

is minimized. The solution of the minimization problem can be succinctly expressed in terms of the data
vectors

20. Show that the intercept b and the slope m of the regression line satisfy the equations below.

21. Find the regression line Y = mx + b for the accident data below and test the equation graphically for
quality of fit.

Years driving 4 6 8 10 12 14

Accidents 10 8 4 8 6 5

22. Assume that the data have been subjected to translations such that
326 5 liNEAR ALGEBRA

Show that the solution to the regression problem is b = 0 and that m is given by the projection formula
(x, Y)
m=--.
(x, x)

23. (Linearity) Prove (x, qYl+C2Y2) = q(x, Yl)+C2(X, Y2)foraliconstantsq,qandvectorsx,Yl,Y2.

24. (Symmetry) Prove the real inner product identity (x, y) = (y, x) for all vectors x and y.
25. (Symmetry) Prove the complex inner product identity (x, Y) = (y, x) for all complex vectors x and y.
26. (Positivity) Prove that in the real or complex case (x, x) 2: 0, and (x, x) = 0 if and only if x = O.
27. (Triangle Inequality) Prove IIx + yll ~ Ilxll + Ilyll from the Schwarz inequality

5.6 Determinants and Cramer's Rule

Cramer's rule for solving 2 x 2 systems Ax = b, familiar from elementary algebra, is
extended in this section to n x n systems. Of course, if A and b are given numerically
then x can be computed by Gauss elimination. Here the goal is different. The unknowns
Xl, X2, ... , Xn are to be calculated algebraically as functions of the n2 parameters all, al2,
... , ann (the components of A) and the n parameters bl , b2 , ... , bn (the components of

b). The resulting formulas are the n-dimensional Cramer's rule. The calculation is based
on n x n determinants, which will be defined first.
The Determinant Function. The n x n determinant function assigns to each n x n
matrix A = (aij) a Single real or complex number, which will be denoted by det (A). An
alternative notation for det(A) is IAI. Here the vertical bars "I" do not denote a norm or
absolute value. The notational distinction is that A = (aij) denotes an n x n matrix, an
array of n2 numbers, while det(A) = I (aij) I denotes a single real or complex number,
the determinant of A, which is to be computed from the numbers all, al2, ... , ann. The
definition of det(A) for general n x n matrices A, and algorithms for computing it, are
given below. The main lines of the theory are suggested by the special cases n = 1, 2,
and 3, which are examined first.
1 x 1 Determinants. Here A = (all) and det(A) = all' The system Ax = b is the single
linear equation all Xl = bl , which is uniquely solvable if and only if det(A) = all =f. O.
2 x 2 Determinants. The determinant of a 2 x 2 matrix is defined by

(1) det(A) __ I all al21 = alla22 - allal2.

a21 an

Moreover, the 2 x 2 system Ax = b is

(2) all Xl+ al2X2 = bl ,
a21Xl + anX 2 = h,
 5.6 DETERMINANTS AND CRAMER'S RULE 327

and if det(A) f. 0, then elementary algebra gives the unique solution

anbl - al2h 1 bl al2
(3) Xl = det(A)
all an - alla2l b2 an

-a 21 bl + allb2 1 all b1
X2 = =- -
det(A) b2
allan - allall all

This is the classical Cramer's rule for the case n = 2. An alternative form of (3) that is
useful for the generalization to n > 2 is given next.
Let Mjh denote thejk-minor of the matrix

(4)

Thus Mjk is the 1 x 1 matrix (or number) obtained from A by deleting row j and column
k. Thus Mll = (an), M12 = (a2l), M21 = (al2), and Mn = (all). Next let

(5) Ajk = jk-cofactor of A = (_ly+k det (Mjk)

and define cof(A), the 2 x 2 cofactor matrix of A, by

(6) cof(A) = (All Al2) =( an

All An -al2

Then the definition (1) can be written in the two equivalent forms

(7)

These are called the first and second row expansions of det(A), respectively Moreover,
(l) implies that 2 x 2 matrices satisfy detCA T ) = detCA). Combining this with (7) gives

(8)

These are called the first and second column expansions of det(A), respectively.
Next, define the adjoint matrix of A to be the transpose of cof(A):

(9) adj(A) = CcofCA)l = (All All).

Al2 A22

Then matrix multiplication and the row expansions (7) give

(10) A adj (A) = ( de~A) o )

det(A)
= det(A)I,
328 5 LINEAR ALGEBRA

where 1= diag(l, 1) is the 2 x 2 identity matrix. It follows that if det(A) -::f:. 0, and hence
A-1 exists, then

(ll) A-I = (det(A))-1 adj(A).

Applying this to Ax = b gives an alternative form of Cramer's rule:

(12)

3 x 3 Determinants. The determinant function for 3 x 3 matrices is defined by

au al2 aD
(13) det(A) = a21 a22 aB
a31 a32 a33

= aU a 22 a 33 + a2l a32 a 13 + a31 a l2 a B

Moreover, the 3 x 3 system Ax = b has the form

(14) aU Xl + al2X2 + al3x3 = bl ,

a2lXI + a22X2 + a23X3 = b2 ,
a31XI + a32X2 + a33X3 = b3 .

In elementary textbooks it is shown after considerable effort that if det(A) -::f:. °then (14)
has the unique solution

bl al2 al3
1
(15)
Xl = det(A) b2 a22 a23 ,
b3 a32 a33

au b1 al3
1
X2 = det(A) a2l b2 a23

a31 b3 a33

au al2 bl
1
X3 = det(A) a21 a22 b2
a31 a32 b3
This is Cramers rule for the case n = 3. It will be verified below by a proof that is valid
for all n ~ 2.
 5.6 DETERMINANTS AND CRAMER'S RULE 329

The cofactor expansions derived above for n = 2 will be formulated next for the
case n = 3. The proofs for this case are omitted here because they are covered by the
calculation for n ~ 2 presented below.
For the 3 x 3 matrix

(16)

the jk-minor Mjk is the 2 x 2 matrix obtained by deleting row j and column k of A. Thus

(17) Mll = (an an), M12 = (a 2l an), etc.

a32 a33 a31 a33
The corresponding jk-cofactor is the number

(18) Ajk = (-1 Y+k det ( Mjk ) .

The cofactor matrix of A is defined by

All
(19) cof(A) = A = ( A2l
A31

Note that the factor (-I)i+k in (18) produces a "checkerboard" pattern of + 1 and -1
over the components of cof (A), with + 1 in the upper left corner. The notation (18) and
the definition (13) of 3 x 3 determinant can be combined to give

(20) detCA) = allAll + anAn + a13A13.

In fact, detCA) can be expanded along any row or column:

(21) det(A) = ajlA;l + aj2A;2 + aj3Aj3, j = 1,2,3,

and

(22)

These equations imply that the adjoint matrix

adj(A) = (cof(A»T
satisfies A adj(A) = (det(A» I (now 3 x 3), and thus if det(A) f: 0, then

A -} = (det1(A») adj(A).
330 5 LINEAR ALGEBRA

Hence Cramer's rule for 3 x 3 systems takes the form

(23)
Aljbl + A 2jb 2 + A 3j b3 .
Xj= det(A) ,)=1,2,3.

Column Vector Representation of det(A). The determinant of an n x n matrix

A = (aij) is a scalar-valued function

(24) det(A) =

of the n2 components aij of A. The function has been defined explicitly for n = 1, 2 and
3. For the extension to n > 3 it will be convenient to think of det(A) as a function of
the n column vectors of A,

(25)

and to write

(26)

Clearly, if aI, a2, ... , an are knOWTl then the n 2 components aij are also knoWTI, and
conversely
The Case n = 2. For this case

(27)

It is easy to verify that V (aI, a2) has the three properties

(28) V (aI, a2) is a linear function in each of the vectors aI,

a2 separately
(29) V(al,al) =0.
(30) V (el, e2) = 1,
where
 5.6 DETERMINANTS AND CRAMER'S RULE 331

are the standard basis vectors in R2 Moreover, we shall show that D (aI, a2) is completely
determined by these three properties. This may be stated as a uniqueness theorem.

Uniqueness Theorem. D (aI, a2) = allan - al2a2l is the only function that has
properties (28), (29) and (30).

Proof. Note first that (28) and (29) imply

(31)

Properties (28) and (29) give

o= D (al + a2, al + a2)

= D (aI, al + a2) + D (a2, al + a2)
= D (aI, al) + D (aI, a2) + D (a2, al) + D (a2, a2)
= D (aI, a2) + D (a2, al),
which gives (31). Next we may write

Thus on using (28)-(31) we get

D (aI, a2) = allD (el, a2) + a2lD (e2, a2),

and

D (el, a2) = al2D (el, el) + anD (el, e2) = a22,

D (e2, a2) = al2D (e2, el) + anD (e2, e2) = -au,
whence

This completes the proof.

The Case n = 3. For this case the determinant

all au al3
(32) det(A) = a21 an a23 = D(al,a2,a3)
a31 a32 a33

is defined by the expression (13), and it is easy, but tedious, to show that D (aI, a2, a3)
has the following three properties.

(33) D (a I, a2, a3) is a linear function in each of the vectors

332 5 LINEAR ALGEBRA

aI, a2, a3 separately.

(34) 1) (aI, aI, a3) = 0 whenever two of the vectors aI, a2,
a3 are equal.
(35) 1)(el,e2,e3) = 1,

where el, e2, e3 is the standard basis in R3 Moreover, 1) (al,a2,a3) is uniquely

determined by these properties. Proofs of these statements are left to the Exercises. A
general theorem, valid for all n x n matrices A with n ::: 2, is given below.
The General Case. A correct definition of n x n determinant, valid for all orders n ::: 1,
is suggested by the cases n = 1,2,3 discussed above. It will be based on the following
basic theorem.

Basic Theorem. For each integer n ::: 1 there is a unique function

where each aj is an n-vector, such that

(36) 1) (aI, a2, ... , an) is a linear function of each of the vectors
a l, a2, ... , an separately;
(37) 1) (aI, al, ... ,an) = 0 whenever two (or more) of the vectors
al, a2, ... , an are equal;

where el, e2, ... , en are the standard basis vectors in Rn (ej is the jth column of In).

The behavior of 1) (al, a2, ... , an) under permutations of the vectors el, e2, ... , en
will be needed in the proof of the theorem. The following notation and facts are needed.
A permutation of the set {l, 2, ... , n} is a one-to-one mapping

(39) a: {l, 2, ... , n} ----+ {I, 2, ... , n}.

The notation a(i) will be used for the image under a of the integer i. The set of all
permutations of n objects is the symmetric group Sn and has n! distinct elements. A
transposition is a permutation that interchanges two integers and leaves the others fixed.
It is known that every a in Sn can be represented as a group product of transpositions.
A permutation a is said to be even (respectively, odd) if it is the product of an even
(respectively, odd) number of transpositions. The notation

1, a even,
sgn(a) = (
-1, a odd,

will be used.
 5.6 DETERMINANTS AND CRAMER'S RULE 333

Properties (36) and (37) imply that V (a], a2, ... ,an) changes sign when two of the
variables a], a2, ... , an are transposed. The proof is the same as for the case n = 2, given
above as equation (31). More generally, we have

(40)

for every permutation a in Sn. The special case of a transposition was proved above. The
general case follows on factoring a into a product of transpositions.

Proof of the Basic Theorem (Optional). Represent the column vectors a], a2, "', an
by means of the basis e], e2, ... , en:
n n n
(41) a] = l..: aill eil' a2 = l..:
i,=1
ai, 2ei 2' an = l..:ainnein.
il=1 in=l

Then if V is any function that satisfies (36)-(38) then substituting (41) into
V (aI, a2, ... , an) and using the linearity (36) gives (nn terms)
n n
l..: L ... L
n

V (a], a2,· .. , an) = ail]ai,2 ... ainn V (eil' ei2" .. ,ei.) .

il =1 h=l in=l

Moreover, the term V(ei, , ei" ... ,eiJ vanishes unless the vectors eil' ei" ... , ei n are
distinct; that is, unless Ci], i 2 , ... , in) is a permutation of (1,2, ... , n). The nonzero terms
correspond to the permutations a of (l, 2, ... , n), so

V (aI, a2, ... ,an) = L(J

aa(l) I aa(2)2 ... aa(n)n V (ea(l), ea(2), ... ,eu(n») ,

which by (40) equals

(42) L sgn(a)aa(J)la(J(2)2'" aa(n)n V (e], e2,"" en),

u

or since V (el, e2, ... , en) = 1 by (38),

(43) V (aI, a2,···, an) = L sgn (a) aa(l) 1 aa(2) 2 ... a(J(n)n'
a

This proves the uniqueness statement of the theorem. We have to show that any function
V (a], a2, ... ,an) that satisfies (36)-(38) must be given by (43). To prove the existence
statement of the theorem we shall show that if V is defined by (43) then V does indeed
have properties (36)-(38). Condition (36) is immediate because each term in the sum (43)
is linear in the components of each of the vectors a], a2, ... , an. Equation (38) also follows
334 5 LINEAR ALGEBRA

because only the identity permutation a(l) = 1, a(2) = 2, ... , a(n) = n places n unit
components of el, e2, ... ,en into the same product. Finally, to prove (37) we have by (43)

= L sgn(a)a (l)la (2)2 ...

r r ar(n)n,
r

where r in Sn differs from a by the transposition (1 2) (interchange of 1 and 2). Thus

sgn(a) = -sgn(r) and so

In particular, if al = a2 then D (aI, a2, ... ,an) = O. The proof for the other pairs of
vectors is similar. This completes the proof of the theorem.

Definition of det(A). For any n x n matrix A = (aij) the determinant of A is the

number

(44)

where aI, a2, ... , an are the column vectors of A and D is the unique function defined
by the basic theorem.
An explicit formula for det(A) is given by equation (43) above. For n = 1,2 and 3 it
is clear from (43) that definition (44) agrees with the definitions of elementary algebra;
see (l) and (13).
Basic Properties of Determinants. The determinant function has the following
useful properties.
Property 1. For all n x n matrices A one has

det(A T ) = det(A),
where AT is the transpose of A.
Property 2. For all n x n matrices A and B one has

det(AB) = det(A) det(B).

Note that det(AB) = det(BA) = det(A) det(B) even if AB i= BA.

Property 3. For all n x n matrices A one has

A is singular (rank(A) < n) if and only if det(A) = O.

 5.6 DETERMINANTS AND CRAMER'S RULE 335

Proof of Property 1. We shall use the product notation

n
k=l
n

Ck = CIC2·· ·C n·

Applying this notation to the products in equation (43) gives

a".(l)l aa(2)k ... a".(n)n = nn

k=l
a".(k)k·

Now, e = a(k) if and only if k = a-I (e) = Tce), where T is the unique inverse
permutation to a. As a runs through the n! permutations in Sn, its inverse T does the
same. Also, sgn(T) = sgn(a- l ) = sgn(a). Thus replacing a by T = a-I in (43) gives

det(A) = Lsgn(T)
r
n n

(=1
alr(l) = Lsgn(T)
r
nn

(=1
a;(t)l = det(A T ),

which completes the proof.

Proof of Property 2. If el, e2, ... , en are the standard basis vectors in Rn (the column
vectors of In), then the rules of matrix multiplication give al = Ael, a2 = Ae2, ... ,
an = Aen . Thus

where hj = Bej. The last function is linear in each hj separately and vanishes if two hjs
coincide. It follows, as in the proof of the basic theorem, that

where c(A) is a constant that is independent of B. Taking B = In gives detCA) = cCA),

which completes the proof.
Proof of Property 3. If rank(A) < n then the column vectors of A are linearly
dependent. Hence a multiple of some column vector is a linear combination of the
others. It follows from the basic theorem that det(A) = O. Conversely, if A is nonsingular
(rank (A) = n), then A-I exists and AA -1 = In. Then Property 2 implies that
detCA) det(A- l ) = det(In) = 1, and hence det(A) :j:. O.
Cofactor Expansions. The calculation of determinants can be Simplified by reducing
them to determinants of lower order. The starting point is the cofactor expansion.
Expansion of det(A) Along Column i. For any n x n determinant detCA), since detCA)
is a linear function of column vector ai, we may write

(45)
336 5 LINEAR ALGEBRA

where Ali, A2i , ... , Ani are suitable constants.

Expansion of det(A) Along Row i. Since detCA) = detCA T ), detCA) is a linear function
of row vector i of A, and we may write

(46)

for suitable constants Ail, Ai2 , ... , A in .

The constant Ajk is called the cofactor of ajk in detCA). Equation C45) is called the
cofactor expansion of detCA) along row i. It is not immediately obvious that the cofactors
in (45) and (46) are the same, but this will be verified below.
Minors and Cofactors. For n = 2 and 3 we have seen how to calculate the cofactors
Aij as determinants of order n - 1. We shall now extend this to all orders n 2: 2. To begin,
let Mij be the Cn - 1) x Cn - 1) matrix obtained by deleting row i and columnj of A.
Then the following theorem holds.

Theorem. For each n x n matrix A (n 2: 2) the cofactors of aij are given by

(47) Aij = (-1) i+"J det ( Mij ) .

Proof. The linearity of V (aI, a2, ... , an) with respect to al gives

(48)

L a kl V (eh,a2, ... ,an)

n

=
k=l

In particular, we have

1 al2 aln
0 an a2n
(49) All =V(el,a2, ... ,an) =

0 an2 ann
 5.6 DETERMINANTS AND CRAMER'S RULE 337

In this determinant, if the first column is multiplied by -al2 and added to the second
column, the value of the determinant is unchanged. This is because the amount added is

1 -al2 al n 1 1 aln
0 0 a2n 0 0 a2n
(50) = -al2 =0.

0 0 ann 0 0 ann

The effect on (49) of adding (50) is to replace al2 by zero (equation (49) is a linear
function of a2!)' Proceeding similarly, we may replace al3, a14, ... , al n by zero. Thus

1 0 0
0 a22 a2n
All =

0 a n2 ann

Now, this last determinant is clearly a linear function of each of the (n - I)-vectors

Moreover, if two of these vectors are equal, then All = O. Finally, if the vectors are
the standard (n - I)-dimensional basis vectors, then All = 1. Hence the uniqueness
statement of the basic theorem gives

All = = det (M ll ) ,

as was to be shown. A similar argument works for the other cofactors. For example,

0 al2 al n 0 al2 aln

1 a22 a2n 1 0 0
A21 = V (e2,a2,··· ,an) = =

0 a n2 ann 0 a n2 ann

1 0 0 al2 an aln
0 al2 al n a32 a33 a3n
= - det (Mu) .

0 a n2 ann a n2 a n3 ann
338 5 LINEAR ALGEBRA

Note that in (47) the factor (-l)i+j arises from moving the coefficient aij in det(A) from
the (i,j)-position to the (1, I)-position by i transpositions ofrows and j transpositions
of columns.

Cofactor and Adjoint Matrices of a Square Matrix. The cofactor matrix of an

n x n matrix A is the matrix cof(A) of its cofactors:

All An

cof(A) = ( A~i A~2

Ani An2

while the adjoint matrix of A is the matrix

All
An
(51) adj(A) = (cof(A))T = ( :
A in

The importance of adj(A) is due to the following theorem.

Theorem. For all n x n matrices A we have

(52) A adj(A) = det(A)In ,

where In = diag(1, 1, ... , I) is the n x n identity matrix.

Proof. The (i,j)-component of the matrix equation (52) is

(53)

To verify this, note that if i = j then (53) is just the cofactor expansion of det(A) along
the ith row. If i #- j then the left side of (53) is the cofactor expansion of a determinant
whose ith andjth rows are equal. This gives zero in accordance with (53).

Corollary. If A is nonsingular (so det(A) #- 0) then

(54) A- i = (det1(A)) adj(A).

Proof. This is immediate from (52).

Cramer's Rule. Equation (54) gives the n-dimensional generalization of Cramer's rule
for n = 2 and 3, given by (12) and (23) above. It states that if A is a nonsingular n x n
 5.6 DETERMINANTS AND CRAMER'S RULE 339

matrix then the unique solution of the system Ax = b is given by

(55) AljbI + A 2j b 2 + ... + Anjbn .
Xj= , J=1,2, ... ,n.
det(A)
This is immediate from (51) and (54). It can be written in the alternative form
det(A ) .
Xj = ---, j

det(A)
J = 1,2, ... ,n,
where Aj is the matrix formed from A by replacing the jth column of A by the column
vector b.
Evaluation of Determinants. How can we compute the value of an n x n
determinant? One answer is to use the combinatorial rule (43):

det(A) = Lsgn(a)a".(l)Ia".(2)2··· a".(n)n·

u

This sum contains n! terms each of which is a product of n terms, for a total of n! x n
multiplications. This count grows rapidly with n. Thus, for example,

1O! x 10=36,288,000.

This rapid growth makes the combinatorial formula impractical for all except very small
values of n. A second choice would be to use the cofactor expansions such as

= L aijAij
n

det(A) (i fixed).
j=l

However, when the cofactors are expanded into lower-order determinants, all n! x n
products are again generated.
A more efficient method is to use elementary row and column operations to transform
det(A) into a new determinant that has many zero entries. The use of the cofactor
expansion then becomes much simpler. The reduction can be based on the following
three rules.

Rule 1. If two rows (or columns) are interchanged then the determinant changes sign.
Rule 2. If a row (or column) is multiplied by a constant then the determinant is
multiplied by that constant.
Rule 3. If a scalar multiple of a row (or column) is added to another row (or column)
then the determinant is unchanged.

For row operations these rules were proved above. The corresponding column
operations follow from the row operations and the rule det(A T ) = det(A).
340 5 LINEAR ALGEBRA

EXAMPLE 1. The use of Rules 1-3 to simplify the computation of a determinant will be
illustrated by the following example:

8 0 0 4 8 0 0 4
9 1 -7 2 #1 9 1 0 -19
det(A) = -8 =
1 14 2 -8 1 0 44
0 0 1 -3 0 0 1 -3
8 0 4 8 0 4
#2 #3
9 1 -19 17 0 -63
-8 1 44 -8 1 44

8 4
~ (_1)2 ~ (-1)2«8)(-63) - (4)(7)) = -572.
17 -63

Here the first equation defines det(A). Step #1 subtracts (-7) times row 4 from row 2
and (14) times row 4 from row 3. Step #2 is the cofactor expansion along column 3.
Step #3 is to subtract row 3 from row 2. Step #4 is the cofactor expansion along column
2. Finally, Step #5 is the cofactor expansion of a 2 x 2 determinant.

Use of Gauss Reduction. A variant of the preceding method may be based on the
Gauss reduction algorithm. In matrix form the algorithm is

(56)

(see Section 5.4, equation (18)), where each Ej is a matrix that describes an elementary
row operation and U is an upper triangular matrix (all entries below the main diagonal are
zero). Taking the determinant of (56) and using the rule det(AB) = det(A) det(B) gives

Moreover, the terms det(Ej ) and det(U) can be found by inspection. Thus if Ej represents
the interchange of two rows then det(Ej ) = -1 (Rule 1); if Ej represents multiplication
of a row by a constant c then det(Ej ) = C (Rule 2); and if Ej represents addition
of a scalar multiple of a row to another row then det(Ej ) = 1 (Rule 3). Moreover,
det(U) = UllU22· .. Unn (product of diagonal elements) by cofactor expansion along the
first column. This is illustrated by application to the determinant of Example 1.
 5.6 DETERMINANTS AND CRAMER'S RULE 341

EXAMPLE 1 (by Gauss Reduction).

1
8 0 0 4 1 0 0 '2
9 1 -7 2 1 1 _Z 2
det(A) = ~ (8)(9)( -8) 9 9 9
-8 1 14 2 1 1 _Z 1
-8 4 -4
0 0 1 -3 0 0 1 -3
1
8 0 0 4 1 0 0 2
9 1 -7 2 0 1 -7 _Z
~ (-576)
-8 1 14 2
~ (-576) (~) (-~) 0 1 14 6
2

0 0 1 -3 0 0 1 -3
1 1
1 0 0 2
1 0 0 2

C:26)
7 7
#4 0 1 -7 -'2 0 1 -7 -'2
17
~ (8)(21) 17
0 0 21 "2 0 0 1 42
0 0 1 -3 0 0 1 -3
1
1 0 0 '2
0 1 -7 _Z
~ (168) 17
2
= (168) (143)
#7
- 42 = -572,
0 0 1 42
143
0 0 0 42

as before. The elementary row operations are as follows. Step # 1: E1 is multiplication of

the 1st row by 8, E2 is multiplication of the 2nd row by 9, E3 is multiplication of the 3rd
row by (-8). Step #2: E4 is subtraction of the 1st row from the 2nd row, Es is subtraction
of the 1st row from the 3rd row. Step #3: E6 is multiplication of the 2nd row by 1/9, E7 is
multiplication of the 3rd row by ( -1/8). Step #4: Es is subtraction of the 2nd row from the
3rd row. Step #5: Eg is multiplication of the 3rd row by 21. Step #6: ElO is subtraction of
the 3rd row from the 4th row. Step #7: Ell is multiplication ofthe 4th row by (-143/42)

I E~~~~i~.~~..~..~ .~
Matrices of order 2. Evaluate det(A) for the following matrices.

1.
( ~ ~)- 4. ( -~ -~ )- 7.
(~ -b)o .
2.
(~ o . -1 ) 5.
(~ ~). 8. ( -A
-1 4 ~ A ).
3.
(~ ~). 6.
(~ -: ). 9.
(~ -~ ).
Matrices of order 3. Evaluate det(A) for the following matrices.
342 5 LINEAR ALGEBRA

10.
(~ ~ ) I
5
5
13.
(r
4
2
2 r)
16.
(~ 0
5
-I)
11.
(~ ~ ) 3
5
5
14.
(~ -~ )0
I
0
17.
(~ -~ )
2
I
0

12.
(~ ~ ) I
0
S
15.
(~ -I)
0
0 18.
(I -~ )
2

Matrices of order 4. Evaluate the determinant of the given matrix and cite the determinant rules used.

(" lL)
n
0 0 0
o b

U
0 12 c 5
19. 21.
I 2 I~ . 0 b
o 0 0 0 0

n un
I

(
a 0 0 6
-I b 0 b
20. 22.
0 -I c c 4
0 0 -I 0 0

Sarrus's Rule for 3 x 3 Determinants.

The rule proceeds by writing the first two rows of A below the matrix. Then the determinant is formed
from the six indicated products, three positive and three negative. The pattern of products in Figure 5 is
det(A) = (a + b + c) - (d + e + j). This agrees with the classical sum definition:

Warning. While Sarrus's rule exists for 2 x 2 and 3 x 3 determinants, there is no rule for 4 x 4 and
higher! In the exercises below, apply Sarrus's rule for n = 3 to evaluate the given determinant.

c FIGURE 5. Sarrus's rule

 5.6 DETERMINANTS AND CRAMER'S RULE 343

23.
(~ 0
5
-1 )- 26.
A=(~ ~)-
2

24.
(~ 3

0 -1
~)- 27.
A= (1 ~)-
(! A= (l :)
2

~)-
4
0
25. 1 28.
-1
-1 -1
Cramer's Rule. Solve for x
in the system Ax = b by Cramer's rule.

29.
(~ =~)x=(~). 31.
(~ 1
5
5
~ )x= (l~)-
30. (_~ ~)x=(~). 32.
(~ 5
~)X=(j)-
(-;
5
0 0

}=m
0
33.
1 2
0 -2

}{;)
0 0

(-!
0
34.
5
0 -5
35. (Definition of DeterminanO Using just the definition of determinant, evaluate

all all al3 al4 a15

an an a23 a24 a25
a31 a32 0 0 0
a41 ~2 0 0 0
a51 a52 0 0 0
36. (Determinant Equations) Solve for x in the determinant equation

2-x
1 3-x =0.

n-x
37. (Determinant Identity) Show by reduction to triangular form that

222 2
233 3 = 1.

234 ... n
344 5 LINEAR ALGEBRA

38. (Number Theory) Use Cramer's rule to prove the following divisibility result: If three numbers each
with three digits are each divisible by an integer m, then the determinant formed from their digits is
divisible by m. For example, the numbers 104, 143, 195 each have three digits and are divisible by 13,
and therefore the determinant

is divisible by 13.
Applications to Geometry. Establish the follOwing results.
39. (Area of a Triangle) Let a triangle have vertices (Xl ,Yl), (X2,Y2), (X3,Y3). Prove that the area A of the
triangle is given by the determinant equation

1 ( Xl X2 X3)
A = - det Yl Y2 Y3 .
2 1 1 1

40. (Equation of a Line) Given a line passing through two distinct points (Xl,Yl), (X2,Y2), prove that the
equation of the line is

XXI X2)
det ( Y Yl Y2 = 0.
1 1 1

41. (Equation of a Plane) Given a plane passing through three points (Xl ,Yl, Zl), (X2,Y2, Z2), (X3,Y3, Z3),
prove that the equation of the plane is

XXI x2 X3)
det ( Y Yl Y2 Y3 = 0.
Z Zl Z2 Z3
1 1 1 1

42. (Volume of a Parallelepiped) The volume of the parallelepiped in space whose concurrent edges are
the vectors

is known to equal the scalar triple product X . (Y x Z). Prove that the volume V is also given by the
determinant equation

Xl x2 X3)
V = det ( Yl Y2 Y3 .
Zl Z2 Z3

Geometric Calculations. Find the given area, volume, or equation.

43. Area of the triangle with vertices at (1, -1), (2,1), (-1, -1).
44. Equation of the line through (1, -1) and (- 2, -1).
45. Equation of the plane through the points (1,0,0), (0, -4,0), (0,0,2).
 5.7 EIGENVALUES AND EIGENVECTORS 345

m m

FIGURE 6. Coupled harmonic oscillators

46. Volume of the tetrahedron with vertices at

(1,1,1), (-1,1,1), (1,-1,1), (1,1,-1).

Hint: The answer is a certain fraction of the volume of the associated parallelepiped.
47. Volume of a parallelepiped whose concurrent edge vectors are

Theoretical Results. Establish the following.

48. (Product Rule) Given A = BC and det(B) = det(C) = 4, then does an equation of the form Ax = b
have a unique solution? Explain.
49. (Product Rule) Let A, B, C be 7 x 7 matrices such that det(A) = 2, det(B) = 6, and B = A2 e.
Determine the value of det(ABC).
50. (Product Rule) GivenA = BC, det(B) = 4, and det(C) = 0, then does an equation oftheformAx = b
have a unique solution? Explain.
Sl. (Product Rule) Given A = B2 C and det(A) = -5, then does CB have an inverse? Explain.

5.7 Eigenvalues and Eigenvectors

A number A is said to be an eigenvalue of an n x n matrix A if there is a vector x such that

(1) Ax = AX and x =I o.
Any vector x that satisfies (1) is said to be an eigenvector for the eigenvalue A. The
problem of determining the eigenvalues and eigenvectors of a given square matrix A is
called the eigenvalue problem for A. The pair x, A is called an eigenpair of (1).
Eigenvalues often appear as the frequencies of vibration of mechanical oscillators.
This will be illustrated by the case of a pair of coupled harmonic oscillators.

ExAMPLE 1. Consider the oscillator shown in Figure 6.

346 5 LINEAR ALGEBRA

Two bodies of mass m slide on a smooth plane and are coupled by springs with
spring constants k. The displacements Xl and Xl are measured from their equilibrium
positions (vertical dashed lines) and are counted as positive when to the right. With these
conventions, Hooke's law and Newton's laws of motion give the following differential
equations (where primes denote time derivatives).
Equations of Motion.

(2) mx~ = - kxl + k(X2 - Xl) = k( -2xl + X2),

mx~ =- k(X2 - Xl) - kx2 = k(Xl - 2X2).

Normal Modes of Vibration. To find the possible modes of vibration let us look for
solutions of the special form

(3)

where Cl and C2 are constants, not both zero. Such solutions are called synchronous
vibrations or normal modes of vibration. Substituting the functions (3) in equations
(2) gives

(4) mT" (t)Cl = kT(t) ( - 2Cl + C2),

mT"(t)c2 = kTCt)(Cl - 2C2).

If TCt), Cl, and C2 are not zero, the variables can be separated to give the equations

(5) T"(t) = ~ (- 2C l + C2) = ~ (Cl - 2C2) = -fL,

TCt) m Cl m C2

where fL is a constant. In particular,

T"(t) + fLTCt) = 0,

and hence

TCt) = a cos(J!it) + b sin(J!it),

fL
provided that > O. Thus the method does predict oscillatory solutions.
Eigenvalue Problem. To complete the construction of the normal modes, we must
determine fL,
Cl and C2 as solutions of (5). These equations can be written as

2Cl - C2 = (fL;) Cl,

-Cl + 2C2 = (fL;) C2,
 5.7 EIGENVALUES AND EIGENVECTORS 347

or Ac = AC, where

(6) A= ( -12 -1),

2
C = (C1)
~
, m
A = kf-L.

Eigenvalues. The equation Ac = AC can be rewritten as

(A - AI)c = 0.

A nontrivial solution c is sought. This is possible if and only if A - AI is singular, or

det(A - AI) = 0. Thus the eigenvalues are solutions of the equation

2 -A -1
det(A - AI) =
-1 2- A
= (A - 2)2 - 1 = (A - l)(A - 3) = 0.
Hence there are two eigenvalues,

and
k k k k
(7) f-L1 = -AI = -,
m m
f-L2 = -A2
m
= 3-.
m
Eigenvectors. The eigenvectors for eigenvalue Al = 1 are the nontrivial solutions of

Clearly, the eigenvectors form a l-dimensional subspace spanned by

Similarly, the eigenvectors for A2 = 3 satisfy

-C2 )
-C2
= (0).
0

Thus the eigenvectors for A2 form a l-dimensional subspace spanned by

Normal Modes of Vibration. The pairs (AI, C1) and (A2, C2) give two normal modes:
348 5 LINEAR ALGEBRA

Mode 1 Xl (t) = X2 CO = Al COS(wt) + Bl sin(wt),

Mode 2 Xl (t) = -X2 (t) = A2 cos( J3wt) + B2 sine J3wt) ,
where by (7),

Recall that w is the frequency of the simple harmonic oscillator with mass m and Hookes
constant k. Mode 1, with the lower frequency, is called the fundamental mode. Note
that for it Xl (t) = X2 (t). Thus the middle spring is not stretched and the two masses
move like a simple oscillator with mass 2m and spring constant 2k. It can be shown that
the general solution of equations of motion (2) is just the sum of modes 1 and 2. Thus

(8) Xl (t) = Al cos(wt) + Bl sin(wt) + A2 cos( J3wt) + B2 sine J3wt) ,

X2(t) = Al cos(wt) + Bl sin(wt) - A2 cos( J3wt) - B2 sin(J3wt),

where At, Bl , A 2 , B2 are arbitrary constants.

Eigenvalues of n x n Matrices. As in the example, a number A is an eigenvalue of
an n x n matrix A if and only if

(9) (A - AIn) x = 0 for some x i= O.

This in turn, by Property 3 of determinants in Section 5.6, is equivalent to the condition

(0) det (A - AIn) = O.

The Characteristic Polynomial. It is easy to check that det (A - AIn) is a polyno-
mial function of A of exact degree n. It is called the characteristic polynomial of the
matrix A, and equation (10) is called the characteristic equation of A.
Necessity of Using Complex Arithmetic. An n x n matrix A may have no real
eigenvalues even if the components of A are all real. A simple example is the 2 x 2 matrix

A= (0 -1)
10'

with characteristic polynomial A2 + 1. The matrix A has eigenvalues Al = i, A2 = -i

T T
and eigenvectors Xl = (i, 1) ,X2 = (-i, 1) . Thus in order to find all eigenvalues and
eigenvectors it is essential to work in the vector space en.
The Existence of Eigenvalues. Gauss' fundamental theorem of algebra states that
every polynomial with complex coefficients and degree n ::: 1 has at least one root (real or
 5.7 EIGENVALUES AND EIGENVECTORS 349

complex). Applying this to the characteristic polynomial, we see that every n x n matrix
A has at least one eigenvalue A]. The corresponding set of eigenvectors is the eigenspace
N (A - A]In) with dimension 1 or more. This is illustrated by Example 1 in Section 5.6.
EXAMPLE 2. The 5 x 5 matrix

0 1 0 0 0
0 0 1 0 0
A= 0 0 0 1 0
0 0 0 0 1
0 0 0 0 0

has characteristic polynomial det(A - AI) =

(_A)5. Hence A] =
0 is the unique
T
eigenvalue. Moreover, the eigenspace N(A) has dimension 1 and basis (1,0,0,0,0) .
More generally, for any n 2: 2, the n x n matrix with l's on the diagonal above the main
diagonal and zeros elsewhere has det(A - AI) = (_A)n and a I-dimensional eigenspace
T
spanned by (1,0, ... ,0) . The simple proofs of these statements are left as exercises.

If the characteristic polynomial has several distinct roots then each root will be an
eigenvalue with at least one corresponding eigenvector. More precisely, the following
theorem holds.

Theorem. Let the n x n matrix A have k distinct eigenvalues A], ... , Ak (so 1 ::s k ::s n).
Then the characteristic polynomial has the factorization

(ll)

where m], ... , mk are positive integers whose sum is n. Moreover, the eigenspaces

(12)

satisfy

(13) 1 ::s dim (Aj) ::s mj-

The exponent mj is called the algebraic multiplicity of eigenvalue Aj, while dim (Aj)
is called its geometric multiplicity. It can be shown by simple examples that dim (Aj)
can be any integer between 1 and mj.
Linear Independence of Eigenvectors. If a matrix A has eigenpairs (A], c]) and
(A2, C2) with A] -# A2 then c] and C2 are linearly independent. For if they were linearly
350 5 LINEAR ALGEBRA

dependent then there would be a nonzero scalar a such that

Also, applying A to the last equation gives

If Al i- 0 then the last two equations imply that

CI = -aC2 = - (aA2/AI) C2,

whence

and so Al = A2 (since a i- 0). The same conclusion holds if A2 =1= O. This contradiction
shows that CI and C2 are linearly independent. The result is extended to more than two
eigenvectors by the following theorem.

Theorem. Let the n x n matrix A have distinct eigenvalues AI, A2, ... ,Ak (so 1 ::: k ::: n)
and corresponding eigenvectors CI, C2, ... , Ck. Then the vectors CI, C2, ... , Ck are linearly
independent.

Proof. We must show that if a linear relation

(14)

holds then the coefficients ai, a2, ... , ak must all be zero. The proof will be by induction
on k.

Case k = 1. In this case aici = 0 and since CI i- 0, (since it is an eigenvector), we have

a] = O.
Case k > 1. We know that ACj = AjCj (j = 1,2, ... , k) and the eigenvalues Aj are
distinct. Multiplying (14) on the left by A gives

(15)

Now subtract Ak times (14) from (15). The last terms cancel and we have

(16)

But the inductive hypothesiS implies that the coefficients in (16) are all zero. Since the
eigenvalues are distinct, it follows that a] = a2 = ... = ak-I = O. Then, returning to
(14), we have akCk = 0 and hence ak = O. This completes the proof.
 5.7 EIGENVALUES AND EIGENVECTORS 351

Corollary. If an n x n matrix A has n distinct eigenvalues then the corresponding

eigenvectors form a basis for Rn (or C).

This result is immediate because n linearly independent vectors in an n-dimensional

vector space always form a basis.
Change of Basis and Similarity. Let A be an n x n matrix and consider the linear
system Ax = h. The solution vector x has the expansion
n
(17) X= (XI,X2, ... ,Xn) = LXkek,
k=l

where {el, e2, ... , en} are the standard basis vectors, defined by

el = (l,o, ... ,of, e2 = (o,l, ... ,of, etc.

It is sometimes possible to simplify the matrix A by introducing a second basis. We shall

show that the search for a better basis leads to the eigenvectors of A.
Let <PI' <P2' ... , <Pn be a second basis for Rn (or Cn) and write
n n
(18) X = LXkeh = LX~<Pe.
k=l e=l
Moreover, define the matrix <l> = (<pjk) by
n
(19) <Pc =L ¢kCek, .e = 1,2, ... , n.
h=l

Note that the order of the indices k, .e is not the same as for a matrix product. Equations
(19) imply that

<PI = (:::), <P2 = (:) , etc.

<Pnl <Pn2
Thus the matrix <l> may be written as

(20)

where the notation means that <l> is the matrix whose first column is <Pl, second column
<P2, etc. Clearly, <l> is nonsingular because the vectors <PI' <P2, ... , <Pn are a basis and
hence are linearly independent.
352 5 liNEAR ALGEBRA

Hence, matching the coefficients of ek gives

n
Xk = L 1>kiX~,
i=l

or in matrix notation,

(21) x = <l>x'.

Similarly,

(22) b = <l>b' ,

where b' is the unique vector defined by b ' = <I>-lb. Substituting (21) and (22) into
Ax = b gives

A<I>x ' = <l>b' , or <I>-lA<I>x' = b'.

Hence if we define A' by

(23)

then

(24) Ax =b if and only if A'x' = b',

where x' and b ' are defined by (21) and (22). Thus to solve Ax = b, it is enough to solve
A'x' = b', where b' = <I>-lb, and then compute x = <l>xl
Matrices A and A' are said to be similar if there is a nonsingular matrix <I> such
that (23) holds. One important property of similar matrices is that they have the same
eigenvalues. More precisely, (A, x) is an eigenpair for A if and only if (A, x') is an eigenpair
for A', where x = <l>x'. To see this, note that

¢=::} A (<I>x') = A ( <l>x')

¢=::} Ax = AX.

Moreover, x -I 0 if and only if x' -I 0 because <I> is nonsingular. Hence A is an eigenvalue

of A' if and only if it is an eigenvalue of A.
 5.7 EIGENVALUES AND EIGENVECTORS 353

Diagonalization. The solution of a linear system Ax = b is expedited if A is similar to

a matrix A' that has simple structure. The most favorable case occurs when A' is diagonal:

The eigenpairs of A are readily seen to be (Aj, ej), where el, e2, ... , en is the standard
basis. Moreover, A is invertible if and only if the eigenvalues are all nonzero. In this case
A -I
1\ = d'lag ('11.1-I ,11.2
,-I ,-I)
, ... , n
II. .

Suppose that A is similar to a diagonal matrix, so

(25) C1AC = A, or AC = CA.

We shall say that A is diagonalizable when (25) holds. If we write

(26) C= (CI C2 ... Cn ) ,

where CI, C2, ... , C n are the columns of C, then the rules of matrix multiplication imply
that

AC = (AcI AC2'" AC n)
=CA
= (AICI A2C2'" AnCn ).

Equating column vectors gives

AC1 =AI C1,

AC2 = A2 C2,

Moreover, each Cj =f:. 0 because C is nonsingular. Hence the column vectors of C are all
eigenvectors of A. Also, they are linearly independent, again because C is assumed to be
nonsingular. Conversely, if A has n linearly independent eigenvectors CI, C2, ... , Cn then
C, defined by (26), is nonsingular, and (25) holds. Hence the following theorem is valid.

Theorem. An n x n matrix A is diagonalizable if and only if it has n linearly independent

eigenvectors.
354 5 LINEAR ALGEBRA

EXAMPLE 3. (Real Eigenvalues and Eigenvectors). Let

Then

det (A - Ah) = (A - l)(A - 3),

so
T
Cl = ( 1, -1 ) ,
T
C2 = ( 1,1 ) ,

hence

Cramer's rule gives

c 1 _ adj
- detC - 2
C_~ (1 -1) 1 1

and

EXAMPLE 4. (Complex Eigenvalues and Eigenvectors). Let

A-( 01)
- -1 0 .

Then

det (A - Ah) = A2 + 1 = (A - O(A + 0,

so

Al = i, Cl = (1, if,
A2 = - i, C2 = (l,_i)T,
 5.7 EIGENVALUES AND EIGENVECTORS 355

hence

Again, by Cramer's rule

c- 1 = -1 (-i -1) = ~ (1 -i)

2i -i 1 2 1 i

and

C-'AC~(~_~).
ExAMPLE 5. (Matrix Not Diagonalizable). Let

A=(~ ~).
Then

so

Moreover, the eigenspace is one-dimensional, spanned by

Cl = (l,of·
Thus A is not diagonalizable.

Diagonalization of Self-Adjoint Matrices. An n x n matrix A, real or complex,

is said to be self-adjoint if it is equal to its adjoint A *. This means that A * = A, or
-T
equivalently, A = A. In particular, if A is real, then it is self-adjoint if and only if
AT = A. Such matrices occur frequently in mechanics (theory of vibrations) and many
other areas of applied science. Here it will be shown that self-adjoint matrices are always
diagonalizable. More precisely, the following theorem holds.

Theorem. Let n be a positive integer and let A denote an n x n self-adjoint matrix, real
or complex. Then A has the following properties:

(27) The eigenvalues of A are all real.

356 5 LINEAR ALGEBRA

(28) Eigenvectors of A with distinct eigenvalues are orthogonal.

(29) A has n orthonormal eigenvectors.

(30) A is diagonalizable.

Property (30) follows from (29) by the preceding theorem. We shall give proofs of (27)-
(29). Only (29) is difficult. It will be proved by induction on the order n, starting with
the simple case n = 1.

Proof of (27). Recall that every n x n matrix A satisfies (Ax, y) = (x, A*y) for all
vectors x and y. In particular, self-adjoint matrices A satisfy (Ax, y) = (x, Ay) for all
vectors x and y in en. In this equation take x = y, where Ax = AX for some (possibly
complex) A and some x =f. O. This gives (Ax, x) = (x, Ax), or A(X, x) = I(x, x), by
the properties of the inner product in en.
This can be written as

since (x, x) = IIx1l2. Since x =f. 0, it follows that A = I, which implies that A is real.

proof of(28). Let (A, x) and (A', y) be eigenpairs for A with distinct eigenvalues Aand A'.
Substituting Ax = Ax and Ay = A'y into (Ax, y) = (x, Ay) gives (Ax, y) = (x, A'y),
or, since A and A' are real,

(A - A') (x, y) = o.

Since A =f. A', it follows that (x, y) = 0 and x and yare orthogonal.

Proof of (29). We proceed by induction on n ~ 1. The result (29) is true when n = 1

(1 x 1 matrices are just scalars). To complete the proof it is enough to show that if the
result is true for (n - 1) x (n - 1) matrices (n ~ 2) then it is also true for n x n matrices.

To begin, recall that every n x n matrix has at least one eigenpair (A I, Xl), so

Next construct an orthonormal basis {YI, Y2, ... , Yn} such that YI = Xl. This can always
be done by the Gram-Schmidt method. Then define Q to be the n x n matrix whose
column vectors are YI, Y2, ... , Yn' The notation

(31)
 5.7 EIGENVALUES AND EIGENVECTORS 357

will be used. Clearly, Q*Q = ((yj , Yk)) = In because (Yj , Yk) = Ojk. The same is true for
QQ*, so

Q*Q = QQ* = In

and Q* = Q-l. Such matrices are said to be orthogonal matrices. It follows from (31) that

and hence

Here

while

for k ::: 2. Thus the matrix A' = Q* AQ has the form

o o
bl,l b1,n-l

bn-1,n-l

where B = (bjk ) is an en - 1) x en - 1) matrix. Moreover,

bjk = (Yj+1 , AYk+l)
= (AYj+l , Yk+l)
= (Yk+l , AYj+l) = bkj .

Thus B is a self-adjoint (n - 1) x (n - 1) matrix. By the inductive hypothesis, B has an

orthonormal basis of eigenvectors in Rn - l (or en-I). Denote it by t\, £'2, ... , £'n-I, so
358 5 LINEAR ALGEBRA

Then the n vectors

(32)

are orthonormal eigenvectors of A'. To see this, note that

while for k = 2, ... , n,

The vectors Zl, Z2, ... , Zn are orthonormal in Rn (or en) because the vectors iI, i 2 , ... ,
i n - l are orthonormal in Rn- l (or en-I). Finally, the vectors
Xj == QZj, j = 1, 2, ... , n,

are orthonormal eigenvectors of A because

whence

Moreover,

(QX, Qy) = (x, Q*Qy) = (x, y)

because Q is an orthogonal matrix. It follows that

(Xj , Xk) = (QZj , QZk) = (Zj, Zk) = 0jk,

and the eigenvectors Xl, X2, ... , Xn are orthonormal. This completes the proof of (29).
Calculation of Eigenvalues and Eigenvectors. It is important for many appli-
cations to have methods for computing eigenvalues and eigenvectors of matrices. In
 5.7 EIGENVALUES AND EIGENVECTORS 359

principle this is straightforward. If A is an n x n matrix, one first determines the eigen-

values 1'1, A2, ... , Ak (1 ~ k ~ n) as the distinct roots of the characteristic polynomial
det (A - AIn). Then for each eigenvalue Aj a basis of the eigenspace N (A - Ajln) is
computed by using Gauss elimination to solve the system (A - A}n) x = O.
For small values of n (n ~ 4) the computations can be carried out by using
classical methods for solving quadratic, cubic, and quartic equations. For larger matrices
it is essential to use machine computation. Computer algebra systems can be used to
solve eigenvalue problems of moderate size. For the large matrices that arise in many
applications, one needs sophisticated numerical methods. Such methods have undergone
intensive development in recent decades. An excellent survey, with FORTRAN programs,
can be found in the text Numerical Recipes; see [Prj.

Exercises 5.7

Eigenvalues. The problem of finding eigenvalues is reduced to factorization of a polynomial of degree n. In

low dimensions n = 2 and n = 3, the problem can be solved by hand calculations. For higher dimensions,
a computer algebra system will be helpful. In each exercise below find the characteristic polynomial of A,
factor it completely, and determine the eigenvalues.

1. A=(_~ ~) 8. A = ( i 0
-1 -~ )-
-1
2. A = ( ~ ~) -5

3. A = ( ~ ~) 9. A = ( ~ 4
0 ~)-
~) ~ ~)-
4. A= ( -25
10. A = ( 0
0
5. A = ( _~ -! ) 0

~
2 ~ ~)-
~)-
11. A = ( 0
6. A = ( 1 0
0 -1
0

7. A = ( ~
2
4
-1 )
-2 . 12. A = ( ~ 4
0 ~)-
6 -3

Eigenvectors. The problem of finding eigenvectors reduces to the computation of at most n reduced row
echelon forms. In each exercise below determine all the eigenvectors. A computer algebra system will be
helpful.
360 5 LINEAR ALGEBRA

13. A = (~ ~) 20. A= (90 -54 3)

3 .
o 0 1
14. A = ( ~ ~)
15. A = ( -21 ~)
21.
A=(~ 0
0 !)
-! ) )
0 0 1
16. A = ( _~ 22. A = ( 0 0 0
0 0 0

)
2
17.A=(~
!)
2

23A~U
0 0
0 -1 2 0

)
0

n
2 -1
18. A = ( ~ 4
6
-2
-3
0 0 -1

A~U
0 0

-~ )
3 0

~
0
24 0 2
19. -1
A= ( 0 0

25. (Multiplicity) Let A be a 7 x 7 real matrix with eigenvalue A of multiplicity 7. Show that the geometric
multiplicity of A is 7 if and only if A = diag(A, ... , A).

26. (Characteristic Polynomial) Prove by means of cofactor expansion that the determinant of A - Un is
a polynomial of degree n with highest order term (_A)n and constant term det(A).

27. (Eigenspace Dimension) Produce five matrices A of size 5 x 5 with all eigenvalues zero and eigenspace
N(A) of dimensions 1,2,3,4 and 5. This shows that if the algebraic multiplicity of an eigenvalue is 5
then the geometric multiplicity can be any integer from 1 to 5.

28. (Similarity) Let A and B be similar matrices. Prove that det(A) = det(B).

29. (Similarity) Do there exist two matrices A and B with the same characteristic polynomial, but A and B
are not similar matrices?

30. (Similarity) Let A and B be similar matrices. Show that they have exactly the same characteristic
polynomial.

31. (Similarity) Let the n x n matrix A be similar to a diagonal matrix D, D = 4>-1 A4>. Show that the
columns of 4> are a basis for Rn It is not assumed that the eigenvalues of A are distinct; e.g., A could be
the n x n identity matrix.

32. (Similarity) Let the n x n matrix A be similar to a diagonal matrix D, D = 4>-1 A4>. Show that any
matrix B similar to A satisfies an equation D = III-I BIII for exactly the same diagonal matrix D.

33. (Eigenvecctors) Let x = (3,1, -1) and y = (1, -1, 1). Define A = xTy (matrix product). Find the
eigenvectors of A.

34. (Trace) Let x and y be two nonzero column vectors in Rn and define A = x T y (matrix product). Show
that the trace of A is an eigenvalue and that all the other eigenvalues are zero.
 5.7 EIGENVALUES AND EIGENVECTORS 361

35. (Separation of Variables) Show that in Example I.

mT"(t)q = kTCt)(-2q + (2),

mT" (t)Q =
kT(t)(q - 2(2)

implies

T" =!!.. (-2 q+(2) =!!.. (q - 2(2) = -11.

T m q m Q

36. (Normal Modes) Show directly that the normal modes of vibration are

Xl = x2 = Al cos(wt) + Bl sinew!)

and

Xl = -X2 = A2 cos(J3wt) + B2 sin(J3wt),

are solutions of the equations of motion

mx~ =
h( -2Xl + X2),
mx~ = h(XI - 2X2)

where w = ,/Wm.
37. (Eigenvector) Establish the formula C2 = 0, -1/ for the second eigenvector in Example 1.
Characteristic Polynomial. Theoretical considerations show that the characteristic polynomial of an n x n
matrix A has n roots in the complex field. The practical problem of determining the polynomial coefficients
is considered in the exercises below.

38. Let A be a 2 x 2 matrix with real coefficients aij. Show that the characteristic polynomial of A can be
written in the form

(_A)2 + trace(A)( -A) + det(A),

where traee(A) = all + a22 is the sum of the diagonal elements of A.

39. Calculate the characteristic polynomial of A = ( _ ~ ~) .

40. Let A be a 3 x 3 matrix with real coefficients aij. Show that the characteristic polynomial of A can be
written in the form

(_A)3 + traee(A) ( _A)2 + (~Mii) (-A) + det(A),

where traee(A) = all + a22 + a33 is the sum of the diagonal elements of A and Mii is the minor of
element aii, 1 :::: i :::: 3.

41. Calculate the characteristic polynomial of A = (~

3
! =~ ).
6 -3
42. Let A be 3 x 3 and upper triangular. Find a formula for the characteristic polynomial of A in terms of
the diagonal entries of A.
362 5 liNEAR ALGEBRA

43. Calculate the characteristic polynomial of A = (~~ ~)

o 0 -1

Companion Matrix. Let xn = an_1Xn- 1 + ... + alx + ao be a polynomial of degree n ::: 1. The matrix
C below is called the companion matrix.

o o o
o o

o )
The five exercises below develop properties of the companion matrix.

44. Describe in words how to obtain the companion matrix from the n x n identity matrix.

45. Show that xn = an_1X n- 1 + ... + al x + aO is the characteristic equation of the companion matrix for
the special cases n = 2 and n = 3.

46. Choose the third row of the matrix A below such that its characteristic polynomial is equal to -A 3 +
3A 2 - 5A + 2.

47. Find a 2 x 2 matrix A that is not similar to any companion matrix.

48. Let t = A be a root of tn = an-l tn- 1 + ... + al t + ao. Show that A is an eigenvalue of the companion
matrix C with corresponding eigenvector

49. (Eigenspaces) Prove that the eigenvectors of an n x n matrix A, corresponding to a given eigenvalue
A, form a subspace of Rn

50. (Column space) Let A be a nonzero eigenvalue of the n x n matrix A. Show that each eigenvector of A
corresponding to A belongs to the column space of A.

51. (Column Space) Let the 3 x 3 matrix A have the property that its three column vectors add to (7,7,7) T
Prove that 7 is an eigenvalue of A.

52. (Eigenvalues of Matrix Products) Let A be an eigenvalue of the matrix product AB of two real n x n
matrices A and B. Prove that A is an eigenvalue of the product BA.

53. (Eigenvalue of the Inverse Matrix) Let A be a nonzero eigenvalue of the n x n nonsingular matrix A.
Show that l/A is an eigenvalue of A -1 .

54. (Eigenvectors of Powers of A) Let A be an ei~envalue of the n x n matrix A with corresponding

eigenvector x. Show that x is an eigenvector of A with eigenvalue Ak, for k = 1, 2, 3, ....
 5.7 EIGENVALUES AND EIGENVECTORS 363

55. (Eigenvectors) Let A be an eigenvalue of the n x n matrix A with corresponding eigenvector x. Show
that Akx is an eigenvector of A corresponding to A, provided that Akx # 0, for k = 1, 2, 3, ..
56. (Eigenvectors of Similar Matrices) Let A be an eigenvalue of the n x n matrix A with corresponding
eigenvector x. Assume that B is similar to A, B = <1>-1 A<I>. Prove that y = <I>-lx is an eigenvector of
B corresponding to eigenvalue A.
CHAPTER 6
 Vector Analysis

Vector analysis is primarily the invention of josiah Willard Gibbs (1839-1903, American
mathematician and physicist). His work on vector analysis first became known through
his widely circulated Yale lecture notes from the years 1881-84. The first definitive
textbook, entitled Vector Analysis, was published in 1901 by Edwin Bidwell Wilson-
a follower of Gibb's Yale lectures. The notation and methods of "Gibbs-Wilson" have
found almost universal use throughout most of the twentieth century
Vector analysis is the theory of geometric vectors in Euclidean space-including
vector algebra and vector differential and integral calculus. The theory has applications
to many problems of geometry and mathematical physics. Some of these are developed
in this and subsequent chapters.
Most first-year calculus textbooks include an introduction to vector analysis. More-
over, science and engineering students are also introduced to concepts of vector analysis
in their introductory physics and engineering courses. For these reasons the present text
will review the basic concepts only briefly before moving on to the main theorems and
applications.
366 6 VECTOR ANALYSIS

6.1 Vector Algebra

The simplest setting for the theory of geometrical vectors is the Euclidean plane E2 of
high-school plane geometry. This idealized surface is suggested by such physical examples
as the large, but finite, plane surface of the ice in a skating rink or the plane of the
ecliptic, defined by the earths orbit around the sun. We shall assume that the classical
constructions and theorems of Euclidean plane geometry hold in E 2.
Displacements. The Euclidean plane E2 is homogeneous, by which we mean that
each point of E2 looks exactly like every other. Thus if E2 is translated parallel to itself
through a fixed distance, then it looks the same as before. Such a displacement is
determined when we mark a single point P before the motion and its location Qafter the
motion. To show this we note that if p' is any other point of E 2, then its location after
the motion is the unique point Q' such that the quadrilateral P'PQQ' is a parallelogram;
see Figure l.
Directed Line Segments. A displacement in E2 may be specified by a directed line
segment, by which we mean a finite line segment in E2 with one endpOint specified as
the initial point and the other endpoint Qspecified as the terminal point. The notation
---+ ---+ ----+
PQ is used for this directed line segment. Directed line segments PQ and p' Q' define the
same displacement of E2 if and only if P'PQQ' is a parallelogram, as in Figure 1. In this
---+ ----+
case we shall say that PQ and P'Q' are equivalent. Note that equivalent directed line
segments generate the same displacement of E 2.
Vectors. The equivalence relation defined above divides the set of directed line
segments into distinct equivalence classes-one for each distinct displacement. Each
such class defines a unique displacement of E 2. It is these classes that we shall call
the vectors of E 2. This definition formalizes the idea that a vector has a direction and
magnitude (= the length of the vector) but no fixed point of application.

Q'

p FIGURE 1. Parallelogram construction of (/

 6.1 VECTOR ALGEBRA 367

Notation. Vectors will be denoted by boldface letters A, B, etc., if in print and by

-+ -+
letters with an over-arrow A , B ,etc., if writing by hand. With this notation we can say
that vectors A and B are equal, A = B, if and only if each directed line segment in class
A is equivalent to each directed line segment in class B.
Addition of Vectors. When two successive displacements of the plane E2 are made,
the result is equivalent to a single displacement. To show this we let the first and second
displacements be described by vectors A and B that are generated by directed line
-+ -+
segments PQ and QR, respectively Then the composite displacement is defined by the
directed line segment PR and the corresponding vector C; see Figure 2.
It is natural to call C the vector sum of A and B and to write

C = A+B.
Multiplication of Vectors by Scalars. In vector analysis a scalar quantity; or scalar,
is a number of the real number field. The product of a vector A and a scalar a is a vector,
denoted by aA, which is defined as follows:
Case 1 (a> 0). In this case aA is the vector parallel to A whose magnitude (= length)
is the product of a and the magnitude of A.
Case 2 (a = 0). In this case aA = OA = 0, where 0 is the zero vector that corresponds
to the trivial displacement pP (no motion).
Case 3 (a < 0). In this case aA is the vector with direction opposite to that of A and
magnitude equal to lal times that of A.
The algebraic properties of the set of geometric vectors in E2 are summarized in the
following theorem.

Theorem. The set of geometric vectors in E2 (and the corresponding set of all
displacements of E2 ) is closed under the operations of vector addition and scalar

p FIGURE 2. Geometric definition of vector sum

368 6 VECTOR ANALYSIS

multiplication and forms a vector space in the sense of Section 5.3. This means that the
following eight properties hold.

(1) A + B = B + A for all vectors A and B.

(2) A + (B + C) = (A + B) + C for all vectors A, B, and C.
(3) A + 0 = 0 + A for all vectors A.
(4) For each vector A there is a vector B such that A + B = B + A = o.
(5) a(A + B) = aA + aB for all scalars a and vectors A and B.
(6) (a + b)A = aA + bA for all scalars a, b and vectors A.

(7) (ab)A = a(bA) for all scalars a, b and vectors A.

(8) (l)A = A for all vectors A.

The proofs of these simple properties, based on plane geometry in E2 , will be left as
exercises. The identification of the geometric vectors in E2 as a vector space allows us to
apply to them the concepts and results of Chapter 5. In particular, we have the following
corollary:

Corollary. The geometric vectors in E2 form a vector space of dimension two.

Proof. Only the dimension statement needs to be checked. The geometry of E2 permits
us to construct pairs of noncollinear line segments and corresponding vectors Al and
A2 . We shall show that any such pair is a basis, so that each vector A in E2 can be written
uniquely as

(9)

where al and a2 are scalars. The proof is geometric; see Figure 3.

If A = 0, we may take al = a2 = 0 in (9). If A "# 0, we construct the parallelogram
-70
PQRS with diagonal A and sides parallel to Al and A2 . Then PQ is a multiple of Al and
-70
PS is a multiple of A2 . The geometric definition of the sum alAI + a2A2 then gives (9).
The Norm of a Vector. The norm of a vector A is its magnitude (= length of the
corresponding line segment). The norm of a vector A is denoted by IAI. The following
simple properties of the norm are implied by the definition and the geometry of E2 .

(10) IAI ::: 0, and IAI = 0 if and only if A = o.

 6.1 VECTOR ALGEBRA 369

FIGURE 3. Construction of a basis

(ll) IA + 81 SiAl + 181 (triangle inequality).

Cartesian Coordinates in E2. A Cartesian coordinate system in E2 is defined by

choosing a point 0 of E2 as origin and two perpendicular lines through 0 as coordinate
axes. The units of distance along the two axes are assumed to be the same. See Figure 4.
Such a system associates to each pointP ofE2 a unique ordered pair of real coordinates
(x, y). The coordinates x and yare constructed by dropping perpendiculars from Ponto
the x- and y-axes, respectively, using standard constructions of plane geometry The
mapping that assigns to each point P of E2 its coordinate pair (x,y) in R2 is denoted by
the notation P *+ (x,y).
The mapping P *+ ex, y) is a bijective map of E2 onto R2, which means that it maps
all of E2 onto all of R2 and is one-to-one. Some texts use this property to define E2 to
be R2. However, this leads to conceptual problems because it appears to give a special
role to a particular Cartesian coordinate system, whereas all such systems should enter
on the same footing.

y p

Q x FIGURE 4. ACartesian coordinate system in E2

370 6 VECTOR ANALYSIS

Y2 Q

Yl p R

L-~------------~--~X

Xl FIGURE 5. The distance formula

The distance between two points in E2 may be computed from their Cartesian
coordinates. Forlet P ++ (Xl ,Yl) and Q ++ (X2,Y2) be any pair of points, as in Figure 5,
and mark the point R ++ (X2,Yl).
PR
Then the distance I I equals IX2 - xII, the distance IRQI equals IY2 - Yll, and hence,
by the Pythagorean theorem,

distance(PQ) = J(X2 - Xl)2 + (Y2 - Yl)2.

Component Representation of Vectors in £2. Each Cartesian coordinate system

in E2 provides a basis for the geometric vectors in E2. These are the unit vectors (= vectors
of unit norm) directed along the positive x- and y-axes. It is traditional to denote them
by i and j, respectively Thus if a geometric vector in E2 is defined by the directed line
--+
segment PQ of Figure 5 then

The real numbers al = X2 - Xl and a2 = Y2 - Yl, uniquely defined by the coordinate

system and

are called the components of A in the given coordinate system. Clearly, the mapping
A ++ (aI, a2) is a bijective map of the vector space of geometric vectors onto R2. Many
textbooks identify the vector space of geometric vectors with the vector space R2 of
Chapter 5. This can lead to confusion because the pair (aI, a2) depends on the choice of
the coordinate system, while A does not.
Euclidean Space E3 . The mathematical setting for Newtonian mechanics, and for
much of the mathematical physics of the eighteenth and nineteenth centuries, is the three-
 6.1 VECTOR ALGEBRA 371

dimensional Euclidean space E3 of high-school solid geometry The theory of E3 and its
vectors closely follows that of E2 . Thus the geometric vectors for E3 are the equivalence
--+
classes of directed line segments PQ, where equivalence is defined as in Figure 1. Vector
addition and scalar multiplication are defined as before, and the set of all geometric
vectors forms a real vector space under these operations. The one real difference is that
the geometric vectors in E2 form a two-dimensional vector space, while those in E3 form
a three-dimensional vector space. Thus if Ai, A2 and A3 are any noncoplanar vectors in
E3 then every vector A can be written uniquely as

Cartesian Coordinates in E3. A Cartesian coordinate system in E3 is defined by

choosing a point 0 of E3 as origin and three mutually perpendicular lines through 0 as
coordinate axes. The units of distance along the three axes are assumed to be the same.
See Figure 6.
Such a system associates with each point P of E3 a unique ordered triple of
real coordinates (x,y, z). The coordinates x, y and Z are constructed by dropping a
perpendicular from P onto the x-axis, y-axis and z-axis, respectively The mapping that
assigns to each point P of E3 its coordinate triple (x, y, z) in R3 is denoted by the notation
P B (x,y,z).
Distance in E3 . The distance between points 0 B (0,0,0) and P B (x,y, z) can be
calculated by two applications of the Pythagorean theorem to Figure 6:

distance(PO) = JX2 + y2 + Z2.

If P B (Xl,Yl, Zl) and Q B (X2,Y2, Z2), then the same method gives

distance(PQ) = J(X2 - Xl)2 + (Y2 - Yl)2 + (Z2 - Zl)2.

Component Representation of Vectors in E3 • Every Cartesian coordinate system

in E3 determines a basis for the geometric vectors in E3. Extending the notation used

x FIGURE 6. A Cartesian coordinate system in E3

372 6 VECTOR ANALYSIS

for E2 , we shall denote by i, j and k the unit vectors along the positive x, y and Z axes,
respectively: With this notation every geometric vector has a unique representation

The numbers al, a2, a3 are called the components of A in the given coordinate system.
--+
By associating A with the directed line segment PQ above, we see that al = X2 - Xl,
a2 = Y2 - Yl, a3 = Z2 - Zl, and hence the vector norm of A satisfies

IAI = Jai + a~ + a~.

The Dot Product. If A and B are two vectors in E3 then the dot product of A and B
is the real number defined by

(12) A B-
. - 1 0 ifA=OandlorB=O,
IAIIBI cose if A and B are both not 0,

where e is the smaller of the two angles between A and B, so 0 :::: e : : 1(; see Figure 7.
The dot product is also called the scalar product or the inner product in some
texts. We shall see that it has many uses in applications of vector algebra to problems of
geometry and physics. Here we shall give several properties of the dot product that are
useful in applications. The most important is described by the following theorem.

Theorem. Let A and B have component representations

(13) A = ali + a~ + a3k,

B = bli + b~ + b3 k,

relative to a Cartesian coordinate system. Then

(14)

A-B
B

A FIGURE 7. Definition of A· B
 6.1 VECTOR ALGEBRA 373

Proof. Equation (14) is clearly true if A = 0 and/or B = O. If neither A nor B is 0 then,

since

we have

(15) IA - BI2 = (al - bl )2 + (a2 - b2)2 + (a3 - b3)2

= IAI2 + IBI2 - 2(al bl + a2b2 + a3b3).
On the other hand, the law of cosines applied to the triangle of Figure 7 gives

(16)

Combining (15) and (16) gives (14).

Additional Properties. The following useful rules follow readily from (12) and (14):

(17) A . B = B . A (symmetry).
(18) (aA + bB) . C = a(A . C) + b(B . C) (linearity).
(19) A . A = IAI2 > 0 unless A = 0 (positivity).
These properties will be used freely in vector algebra calculations.
Position Vectors in E3 • For geometrical calculations it is often convenient to
represent the points of E3 by vectors in E3 . We shall illustrate the method by applying
it to a description of the lines and planes in E3 . To begin, we introduce a Cartesian
coordinate system and write

0# (0,0,0) (origin of the coordinate system),

P # (x,y, z) Cany point of E3)'

Then the position vector of P is the vector

-+
(20) r = OP = xi + yj + zk,

as illustrated in Figure 8.
Lines in E3. A line L in E3 is specified by two distinct points on it or, equivalently, by
a point on it and a nonzero vector parallel to it. The second deSCription gives the formula

(21) r = ro + tA, -00 < t< 00,

where rand ro are the position vectors of any two points on L and A =f. 0 is parallel to
L; see Figure 9.
374 6 VECTOR ANALYSIS

x FIGURE 8. Position vector of a point P ++ (x,y, z)

To find a coordinate representation of L, let

ro = .xo i + yoj + zok,

A = ali + aD + a3k.
Then (21) implies the parametric equations

(22)

y = Yo + ta2,
Z = Zo + ta3,
where (al, a2, a3) =f. (0,0,0). A nonparametric description is obtained by eliminating t
to get

x - Xo Y - Yo Z - Zo
(23) - - = - - = - - (= t).
al a2 a3
Here it is understood that if, for example, a2 = 0 then y = yo. Thus if al = a2 = 0 in
(23) then L is the line described by x = Xo, Y = Yo, and Z = Zo + a3t. This is the line
through (xo,Yo, Zo) that is parallel to the z-axis.
Planes in £3. A plane Il in £3 is specified by three points on it or, equivalently, by
a point on it and two linearly independent vectors that are parallel to it. The second

o FIGURE 9. The line L through fo and parallel to A

 6.1 VECTOR ALGEBRA 375

construction gives the formula

(24) r = TO + sA + tB,
where TO is on n and A and B are linearly independent, so that neither is a multiple of
the other; see Figure 10.
To find a coordinate representation of n, let
TO = xoi + yoj + zok,
A = ali + a2.i + a3k,
B= bli + b2j + b3 k.

Then (24) implies the parametric equations

x = Xo + sal + th,
Y = Yo + sa2 + tb 2,
Z = Zo + sa3 + tb3·
A nonparametric representation can be found by introducing the vector N orthogonal to
A and B, so

N . A = N . B = 0 (and N f. 0).

Combining this with (24) gives the equation

N . (r - ro) = O.

In particular, if

N = ai + bj + ck
then an equation for n is
a(x - xo) + bey - Yo) + c(z - zo) = 0,
or

ax + by + cz = d,

A
ro
FIGURE 10. Plane n through ro and parallel to
A and B
376 6 VECTOR ANALYSIS

where (a, b, c) =1= (0,0,0) and

d = axo + byo + CZo·

The Cross Product. If A and B are two vectors in E3 , then the cross product of A
and B is the vector in E3 defined by

(25) AxB= I0
nlAllBI sin e
if A = 0 andJor B = 0,
if A and B are both not 0,

where e is the angle between A and B, chosen so that 0 ::: e ::: n, and n is a unit vector
perpendicular to A and B.
There are two unit vectors that are perpendicular to A and B, each being the negative
of the other. The desired vector n is determined by the following rule.
Right Hand Rule. The direction of n is the direction of advance of a right-threaded
screw, perpendicular to the plane of A and B, if the screw's head is rotated from A to B
through the angle e; see Figure 11.
The cross product is also called the vector product or the outer product in some
texts. Some of its applications to geometry and physics are developed in this and
subsequent chapters.
Properties of the Cross Product. The following properties can be proved directly from
the definition of the cross product:

(26) B x A = -A x B (antisymmetry).

(27) (aA + bB) x C = a(A x C) + b(B x C) (linearity).

Note that (26) implies that A x A = O. Only the linearity property is difficult to prove
from the definition (25). For a proof see standard texts on vector calculus such as [D-SJ.
The details will not be reproduced here.

AxB
AxB

~ ~~
.... .... .. •....•...•••....•.•...............

FIGURE 11. Definition of Ax B

4:~
 6.1 VECTOR ALGEBRA 377

The final property of the cross product is the coordinate representation

= a2b3 - a3b2,
(28)
= a3bl - al b3,

and where ai, a2, a3, bl , b2, b3 are the components of A and B, defined by (13). This is
readily derivable from the linearity property (27) and the elementary products

i xi =j xj = k x k = 0,
i xj =-j xi = k,
j x k = - k x j = i,
k x i =- i x k = j.
The latter follows directly from (25). Indeed,

Axi = (ali + a~ + a3k) x i = -a2k + a3j,

A xj = (ali + a2j + a3k) x j = alk - a3i,
A x k = (ali + a~ + a3k) x k = -ad + a2i.
Multiplying these three equations by bl , h b3 , respectively and adding and using (27)
gives (28).
It is left for the Exercises to show that if (28) is taken as a definition then properties
(25), (26), (27) follow as consequences.
Symbolic Representation. The rules for evaluating determinants and equation (28)
imply the symbolic representation

j k
(29) A x B= al a2 a3
bl b2 b3

Of course, (29) is not a true determinant because the first row consists of vectors rather
than scalars. However, if (29) is expanded in minors along the first row then equation
(28) is recovered. Equation (29) is a useful mnemonic for remembering (28).

Triple Scalar Product. This product, defined by

(30) [A,B,C] = A· (B x C),

378 6 VECTOR ANALYSIS

assigns a scalar to each ordered triple of geometric vectors. From (29) it is easy to see
that if C = cli + c2.i + C3k then

Cl C2 C3
(31) [A,B,CJ = al a2 a3
bl b2 b3
Conversely, the properties of the cross product are all easily derivable from (30) and (31).
For example, (26) follows from (31) and the properties of determinants. Derivations from
(31) of other properties are left as exercises.
Vector Identities. The following identities are frequently useful in cross product
calculations.

(32) A x (B x C) = (A . C)B - (A . B)C,

(33) (A x B) x (C x D) = [A,C,DJB- [B,C,DJA,
(34) (A x B) . (C x D) = (A· C)(D . B) - (B . C)(D . A).

Verification of these identities are straightforward but tedious. More concise proofs can
be found in [D-SJ and other vector calculus texts.
Notational Conventions. Some references use A = xi + yj + zk and the triple
notation A = (x,y, z) to mean the same thing. This is valid when a fixed basis i, j, k is
used in calculations. The notation is used in computer algebra systems to apply linear
algebra methods to vector analysis problems. We will adopt this notational convenience
when the context makes it clear that a fixed basis i, j, k is being used:

A = (x,y, z) means A = xi + yj + zk.

1. (Parallelogram Diagonals) Given an origin 0 above the plane of parallelogram ABCD, form vectors
a = oA, b = aB, c = OC, d = 00. The diagonals of the parallelogram intersect at a point P. Find a
vector formula for oP.
2. (Parallelogram) Let A, B, C, D be the consecutive vertices of a parallelogram. Let 0 be an origin not in
-+
the plane of the parallelogram ABCD. Define a = OA, etc. Explain geometrically why b - a = c - d.
3. (Equality of Vectors) Show that two vectors a, b are equal if and only if la - bl = O.
4. (Zero Vector) Is the zero vector 0 a directed line segment? What is the magnitude of the zero vector?
5. (Parallel Vectors) Find a condition for two nonzero vectors a and b to be parallel.
6. (Magnitude) Let a, b be nonzero vectors and define c = Ibi/ial. Find the magnitude of ca.
 6.1 VECTOR ALGEBRA 379

7. (Midpoint) Let a, b be vectors joining the origin 0 to points A, B respectively Show that i(a + b)
corresponds to the midpoint of segment AB.
8. (Normal to a Plane) How many distinct unit vectors are orthogonal to a given plane'

9. (Triangle Inequality) Prove that lal ~ Ibl :::: la ~ bl·

10. (Identities) Describe the nonzero scalars q, C2, C3 such that

qa + c2(a + b) + c3(a ~ b) = 0
for all vectors a and b.

11. (Magnitude) Given vector a such that lal = n, find 14al.

Scalar and Vector Projection. Dot Product. The scalar projection of u onto the direction of v and the
vector projection of u onto the direction of v are defined respectively by the formulas

Projv(u) = u·v
-~,
~
Projv(u) = u·v
-~ v.
v·v
Ivl

Let u = (t, et , e2t ), v = (t + I, et - l ,e l - t ) and let t = 0.6. Compute the requested quantity and display the
result with 3-digit accuracy.

12. Find the vector projection of u onto the direction of v.

13. Find the three altitudes of the triangle determined by u and v.

14. Find the acute angle between u and v.

15. Find all angles in the parallelogram determined by u and v.

Cross Product and Scalar Triple Product. Let

u = (t 2 ,et ,e- t ),
v= (t " e2t e- 2t ) ,
w = (I,~I,O),

and let t = 0.5. Compute the following answers and display the result with 3-digit accuracy
16. Find the area of the parallelogram determined by u and v.
17. Compute the volume of the parallelepiped determined by u, v, w.

18. Find the cross products u x wand v x w.

19. Find the magnitude of the cross product of the diagonals of the parallelogram determined by u and v.
20. (Cross Product) Find a unit vector x that is perpendicular to both (2, I, I) and (1, ~I, 2).

21. (Angle in a Triangle) A triangle in R2 has vertices P = (3, ~l), Q = (4,5) and R = (~I, ~2). Find
e
the cosine of the angle at the vertex P.

22. (Scalar Triple Product) Show that if any two vectors in a scalar triple product are equal, then the value
of the scalar triple product is zero.
23. (Coplanar Points) It is known that three vectors are coplanar if and only if their scalar triple product
is zero. Use this fact to find x such that the three vectors below are coplanar.

(~I, ~I, 1), (2, ~l,l), (x, ~l,x).

380 6 VECTOR ANALYSIS

24. (Vector Triple Product) Let a = 0,2, -3), b = (2, -1,1) and c = (1, -1, 1). Compute the vector
triple product a x (b x c).
25. (Scalar Triple Product) Given three vectors a, band c, simplify (a + b) . (b + c) x (a + c).

26. (Vector Triple Product) Prove the second of the fundamental identities

a x (b x c) = (a . c)b - (a· b)c,

(a x b) x c = (a . c)b - (b . c)a

from the first.

27. (Vector Line Equation) Find the vector equation of the line through 0, -1, 1) and parallel to 0, 2, -1).
28. (Vector Line Equation) Find the vector equation ofthe line through the points (1, -1, 1) and (-1,0,1).

29. (Vector Line Equation) Find the point in the plane x + 5y + 6z = -1 that is also on the line through
(0,1,1) in direction (-I, -1,2).

30. (Near Point) Find the point R nearest to the origin on the line L through P = 0, -4, 6) and Q =
(3,1,2)

31. (Equation of a Plane) Find the equation of the plane passing through the three points P = (t, t2 , t3 ),
Q = (t sin(t), t cos(t) , t2 ), R = (sin(t), cos(t) , t), where t = 0.886.
32. (Intersection of Planes) Find an algebraic description of the intersection of the two planes 3x-y+4z =
3 and -4x - 2y + 7z = 8. Plot the two planes on one set of axes and show the intersection set.
33. (Angle Between Planes) Find the angle in degrees between the two planes 3x - y + 4z = 3 and
-4x - 2y + 7z = 8.
Geometry. Prove by means of geometrical definitions the follOwing properties of geometrical vectors. In
the proofs, use only methods of plane geometry
34. A + B = B + A for all vectors A and B.
35. A + (B + C) = (A + B) + C for all vectors A, Band C.
36. A + 0 = 0 + A for all vectors A.

37. For each vector A there is a vector B such that A + B = B + A = O.

38. a(A + B) = aA + aB for all scalars a and vectors A and B.
39. (a + b)A = aA + bA for all scalars a, b and vectors A.

40. (ab)A = a(bA) for all scalars a, b and vectors A.

41. O)A = A for all vectors A.

Cross Product Equivalences. In the three exercises below, take as the definition of cross product

(35) A x B = qi + c2j + C3k, where

q = a2b3 - a3 b2,
C2 = a3bl - alb3,
C3 = alb2 - a2bl·

42. Assume (35) to prove

(36) B x A = -A x B (antisymmetry).
 6.2 VECTOR CALCULUS OF CURVES IN SPACE 381

43. (Cross Product) Assume (35) and prove

(37) (aA + bB) x C = a(A x C) + b(B x C) (linearity).

e
44. (Cross Product) Assume (35) as a definition. Let A and B be two vectors in E3. Define to be the angle
between A and B, chosen so that 0 :5 e :5 Jr. Let n be a unit vector perpendicular to A and B chosen
by the right hand rule. Prove the properties below.

(38) A B= {0 if A = 0 and/or B = 0,
x nlAllBI sine if A and B are not O.

6.2 Vector Calculus of Curves in Space

This section presents applications of vector algebra and calculus to the study of space
curves (= curves in E3). The essential property of a space curve is its one-dimensionality.
This is made precise below by requiring the curve to have a parametric representation
in terms of a single real parameter. First, we shall review some simple properties of
vector-valued functions that are needed for a study of space curves.
Vector-Valued Functions. Consider a function that assigns to each number t from
an interval a ::: t ::: b a vector fCt) in E3 . Such functions occur frequently in geometry
and physics, where f(t) may represent a displacement, velOCity, acceleration, electric
field, magnetic field, gravitational field, etc. Such functions will be called vector-valued
functions of the variable t. Relative to an orthogonal basis i, j, k, the vectors f (t) can be
written as

(1) f(t) = fl(t)i + h(t)j +h(t)k, a::: t ::: b,

wherefl(t) = f(t) . i,f2(t) = f(t) . j andh(t) = fCt)· k. Thus a vector-valued function
may be represented by an ordered triple

of real-valued functions.
Space Curves in E3 • Relative to an origin 0 in E3 , the points pet) in E3 , defined by
-:=+
(2) OP(t) = f(t), a::: t::: b,
represent a point set in £3' The space curves treated in this section will be constructed
in this way. As preparation for the construction, some simple concepts of vector calculus
are reviewed.
382 6 VECTOR ANALYSIS

Limits of Vector-Valued Functions. Let f(t) be a vector-valued function defined

on a::: t ::: b, and let A be a fixed vector in E3. Then if a < to < b we say that

(3) lim Ht)

~~
=A if and only if lim IHt) - AI = O.
~~

Moreover, if (1) holds and A = ali + a2.i + a3k then

and it follows readily that

(5) lim f(t)

~~
=A if and only if lim A(t) =
~~
ak for k = 1,2,3.
Continuity of Vector-Valued Functions. In accordance with elementary calculus,
a vector-valued function f (t) is said to be continuous at t to if =
(6) lim f(t)
t->to
= f(to).
Note that from (4) with A = f(to), f(t) is continuous at t = to if and only if the
components fk(t) are continuous at t = to for k = 1,2,3.
Derivatives of Vector-Valued Functions. Derivatives of vector-valued functions
are defined by the difference quotient formula, just as in the scalar case. Thus we define

(7) f'(t) = lim f(t + h) - Ht)

h->O h
whenever the limit exists. It follows that f(t) has a derivative at t = to if and only if each
componentjk(t) has a derivative R(t) for k = 1,2,3. Moreover,

(8) f' (t) = f{ (t)i + j~ (t)j + j~ (t)k.

Linearity of the Derivative. Let Ht) and get) be vector-valued functions defined
on a ::: t ::: b, and let c and d be scalars. Also, let Ht) and get) be differentiable. Then
the linear combination d(t) + dg(t) is also differentiable and

(9) (d(t) + dg(t))' = d'(t) + dg'(t).

The proof is the same as in the scalar case.
Derivatives of Products. The calculus rule for the derivative of a product of two
scalar functions,

[f(t)g(t)]' = j'(t)g(t) + j(t)g'(t),

 6.2 VECTOR CALCULUS OF CURVES IN SPACE 383

extends immediately to the various products involving vectors. Thus if aCt) and f(t) are
scalar and vector functions, respectively, then

(10) [a(t)f(t)]' = a'(t)f(t) + a(t)f'(t).

This follows if we write f in component form (1) and use (8) and (9). Similar product
rules, provable by the same method, are

(ll) [Ht) . g(t)]' = f'(t)· g(t) + f(t)· g'(t)

and

(12) [f(t) x g(t)]' = f'(t) x get) + f(t) x g'(t).

Care must be taken in (12) to preserve the order of the vector factors because the vector
product is not commutative.
Space Curves. The space curves studied below are subsets C of Euclidean space E3
that have parametric representations

(13) r = r(t) = x(t)i + y(t)j + z(t)k, a::: t ::: b,

where x(t), yet) and z(t) are real-valued functions of t, defined for a ::: t ::: b. We
shall assume that the correspondence t --+ r(t) is one-to-one (or bijective). This means
that each point of C corresponds to exactly one point of the interval a ::: t ::: b. The
correspondence is also assumed to be continuous, which implies that C is a connected
set of points.
Equivalent Parameterizations. Notice that the parameterization of C defined by
(13) is in no way unique. For let 1/I(t) have a positive derivative 1/1' (t) and let t = 1/1 (.)
map a' ::: • ::: b' onto a ::: t ::: b. Then r = r(1/I(.» will map a' ::: • ::: b' one-to-
one onto C and will be a second parameterization of C. We shall call it an equivalent
parameterization of c. Clearly, a given space curve has infinitely many eqUivalent
parameterizations. Each one of them determines the curve C as a set in E3 , and its
orientation rea) is the initial point of C, and reb) is its final point. A point ret) is between
r(tl) and r(t2) if and only if tl < t < t2.
Smooth Space Curves. The class of space curves defined above is too large to
permit the use of methods of vector differential calculus. We shall restrict our study
to parametric curves that have at least three continuous derivatives. This suggests the
follOwing definition.
Definition. A space curve C is said to be a smooth space curve if and only if it satisfies

(14) t --+ ret) is one-to-one from a < t < b to C.

384 6 VECTOR ANALYSIS

(15) The derivatives r(k) (t) = dkr(t)!dt k exist and are

continuous for k = 1,2,3 and a :s t :s b.
(16) r'(t) ::f. 0 for a :s t :s b.
The Tangent Vector. Condition (16) and the difference quotient definition of the
derivative imply that the vector

dr ,
(17) dt = r (t)

is a tangent vector to C at the point ret). Thus conditions (14)-(16) guarantee the
existence of a continuously turning tangent line at each point of C.
Elementary Examples. The following three examples of smooth space curves illustrate
the concepts defined above.

EXAMPLE 1. A Straight Line. Let ro and rl be the position vectors of two distinct points
in E3 . Then

(18) ret) = (1 - t)ro + trl, O:s t :s 1,

describes the straight line segment joining ro and rl. In particular, reO) = ro and rO) =
fl. Also, f' (t) = fl - fO is a nonzero constant vector, parallel to the line segment r.

EXAMPLE 2. An Ellipse. Consider an ellipse with semiaxes a and b (a > b > 0). If it lies
in the z-plane of a Cartesian coordinate system and has axes along the x and y axes then
r = xi + yj + zk is on the ellipse if and only if

This may be viewed as an elliptic space curve with parametric representation

(19) ret) = (a cos t)i + (b sin t)j, O:s t :s 2n.

Note that r(2n) = reO) = ai, and the curves closes when t = 2n.

EXAMPLE 3. A Helix. A helix is the uniform spiral of DNA fame. A parametric

representation of a helix can be written as

(20) ret) = (a cos t)i + (a sin OJ + (bt)k, O:s t :s 4n,

where a > 0 is the radius and b > 0 is the pitch of the helix, which measures the spacing
of successive turns of the spiral; see Figure 12.
 6.2 VECTOR CALCULUS OF CURVES IN SPACE 385

x ~>y

FIGURE 12. A helix

Length of a Smooth Curve. In elementary calculus courses the length f. of a smooth

curve (13) is defined by

(2l) f. = 1b Jr' (t)Jdt.

This is usually motivated by deriving (21) as a limit of the lengths of inscribed polygonal
curves. Also, (13) implies that

(22) r' (t) = x' (Oi + y' (OJ + z' (t)k, a ~ t ~ b,

which gives the alternative equation

(23) f. = 1b J(X' (t»2 + (y' (t»2 + (z'(t»2dt.

Arc Length as a Parameter. If s is the length of the portion of C for which a ~ r ~ t
(where a ~ t ~ b) then (21) gives

(24) s= 1t Jr'(r)Jdr == rp(t),

°
where rp(O maps a ~ t ~ b one- to-one onto ~ s ~ f.. Moreover, if C is a smooth
space curve then rp(O has continuous derivatives through order three. Also, by (16),
ds r--~--~~~~~~
(25) dt = rp' (t) = Jr'(t)J = V(X' (t))2 + (y'(t))2 + (Z'(t»2 > 0,

and we see that (24) is invertible and t = rp-l(S) has derivatives of order three or more
°
for ~ s ~ f.. Thus r(rp-l(s» defines an equivalent parameterization with arc length s
as parameter. A particular consequence of equation (25) is the theorem that the tangent
vector r'(O is a unit vector if and only if s = t - a. We shall see below that the arc length
parameterization is very useful in studying the structure of smooth space curves.
The Geometry of Space Curves. The information given above will now be used to
show that the shape of each smooth space curve is determined by two scalar functions
386 6 VECTOR ANALYSIS

of the arc length parameter. We begin with the arc length representation of C as

(26) res) = x(s)i + y(s)j + z(s)k, 0::: s ::: c,

where xes), yes) and z(s) have continuous derivatives of order three. We know that

(27) T(s) = dr = r'(t)

ds Ir'COI
is the unit tangent vector to C at res) and so T(s) . T(s) = 1 for 0 ::: s ::: C. On
differentiating this equation with respect to s via the rule (11) above, we find that
dT(s)
(28) T(s) . ~ = 0 for 0::: s ::: e.
It will be convenient to distinguish three cases.

dT(s)
Case 1. ~ = 0 for 0 ::: s ::: C.

dT(s)
Case 2. d;- =f:. 0 for 0 ::: s ::: e.
dT(s)
Case 3. - - = 0 for some values of s but not all.
ds
In Case 1, integration gives T(s) = To = constant. It follows easily that C is a straight
line parallel to To. In Case 2 we have

dT(s) I >
(29) K(S) = Id;- 0 for 0::: s ::: e.
This will be called the regular case. The function K(S) is called the curvature of C at
res). In Case 3, K(S) is zero at some values of s and not zero at others. If we work between
zeros of K(S) we revert to the regular case. The study of the behavior of C near the zeros
of K(S) is too difficult to study here. It is studied in advanced geometry texts.
The Regular Case. In this case we have, by (29),

dT(s)
(30) d;- = K(s)N for 0::: s ::: C,

where N = N(s) is a unit vector orthogonal to T by (28). Thus N lies in the normal
plane to Cat res); N is called the principal normal at res). To span the normal plane
at res) we need another normal vector that is orthogonal to both T and N. An obvious
choice is the binormal at res), defined by

(31) B(s) = T(s) x N(s), 0::: s ::: e.

 6.2 VECTOR CALCULUS OF CURVES IN SPACE 387

Since T and N are orthogonal unit vectors, we see that B is a unit vector and

(32) B is orthogonal to T and N.

In fact, for each s, the vectors T, Nand B form an orthogonal basis for all vectors in E3 .

Torsion of a Space Curve. On differentiating the identities B . B = 1 and B . T = 0

and using (30), one finds that dBlds is orthogonal to Band T and hence is parallel to N.
Thus there is a scalar function r = res) such that

dB
(33) - = -rN
ds
for 0<
-
S <
-
.e.
The coefficient res) is called the torsion of Cat res). (The use of the minus sign in (33)
is a convention and is not essential.) A fundamental theorem of elementary geometry
states that the two scalar functions K(S) and res) determine C completely except for its
position in space. An explanation of this fact is sketched below.

The Frenet-Serret Formulas. The derivatives dT/ds and dBlds are written in terms
of the orthogonal basis T, N, B in (30) and (33). To complete the triad we will show that

dN
(34) - = -KT+ rB.
ds
One way to do this is to differentiate the relation

(35) N=BxT.

Application of identity (12) to (35) gives

dN dB dT
-=-xT+Bx-.
ds ds ds
Combining this with (30) and (33) gives (34). The three vector relations

dT
(36) -=KN
ds '
dN
- KT + rB,
ds
dB
-rN
ds
are called the Frenet-Serret formulas after their discoverers F Frenet (1852) and JA.
Serret (1851).
388 6 VECTOR ANALYSIS

Calculation of K(S) and res) From C. If C is given in the arc length representation (26)
then differentiation and the definition (29) give

(37) ( d2x)2 (d2y)2 (d2z)2

ds 2 + ds 2 + ds 2
The calculation of res) is more difficult. From (36) we have
dN dN
r = B . a; = T x N· a;'
Combining this with

N= ~ dT and dN = ~ (~) dT + ~ d2T

K ds ds ds K ds K ds 2
gives

or, in determinant form,

dx !!l ~
dS tis ds
1 d2x !b.2 d2 z
(38) res) = K(S)2 tis 2 ds ds 2
d3 x
ds3
:!2
ds 3
d3 z:
ds 3

Calculation of C From K(S) and res). It is enough to calculate T(s), since integration
then gives Cas

res) = reO) + 15 T(a)da.

Now, the Frenet-Serret formulas, written in terms of components

(T,N,B) = (T1 , T2 , T3,Nl,N2,N3,Bl,B2,B3),

constitute a 9 x 9 first order system of linear ordinary differential equations. The

coefficients are determined by K(S) and res). It follows from the basic existence and
uniqueness theorems (Chapter 3) that the Frenet-Serret system (36) has a unique solution
for each choice of initial values TCO), N(O), B(O). In advanced geometry courses it is
shown that the choice of the initial values affects only the position in space of the solution
curve C.
Line Integrals of Scalar Functions. Consider the problem of calculating the mass
M of a slender nonuniform wire whose mass density is known. Let us model the wire by
 6.2 VECTOR CALCULUS OF CURVES IN SPACE 389

a smooth space curve C with representation

(39) ret) = x(t)i + y(t)j + z(t)k, a S t S b.

The mass density of the wire (mass per unit length of C) is a positive function p(x, y, z).
The usual approximation procedure gives the approximate Riemann sums

L p(x(U,y(ti), z(ti))lr'(ti)l~ti'
n

M ~
i=1

Passing to the limit as the partition norms tend to zero gives the definite integral

(40) M = lab p(x(t),y(t),z(t))lr'(t)ldt.

As before, equivalent parameterizations give the same value for M. In particular, if the
parameter is the arc length s along C, then ds = Ir'(t)ldt, and we have

(41) M = 1£ p(X(S),y(S), z(s))ds.

The common value of the integrals in (40) and (41) is denoted by

(42) M = [P(X,y, z)ds.

It is usually called the line integral over C of the scalar function p.

Line integrals of scalar functions occur frequently in physics. For example, if p(x, y, z)
represents the electric charge density along C then the integral (42) gives the electric
charge on C. Additional applications may be found in connection with Green's and Stokes'
theorems below.

I. Exercises 6.2 . ~
",~ bi, 1 R" ~nw' w"","'~' ~,,"'" ,'"

1. (Vector Form of a Line) Find a vector representation ret) = ro + tf) of the line that passes through the
points (3,4,7) and (1, -1, 3)
2. (Parametric Form of a Line) Use vector methods to find a parametric representation

x = xo + t(Xl - xo),
Y = YO + t(Yl - YO),
Z = Zo + t(Zl - zo)
of the line that passes through the point (3,4,7) in the direction of the vector i - j + 2k.
3. (Symmetric Form of a Line) Use vector methods to find a symmetric representation
x - Xo Y - YO Z - Zo
Xl - Xo YI - YO Zl - Zo
390 6 VECTOR ANALYSIS

of the line that passes through the point (3, 4, 7) in the direction of the vector 2i - j + k.
4. (Equations of a Line) Use vector methods to find the vector, parametric, and symmetric representations
of the line that passes through the point (3,4,7) and is parallel to the line of intersection of the planes
defined by the equations x + 2y + llz = 0 and 5x + lly + 58z = O.
5. (Circle Parameterization) Find the parametric equations for a circle with center (1,2,5) of radius 7 in
the plane z = 5.
6. (Parameterizations) Show that ret) = cos(t)i + sin(t)j, 0 ::: t ::: 7T12, is a parameterization of a quarter
circle and so is r(u) = (0 - u 2 )i + 2uj)/(1 + u 2 ), 0 ::: u ::: l.
7. (Triangle Parameterization) The boundary of the triangle with vertices 0,0,0), 0,1,0), and (2,3, 1)
is the image of a parametric curve r(t). Find one such parametric curve r(t).
8. (Tangent Vector) Find the velocity vector r' and unit tangent vector r'llr'l to the curve ret) = cos(t)i +
sin(t)j at t = 7T12.
9. (Constant Tangent) Let ret) = ti + (2t + 10)j + 5tk, which represents a line in space. Find the unit
tangent T and verify that it is constant, i.e., dT/dt = O.
10. (No Tangent) Find the points on the curve ret) = tj + Itlj that fail to have a tangent defined.
11. (Velocity and Speed) Let ret) = cos(t)(i - j) + sin(t)(i + j) + tk describe the motion of a particle.
Find the velOcity r' and speed Ir'l of the particle.
12. (Particle Dynamics) Let ret) describe the path of a moving particle that has constant speed (Ir'(t)1 =
constant). Does it follow that rl/ (t) = O? Explain.
13. (Angular Speed) A particle moves around the circle x 2 + y2 = 4 with angular velOCity dOfdt a constant
10 radians per second. Find a vector formula ret) for the motion of the particle and compute the particle
speed Ir' (t) I·
14. (Polar Equations) A plane motion ret) is given in polar coordinates p, 0 by the equations p = 1 + sin t,
o= -1 + e- t . Find the velocity r' (t) and the acceleration rl/ (t) at t = O.
15. (Polar Motion) A particle moves in the plane by the equation ret) = (1-cosO(t»(cos O(t)i +sin O(t)j),
where 0' (t) = 4. Find the velOcity r' (t) and the acceleration rl/ (t).
16. (Continuity) Prove that the dot product f(t) . get) and cross product f(t) x get) of continuous vector
functions f(t) and get) are continuous functions of t.
17. (Dot Product) Let aCt) and bet) be two continuously differentiable vector functions. Prove that
(a(t) . b(t»' = a' (t) . bet) + a(t) . b' (t).
18. (Cross Product) Let aCt) and bet) be two continuously differentiable vector functions. Prove that
(a(t) x b(t»' = a'(t) x bet) + aCt) x b'(t).
19. (Arc Length) Find the length of ret) = cos ti + sin tj + 2tk on 0::: t ::: 27T.
20. (Arc Length) Find the length of ret) = ti + sin(27Tt)j + cos(2m)k on 0 ::: t ::: 1.
21. (Arc Length) Find the length of the plane curve r(t) = (cos 3t)i + (sin 3t)j on 0 ::: t ::: 7T.
22. (A Formula for dTlds) Given a twice continuously differentiable vector function ret), let s
f~ Ir'(r)ldr and let T = drlds be the unit tangent. Prove the decomposition formula (below' = dldt)

I/() 1'( )12 dT(s) r'(t)· rl/(t) '()

r t=rt --+ rt.
ds Ir'(t)1 2
 6.2 VECTOR CALCULUS OF CURVES IN SPACE 391

23. (Normal and Tangential Components) Given a twice continuously differentiable vector function ret),
let s = I~ Ir' (r)ldr. Define as usual the unit tangent T = dr/ds, the curvature K = IdTldsl, and the
standard unit normal N = K-1dTlds. Prove the decomposition formula (below' = dldt)

r' . r"
aT = y'
24. (Curvature) Given s = I~ Ir'(r)ldr, prove the curvature formula

K(S) = Ir'(t) x r"(tWlr'(t) 13.

25. (Curvature) Show that the curvature ofr(t) = (3t-t 3 )i+3t2j+(3t+t 3 )kisK(s) = 1/(3(1+t2 )2).
26. (Tangential and Normal Components) In the acceleration representation r" (I) = aT T + aN N, which
says that r" is in the plane of T and N, the coefficients aT and aN are called the tangential and normal
components of acceleration. Find aT and aN for the helix ret) = (cos t)i + (sin t)j + 2tk.
27. (Circular Helix Tangents and Normals) Sketch the circular helixf(1) = cos2ti + sin2tj + tk for
o :::: t :::: 4rr and draw on the curve aU = rr/4 the unit tangent vector T and the standard normal vector N.
28. (Binormal Magnitude) Prove directly from the definition B = T x N that the binormal B has unit length.

29. (Formulas for T, N and B) Show that the following formulas are valid for a twice-differentiable vector
function r(t).

T = 1:1' x = r' (I),

B= 1:1' x = r'(t) x r"(t),

N= 1:1' x = (r' (I) x r" (t» x r' (t).

30. (Helix Formulas) Verify the following formulas for the circular helix ret) = (a cos t)i + (a sin t)j + btk,
where a > 0, b > °and R= Ja2 + b2
-a a b
T = -(sint)i + -(cost)j + -k,
R R R
b -b a
B = -(sint)i + -(cost)j + -k,
R R R
N = - (cos t)i - (sin OJ.

31. (Trihedral Formulas) Verify the formulas B = T x N, N = B x T, T = N x B for the special case of
the helix ret) = cos ti + sin tj + tk.

32. (Binormal) Prove that the binormal B and its arc length derivative dBlds are orthogonal

33. (Reciprocal Basis) Let aI, a2 and a3 be given vectors that are nonzero and noncoplanar. Define

a2 x a3 al x a2
hI = _=-----0..._ b3 = -=---::..-
al . a2 x a3 al . a2 x a3

Compute hj, h2 and h3, given al = i +j, a2 = i -j, a3 = i +j + k.

34. (Reciprocal Basis) Let aj, a2 and a3 be given orthogonal unit vectors. Prove that the reciprocal basis
is given bybj = aj, h2 = a2, h3 = a3·
Line Integrals. Evaluate the line integral Ie p(x, y, z)ds over the given curve C, using the given orientation.
392 6 VECTOR ANALYSIS

35. The curve C is the straight line segment C from (1,0, 1) to 0,3,2). The functionpis defined by

p(x,y,Z) = 3y + xz - x.

36. The curve C is the semicircle y = J1 - x 2 clockwise from ( - 1, 0) to 0, 0). The function p is defined by

p(x,y) =x 2 +x2/.

37. Compute the line integral of the tangential component of the force F = xi + x 2yj along the straight
line from (-1,0) to 0,0). Hint: The answer is the line integral fcp(x,y,z)ds, where p(x,y,z)ds =
F· dr = F(r(t» . r'(t)dt.
38. Compute the line integral of the tangential component of the force F = xi + x 2yj along the polygon
from (-1,0) to (1,1) to (1,0).

6.3 Vector Calculus of Surfaces in Space

In this section the methods of vector analysis are used to develop the elementary
geometric properties of surfaces in E3 . Tangent planes and normal lines at a surface point
are constructed and surface areas are calculated. Finally, surface integrals of scalar fields
are defined and calculated. These results are needed for the study of Green's and Stokes'
theorems in Section 6.5 below.
In Section 6.2 space curves were represented parametrically as vector-valued func-
tions of one real parameter. Here we shall represent surfaces 5 in E3 as vector-valued
functions of two independent real parameters. Thus 5 is a subset of E3 whose position
vectors f have the form

(1) feU, v) = X(U, v)i + y(u, v)j + z(u, v)k,

where (u, v) ranges over a domain D in the (u, v)-plane. Before we discuss the general
properties of such surfaces, we shall give three simple examples.

EXAMPLE 1. Planes. Let fO, a and b be three vectors such that

(2) a x b i= o.
Then

(3) feU, v) = fO + ua + vb, -00 < u, v < 00,

represents the plane through fO that contains the straight lines f = fO +ua and f = fO +vb
(see Section 6.1). Note that if condition (2) is violated then a and b are linearly dependent
and the plane degenerates to a line or point.

EXAMPLE 2. Cones. Consider the upper nappe 5 of the cone with vertex at the origin
and axis along the positive z-axis (see Figure 13).
 6.3 VECTOR CALCULUS OF SURFACES IN SPACE 393

o
FIGURE l3. Cone with vertex angle ex

If ex is the cone half-angle, then from the figure,

Jx2 + y2
(4) -'-----'-- = tan ex,
Z
and then the cone has the parametric form

(5) x = (tanex)vcosu,
y = (tan ex)v sin u,
z = v,
with 0 :::: u :::: 2Jr, v ::: O. In vector form,S has the representation

(6) r(u, v) = {casu i + sin uj} (tanex)v + vk.

EXAMPLE 3. Unit Sphere. A scalar equation for the unit sphere is

(7)

One parametric representation is given by the defining equations for spherical coordinates
with r = Jx 2 + y2 + Z2 = 1, so (see Figure 14)

(8) x = sin u cos v,

y = sin u sin v,
Z = casu,

where 0 :::: u :::: Jr, 0 :::: v :::: 2Jr.

In vector form this is

(9) r(u, v) = (cos v i + sin v j) sin u + (cos u)k.

Nonuniqueness of Parametric Equations. As with space curves, parametric
equations of surfaces are never unique. Instead, each surface can be represented in many
ways. Thus in Example 1 we may replace a and b by two (linearly independent) linear
combinations of them to obtain equivalent equations.
394 6 VECTOR ANALYSIS

FIGURE 14. Unit sphere with polar angles

In Example 2, an elementary equation of the cone is

(0)

This suggests the representation

(ll) x=u,
y =v,
Z = (cota)Ju 2 + v2 ,

where -00 < u, v < 00.

In Example 3, the unit sphere, the analogue of (1), is

(12) x=u,
y =v,
Z = J'I---u2,-------v-,-2,
where 0 :s u2 + v2 :s l. Note that this gives the upper half of the sphere. The lower half
is obtained by changing the sign of the square root in (2).
Stereographic Projection. Another parametric representation of the unit sphere is
obtained by mapping each point (x,y, z) of the sphere from the north pole (0,0,1) onto
the equatorial plane (u, v, 0); see Figure 15.
Equating the unit vectors for the displacements from (0,0,1) to (x,y, z) and from
(0,0,1) to (u, v, 0) gives the vector equation
xi + yj + (z - l)k ui + vj - lk
(13)

Equating the coefficients ofi,j, and k and solving for x,y, and Z gives, after some algebra,
2u
(14) x = u2 + v2 + l'
 6.3 VECTOR CALCULUS OF SURFACES IN SPACE 395

z
(0,0,1)

(u, v, 0)

FIGURE 15. Stereographic map of the unit

x sphere

2v
Y = u2 + v2 + l'
u2 + v2 - 1
z - --::---::---
- u2 2 + v + l'
for -00< u, v < 00.
This representation is used in complex analysis. Note that the unit sphere with the
north pole (0,0,1) removed is mapped one-to-one onto the entire finite (u, v)-plane.
Smooth Surfaces in Space. Methods of vector calculus are applicable to parametric
surfaces (1) that have several continuous partial derivatives. We shall find the following
definition useful.
Definition. A surface 5 in E3 is said to be a smooth surface if it satisfies (with D a
domain in the Cu, v)-plane)

(15) (u, v) ~ r(u, v) is one-to-one from D to S.

(16) The partial derivatives ak+lr(u, v)/akualv exist and
are continuous to order k +.e = 3 for all (u, v) in D.
(17) ru x rv =f. 0 for (u, v) in D.
It is not difficult to see that planes, cones and spheres are smooth surfaces with suitable
choices of D. For cones, the vertex is a Singular point and must be avoided.
Limits, Continuity, Derivatives. In Section 6.2 we have already applied these
terms to vector-valued functions of one real variable. It is worth noting that they are also
applicable to vector-valued functions of any number of variables. The results stated in
Section 6.2, equations (1)-(12), are easily extended to any number of variables.
Geometry of Surfaces in Space. We shall consider only smooth surfaces 5 as
defined by (1), (15)-(17). Our first step will be to calculate the tangent vectors to S.
396 6 VECTOR ANALYSIS

Notice that if Co is any smooth curve in D then it is lifted by the mapping (1) to a space
curve C that is embedded in s!
u-Curves in S. To begin, let (uo, va) lie in D and consider the curve in 5 defined by

(18) f = feU, va), Iu - uol < E.

We shall call it the u-curve through r(uo, va), since only u varies on it. The results of
Section 6.2 imply that the vector

(19)

is a tangent vector to the u-curve at fa = r(uo, vo). It is therefore also a tangent vector to
5 at fO.
v-Curves in S. Similarly, define the v-curve through fa by

(20) f = f(UO, v), Iv - vol < E.

Its tangent vector at fO is a second tangent vector to 5 at fa, namely

(21)

Note also that condition (17) implies that Tu and Tv are linearly independent.
Tangent Vectors to 5 at roo We shall show that every tangent vector to 5 at ro is a
linear combination of Tu and Tv. For this purpose define a curve Co in D by

(22) u(t) = Uo + u~t,

vet) = va + v~t, It I < E.

The image of Co in 5 is the curve C in 5 defined by

(23) r = f(U(t), vet)), It I < E.

By Section 6.2, the tangent to C at fO is obtained by taking the t-derivative and setting
t = 0 (so (u, v) = (uo, vo». The t-derivative of (23) at t = 0 is given by the chain rule as

Since we can choose any values for u~ and v~, it is clear that we obtain the entire tangent
plane in this way.
The Tangent Planes to S. The preceding result shows that the tangent plane to 5
at any point ro = f(UO, vo) passes through that point and is spanned by the vectors
fu (uo, vo) and fv(UO, va). Thus, by Example 1, the tangent plane has parametric equation

(24) f = fO + Ufu(UO, vo) + Vfv(Uo, vo), -00 < u,v < 00.
 6.3 VECTOR CALCULUS OF SURFACES IN SPACE 397

The Normal Lines to S. The normal vectors to 5 at a point r(uo, vo) are orthogonal to
the tangent plane at r(uo, vo). Thus they are just the scalar multiples of the normal vector

(25)

With this notation, a scalar equation for the tangent plane is

(26) N(uo, vo) . {r - r(uo, vo)} = O.

Surface Area. Two formulas for areas of surfaces are developed in most first-year
calculus courses. The simplest formula is applicable to surfaces of revolution. It states
that if x, y and z are Cartesian coordinates and if the graph of

y = J(x), a ~ x ~ b,

is revolved around the x-axis, then the area A of the resulting surface of revolution is
given by (see page 324 in [V-P])

(27) A = 2n lab J(x)) 1 + [j'(x)]2 dx.

The second formula is more general, applying to surfaces that are graphs of an equation

z = J(x,y), (x,y) in a domain D.

It states that the area A is given by (see [V-P, p. 802])

(28) A= fL/J1 + f} + 1 dxdy.

In most texts the proofs offered for these formulas are heuristic, being based on estimates
of terms in the corresponding Riemann sums. Here we shall derive a formula for the
area of any smooth parametric surface 5 as defined by (15)-(17). The proof will again be
heuristic. Finally, the generality of the new formula will be shown by deriving (27) and
(28) from it.
The Area Formula for Parametric Surfaces. The formula states that if 5 satisfies
(15)-(17), then its area is

(29) A(S) = f L,ru(u, v) x rv(u, v)1 dx dy.

Heuristic Derivation of (29). First we must realize that no rigorous proof of (29)
is possible because no independent definition of A(S) has been given. Instead, our
"derivation" is a heuristic argument meant to motivate equation (29).
398 6 VECTOR ANALYSIS

FIGURE 16. Decomposition of 5

For the derivation, D is assumed to be the rectangle defined by a SuS b, c S v S d.

This is done only to simplify the notation. The general case is similar. Next introduce the
uniform partitions

b-a
a = Ul < U2 < ... < U m+1 = b, f},Ui = f},u = --,
m
d-c
c = Vl < V2 < ... < Vn+l = d, f},vi = f},v = - - .
n
Note that the particular lines U = Ui (a v-curve) and v = Vj (a u-curve) decompose S
into small surfaces Sij, defined by (l) with Ui SuS Ui+l, Vj S v S Vj+l, i = 1, ... , m,
j = 1, ... ,n; see Figure 16.
Then the additivity of area gives

LL A(Sij)'
m n
A(S) =
i=1 j=1

Moreover, and this is the main idea behind equation (29), Sij is approximated by the
parallelogram at f(Ui, Vj) with edges f},u fu(Ui' Vj) and f}, v fv(Ui, Vj); see Figures 17 and 18.

Thus

FIGURE 17. Approximation of Sij

by a parallelogram S7j
 6.3 VECTOR CALCULUS OF SURFACES IN SPACE 399

FIGURE 18. Area calculation of parallelogram S;j

with the approximation improving as t.u -+ 0, t.v -+ 0. Combining the last two
formulas gives

L L Iru(u;, Vj) x rv(ui.vj)lt.ut.v.

m n
(30) A(S) ~
i=! j=!

This last sum is a Riemann sum for the double integral on the right in (29). Thus passing
to the limit (m, n) -+ (00,00) gives formula (29).
Application to Surfaces of Revolution. We shall show that formula (27) is a special
case of (29), thus giving a second derivation of (27). Our method is based on the
parameterization

(31) r(u, V) = feu) cos v i + feu) sin vj + uk,

obtained by rotating the graph of y = fez) about the z-axis. The elementary identity (see
Section 6.1, equation (34»

(32)

will also be useful. Differentiating (31) gives

ru = feu) cosvi +feu) sin vj + k,

rv = - feu) sin vi + feu) cosvj,

and so

Thus taking A = r u, B = Tv in (32) gives

Substituting this into (29) gives (27).

Application to the Graph of z = f(x,y). Here we will show that formula (28) is also a
special case of formula (29). The method of proof is to use the parameterization

(33) r(u, v) = u i + v j + feu, v) k.

400 6 VECTOR ANALYSIS

Differentiating gives

ru = i + Iuk,
rv = j + Iv k,
whence after a short calculation we get

ru x rv = -Iu i - Id + k
and

Iru x rvl = JIJ +I} + l.

Substituting this into (29) gives (28).
Uniqueness of Surface Areas. It was emphasized above that a given surface 5 in E3
can be parameterized in many ways. The integral formula (29) for the area of 5 contains
explicitly the functions of the parameterization and the question arises whether the
area can change with the parameterization. That it does not is implied by the following
theorem.

Theorem. Let 5 be a smooth surface in E3 and let

(34) f(u, v) = !ICU, v)i + hCu, v)j +hCu, v)k, Cu, v) in D,

and

(35) g(u' , Vi) = gi (u' , v')i + g2 CUi, v')j + g3 CUi, v')k, (u' , Vi) in D'
be any two parameterizations representing S. Then

(36) fi Ifu x fvldudv = fi Igu' X gv'lduldv' .

The proof, which is based on a theorem of advanced calculus for changing variables
in a double integral, is too long to reproduce here.

Notation. The invariance result (36) makes it natural to suppress the variables u, v and
write

dA = Iru x rvldudv

and

A(S) = fIs dA.

We will use this notation below when convenient.

 6.3 VECTOR CALCULUS OF SURFACES IN SPACE 401

Surface Integrals of Scalar Fields. If a smooth surface S is used to model a thin

lamina, and if the mass distribution in S (mass per unit area in S) is given by a mass density
function f(x, y, z), then we may estimate the total mass by a Riemann sum (see (30) above)

Mass of S ~ Lf(x(Ui, Vj),Y(Ui> Vj), Z(Ui> vj))lru(ui> Vj) x rvCui> Vj)l~u~v.

ij

Passage to the limit gives the formula

Mass of S = fif(X(U, v),y(u, v), z(u, v))lru x rvl~u~v.

Again, the integral is independent of the parameterization and we may write

(37) Mass of S = fif(X,Y,Z)dA.

The last integral is called the surface integral off over S. This concept plays an important
part in our analysis of Green's and Stokes' theorems in Section 6.5 below.

IExercises 6.31
1. (Partial Derivatives) The equations x = u2 , y = v2 , Z = uv define a surface S. Find the vector panials
ru and rv for the vector parameterization r(u, v) = xi + yj + zk.
2. (Plane) Find a vector normal to the plane r(u, v) = ui + vj + (- 3u + 2v)k.
3. (Plane) Find a vector parameterization r(u, v) for the plane 3x - 2y + 5z = 7.
4. (Plane) Find a vector equation r(u, v) for the plane parallel to the plane 3x-2y+5z = 7 and containing
(0,1, -1).

5. (Plane) LetP = (1,-1,0), Q = (1,1,-1), R = (0,1,-1). Find a vector equation r(u,v) for the plane
determined by P, Q, R.
6. (Cone) Represent the cone x 2 + Z2 = y2, Y :::: 0, as a vector parameterization r(u, v) on an explicit
(u, v)-domain.

l
7. (Cone) Represent the section of the cone x 2 + = 3z2 between the planes z = 1 and z = 3 as a vector
parameterization r(u, v) on an explicit (u, v)-domain.

8. (Hemisphere) Represent the upper hemisphere x 2 +y2 + z2 = 100, z :::: 0 as a vector parameterization
r(u, v) on a rectangle.

9. (Spherical Cap) Represent the spherical cap above the plane z = 2 on the sphere x 2 + l + Z2 = 16
by a vector parameterization rCu, v) on a rectangle.

10. (Stereographic Projection) Describe graphically the curve in the (u, v)-plane that corresponds to the
circle on the unit sphere in the plane z = 0.5 according to the stereographic projection equations.

ll. (Limits) Determine the limit at u = v

y = usin(l/u), z = uv.
= °for the surface 5 defined by the equations x = v cos(1/v) ,
402 6 VECTOR ANALYSIS

12. (Smooth Surfaces) The equations

x = u 3/(u 2 + v2 ), Y = v3 /(u 2 + i), z = 1

define a surface 5 if at u = v = 0 we define x = y = O. Is this a smooth surface 7

13. (Smooth Surfaces) The equations x = e- 1/u2 , y = e- 1/v2 , z = 1 define a surface 5 with domain
o ::s u ::s 1,0 ::s v ::s 1, provided that at u = v = 0 we define x = y = O. Is this a smooth surface?
14. (Smooth Conical Surface) Show that the section of the cone x 2 + y2 = 3z 2 between the planes z = 1
and z = 3 has a smooth vector parameterization r(u, v).
15. (Normal to a Surface) Let r = ui + vj + (u 2 + v2)k. Compute the normal n = ru x rv at u = 1, v = 2.
16. (Normal to a Surface) Find a vector parameterization r(u, v) and the outer unit normal n = ±N/INI.
N = ru x rv, for the surface z = x 2 + i·
17. (Tangent Plane) Find by vector methods the equation of the tangent plane to the surface z = 3x 2 + i
at the point (1, 1,4).
18. (Regular Surface) A point on a surface is said to be regular if the partials ru and rv at that point span
a plane, or, more SUCcinctly; ru x rv 1= O. Find the nonregular points for r = vi + vj + uvk.
19. (Space Triangle Parameterization) Let a space triangle PQR be given by three vertices P, Q, R. Define
edge vectors a = PQ, b = PR. Show that a parameterization of triangle PQR is
r(u, v) = oP + ua + vb
with domain

D = (cu, v) 0 ::s u .s: 1, O::s v ::s 1 - u}.

20. (Space Triangle Area) Let a space triangle PQR be given by three vertices P, Q, R. Define edge vectors
a = PQ, b = PR. Show that the area of triangle PQR is equal to

Surface Area. For each exercise below, identify a vector parameterization r(u, v) and its domain D.
Compute the partials ru, rv and Iru x rvl 2 = Irul21rvl2 - (ru . rv)~. Finally, compute the surface area

fi ,ru x rvl dudv .

21. The surface x = sinucosv,y = sinucosv, z= cosu, O::s u::s rr/2, O::S v::S rr/3.
22. The portion of the unit sphere at the origin that is above the plane z = 1/2.
23. The section of the surface x = u2 , Y = uv, z = v2/2 corresponding to the region bounded by the curves
u = 0, u = 1, v = 0 and v = 3.
24. The triangle with vertices 0,0,0), (0,1,0), (1, 1, 1).
25. An open box (no top) determined by the edge vectors i, 2j and 4k.
26. The closed tetrahedral surface formed from base triangle (0,0,0), (1,0,0), (0, 1,0) by connecting lines
to the fourth point (1, 1,2).
27. The cone x = vsinacosu, y = vsinasinu, z = vcosa, for 0 < a < rr/2 fixed and 0 ::s u < 2rr,
O::s v::S h.
 6.4 CALCULUS OF SCALAR AND VECTOR FIELDS 403

28. The ellipsoid with semiaxes a, b, c described by x = a sin u cos v, y = b sin u sin v, Z = c cos u,
0::0 u ::0 n, 0 ::0 v ::0 2n. Give the surface area in unevaluated integral form.

29. The cylinder x = a cos u, y = a sin u, Z = v, for 0::0 u ::0 2n, hI ::0 v ::0 h2, with a > 0 fixed.

30. The paraboloid x = av cos u, y = av sin u, Z = v2 , 0 ::0 u ::0 2:7r, 0 ::0 v ::0 h, with a > 0 fixed.
31. The section of the paraboloid Z = x 2 +l within the unit square 0 ::0 x ::0 1, 0 ::0 Y ::0 1. Give the
surface area in unevaluated integral form.

32. The surface of revolution x = v cos u, y = v sin u, Z = f( v), where f is a function of one variable defined
for VI ::0 v ::0 V2 and 0 ::0 u < 2n. Give the answer in integral form.

33. (Center of Mass) Let f(x, y, Z) = x + y + z represent the mass density per unit of surface area on the
surface 5 described by the equations x = u, y = v, Z = 1 - 2u - v, 0 ::0 u ::0 1,0 ::0 v ::0 2. Find the
center of mass (x,y, z) from the formulas

X= ~f Is xf(x,y,z)dS, z= ~ f Is ~(x,y, z)dS,

y= ~f IsYf(x,y,Z)dS, m = f Isf(x,y,Z)dS.

6.4 Calculus of Scalar and Vector Fields

A scalar field in E3 is a function J whose domain of definition D is a set of points ex, y, z)
in E3 and whose values Jex,y, z) are real numbers. Similarly, a vector field in E3 is a
function F whose domain of definition D is a set of points ex,y, z) in E3 and whose
values Fex,y, z) are vectors in E3 . Scalar and vector fields in the Euclidean plane E2 are
defined Similarly by replacing E3 by E2 and deleting the coordinate z.
Scalar fields have been used to model physical phenomena at least since the
work of L. Euler in the mid-eighteenth century Examples of scalar fields in physics
are temperature, pressure, mass density, charge denSity, gravitational potential and
electrostatic potential. It is easy to lengthen this list.
Vector fields as models for physical phenomena were popularized by the work of
M. Faraday and].C. Maxwell in the mid to late nineteenth century. Examples include
electric fields, magnetic fields, gravitational fields and velocity fields in fluid dynamics.
Visualizing Scalar Fields. Scalar fields Z = J(x,y) in E2 can be visualized by their
graphs in (x,y, z)-space. Another popular way to visualize such fields is by their level
curves

(1) J(x,y) = c.

This notion leads to the topographic maps of geography and the isothermal maps and
isobaric maps of weather prediction. A simple example is shown in Figure 19.
404 6 VECTOR ANALYSIS

c=3

FIGURE 19. Level curves ofj(x,y) = Jx 2 +y2

A scalar field w = j(x,y, z) in E3 has a graph in (x,y, z, w)-space (four dimensions)

and cannot be visualized directly. Instead, some geometrical insight may be gained by
constructing the level surfaces of j in E3 , defined by

(2) j(x,y, z) = c.

An example is shown in Figure 20.

Visualizing Vector Fields. This is more difficult. let us consider first a vector field
in E2 :

(3) F(x,y) = F1(x,y) i + F2 (x,y)j

If a grid of points is constructed in the domain D of definition of F, and if for each
grid point the vector F(x,y) is constructed with initial point (x,y), then the resulting
diagram represents the vector field graphically. This diagram is just the direction field
of the 2 x 2 system of differential equations

dx
(4) -;it = F1(x,y),
dy
dt = F2 (x,y).

FIGURE 20. Level surfaces of j(x,y, z) = Jx2 + y2 + Z2

 6.4 CALCULUS OF SCALAR AND VECTOR FIELDS 405

For greater accuracy we may integrate system (4) to find the flow-lines of the vector field.
The term "flow-line" is motivated by the case where F represents the velocity field in a
fluid and the integral curves are the streamlines. Physically, these are the paths followed
by a light particle that drifts with the fluid.
For vector fields F(x,y, z) in £3, the direction fields lie in £3 and are much harder to
display. Computer graphics can be used to construct perspective views of the direction
fields. Skill is needed to produce helpful diagrams. The flow-lines can also be constructed
and displayed in the same manner. The simple example

dx
(5) - = a(y -x)
dt '
dy
- =px-y-xz (a, p, f3 constants)
dt '
dz
-=-f3z+xy
dt
shows how complicated the flow-lines can become. System (5) is the well-known Lorenz
system, whose flow lines exhibit chaotic behavior (see [Sch]). A computer-generated
trajectory is shown in Figure 21, with a = lO, p = 28, and f3 = 8/3.

FIGURE 21. The Lorenz attractor

406 6 VECTOR ANALYSIS

Directional Derivatives and Gradients. Let fer) denote a scalar field whose
first -order partial derivatives

a~~) = fAr), a~~) = fyCr) , a~~) = fz(r),

exist and are continuous. Let u be a unit vector. Then df(ro + tu)ldt is the derivative
along the line r = ro + tu. In particular, we shall write

(6) Df(ro, u) = dd f(ro + tU)1

t t=O

and call (6) the directional derivative of fer) at ro in the direction u. An important
theorem of vector analysis states that (6) is a linear function of u, namely,

(7)

In fact, (7) is a consequence of the chain rule of calculus. We shall also write

(8) gradf(r) = fx(r) i + fyCr) j + fz(r) k

and call grad f(r) the gradient off at r. With this notation the directional derivative can
be written as the scalar product

(9) Df(r, u) = grad f(r) . u.

The function that assigns to each r the vector grad f(r) is a vector field. It is called the
gradient field of f(r).
Interpretation of the Gradient Vector. If gradf(r) = 0, then DfCr, u) = 0 for all
unit vectors u. If gradf(r) -=1= 0, then we may write for lui = 1,

(10) Df(r,u) = Igradf(r)lcose,

where e is the angle between gradf(r) and u CO ~ e ~ Ji). In particular, for e = 0 we
see from (0) with cose = 1 that
gradfCr)
u = ~---':.--
I gradfCr)I
is the direction of maximum increase of fCr) at r, while I grad fCr) I is the maximum
rate of increase. Similarly, on taking e = Ji we see from (0) with cos e = -1 that
gradfCr)
u= - ~---':.-,.....
I gradf(r)I
is the direction of maximum decrease of fer) at r. Finally, if e = Ji/2 then (9), (10)
imply that u is orthogonal to grad fCr).
 6.4 CALCULUS OF SCALAR AND VECTOR FIELDS 407

Normal Vectors to a Level Surface of a Scalar Field. Let j(r) be a scalar field
and consider a point ro where

grad j(ro) ::j= O.

Let S be the level surfacej(r) = j(ro). We wish to show that

(ll)

is a normal vector to the level surface S. To do this, let res) be a parametric curve with
arc length parameter s, chosen in such a way that

reO) = ro, r'(O) = T,

where T is any unit tangent vector to S at roo Then Jeres)) = j(ro) and the chain rule gives

(12) :J(r(s))is=o = gradj(ro) . T = O.

Since T can be any tangent vector to S at ro, we see that (11) is a normal vector to S at
roo This result often gives an easy way to find a normal vector to a surface.
The Divergence Operator and the Divergence Field. Consider a vector field F
defined in a domain D in E3 , and write

(13) F(x,y, z) = Fl (x,y, z) i + F2 (x,y, z) j + F3 (x,y, z) k.

Let S be a smooth surface in D with unit normal field n(x,y, z), and consider the surface
integral

(14)

defined as in Section 6.3, equation (37).

Orientable Surfaces. There are two possible choices for the unit normal n(x,y, z),
each the negative of the other. We shall assume that n can be chosen so as to be continuous
on S. This makes S an orientable surface and eliminates such nonorientable surfaces as
the well-known Mobius strip. An orientable surface is oriented by choosing the side of
S on which n lies. This side will be called the positive side, the other side being called
the negative side.
An Interpretation of the Oriented Surface Integral (14). We wish to show that

I
if F is the velocity field of fluid and S is a smooth oriented surface then
volume of fluid per unit time crossing}
(15) fIs F . n dA = ~;; the negative side to the positive side
408 6 VECTOR ANALYSIS

It is to be understood that if the integral (14) is negative then the net flow is from the
positive to the negative side of S.
Derivation of (15). Partition S into a union of small surface elements and consider the
flow across a typical element ~S. The volume of fluid crossing ~S in time ~t fills a
cylinder of height F . n ~t; see Figure 22.
Thus, writing ~A for the area of ~S,

(F . n) ~A ~t = net volume crossing ~S in time ~t,

and

(F· n) ~A = net volume per unit time crossing ~S.

Equation (15) now becomes apparent if we sum over all surface elements of the partition
and make the norm of the partition tend to zero.
Net Flow out of a Parallelepiped. Let us consider a small parallelepiped V in the
domain of the vector field F:

V = (cx,y, z) : 0 ~ x - Xo ~ ~x, 0 ~ y - Yo ~ ~y, 0 ~ z - Zo ~ ~z} .

We shall calculate the net flow rate per unit volume out of V, in the limit where

~X2 + ~i + ~Z2 ~ O.

The boundary S of V consists of six rectangles, in the coordinate planes x = Xo, x =

Xo + ~x, y= Yo, Y = Yo + !1y, z = Zo, Z = Zo + !1z, with exterior unit normals n = -i,
n = i, n = -j, n = j, n = -k, n = k, respectively; see Figure 23.
Thus for small ~X2 + ~y2 + ~Z2 we have

fIs F· ndS;:;:; [F1(xo + !1x,yo,Zo) - F1(xo,YO,Zo)] ~y~z

+ [F 2(xo,YO + ~y, Zo) - F2(xo,yo, Zo)] ~x~Z
+ [F3(xo,YO,zO + ~Z) - F3(xo,yo,Zo)] ~x~y

FIGURE 22. Volume crossing AS in time At

 6.4 CALCULUS OF SCALAR AND VECTOR FIELDS 409

FIGURE 23. Parallelepiped with exterior unit normals

and if IV I = volume of V = ~x~y ~z then

-1
IVI
11 5
F·ndS~-+-+-.
~Fl
~x
~F2
~y
~F}
~z

It follows that if ~X2 + ~y2 + ~Z2 """"* 0 then

(16) . -1
hm
IVI ..... O IVI
11 5
oF=
F·ndS
l
-oF
+-+
2 of}
-
ox oy oz '
evaluated at (xo,Yo, zo)· This last expression is called the divergence of Fat (xo,Yo, zo)
and it is written

.
dI OFI OF2 of}
(17) vF=-+-+-.
ax ay az
Equation (16) implies that the net flow rate per unit volume out of parallelepiped V is
div F(xo,yo, Zo). Note that the divergence operator diy, defined by (17), carries a vector
field F into a scalar field div F. Also, from (16), div F(xo,yo, zo) > 0 means that the fluid
is expanding near (xo,Yo, zo), while div F(xo,yo, zo) < 0 means that it is compressing.

The Curl Operator and the Curl Field. let C be a closed curve lying in the domain
D of F and consider the line integral

(18) l F . dr = l F . T ds,

where s is the arc length parameter on C and T = dr/ds is the unit tangent vector.
Integral (18) is called the circulation of F around C. If F is the velocity field of a fluid
then the circulation of F measures the rate at which the fluid is rotating. Note that for
fixed magnitude IFI the circulation (18) is largest when F is parallel to T, while it is zero
when F is orthogonal to T.
410 6 VECTOR ANALYSIS

The circulation is always zero for a gradient flow F = grad f. To see this, note that
by the chain rule,
dr
F . T = gradf . ds
= afdx + afdy + afdz
ax ds ay ds az ds
df
= ds'
and hence if £ is the length of C, then

1 1
C
F·dr=
1df
-ds
0 ds
= f(x(£),y(£), z(e) - f(x(O),y(O), z(O»
=0

because C is closed.
Circulation Around a Small Rectangle. We shall estimate the circulation around
a small rectangle Cz in the (x,y)-plane with venices

(xo,Yo, zo), (xo + ~x,Yo, zo), (xo + ~x,yo + ~y, zo), (xo,Yo + ~y, Zo).
Rectangle Cz is oriented counterclockwise as seen from the positive z-axis; see Figure 24.
For ~X2 + ~y2 small we can estimate

I c,
F. dr ~ [Fl(xo + ~X,Yo,Zo) - Fl(xo + ~X,Yo + ~y,Zo)] ~x
+ [F 2 (xo + ~X,Yo + ~y, zo) - F2 (Xo,Yo + ~y, Zo)] ~y,
so if Sz = ~x~y = area enclosed by Cz , then
~ 1
Sz c,
F· dr ~ aF2
ax
_ aFl.
ay

Y
x
(Xo + t:.x, Yo + t:.y , <0) FIGURE 24. Oriented rectangle Cz
 6.4 CALCULUS OF SCALAR AND VECTOR FIELDS 411

Passing to the limit as .6.x 2 + .6.y2 -+ 0 gives

(19) . -1 F . dr =
hm
5,-+0 Sz
i Cz
-aF - - .
ax
2 aF l
ay
Similarly, with rectangles Cx (resp. Cy) in the (y, z)-plane (resp. (x, z)-plane) we get

.
hm -
1 i F·dr= - - - ,
aF3 aF2

i
Sx
5x -+O Cx ay az
. -1 F . dr = -aF - -
aF.
l 3
hm
5y --+0 Sy c, az ax
These three limits determine the vector field curl F, defined by

(20) curl F = ( aF3 _ aF2) i + (aF l _ aF3)j

ay az az ax
+ ( aF2 _ aFl) k.
ax ay
In Section 6.5, after Stokes' theorem has been established, we shall be able to show that
if C is a rectangle with a vertex at (xo,yo, Zo) and with unit normal n, then

lim ~ { F· dr = (curl F) . n,
5--+oS1c
where 5 is the area of C.

EXAMPLE 1. A Rotating Fluid. Consider a fluid that is rotating about the z-axis with
constant angular velocity w radians per second. The fluid velocity F = F (r) at each point
in space is determined by the vector w = wk; see Figure 25.
We have

IFI = wlrl sine = Iwllrl sine = Iw x rl·

o FIGURE 25. Fluid rotating with constant angular velocity

412 6 VECTOR ANALYSIS

Moreover, F and w x r have the same direction, by the right hand rule; see Figure 25. Thus

(21) F =W x r

gives the velocity field in a fluid that rotates as a rigid body about the z-axis with angular
velocity w. A simple calculation, using (20), gives

(22) curl F = 2w.

The factor 2 is correct. Thus the rotation vector w is one-half the curl of the velocity field.
The Del Notation. An alternative notation for the differential operators grad, div
and curl is provided by the vector differential operator
a a a
(23) V=i-+j-+k-.
ax ay az
The symbol V is called "del" or, in some texts, "nabla." In applying V to a scalar or vector
field it is understood that the differentiations are applied to all terms to the right of them,
unless they are delimited by parentheses. With this convention we have the equivalence

(24) Vj = i aj + j aj + k aj = gradf.
ax ay az
Furthermore, if

then

(25)

and

(26)

= curl F.

In addition, we have the operator identities

(27) V x (Vf) =0
and

(28) V· (V x F) = 0,
 6.4 CALCULUS OF SCALAR AND VECTOR FIELDS 413

provided that I and F have continuous second derivatives. Thus, to verify (28), we have

az
ay - aF2) + aya (aFr ~
v . (V x F) = axa (aF3 aF3) az -
+ ~ (aF2 _ aFr)
az ax ay
a2F3 a2F2 a2Fr
=---+-
axay axaz ayaz
a2F3 a2F2 a2Fr
--+---
ayax azax azay
=0,

provided that the mixed second-order partials are equal. This holds if the second-order
derivatives are continuous.
Additional Operator Identities. The following identities are often useful in vector
calculations. They can be proved by straightforward calculation; see page 155 in [D-S].

(29) V (fg) = IVg + gVf.

(30) V· (f F) = I V . F + VI . F.
(31) V x (fF) = I V x F - VI x F.
(32) V . (F x G) = G· (V x F) - F . (V x G) .
(33) V x (F x G) = (G . V) F - (F· V) G. + (V . G) F - (V . F) G.
(34) V x (V x F) = V (V· F) - V2F.
(35) V (F . G) = (F . V) G + (G . V) F + F x (V x G) + G x (V x F) .

Exercises 6.4

Scalar Fields in R2. For each exercise, construct from the given scalar field two figures: (1) the surface,
and (2) a contour diagram.

1. The paraboloid equation z = x 2 + 4yl, -1 ::: x ::: 1, -1 ::: y ::: 1. Plot the curves of constant altitude
z = 0.2 to 1.8 in steps of 0.4.
2. The pendulum energy equation E = 1 - cose + y 2/2, -rr/2 ::: e ::: rr/2, 0 ::: y ::: 1. The variable e is
the angle the pendulum makes with the venical, while y is the velocity of the pendulum bob. Plot the
curves of constant energy E = 0 to E = 1.4 in steps of 0.2.
3. The saddle surface z = x 2 -yl, -3::: x::: 3, -3::: y ::: 3. Plot the curves of constant altitude for
z = -4 to 4 in steps of 2.
4. The temperature function T = 100 - x 2 - y2 on the plate -10 ::: x::: 10, -10 ::: y ::: 10. The T-axis
scale is degrees Celsius. Plot the isothermal curves for T = 0 to 100 in steps of 25°.
414 6 VECTOR ANALYSIS

top

B FIGURE 26. A topographical map

5. The traveling wave h = 4 + cos(x - tl2), 0 ::: x ::: 2n, 0 ::: t ::: 2n, where x is position and t is time.
Plot the curves of equal displacement for h = 3 to h = 5 in steps of 0.5.
Topographic Maps. The hill in Figure 26 has two walking paths to the top. A similar but different hill is
pictured in Figure 27. The following exercises discuss the topographic map and the 3-D figure of the nearby
hill.
6. Which path (A or B) in the topographic map of Figure 26 is the steepest climb? Explain your reasoning.
7. Make a rough 3-D model of the hill directly from the topographic map of Figure 26.
8. Make a rough contour map with three curves (z = 100, 150, 200) using the 3-D figure of the nearby
hill; see Figure 27. Assume that the contours were generated fromy = O.Olzsin(nx) with z fixed.
Flux of a Constant Field. For each exercise below, find the flux for the constant field v = i + j + 4k
through the given surface.
9. A disk of radius 3 in the (x,y)-plane with normal k.
10. A square of side 1 in the plane z= 1 with normal k.
11. A rectangle with vertices (6,0,0), (6,1,0), (6,1,2), (6,0,2) oriented in the positive x-direction.
12. A tilted rectangle with vertices (1,0,2), (1,1,0), (0,1,0), (0,0,2) and normal in the direction of
(j - 2k) x (-i).
13. A triangular plate of area 6 in the (y, z)-plane with normal i.
14. A triangular plate with vertices (1,0,0), (0,1,0), (0,0,1) and normal pointing away from (0,0,0).

FIGURE 27. Anearby hill

 6.4 CALCULUS OF SCALAR AND VECTOR FIELDS 415

Flow Lines of Vector Fields. In the exercises below, solve for the parametrized curve ret) satisfying
r'(t) = F(r(t».
15. F(x,y) = 2i + 3j, reO) = 2i + 4j.
16. F(x,y) = xi + yj, reO) = 2i + 4j.
17. F(x,y) = y 2 i + 2xj, reO) = i +j. Find ret) for t = 0 to t = 1 in steps of 0.2 by Euler's numerical method.
Tangent Lines and Tangent Planes. In the exercises below, find the equation of the tangent line or the
tangent plane to the given curve or surface using gradient methods for level curves and level surfaces.
18. Find the tangent line to x- i = 4 at x = 5, Y = 1.
19. Determine the equation ofthe tangent plane to the surface 2x2 +3i = liz at the point X = l,y = -1,
z = 115.
20. Determine the equation of the tangent plane at (-1,1,1) for the surface x 2 + y4 + x 2 Z2 = 3.
21. (Electric Field) The electric field produced by a charge placed at the origin is given by E = (xi + yj +
zk)/(x 2 + i+ Z2 )3/2. Find div E.

22. (Magnetic Field) A magnetic field B must have zero divergence. Is it possible that B = yi-xj+(x+y)k
is a magnetic field?
23. (Point Source) Let F = j(r 2 )(xi + yj), where r2 = x 2 + i
andj(u) = l/u. Show that F has a point
source at the origin, i.e., the divergence is zero away from (0,0), and its direction is always away from
(0,0)

24. (Point Sink) Let F = j(r2) (xi + yj), where r2 = x 2 + i

andj(u) = -l/u. Show that F has a point
sink at the origin, i.e., the divergence is zero away from (0,0) and its direction is always towards (0,0).
25. (Leaf Motion) It is reasonable to expect that a leaf placed in a fluid with zero velocity will remain
motionless. Where should a leaf be placed for the field F = -yi + xj in order to remain motionless?
Can you see this from the field plot?
26. Orrotational Field) A vector field is said to be irrotational if its curl is zero. Explain in words what
this means, in terms of a test paddle wheel inserted into a fluid.
27. (Floating Boat) A boat is floating down a stream. The stream makes a 45° tum and then continues
straight. Explain in terms of fluid velocities why the boat must go through a 45° rotation.
Curl of a Vector Field. In the exercises below compute the curl of the vector field v.
28. v = (x 2 - i)i + 2xyj.
29. v = (xi + yj + zk)/(x 2 + y2 + z2)lI2.
30. v = (y - xli + (y + z)j + (z - x)k

31. v = x 2i + i j + z4 k
32. v = yzi + xzj + xyk
33. v = 3yzi + 5xzj + 9xyk

34. v = e"i + yj +xzk

35. v = sinCy)i + cos(x)j + xzk
36. v = -wyi + wxj.
37. (Harmonic Functions) Let u be a solution of the partial differential equation Uxx + Uyy + u zz = 0 and
let v = 'V u. Show that curl v = 0 and div v = o.
416 6 VECTOR ANALYSIS

amount
8000
...... r---
......

-
level curves
r--- ...... ......r--
r-- ...... ...... r--" for payment P
6000
-r-- -r-- -r-- -,...
-,... r--_
r--_ p= 140
r--_ p= 120
r--_I- I-r-- - l -
P = 100
4000
I--
-I- -,... 1'--
1'--1- 1-- P=80

2000
-I-
- P = 60
3 5 7 9 11 13 15 interest

FIGURE 28. Interest rate versus loan amount

38. (Curl Identity) Apply the relation curl(w x r) 2w, valid for r = = xi + yj + zk, to solve the equation
curl v = 2i - 4j + 12k forthe vector v = v(x,y, z).

39. (Curl Identity) Let v be a vector field with continuous second partials. Show that div(curl v) = O.
40. (Loan Payments) In Figure 28, the contour graph shows the monthly payment and loan amount on a
5-year car loan. Assume a loan amount of $6000 dollars at 14% interest. If the car is refinanced at 9%
within the first 30 days, then by how much is the monthly payment reduced, approximately?

41. (Fluid Flow) The fluid velocity flow F = 2x i + y j + z k acts in the cube V described by 0 ~ x ~ 1,
o ~ Y ~ 1, 0 ~ z ~ 1. Find the volume of fluid per unit time flowing out of V.
=
42. (Fluid Flux Across a Surface) The fluid velocity flow F xj + z k acts on the surface 5, which is the
=
portion ofthe plane 3x + 2y + z 6 in octant I (x,y,z > 0), oriented by the vector 3i + 2j + k. Find
the volume of fluid per unit time flowing across 5.

43. (Heat Flux) Let 5 be the surface bounded by the spherex2 +i +z2 = 16 with n the unit outer normal
i
to 5. Given the temperature function T = x 2 + + Z2 and the temperature gradient F = -'VT, find
the heat flux out of the sphere,

44. (Fluid Flux Across a Surface) The fluid velocity flow F = ecosx i + j + eX k acts on the surface 5 2i
given by r = ui + vj + v2 k on the rectangle D: 0 ~ u ~ 3, 0 ~ v ~ 2. Find the volume of fluid per
unit time flOwing across 5 in the direction of N ru x rv. =
45. (Net Flow) Let 5 be the surface bounded by Z = x 2 + i and z= 1 with n the unit outer normal to
5. Find the net flow

of the fluid velocity field F = i i + x 2 j + 2z k across 5 in the direction of n.

 6.5 INTEGRAL THEOREMS OF VECTOR CALCULUS 417

6.5 Integral Theorems of Vector Calculus

The Fundamental Theorem of Calculus states that if j(x) and l' (x) are continuous on
an interval a .::; x .::; b then

(1) lb 1'(x) dx = j(b) - j(a).

The goal of this section is to extend the fundamental theorem to integrals of scalar and
vector fields over curves, surfaces and volumes in E2 and E3. Several theorems will be
given that involve V j, V . F and V x F.

A Gradient Theorem. Let C r = ret), a.::; t .::; b, be a regular (= sectionally smooth)

oriented curve in E3 and let j(r) be a scaJar field with gradient V j(r) that is defined and
continuous in a domain D containing C. Then

(2) i V j(r) . dr = Jereb)) - j(r(a)).

For the proof, the chain rule gives

d
(3) dJ(r(t)) = V j(r(O) . r' (0.

Integrating (3) over a.::; t .::; b and using (1) gives (2).

Green's Formulas. It was discovered in 1828 by the English mathematician and

physicist George Green (1793-1841) that if Q(x,y) and Qx(x,y) are continuous in a
domain D of E2 with boundary C then

(4) fin Qx(X,y) dxdy = [ Q(x,y) dy.

Here C is an oriented curve in E2 and it is required that the orientation be chosen to

make domain D lie to the left as C is traced in the positive sense. Similarly, Green found
that if P(x,y) and PyCx,y) are continuous in D, then

(5) fin Py(X,y) dxdy = - L P(x,y)dx,

where C is as before. If both (4) and (5) hold, then by subtracting (5) from (4) one gets
the formula

(6)
418 6 VECTOR ANALYSIS

Here we shall verify (4) and (5) for some domains of simple type. Then we shall extend
the class of domains for which (4) and (5) hold to a larger class that includes most of the
cases that are needed in applications.
x-Domains. A domain D in E2 is said to be an x-domain if and only if there is an
interval a :'S y :'S b and smooth functions g(y) and h(y) such that points in D satisfy

g(y) :'S x :'S h(y) for a:'S y :'S b;

see Figure 29.

If oriented curves C1 , C2 , C3 , C4 are defined as in Figure 29 then

is the oriented boundary of D. The four components of C have the following parametric
representations, correctly oriented.

(7) C1 : x = h(t), y = t, a:'S t :'S b,

C2 : x = - t, Y = b, -h(b):'S t:'S -g(b),
C3 : x = g(-t), y = -t, -b:'S t:'S -a,
C4 : x = t, Y = a, g(a):'S t :'S h(a).

We can now prove the following theorem.

Theorem. Green's formula (4) holds for x-domains.

Proof. We shall begin with the left-hand side of (4) and show by direct calculation that
it equals the right-hand side. The first step is justified by the standard calculus method
of evaluating double integrals as iterated integrals. Thus

(8) fi Qxdxdy = 1b (l:~) Qx(X,Y)dx) dy

= 1b [Q(h(y),y) - Q(g(y),y)] dy.

b
C2
...
C3 19(Yl Cl J~h(Yl
a
C4
'" x
FIGURE 29. An x-domain
 6.5 INTEGRAL THEOREMS OF VECTOR CALCULUS 419

Next we have the four integrals

I) = [ Q(x,y)dy, 12 = [ Q(x,y)dY,h = [ Q(x,y)dy, 14 = [ Q(x,y)dy.

lc) lC 2 lC 3 lC 4

Then

(9) 12 = 14 = 0,
I) = lab Q(h(t), t) dt, 13 = -lab Q(g(r), r)dr,
because y = constant on C2 and C4 implies h = 14 = 0, while
I) = lab Q(x(t),y(t))y'(t)dt = lab Q(h(t), t)dt,
and

13 = i-a -b
Q(x(t),y(t))y'(t)dt = i-a
-b
Q(g( -t), -t)( -l)dt

= l a
Q(g(r), r)dr = -lab Q(g(r), r)dr.
Combining (8) and (9) gives (4).

y-Domains. In analogy with x-domains we shall say that a domain D in E2 is a y-

domain if there is an interval a :s x :s b and smooth functions g(x) and hex) such that
points in D satisfy

g(x) :s y :s hex), a:s x :s b;

see Figure 30.

Proceeding in analogy with the case of x-domains we can prove the following theorem.

y = hex)
y

y = g(x)

a b FIGURE 30. Ay-domain

420 6 VECTOR ANALYSIS

Theorem. Green's formula (5) holds for y-domains.

The proof will be left as an exercise. Careful work will verify that the minus sign
that appears in (5) is correct. In fact, it is a consequence of the way we have oriented the
boundary C of D.
The restriction to x-domains and y-domains is quite severe. Here we shall define a
much larger class, called regular domains, for which Green's formulas also hold.
Definition of Regular Domains. For the definition, Cl, C2 , ... , Cn denote n disjoint
simple closed curves such that C2 , ... , Cn lie inside C1 . The domain D then consists
of the points that are inside C1 and outside C2 , ... , Cn. The domain D is said to be a
regular domain if it has this form and each curve is regular (= sectionally smooth).
Orientation. In accordance with Green's rule we orient C1 , C2 , ... , Cn such that C1 is
counterclockwise and C2 , ... , Cn are clockwise. With this understanding we have the
following theorem.

Green's Theorem. Let D be a regular domain and let Q, Qx, P, and Py be continuous
in D. Then Green's formula (6) holds with

A formal detailed proof will not be given. However, the idea of the proof is Simple. We
must show that each regular domain is the union of a finite number of nonoverlapping x-
domains. Then equation (4) will hold for each x-domain. Adding these equations gives (4)
for D. To get (5) for D we represent D as a finite union of y-domains. Note that when two
x-domains, or two y-domains, share a common boundary arc, then the corresponding
terms in the integral over C cancel in pairs because they occur with opposite orientations.

In specific cases it is usually easy to write a regular domain as a finite union of x-

domains or y-domains. An example is depicted in Figure 31, where the horizontal lines
decompose D into seven x-domains.

y
.............. ·0···
........ ·.. ·
C2
.....................................
...... ·· .. 0·····
.. ·· .... ·.. ··· .. ····
C3
............................

L----------~x FIGURE 3l. Decomposition of a regular domain into x-domains

 6.5 INTEGRAL THEOREMS OF VECTOR CALCULUS 421

Several formulas of vector integral calculus can be derived from Greens theorem by
suitable choices of the functions P(x,y) and Q(x,y). Two will be derived here.
The Divergence Theorem in E2. Let D be a regular domain in E2 and let
(0)

be continuous and have a continuous divergence

oF l oF2
\7·F=-+-
ox oy
in D. Then

(ll) flo \7. FdA = 1 F· nds,

where C is the oriented boundary of D, and n is the exterior unit normal vector to D.
For the proof we apply Green's theorem with Q = Fl and P = -F2 . Then in Green's
formula (6) we have
oF oF
Qx - Py = ax-
l
+ ay = \7 . F,
2

which gives the left-hand side of (11). For the right-hand side we have

dy dx)
(12) Pdx + Qdy = ( Fl ds - F2 ds ds.

Now, writing
dx dy
= r (s) = -ds i + -ds j
!
T

for the unit tangent vector to C with positive orientation, we have

dy dx
n=Txk=-kxT=-i--j
ds ds'
and hence nl = dy/ds, n2 = -dxlds. Then (12) becomes
(13)

and Green's formula gives (11).

If in Green's formula we take P = F l , Q = F2 and define the vector field F in E2 by
(10) as before, then we may write

(\7 x F) . k =-
aF2 - -oF l = \2x - P .
ox oy y
422 6 VECTOR ANALYSIS

Moreover,

Pdx + Qdy = ( F 1 -dx + F2-dY ) ds = (F· T) ds.

ds ds

Thus we have the following theorem.

Stokes' Theorem in E2 • Let D be a regular domain in E2 and let '\1 x F be continuous
in D. Then

(14) fi ('\1 x F) . k dA = [(F, T) ds.

The proof is an immediate consequence of Green's theorem. We shall see below that (14)
is a special case of Stokes' Theorem in E3 .
The Divergence Theorem in E3 . This theorem, also called Gauss' theorem, is the
three dimensional analogue of equation (11). Thus it states that

(15) ffi '\1. F dV = f1 F . n dS,

where D is a domain in E3 , 5 is its oriented boundary surface with outward unit normal
vector field nand F is a vector field on D. We shall seek conditions on D and F that
are sufficient to guarantee that (15) holds. As preparation let us note that if Cartesian
coordinates are introduced, so that

then (15) is equivalent to the equation

(16)

Moreover, the functions FI, F2 and F3 can be chosen independently, and hence (16) is
equivalent to the three equations

(17) f f i C:l) dxdydz = f1 (F1nl)dS,

(18) f f i C~2) dxdydz = f1 2 (F n2)dS,

(19) f f i C::) dxdydz = f1 (F 3n3)dS.

 6.5 INTEGRAL THEOREMS OF VECTOR CALCULUS 423

The divergence theorem (15) is easy to prove for some simple domains D such as
parallelepipeds, cylinders and spheres. We shall prove it for a large class of domains that
includes those that usually occur in applications.
Definition. A domain D in E3 is said to be a z-domain if there exist a domain R in E2
and sectionally smooth functions g(x,y) and h(x,y) with g(x,y) ~ h(x,y) for all (x,y)
in R such that

(x,y, z) is in D if and only if(x,y) is in Rand g(x,y) ~ Z ~ h(x,y).

Theorem. Equation (19) holds whenever D is a z-domain and the component F3 (x,y, z)
is a smooth function in D.

Proof. From the definition of z-domain, the left-hand side of (19) is

(20) II JD{( f3) i1

3z
dxdydz = IJD((fhCX'Y)
g(x,y)
3F3 dZ) dxdy
3z

= Ik F3(x,y,h(x,y»dxdy

-Ik F](x,y,g(x,y»)dxdy.

Let us compare this with the right-hand side of (19). Note that the boundary 5 of D
consists of three parts:

See Figure 32.

Thus

(2l) / 1 F3 n3 d5 =/ L F3 n3 d5 + / 1, F]n3 d5 + / LF3 n3 d5.

Next we calculate n3d5 for each of the three components 51, 52 and 53. For 51 we have
the parametric representation

51 : x = u, y = v, Z = h(u, v) with (u, v) in R.

Writing

r(u, v) = ui + vj + h(u, v) k,
we get
424 6 VECTOR ANALYSIS

o
FIGURE 32. A z-domain

so

Next,

d5 = Iru x rvldudv = Jh~ + h~ + 1 dudv,

so

(22)

Proceeding in the same way we get

52 : x = u, y = v, z = g(u, v) with Cu, v) in R.

The only difference for 52 is that now n3 < 0 and

(23)
 6.5 INTEGRAL THEOREMS OF VECTOR CALCULUS 425

Next,

(24)

because n3 = 0 on 53. Substituting (22), (23), (24) in (21) gives

(25) fiF3n3d5= ffRF3(U,v,h(u,V»dUdV- fkF3(U,V,g(u,V»dUdV.

Finally, (20) and (25) give (19), completing the proof.

Corollary. The preceding theorem extends immediately to domains D that are finite
unions of z-domains. We shall not develop this in detail here, but it is clear that in such
a union, where portions of the boundary surfaces coincide, these portions occur in pairs
with opposite orientation and thus cancel from the boundary of the composite domain.

x-Domains and y-Domains. These special domains are defined in strict analogy
with z-domains: Just permute the coordinates x, y, Z in an obvious way. Thus we see
that equation (17) holds in unions of x-domains, and equation (18) holds in unions of
y-domains. This suggests the following definition.
Regular Domains in E3 . A domain D in E3 will be said to be a regular domain if
and only if

D is a finite union of x-domains, and

D is a finite union of y-domains, and

D is a finite union of z-domains.

With this terminology we have the following result.
Criteria for the Validity of the Divergence Theorem. Let D be a regular domain
in E3 and let F be a smooth vector field in D. Then (15) holds:

f f In V· F dV = f iF, n d5,

where n is the outward unit normal on 5.

The class of regular domains is very large and includes most of the domains that arise
in applications. It is interesting to search for domains with sectionally smooth boundaries
that are not regular, but we shall not pursue this further here.
Stokes' Theorem in E3 • This theorem, the last vector integral theorem to be discussed
here, is named for its discoverer, the Irish mathematician and physicist Sir George Gabriel
426 6 VECTOR ANALYSIS

FIGURE 33. Orientation of 5 and C for Stokes' theorem

Stokes (1819-1903). Briefly, the theorem states that

(26)

where 5 is an oriented surface in E3 with positive unit normal vector nand C is the
oriented space curve that bounds S.
It is understood that C is oriented such that an observer standing at a boundary point
with his hand in the direction of n and facing in the positive direction on C will find the
surface 5 to his left; see Figure 33.
In (26), F (x, y, z) denotes any smooth vector field that is defined in a neighborhood
of S. In words, (26) says that the flux of V x F across 5 Cleft-hand side of (26) equals
the circulation around C (right-hand side of (26)).
The validity of Stokes' theorem for plane surfaces 5 was shown above using Green's
theorem. Complete proofs for the general case are long. The details will not be given
here. Instead, a beginning of a proof will be given from which interested readers can
complete the proof. Details can be found in [D-S]; see Figure 34 for notation.
The idea of the proof is to use the chain rule

dr = ru du + rv dv

(x. y. 7) -space

(II . u)-plane
FIGURE 34. Notation for the proof of
Stokes' theorem
 6.5 INTEGRAL THEOREMS OF VECTOR CALCULUS 427

to write

iF, i dr = (F· ru) du + i (F· rv) dv

and then use Green's theorem in the (u, v)-plane to get

Jer F· dr = JEr [~
au
f(F . ru) - ~ (F . rv)]
av
du dv.

The remainder of the proof is to simplify the last integral to get (see page 271 in [D-S])

f1 (V' x F) (ru x rv) dudv = fis (V' x F)· ndS.

Line Integrals. Convert each line integral below by means of a parameterization of C into an ordinary
calculus integral and then evaluate.

1. Ie x dx + x 2 dy +l dz, ret) = t(i + j + k), 0 :::: t :::: 1.

2. fc(x + y)dx + dy, ret) = t2 i + tj, 0:::: t :::: 1.

3. IeYdx, ret) = costi+ sintj, 0:::: t:::: 11:.

4. Ie dx + dy, ret) = cos Ii + sin Ij, 0 :::: t :::: 2Jr.

5. Ie ~:~L ret) = cos ti + sin tj, 0 :::: t :::: Jr.

6. Ie e" dy, r(t) = ti + t2j, 0 :::: I :::: 1.
7. Ie sin x dy, r(t) = ti + sin tj, 0 S t S Jr.

8. Ie tanydx, ret) = sin Ii + Ij, 0 S t :::: Jr/4.

9. Ie l~~i~X dx, ret) = Ii + tj, 0 SIS Jr12.

Fundamental Theorem for Line Integrals. Use the result

[ 'V f . dr = feb) - f(a),

valid for curves C connecting a to b, to evaluate the following line integrals.

10. Iexdx + (y+ l)dy, Cjoins (0,0) to (1, 1).
11. Ie (sin y - y sin(xy»dx + (x cos y - x sin(xy»dy, C joins (0,0) to (1, JrI2).
12. Ie y dx + (x + l)dy, C joins (0,0) to 0, 1)
I3. Ie y dx + x dy, C joins (0,0) to (1,2)
Oriented Boundary Curves. In the following exercises, find a vector parameterization for each curve,
including its parameter domain, oriented as prescribed.
428 6 VECTOR ANALYSIS

14. The line segment from P = (1,1,0) to Q = 0,3,5).

15. The line segment joining 0,2, -1) to the nearest point of the plane x + y + z = 1.
16. The circle x 2 +i = 16 in the plane z= 0, oriented counterclockwise when viewed from above.

17. The circle x 2 + z2 = 16 in the plane y = 1, oriented counterclockwise when viewed from (0,2,0).
18. The ellipse x 2 + 4i = 4 in the plane z = 0, counterclockwise as viewed from the point (0,0,1).

19. The ellipse i + 4z 2 = 8 in the plane x = 2, counterclockwise as viewed from the point (4,0,0).
20. The square bounded by the lines x = ±1, Y = ±1 in the plane z = 0, counterclockwise when viewed
from (0,0,1).

21. The boundary of the triangle determined by the portion of the plane x + y + z = 1 in the first octant,
counterclockwise when viewed from (2,2,2).
Decomposition of a Regular Domain in E2. In the following exercises, decompose the given region D in
E2 into a finite number of x-domains or y-domains. Report a parameterization for each subdomain, including
its outward normal n. The decomposition is to be a regular domain, which implies careful definition of the
normals.

22. The set D is the region in E2 bounded by the x-axis, the line x = 1, and the curve y = x 3
23. The set D is the unit square in the first quadrant.

24. The set D is the rectangle bounded by x = 0, x = 1, Y = 0, Y = 2.

25. The set D is the disk at (0,0) with radius 1 in the upper half-plane plus the region inside the rectangle
y = 0, y = -1, x = -1, x = 1 in the lower half-plane.
26. The set D is the circle with center 0, -2) and radius 4.
27. The set D is the region remaining when the triangle bounded by x = 0, x = 2 andy = x is removed
from the square bounded by x = 0, x = 3, Y = 0, Y = 3.
Decomposition of a Domain in E3. In the following exercises, decompose the given domain D in E3 into
a finite number of z-domains. Report with each z-domain the outward normal n to the boundary surface.
28. Domain D is the parallelepiped bounded by the planes x = 0, x = 1, Y = ±1, z = ±l.

29. Domain D is the parallelepiped bounded by the planes x = 0, x = 1, Y = ±1, z = ±2.

30. The domain D is the portion of the cylinder y2 + z2 ::: 1 cut off by the planes x = 0 and x = I.

31. The domain D is the region between the cone 4x 2+4i = Z2 and the circular cylinder (x-l)2 +y2 = 1
in the half-space z ::: o.
Green's Theorem. In the following exercises, compute directly the two sides of Green's formula

h= fl(Qx-Py)dxdY , 12= ipdx+QdY

and verify equality of II and 12. In each let D be the set x 2 +i ::: 1 and take C to be the circle ret) =
cos t i + sin t j, 0 ::: t ::: 2T(.
32. P = y, Q = -x.
33. P = -y, Q = 0
 6.5 INTEGRAL THEOREMS OF VECTOR CALCULUS 429

34. P =x 2y, Q =xy2

35. P = x 2y, Q = _xy2
Divergence Theorem in E2. In the following exercises, compute directly the two sides of the Divergence
theorem formula

h = fin v· FdA, 12 = iF, n ds

and verify that hand 12 equal. In this formula, C is the oriented boundary of D and n is the outward unit
normal vector to D.
36. The vector field is F = (y - x)(i - j) and D is the square bounded by x = 0, x = 1, Y = 0, Y = l.
37. The vector field is F = (4y +x 2 )i +ij and Dis the square bounded byx = O,x = 1,y = O,y = l.
38. The vector field is F = (x + y)i + x 2j and D is the triangle bounded by y = 0, x = 2, Y = x.
39. The vector field is F = xi + (x + y)j and D is the triangle bounded by y = 0, x = 2, Y = x.
Stokes' Theorem in E2. In the following exercises, compute directly the integrals on the two sides in
Stokes' theorem

h= fin (V x F) . k dA, 12 = i (F . T) ds

and verify by direct integration that Ii and 12 are equal. In this formula, C is the oriented boundary of D
and T is the unit tangent vector to C.
40. The vector field is F = (y - x)(i + j) and D is the square bounded by x = 0, x = 1, y = 0, y = 1,
oriented counterclockwise as viewed from (0,0,1).
41. The vector field is F = (4y + x 2 )i + yj and D is the square bounded by x = 0, x = 1, Y = 0, Y = 1,
oriented counterclockwise as viewed from (0,0,1).
42. The vector field is F = (x + y)i - xj and D is the triangle bounded by y = 0, x = 2, y = x, oriented
counterclockwise as viewed from (0,0,1).
43. The vector field is F = xi + (x + y)j and D is the triangle bounded by y = 0, x = 2, Y = x, oriented
counterclockwise as viewed from (0, 0, 1).
Divergence Theorem in E3. In the following exercises, compute the surface integral in the Divergence
theorem of Gauss,

by direct evaluation of the triple integral

f f In div FdV.

In this formula,S is the oriented boundary of D and n is the outward unit normal vector to D.
44. The vector field is F = xi + yj + zk and D isx 2 + i + z2 ~ 1.
45. The vector field is F = 2xi - yj + zk and D is x 2 + i + z2 ~ 4.
46. The vector field is F = xzi + xyj + xyk and D is the parallelepiped bounded by the planes x = 0,
y=O,z=O,x= l,y= 1,z=2.
430 6 VECTOR ANALYSIS

47. The vector field is F = xyi + xzj + yzk and D is the parallelepiped bounded by the planes x = 0,
y = 0, Z = 0, x = 1,y = 1, z = 4.
Divergence Theorem in E3. In the following exercises, compute the triple integral and the surface integral

in the Divergence theorem of Gauss to show without appeal to the theorem that 11 = 12. In this formula, S
is the oriented boundary of D, and n is the outward unit normal vector to D.

48. The vector field is F = 2yj + zk and D is the interior of the sphere x 2 + i + z2 = 1.

49. The vector field is F = 3yj and D is the interior of the sphere x 2 + i + Z2 = 4.

= yzj + xyk and D is the parallelepiped bounded by the planes x

50. The vector field is F = 0, y = 0,
z=O,x= 1,y= l,z=2.
51. The vector field is F = xyi + yzk and D is the parallelepiped bounded by the planes x = 0, y = 0,
z=0,x=l,y=l,z=4.
Stokes' Theorem in E 3. In the following exercises, compute the oriented line integral

in order to evaluate indirectly the surface integral

fIs (V x F) . n dA

in Stokes' theorem. In this formula,S is a smooth oriented surface in E3 with boundary curve C oriented
counterclockwise relative to the outward unit normal vector n to the surface S.

52. The vector field is F = yi + zj and the oriented parabolic surface S is r(u, Y) = v cos u i + v sin u j +
(I - v2 )k on D: °
direction of ru x rv.
°
:<:: u :<:: 2n, :<:: v :<:: I, with boundary curve C x 2 + i
= 1 and normal in the

53. The vector field is F = xj and the oriented parabolic surface S is r(u, v) = v cos u i +v sin u j + (1- y2)k
° ° +i
on D: :<:: u :<:: 2n, :<:: v :<:: 2, with boundary curve C x 2 = 2 and normalin the direction of ru x rv.
54. The vector field is F = xj and the oriented surface S is the cylinder x 2 +
top x 2 + i :<:: 1, z = I, with normal n equal to k on the top.
i = I, °:<:: z :<:: I, plus its

55. The vector field is F = zyi, and the oriented surface S is the cylinder x 2 + i = I, 0:<:: z :<:: 2, plus its
bottom x 2 + i :<:: I, z = 0, with normal n equal to - k on the bottom.
Stokes' Theorem in E3. In the following exercises, compute the oriented line integral and the surface
integral

in Stokes' Theorem, in order to verify lj = b without appeal to the theorem. In this formula, S is a smooth
oriented surface in E3 with boundary curve C oriented counterclockwise relative to the outward unit normal
vector n to the surface S.

56. The vector field is F = yi, and the oriented parabolic surface S is r(u, v) = v cos u i + v sin u j + (1- v2 )k
onD:O:<:: u:<:: 2n,0:<:: v:<:: 1,withboundarycurveCx2 +i = landnormalinthedirectionofruxrv.
 6.6 X-RAY DIFFRACTION AND CRYSTAL STRUCTURE 431

57. The vector field is F = xj, and the oriented parabolic surface 5 is r(u, v) = v cos u i +v sin u j +0- v2 ) k
on D: 0 :::: u :::: 2n,0 :::: v :::: 2, with boundary curve C:x 2+i
= 4 and nOTIllal in the direction of ru xrv.
58. The vector field is F = yi and the oriented surface 5 is the cylinder x 2 +i = 1, 0 :::: z :::: 1, plus its
top x 2 + i ::1, Z = 1, with outward normal n equal to k on the top.
59. The vector field is F = 2xj and the oriented surface 5 is the cylinder x 2 +i = 1,0 :::: Z :::: 2, plus its
bottom x 2 + i ::
1, Z = 0, with normal n equallo k on the bottom.
Green's Area Formulas. The next four exercises derive and apply the basic area formulas of Green, which
reduce two-dimensional area problems in E2 to a one-dimensional calculation of a line integral.
60. Verify that the area of region D in E2 is given by

Area of D = ~ { x dy - ydx.
2lc
61. Verify that the area of region D in E2 is given by

AreaofD= !cXdY.
62. Verify that the area of region D in E2 is given by

Area of D = !c -y dx.
63. Verify that the area enclosed in E2 by the astroid C defined by r = (cos 3 t)i + (sin 3 OJ, 0 :::: t :::: 2n, is

Area = !c xdy = 3;.

Additional Proofs. The text omits some details, which are filled in by the follOwing exercises.
64. (Green's Formula any-Domains) Let P(x,y) and Py(x,y) be continuous in a y-domain D of E2 with
boundary C. Prove Green's formula

(27) J10 Py(x,y) dx dy = - [ P(x,y)dx.

65. (Green's Theorem Cancellations) Let D be a regular domain and let Q, Qx, P and Py be continuous
in D. Prove that when two x-domains, or two y-domains, share a common boundary arc then the
corresponding terms in the integral

[Pdx+Qdy

cancel in pairs.

6.6 X-Ray Diffraction and Crystal Structure

This section presents applications of vector algebra in E3 to the description of crystal
structure and its calculation from X-ray diffraction data. The idea of modeling crystals as
three-dimensional periodic arrays of atoms was advanced by A. Bravais in 1850. However,
432 6 VECTOR ANALYSIS

a3

~
: .... ........ ... :::::;. :.........::..... ::::: FIGURE 35. A crystal lattice with atoms at
R = n,8, + nz8z + n383

the spacing of the atoms was much too small to be resolved by optical microscopes. In
1895 We. Roentgen discovered X-rays, which proved to be electromagnetic waves of
very short wavelengths. In 1912 the German physicist Max von Laue validated Bravaiss
model by showing that crystals diffract X-rays like a three-dimensional diffraction grating.
For example, it was found that salt (Nact) has a cubic lattice with spacing 4.08 A (where
1 A = 1 angstrom unit = 10-8 cm). Thus a 1 cm3 salt crystal has about 24.5 million
atoms along each edge, for a total of about 1.47 x 10 22 atoms. Here we shall sketch how
X-ray diffraction data for a crystal can be used to determine its crystal structure.
Crystal Lattices. A crystal lattice is a set of vectors in E3 of the form

where aI, a2, a3 are linearly independent (so al x a2 . a3 i- 0) and n1, n2, n3 run through
the integers. A portion of a crystal lattice is shown in Figure 35.
Figure 36 shows a portion of a 2-dimensional lattice and three families of parallel
lines that pass through the lattice points. The same phenomenon occurs in 3-dimensional
lattices but is more difficult to depiCt.

.
: __ F;ll1ilY
~ .....•.
1__ : ___ :_

. ..
~~---

.-.
. , , ,
, : : : :
---

--~-- -

" ----e

e ----e-- -----. -- -. ----. ----.

Family 2 Family 3 FIGURE 36. Lines in a 2-dimensionallattice
 6.6 X-RAY DIFFRACTION AND CRYSTAL STRUCTURE 433

The calculation of a crystal lattice from X-ray diffraction data of a sample depends
on two facts.

(a) The atoms is a crystal lattice can be arranged in stacks of parallel equispaced families
in (infinitely) many different ways.
(b) Each such stack of parallel lines will reflect a beam of X-rays of a suitable wavelength.

The remainder of this section will show how this can be done.

Nonorthogonal Bases in E3 and Their Duals. Given a crystal lattice determined

by a basis aI, a2, a3 in E3 , we wish to calculate those planes in E3 that contain an infinite
periodic subset of the lattice. This problem can be simplified by introducing a second
basis a I , a 2, a 3, called the dual basis of aI, a2, a3. The dual basis vectors a I , a 2, a 3 are
determined by their orthogonal projections onto the vectors aI, a2, a3, namely,

(2) a l . al = 1, a l . a2 = 0, a l . a3 = 0,
a 2 . al = 0, a 2 . a2 = 1, a 2 . a3 = 0,
a 3 . al = 0, a 3 . a2 = 0, a3 . a3 = 1.
It is easy to see that these conditions determine a I , a 2, a 3 uniquely. To find them, note
that the vector a2 x a3 is orthogonal to a2 and a3, and hence

Next, the scalar product with al gives Cl = al . a2 x a3, and hence a l =

(a2 x a3)/(al . a2 x a3). Continuing in the same way, we find that the dual basis to al,
a2, a3 is given by

(3) a l = Ca2 x a3)/Cal . a2 x a3),

a 2 = (a3 x al)/(al . a2 x a3),
a 3 = (al x a2)/(al . a2 x a3).

Two remarks on the construction (3) are noteworthy. First, if

is orthonormal, then
434 6 VECTOR ANALYSIS

so orthonormal bases are self-dual. Secondly, equations (2) are symmetric in (aI, a2, a3)
and (aI, a 2, a 3). Hence we may interchange the two bases in (2), which implies that

(4) al = (a 2 x a 3)/(a l . a 2 x a 3),

a2 = (a 3 x al)/(a l . a 2 x a 3),
a3 = (a l x a 2)/(a l . a 2 x a 3).

Components Relative to a Basis. If al . a2 x a3 #- 0 then (aI, a2, a3) is a basis in

E3 and every vector A in E3 can be written as

(5)

To calculate the coefficients (A I, A 2 , A3 ) we take the scalar product of A with the dual
basis vectors and use (2). This gives

al' A = AI(a l . al) +A2(a l . a2) +A 3(a l . a3) = AI.

More generally, the method gives

(6)

Calculation of Crystal Planes-Miller Indices. Consider a crystal lattice with

basis a I, a2, a3, so that the points

are the lattice points when nl, n2, n3 are integers. Let aI, a 2, a 3 be the dual basis. Then

(7)

defines a second lattice, called the dual lattice. The dual lattice is important for the
calculation of crystal structure because of the following theorem.

Theorem. Every dual lattice vector N = N(ml, m2, m3) is a normal vector to a family
of parallel crystal planes.

Proof. For the proof, consider a plane

(8)

where L is an integer. In particular, the three lattice points

(9)
 6.6 X-RAY DIFFRACTION AND CRYSTAL STRUCTURE 435

lie on a plane if

(0)

If L is any common multiple of ml, m2, m3 (so L is divisible by each) then (0) has
integral solutions fll, fl2, fl3 and the plane contains three lattice points. But then, by (0),

and similarly,

Hence the lattice points

are on the same plane for all integers k, e, m.

Spacing of the Crystal Planes. Let

be a dual lattice vector whose components (ml, m2, m3) have no common factors. Then
the family of parallel crystal planes with normal N = N(ml, m2, m3) is given by the
solutions of

In particular, for L = 1,
1
(ll) d = d(ml,m2,m3) = ------
IN(ml,m2,m3)1
is the distance between adjacent parallel planes.
For the proof, we have
N 1
-·R=-=d.
INI INI
X-Ray Beams and Their Vectors. In physical optics a collimated beam of X-rays
may be characterized by a unit vector u (beam direction) and a wavelength A. However,
it is customary to use instead
2IT
k= - = the wave number
A
436 6 VECTOR ANALYSIS

and

k = ku = the wave vector.

For X-ray diffraction patterns of crystals, typical values of the wavelength are

0.2A:s A :s lo2A.
Reflection of an X-Ray Beam by a Crystal Plane. Let an X-ray beam with wave
vector k be reflected from a crystal plane with normal N. Then by the laws of geometrical
optics a reflected beam is produced with wave vector k', where k, N, and k' lie in a plane
e
and the angle of incidence equals the angle ofreflection; see Figure 37.
This can be written as

(12) k' - k = aN (a> 0).

The Bragg Condition. Crystal planes occur as stacks of parallel planes. In general,
reflected beams from two or more parallel planes will tend to cancel by interference,
producing a weak reflected beam. A strong reflected beam will occur only when reflected
beams from adjacent planes are in phase and reinforce one another. A criterion for this
to happen is the equation

(13) 2dcose = nA,

where d is the crystal plane spacing, e is the reflection angle (see Figure 37), and n is a
positive integer. Equation (13) is called the Bragg condition after Sir William Bragg and
Sir Lawrence Bragg, a father-and-son team of physicists who pioneered X-ray analysis of
crystal structure.
The correctness of the Bragg condition is apparent from Figure 38.
Clearly, the lower beam in the figure travels 2d cos e units farther than the upper
beam. The Bragg condition requires that the increment be an integer n times the
wavelength A.

., ....
k .... ..... k'

FIGURE 37. Reflection of an X-ray beam by a plane

 6.6 X-RAY DIFFRACTION ANO CRYSTAL STRUCTURE 437

t
d
adjacent
crystal
•
~
planes
FIGURE 38. Reflection by adjacent
+ crystal planes

The Bragg Condition in Vector Form. We may write the Bragg condition as
k cos e = nrr/d, or (see Figure 37)
k'.~=-k.~= nrr
INI INI d'
or, since d = 1I1NI,

k' . N = -k . N = IN1 2nrr.

Combining this with k' - k = aN gives

or

a = 2rrn, n = 1,2,3,4, ....

Thus we have the following theorem.

Theorem. The Bragg condition (13) is equivalent to the vector condition

(14) k' = k + 2rrnN, n = 1,2,3, ... ,

where

(15)

is any vector of the dual lattice.

Crystal Structure Determination. The vector Bragg condition (14) gives all
possible diffracted X-ray beams k' corresponding to a given incident beam k and a given
438 6 VECTOR ANALYSIS

FIGURE 39. X-ray diffraction experiment

crystal lattice. Figure 39 gives a schematic diagram of a diffraction experiment with

several diffracted beams.
The central problem of X-ray diffraction analysis of crystals is the inverse problem
of finding the crystal lattice from measured values of k and kj. The Bragg condition
(14) plays an important part, but the problem is overdetermined and difficult to solve
accurately. The problem has given rise to a large literature that is beyond the scope of this
book. A thorough discussion, with references to the literature, may be found in [Zal.

IExercises.6.6. !
,. ""." . .·.,h .. , < ......~

1. (Counting Salt Crystal Atoms) Find approximately the number of atoms in a salt crystal having the
shape of a parallelepiped with edges i + j, i + 2j, j + 3k, all units in centimeters.

2. (Modeling of a Crystal Lattice) A portion of a crystal lattice with orthogonal basis vectors can be
modeled by stacking a number of identical blocks (parallelepipeds). How should the dimensions of a
block be related to the basis vectors aj, a2, a3 of the crystal lattice?
 6.6 X-RAY DIFFRACTION AND CRYSTAL STRUCTURE 439

3. (Dual Basis of a Crystal Lattice) A crystal lattice is defined by the set of orthogonal vectors a1 = i +j.
a2 = i - j, a3 = k. Compute the dual basis a I , a 2 , a 3 .

4. (Dual Basis of a Crystal Lattice) A crystal lattice is defined by the set of vectors a1 = i +j + k, a2 = i -j,
a3 = k. Compute the dual basis aI, a 2 , a 3
5. (Crystal Plane Equation) A crystal lattice is defined by the set of vectors a1 = i + j, a2 = i - j, a3 = k.
Find a crystal plane N . r = L containing aI, 3a2 and 4a3.
6. (Crystal Plane Equation) A crystal lattice is defined by the set of vectors a1 = i + j + k, a2 = i - j,
a3 = k. Find a crystal plane N . r = L containing Sa], 3a2 and 4a3.
7. (Family of Parallel Crystal Planes) A crystal lattice is defined by the set of vectors a1 = i +j, a2 = i -j,
a3 = k. Find a family of parallel crystal planes N . r = L (L an integer) containing aI, 3a2 and 4a3.
8. (Family of Parallel Crystal Planes) A crystal lattice is defined by the set of vectors a1 = i +j + k, a2 =
i - j, a3 = k. Find a family of parallel crystal planes N . r = L (L an integer) containing Sa 1, 3a2 and 4a3.

9. (Reflected X-Ray Beam) A family of parallel crystal planes is given by 12n1 + 3n2 + 4n3 = L, where L
is an integer, for the crystal lattice defined by the set of vectors a1 = i + j + k, a2 = i - j, a3 = k. Find
a reflected X-ray beam k' given incident beam k = 2a 1 + a2, assuming that the Bragg condition holds.
10. (Reflected X-Ray Beam) A family of parallel crystal planes is given by 12n] + 20n2 + ISn3 = L, where
L is an integer, forthe crystal lattice defined by the set of vectors a1 = i +j +k, a2 = i -j, a3 = k. Find a
reflected X-ray beam k' given incident beam k = rra] + 2rra2, assuming that the Bragg condition holds.
CHAPTER 7
 Partial Differential Equations
of Mathematical Physics

Chapters 2 through 4 of this text developed solution methods for physical problems that
are governed by ordinary differential equations. The purpose of Chapters 7 through 9 is
to extend these methods to problems that are governed by partial differential equations.
Partial differential equations have been the subject of vigorous mathematical research
for over 250 years and remain so today. A systematic and complete coverage of this
subject is far beyond the scope of this text. Instead, the equations to be solved here will
be introduced in the context of several explicit applications. This has the advantage of
leading to both the differential equations and the auxiliary conditions (initial conditions,
boundary conditions) that are needed to Single out the unique solutions to specific
physical problems. These ideas are developed in the present chapter. The actual solution
procedures are presented in Chapters 8 and 9.
442 7 PARTIAL DIFFERENTIAL EQUATIONS OF MATHEMATICAL PHYSICS

7.1 Vibrating Strings: D'Alembert's Wave Equation

Historically, the first physical problem to be solved by means of a partial differential
equation was that of a vibrating taut string such as those used in musical instruments. In
1746 J. d'Alembert 1 derived his now well-known wave equation and solved it to predict
the possible motions of such strings. Here d'Alembert's equation will be derived by an
alternative method.
Formulation of the Problem. Consider a thin elastic string that is fastened under
tension between two points, as in Figure 1.
Most people know by experience that if such a taut string is plucked or struck, it will
vibrate and produce a musical tone. The goal of this section is to show how Newton's
laws of motion can be used to derive d'Alembert's equation. The analysis will be based
on the following assumptions:

(I) The string is thin and uniform.

(II) The vibrations are lateral and take place in a plane.
(III) The string is perfectly flexible.
(IV) The angular deviation of the string from its rest position is small.

These assumptions will now be discussed and made more precise. First, the thinness
assumption of (I) is interpreted to mean that the string position at any time can be
modeled as a mathematical curve in space. The uniformity assumption means that the
physical characteristics of the string (diameter, mass density, elastiCity, etc.) are constant
over the length of the string. It will be shown that such a string is characterized by three
parameters:

(1) L = the length of the stretched string,

T the tension in the string,
p the mass per unit length (mass density).

x=O x=L
FIGURE l. Small transverse vibrations of an elastic string

lJean Ie Rond d'Alembert 0717-1783), one of the leading French mathematicians of the eighteenth century.
 7.1 VIBRATING STRINGS 443

To interpret assumption (II), let (x,y) be rectangular coordinates in the plane of the
motion, chosen so that when the string is at rest it occupies the interval 0 S x S L
of the x-axis. Then a lateral vibration of the string is one for which each particle of the
string with rest position (x,O) moves in time t to a new position (x, u(x, t)) with the
same x-coordinate. Thus each particle moves orthogonally to its rest position.
The Wave Function u(x, t). If observation of a vibrating string begins at time t = 0
then the prediction of its future motion is equivalent to computing the function u(x, t) of
the two independent variables x and t for all x with 0 S x S L (all points of the string)
and all times t > 0 (all future times) For fixed Xo the graph of y = u(xo, t), t :::: 0, gives
a history of the motion of particle Xo. In particular, the partial derivatives Ut(xo, t) and
Utt(xo, t) are the velocity and acceleration of that particle. For a fixed time to, the graph
of y = u(x, to), 0 s x S L, is a stroboscopic image, or "snapshot," of the position of the
string at time to. In particular, the partial derivative uxCx, to) is the slope of the string at
particle x and time to.
The Tension Vector T(x, t). The forces on any particle of the string are those
transmitted by their neighbors through the constant tension T. It will be convenient to
write

(2) T(xo, t) = force of portion x :::: Xo of the string on the portion x S Xo.

Then T(x, t) lies in the (x,y)-plane and its magnitude is the prescribed tension T:

(3) IT(x, 01 = T = constant

for 0 S x S L, t :::: O. Assumption (III) can now be interpreted. Perfect flexibility implies
that T(x, t) must be tangent to the string at point y = u(x, t), since the string does not
resist bending. Thus

(4) T(x, t) = (TCOS8, TSin8),

where

(5) Ux(x, t) = tan 8.

Finally, assumption (IV) will be interpreted to mean that the slope (5) remains so small
that its square u; is negligible in comparison with 1(u; « 1) Combining this and
Taylor's theorem gives

(6) cos8
sin 8 Ux cos 8 = Ux + O(u~) ~ Ux .
444 7 PARTIAL DIFFERENTIAL EQUATIONS OF MATHEMATICAL PHYSICS

In this approximation the tension vector T is given by

(7) T(x, t) ~ (T, TuxCx, t));

see Figure 2.
Application of Newton's Second Law. Newton's second law states that a
mechanical system with momentum M(t) and net external force F(t) satisfies
d
(8) dtM(t) = F(t).

This will be applied to an arbitrary portion [xo, xII of the string, where 0 :s Xo < Xl :s L.
A small segment [x,x + ~xl has mass p~x, velocity Ut(x, t) and momentum pUt~x.
Thus the usual calculus argument gives the integral

(9) M(t) =
IXo
XI pUt (x, t)dx

for the momentum of the portion [xo, xII. This momentum is entirely in the y-direction,
since the displacement of each particle is orthogonal to the x-axis by assumption (II).
The y-component of the net external force on the portion [xo, xII is

(10)

by (7). See Figure 3.

Hence (8) gives

(ll)

or

(12)

string

x
FIGURE 2. Tension vector for a deflected
xo string at fixed time t
 7.1 VIBRATING STRINGS 445

y = u(x, t)

T
T string

TUx(XQ, t)
x
FIGURE 3. Vertical forces on an
XQ arbitrary string portion [xo, x,]

D'Alembert's wave equation is obtained by differentiating (12) with respect to Xl:

(13)

where Xl is any point of the interval 0 :::: Xl :::: Land t ::: O.

This may be written in the more usual form as

(14)

If.
where

(15) c=

Some Product Solutions. It is easy to see that d'Alembert's equation (14) has
infinitely many solutions. In particular, if J.L is any positive constant, then the products

(16) sin(fLX) sin(fLct), sin(fLx) COS(fLCt),

COS(fLX) sin(fLct), COS(fLX) COS(fLct),

are all solutions of (14).

The Boundary Conditions. For the vibrating string problem the ends of the string
are fixed. This means that the wave function u(x, t) must satisfy the boundary conditions

(17) u(O, t) = 0, u(L, t) = 0 for t::: O.

Only the first two of the product solutions (16) satisfy u(O, t) = O. These also satisfy
u(L, t) = 0 if and only if J.L = mr/L, where n is a positive integer. This gives the following
family of solutions of (14), (17) for n = 1, 2, 3, ... :

(18) sin C~X) cos (n:ct) , sin C~X) sin (n:ct) .

These were recognized by eighteenth-century scientists as the normal modes of vibration
of a taut string. They are easy to observe experimentally.
446 7 PARTIAL DIFFERENTIAL EQUATIONS OF MATHEMATICAL PHYSICS

Linearity of d'Alembert's Equation. D'Alembert's equation (14) is linear and

homogeneous. This means that sums of solutions and constant multiples of solutions
are also solutions. More generally, if u(x, t) and vex, t) are solutions and if a and bare
constants, then

(19) W(x, t) = au(x, t) + bv(x, t)

is also a solution. The boundary conditions (17) also have this property.
Superposition. The linearity property implies that the normal mode solutions (18)
can be used to build more general solutions of the form

L sin (wnx) {an cos (wnct) + bn sin (wnct)} ,

00

(20) u(x, t) =
n=l

where the an's and bn's are constants and Wn = mr/L. The corresponding string velocity
is obtained by t-differentiation:

L (wnc) sin (wnx) {bn cos (wnct) -

00

(21) ut(x, t) = an sin (wnct)} .

n=l

For any choice of the an's and bn's (such that the series converge), (20) defines a possible
mode of vibration of the string. This is called the superposition principle.
Need for Initial Conditions. Experience suggests that to predict the future states
of the string it is necessary to know the initial state:

(22) u(x,O) = j(x) = initial position of string,

Ut(X,O) = g(x) = initial velocity of string.
Thus plucking the string means to choose an initial shape j(x) -=f. 0 and initial velocity
g(x) = 0, while striking the string means to choose initial shape J(x) = 0 and initial
velocity g(x) -=f. O. The general case is to do both simultaneously: One way to do this is
to use the normal mode representation (20). This evidently works if
00

(23) j(x) = Lansin C~X) for 0 ~x ~L

n=l

and

(24) g(x) = f: (nnc) b

n=l L
n sin (nnx)
L
for 0 ~ x ~ L.

If an's and bn S can be found such that these equations hold then by inserting them in (20)
the solution u(x, t) is obtained. It is natural to ask which functions can be so represented.
 7.1 VIBRATING STRINGS 447

The surprising answer is, essentially any function. This is shown in Chapter 8, where a
method for finding the an's and bn's is derived.

The Boundary Value Problem for u(x, t). The preceding discussion suggests that
the wave function u(x, t) for the vibrating string is characterized by the following
properties:

(25) U(x, t) is defined for °: : x ::: Land t ~ 0,

(26) Utt = c2 uxx for °: : x ::: Land t ~ 0,
(27) U(O, t)= ° and u(L, t) = ° for t ~ 0,
(28) u(x,O) = j(x) and Ut(x,O) = g(x) for °: : x ::: L,

where j(x) and g(x) are prescribed. This is an example of a boundary value problem.
A solution of the partial differential equation in the strip °: :
x ::: L, t ~ is sought °
that satisfies the boundary conditions (27) and (28) along the boundary of the strip.
It is suggested above that these conditions determine u(x, t) uniquely However, it is
conceivable that other conditions necessary for uniqueness have been missed. It will be
shown next that this is not the case by proving the following uniqueness theorem.

Uniqueness Theorem. The boundary value problem (25)-(28) has at most one solution.
This means that if Ul (x, t) and Uz (x, t) are any two solutions of (25)-(28) with the same
j(x) and g(x) then

(29)

In other words, the two solutions are identical.

The uniqueness theorem is proved below. The proof is based on the following energy
calculation.

Total Energy of the Vibrating String. The energy of the string is a sum of the
kinetic energy (KE.) due to the velOcity v = Ut and the potential energy (PE.) due
to the stretching of the string from rest against the tension T. To calculate them, let
(x, x + ilx) be an element of the string. Then its mass is m = pilx and its velOCity is
v = Ut. Hence its K E. is ~ mv 2 = ~ pu~ ilx. Integrating over the string gives

(30) Total KE. of string =- IlL

2 0
pu~dx.
448 7 PARTIAL DIFFERENTIAL EQUATIONS OF MATHEMATICAL PHYSICS

For the same element (x,x + llx) the stretched length is ds - llx, where ds is given by
elementary calculus as

(31) ds =)1 + u;llx = (I + ~u; + O(U;)) llx.

Thus the work done in the stretching against I is PE. = T(ds - ~x)
i
O(u;)) ~x ~ Iu;, since u; « 1. Integrating over the string gives

(32) Total PE. of string = -liL Iu;dx.

2 0
Finally, the total energy of the string at time t is

(33) E(t) = K.E. + PE. = -liL (2

pUt + Tux2) dx.
2 0

We can now prove the following theorem.

Conservation of Energy Theorem. The solution of the boundary value problem for
the vibrating string satisfies

(34) E(t) = E(O) = constant for all t ::: O.

Thus the total energy is constant in time!

For the proof, the derivative E'(t) will be calculated and shown to be zero for t ::: O.
Differentiating (33) and using U tt = (2Uxx and (2 = Tip gives

(35) E'(t) = i L
(pUtUtt + IUxuxt) dx

i L
(Iutu xx + Tuxuxt ) dx

T i L
(Utux)xdx

Now, the boundary condition (27) implies that Ut(O, t) = 0 and ut(L, t) = O. Hence
E' (t) = 0 and E(t) = constant, as stated.

Proof of the Uniqueness Theorem. Let Uj (x, t) and U2(X, t) be the two solutions of the
theorem and define

(36)
 7.1 VIBRATING STRINGS 449

Then the linearity of the differential equation and boundary conditions implies that
u(x, t) is also a solution of the boundary value problem (25)-(28) but with zero initial
values for 0 :::; x :::; L:

(37) U(x,O) Ul (x, 0) - U2 (x, 0) = j(x) - j(x) = 0,

aUl OU2
Ut(x,O) -(x,O) - -(x, 0) = g(x) - g(x) = o.
at at
Now the energy for the solution u(x, t) is

(38) E(t) = ~ {L {pu~(x, t) + Tu;(x, t)} dx, t ~ O.

2 10
In particular, setting t = 0 and using (37) gives E(O) = O. (Note that u(x,O) = 0 for
o :::; x:::; L implies that ux(x, 0) = 0 for 0 :::; x :::; L.) Thus the energy theorem gives
E(t) = E(O) = 0, or

(39) P i L
u;(x, Odx +T i L
u;(x, Odx =0 for t ~ O.
Here the two integrands are nonnegative and hence it must be that

(40) ut(x,t) =0 and ux(x,t) =0 for 0:::; x:::; L, t ~ O.

This clearly implies that u(x, t) = k (a constant) for 0 :::; x :::; L, t ~ O. But u(x, 0) =0
and hence k = O. This shows that u(x, t) = Ut (x, t) - U2(X, t) = 0, or

(41) Ut(X,t)=U2(X,t) for O:::;x::::L, t:::O.

Thus the two solutions are in fact identical, as was to be proved.

Other Boundary Conditions. A vibrating string can be supported at the ends in

other ways than being fixed. Imagine that a metal rod is fixed along the y-axis and attach
the string to it by a metal ring that slides on the rod without friction, as in Figure 4 below.
The analysis above shows that the string exerts on the rod a tension force whose y-
component is Tux (0, 0. This must be zero because the ring can exert no force along the

x=O x=L

FIGURE 4. String fixed at the right end, free to slide at the left
450 7 PARTIAL DIFFERENTIAL EQUATIONS OF MATHEMATICAL PHYSICS

rod. Since T =f. 0, the string must satisfy the boundary condition
(42) ux(O, t) = ° for t 2: 0.

Similarly, the other end of the string may be fixed as before (so u(L, t) = 0) or may also

°
be free to slide along a rod (so ux(L, t) = 0).
Still another way to attach the string at x = is to let it slide along a vertical rod,
as above, and attach it to the origin with a spring with Hooke constant k. Then the y-
°
component of the tension at x = must balance the restoring force of the spring, giving
the boundary condition

(43) TuxCO, t) = ku(O, t) for t 2: 0.

The same arrangement at x = L gives the boundary condition

(44) - Tux(L, t) = ku(L, t) for t 2: 0.

The vibrations produced by these alternative boundary conditions will be analyzed in

the Exercises.

Nonuniform Strings. Strings with nonconstant mass density p and/or tension T

arise in some applications:

(45) p = p(x), T = T(x).

Thus if a string is tapered, or wrapped over part of its length to reduce the frequency
of vibration, then p may vary with x. Variable T occurs in the study of hanging strings,
where T(x) is proportional to the mass of the string below point x.
D'Alemberts wave equation (14) can be generalized to cover such nonuniform
strings. In fact, equation (12) becomes

(46)

and differentiating with respect to x gives

(47)

The possible boundary conditions are the same as for the uniform strings, as are the energy
theorem and the uniqueness theorem. Of course, it is more difficult to find solutions of
(47) if P and/or T are not constants.
 7.2 HEAT DIFFUSION IN RODS: FOURIER'S HEAT EQUATION 451

I Exercises 7.1 I
1. Use Taylors theorem to show that if Ux = tane then cose = 1 + O(u~) and sine = Ux + O(u~).
2. Verify that each of the four products in equation (16) is a solution of d'Alembert's equation (26)
3. Show that the normal mode solutions of the vibrating string problem, defined by (18), are periodic
functions of t with common period P = 2L./plf.

° ° °
4. Prove the uniqueness theorem forthe vibrating string for the case where the boundary condition u(O, t) =
is replaced by ux(O, t) = and the boundary condition u(L, t) = is unchanged.

°
5. Prove the uniqueness theorem for the vibrating string for the case where the boundary conditions are
given by Tux (0, t) = ku(O, t) and ux(L, t) = 0, with k >
6. Prove the uniqueness theorem for the nonuniform string, governed by (47), with boundary conditions
u(O, t) = 0, u(L, t) = 0. Suggestion: show that

1 (L
E(t) = 2: 10 {p(x)ut(x,t) + T(x)u~(x,t)} dx = constant.

7.2 Heat Diffusion in Rods: Fourier's Heat Equation

After d'Alemberts solution of the vibrating string problem, the next major advance in
the use of partial differential equations was the mathematical theory of the diffusion of
heat of ].B. Fourier 2 In work announced in 1807, Fourier formulated the basic laws
governing heat diffusion, used them to derive a partial differential equation (later called
Fourier's heat equation), and then solved the equation to compute temperatures in heat-
conducting bodies. Here Fourier's laws are formulated in the context of one-dimensional
heat flow in a rod and used to derive the heat equation. Boundary value problems are
then formulated, and the uniqueness of their solutions is shown. The solutions to the
problems are constructed in Chapter 9.
Formulation of the Problem. Consider a cylindrical rod made of a heat -conducting
material, as in Figure 5 below.
The transient diffusion of heat in the rod will be analyzed under the following
assumptions.

(I) The rod has length L and x is the distance along the rod axis measured from
one end, so that 0 ::: x ::: L.
(II) The temperature T in the rod is a function of the time t and the coordinate x only:

2Jean Baptiste Joseph Fourier (1768-1830), French physicist and mathematician. His book Theone analytique de
la Chaleur, published in 1822, had a decisive influence on the development of mathematics and mathematical physics in
the nineteenth century.
452 7 PARTIAL DIFFERENTIAL EQUATIONS OF MATHEMATICAL PHYSICS

(1) T = u(x, t), 0::: X ::: L, t::: O.

(III) The lateral surface of the rod is thermally insulated.

(IV) The rod is homogeneous.
(V) No heat is generated in the rod.
(VI) The two ends of the rod are held at temperature zero:

(2) u(O, t) = 0 and u(L, t) = 0 for t::: O.

(VII) The initial temperature distribution is known:

(3) u(X,O) = j(x) for 0::: X ::: L,

where j(x) is a prescribed function. Assumptions (I) and (II) define the temperature
function u(x, t). The goal is to calculate it. Assumption (III) is needed for consistency.
If there is heat flow across the lateral boundary then T = u(x,y, z, t) will vary with the
three spatial coordinates and t. Assumption (IV) means that the physical parameters
that characterize the rod material are all constants, independent of X and t. They are
listed below. Assumption (V) eliminates heat sources that, if present, would modify the
temperature distribution.
Fourier's Laws of Heat Conduction. Fourier's analysis of heat diffusion was based
on two physical laws that he discovered. To formulate them, let q(x, t) denote the heat
flux in the rod at position x and time t. Thus q(x, t) is the number of calories of heat
energy per unit area per unit time that are passing point x at time t. Then Fourier's first
law is the following:
Fourier's Heat Flux Law.

(4) q(x, t) = -Kux(x, t),

where K is a positive constant, called the thermal conductivity. In words, q(x, t) is

proportional to the temperature gradient and in the opposite direction (minus sign!).
Next, let Q(t,xo,XI) denote the heat content in calories of the portion of the rod
Xo ::: x ::: Xl. Then Fouriers second law is as follows:

~u)~ ________________~)~
FIGURE 5. Homogeneous heat-conducting rod
x=o and coordinate x
 7.2 HEAT DIFFUSION IN RODS: FOURIER'S HEAT EQUATION 453

Fourier's Heat Conservation Law. For all Xc, Xl,

d
(5) -Q(t,XO,XI) = Aq(xo, t) - Aq(XI, t).
dt
This is a conservation law for heat energy because it asserts that the time rate of change
of the heat content Q(t, Xc, Xl) is equal to the net flux into the portion [Xo, xIl. The factor
A in (5) is the cross-sectional area of the rod. It is needed because q was defined as flux
per unit area. See Figure 6.
Fourier's Heat Equation. Fourier's equation will be obtained by combining Fourier's
two laws. First Q(t, Xc, Xl) must be expressed in terms of u(x, t). For this the mass density
p and specific heat c of the rod are needed. Assumption (IV) above is interpreted to
mean that the parameters K, A, p and c are all constants. Recall that c is the number of
calories needed to raise the temperature of one gram of the rod material by 1 Celsius. 0

A slice [x,x + ~xl of the rod has volume A~x, mass pA~x and hence cpA~x is the
number of calories needed to raise its temperature 10. Thus the heat content of the slice
at temperature u(x, t) is u(x, t)pcA~x and integration over Xo S X S Xl gives

(6) Q(t, xo, Xl) = lXo

X'
U(X, t)cpA dx

with t-derivative

(7) d Q(t,xo,XI) =
-d
t
lX'
Xo
Ut(x,t)cpAdx.

Substituting (7) and (4) into (5) gives

(8)

°
for all Xc, Xl with S Xo S Xl S L. Differentiating (8) with respect to Xl and then
making Xl = X gives, after division by cpA,

(9) Ut = KUxx for ° S X S L, t::: 0,

-
Xo

x=O -
Aq(xo , b)
x=L

FIGURE 6. Net heat flux into a segment Xo ::: x ::: XI

454 7 PARTIAL DIFFERENTIAL EQUATIONS OF MATHEMATICAL PHYSICS

where
K
(10) K=-.
Cp

Equation (9) is Fourier's partial differential equation for heat diffusion. The positive
parameter K is known as the thermal diffusivity. It is a function of the rod material.
The Boundary Value Problem for the Rod. The temperature in the rod can now be
calculated by solving the boundary value problem of finding a function u(x, t) such that

(11) U(x, t) is defined for 0 S x S L, t :::: 0,

(12) Ut = KU xx for 0 S x S L, t :::: 0,
(13) U(O, t) =0 and u(L, t) =0 for t > 0,
(14) u(x,O) = j(x) for 0 S x S L,

where j(x) is prescribed. These properties completely determine the function u(x, t).
This is confirmed by the following uniqueness theorem.

Uniqueness Theorem. The boundary value problem (11)- (14) has at most one
solution.

Proof. As in the case of the vibrating string, consider two solutions of (11)-(14), say
Ul (x,t) and U2(X, t), having the same initial values j(x). To show that Ul (x, t) and U2(X, t)
must be identical, it is enough to show that the function u(x, t) = Ul (x, t) - U2 (x, t) is
zero for 0 S x S Land t :::: O. Now, u(x, t) is a solution of (l1)-(14) with initial values
u(x,O) = O. Thus if

(15) let) = ~ (u 2 (x, t) dx

2 10
then let) :::: 0 and J(O) = O. Next, taking the time derivative of J(t) and using the heat
equation gives

(16)

An integration by parts in the last integral gives

L L
(17) I' (t) = KUUx I0 -K 10 u~ dx.
 7.2 HEAT DIFFUSION IN RODS: FOURIER'S HEAT EQUATION 455

The integrated terms are zero by the boundary condition (13). Thus

(18) l' (t) = -K it u~dx ::: O.

It follows (since J(O) = 0) that

(19) let) = it l' (r)dr ::: o.

But J(t) ::: 0, and hence let) = 0 for 0 ::: t < 00. This can only happen if u(x, t) = 0 for
o ::: x ::: L, t ::: 0, which completes the proof.
Other Boundary Conditions. Other physical arrangements at the ends of the bar
lead to other boundary conditions. Thus if the bar is insulated at x = 0, then there is
no heat flow there. This means that q(O, t) = 0, which, by Fourier's heat flux law (4), is
equivalent to the boundary condition

(20) Ux(O, t) = 0 for t::: O.

Another possibility is that heat is radiated out at x = 0 following Newton's law of cooling.
The law states that Flux out = k[Uin - uoutl where k > 0 is a material constant. Applied
at x = 0 with U out = 0 the law gives

(21) KuxCO, t) = ku(O, t) for t::: o.

It is noteworthy that this is isomorphic to the boundary condition (43) of Section 7.1
for the vibrating string. The corresponding boundary condition at x = L is

(22) - KUx(L, t) = ku(L, t) for t::: 0,

which is isomorphic to (44) in Section 7.l. Clearly these boundary conditions can be
prescribed independently at the two ends of the rod, leading to six different boundary
value problems. It can be shown that the uniqueness theorem and its proof extend to all
of them. Several will be solved in Chapters 8 and 9.
Nonuniform Rods. The heat equation must be modified if one or more of the bar
parameters c, p, A, or K varies with x. As an example, suppose that the cross-sectional
area A is variable:

(23) A = A(x) for 0::: x ::: L.

Thus the energy conservation equation (8) becomes

(24)
456 7 PARTIAL DIFFERENTIAL EQUATIONS OF MATHEMATICAL PHYSICS

Differentiating with respect to Xl and putting Xl = X as before gives

(25)

where K = K/(cp). Similar equations arise if c, p, or k is variable. The boundary conditions

are the same as in the uniform case.

Exercises 7.2

1. Show that for every A, B, and /L > 0, the products

u(x, t) = (A cos (/LX) + B sin(/Lx))e-/L 2Kt

are solutions of the heat equation Ut = KUxx.

2. Show that nonzero product solutions of Exercise 1 satisfy the boundary conditions u(O, t) = °and
° °
u(L, t) = if and only if A = and /L = mrlL, where n is a positive integer.

3. Prove the uniqueness theorem for heat diffusion in a rod if the boundary conditions are u(O, t) = ° and
ux(L,t) = 0.

4. Same exercise as 3 but with u(O, t) = °

and Kux(L, t) + ku(L, t) = °with K and k positive.
5. Prove the uniqueness theorem for the nonuniform rod governed by (25) with A(x) > °and u(o, t) =
0= u(L, t). Suggestion: Show that I'(t) ~ 0, where

I(t) = -1
2 0
la L
u2 (x, t)A(x) dx

and then proceed as for the uniform case.

7.3 Heat Diffusion in Plates

In this section Fouriers laws are extended to heat flow in plates. The analysis follows
closely the plan of Section 7.2. The main difference is that the temperature is now a
function of three independent variables.
Formulation of the Problem. Consider a thin flat plate made of a heat-conducting
material as in Figure 7 below.
The region Q is assumed to be connected, because more general regions may be
treated by solving the problem on each connected component separately. The diffusion
of heat in the plate will be analyzed under the following assumptions.

(I) Rectangular coordinates X and yare defined in the plane of the plate. The plate
occupies a connected region Q in the (x,y)-plane.
(II) The temperature T in the plate is a function of x, y and the time t:
 7.3 HEAT DIFFUSION IN PLATES 457

(1) T = u(x,y,t) for (x,y) in n,t::: O.

(III) The faces of the plate are thermally insulated.

(IV) The plate is homogeneous.
(V) No heat is generated in the plate.
(VI) The edgesof the plate, described by the boundary an of the plate region, are
held at temperature zero:

(2) U(X,y, t) = 0 for (x,y) in an and t ::: o.

(VII) The initial temperature distribution in the plate is known:

(3) U(X,y, O) = j(x,y) for (x,y) in n,

where j(x, y) is a prescribed function. Assumptions (I) and (II) define the temperature
function u(x,y, t). Assumption (III) is again needed for consistency Assumption (IV)
means that all plate parameters are constant. Assumption (V) eliminates heat sources
that would modify the problem.
Fourier's Heat Flux Law. For plates the heat flux is a plane 'vector q(x,y, t) and
Fourier's heat flux law takes the form

(4) q(X,y, t) = -KVu(x,y, t) = -K(u x , uy ),

where K is the (constant) thermal conductivity.

Fourier's Heat Conservation Law. To express the conservation law, let D denote
any subset of the plate n and let Q(t, D) be the heat content of D at time t, measured
from temperature zero. Then Fourier's second law states that
d
(5) dt Q(t, D) = Net flux of heat into D.

x
FIGURE 7. Homogeneous heat-conducting plate Q in E2
458 7 PARTIAL DIFFERENTIAL EQUATIONS OF MATHEMATICAL PHYSICS

A partial differential equation for u(x,y, t) will be derived by expressing the two sides of
(5) in terms of u. On the left,

(6) Q(t, D) = Ii u(x,y,t)cpAdxdy,

in analogy with (7), Section 7.2, where A is the constant plate thickness, c is the specific
heat and p is the mass density (mass per unit volume). On the right, the net flux is obtained
by integrating the normal component of q over aD, the boundary curve of the region D.
If n = n(x,y) denotes a unit normal vector to aD at (x,y) pointing out of D, then

(7) Net flux of heat into D =- ( q. nA ds.

laD
The minus sign is needed because q . n measures the flux out of D. Substituting (6) and
(7) into (5) and using (4) gives

(8) Ii UtcpAdxdy= lD K(Vu·n)Ads.

Now the divergence theorem of Chapter 6 implies that

(9) lD Vu· nds = Ii V 2udxdy,

where

(10)

Thus equation (8) can be written (after canceling A) as

(II) Il (CPUt - KV 2 u) dx dy = 0.

The equation must hold for all subsets D of the plate n and all times t ::: O. This can
happen only if the integrand is zero. For if it were positive at some point (xo,Yo) in n
then by continuity it would be positive in some disk D centered on (xo,Yo) and (ll)
would fail. The same argument holds if it is negative. Thus (11) implies Fourier's heat
equation for plates:

(12) Ut = K(Uxx + Uyy) for (x,y) in n, t ::: 0,

where again K = KJ(cp) is the thermal diffusivity.

The Boundary Value Problem for the Plate. The temperature function u(x,y, t)
for the plate is characterized by the properties

(13) U(X,y, t) is defined for (x,y) in n, t ::: 0,

 7.3 HEAT DIFFUSION IN PLATES 459

(14) ur = K(Uxx + Uyy ) for (x,y) in n, t 2: 0,

(15) u(x,y, t) = 0 for ex,y) in an, t 2: 0,
(16) u(x,y,O) = f(x,y) for (x,y) in n,
where fex, y) is a prescribed initial temperature distribution in the plate. The correctness
of this statement is confirmed by the following uniqueness theorem.

Uniqueness Theorem. The boundary value problem (13)- (16) has at most one
solution.

Proof. The proof is almost the same as that in Section 7.2. If Uj and U2 are two solutions
of (13)-(16) with the samefex,y) then U = Uj -U2 is also a solution but withfex,y) = 0
in n. Hence it is enough to show that iff = 0 then U = O. To do so, consider the double
integral

(17) let) = ~ 11 u 2 (x,y, t) dx dy,

so l(t) 2: 0 and 1(0) = O. Taking the t-derivative of let) and using the heat equation (14)
gives

(18) 1'(t) = 11 uU r dxdy =K 11 U(Uxx + Uyy ) dxdy.

If 1'(t) :s 0 then the proof may be completed by the argument following (19) in Section
7.2. To verify this, the last integral in (18) will be transformed by the divergence theorem.
By the chain rule, one has

(19) V . (uVu) = Vu . Vu + u(V . Vu) = IVuI 2 + u(uxx + Uyy ).

Integrating over n gives, on the left,

(20) 11 V·(uVu)dxdy= lo. u (vu.n)ds=o

because u = 0 on an. Combining this and (19) gives

(2l) 11 U(U xx + Uyy ) dx dy = - 11,vu ,2 dx dy :s 0,

which completes the proof.

460 7 PARTIAL DIFFERENTIAL EQUATIONS OF MATHEMATICAL PHYSICS

Other Boundary Conditions. If the plate boundary an is thermally insulated then

the boundary condition (15) of Section 7.2 must be replaced by q . n = 0, or by

au
(22) Vu . n = - =0
an
for (x,y) in an, t :::: 0,
where n = n(x,y) is the outward unit normal to an at (x,y). Equation (22) defines
au/an as the normal derivative of u at point (x, y) in an. Still another boundary condition
requires

(23) u o for(x,y)onC1,t::::0,
au
o for (x,y) on C2 , t :::: 0,
an
where an = C1+C 2 . This is called a mixed boundary condition. Yet another boundary
condition is to let the boundary radiate into an exterior at zero temperature. In this case
Newton's law of cooling gives the boundary condition

au
(24) K- + ku = 0
an
for (x,y) in an, t :::: 0,
where k > 0, in analogy with (22), Section 7.2. The uniqueness theorem extends to all
of these cases. The proofs are left for the Exercises.

I Exercises 7.3
;;0: ~
t
<4 G/WL, % ~~,~ lvrn"",,==

1. Prove the uniqueness theorem for heat diffusion in plates if the boundary condition (15) is replaced by
au/an = 0 on the boundary, t ::: 0
2. Same exercise as 1 if (15) is replaced by (23).

3. Same exercise as 1 if (15) is replaced by (24) with K and k positive.

4. Show that the heat equation for plates, Ut = K(Uxx + Uyy), has product solutions

where)., and fL are positive constants.

5. Show that if n is the unit square, defined by 0 ::: x ::: 1, 0 ::: y ::: 1, then the product solutions of
Exercise 4 satisfy the zero boundary condition on an if and only if )., = mr and fL = mJ!', where nand
m are positive integers.
 7.4 STEADy-STATE HEAT DIFFUSION IN PLATES 461

7.4 Steady-State Heat Diffusion in Plates:

The Laplace Equation
A steady-state heat distribution in a plate is one that is independent of the time t, so that
the temperature T = u(x, y). Physically, such distributions represent the equilibrium
temperatures that occur when fixed boundary values are maintained for a long time.
Now, u(x, y), as a temperature distribution, must still satisfy Fourier's heat equation (14)
of Section 7.3. Hence, since Ut = 0, u(x,y) must satisfy
(1) Uxx + Uyy = O.

This is the well-known Laplace equation in two space dimensions.

Boundary Value Problem for u(x, y). If boundary values are given then u(x, y) is
a solution of the boundary value problem

(2) u(x,y) is defined for (x,y) in the connected region Q,

(3) UXX + Uyy = 0 for (x,y) in Q,
(4) u(x,y) = j(x,y) for (x,y) on aQ,

where j(x,y) is a prescribed function on the boundary aQ. The following theorem
guarantees that these properties completely determine u(x, y).

Uniqueness Theorem. The boundary value problem (2) - (4) has at most one solution.

Proof. As in Section 7.3, it will suffice to show that if j(x,y) = 0 on aQ then u(x,y) = 0
in Q. To verify this, apply equation (19) of Section 7.3 to u. Since (3) holds, one gets

(5) V' . (uV'u) = IV'uI 2 for (x,y) in Q.

Integrating this over Q and using the divergence theorem gives

(6) fL IV'uI 2 dx dy = LI. u(V'u . n) ds =0

because U = 0 on aQ. Equation (6) implies that IV'ul = 0 and hence V'u = (u x , uy) =
(0,0) in Q. But this implies that u(x,y) = c = constant in Q. Finally, c must be zero
because by assumption, u(x,y) = 0 for (x,y) on aQ. Thus u(x,y) = 0 for (x,y) in Q,
as stated.

The boundary value problem (2)-(4) is usually called the Dirichlet problem. It will
be solved in Chapter 9 for plates of several special shapes including rectangles, circular
462 7 PARTIAL DIFFERENTIAL EQUATIONS OF MATHEMATICAL PHYSICS

disks and related shapes. For more general shapes the problem is difficult and often can
be solved only by numerical methods.
Other Boundary Conditions. The boundary value problem obtained by replacing
the Dirichlet boundary condition (4) by the condition

(7)
au = j(x,y)
- for (x,y) on an
an
is called the Neumann problem. Physically, (7) is equivalent to specifying the heat flux
across the boundary an. A third possibility is

(8) u = jl(X,y) for (x,y) on C1 ,

au
an = h(x,y) for(x,y) on C2 ,

where an = C1 + C2 • This is called the mixed problem. Its physical interpretation is

clear. Finally, the boundary condition

(9) K-
au + ku = kj(x,y) for (x,y) on an
an
defines the Robin problem. It corresponds phYSically to heat diffusion with radiation
into a medium with temperature j(x,y). The uniqueness theorem holds for the mixed
and Robin problems. The solution of the Neumann problem is not unique, but any two
solutions differ only by a constant. These questions are taken up in the Exercises.
Other Applications. The Laplace equation and the boundary value problem defined
above occur in many other applications. These include the analysis of gravitational,
electric and magnetic potentials, and theories of ideal fluid flow.

IExercises 7.41
1. Supply the details for identity (5).
2. Apply the divergence theorem carefully to obtain from identity (5) the resulting equation (6).
3. For the Neumann problem, show that if Ul(X,y) is a solution and c is a constant then U2(X,y) =
U1 (x, y) + c is a second solution. Hence the Neumann problem does not have a unique solution.

4. For the Neumann problem, show that if Ul (x,y) and U2(X,y) are two solutions with the same f(x,y)
then there is a constant c such that U2(X,y) - Ul (x,y) = c.
5. Prove the uniqueness theorem for the mixed problem.
6. Prove the uniqueness theorem for the Robin problem (assuming that K and k are positive).
 7.5 VIBRATIONS OF DRUMS 463

7.5 Vibrations of Drums

A mathematical theory of the vibrations of drums was initiated by Euler in 1757, only
eleven years after d'Alembert's solution of the vibrating string problem. The drum problem
proved to be substantially more difficult than the string problem and Euler was able
to give only a partial solution. Full understanding of the problem came much later, in
the nineteenth and early twentieth centuries. In this section Euler's partial differential
equation for the vibrating drum is derived from Newton's laws and a uniqueness theorem
is given. The analysis closely parallels that for the vibrating string in Section 7.1.
Formulation of the Problem. Consider a thin elastic membrane whose boundary
is fastened under tension along the points of a plane curve and that occupies a connected
region Q, as in Figure 8 below.
At equilibrium the membrane is at rest in the plane of its boundary curve. If struck it
vibrates out of that plane. This motion will be analyzed under the following assumptions.

(I) The membrane is thin and uniform.

(II) The vibrations are perpendicular to the equilibrium plane.
(III) The membrane is perfectly flexible.
(IV) The angular deviation of the membrane from its rest position is small.

The thinness assumption of (I) is interpreted to mean that the membrane position at
any time can be modeled as a mathematical surface in space. The uniformity assumption
means that the physical parameters of the membrane (mass denSity, tension) are constant.
To interpret assumption (II) let (x,y, z) be Cartesian coordinates in space, chosen such
that the membrane at eqUilibrium occupies a domain Q in the (x,y)-plane. Then (II) is
interpreted to mean that each point (x,y, 0) of the resting membrane is found at time t at
a point (x, y, u(x, y, t)). The problem of predicting the membrane motion then becomes
the mathematical problem of calculating the wave function u(x,y, t). The condition that
the membrane boundary be fixed becomes the boundary condition u(x,y, t) = 0 for

x FIGURE 8. Thin elastic membrane stretched over plane

domain Q
464 7 PARTIAL DIFFERENTIAL EQUATIONS OF MATHEMATICAL PHYSICS

(x, y) on an and t ::: O. Assumption (III) is interpreted to mean that the membrane offers
no resistance to either bending or shearing. This hypothesis is used below to calculate
the tension. Finally, assumption (IV) will be interpreted to mean that the membrane
displacements are so small that the deviation of the surface normal vector

(1) N = ( -Ux , -uy, 1)

from the vertical satisfies u; + u; « 1. To this approximation one has

(2)

and N is the unit normal vector to the membrane, with positive z-component.
The Tension Vector T(x,y, t). Fix a time t and consider on the membrane surface
z = u(x,y, t) a small element of arc C with length ds and unit tangent vector t. If the
membrane were cut along C, the two sides would pull apart. Thus if the two sides of C
are denoted by 1 and 2, then side 1 exerts a force T(x,y, t)ds on side 2, and side 2 exerts
the balancing force - T(x,y, t)ds on side 1. It will be shown that the four assumptions
listed above determine T(x,y, t) completely. First, T(x,y, t) must lie in the tangent plane
to the membrane at the point (x,y, u(x,y, t)), because the membrane does not resist
bending This property is equivalent to the condition that T(x,y, t) is orthogonal to the
normal vector N at the point. Second, T(x,y, t) must be orthogonal to the unit tangent
vector t, because the membrane does not resist shearing. Finally, IT(x,y, t)1 = To is a
fixed positive constant by the uniformity assumption. It follows that

(3) T(x,y, t) = To(t x N),

since t and N are orthogonal unit vectors; see Figure 9 below.

To calculate T from the derivatives Ux and uy let

Co: roes) = (x(s),y(s),o)

be the projection of C onto the (x,y)-plane, with arc length s as parameter (see Figure
9). Then

(4) to = (x'(s),y'(s),o)
and

(5) n = (y'(5), -x'es), 0)

 7.5 VIBRATIONS OF DRUMS 465

FIGURE 9. Deflected membrane tension (3)

are the unit tangent and normal vectors to Co, respectively It follows that C has the
representation

(6) res) = (x(s),y(s),u(x(s),y(s),t»),

and hence

(7) t = r'(s) = (xl(s),y'(s), uxx'(s) + uyy'(s)).

The vector t is also a unit vector because X'(S)2 + y'(s)2 = 1 and hence (uxX'(S) +
uyy'(S))2 « 1 by assumption (IV). A short calculation using u~ + u; « 1 gives

where aulan = Uxnl + uyn2 is the normal derivative of u on Co. Combining this with (3)
gives

(9) T(x,y, t) = (Ton 1, Ton2, To :~).

Application of Newton's Second Law. Euler's wave equation for the vibrating
drum will now be derived by applying Newton's second law to an arbitrary portion of the
membrane. For this purpose let D be any portion of Q whose boundary aD is a simple
closed curve. The momentum in the z-direction of the portion of the membrane lying
466 7 PARTIAL DIFFERENTIAL EQUATIONS OF MATHEMATICAL PHYSICS

over D is

(10) M(t) = fin put(x,y, t) dx dy,

where p is the mass density of the membrane. By (9), the z-component of the net force
on the membrane portion over D is

(ll) F(t) = 1
aD
au
To- ds,
an

or, by the divergence theorem,

(12) F(t) = fin To(uxx + Uyy ) dx dy.

Thus Newton's second law dM(t)ldt = F(t) gives

(13)

By a now-familiar argument, this can hold for all subsets D of Q only if the integrand is
zero, or

(14)

This is Euler's wave equation for the vibrating drum.

The Boundary Value Problem for the Vibrating Drum. The formulation of this
problem parallels that for the vibrating string. Thus the drum motion is characterized by
a wave function u(x,y, t) that satisfies the conditions

(15) u(x,y, t) is defined for (x,y) in Q and t 2: 0,

(16) Utt = [2(u xx + Uyy ) for (x,y) in Q and t 2: 0,
(17) u(x,y, t) = 0 for (x,y) on aQ and t 2: 0,
(18) u(x,y,O)=j(x,y) and Ut(x,y,O)=g(x,y) for(x,y)inQ,

where j(x,y) and g(x,y) are the initial displacement and velocity of the membrane,
respectively The completeness of these properties will be verified by proving the
uniqueness theorem below.

Uniqueness Theorem. The boundary value problem (15)-(18) has at most one solution.
 7.5 VIBRATIONS OF DRUMS 467

Proof. As for the vibrating string problem, it is enough to show that if (15)-(18) hold
withf = g = 0 then U = O. This may be done by showing that the energy integral
(19) E(t) = ~ fin (PU~ + To(u; + U~») dx dy

is constant for all t ::: 0 and is zero iff = g = O. First, taking the t-derivative of E(t) and
using the wave equation (16) gives

(20) E'(O = fin {pUtUtt + To(uxu xt + UyUyt)} dx dy

To fin {Ut(Uxx + Uyy ) + UxUxt + UyUyt} dx dy

To fin {(UtUx)x + (UtUy)y} dx dy

To ( ut(V'u· n) ds,
Jar:!.
where the last step follows from the divergence theorem. Now (17) implies that
Ut(x,y, t) = 0 for (x,y) on aQ and t ::: O. Thus the last integral is zero and so E'(t) = 0
for t ::: O. Also

(2l) E(O) = ~ fin {pg2(X,y) + To (P(x,y) + j}(x,y»)} dxdy,

and in particular, E(O) = 0 iff = g = O. This proves that iff = g = 0 then E(t) = 0
for all t ::: o.

It follows that

(22) Ut = Ux = uy = 0 for all (x,y) in Q and t ::: 0,

and so u(x,y, t) = k (a constant). But u(x,y, 0) = 0, by assumption, and hence k = 0

and u(x,y, t) = 0 for all (x,y) in Q and t ::: 0, as was to be shown.
In Figure 10, one of the normal modes of vibration of a circular drumhead is depicted.

FIGURE 10. Vibrations of a circular drumhead

468 7 PARTIAL DIFFERENTIAL EQUATIONS OF MATHEMATICAL PHYSICS

In Chapter 9 the vibrating drum problem will be solved for rectangular and circular
drums and some related shapes. Solution methods for more general shapes may be found
in more advanced texts such as [Gal, [Ke].

1Exercises 7.51'
1. Show that Euler's wave equation (14) has product solutions

U(x,y, t) = [A cos(wct) + B sin(wct)] sin (Ax) sin(/1Y),

where A and /1 are positive constants and w = )).,2 + /1 2
2. Show that it Q is the unit square, defined by 0 ::: x ::: 1, 0 ::: y ::: 1, then the product solutions of
Exercise 1 satisfy the zero boundary condition on 3Q if and only if A = mr and /1 = mrr, where nand
m are positive integers

3. Prove the uniqueness theorem for the case that the Dirichlet boundary condition (17) is replaced by the
Neumann condition
3u(x,y, t) f
_--'.- = 0
an or Cx,y) on aQ, t >
- o.
4. Same as exercise 3 but with the Robin boundary condition

au
-dn + hu = 0 for (x,y) on aQ, t ::: 0,

where h > O.

7.6 Heat Diffusion in Solids

This section extends the analysis of Section 7.3 to 3-dimensional heat flow in solid bodies.
The theory closely parallels that of Section 7.3 and only the main steps are presented.
Details are left for the Exercises.
Formulation of the Problem. The diffusion of heat in a 3-dimensional heat
conducting body will be analyzed under the following assumptions.

(I) Rectangular coordinates in 3-dimensional Euclidean space are given by (x,y, z).
The body occupies a connected region n in (x,y, z)-space.
(II) The temperature T in the body is a function of x,y, z, and the time t.
(III) The body is homogeneous.
(IV) No heat is generated in the body
(V) The boundary an of the body is held at temperature zero.
(VI) The initial temperature is known.
 7.6 HEAT DIFFUSION IN SOLIDS 469

Assumptions (I) and (II) define the temperature function u(x,y, z, t). Assumption
(III) means that the parameters K, p, and c for the body are constants. The remaining
assumptions are self-explanatory.
Fourier's Heat Flux Law. For 3-dimensional bodies the heat flux is a vector
q (x, y, z, t) in 3-dimensional space and Fourier's heat flux law takes the form

(1) q(x,y, z, t) = -KVu(x,y, z, t) = -K(ux , uy , u z ).

Fourier's Heat Conservation Law. If D is any subset of the body n, then Q(t, D),
the heat content of D at lime t, is given by the triple integral

(2) Q(t,D) = ff1 u(x,y,z,t)cpdxdydz;

see (6) in Section 7.3. Moreover,

(3) Net flux of heat into D = - flaD[ q. nds,

where aD is the boundary surface of D and n is a unit normal vector on aD pointing

out of D. Hence Fourier's second law (5) of Section 7.3, together with (1) and the
3-dimensional divergence theorem, gives

(4) ff1 (cPUt - KV 2 u) dx dydz = 0,

in analogy with (11), Section 7.3, where

(5)

is the 3-dimensional Laplace operator. The validity of (4) for all subsets D of n gives
Fourier's heat equation for 3 dimensions:

(6) Ut = K(U xx + Uyy + uzz ), K = K/(cp).

The Boundary Value Problem for Solids. The temperature function u(x,y, z, t)
for a body n is characterized by the properties

(7) U(x,y, z, t) is defined for (x,y, z) in n, t :::: 0,

(8) Ut = K(Uxx + Uyy + uzz ) for (x,y, Z) in n, t :::: 0,
(9) u(x,y, z, t) = ° for (x,y, z) on an, t :::: 0,
(10) u(x,y, z, 0) = J(x,y, z) for (x,y, z) in n,
470 7 PARTIAL DIFFERENTIAL EQUATIONS OF MATHEMATICAL PHYSICS

where f(x,y, z) is a prescribed initial temperature distribution in Q. As in Section 7.3,

the completeness of these properties is confirmed by a uniqueness theorem.

Uniqueness Theorem. The boundary value problem (7)-(10) has at most one solution.

As before, it is enough to show that if (7)-(10) hold with f = 0 then u = O. The

proof is left for the Exercises.
Other Boundary Conditions. The alternative boundary conditions (22)-e24) of
Section 7.3 are also applicable to heat diffusion in three dimensions. The uniqueness
theorem holds in all these cases. Proofs are left for the Exercises.

I,Exercises 7.6J
1. Prove the uniqueness theorem for the boundary value problem (7)-00) by the method of Section 7.3.

2. Prove the uniqueness theorem for the case that the Dirichlet condition (9) is replaced by the Neumann
condition au/an = 0 on the boundary

3. Prove the uniqueness theorem for the case that the Dirichlet condition (9) is replaced by the Robin
condition Kau/an + ku = 0 on the boundary (with K and k positive).

4. Show that the heat equation for solids, Ut = K(Uxx + Uyy + uzz), has product solutions

where A, fJ." and v are positive constants.

5. Show that if Q is the unit cube, defined by 0 :::: x :::: 1, 0 :::: y :::: 1, 0 :::: z :::: 1, then the product solutions
of Exercise 4 satisfy the zero boundary condition on aQ if and only if A = mr, fJ., = m1f, v = p1f, where
n, m and p are positive integers.

7.7 Steady-State Heat Diffusion in Solids

A steady-state temperature distribution in a 3-dimensional body Q is defined by a
function T = uex,y,z) that satisfies the 3-dimensional heat equation and Ut = O. This
gives the 3-dimensional Laplace equation

(1) U xx + Uyy + U zz = O.
Three-dimensional analogues of the boundary value problems of Section 7.4 can be
formulated, and the uniqueness of the solutions proved, just as in Section 7.4. These are
developed in the Exercises below.
 7.7 STEADy-STATE HEAT DIFFUSION IN SOLIDS 471

The Dirichlet Problem. The Dirichlet problem for a connected domain n in 3-dimensional Euclidean
space asks for a function u(x, y, z) that satisfies

(2) U(x,y, z) is defined for (x,y, z) in n,

(3) Uxx + Uyy + uzz = 0 for (x,y, Z) in n,
(4) u(x,y, z) = j(x,y, z) for (x,y, z) on an,

wherej(x,y, z) is a prescribed function on an.

1. State and prove a uniqueness theorem for the Dirichlet problem. Suggestion: Follow the plan of Section
7.4.
2. The Neumann problem in three dimensions is obtained from (2)-(4) by replacing (4) by au/an = j(x, y, z)
on an. Show that if UI (x, y, Z) is a solution of this problem and c is any constant then U2 (x, y, z) =
U I (x, y, z) + c is another solution.

3. Show that if Ul (x,y, z) and U2 (x,y, z) are any two solutions of the Neumann problem with the same j
then there is a constant c such that Ul (x,y, z) = Ul (x,y, z) + c.
4. Prove the uniqueness theorem for the mixed problem in three dimensions.
5. Prove the uniqueness theorem for the Robin problem in three dimensions.
CHAPTER 8
 Fourier Analysis and
Sturm-Liouville Theory

The principal method for solving the boundary value problems of mathematical physics
is the method of separation of variables. The method will be developed and applied
to a variety of problems in this chapter and Chapter 9.
Several examples of the method of separation of variables were given by D. Bernoulli,
L. Euler and others in the late eighteenth century. However, it was ].B. Fourier in his
book of 1822 who revealed the true scope of the method. Fourier based his method
on the convergence of certain infinite series, which are now called Fourier series, but
full proofs of the convergence of these series eluded him. The first rigorous proof of the
convergence of Fourier series was given by P.G.L. Dirichlet in 1829. The scope of the
method was extended to heat diffusion in nonhomogeneous media by C.E Sturm and
]. Liouville in the 1830s and 1840s. An interesting history is given in ]esper Llltzen's
article in the book Studies in the History of Mathematics, Volume 26 (1987), page 242,
published by the Mathematical Association of America.
This chapter is divided into three major parts. Part I treats heat diffusion in uniform
homogeneous rods and the corresponding theory of Fourier series. Part II treats heat
diffusion in the idealized, but important, problem of infinite rods. This leads to Fourier's
integral theorem. Part III treats heat diffusion in nonhomogeneous rods and Sturm and
Liouville's generalization of Fourier series.
474 8 FOURIER ANALYSIS AND STURM-lioUVILLE THEORY

Part I Fourier Series

Part I develops the basic properties of Fourier series in the context of heat diffusion
in a homogeneous rod of finite length. Fourier sine series, Fourier cosine series and
the complete Fourier series are shown to arise from different choices of the boundary
conditions.

8.1 Dirichlet Boundary Conditions and

Fourier Sine Series
To begin we shall consider linear heat diffusion in a rod (see Section 7.2) with zero
temperature at the two ends (Dirichlet boundary conditions); see Figure 1 below.
For simplicity let the rod have length L = n and thermal diffusivity K = 1. Then the
temperature distribution u(x, t) in the rod is determined by the boundary value problem
of finding a function u(x, t) that satisfies

(1) U(x, t) is defined for 0 :::: x :::: n, t ~ 0,

(2) Ut= Uxx for 0 :::: x :::: n, t ~ 0,

(3) u(O, t) = 0 and u(n, t) = 0 for t ~ 0,
(4) u(x,O) = f(x) for 0:::: x:::: n,

where f(x) is a prescribed temperature distribution. It was shown in Section 7.2 that
there is at most one such function u(x, t). The problem is now to determine for which
f(x) (if any) the boundary value problem has a solution u(x, t) and to find a method for
calculating it. Of course, ifj(x) is the zero function then u(x, t) = 0 is the unique solution.
If f(x) I- 0 then it is not clear whether or not the problem is solvable. The solvability
problem will be approached here by Fourier's method of separation of variables. The
solvability will be shown by a construction of the solution.
Product Solutions. To begin, the initial condition (4) will be ignored, and the
solutions of conditions (1)-(3) of the special form

(5) U(x, t) = X(x)T(t)

~u)~ ________________~)~
FIGURE 1. Heat diffusion in a finite rod; Dirichlet
x=o x=L boundary conditions
 8.1 DIRICHLET BOUNDARY CONDITIONS AND FOURIER SINE SERIES 475

will be constructed. These are called product solutions. Substitution of (5) into the heat
equation (2) gives

(6) X(X)y'(t) = X"(x)T(t) ,

where primes denote ordinary derivatives. Division by X(x)T(t) gives
yl(t) X"(x)
(7) -- = --
T(t) X(x)
for a -< x< n
- ,
t >_ a.

Now, the left side of (7) is independent of x, and the right side is independent of t.
Hence the two sides must equal one and the same constant. Thus separate equations are
obtained for T(t) and X(x), namely

(8) y' (t) = -A T(t) for t ~ a

and

(9) X"(x) = -AX(X) for a:::: x :::: n,

where A, called the separation constant, is an undetermined real number. If T(t) and
X(x) are any solutions of these equations then the product (5) is a solution of the heat
equation. The factors X and Y in (5) are linked by the separation constant which must
be the same for both.
Next consider the two boundary conditions

(10) u(a, t) = X(a)T(t) = a,

u(n, t) = X(n)T(t) = a, for t ~ a.

If T(t) is not the trivial solution of (8) then (10) holds only if x(a) = a and X(n) = a.
This leads to the following eigenvalue problem for A and X(x).
Dirichlet Eigenvalue Problem. Find the special values of A such that

(11) X" + AX = 0 for 0:::: x:::: n,

(12) X(O) = 0 and X(n) = 0,

has a nontrivial solution. This problem is analogous to the matrix eigenvalue problem of
Chapter 5. It will be shown to have infinitely many distinct eigenvalues and associated
eigenfunctions. The method of finding them is to write the general solution of (1),
containing A, and then to use the boundary conditions to find a nontrivial solution X(x).
The cases A < 0, A = 0 and A > 0 will be treated separately.
476 8 FOURIER ANALYSIS AND STURM-lioUVILLE THEORY

Case 1 (). < 0). Write)' = -IL 2 (IL > 0) Then (11) becomes X" - IL 2 X = °with
general solution

(l3)

where (] and (2 are arbitrary constants. The two boundary conditions give a pair oflinear
equations for (1 and (2:

(14) (1 + (2 = 0,

and A = -IL 2 is an eigenvalue if and only if (14) has a nontrivial solution «(], (2) =I- (0,0).
This is possible if and only if system (14) has a zero determinant:

1 1
(15)

But this is impossible for IL > 0, because e- J1T[ < 1 < el"Jr. Hence the eigenvalue problem
has no solution for A < 0.
Case 2 (A = 0). Proceeding similarly, (11) becomes X" = 0 with general solution
X(x) = (1 + C2X. Hence X(O) = Cl, X(n) = Cl + C2n, and A = 0 is an eigenvalue if and
only if

C] + n(2 = 0,

°
with determinant n =I- O. Thus A = is not an eigenvalue of problem (1)-(12).
°
Case 3 (A> 0). Write A = -IL 2 (IL > 0), so (11) becomes X" + IL 2X = with general
solution

(16) X(x) = (1 COS(ILX) + (2 sin(ILx).

The boundary conditions yield the linear system

(17) (] = 0,

with determinant

(18)
 8.1 DIRICHLET BOUNDARY CONDITIONS AND FOURIER SINE SERIES 477

The determinant is zero if and only if sin(ltn) = a or It = Itn = n, where n is any

positive integer. System (17) with It = n has the solution (c], C2) = (0,1). Thus the
eigenvalues and eigenfunctions of (11)-(12) are

(19)

X(x) = Xn(x) = sin(nx), n = 1,2,3, ....

The corresponding factors T(t) = Tn(t) are defined by (8) with A = n 2 , and we take

(20) T(t) = Tn(t) = e- n1t , n = 1,2,3, ....

Thus a complete list of the product solutions of (1)-(3) is

(21) Un (x, t) = Xn(x)Tn(t) = sin(nx)e- n2t , n = 1,2,3, ....

Superposition. A product solution Un (x, t) is a solution of the boundary value
problem (1)-(4) only if the initial temperature j(x) is an eigenfunction sin(nx). To
construct solutions for more general j(x) , we note that the heat equation (2) and
the boundary conditions (3) are linear and homogeneous in u. It follows that linear
combinations

L b sin(nx)e-
N
(22) u(x, t) = n n2t

n=]

also satisfy (1)-(3). This function is a solution of the boundary value problem, provided
that j(x) is the linear combination of the eigenfunctions sin(nx)
N
(23) j(x) = I)n sin(nx).
n=]

More generally, if j(x) can be represented by a convergent infinite series

L b sin(nx),
00

(24) j(x) = n O::s x ::s n,

n=]

then the corresponding solution is

L b sin(nx)e-
00

(25) u(x, t) = n n2t , O::s x ::s n, t 2: O.

n=]

Constructing solutions from the product solutions in this way is called superposition.
478 8 FOURIER ANALYSIS AND STURM-LIoUVILLE THEORY

I~Exercises 8.1 f:

Heat Diffusion and Graphs. Solve the linear heat diffusion problem and graph the solution surface
z = u(x, t) for 0::: x ::: Jr, 0 ::: t ::: 2.

1. j(x) = sin(2x)
2. j(x) = 2 sin(2x) - sin(x).
3. j(x) = sin(2x)cos(2x). Hint: 2sinecose = sin2e.
4. j(x) = sin(4x). Describe geometrically the difference between this solution surface and the solution
surface forj(x) = sin(2x), on 0::: x::: Jr, 0::: t ::: 2.
Heat Diffusion and Proofs. Solve the linear heat diffusion problem and verify directly that the formula
obtained for u(x, t) is a solution of the problem.
5. j(x) = sin(3x).
6. j(x) = 2 sin(3x) + sin(2x).
7. j(x) = sin(2x)(l + 2 cos(2x»
8. (Trigonometry) Establish the trigonometric identity 2 sin (a) sin(b) = costa - b) - costa + b), assuming
the sum identity cos(x + y) = cos x cos y - sin x sin y.
9. (Integration) Show that for m =1= n,

fa" sin(mx) sin(nx)dx = O.

10. (Calculus) Derive If: sin2 (nx) dx = n12, for all integers n > O.
11. (Calculus) Derive If: cos 2 (nx) dx = Jr/2, [or all integers n > 0
12. (Integration Theory) Explain from basic rules of integration why

1b(N];fnC ) x ) dx = ] ;
N
1b fnCx)dx.

13. (Heat Solution) Let

U(X, t) = L bn sin (nx) e-n

N 2
t.
n=l
Find explicitly the partial derivatives Ut(x, t) and uxx(x, t) and compare the answers for equality to show
that Ut = uxx.
14. (Dirichlet Conditions) Show that the sum

U(X, t) = L b sin(nx)e- n
N
n
2
t
n=l
satisfies the homogeneous boundary conditions u(O, t) = 0 = urn, I).
15. (Initial Heat Distribution) Let J be a finite sum of trigonometric functions,
N
j(x) = L bn sin(nx).
n=l
 8.2 ORTHOGONALITY ANO FOURIER COEFFICIENTS 479

Define u(x, t) = L~=l bn sin(nx)e~n2t Show that u(x, 0) = j(x).

16. (Heat Equation) Show that the series L~=1 Xn(x)Tn (I) is a solution of the partial differential equation
Ut = u xx , provided that each product term Xn (x)Tn(t) is a solution of Ut = Uxx.

17. (Dirichlet Conditions) Show that the series L~=l Xn(x)Tn(t) satisfies the homogeneous bound-
ary conditions, provided that each product term Xn (x)Tn (t) satisfies the homogeneous boundary
conditions.

8.2 Orthogonality and Fourier Coefficients

When can a function j(x) be represented as the sum of a sine series,

L b sin(nx),
00

(1) j(x) = n 0 S x S :rr?

n=1

Fourier's surprising discovery is that essentially any function can be so represented. The
key to the representation is the observation that the eigenfunctions sin(nx), n = 1,2,3,
... , satisfy

(2) 1" sin(mx) sin(nx)dx = 0 for m =j:. n.

Check this by integration. We shall refer to this property by saying that the functions
{sin(nx)} are orthogonal on the interval 0 S x S :rr.
The orthogonality relation (2) and the integral identity

(3) [" sin 2 (nx)dx = ~ for n = 1,2,3, ...

10 2
imply that the coefficients bn in (1) must be given by the integrals

(4) bn = - 210" j(x) sin(nx)dx,

:rr 0
n=1,2,3, ....

To verify this we multiply (1) by sin(mx), where m is any positive integer, and integrate
over 0 S x S :rr. The result is

(5) 1"o
j(x) sin(mx)dx = L b 1" sin(mx) sin(nx)dx.
00

n=1
n
0

Moreover, in the sum on the right of (5), all but one of the integrals are zero by the
orthogonality property: The single nonzero term occurs when n = m. It is equal to
(:rr/2)b m by (3). Thus (5) implies (4). The numbers bn defined by (4) are called the
Fourier sine series coefficients ofj(x).
480 8 FOURIER ANALYSIS AND STURM-lioUVILLE THEORY

l~xerci~~s8.2 ~
Fourier Sine Series. Compute the Fourier sine series coefficients bn for the given function j(x) on [0, n].
1. j(x) = 2 sin(2x) + sin(3x)(l + cos(3x)).
2. j(x) = x.
3. j(x) = e- X

4.j(X)={~1 °nl2:s x :s :sn/2,n.

< x

8.3 Convergence of Fourier Sine Series

We have seen that if

L b sin(nx)
00

(1) f(x) = n
n=l

then the coefficients bn are given by Fourier's formula

(2) bn = -
n
21"
0
f(x) sin(nx)dx, n = 1,2,3, ....

When is the converse true? That is, if j(x) is given and if the coefficients bn are defined
by (2), then does (1) hold? Dirichlet showed that this is true when j(x) and rex) are
sectionally continuous on 0 :S x :S n as described by the following definition.
Sectionally Continuous Functions and Derivatives. A function f(x) defined for
a :S x :S b is said to be sectionally continuous if there is a partition of a :S x :S b into
finitely many subintervals,

(3) a = Xo < Xl < X2 < ... < Xm-l < Xm = b,

such that

(4) f(x) = Nx) for Xi-l < x< Xi (i = 1,2, ... , m)

and

(5) Nx) is continuous for Xi-l:S x :S Xi (i = 1,2, ... , m).

The values j(a) ,f(XI), ... ,f(Xm-l) ,f(b) may be assigned in any way. If in addition

(6) f((x) is continuous for Xi-l:S x :S Xi (i = 1,2, ... , m),

then rex) is said to be sectionally continuous. Note that the definition prescribes a
procedure for constructing these functions. One need only select the partition, the
 8.3 CONVERGENCE OF FOURIER SINE SERIES 481

a
FIGURE 2. A sectionally continuous function

functions.li Ci = 1,2, ... , m), and the values fCa),fCxl), ... ,f(xm~l),fCb). An example
is shown in Figure 2.
Right-Hand and Left-Hand Limits and Derivatives. The right-hand limit off
at any point x is defined to be the limit of fCt) as t approaches x from the right. It is
denoted by f(x + 0). Thus

(7) fCx + 0) = lim fCx + h),

h~O+

where h -+ 0+ means that h approaches 0 through positive values. Similarly, the

left-hand limit off at x, denoted by fCx - 0), is defined by

(8) f(x - 0) = lim f(x - h).

h~O+

For functions f with right-hand and left-hand limits, the right-hand and left-hand
derivatives off at a point x, denoted here by f~ ex) and f~ ex) respectively, are defined by

(9) f +' ex ) -- l'

1m
f(x + h) - f(x + 0)
h
h--+O+

and

(10) f '( x) = l'1m fCx-h)-f(x-O) .

~ h--+O+ -h
Sectionally continuous functions f on a:::: x :::: b have right-hand and left-hand limits at
every point of a < x < b, plus a right-hand limit at x = a and a left-hand limit at x = b.
Moreover, if x is not a partition point then f is continuous at x and

(1) f(x - 0) = f(x) = f(x + 0).

At partition points we have by (4), (5),

(12) f(Xi - 0) = ji(Xi) ,

feXi + 0) = Ji+l (Xi)
482 8 FOURIER ANALYSIS AND STURM-LIoUVILLE THEORY

and the numbers f(Xi - O),f(Xi), andf(xi + 0) may be all distinct. Iff' is also sectionally
continuous on a::: x::: b, thenf has right-hand and left-hand derivatives at every point
of a < x < b, plus a right-hand derivative at x = a and a left-hand derivative at x = b.
Moreover, if x is not a partition point then l' (x) exists and

(13) f~(x - 0) = rex) = f~(x + 0).

At partition points one has, by (4), (6),

(14) f~(Xi) = f(Xi),

f~(Xi) = f(+l(Xi)'

With these preliminaries, Dirichlets convergence theorem for Fourier sine series may be
formulated as follows.
Convergence Theorem for Fourier Sine Series. Let f and l' be sectionally
continuous on 0 ::: x ::: Jr and let

(15) bn = - 217r f(x) sin(nx)dx, n = 1,2,3, ....

Jr 0

Then the Fourier sine series off converges for every x and it has the sum
00 1
(16) I)n sin(nx) = "2 [r(x + 0) + f(x - 0)] for 0 < x< Jr.
n=l

In particular, if x is not a partition point then

L b sin(nx) = f(x)
00

(17) n for 0 <x< Jr.

n=l

A proof of the convergence theorem is given in Section 8.6 below.

Solution of the Dirichlet Boundary Value Problem. The convergence theorem
provides a solution of the boundary value problem (1)-(4) of Section 8.1 for any initial
distribution f(x) such that

(18) f and l' are sectionally continuous,

f(O) = 0, f(Jr) = 0, and
f(x) = ~ [r(x + 0) + f(x - 0)] for 0 < x< Jr.

Recall that the last equation holds automatically except at the finitely many points where
f(x) is discontinuous. For such f(x) the problem is solved by the following two steps.
 8.3 CONVERGENCE OF FOURIER SINE SERIES 483

First, compute the Fourier sine series coefficients bn defined by (15) and write
00

(19) f(x) = I)n sin(nx), 0 S x S Jr.

n=1

Second, construct u(x, t) for t > 0 by

L b sin(nx)e-
00

(20) u(x, t) = n n2r .

n=1

Note that after f(x) is expanded as a sine series (19), then u(x, t) is obtained by multiplying
each term bn sin(nx) by the factor e- n2t .
Examples. This section is concluded with four examples that illustrate the theory
developed above.

EXAMPLE 1. (Constant Initial Temperature). Letf(x) = 1 for 0 < x < Jr. Then

bn = -211f 1 . sin(nx)dx = -2 (1 - cos(nJr» ,

Jr 0 nJr
or

hn =0, b2»+1 -_ (2n

4
+ I)Jr' n = 0, 1,2, ....

The convergence theorem implies that

f(x) = ~
Jr
f
n=O
sin(2n + l)x ,
2n +1
0 < x < Jr.

Since f(x) = 1, this relation is equivalent to

Jr. sin(3x) sin(5x) sin(7x)
(2l) - = sm(x) + - - - + - - - + - - - + . . . 0< x < Jr.
4 3 5 7 '
This is a remarkable result. First, the convergence theorem guarantees that the series
converges for every x. This is not obvious from direct inspection. Second, the sum of the
series is independent of x for 0 < x < Jr. As a special case, if x = Jr/2 then

sin ((2n + 1)~) = sin(n + ~)Jr = cos(nJr) = (_1)n

and we have
Jr 1 1 1 1 1
(22) -=1--+---+---+···.
4 3 5 7 9 11
The convergence of this alternating series is clear from Leibnitz's alternating series con-
vergence theorem. However, it is difficult to show that the sum is Jr/4. The convergence
is very slow.
484 8 FOURIER ANALYSIS AND STURM-lioUVILLE THEORY

EXAMPLE 2. (Constant Temperature Gradient). Take f(x) = x for 0 < x < rr, so
l' (x) = 1 is a constant, and f(O) = f(rr) = O. Then f and l' are sectionally continuous and

bn = -
rr
2i 0
Jf

x sin(nx)d.x =
2(_l)n-l
n
, n=1,2,3, ....

Application of the convergence theorem gives

. sin(2x) sin(3x) sine 4x) )
x = 2 ( sm(x) - - - + - - - - - + ...
234
for 0 < x < rr. In particular, taking x = rr/2 gives

~=2(1-~+~-~+"-)
as before.

EXAMPLE 3. (Triangle Function). Let the graph of f(x) on 0 S x S rr be the triangle

with vertices (0,0), (rri2, 1) and (rr, 0), as in Figure 3, so

f(x) = ~ Ix for 0 S x S rr/2,

rr rr - x for rr/2 < x S rr.

Integration gives the Fourier coefficients

2i
bn = -
rr 0
Jf
8 sin(nrr/2),
f(x) sin(nx)d.x = 22
n rr
n=1,2,3, ....

In particular, b2n = 0 and

b _ 8(-l)n
n=0,1,2, ....
2n+l - rr2(2n + 1)2 '
This gives the sine series representation

y (nl2,1)

o FIGURE 3. The triangle function

 8.3 CONVERGENCE OF FOURIER SINE SERIES 485

for 0 :::: x :::: n. In particular, x = nl2 gives

1= ~
n2
(1 + ~ + ~ + ~ + ... )
32 52 72
.

EXAMPLE 4. (Parabolic Profile). Let

j(x) = x(n - x), 0:::: x S n.

The graph of j(x) is a parabola through the points (0,0) and (n,O) with vertex at the
point (nI2,n 2/4). Integration gives the Fourier coefficients

bn =-
n
21 0
1C
x(n - x) sin(nx)dx = -43 (1 -
nn
(-It) .

Thus b2n = 0, b2n+1 = (8In)/(2n + 1)3 and the convergence theorem gives
x(n - x) = ~ (Sin(X) + sin(3x) + sin(5x) + ... )
n 13 33 53 '
°: : x :::: n.

Taking x = nl2 gives, after multiplication by n18,

n3 1 1 1
-=1--+---+···.
32 33 53 73

Exercises 8.3

Linear Heat Diffusion. Solve the linear heat diffusion problem for the given initial heat distribution f and
Dirichlet boundary conditions. Display at least the first five terms of the series solution.
1. f(x) = 3.
2. f(x) = sin (x) - 2 sin(3x).
3. f(x) = 1 +x.
4. f(x) = sin(3x) cos(3x).
5. j(x) = x(rr - x)2.
x2 o :'S x :'S rr/2,
6. j(x) = { (rr _ x)2
rr/2 :'S x :'S rr.
Graphs of Heat Diffusion Solutions. Solve the linear heat diffusion problem with Dirichlet boundary
conditions for the given initial temperature distributionj. Graph the temperature surface on 0 :'S x :'S rr,
o :'S t :'S 2. For infinite series, graph the first three terms of the series (n = 1,2,3).
7. f(x) = 2.
8. j(x) = sin(3x).

9. f(x) = g o :'S x :'S rr/2,

rr/2 :'S x :'S rr.
486 8 FOURIER ANALYSIS ANO STURM-lioUVILLE THEORY

Snapshots of Heat Diffusion Solutions. Solve the linear heat diffusion problem with Dirichlet boundary
conditions for the given initial temperature distribution f. Graph the snapshot u(x, t) on 0 :S x :s :rr for
t = 0, t = 0.05, t = 0.1. For infinite series, graph the first three terms (n = 1,2,3).
10. f(x) = sine 4x)

11. f(x) = x 2 (n - x)

I 0 :s x :s :rr12,
12. j(x) = { 0
:rr12 :s x :S n.
Numerical Series Identities. Compute the Fourier sine series for the given function f. Substitute a value
of x in 0 :S x :S :rr into the sine series and apply Dirichlet's convergence theorem to obtain the given numerical
series identity

13. f(x) = {:rr12 o :S x :S :rr12, ~ = 2::00 n(2k+1)(_l)k+2

:rr-x :rr12 :S x :S :rr, 2 k=O (2k+l)2 .

3n 3 J2 _
14. pCx) -- x(:rr - x), ---r28 - 1 + J3 - 53 - f3
1 1 1
+ 931 + j'j3
1
- .. '.

8.4 Neumann Boundary Conditions and

Fourier Cosine Series
This section treats linear heat diffusion in a rod (see Section 7.2) with zero boundary
temperature conditions (Dirichlet boundary conditions) replaced by zero heat flux
conditions (Neumann boundary conditions). The temperature distribution in the rod is
then determined by the boundary value problem of finding u(x, t) such that

(1) U(x, t) is defined for °: : x ::: rr, t ~ 0,

(2) Ut = Uxx for °: : x ::: rr, t ~ 0,
(3) Ux(O, t) = ° and ux(rr, t) = ° for t ~ 0,
(4) u(x,O) = j(x) for °: : x::: rr,

where j(x) is a prescribed function. From Section 7.2 we know that the problem has
at most one solution. The solution will be constructed by the separation of variables
methods of Section 8.1. Only the boundary condition (3) is different. Hence, only the
main steps are given. Details are left for the Exercises.
The product solutions of (1)-(3) satisfy the same differential equations as before,
(8)-(9) of Section 8.1. The new boundary conditions (3) for a product solution u(x, t) =
X(x)T(t) are

(5) UxCo, t) = X'(O)T(t) = 0,

ux(rr, t) = X'(rr)T(t) = 0, for t ~ 0.
 8.4 NEUMANN BOUNDARY CONDITIONS AND FOURIER COSINE SERIES 487

For nontrivial solutions, X(x) must satisfy X/CO) = 0, X/(n) = O. This leads to a new
eigenvalue problem.
Neumann Eigenvalue Problem. Find the values of A such that

(6) X" + AX = 0 for 0:::: x :::: n,

(7) X/CO) =0 and X/(n) = 0,
has a nontrivial solution.
Again, the three cases A < 0, A = 0 and A > 0 must be considered separately. The
work is similar to the Dirichlet case. The results are

(8)

X(x) = Xn(x) = cos(nx),

T(t) = Tn(t) = e- n2t , n = 0, 1,2, . ...

The details are left for the Exercises. Note that this time AO = 0 is an eigenvalue with
eigenfunction Xo(x) = 1 (a constant).
Superposition. The product solutions are

(9)

Solutions of the complete problem will be built by superposition. Thus we try

00

(10) u(x, t) = Lan cos(nx)e-n2t

n=O

and the constants an are determined by the initial values

00

(ll) U(X,O) = f(x) = Lan cos(nx) for 0:::: x :::: n.

n=O

Orthogonality. The coefficients {an} can be determined by an orthogonality relation,

as in the previous case. Here the relation is given by the elementary integrals

(12) l 1C
cos(mx) cos(nx)dx =0 for m =1= n,

(13) lITo cos 2 (nx)dx = -rr

2
for n = 1,2,3, ... ,

(14) lIT 12 dx=n for n =0.

488 8 FOURIER ANALYSIS AND STURM-lioUVILLE THEORY

Multiplying (11) by cos(mx) and integrating over 0 ::s x ::s 7r gives the coefficients
(15) ao = -liT( J(x)dx,
7r 0

(16)
2iT( J(x) cos(nx)dx,
an = - n = 1,2,3, ....
7r 0

The series in (11) with these coefficients is called the Fourier cosine series for J(x) on
o ::s x ::s 7r. The convergence theorem for the cosine series is analogous to that for the
sine series.
Convergence Theorem for Fourier Cosine Series. Let J(x) and J' (x) be section-
ally continuous on 0 ::s x ::s 7r and let the Fourier cosine series coefficients of J(x) be
defined by (15)-(16). Then the Fourier cosine series ofJ(x) converges for every x and it
has the sum

I>n cos(nx) = ! [J(x + 0) + J(x -

00

(17) 0)] for 0 < x < 7r.

n=O
This theorem is verified in Section 8.6 below.
Solution of the Neumann Boundary Value Problem. The preceding convergence
theorem provides a solution ofthe boundary value problem (1)-(4) for any J(x) such that

(18) J and J' are sectionally continuous,

J'(O) = 0, fe7r) = 0, and
J(x) = ![Jex + 0) + J(x - 0)] for 0 < x < 7r.

For suchJ(x) the problem is solved in two steps.

First, compute the coefficients {an} by (15)-(16) and write
00

(19) J(x) = ~::>n cos(nx), 0 < x < 7r.

n=O
Second, construct u(x, t) for t > 0 as
00

(20) u(x, t) = Lan cos(nx)e-n 2t

•

n=O
Examples. The following five examples illustrate the theory.

EXAMPLE 1. (Constant Initial Temperature). LetJex) = 1 for 0 < x < 7r. Then ao = 1
and

an = -
2iT( cos(nx)dx = 0 for n:::: 1.
7r 0
 8.4 NEUMANN BOUNDARY CONDITIONS AND FOURIER COSINE SERIES 489

Thus f(x) = 1 = ao and the series has only one nonzero term.
EXAMPLE 2. (Constant Initial Gradient). Define f(x) = x for 0 < x < n so l' (x) = 1.
Then f and l' are continuous on 0 :::; x :::; nand

1 (" x21ll'
ao = -; 10 x dx = 2n 0 2'

_21" _
while for n :::: 1,
2(cos(nn) - 1)
an - - x cos(nx)dx - 2 .
n 0 nn
Hence a2 n = 0 and az n+! = -( 4In)/(2n + 1)2 Thus the convergence theorem gives

(21) x= ~ - .!.
(COS(x) + cos(3x) + cos(5x) + ... ) 0 :::; x :::; n.
2 n 12 32 52 '

Taking x = 0 gives the equation

n2 1 1 1
- = 1 + - + - + - + · ...
8 32 52 72
The value x = n leads to the same equation. For x = nl2 equation (21) is verified by
noting that cos ((2n + l)nI2) = 0 for n = 0,1,2, ....

EXAMPLE 3. (A Sectionally Constant Function). Define

(22) j(x) = 11o for

for
0:::; x :::; nl2,
nl2 < x :::; n.

Thenf and1' are sectionally continuous. Also,

ao = 2:. ("/2 dx = ~
n 10 2
and

an = ~ ("12 cos(nx)dx = ~ sin (n~) , n:::: l.

n 10 nn 2
Hence a2n = 0 for n :::: 1, while
2 . 1 2( -1)"
a2n+l = n(2n + 1) SIn ((n + 2)n) = n(2n + 1)' n = 0, 1,2, ....

The convergence theorem gives

1 2 ~ (_1)n cos(2n + 1)x
f (x) = - +- ~
2 n n=O 2n + 1
490 8 FOURIER ANALYSIS AND STURM-lioUVILLE THEORY

for 0 < x < nl2 and nl2 < x < n. At the point x = nl2 one has

and cos ((2n + l)nI2) = O. Hence the convergence theorem (17) is verified at this point.

EXAMPLE 4. (Triangle Function). Define

for 0::: x ::: nl2,

(23) f(x) = -2jx
n n-x for nl2 < x ::: n.

Then

ao =~
n 10r f(x)dx = ~n (::)
2
= ~.
2
Moreover, integration gives, after some work,

21][
an = -
4
f(x) cos(nx)dx = 22 [2 cos(nn/2) - cos(nn) - 1]
non n
for n = 1,2,3, .... This implies that a2n+l = 0, while
4
a2n = n 2(2n)2 [2 cos(nn) - 2] ,
so
-16 -4
'4n+2 = n2(4n + 2)2 = n 2(2n + 1)2 .
Since f(x) is continuous, and l' (x) is sectionally continuous, the convergence theorem
gives

f( x) = -1 - -4 [ cos ( 2x) + cos(6x) + cos(lOx) + ... ]

2 n2 32 52

for 0 ::: x ::: n. In particular, taking x = 0 gives

n2 1 1 1
-=1+-+-+-+
8 32 52 72
.. · '
as was shown above.

EXAMPLE 5.The functionf(x) = sin(x), 0 ::: x ::: n, is its own Fourier sine series (so
bl = 1and all other bn = 0). Moreover,f(x) = sin (x) and1'(x) = cos(x) are continuous
and hence sin(x) also has a Fourier cosine series on 0 ::: x ::: n. Integration gives

ao = -1 17r sin(x)dx = -,
2
non
 8.4 NEUMANN BOUNDARY CONDITIONS AND FOURIER COSINE SERIES 491

a] = -2fT( sin (x) cos(x)d.x = 0,

Jr 0
while, for n :::: 2,
2fT( . 2(1 + (_l)n)
an = -Jr 0
sm(x) cos(nx)d.x = (
Jrl-n
2)'

Thus a2n+l = 0 and a2n = (4IJr)/(1 - 4n 2). It follows that

sin (x) =~- ~

Jr Jr
f
n=l
cos(2nx)
4n 2 - 1
for 0 ~ x ~ Jr.

It is clear that this equation fails for x < 0 since sin (x) is odd and cos(2nx) is even. It is
not difficult to see that the sum of the above series is I sin (x) I for all x.

I Exercises 8.4 ~

Linear Heat Diffusion. Solve the linear heat diffusion problem for the given initial heat distribution j and
Neumann boundary conditions. Display at least the first five terms of the series solution.
1. j(x) = -1.
2. j(x) = cos (x) + 3 cos(2x)
3. j(x) = 1 +x
4. j(x) = cos(3x) cos(2x).
Hint: cos(a - b) - cos(a + b) = 2 cos(a) cos(b).
5. j(x) = x(n - x)2.
x2 o ~ x ~ nil.
6. j(x) = { (n _ x)2
nl2 ~ x ~ n.
Graphs of Heat Diffusion Solutions. Solve the linear heat diffusion problem with Neumann boundary
conditions for the given initial temperature distributionj. Graph the temperature surface on 0 ~ x ~ n,
o ~ t ~ 2. For infinite senes, graph the first three terms (n = 0, 1,2).
7. j(x) = 3.
8. j(x) = cos(5x).
o~ x
9. j(x) = {~l ~ n12,
nl2 ~ x ~ n.
Snapshots of Heat Diffusion Solutions. Solve the linear heat diffusion problem with Neumann boundary
conditions for the given initial temperature distribution j. Graph the snapshot u(x, t) on 0 ~ x ~ n for
t-values t = 0, t = 0.05 and t = 0.1.
10. j(x) = cos(3x).

11. jex) = x 2 (n - x)
l 0~x~nI2,
12. j(x) = { 0
nl2 ~ x ~ n.
492 8 FOURIER ANALYSIS AND STURM-lioUVILLE THEORY

Numerical Series Identities. Compute the Fourier cosine series for the given function f. Substitute a
value of x in 0 :::: x :::: 7f into the cosine series and apply Dirichlet's convergence theorem, to obtain the given
numerical series identity

13. j(x) = {~/2 0:::: x :::: 7f/2,

nl2 < x < 7f
1'{ _ 00
4" - Lk=O 2k+l .
(-Ok

14. j(x) = X(7f - x), 7f 3/6 - L~l ~ cos(2hx)f = 1 + 1r + tr + ~ + ...

15. (Neumann Problem Details) Derive the result Xn = cos nx, An = n 2 for the Neumann problem
X" +AX = 0, 0:::: x:::: 7f, X'(O) = X'(n) = 0, by considering the three cases A> 0, A = 0 and A < O.

8.5 Periodic Boundary Conditions and the

Complete Fourier Series
An interesting variant of the linear heat diffusion problem of Section 7.2 is obtained
by bending the rod into a circle and welding the ends together to make a ring. This
problem, which was one of the first to be studied by Fourier, is often called Fourier's
ring problem. It is solved below by the method of separation of variables.
For simplicity let the radius of the ring be 1. Then the circumference is L = 2rr, and
the length of any arc of the ring coincides with the angle, in radians, that it subtends at
the rings center. In particular, the points of the ring may be identified by their angular
distance from a fixed point; see Figure 4.
The temperature distribution in the ring at time t will be characterized by a
temperature function u(x, t), defined for 0 :s x :s 2rr and t 2: O. Our goal is to compute
this function.
Periodic Boundary Conditions. The function u(x, t) will satisfy Fourier's heat
equation Ut = Uxx and an initial condition u(x, 0) = j(x) , as in the preceding problems.
However, there seem to be no obvious boundary conditions, since the ring is closed. To
remedy this lack we note that the angle x is defined only up to multiples of 2rr; that is,

FIGURE 4. Circular ring and angular coordinate x

 8.5 PERIOOIC BOUNDARY CONDITIONS AND THE COMPLETE FOURIER SERIES 493

x and x + 2n define the same point on the ring. Hence, if x is unrestricted, then
(1) u(x + 2n, t) = u(x, t), for all x and all t ::: 0.

The same is true of the partial derivatives of u(x, t). Condition (1) states that u(x, t) is a
periodic function of x with period 2n. In particular,

(2) U(O, t) = u(2n, t) and ux(O, t) = ux (2n, t) for t::: 0.

These conditions are called periodic boundary conditions. We shall show that the
temperature distribution in the ring is determined by the boundary value problem of
finding a function u(x, t) that satisfies

(3) U(x, t) is defined for °: : : x ::::: 2n, t ::: 0,

(4) Ut = Uxx for °: : x ::::: 2n and t ::: 0,

(5) U(O, t) = u(2n, t) and UX(O, t) = ux (2n, t) for t ::: 0,
(6) u(x,O) = j(x) for °: : x ::::: 2n,

where j(x) is a prescribed initial temperature distribution. This problem has at most one
solution. This uniqueness theorem is readily verified by the method of Section 7.2. The
proof is left for the Exercises. Here the separation of variables method will be used to
construct the solution.
The product solutions of (3)-(5) satisfy the same differential equations as before.
Only the boundary conditions (5) are different. Thus T(t) = c- At and X" + AX = 0
where A is the separation constant. To determine the values of A and X(x) we have a new
eigenvalue problem, which is formulated as follows.
Periodic Eigenvalue Problem. Find the values of A such that

(7) X" + AX = ° for 0::::: x::::: 2n,

(8) X(2n) = XeD) and X'(2n) = X/CO),
has a nontrivial solution. The three cases A < 0, A = ° and A > ° must be treated
separately, as before.
Case 1 (A < 0). Write A = _fI2 with fI > 0. Then

(9) X(x) = c) eIU + C2e-Jlx,

x/ex) = fI (CjC JlX - C2C-JlX) ,

where C), C2 are arbitrary constants. For an eigenvalue we must choose fI and (c), C2) =f:.
(0,0) such that (9) satisfies the boundary conditions (8). Substituting (9) into (8) gives
494 8 FOURIER ANALYSIS AND STURM-lioUVILLE THEORY

a pair of linear homogeneous equations for (Cl, C2) that may be written

(10) (e21r/-L - 1) Cl + (e- 21r/-L - 1) C2 = 0,

(e 21rfL -1)Cl - (e- 21rfL -1)c2 = O.

There is a solution (Cl, C2) =1= (0,0) if and only if the system has determinant zero. But
the determinant is

(11)
(e 21rfL - 1) (e- 2JrfL - 1)
(e l1rfL - 1) - (e- 21rfL - 1)
°
which is never zero for p., > (since e21rfL > 1 for p., > 0). Hence, there are no negative
eigenvalues for problem (7)-(8).
Case 2 (A = 0). Here X(x) = Cl +C2X, and the boundary conditions are X(21l') - X(O) =
21l'C2 = 0, X' (21l') - X' (0) = C2 - C2 = 0 with one nontrivial solution (Cl, C2) = (1,0).
This gives the pair

(12) A = Ao = 0,
X(x) = Xo(x) = l.

Case 3 (A > 0). Write A = p.,2 with p., > O. Then

(13) X(x) = Cl sin(p.,x) + C2 cos(p.,x) ,

X' (x) = p., [Cl cos(p.,x) - C2 sin(p.,x) 1.

The boundary conditions (8) give, after simplification, the linear homogeneous equations

(14) sin(21l'p.,)Cl + (cos(21l'p.,) - 1)C2 = 0,

(cos(21l'p.,) - I)Cl - sin(21l'p.,)c2 = O.

The determinant of this system is

sin(21l'p.,) (cos(21l'p.,) - 1)
(15) = 2[cos(2rrp.,) - 1].
(cos(21l'p.,) - 1) - sin(21l'p.,)

Thus the eigenvalues A = p., 2 satisfy cos(21l' p.,) = l. The positive roots of this equation
are the positive integers p., = n (n = 1,2,3, ... ). Moreover, for p., = n the matrix of
system (n) has rank 0 (all four entries are zero), and hence there are no restrictions on
Cl and C2. This means that the solution space of (7)-(8) for the eigenvalue A = n2 is
two-dimensional. The eigenfunctions are sin(nx), cos(nx) and all their nontrivial linear
combinations. The positive eigenvalues, and their eigenfunctions, are

(16)
 8.5 PERIODIC BOUNDARY CONDITIONS AND THE COMPLETE FOURIER SERIES 495

X(x) = Xn(x) = an cos(nx) + bnsin(nx), n = 1,2,3, ... ,

where an and bn are arbitrary constants not both zero. The associated factors TCt) = Tn (t)
satisfy T~ = -n 2 T" and hence Tn(t) = e- n2t A complete list of product solutions may
be written

(17)

°
for n = 0, 1,2, .... For n = this reduces to the constant function uo(x, t) = ao, while
for each n ~ 1 the product solution contains two parameters, an and bn.
Superposition. Addition of product solutions gives solutions of the form

L (an cos(nx) + b sin(nx)) e-

00

(18) u(x, t) = ao + n
n2t •

n=l

This provides a solution of the Fourier ring problem (3)-(6) if the initial values j(x) satisfy

= j(x) = ao + L
00

(19) U(x, 0) (an cos(nx) + bn sin(nx)

n=l

for°: : x :::: 2](.

Orthogonality. The key to writing j(x) as a series (19) is again orthogonality In this
case the set of eigenfunctions

(20) {l, sin(x), cos(x), sin(2x), cos(2x), ... }

°: :
°: :
of (7)-(8) is an orthogonal set on the interval x :::: 2](. This means that the product
of any two distinct members of this set integrates to zero over x :::: 2](. Explicitly,

(21) f027f 1 . sin(nx)dx = 0, for n = 1,2,3, ... ,

f027f 1· cos(nx)dx = 0, for n = 1,2,3, ... ,
ft~7f sin(mx) . sin(nx)dx = 0, for m i- n, m, n = 1,2,3, ... ,
f027f sin(mx) . cos(nx)dx = 0, for m, n = 1,2,3, ... ,
f027f cos(mx) . cos(nx)dx = 0, for m i- n, m, n = 1,2,3, ....

All of these integrals are elementary Moreover, the integrals of the squares of members
of the set (20) are

(22) f027f 12 dx = 2](,

fo27r sin 2 (nx)dx = f02][ cos 2 (nx)dx = ](, n = 1,2,3, ....
496 8 FOURIER ANALYSIS AND STURM-lioUVILLE THEORY

If (19) is multiplied successively by the eigenfunctions (20) and integrated over 0 :s x :s

2n, then the integrals (21)-(22) yield Fouriers formulas for the coefficients:

(23) ao = -1
2n
1 0
2Jr
j(x)dx,

an = -1 1 2Jr
f(x) cos(nx)dx,

1
n 0

2Jr
bn = -1 f(x) sin(nx)dx, n=1,2,3, ....
n 0

The series in (19), with coefficients an, bn defined by (23), is called the complete Fourier
series for f(x) on 0 :s x :s 2n. The following analogue of the convergence theorem for
the sine and cosine series is proved in Section 8.6 below.

Convergence Theorem for Complete Fourier Series. Let f(x) and j'(x) be
sectionally continuous on 0 :s x :s 2n and let the Fourier coefficients off be defined by
(23) above. Then the complete Fourier series off converges for every x and has the sum

L {an cos(nx) + bn sin(nx)} = ![j(x + 0) +f(x -

00

(24) ao + 0)]
"=1

for 0 < x < 2n.

Solution of Fourier's Ring Problem. The preceding convergence theorem provides

a solution of the Fourier ring problem (3)-(6) for any f(x) such that

(25) f and l' are sectionally continuous on 0 :s x :s 2n,

f(2n) = f(O) and1'(2n) = 1'(0), and
f(x) = ! [[(x + 0) + f(x - 0)] for 0 < x < 2n.

The solution is given by (18), where the an and bn are the Fourier coefficients off, defined
by (13)

Examples. The following two examples illustrate the theory. Additional examples are
given in Section 8.6 and in Chapter 9.

EXAMPLE 1. Define

X for o :s x :s n,
f(x) = [
0 for n < x :s 2n.
 8.5 PERIODIC BOUNDARY CONDITIONS AND THE COMPLETE FOURIER SERIES 497

Then J and l' are sectionally continuous and j(x) is continuous except at x = re. The
Fourier coefficients (23) are

ao = -
1 12;[ J(x)dx = -
1 121!" x dx = re
-
2re 0 2re 0 4
and, for n ::: 1,

an = -
re
11 0
2 ][
J(x) cos(nx)dx =
(-l)n-l
ren
2

and
1 121!" (_1)n+1
bn =- J(x) sin(nx)dx = ,
re 0 n
by elementary integration. Details are left to the Exercises. The complete Fourier series
can be written. It is
~ + (Sin(X) _ sin(2x) + sin(3x) _ sin(4x) ... )
4 1 2 3 4
2 (cosex) cos(3x) cos(5x) cos(7x) )
- -; -1-2- + 32 + 52 + 72 + . .. .
The convergence theorem implies that the sum of this series is J(x) for all 0 ::s x ::s 2re
except possibly at x = re. For example, taking x = 0 gives the equation

o = ~4 - 3.re (~
12
+~
32
+~
52
+~
72
+ ...) '

or
re 2 1 1 1 1
8 = 12 + 32 + 52 + 72 + ... ,
as shown above in Section 8.3 and Section 8.4. For x = re one has! [J(re+O)+J(re-O) 1=
rel2 and hence the convergence theorem gives

which leads back to the same equation. As a third special case take x = re12. Then
J(rel2) = rel2, and all the cosine terms are zero. Hence the convergence theorem gives

~ = ~ + (~- ~ + ~ - ~ + .. -),
as was shown above in (5) of Section 8.3.
498 8 FOURIER ANALYSIS AND STURM-lioUVILLE THEORY

EXAMPLE 2. Define

f(x) = !sin(X) for 0::::: x ::::: n,

o for n < x ::::: 2n.

Clearly,f is continuous and J' is sectionally continuous, so that the convergence theorem
applies. Elementary integration gives

ao = 1
-
lJ( sin(x)dx = -,1
2n n
I1J(
0

al = - sin (x) cos(x)dx = 0

n 0

and

an
1
= -no
lJ( sin(x) cos(nx)dx
1 + (_1)n
= nI-n
( 2) for n::: 2.

Thus a2n+1 = 0, while

-2
a - for n >_ 1.
2n - n(4n2 - 1)

Similarly

bn = -
1 lJ( sin(x) sin(nx)dx,
n 0
so
1
bi = 2' bn =0 for n::: 2.

Thus the convergence theorem gives, for 0 ::::: x ::::: 2n,

1 1. 2 ~ cos(2nx)
f(x) = - + - sm(x) - - ~ .
n 2 n n=1 4n 2 - 1

Making x ~ 0+ gives
1 2 1
L 4n2 _ 1 '
00

o= ; - ;
n=l

or
1 00 1
2 = ~ (2n - 1) (2n + 1)
I I I
=-+-+-+
1·3 3·5 5·7
....
 8.5 PERIODIC BOUNDARY CONDITIONS AND THE COMPLETE FOURIER SERIES 499

This last equation is often proved in the first year calculus course. The choice x = nl2
leads to
1 1 2 (_l)n
L 4n
00

1 = -; + 1 - -; 2 - 1'
n=l

or

~ _ 1 = 2 (_1___1_ + _1___1_ + ... ) .

2 1·3 3·5 5·7 7·9

Complex Form of the Complete Fourier Series. The convergence theorem (24)
implies that ifj(x) satisfies (24) thenj(x) has the representation

L {an cos(nx) + bnsin(nx)}

00

(26) j(x) = ao + for 0:::: x:::: 2n,

n=l

where the coefficients an and bn are defined by (23). The expansion (26) takes a much
Simpler form if we use Euler's formula to express sin(nx) and cos(nx) in terms of the
complex trigonometric functions e±inx. Indeed,

1 ( 'nx
cos(nx) =1 e. + e- lnx
. )
,

1 ( 'nx
sin(nx) = 1 e. - e- lnx
. )
,

and so

where
1
(27) Co = lao,
1
Cn = 1 (an - ibn) ,
en = C_ n ,

for n = 1,2,3, .... It follows from (26) that

L {cneinx + cne- inx } = + L

00 00

(28) j(x) = Co + cneinx .

n=l n=-oo
500 8 FOURIER ANALYSIS AND STURM-lioUVILLE THEORY

This formula is called the complex Fourier series for j(x). Notice that by (23),

(29) en = 2~ 1 2Jr
j(y)e- iny dy, n=O,±1,±2, ....

This can also be seen by using the orthogonality relation

1 2Jr
eimy einy dy = 1 2Jr
ei(m-n)y dy = 27l'omn,
where omm = 1 and omn = ° if m =I n.

IExercises 8.5 ~
JLV"S ~"b~" dt:Lii£i%bdi ¥iJih%'"

1. (Fourier Series) Given the function

X for 0::: x ::: n,

f(x) = {0 for n < x ::: 2n,

verify by direct integration of the Fourier coefficient formulas for an, bn that the complete Fourier series
off is

2. (Fourier Series) Given the function

f(x) = {~n(x) for 0::: x::: n,

for n < x ::: 2n,

verify by direct integration of the Fourier coefficient formulas for bn , an that the complete Fourier series
off is

2. + 2. sin (x) _ ~~ cos(2nx) .

n 2 ~ 4n 2 - 1
n n=l

Convergence Theorem. Apply the Fourier convergence theorem to the functions f of the two preceding
exercises in order to verify the following results.
1 1 1 1
1" + 3I + 52 + 7f + ....
Jr2
3. 8' =

4. ~ = f - ~ + ! - t + " ..
1 ,",00 1
S. '2 = L-n=l (2n-l)(2n+l)'

6. 1= 1 + )23 - 325 + 5\ - 19 + " ..

7. (Uniqueness for the Fourier Ring Problem) Consider the Fourier ring problem

U(x, t) is defined for 0 :'S x ::: 2n, t ::: 0,

 8.6 PROOFS OF THE CONVERGENCE THEOREMS 501

Ut = Uxx for 0::: x::: 2rr and t ~ 0,

u(O, t) = u(2rr, t) and ux(O, t) = ux(2rr, t) for t ~ 0,
u(x,O) = j(x) for 0::: x::: 2rr,

where j(x) is a prescribed initial temperature distribution. Prove that this problem has at most one solution.

8.6 Proofs of the Convergence Theorems (Optional)

The present section presents a proof of the convergence theorem for the complete Fourier
series. The corresponding theorems for Fourier sine and cosine series are then obtained
as simple corollaries.
Periodic Functions. The periodicity of the component functions cos(nx) and sin(nx)
(with n an integer) plays an important part in the proof of the convergence theorem.
Recall from Chapter 4 that a function f(x) has period p if and only if

(1) f(x + p) = f(x) , -00 <x< 00.

It follows on replacing x by x - p that

f(x - p) = f(x) , -00 <x< 00.

Similarly, for any integer n,

f(x + np) = f(x), -00 < x < 00.

This last relation may be derived from (1) by mathematical induction.

A periodic function is determined by its values over one period. More precisely, if
f(x) has period p and fa is a given function on 0 :::: x :::: p such that

(2) f(x) = fa (x) for 0:::: x < p

then one has

f(x) = f(x - p) = fa (x - p) for p :::: x < 2p,

f(x) = f(x - 2p) = fa (x - 2p) for 2p :::: x < 3p,

This procedure generates the unique p-periodic function f whose restriction to 0 :::: x < p
is fa. Conversely, any p-periodic function can be generated in this way from its values on
o :::: x < p or any other interval of length p.
502 8 FOURIER ANALYSIS AND STURM-lioUVILLE THEORY

A useful property of periodic functions is

(3) l a p
+ f(x)dx = i P
f(x)dx for all real a.

To verify this we note that if lea) is the function defined by

lea) = l
a
a+p
f(x)dx,

then

I'(a) = f(a + p) - f(a) = 0

for all a by the periodicity property (1). Thus lea) = leO) (a constant) for all a.
Now letfo be defined for 0::: x < 2n and letf be its 2n-periodic extension. Suppose
that

= ao + L
00

fo(x) (an cos(nx) + bn sin(nx)) for 0::: x < In.

n=l

Then

+L
00

(4) f(x) = ao (an cos(nx) + bn sin(nx)) for all real x.

n=l

This follows because both f(x) and the sum of the Fourier series are 2n-periodic
extensions offo(x), and we have seen that such an extension is unique.
The coefficients an and bn above are the Fourier coefficients of fo on 0 ::: x < 2n.
Since fo(x) = f(x) on this interval, we may write

ao = -
1 l2Jr f(x)dx,
2n 0

an = -
1 l2Jr fCx) cos(nx)dx, n ::: 1,
n 0

bn = -
1 l2Jr fCx) sin(nx)dx, n ::: 1.
n 0

Equation (3) with P = 2n implies that we may replace the integration interval [0,2nl
with any interval of length 2n. This freedom to change the interval will be used below.
Frequently, a convenient choice is the symmetric interval [-n, n I:

(5) ao = -
1 fJr f(x)dx,
2n -Jr
 8.6 PROOFS OF THE CONVERGENCE THEOREMS 503

1 f11:
an = - JCx) cos(nx)dx, n ::: 1,
-11:
:rr:

bn = -1 f11: f(x) sin(nx)dx, n ::: 1.

:rr: -11:
Partial Sums. The nth partial sum of the Fourier series (4) for f(x) will be denoted
by Sn(x). Thus

L (ak cos(kx) + bk sin(kx)).

n
(6) Sn(x) = ao +
k=l

The convergence theorem of Section 8.5 will be proved by showing that iff and l' are
sectionally continuous then

(7) lim Sn(x)

n->!Xl
= ~[f(x + 0) + f(x - 0)]

for all x. The first step is to replace ak and h in (6) by the integrals of (5) and simplify.
This gives

ak cos(kx) + bk sin(kx)
= (~ l>(U) COS(kU)dU) cos(kx) + (~ l>(U) Sin(kU)dU) sin(kx)
1 f11:
= - JCu) (cos(ku) cos(kx) + sin(ku) sin(kx» du
:rr: -11:

= -1 f11: feu) cos k(u - x)du.

:rr: -11:
Substituting this and the integral formula for ao into (6) gives

(8) 1
Sn(x) = - f11: feu)
:rr: -11:
(1 + L
-
2
n

k=l
cos k(u - x) ) duo

A Trigonometric Formula. The next step is to evaluate the sum of cosines in (8).
The result is
1 ~ sin(n + ~)t
(9) - + ~ cos(kt) = . I ' n=1,2,3, ....
2 k=l 2sm(2:t)
To verify this formula we recall the identity

2 sine ~t) cos(kt) = sin(k + ~)t - sin(k - ~ )t, k = 0,1,2, ....

Summing this formula over k = 1, 2, ... , n gives

2sin(~t) (cos(t) + ... + cos(nt» = (sin(~t) - sin(~t»)

504 8 FOURIER ANALYSIS AND STURM-lioUVILLE THEORY

+ (sin(~t) - sin(N)
+ ... + (sin(n + ~)t - sin(n - ~)t)
= sin(n + ~)t - sin(~t),

since all but the last two terms cancel (a telescoping sum). Dividing by 2sin(~t) and
adding 1/2 gives (9).

Integral formula for Sn(x). Combining (8) and (9) gives

_ 1
Sn(x) - -
lr
l Jr

-Jr
J(u)
sin(n + ~)(u - x)
. 1
2sm2:(u-x)
duo

The change of variables t = u - x in the integral gives

Sn (x) = -
1 l Jr - x
J(t + x)
sin(n
.
+ !)t
12 dt.
lr -Jr-X 2 sm( 2:t)

The integrand in this integral is periodic in t with period 2lr (check this). Hence property
(3) gives

1 jJr sin(n + !)t

(10) Sn(x) =- J(t + x) . 12 dt.
lr -Jr 2sm(2:t)

A similar integral formula can be written for the difference

Sn(X) - ~[J(x + 0) +J(x - 0)].

To find it note that integration of (9) over [-lr, Oland [0, lr 1 gives

2
r
~ = .. !:. sin(n + ~ )t dt
lri-Jr 2sin(~t) ,

~ =..!:. r sin(n + ~)t dt.

2 lr Jo 2 sin(~t)

If we multiply these equations by J(x + 0) and J(x - 0), respectively, and subtract from
(10) then we get the representation
(11)
Sn(x)-~[J(x+O)+J(x-O)l = -
i
1 Jr J(x + t) - J(x + 0) . ( 1) d
sm n + - t t
lr a

+ -1
lr
1
2sin(~t)
0

-Jr
J(x
2

+ t) .-
2 sm(2:t)
J(x - 0) . (
1
d
sm n + 2: t t.
1)
 8.6 PROOFS OF THE CONVERGENCE THEOREMS 505

The convergence theorem will now be verified by showing that the last two integrals
have limit 0 as n -+ 00. The proof will be based on a property of the Fourier coefficients
called Bessel's inequality:
Bessel's Inequality. Letfex) be sectionally continuous on -1T :::; X :::; 1T, and let an,
bn be its Fourier coefficients, defined by (5). Then

(12) 2a~ + 8 (a~ + bD : :; -;; fT(_J ex)dx.

(Xl 1
2

Proof. This result is called Bessel's inequality. To prove it, let Sn (x) be defined by (6),
as before, and consider the equality

The left side is nonnegative, giving

(13) 2 i>ex)SneX)dx - i: S~(x)dx:::; i:fex)dx.

Next, the definition of ak and bk and orthogonality esee (21)-(22), Section 8.5) imply that

(14) i>(X)Sn(X)dx = i: S~(x)dx = 21Ta6 + 1T ~ (a~ + b~).

Combining (13) and (14) gives

(15) n=1,2,3, ....

Thus the partial sums of the infinite series in (12) form an increasing sequence with finite
upper bound. It follows that the series converges and satisfies Bessel's inequality (12).
The nth term in a convergent series must tend to zero as n -+ 00. Thus Bessel's
inequality implies that a~ + b~ -+ 0 as n -+ 00. This result is called Riemann's theorem.
It can be formulated as follows.
Riemann's Theorem. If fex) is sectionally continuous on -1T :::; X :::; 1T then

(16) n1i~ i>CX) sinCnx)dx = }!..~ i>ex) cosenx)dx = O.

i: i:
The integrals in (11) have the form

get) sin(n + ~)t dt = (g(t) sine ~t)) cos(nt)dt

506 8 FOURIER ANALYSIS AND STURM-LIOUVILLE THEORY

+ i:
Combining this with Riemann's theorem gives the following result.
(g(t) cos(!t)) sin(nt)dt.

Corollary. If get) is sectionally continuous on -7r ::: t ::: 7r then

(17) lim
n-+-oo
11< get) sin(n + !)t dt = O.
-Jr

Equation (11) for Sn (x) can be written as

(18) Sn (x) - ! [J(x + 0) + J(x - 0)] = - 111< g(x, t) sin(n + !)t dt,
7r -1<

I
where g(x, t) is defined for all x and for 0 < t ::: 7r by

(19) (x t) = ! [J(x + t) - J(x + 0)]1 sin(!t), o < t ::: 7r,

g, ![J(x+t)-J(x-O)]!sin(!t), -7r ::: t < o.
Application of Riemann's theorem to (18) gives the follOwing extended version of the
convergence theorem of Section 8.5.
Convergence Theorem for Periodic Functions. Let J(x) have period 27r and let
J(x) andf'(x) be sectionally continuous on the interval 0:::: x::: 27r (and hence on any
finite interval). Then the complete Fourier series of J converges for every real x and it
has the sum

L (an cos(nx) + bn sin(nx)) = ! [J(x + 0) + Jex -

00

(20) ao + 0)].
n=l

Proof. For each fixed x we apply Riemanns theorem, in the form (17), to the integral
in (18). The hypothesis that J is sectionally continuous implies that g(x, t), as a function
of t, is sectionally continuous except possibly at t = o. the hypothesis thatJ andf' are
sectionally continuous implies that
Jex + t) - Jex + 0) J(x + t) - Jex + 0) t
2sin(!t) 2sin(!t)
has the limit
lim Jex + t) - J(x + 0) =l' (x)
HO+ 2 sine It)
2
+

at every point x. Similarly,

lim J(x + t) - J(x - 0) =l' (x).

HO- 2sin(1t) -
2
 8.6 PROOFS OF THE CONVERGENCE THEOREMS 507

Hence if we define g(x, 0 ± 0) = f~ (x) then g(x, t) is in fact sectionally continuous for
-Jr :::: t :::: Jr and the proof is complete.

The convergence theorems for Fourier sine and cosine series of Section 8.3 and
Section 8.4 may be proved as corollaries of the convergence theorem of this section. The
proof makes use of the notion of even and odd functions.

Even and Odd Functions. Consider functions f(x) that are defined on a symmetric
interval -L :::: x :::: L We say that

(21) f(x) is even {} f( -x) = f(x) for - L :::: x :::: L,

f(x) is odd {} f( -x) = -f(x) for - L :::: x :::: L

Examples of even functions are cos(nx), Ixl and x2n. Examples of odd functions are
sin(nx), X 2n+ 1 and x cos(nx). Several simple but useful rules are listed next. The proofs
are left for the Exercises.
Rule 1. Linear combinations of even (respectively, odd) functions are also even
(respectively, odd)
Rule 2. A product of two even functions is even. A product of two odd functions is even.
A product of an even function and an odd function is odd.
Rule 3.

(22) f(x) is odd => l>cx)dx = 0,

f(x) is even => jL f(x)dx

-L
= 2 [L fCx)dx.
10
Rule 4. Iff(x) is even and1'(x) exists then1'(x) is odd. Uf(x) is odd and1'(x) exists
then l' (x) is even.
Rule 5. Every function f(x) on 0 :::: x :::: L has a unique even extension g to - L :::: x :::: L,
defined by

C) = /f(X) for 0:::: x:::: L,

(23)
gx fe-x) for -L:::: x:::: o.

Similarly; every function f(x) on 0 :::: x :::: L that satisfies f(O) = 0 has a unique odd
extension h to -L :::: x :::: L, defined by

hex) = /f(X) for 0:::: x:::: L,

(24)
-fe-x) for -L:::: x:::: o.
508 8 FOURIER ANALYSIS AND STURM-lioUVILLE THEORY

Next, let J(x) satisfy the conditions of the convergence theorem for periodic functions
and letJ be even. Then for L = rr,

bn = -1 1][ J(x) sin(nx)dx =0 for n = 1,2,3, ... ,

rr _IT

by Rules 2 and 3. Hence the Fourier series of J is a pure cosine series:

00

(25) ao + Lan cos(nx).

n=l

Moreover, Rules 2 and 3 give

(26) ao = -1
2rr
1][ J(x)dx = -1
rr
1" J(x)dx,

1][ 21][ J(x) cos(nx)dx.

_IT 0

an = -1 J(x) cos(nx)dx = -
rr _IT rr 0

These equations define the Fourier cosine series of Section 8.4.

Similarly, let J(x) satisfy the conditions of the convergence theorem and let J be odd.
Then for L = rr,

an 1
= -
rr
1][ _IT
J(x) cos(nx)dx = 0 for n = 1,2,3, ... ,

by Rules 2 and 3. Hence the Fourier series of J is a pure sine series:

L b sin(nx).
00

(27) n
n=l

And again by Rules 2 and 3,

(28) bn = -
1 1][ J(x) sin(nx)dx = -
21][ J(x) sin(nx)dx.
rr _IT rr 0

These equations define the Fourier sine series of Section 8.3.

By reversing the arguments given above we may prove the convergence theorems of
Section 8.3 and Section 8.4 as corollaries of the convergence theorem of this section.
Proof of the Convergence Theorem for Fourier Sine Series. Given J(x) on
o ::: x ::: rr with J and l' sectionally continuous, we extend J to be odd and periodic with
period 2rr (unique extension). Then the Fourier series of (20) reduces to the sine series
of (16), Section 8.3, and the coefficients bn given by (28) are the same as in Section 8.3.
Proof of the Convergence Theorem for Fourier Cosine Series. For this
problem extend J to be even and periodic with period 2rr. Then (20) reduces to the cosine
 8.6 PROOFS OF THE CONVERGENCE THEOREMS 509

series of (17), Section 8.4, and the coefficients an given by (26) are the same as in Section
8.4.
Other Periods. The preceding results are easily extended to functions f(x) having
any period p > O. It will be convenient to write p = 2L, so that

(29) f(x + 2L) = f(x) for all x.

The trigonometric functions with this period are sin(nrrxlL) and cos(nrrxlL). Hence we
expect the Fourier expansion of f(x) to have the form

8 (an
00

(30) f(x) = ao + cos C~x) + bnsin C~x)).

The theory of these series is readily derived from the convergence theory proved above.
To do this introduce a new independent variable v by
rrx Lv
(31) v=- x=-
L rr'
and define

(32) g(v) = f (rr;) = f(x).

Then g(v) has period 2rr:

(33) g( v + 2rr) = f (~ + 2L ) = f (~) = g(v).

Hence if g and g' are sectionally continuous then

(34)
1
2: [g(v + 0) + g( v - 0) I = ao + L (an cos (nv) + bnsin (nv)) ,
00

n=l

where

(35) ao = - 1
2rr
l Jr
g(v)dv,

l
-Jr

an = -1 Jr
g(v) cos(nv)dv, n ::: 1,
rr

l
-Jr

Jr
bn = -1 g(v) sin(nv)dv, n ::: 1.
rr -Jr

Making the change of variable (31) in (34) and (35) gives

L (an cos Crrx ) + bn sin C

00

(36) ! [f(x + 0) + f(x - 0) I = ao +

n=l L
rrx ) ) ,
L
510 8 FOURIER ANALYSIS ANO STURM-LIOUVILLE THEORY

where

(37) ao = 2~ l>(X)dx,
an =
L
i
~ Lj(x) cos (n1l'X) dx,
L
n ~ 1,

i
-L

bn = ~L L

-L
j(x) sin C1l'X) dx,
L
n ~ l.

This implies the follOwing generalization of the convergence theorem.

Convergence Theorem for Period p = 2L. Let j(x) have period p = 2L and let
j(x) andj'(x) be sectionally continuous on the interval -L ::: x ::: L (and hence on any
finite interval). Then for every real number x, equation (36) holds, where an and bn are
given by (37).

Exercises 8.6

1. (Periodicity) Letf(x + p) = f(x) for allx. Show thatf(x + np) = f(x) for all x and all integers n.
2. (Partial Sum Identity) Verify the identity Sn(x) - i [{(x + 0) + f(x - O)J = II + 12, where

I 1 = -1 !O'" fex + t) - fex

1
+ 0) . ( + 1) d
sm n 2: t t,
7r 0 2 sin ( 2:t)

h=-
1 fO f(x + t) - f(x - 0) . (
1
1) d
smn+2: t t.
7r -1C 2sin(2:t)

Even and Odd Functions. Prove the follOwing rules.

3. (Rule 1) Iff and g are odd functions and C1 and C2 are constants then cif + c2g is odd.
4. (Rule 2) Iff and g are odd functions thenfg is even.
5. (Rule 2) Iff and g are even functions thenfg is even.
6. (Rule 2) Iff is odd and g is even then fg is odd.

7. (Rule 3) Iff(x) is odd then I~d(x) dx = O.

8. (Rule 3) If f(x) is even then lLf(x) dx = 2 It f(x) dx.

9. (Rule 4) Letf(x) be even and assume thatj'(x) exists. Show thatj'(x) is odd.
10. (Rule 4) Letf(x) be odd and assume thatf'(x) exists. Show thatf'(x) is even.
Convergence Theorems. Prove the following results about the convergence of Fourier series.
11. (Functions of Period 2L) Let f(x+ 2L) = f(x) and assume that f(x) and f' (x) are sectionally continuous
on (-00, (0). Let w,. = n7r/L. Verify that for every real number x,

+ 0) + f(x -
f(x
2
0)
= ao + L (an) cos (lLIn
00
X) + bn sin (w,.x),
n=1
 PART II FOURIER INTEGRALS 511

where

ao = ~ fL j(x)dx, an = ~ fL j(x) cos (WnX) dx,

2L -L L -L

bn = ~ fL j(x) sin (wnx) dx, n:::: l.

L -L

Fourier Convergence Theorem. The following exercises establish a special case of Fouriers theorem,
given below The main feature of this result is an explicit convergence rate at points where F is very smooth.

Theorem. Let F be piecewise continuous and 2Jr-periodic. Assume x given and define

C(u) == (F(x + u) - F(x»1 sin(ul2).

If F is twice continuously differentiable at x then the partial sums Sn of the Fourier series of F
satisfy the following identities:

I fIT (
sn (x) = - Fx + u)sin «2n + l)ul2)d
u,
2JT _" sin (ul2)
sn(x) - F(x) = - I
2Jr _"
f" C(u) sin«2n + l)ul2)du,

=
(2n +1 l)Jr f"-IT
C' (u) cos«2n + 1)ul2)du.

12. Establish Dirichlet's trigonometric identity

sin«2n+l)ul2) I ~ (h
Dn ()
u == = - + L..J cos u).
2 sin (ul2) 2 k=l

13. Integrate Dirichlet's trigonometric identity over [-Jr, Jr] to obtain f~" Dn (u)du = Jr.
14. Assume that x is fixed and C(u) == [F(x+u) - F(x)]/ sin(ul2) is continuously differentiable on [-Jr, Jr]
Establish using integration by parts that

-
1 f" C(u) sin«2n + l)ul2)du =
~ J~" c' (u) cost (2n + 1 )ul2)du
.
2Jr -IT 2n + 1

M
15. Conclude from the preceding exercises that in the convergence formula F(x) = sn(x) + - - the
2n +1

L:
value of M is the integral

M = ~1 c' (u) cos«2n + 1)ul2)du,

which is bounded by 2 max IC' (u) I on lui:::: Jr.
16. Verify that the last equality of the theorem can be written as F(x) = Sn (x) +M/(2n+ 1), with M a bounded
quantity. Explain why this equality implies that Sn uniformly approximates the graph of F as n -+ 00.

Part II Fourier Integrals

Part II develops the basic properties of Fourier integrals in the context of heat diffusion
in homogeneous rods of infinite length. Fourier sine integrals arise in the analysis of
512 8 FOURIER ANALYSIS ANO STURM-LIOUVILLE THEORY

heat flow in a semi-infinite rod with Dirichlet boundary conditions at the free end.
The analogous problem with Neumann boundary conditions leads to the Fourier cosine
integral. The doubly infinite rod leads to the full Fourier integral with both sine and
cosine terms.

8.7 Heat Diffusion in an Infinite Rod

Consider the problem of predicting the linear flow of heat in a long thin rod when
attention is restricted to points far from the two ends. A useful model is the idealized
problem of linear heat flow in a doubly infinite rod. Thus if x is a distance coordinate,
measured from a fixed point on the rod, then the temperature distribution in the rod at
time t is a function u(x, t) that satisfies

(1) u(x, t) is defined for -00 < x < 00, t ::: 0,

(2) Ut = Uxx for -00 <x< 00, t ::: 0,
(3) U(x,O) = j(x) for -00 <x< 00,

where j(x) is a prescribed initial temperature distribution. For this problem there are no
explicit boundary conditions. However, the following assumptions are made.

(4) U(x, t) is bounded for -00 < x < 00, t ::: 0,

(5) f~oo If(x)ldx < 00.

Condition (4) means that

Umin S u(x, t) S Umax for - 00 < x< 00, t::: O.

To interpret (5) let c be the specific heat, p the mass density and A the cross-sectional
area of the rod. Then, as in Section 7.2, (equation (6)),

cpA lb u(x, t)dx

is the heat content of the portion a S x S b of the rod. Hence, (5) means that the heat
content of the rod at time t = 0 is finite.
Separation of Variables. As before, the search for a solution begins with the
construction of all product solutions u(x, t) = X(x)T(t) that satisfy 0), (2) and (4).
Conditions (3) and (5) are left for the superposition step. The usual separation argument
shows that T/(t) = -AT(t) and XI/(x) = -AX(x), where A is the separation constant.
This suggests the following eigenvalue problem.
 8.7 HEAT DIFFUSION IN AN INFINITE Roo 513

Eigenvalue Problem. Find the values of A for which

(6) X"(x) + AX(X) = 0, for - 00 < x < 00,

has nontrivial bounded solutions.

To solve the problem it is necessary to consider the usual three cases.
Case 1 (A < 0). If A = _fL2 < 0 with fL > 0 then

X(X) = A el1 x + Be- l1x .

It is clear that X(x) is bounded for x ---+ 00 only if A = O. Similarly, X(x) is bounded for
x ---+ -00 only if B = O. Thus X(x) = 0 and there are no eigenvalues for A < O.
Case 2 (A = 0). Here

X(x) = A + Bx,
which is bounded if and only if B = O. Thus A = 0 is an eigenvalue with eigenfunction
X(x) = 1 (a nonzero function).

Case 3 (A > 0). If A = fL2 with fL > 0 then

(7) X(x) = A cos fLX + B sin fLX,

which is bounded for any choice of A and B.
To summarize, the eigenvalues and eigenfunctions of the problem are

(8) A= fJ2,

X(x) = A(fL) cos fLX + B(fL) sin fLx, fL 2: 0,

where A(fL) and B(fJ) are arbitrary functions of fL. The set of eigenvalues, often called
the spectrum of the problem, is the set of all nonnegative real numbers.
Product Solutions. The T(t) associated with (8) satisfies T'(t) = _fJ2T(t) and hence
2
T(t) = e- 11 t Thus the product solutions are the one-parameter family of functions

(9)
Superposition. It is natural to seek a solution of the boundary value problem (1)-(5)
by integrating over the parameter fJ in (9). Thus we seek to determine functions A(fL)
and B(fL), fL 2: 0, such that the problem has the solution

(10) u(x, t) = E OO
[A(fJ) cos(fJx) + B(fL) sin(fLx)]e-112t dfJ.
If A(fL) and B(fL) are integrable over 0 :::: fL < 00 then (10) satisfies all conditions of the
problem except the initial condition u(x, 0) = f(x). This will be satisfied as well if f(x)
514 8 FOURIER ANALYSIS AND STURM-LIOUVillE THEORY

has a representation

(ll) j(x) = 1 00
[A(JL) cOS(JLX) + B(JL) sin(JLx)] dJL, -00 <x< 00.

When is this possible? Fourier suggested that most integrable functions j(x) can be so
represented and he gave a method to calculate A(JL) and B(JL) whenj(x) is known. A
rigorous theorem and proof were first given by Dirichlet. His theorem will be stated and
proved below.

8.8 Orthogonality Calculation

In Section 8.5 the orthogonality on 0 ::::: x ::::: 277: of the eigenfunctions cos(nx), sin(nx)
was used to calculate the Fourier coefficients of a function j(x). In this section an
analogous calculation is used to calculate the coefficient functions A(JL), B(JL) in the
equation

(1) j(x) = 1 00
[A(JL) cOS(JLX) + B(JL) sin (JLx) ] dJL, -00 < x< 00.

lt is assumed that j(x) is sectionally continuous and If(x) I is integrable over -00 < x <
00. It follows that the integrals

(2) L:j(X) cos(vx)dx, L:j(X) sin(vx)dx

are finite for all real values of v. They will be calculated by multiplying (1) by cos(vx) or
sin(vx), integrating over -N ::::: x ::::: N and then calculating the limit as N -+ 00. The
first step gives

1
i:
00
(3) Lj(X) cos(vx)dx = (A(JL) L: cos(JLx) cos (vx) dx) dJL

+ 1 00
(B(JL) sin (JLx) cos (vx) dx ) dJL.

Now, cos(vx) and sin(vx) are even and odd functions, respectively. Hence

j
-N
N cos(JLx) cos( vx)dx = 2 r
10
cos(JLx) cos( vx)dx

= sin(JL - v)N
JL-V
sin(JL + v)N
+--'---
JL+V
 8.8 ORTHOGONALITY CALCULATION 515

and

I -N
N sin(j.tx) cos( vx)d.x = o.
Substituting these into (3) gives

I N
j(x) cos(vx)d.x =
100

0
A(j.t)
sin(j.t - v)N
j.t-V
dj.t

1
~

)sin(j.t + v)N d
+ 00 (
A j.t j.t.
o j.t+V
If we define A(j.t) = 0 for j.t < 0 then the last two integrals have the form
(4) 1 00

-00
(
A j.t
) sin N(j.t - x) d
j.t - x
j.t, x = ±j.t.

Limits of such integrals were first studied by Dirichlet. The following theorem will be
proved in Section 8.9 below.

i:
Dirichlet Limit Theorem. Let A(j.t) and A'(j.t) be sectionally continuous and

IA(j.t)ldj.t < 00.

Then

(5) lim
N..... oo
1 00

-00
A(j.t)
sin N(j.t - x)
j.t - x
dj.t
n
= -2 [A(x + 0) + A(x - 0) J.

In particular, if A(j.t) is continuous at j.t = x then the limit is nA(x).

Apply this theorem to the right-hand side of (4) with v > 0, A(j.t) continuous for
j.t > 0 and A(j.t) = 0 for j.t < O. The result is

(6) lim
N-+oo
IN j(x) cos(vx)d.x = nA(v).
-N

An exactly parallel calculation with cos(vx) replaced by sin(vx) gives

(7) lim
N-+oo
IN j(x) sin(vx)d.x = nB(v).
-N

The finiteness of the integrals (2) permits us to rewrite these equations as

(8) A(v) =; 11 00
_ooj(x) cos(vx)d.x,

B(v) = - 1/
n
00

-00
j(x) sin(vx)d.x.
516 8 FOURIER ANALYSIS AND STURM-LIOUVILLE THEORY

In summary, we have shown that if J(x) has the representation

(9) J(x) = 1 00
[A(/L) cos (/LX) + B(/L) sin (/Lx) ] d/L, -00 <x < 00,

then A and B must be given by (9). In the next section the converse problem is solved.

8. 9 The Fourier Integral

The integral

1 00
[A(IL) COS(ILX) + B(/L) sin(/Lx)] d/L, -00 < x< 00,

where A(/L) and B(/L) are defined by (8) in Section 8.8, is called the Fourier integral of
J(x). The main result of this section is the follOwing theorem.
Fourier Integral Theorem. LetJ(x) satisfy the conditions

(1) J(x) andf'(x) are sectionally continuous for -00 < x < 00,
(2) J(x) = ~ [J(x + 0) +J(x - 0)] for -00 <x < 00,

(3) J~oo [f(x)ldx < 00.

Then for -00 < x < 00,

(4)

where

(5) B(/L) = - 11
:n: -00
00
J(x) sin(ILX)dx.

A function is said to be sectionally continuous on an infinite interval if it is so on

any finite subinterval. Thus (1) implies thatJ(x) andf'(x) have only a finite number of
jump discontinuities in any finite interval. Condition (2) holds automatically at points of
continuity ofj. Thus the condition is simply a normalization that simplifies the statement
of the theorem. In particular, it does not alter the values of A(/L) and B(IL). Finally, an
alternative version of the Fourier integral (4) is given by the follOwing corollary.
Corollary. Under conditions (1)-(3) of the Fourier integral theorem, one has

(6) J(x) = ~ 1 (i:

00
J(I;) cos IL(X - I;)dl;) dIL

for -00 < x < 00.

 B.9 THE FOURIER INTEGRAL 517

Proof of the Corollary. The integral representations (5) for A(JL) and B(JL) give
1 [CO
A(JL) cos(JLx) + B(JL) sin(JLx) = ~ Lco fW [cos(JL~) cos(JLx) + sin(JL~) sin(JLx)] ~
1 [CO
= ~ J_cof(~) COSJL(x - ~)~.

Combining this and (4) gives (6).

A proof (optional) of the Fourier integral theorem is given at the end of the section.
First, several examples will be discussed.

EXAMPLE 1. Consider the Single-pulse square wave defined by

I Ixi < 1,
(7) f(x) = { 1/2 Ixi = 1,
o Ixi > 1.

See Figure 5. It satisfies conditions (1)-(3) (check this). Moreover,f(x) is even, and so

(8) A(JL) =- 111

n -1
cos(JLx)dx = -2 --,
n
sinJL
JL
JL > 0,

and B(JL) == O. Thus the convergence theorem implies that

(9) f(x) =~
n
l
0
co
sin JL cos(JLX) dJL
JL
for - 00 <x< 00.

Taking x = 0 gives the equation

(10) [CO sin JL dJL = ~.
Jo JL 2
Independent derivations of this integral formula can be found in many advanced calculus
books.
The Fourier integral (9) will now be checked directly; starting from (lO). Note that
since both f(x) and cos(JLx) are even functions, it will suffice to check (9) for x > O.
The evaluation of the integral in (9) will make use of Si(x), the sine integral function,

f x
FIGURE 5. Asquare wave on (-00,00)
518 8 FOURIER ANALYSIS AND STURM-lioUVILLE THEORY

defined by

(ll) Si(x) == io
x sinlL
--dlL.
IL
It is an odd function and it satisfies
7r
(12) Si(O) = 0, lim Si(x) =
X~±OO
±-,
2
by (10).
Equation (9) means that

(13)

where

(14) IN X
( ) =_ -2iN sin IL COS(ILX) dIL.
7r 0 IL
To prove (9) recall that

2 sin(lL) COS(ILX) = sin 1L(l + x) + sin 1L(l - x).

Thus

(15) IN(x) =..!:.

7rio
r sin 1L(l
IL
+x) dlL +..!:.
trh
r sin 1L(l - x) dlL.
IL
The change of variables IL( 1 + x) = t (respectively, IL( 1- x) = t) in the first (respectively,
second) integral gives

(16) IN(x) =-
1 iNCl+ X) sin(t)
--dt + -
1 iNO-X) --dt
sin(t)
7r 0 t 7r 0 t
1 1
=- Si(N(l + x)) + - Si(N(l - x)).
7r 7r
The existence of the limit as N ~ 00 is now apparent from (12). To verify (13) for x ~ 0
three cases arise, as follows.
17r 17r
O::;x<l: lim IN(x) = - - + -- = 1 = j(x),
N~oo 7r 2 7r 2
1 7r 1 1
x = 1: lim IN(x) = --
N~oo 7r 2
+ -·0
7r
= -
2
=j(x),

1 7r 1 -7r
x> 1: lim IN(x) = - - + - - = 0 = j(x).
N ..... oo 7r 2 7r 2
This completes the verification.
 B.9 THE FOURIER INTEGRAL 519

Gibbs Phenomenon. In Figures 6, 7 and 8 below, IN(x) is plotted for N = 16,32

and 64.
1.0
n,
O.S O.S O.S

0.6 0.6 0.6

0.4 0.4 0.4
0.2 0.2 0.2
VVI
-1 -0.5 0.5 1 -1 -0.5 0.5 1 -1-0.5 0.5 IV Vv

It is seen that as N increases, there are rapid oscillations near the discontinuities of
f(x) at x = ±1. This behavior is called the Gibbs phenomenon. It can be understood
from the representation (16). The maximum and minimum of Si(x) are easily seen to be

Si(±n) ~ ±1.8519.

Thus
1
= -[Si(2N - n) + Si(n)]
n
~~ [i + Si(n)] = 1.0895

for N --+ 00, and similarly,

1
IN (1 +~) = - [Si(2N + n)
n
+ SiC -n)]

~~ [i - Si(n)] = -0.0895.
Thus the maximum and minimum of IN(x) overshoot those of f(x) by a fixed amount
and they occur at the points 1 ± nlN, which approach 1 as N --+ 00. The total overshoot
in dropping fromf(1 - 0) = 1 to fO + 0) = 0 (left- and right-hand limits) is

IN ( 1 - ~) - IN ( 1 + ~) ~ 1.1790,
which exceeds the jump fO - 0) - fO + 0) = 1 by 17.9 percent.
EXAMPLE 2. Define f(x) by

f(x) = jne-x x> 0,

~/2 x=O,
x < O.
520 8 FOURIER ANALYSIS AND STURM-lioUVILLE THEORY

Then j satisfies the conditions of the convergence theorem. The Fourier coefficients of
j(x) are given by

A(fL) = -1
n
1 00
j(x) COS(fLX)dx = 1 0
00
e- X cos(fLx)dx,

1 1
-00

B(fL) = -1 00
j(x) sin(fLX)dx = 00
e-X sin(fLX)dx.
n -00 0

These integrals can be calculated from tables. In fact, they were worked out in Chapter 4
as the Laplace transforms

(17) .c{COS(fLt)} = 1 o
00
e- SX COS(fLX)dx =
s
-2--2'
S + fL
.c{sin(fLt)} = roo e-
10
sx sin(fLx)dx = A.
+S fL
Setting S = 1 gives the results

B( ) _ _ fL_
fL - 1 fL2 +
and the convergence theorem gives

1 o
00 COS(fLX) + fL sin (fLX) d
1 + fL2
fL =
£()
J X =
I ne-
n12
°
x
x > 0,
x = 0,
X < 0.

EXAMPLE 3. Define j(x) = e- alxl for all x and a > 0. Then j is even, so B(fL) °
== and

A(fL) =- 21
e- ax COS(fLX)dx = - 2 a 2
00

n o n a +fL
2
by (17) with s = a. Hence the convergence theorem gives

1 !:..
00 a COS(fL
2
o n a +fL
X) d = £( ) = -alxl
2 fL JX e , -00 < x< 00.

EXAMPLE 4. Define j(x) by

l~-ax
for x> 0,
j(X) = for x =0,
_eax for x < 0,
 8.9 THE FOURIER INTEGRAL 521

where a > O. Then f satisfies the conditions of the convergence theorem and f is odd.
Hence A(/I) = a and
B(/I) = -21
IT 0
00
e- ax sin(/Ix)dx = 2
- 2 /I 2
ITa +/I
by (17) with 5 = a. Hence

roo ~ /I sin (/Ix) d/I = f(x) = I~-ax : : ~:

10 IT a2 + /I 2 -e
ax x < O.

EXAMPLE 5. Define f(x) = e- Ixl cos(x) for all x. Then f satisfies the conditions of the
convergence theorem and f is even. Hence B(/I) = a and
A(/I) = - 21°C e- x cos(x) cos(/Ix)dx.
IT 0

The identity

2 cos(x) cos(/Ix) = cos(x + /Ix) + cos (x - /Ix)

implies that

A(/I) = - 11
IT 0
00
e- X [cos(x + /IX) + cos(x - /IX)] dx.

This may be evaluated by means of the Laplace transform

L (cosO ± fL)t) = 1o
00
e- st cosO ± /I)tdt =
5
2
5

+ (1 ± fL)
2'

The result is
1 1 1 1
A(fL) = - + -----
IT 1 + (l + /I)2 IT 1 + (l - /I)2 '
or, after some algebraiC manipulation,

Thus, by the convergence theorem,

1
o
00
--4--
IT/I
+
2 /I 2 2
+4
cos(/Ix)d/I = f(x) = e- Ixl COS(X), -00 < X < 00.
522 8 FOURIER ANALYSIS AND STURM-lioUVILLE THEORY

EXAMPLE 6. Define f(x) to be the odd extension of e- X cos(x), so

e-X cos(x) x > 0,

f(x) = { 0 x = 0,
-r cos(x) x < o.

Then f satisfies the conditions of the convergence theorem and f is odd. Hence A(JL) = 0
and

B(JL) = - 21
rr 0
00
e- X cos (x) sin(JLx)dx.

The identity

2 cos(x) sin(JLx) = sin(JLx + x) + sin(JLx - x)

gives

B(JL) =-11
rr 0
00
e- X [sin(JLx + x) + sin(JLx - x) 1dx.

Hence the Laplace transform

.c (sin(JLt ± t» = 1 o
00
e- st sin(JL ± l)t dt = 2
s+JL±l
JL±l
( )2

gives
1 JL+1 1 JL-1
B(JL) = - +- ,
rr 1 + eJL + 1)2 rr 1 + eJL - 1)2
or, after algebraic simplification,

Finally,

1
e-X cos(x) x> 0,
2 JL3 {
00
- - 4- sin(JLx)dr = fex) = o x =0,
o rrJL +4
-r cosCx) x< o.

EXAMPLE 7. Let a> 0 and define

2
fCx) = e- ax for - 00 < x < 00.
 8.9 THE FOURIER INTEGRAL 523

Thenf is even, whence B(p,) = 0 and

A(p,) = - 21 00
JT 0
e- ax 1 cos(p,x)dx.

This integral is not elementary However, it can be evaluated by an indirect method.

Differentiating A(p,) with respect to p, gives

~A'(p,) = _ roo e-ax\sin(p,x)dx.

2 10
Then integration by parts with

u = sin(rx), du = p, cos(p,x)dx,
1 _ax"
V= -e
2a '
gives

~A'(r) = (
sin(p,x)e-ax2 00) - 100 (e-
--
ax2 )
p, cos(p,x)dx
2 2a 0 2a

1
o

- -p,
00 e- ax 2 cos(p,x)dx
2a 0
JTp,
- -A(p,).
4a
Thus

A' (p,) = (- :a) A(p,) (separable),

so

Now,

A(O) = -
JT
2100 0
e- ax2 dx,

and the change of variables y = x,Ja gives

2 1 roo _ 2 2 J1i 1
A(O) = ~,Ja 10 e Y dy = JT,Ja2 = Fa'
whence
524 8 FOURIER ANALYSIS AND STURM-LIOUVILLE THEORY

Finally, the convergence theorem gives

('" _1_ e- JL 2/(4a) cos(f.,Lx)df.,L = e-ax 2, -00 <x < 00.

10 Fa
Proof of the Convergence Theorem for Fourier Integrals (Optional). The
Fourier integral representation of a function f(x) may be written as the limit relation (see
equation (6»

lim -
M-+oo 11:
11 l°O0
M

-00
fm cOS(f.,LX - f.,L~)d~df.,L = f(x).

Moreover,

.!. 10(M 1 f(~) cOS(f.,LX - f.,L~)~df.,L

i:
00

11: -00

= ~ fm (1 M
cOS(jLX - f.,L~)df.,L) ~
= .!.1°O f(~) sin(Mx - M~) ~

11
11: -00 X - ~

=-
00
)'x+t
"( )sin(Mt)dt.
11: -00 t
Thus to prove the convergence theorem it will suffice to prove that

(I8) lim -
1
M-+oo 11:
1 -00
00
f(x + t) sin(Mt)
t
dt = f(x) for -oo<x<oo

whenever f satisfies (1)-(3). Moreover, this relation is equivalent to the Dirichlet limit
theorem of Section 8.8, differing only in the letters used. Hence both theorems will be
proved when (18) is verified.
The proof of (18) given below is based on the Riemann-Lebesgue theorem. This
extension of the Riemann theorem of Section 8.6 can be formulated as follows.

i:
Riemann-Lebesgue Theorem. Let g(x) satisfy

(19) Ig(x)ldx < 00.

Then

(20) lim
M->oo
1 00

-00
g(x) sin(Mx)dx = M-+oo
lim 1 00

-00
g(x) cos(Mx)dx = o.
 8.9 THE FOURIER INTEGRAL 525

A proof of this theorem is outlined below, at the end of this section. First it will be
used to prove the convergence theorem.
To prove (18) it will be convenient to write

(21)
1
-
/00 f(x + t) sin(Mt) dt - 1
- [f(x + 0) + J(x - 0) I
n -00 t 2
= II (M) + heM) + I3 (M) + I4 (M),

where

(22)
1
I 1 (M) = -
[00 f(x + t) sin(Mt) dt,

(23) I 2 (M) = -
n
nIt
1 /-1-00 f(x + t) sin(Mt) dt,
11
t
1 sin(Mt) 1
(24) I3 (M) = - f(x + t) dt - -f(x + 0),

(25)
not

I4 (M) = -
1
n
/0
-1
f(x
2

+ t) sin(Mt) dt -
t 2
1
-f(x - 0).

Then to prove (18) it will be sufficient to show that each of the functions II, ... , 14 has
limit zero as M --+ 00. The Riemann-Lebesgue theorem is directly applicable to II and

i:
12 . Indeed,

11 (M) = get) sin(Mt)dt,

where

get) = g(X + t)/(nt) for t::: 1,

for t < 1,

and

/ -0000 Ig(t) Idt = ~1 /00 dt 1 /00

-00 If(x + t) ItS ~ -00 If(x + t) Idt < 00
by hypotheSiS (3). The integral heM) may be treated similarly
To show that I3 (M) --+ 0, note that if Mt = s, then

(26) -
1 11 sin(Mt)
dt = -
1 1M sines)
--ds = -Sl(M) --+-
1. 1
no t no s n 2
526 8 FOURIER ANALYSIS AND STURM-LIOUVILLE THEORY

as M --+ 00. Moreover, we may write

(27) MM) = -
111
If(x + t) - f(x + 0)1
sin(Mt)
dt

+ (-111
n o t
sin(Mt)dt - -I)Ji(x+ 0),
not 2
and the last term has limit zero as M --+ 00 by (26). It remains to show that

(28) ·
I1m
M~oo
11
0
f(x + t) - f(x + 0) sm
t
. (M)d 0
t t= .

Hypothesis (1) (f and f' sectionally continuous) implies that the right -hand derivative

f' (x) = lim f(x + t) - f(x + 0)

+ t~O+ t
exists for every x. It follows that if

If(x + t) - f(x + O)]lt for 0 < t:s 1,

get) = f~(x)
{ for t = 0,
o otherwise,

then get) is sectionally continuous and Ig(t) I is integrable over -00 < t < 00. Hence (28)
follows from the Riemann-Lebesgue theorem. This proves that I3(N) --+ 0 as M --+ 00.
A similar proof works for I4 (M). This completes the proof of the convergence theorem
for Fourier integrals.
Proof of the Riemann-Lebesgue Theorem. Only the first equality of (20) will be
proved. To do this it must be shown that given any E > 0 there is a number Mo = MO(E)
such that

(29) Ii: I
g(x) sin (Mx) dx < E for M > Mo.

For this purpose it will be convenient to introduce a parameter N > 0 and an auxiliary

i:
function ¢(x), - N :s x :s N. With this notation we may write
g(x) sin(Mx)dx = MM,N) + h<M,N) + I 3 (M,N) + I 4 (M,N),
where

II (M, N) = LOO
g(x) sin(Mx)dx,
 B.9 THE FOURIER INTEGRAL 527

i:
i:
h(M,N) = g(x)sin(Mx)dx,

i:
13 (M, N) = (g(x) - ¢(x)) sin(Mx)dx,

14 (M, N) = ¢(x) sin(Mx)dx.

It will be shown, for suitable choices of N and the function ¢(x) , that each of these four
integrals is less than E/4 for all M greater than some Mo = Mo(E).
Consider first the integrals II and h. The integrability condition (19) implies that
there is an N = N(E) such that

1N
00 Ig(x)ldx <:. and
4
j-N
-00
Ig(x)ldx < :..
4
It follows, since I sin (Mx) I ::s 1, that
E E
(30) IItCM,N)1 < 4' Ih(M,N)1 < 4
for all M. Next, a fundamental theorem of integration theory states that if Ig(x) I is
integrable on -N ::s x ::s N then there is a function ¢(x) with a continuous derivative
on - N ::s x ::s N such that

(31) j
-N
N
Ig(x) - ¢(x)ldx < -
E

4
and hence

(32) 113 (M,N)1 ::s jN

-N
Ig(x) - ¢(x)ldx < :.
4
for all M. Finally, integration by parts gives

14 ( M,N) = [¢(-N) - ¢(N)] cos(MN)

+ -1 jN ( )¢ '(cos Mx x
)dx
.
M M -N

This implies that if N = NO(E) is fixed as above then there is an Mo = MO(E) such that
(33)

for all M ~ Mo(E). Combining (30), (32) and (33) gives

II:g(x)sin(Mx)dxl ::s IItCM,N)1 + 112 (M,N)1 + 113 (M,N)1 + II4 (M,N)1

<E
528 8 FOURIER ANALYSIS AND STURM-lioUVILLE THEORY

for all M ::: Mo(E), which completes the proof.

The only unproved step in the proof is (31). This result is proved in textbooks on
integration theory; see, for example H.R. Pitt, Measure and Integration for Use, Clarendon
Press, Oxford (1985), page 43.
Complex Form of the Fourier Integral. The Fourier integral theorem implies that
if f(x) satisfies conditions (1), (2), (3) then f(x) satisfies the integral identity (9). This
theorem takes a Simpler form if the cosine is replaced by a complex exponential. To derive
the result note that since the cosine and sine are even and odd, respectively, we have

(34) f(x) = ~
2n
1 (1 f(~)
-00
00 00

-00
cos fleX - ~) d~) dfl,

and

(35) o= ~
2n
1 (1 f(~)
00

-00 -00
00
sin fleX - ~) d~) dfl.
Multiplying (35) by i and adding the result to (34) gives

(36) j(x) = ~
2n
1 (1 f(~)ei/L(X-~) d~)
00

-00
00

-00
dfl,

or

(37)

This is the complex form of the Fourier integral representation of f(x). Note that if we
define J(fl) by

then

J
The function (x) is called the Fourier transform ofJ(x). Large tables of complex Fourier
transforms are available in [ER].

I~~~rcises 8.9J
Fourier Integral Theorem. Below, verify for the given f the three hypotheses of the Fourier integral
theorem.
 8.10 FOURIER SINE ANO COSINE INTEGRALS 529

1. j(x) = xe- Ixl on -00 <x< 00.

2. j(x) = x 2e- lxl on -00 <x< 00.

3. j(x) = x 2 for Ixl < l,f( -1) = 0.5,f(l) = 0.5,f(x) = 0 elsewhere.

4. j(x) = _x2 for Ixl < l,f( -1) = -0.5,f(1) = -0.5,f(x) = 0 elsewhere.
Calculation of A(/t) and B(/t). &Iow, determine by direct integral evaluation explicit formulas for the
Fourier integral coefficient functions

A(/t) = ~
7f
1 00

-00
j(x) cos(/tx)dx, B(/t) = ~
7f
1
00

-00
j(x) sin(/tx)dx.

5. j(x) = xe- 2lxl on -00 < x < 00.

6. j(x) = x2e-2lxl on -00 <x < 00.

7. j(x) = 2x for Ixl < l,j( -1) = -l,f(l) = l,f(x) = 0 elsewhere.

8. j(x) = 1 +x2 for Ixl < 1,f(-1) = l,f(l) = l,j(x) = 0 elsewhere.
Fourier Integral Representation. Below, verify, by direct integration of the integral, the representation

j(x) = 10 00
(A(/t) cos(/tx) + B(/t) sin(/tx)) d/t.
9. A(/t) = 4sin/tI/t, B(/t) = O,j(x) = 27f on Ixl < l,f(-l) = j(l) = 7f,f(x) = 0 elsewhere.

10. A(/t) = ~,B(/t) = 1-~~S/1- ,f(x) = 1 for 0 < x < 1,f(O) = j(l) = 1/2,f(x) = 0 elsewhere.
Fourier Integral Theorem. Apply the Fourier integral theorem in the form

j(x) = 10 00
(A(/t) cos(/tx) + B(/t) sin(/tx)) d/t

to obtain the follOwing new integration formulas, which are not easily obtained by any other method.
11. For x> 0,

10rOO (4COS(/tX) + 2/tsin(/tX)) d

----'--4-+-/t""2,------'--- /t = 27fe
-2x
,

while for x = 0 the integral evaluates to 7f, and for x < 0 it evaluates to O.
rOO (2COS(j.LX))
12 . For aII x, JO 1+/1-2 d/t = 7fe-Ixl .

13. For 0 < x < 7f roo (cosax+cosa(lI'-x)) da = sinx and for all other x it evaluates to O.
- - 'JO 1-a2 '

8.10 Fourier Sine and Cosine Integrals

This section treats linear heat flow in a long thin rod when attention is restricted to points
near one end but far from the other. A natural model is the idealized problem of heat
flow in a semi-infinite rod (Figure 9).
530 8 FOURIER ANALYSIS AND STURM-LIOUVILLE THEORY

~~)----------------~I_~ FIGURE 9. Heat diffusion in a semi-infinite rod,

x =o Dirichlet boundary conditions

°
Thus if x is a distance coordinate, measured from the near end, then the temperature
is a function u(x, t) defined for all x :::: and t :::: 0. If the near end is held at zero
temperature then u(x, t) is the solution of the following boundary value problem.
Heat in a Semi-Infinite Rod (Dirichlet Boundary Condition). Find u(x, t) such
that

(1) U(x, t) is defined for °: : x < 00, t :::: 0,

(2) = °: : x
for < 00, t :::: 0,
° t :::
Ut U xx

(3) U(O, t) = for 0,

(4) u(x,O) = f(x) °: : for x < 00,

where f(x) is a prescribed initial temperature distribution. In analogy with the case of
the infinite rod, we shall seek a solution under the follOwing additional assumptions.

(5) U(x, t) is bounded for all x :::: 0, t :::: 0,

(6) fooo If(x)ldx < 00.

This problem can be solved by separation of variables. The steps parallel those for the
infinite rod in Section 8.7. The result is

(7) u(x, t) = 1 00
B(jL) sin(jLx)e-/12tdjL ,

where

(8) f(x) = 1 00
B(jL) sin(jLx)djL.

Details of the derivation are left for the Exercises. To complete the solution, (8) must be
solved for B(jL). This can be done by an orthogonality calculation, following the plan of
Section 8.8. However, a more direct approach may be based on the full Fourier integral
of Section 8.9. The key idea is to consider the odd extension ofj(x) , defined by

°
f(x) x> 0,
{
(9) rex) = x = 0,
-fe-x) x < 0.
 8.10 FOURIER SINE AND COSINE INTEGRALS 531

ThenJ*(-x) = -J*(x) and hence the Fourier integral ofJ* is given by

(10) A*(fL) = -;:;11 00

_ooJ*(x) cos(fLx)dx = 0,
B*(fL) =- 1lOQ rex) sin(fLX)dx = -
Jr -DO
21
Jr 0
00
rex) sin(fLX)dx

and

(ll) rex) = -
Jr
21 0
DO
B*(x) sin(J.Lx)dJ.L

for all x. In particular, we have J*(x) = J(x) for x ::::: 0, and (ll) becomes (8) with

(12) B(fL) = - 21
Jr 0
DO
J(x) sin(fLx)dx.

This procedure leads to the following theorem.

Fourier Sine Integral Theorem. Let J(x) satisfy the conditions

(13) J and l' are sectionally continuous for 0 :::: x < 00,
(14) J(O + 0) = O,f(x) = ~[J(x + 0) + J(x - 0)] for 0 < x < 00,

(15) IoDO If(x)ldx < 00.

Then

(16) J(x) = 10= B(fL) sin(fLx)dfL for x::::: 0,

where B(fL) is defined by (12).

This theorem is a direct corollary of the Fourier integral theorem of Section 8.9. We
need only verify that if J satisfies (13)-(15) then J*, the odd extension of J, satisfies (1)-
(3) in Section 8.9. The details are left as an exercise. An alternative formulation is given
by the following corollary
Corollary. Under the hypotheses of the Fourier sine integral theorem we have

(17) J(x) = - 21
Jr 0
DO
sin (fLx) 1°C
0
J(~) sin(fL~)d~dfL for x::::: o.

Returning to the problem of heat diffusion in a semi-infinite rod, if the boundary

condition of zero temperature at the near end (Dirichlet boundary condition) is changed
to the boundary condition of zero heat flux (Neumann boundary condition) then the
temperature u(x, t) is the solution of the following boundary value problem.
532 8 FOURIER ANALYSIS AND STURM-LIOUVILLE THEORY

Heat Diffusion in a Semi-Infinite Rod (Neumann Boundary Condition).

Find u(x, t) such that

(18) U(x, t) is defined for 0 :::: x < 00, t 2: 0,

(19) Ut = Uxx for 0 :::: x < 00, t 2: 0,
(20) uAO, t) = 0 for t 2: 0,
(2l) u(x,O) = f(x) for 0 :::: x < 00,

where f(x) is a prescribed initial temperature distribution. Conditions (5), (6) are again
assumed. In this case separation of variables gives the solution

(22)

where

(23)

The last equation can be solved for A(fL) by applying the Fourier integral theorem to the
even extension off. This leads to the following theorem.
Fourier Cosine Integral Theorem. Let f(x) satisfy the conditions

(24) f andf' are sectionally continuous for 0:::: x < 00,

(25) fCO) = f(O + 0) ,fcx) = ! [fCx + 0) + fCx - 0) 1 for 0 :::: x < 00,

(26) 10 00
If(x)ldx < 00.

Then

(27) f(x) = 1000

A(fL) COS(fLX)dfL for x 2: 0,

where A(fL) is defined by

(28) A(fL) = ~ 10 00
f(x) cosCfLX)dx.

Details of the proof are left for the Exercises. Substitution of (28) into (27) leads to
the following corollary
Corollary. Under the hypotheses of the Fourier cosine integral theorem we have

(29) f(x) = - 21
:rr 0
00
COS(fLX) 1
0
00
f(~) COS(fL~)d~dfL for x 2: O.
 8.11 HEAT DIFFUSION IN NONHOMOGENEOUS RODS 533

IExercises 8.10 I
l. (Heat Diffusion) Derive the solution (7)-(8) of the semi-infinite rod problem with Dirichlet boundary
conditions by the method of separation of variables.

2. (Heat Diffusion) Derive the solution (22)-(23) of the semi-infinite rod problem with Neumann boundary
conditions by the method of separation of variables.

3. (Fourier Sine Integral) Derive the representation (16), where B(IL) is defined by (12), using the Fourier
sine integral theorem.

4. (Fourier Cosine Integral) Derive the representation (27)-(28) using the Fourier cosine integral theorem.

Part III Sturm-Liouville Theory

Sturm-Liouville theory is a generalization of Fourier theory that greatly extends the scope
of the method of separation of variables. Part III develops this theory in the context of
the boundary value problem of heat diffusion in a nonhomogeneous rod. Applications
to other boundary value problems of mathematical physics are developed in Chapter 9.
Part III consists of Sections 8.11 through 8.14. In Section 8.11 the boundary value
problem is formulated, and its solution by separation of variables is begun. Presented
in Section 8.12 is the basic theory of Sturm-Liouville eigenvalue problems, including
the orthogonality property and a convergence theorem for eigenfunction expansions.
Developed in Section 8.13 is a method for calculating the eigenvalues and eigenfunctions
and it is applied to some representative examples. Finally, Section 8.14 discusses singular
Sturm-Liouville problems in which the differential equation has a Singular point at an
endpoint of the interval. It will be shown in Chapter 9 that such Singular problems occur
naturally in applications.

8.11 Heat Diffusion in Nonhomogeneous Rods

A nonhomogeneous rod that occupies the interval a .:::: x .:::: b will be considered. The
temperature of the rod will be a function u(x, t) with the following properties.

(1) u(x, t) is defined for a .:::: x .:::: b, t ::: 0,

(2) w(x)Ut = (p(x)ux)x + q(x)u for a .:::: x .:::: b, t ::: 0,

(3) ! aIu(a, t)+ a2uAa, t) = 0,

+ b2uAb, t) = 0, t::: 0,
bi u(b, t)
(4) u(x,O) = j(x) for a :s x :s b,
534 8 FOURIER ANALYSIS AND STURM-LIOUVILLE THEORY

where j(x) is a prescribed initial temperature distribution. The partial differential equation
(2) is equation (24) in Section 7.2 with coefficients p, q, w that are related to the physical
parameters p, c, A, K, H by

(5) w=pcA, p=AK, q=-H.

In particular, w > 0, p > 0 and q :::: o.
The two boundary conditions of (3) are specified by the number pairs (aI, a2)
and (b l ,b2 ). It is assumed that neither pair is (0,0). Note that either condition can
be multiplied by a nonzero constant to give an equivalent boundary condition. The
boundary conditions (3) include all the examples of Section 7.2. Thus the Dirichlet
condition at x = a is (aI, a2) = (1,0); the Neumann condition at x = b is (b l , b2) =
(0, 1), etc. Newtons law of cooling is (aI, a2) = (k, -K) at x = a and (b l , b2) = (k, K)
at x = b, where k and K are positive.
Separation of Variables. The first step in applying the method of separation of
variables to the diffusion problem (1)-(4) is to find the product solutions of (1)-(3). By
the now familiar argument, u(x, t) = X(x)nt) is a solution if nt) satisfies

(6) T'(t) + Ant) = 0,

where A is a separation constant and X(x) is a solution of the follOwing eigenvalue

problem:

(7)
(p(x)X')' + [q(x) + AW(x)]X = 0 for a:::: x:::: b,
aIX(a) + a2X'(a) = 0, and bIX(b) + b2X'(b) = O.
This problem is called a Sturm-Liouville problem. The principal facts concerning such
problems are stated in the next section and used to solve the heat diffusion problem
(1)-(4).

B.12 Sturm-Liouville Problems: Basic Theory

The main facts of the theory are suggested by the theory of Fourier series (Sections
8.1 through 8.3). They include the existence of an infinite number of eigenvalues and
corresponding eigenfunctions, an orthogonality property of the eigenfunctions, and
the completeness of the eigenfunctions, as expressed by a convergence theorem for
the eigenfunction expansions. Such results have been extended to the Sturm-Liouville
problem (7) of Section 8.11 with a variety of classes of coefficients p, q, wand boundary
conditions. The class defined in the follOwing theorem is applicable to many boundary
value problems of mathematical physics.
 8.12 STURM-liouVILLE PROBLEMS: BASIC THEORY 535

Existence Theorem for Eigenvalues and Eigenfunctions. Let the coefficients

p(x), q(x) , w(x) and the constants aI, a2, bl , b2 satisfy the conditions

(1) p(x), p'(x) , q(x) and w(x) are continuous for a::: x ::: b,
(2) p(x) > 0 and w(x) > 0 for a ::: x ::: b,
(3) (aI, a2) f. (0,0) and (b l , b2 ) f. (0,0).

Then the Sturm-Liouville problem

(p(x)X')'+ [q(x) + AW(x)]X = 0 for a::: x ::: b,

alX(a) + a2X'(a) = 0, and blX(b) + b2X'(b) = 0,

has an infinite sequence of eigenvalues An and corresponding eigenfunctions Xn(x),

n = 1,2,3, .... The eigenvalues are real and can be enumerated in increasing order:

(4)

and

(5) lim An =
n-+oo
+00.
Moreover, the eigenspace for any An is one-dimensional. This means that any two
eigenfunctions of the eigenvalue An differ at most by a constant factor.
Proofs of the existence theorem depend on methods of advanced calculus; see, for
example, [B-R]. Methods for calculating An and Xn(x) are given in Section S.13, together
with examples.
The success of the method of separation of variables depends on the orthogonality
of the eigenfunctions Xn(x). For the Sturm-Liouville problem of the above existence
theorem, the correct condition is described by the following theorem.
Orthogonality Theorem. Under the conditions of the existence theorem, the
eigenfunctions satisfy the equation

(6) lab Xm (x)Xn (X) W (X) dx = 0 for m f. n.

Condition (6) will be expressed by saying that Xm and Xn are orthogonal with weight
w(x) on a ::: x ::: b. When w(x) = 1, it reduces to the usual orthogonality of Fourier
theory. The follOwing proof of (6) is based directly on the Sturm-Liouville differential
equation and boundary conditions.
536 8 FOURIER ANALYSIS AND STURM-lioUVILLE THEORY

Proof of the Orthogonality Theorem. The eigenpairs (A'm, Xm) and (An, Xn) satisfy

(pX~)' + [q + AmwlXm = 0,
(pX~)' + [q + An W lXn = O.

If the first equation is multiplied by Xn , the second by Xm , and the second is subtracted
from the first, then the terms with q cancel. The result is

(7) (pX~)'Xn - (pX~)'Xm = (pX~Xn - pX~Xm)'

= (An - Am)XmXnw.
If the last equation is integrated over a ~ x ~ b then the result can be written

(8)

where

It = p(a)[X~ (a)Xn(a) - X~ (a)Xm(a) 1 and

12 = p(b)[X~(b)Xn(b) - X~(b)Xm(b)l.

Now, the boundary condition at x = a implies that

(9)

if al i- O. These equations imply that II = O. If al = 0 then a2 i- 0 and

which again implies that It = O. Similarly, the boundary condition at x = b implies that
h = O. Thus
(10) (An - Am) lb Xm (x)Xn (x) w(x) dx = O.
If n i- m then An i- Am and (10) implies (6). This completes the proof.
Sturm-Liouville Series. The final ingredient in the method of separation of variables
is a convergence theorem that states that rather general functions f(x) have expansions
into an eigenfunction series:
00

(11) f(x) = I:>nXn(X) for a ~ x ~ b.

n=1
 8.12 STURM-liouVILLE PROBLEMS: BASIC THEORY 537

If (11) is multiplied by Xm(x)w(x) and then integrated over a'::: x .::: b then orthogonality
condition (6) implies that

(12) =
f: f(x)X n (x)w(x)dx
X; (x)w(x)dx.
Cn b
fa
The series in (ll) with coefficients (2) is called the Sturm-Liouville series of/ex).
When is f(x) equal to its Sturm-Liouville series? Various sufficient conditions are known.
One classical theorem states that if

(13) f and1' are continuous on a'::: x.::: b, and

ad(a) + a21'(a) = 0, bd(b) + bd'(b) = 0,
then (1) holds with Cn given by (12); see page 197 in [C-L]. More generally, iff(x) and
l' (x) are only sectionally continuous then (11) holds if at points of discontinuity ofj(x)
one has
1
f(x) = 2" [f(x + 0) +f(x - 0) 1 for a< x < b

and

fCa) = f(a + 0), feb) = feb - 0).

More General Boundary Conditions. The preceding theorems do not include the
case of periodic boundary conditions
(14) X(a) = X(b), X'(a) = X'(b).
It can be shown that most of the preceding theory holds in this case. An exception is
in the statement that the eigenvalues are simple (have a one-dimensional eigenspace).
In fact, for the full Fourier series it was shown that the eigenvalues An = n 2 > 0 had
two-dimensional eigenspaces with orthogonal eigenfunctions sin(nx) and cos(nx).
The separated boundary conditions of (7) and the periodic boundary conditions of
(14) are examples of more general linear homogeneous boundary conditions of the form

(IS) MllX(a) + M 12P(a)X'(a) + NllX(b) + N 12P(b)X'(b) = 0,

M 21 X(a) + MnP(a)X'(a) + N 21 X(b) + NnP(b)X'(b) = O.

It is known that the orthogonality theorem holds for the Sturm-Liouville problem with
boundary conditions 05) if (and only iD the 2 x 2 coefficient matrices M and N satisfy

rank (M I N) = 2 and detM = detN.

For details see [C-L].

538 8 FOURIER ANALYSIS AND STURM-LIOUVILLE THEORY

Applications to Heat Diffusion. Consider the problem of heat diffusion in a

nonhomogeneous rod, as studied in Section 8.11:

(16) U(x, t) is defined for a.:::: x .:::: b, t ~ 0,

(17) w(x)Ut = (p(x)Uxt + q(x)u for a .:::: x .:::: b, t ~ 0,

a1u(a, t) + a2ux(a, t) = 0,
(18) (
b1u(b, t) + b2 ux (b, t) = 0, t ~ 0,
(19) u(x,O) = f(x) for a .:::: x .:::: b,

where f(x) is a prescribed initial temperature distribution. The solution of the boundary
value problem (16)-(19) by the method of separation of variables can now be completed.
The product solutions of (16)-(18) are the functions

Un(X,t)=Xn(x)e-Ant, n=1,2,3, ... ,

where An and Xn are the eigenvalues and eigenfunctions of the problem (7), Section 8.11.
The superposition step is to build the complete solution as a sum

L cnXn(x)e-Ant.
00

(20) u(x, t) =
n=l

To satisfy the initial condition (19) we need to represent f(x) as a Sturm-Liouville series
ell). This is possible if f(x) satisfies the conditions of a convergence theorem such as
(13). Then the solution is given by (20) with Cn defined by (12).

[~~ercises ~:12~
Weight Functions. Determine for the given problem the weight function w(x) and the interval of definition
of w(x).

1. (xy')' + x(x + h = 0, yeO) = y(2) = o.

3 )y

2. xy" + (x + Asinx)y = 0, y(n/2) = yen) = o.

3. y" + y' + (x + A + h)y = 0, yeo) = y(l) = o.

4. y" - y' + (x + A + h)y = 0, yeo) = y(l) = O.

Eigenvalues and Eigenfunctions. Verify that the given sequence of eigenpairs (An,Xn), n = 1,2,3, ... ,
solves the stated Sturm-Liouville problem. Verify that all solutions are represented.

S. An = n2 , Xn = sin nx, X" + AX = 0, X(O) = X(n) = o.

6. An = n2 , Xn = cos nx, X" + AX = 0, X'(O) = X'(rr) = o.
 8.13 CONSTRUCTION OF EIGENVALUES AND EIGENFUNCTIONS 539

7. An = n 2rr2, Xn = sin(nrr lnx), x2X" + xX' + AX = 0, XO) = X(e) = O.

8. An = n 2rr2, Xn = cos(nrr lnx), x2X" + xX' + AX = 0, X'(l) = X'(e) = O.

Orthogonality. Verify by direct integration that the given sequence of functions {Xn} is orthogonal on
la, b] with weight function w(x)
9. Xn(x) = sinnx, a = 0, b = rr, w(x) = 1, n = 1,2,3, ... .

10. Xn(x) = cosnx, a= 0, b = rr, w(x) = 1, n = 1,2,3, ... .

11. The Legendre polynomials Xj (x) = x, X2(X) = (3x 2 - 1)/2, X3(X) = (5x 3 - 3x)/2, a = -1, b = 1,
w(x) = 1.

12. The Legendre polynomialsX2(x) = (3x 2 -1)12, X3(X) = (5x 3 -3x)!2, X4(X) = (35x 4 - 30x 2 +3)/8,
a = -1, b = 1, w(x) = 1.

13. (Separation of Variables) Let u(x, t) = X(x)T(t) be a product solution for the heat diffusion problem
in a nonhomogeneous rod. Supply the details in the method of separation of variables that show that
Tet) and X(x) satisfy, for some constant A,

T'et)+ H(t) = 0, T(O) =I- 0,

+ Iq(x) + Aw(X)]X = 0 for a:S x :s b,
(p(x)X')'
ajX(a) + a2X' (a) = 0, bjX(b) + b2X'(b) = o.

14. (Orthogonality Theorem) Let (X n , An) and (Xm , Am) be two eigenpairs, as in the orthogonality
theorem. Supply the details in the derivation of the relation

8.13 Construction of Eigenvalues and Eigenfunctions

The existence theorem of Section 8.12 does not show how the eigenvalues and eigenfunc-
tions can be found. Here a method is presented that is applicable whenever a solution
basis of the differential equation

(1) (p(x)X'(x))' + [q(x) + )..w(x)]X(x) = °

is known. Let Vj (x,)..), V2 (x,)..) be such a solution basis. Then the general solution of
(1) can be written as

(2)

where Cj and C2 are arbitrary constants. The boundary conditions

0) ajX(a,)..) + a2P(a)X' (a,)..) = 0,

bjX(b,)..) + b2P(b)X'(b,)..) = 0,
540 8 FOURIER ANALYSIS AND STURM-lioUVILLE THEORY

are assumed with (al,a2) i= (0,0) and (b l ,b2) i= (0,0). If (2) is substituted into (3)
then the result is a homogeneous linear system for (Cl, C2),

(4) All (A)Cl + Al2 (A)C2 = 0,

An (A)Cl + All (A)c2 = 0,
where

(5) All (A) = al VI (a, A) + a2 V~ (a, A),

Al2 (A) = al V2(a, A) + a2 V~(a, A),
An (A) = bl VI (b, A) + b2 V~ (b, A),
Al2 (A) = bl V2 (b, A) + b2 V~(b, A).
A nontrivial solution of (4) is needed. This exists if (and only iO the determinant

(6) D(A) = All (A) Al2 (A)

An (A) All (A)

is zero. Thus D(A) = 0 is an eigenvalue equation. The eigenvalues AI, A2, A3, ... , which
are guaranteed by the theory, are precisely the roots of this equation. In simple cases the
roots can be found exactly If this is not possible, then numerical methods must be used.
Examples of both types are given below. The corresponding eigenfunctions are given by

(7)

where (Cl, C2) is any nontrivial solution of (4). The solution (Cl, C2) is unique up to scalar
multiples. Indeed, the existence theorem implies that for each eigenvalue An one has

(8) rank ( All (A)

All (A)
AuC A
All (A)
») = 1,
and hence the solution space of (4) is one-dimensional. To see this, note that if the rank
were 2, then (4) would have only the trivial solution when A = An, and An would not be
an eigenvalue, while if the rank were 0, then the eigenspace would have dimension 2.
TheeigenfunctionsXn(x) can be found by setting A = An in the system (4) and solving
one of the equations for the ratio Cl le 2 or C2lel. Condition (8) guarantees that this is always
possible. The remainder of this section presents several examples of this method. The first
two are the Dirichlet and Neumann problems of Section 8.1 and Section 8.4, respectively
They will be reviewed briefly as simple examples of the method of this section.
 8.13 CONSTRUCTION OF EIGENVALUES AND EIGENFUNCTIONS 541

EXAMPLE 1. Consider the Dirichlet problem

(9)
X" + AX = 0 for 0::: x::: rr,
X(O) =0 and X(rr) = o.

This is the special case of the general Sturm-Liouville problem (7) with P = w = 1, q = 0,
(a, b) = (0, rr), (aI, a2) = (1,0) and (b l , b2) = (1,0). As in Section 8.1, theform ofthe
solution basis VI (x, A), V 2 (x, A) changes with the sign of A. For A > 0 a suitable basis is

(10)

Substituting this into (5) gives

(ll) All(A) = 0, A12 (A) = 1,

A21 (A) = sin J):, A22 (A) = cos J):,

whence

(12) D(A) 0 1 = _ sin 'A.

= sin JI cos JI "1/ A

This yields the eigenpairs An = n 2 , Xn(x) = sin(nx) as before. The cases A = 0 and
A < 0 yield no further eigenvalues.

EXAMPLE 2. Consider the Neumann problem

X" + AX = 0 for 0 S x S rr,

(13)
X'(O) = 0 and X'(rr) = o.

This is the same as the preceding example except that the boundary conditions are
changed to (aI, a2) = (0,1) and (b l , b2 ) = (0,1). For A > 0 the solution basis (10)
may again be used and

(14) V~(X,A) = JIcos(JIx), V~(X,A) = -JIsin(J):x).

Substitution into (5) gives the matrix elements

(15) All(A) = J):, A12 (A) = 0,

A21 (A) = J): cos J):, A22 (A) = -J): sin J):,

whence

(16) DA
() = I JIcosJI
JI 0
-JIsinJI
= - 'Sl·U "I/A
I\.
".
542 8 FOURIER ANALYSIS AND STURM-lioUVILLE THEORY

This readily gives the eigenpairs ()..n,Xn(x» = (n 2 ,cos(nx», for n = 1,2,3, .... The
case A = 0 gives the additional eigenpair (AO,XO(X» = (0,1). The case A < 0 gives no
eigenvalues, as was shown in Section 8.4.

EXAMPLE 3. Consider the Sturm-Liouville problem

(17)
X" + AX = 0 for 0::::: x ::::: 1,
X(O) = 0 and X(1) + X'(l) = o.
Physically this corresponds to heat diffusion in a uniform rod with left end held at
temperature zero and Newton's law of cooling at the right end. The boundary conditions
are (aI,a2) = (1,0) and (b I ,b2) = (1,1).

To find convenient solution bases VI(X,A), V2 (X,A), the three cases A < 0, A = 0
and A > 0 are again treated separately:

Case 1 (A < 0). Write A = _J,t2 with J,t > O. Then a suitable basis is

VI (X, A) = efl-X, V2(X,A) = e-fl-X,

V~(X,A) = J,tefl-X, V~(X,A) = -J,te-fl-x,

whence

All (A) = 1, AdA) = 1,

A21 (A) = cfl- + J,tcfl-, A 22 (A) = c-fl- - J,te-fl-.

A simple calculation gives

1 1
D(A) =
c fl- + J,te /L e-fl- - J,tc-fl-,

where A = _J,t2 Hence D(A) < 0 for all A < 0 and so there are no negative eigenvalues.
Case 2 (A = 0). In this case the differential equation is X" = 0 and a basis is

= 1,
VI (x, 0) V2 (x,0)= x,
V~(x,O) = 0, V~(x,O) = 1,

and so

AIl(O) = 1, A 12 (O) = 0,
A2l (0) = 1, A 22 (O) = 2,
 8.13 CONSTRUCTION OF EIGENVALUES AND EIGENFUNCTIONS 543

and

1 0
D(A) = = 2 =F O.
1 2

Hence A = 0 is not an eigenvalue of problem (17).

Case 3 (A > 0). Write A = f.L2 with f.L > O. Then we may take

VI (x, A) = sin (f.Lx) , V 2(x, A) = cos(f.Lx),

V;(X,A) = f.Lcos(f.Lx), V~(X,A) = -f.Lsin(f.Lx),
which gives

All(A) = 0, A 12 (A) = 1,

A2l (A) = sin f.L + f.L cos f.L, An (A) = cos f.L - f.L sin f.L,

and hence

D(A) =
o 1
= - (sin f.L + f.L cos f.L).
sin f.L + f.L cos f.L cos f.L - f.L sin f.L

Note that if cos f.L = 0 then D(A) = ±l. Hence, the eigenvalue equation is sin f.L +
f.L cos f.L = 0, or, equivalently,

(18) f.L + tan f.L = O.

Graphical Solution. The existence theorem for Sturm-Liouville eigenvalue prob-

lems implies that (18) has infinitely many positive roots. This may be verified by plotting
the curves y = tan f.L and y = - f.L in the (f.L, y)-plane (see Figure lO). The roots f.L1 =
J):;, f.L2 = J):;., ... are the projections onto the f.L-axis of the points of intersection of
the two curves.

u_ 3rr
r- 2 FIGURE 10. Crossings of y = tan /1 and y = -/1
544 8 FOURIER ANALYSIS AND STURM-lioUVILLE THEORY

It is clear from the figure that

3n ;,
~ < ;;:; < n, - < VA2 < 2n,
2 2
and

A ~ (n - ~)n for n ---+ 00.

Moreover, a given eigenvalue An can be computed to any desired degree of accuracy by

applying a standard numerical method to equation (18).
The eigenfunctions Xn (x) are given by (7), where (Cl, C2) #- (0,0) satisfies (4). In the
present case All = 0, A 12 = 1, and the first equation of (4) gives a particular nonzero
solution (Cl, C2) = (1,0). The second equation gives the same solution, by (8). Thus the
eigenfunctions are given by

(19) Xn(x) = sin(Ax), n = 1,2,3, ....

As a direct check, using the boundary conditions, note that Xn(O) = °
and

by (18). The orthogonality property of the eigenfunctions is

(20) 11 sin(Fmx) sin(Ax)dx = ° for m #- n.

This can be checked by direct integration and use of the eigenvalue equation (18).
However, this is unnecessary, because the orthogonality is guaranteed by the general
proof of Section 8.12.
The remaining three examples involve Sturm-Liouville equations with variable
coefficients p(x) , q(x) and w(x). In most cases the most difficult step in solving such
problems is to find a solution basis VI (x, A), V 2 (x, A).

EXAMPLE 4. The eigenvalue problem

(2l)
(elxX')' + (e lx + Aelx ) X = ° for 0::::: x::::: n,
X(O) = 0, X(n) = 0,
is a Sturm-Liouville problem with p(x) = q(x) = w(x) = elx , (a,b) = (O,n),
(al,a2) = (1,0) and (h,b 2) = (1,0). The differential equation is equivalent to the
constant-coefficient equation

(22) X" + 2X' + (1 + A)X = 0,

 8.13 CONSTRUCTION OF EIGENVALUES AND EIGENFUNCTIONS 545

which can be solved by Euler's substitution X = e"x. It is easy to verify that there are no
°
eigenvalues A :s (Cases 1 and 2). For Case 3, write A = f-L2 with f-L > 0. Then Euler:S
method gives the solution basis

Substituting into (5) gives the matrix elements

AJl (A)=1, A12 (A) =0,

A2l (A) = e-" cos(rrf-L), A22 (A) = e-" sin (rrf-L),

whence

D(A) =
elf cos(rrf-L)
1 °
e- rr sin (rrf-L) ,

where A = f-L 2. Thus the eigenvalues A = f-L 2 are given by

1 2 2 2 n = 1,2,3, ....
rrf-Ln = rrn or An = f-Ln = rr n ,

Moreover, for A = An the system (4) has the solution (Cl, (2) = (0, 1) and hence

Xn (x) = e- x sin (rrx), n = 1,2,3, ....

The orthogonality relation on °:s x :s rr with weight = is w(x) e2x

1" Xm (x)Xn (x)e2xdx = 1lf sin(mrrx) sin(nrrx)dx = ° for m i= n,

which is easy to check by direct integration.

EXAMPLE 5. The eigenvalue problem

(23)
(xX')' + ).,x-IX = for ° 1:s x :s e,
XO) = 0, X(e) = 0,
is a Sturm-Liouville problem with p(x) = x, q(x) = 0, w(x) = 1/x, (a,b) = (1,e),
(aI, a2) = (1,0) and (b l , b2) = (1,0). The problem is solvable because the differential
equation is equivalent (use (xX')' = xX" + X') to the Euler equation

(24)

The equation may be solved by the substitution X = x", which produces a solution if
0i 2 + A = 0. It is easy to check that there are no eigenvalues A :s 0. For A = f-L 2 (f-L > 0)
546 8 FOURIER ANALYSIS AND STURM-LIOUVILLE THEORY

the roots ex = ±ijL give the solution basis

The corresponding matrix elements are

All(A)=I, A 12 (A) =0,

A21 (A) = COS(jL) , A22 (A) = sin(jL),
and hence the eigenvalue equation for A = jL 2 is

D(A) = 1
cOS(jL)
°
sin(jL)
= sin(-v'A\

This gives the eigenvalues and eigenfunctions

Xn(x) = sin(nnlnx), n = 1,2,3, ....

The corresponding orthogonality relation is

Ie
1
Xm(x)Xn(x)x-1dx = Ie
1
dx
sin(mnlnx) sin(nnlnx)-
X
= 0, m#n.

The relation is easily verified by direct integration using the substitution t lnx,
dt = x-ldx.

EXAMPLE 6. The eigenvalue problem

(25)
(X2X')' + Ax- 2X = ° for I/2:s x :s 1,
X(l/2) = 0, X(l) = 0,

is a Sturm-Liouville problem with p(x) = x 2 , q(x) = 0, w(x) = I/x 2 , (a, b) = (112,1),

(al, a2) = (1,0) and (b 1, b2) = (l, 0). An equivalent differential equation is

(26)

This has variable coefficients and is not an Euler equation. Hence a solution basis is not
obvious. However, the problem may be Simplified by a change of independent variable.
To do this, define new independent and dependent variables y and Y(y) by
 8.13 CONSTRUCTION OF EIGENVALUES AND EIGENFUNCTIONS 547

Then careful application of the chain rule gives

X'(x) = Y'(y) ~ = _x- 2y'(y),

XI/(x) = x-4yl/(y) + 2x- 3 y'(y),

whence

Thus the transformed equation is Y" + AY = 0. Moreover, X(1I2) = Y(2) = 0, and

X(I) = Y(l) = 0. Thus the transformed function Y(y) is the solution of the simple
Sturm-Liouville problem

Y"+AY=O,
(27)
Y(l) = 0, Y(2) = 0.
The original problem will be solved by solving (27) and then transforming back.

It is easy to verify that (27) has no eigenvalues A ::::: 0. For A = ft2 (ft > 0) a solution
basis for (27) is

and the corresponding matrix elements are

All (A) = COS(/-L), AdA) = sin(/-L),

A21 (A) = cos(2/-L), A 22 (A) = sin(2ft).
The eigenvalue equation is

by the trigonometric identity cos a sin b - sin a cos b = sin (b - a). Thus the eigenvalues
are

Moreover, for A = An the coefficient matrix in system (4) is

cos(nn) Sin(nn)) ( (_l)n

(
cos(2nn) sin(2nn) - I
548 8 FOURIER ANALYSIS AND STURM-LIOUVILLE THEORY

Hence we may take ((1,(2) = (0,1) in system (4), which gives the eigenfunctions of
(27) as

YnCy) = sin(n1lY), n = 1,2,3, . ...

Returning to the original variables x and X(x) gives the eigenpairs

An = rr 2n2 ,
Xn(x) = sin (7), n = 1,2,3, . ...
The corresponding orthogonality condition is

1 m
1
X m (x)X n (x)x- 2dx = 11
m
(mrr) sin (nrr)
sin -
x
- dx = 0,
2"
x x
m:f. n.

This may be verified by the change of variable y = X-I followed by integration.

Exercises 8.13

Sturm-Liouville Problems. Solve the given Sturm-Liouville eigenvalue problem with Dirichlet conditions
for the eigenvalues and eigenfunctions and explicitly write out the orthogonality relations.
1. X" (x) + )'X(x) = 0, X( -rr) = 0, X(rr) = O.
2. (x-IX')' + x- 3(1 + )')X = 0, X(l) = 0, X(e) = O.

3. (xnX')' + h n- 2X = 0, X(a) = 0, X(b) = 0, 0 < a < b, n real. Do only the special case n = 1, a = 1,
b = 2.

4. (ix-Ix')' + 2)'xX = 0, X(l) = 0, X(-Ji) = 0.

Hint: Make the change of variable y = x 2 .

S. X" + 4AX = 0, X(O) = 0, X(rr) = 0.

6. X" + 16)"X = 0, X(O) = 0, X'(rr) = 0.

8.14 Singular Sturm-Liouville Problems

A Sturm-Liouville problem

(p(x)X')' + [q(x) + AW(x)]X = 0 for a::: x::: b,

aIX(a) + a2X'(a) = 0,
b1X(b) + b2X'(b) = 0
 8.14 SINGULAR STURM-LIOUVILLE PROBLEMS 549

is said to be singular if either a = -00 or b = +00 (or both) or if pea) = 0 or pCb) = 0

(or both). Problems on infinite and semi-infinite intervals were illustrated by the theory
of Fourier integrals in Part II above. For those problems, boundary conditions at infinity
were replaced by a boundedness condition, and the problems had continuous spectra.
Similar phenomena may occur if a and b are finite and pCa) = 0 and/or p(b) = O. Thefull
theory of singular Sturm-Liouville problems is too technical to be presented here. Instead,
three illustrative examples are discussed. In Chapter 9 each of these examples is used to
solve boundary value problems for partial differential equations of mathematical physics.
The Sturm-Liouville differential equation can be written

(1) p(x)X" + p'(x)X' + [q(x) + AW(X)]X = 0 for a:::: x:::: b.

If pea) = 0 then this equation has a singular point at x = a, in the sense of Chapter 3, and
solutions X(x) and their derivatives X' (x) may not have limits when x -+ a. In such cases
the boundary condition may be replaced by the weaker condition that XCx) and X'(x)
remain bounded as x -+ a. Recall the role of the boundary condition ajX(a) +a2X'(a) =
o in the proof of the orthogonality condition. It implied that (see (8) in Section 812)
pea) [X~(a)Xn(a) - X~(a)Xm(a)] = 0,

or, equivalently,

(2) limp(x) [X~(X)Xn(X) - X~CX)Xm(X)] = O.

x--+a

If p(x) -+ 0 and Xm(x), X~(x), Xn(x) and X~(x) remain bounded as x -+ a then (2)
again holds and the orthogonality theorem remains true.

EXAMPLE 1. Fourier-Bessel Series. Consider the eigenvalue problem

(3) (xX')' + AXX = 0,0 < x:::: 1,

(4) X(x) and X'(x) are bounded for 0 < x:::: 1,
(5) X(l) = o.
°
It is a Sturm-Liouville problem with p(x) = w(x) = x, q(x) = and Ca, b) = (0,1).
It is singular because pCO) = O. In Chapter 9 the problem arises in the analysis of the
vibrations of a circular drum head. Here it will be shown that the problem has an infinite
sequence of positive eigenvalues and corresponding orthogonal eigenfunctions.

Positivity of the Eigenvalues. It will be convenient to show first that any eigenvalue
must be positive. To do this let (A, X(x») be any eigenpair, then multiply (3) by XCx) and
550 8 FOURIER ANALYSIS AND STURM-lioUVILLE THEORY

integrate over 0 ::s x ::s 1. The result can be written

(6)

Moreover,

11 xX 2 (x)dx > 0

because X(x) is a nontrivial solution of (3). Hence, to prove that A > 0, it will suffice
to show that the last integral in (6) is negative. This can be done by integration by parts
and use of conditions (4) and (5). The result is

11 X(x) (xX'(x)), dx = xX(x)X'(x)l~ -1 1

x (X'(x))2 dx.

Now, the integrated terms give zero because (l)X(l)X'(l) = 0, by (5), while

lim xX(x)X' (x) = 0

X"'" 0

by (4). Combining these results gives

A11 x (X(x)/dx = 11x(X'(X))2dx> 0

and hence A> o.

General Solution of the Differential Equation. Equation (3) is equivalent to the
equation

xl/ +X~lX' + AX = O.
For A = 1 this is the Bessel equation of order zero. To find the general solution for any
A > 0, make the change of variable

X(x) = U(y), where y = J)..x.

Then the chain rule of calculus gives

X'(x) = J).. U'(y), XI/(x) = A UI/(y)

and the differential equation (3) is equivalent to the Bessel equation

UI/(y) + y~lU'(y) + U(y) = O.

 8.14 SINGULAR STURM-lioUVILLE PROBLEMS 551

It follows that

where Jo(y) is the Bessel function of order zero, defined in Chapter 3, and Yo(y) is a
second independent solution. Hence the general solution of (3) for A > 0 may be written

(7)

Calculation of the Eigenvalues and Eigenfunctions. It was shown in Chapter 3

that the solution (7) is bounded as x ~ 0 if and only if C2 = O. Thus (3) and (4) imply
that X(x) = cJo ("fix). Finally, (5) gives the eigenvalue equation

(8) XCl) = cJo("fi) = O.

For a nontrivial solution we must have Cl i= 0 and hence "fi must be a positive zero
ofJo(x). It was seen in Chapter 3 that there are infinitely many positive zeros ofJo(x),
listed in increasing order as

o< SI < S2 < S3 < ... < Sn < ... ,

and

Sn~(n-~)1T for n~oo.

Thus the eigenpairs for probJem (3)-(5) are

(9) An = s~, Xn(x) = JO(snx), n = 1,2,3, ....

Orthogonality of the Eigenfunctions. The weight function for the Sturm-Liouville

problem (3)-(5) is w(x) = x. Thus the orthogonality relation is

(10) 11 Xm(x)Xn(x)xdx = 11Jo(smx)Jo(snx)xdx = 0 for m i= n.

The simplest verification is to repeat the calculation of Section 8.12, equation (8). The
details are left for the Exercises.
Fourier-Bessel Series. If a function j has a representation

L
00

(11) j(x) = cnlO(snx ) for O:s x :s 1,

n=1
552 8 FOURIER ANALYSIS AND STURM-LIOUVILLE THEORY

then the orthogonality property (10) implies that

fo1 j(x)JO(snx)x dx
(12) en = 1
fo 15 (SnX)X dx
The series in (11) with coefficients (12) is called the Fourier-Bessel series of j(x). The
convergence of the Fourier-Bessel series to its defining function is known to hold under
the same conditions as for regular Sturm-Liouville series. Convergence theorems and
full proofs are given in [Wal, Chapter 18.

EXAMPLE 2. Legendre Series. Consider the eigenvalue problem for the Legendre
equation

(13) ((1- X2)XI)' + AX = :s x :s 1,

0, for -1
(14) X(x) and X' (x) are bounded for -1 :s x :s 1.

It is a Sturm-Liouville problem with p(x) = 1 - x 2 , q(x) = 0, w(x) = 1 and

(a, b) = (-1, 1). It is clearly a singular problem with Singular points at both x = -1
and x = 1, because p(±l) = O. In Section 3.4 the power series method was used to
construct a basis of (13) that was valid for every A. In particular, it was shown that
the boundedness condition (14) can be satisfied if and only if A = n(n + 1), where n
is a nonnegative integer. Moreover, the corresponding eigenspace is one-dimensional,
and the eigenfunction Xn(x) is a polynomial Pn(x) of exactly degree n. A method for
generating Legendre polynomials was given in the Exercises for Section 3.4.

Orthogonality. Since w(x) = 1, the orthogonality relation for Legendre polynomials

is

(15) i~ Pm(x)Pn(x)dx = 0 for m i- n.

The proof is the same as in Section 8.12, equation (8).

Legendre Series. The representation

L cnPn(x)
00

(16) j(x) =
n=O

with coefficients

(17) n = 0, 1,2, ... ,

 8.14 SINGULAR STURM-LIOUVILLE PROBLEMS 553

is called the Legendre series representation of j(x). It is known to be valid for -1 <
x < 1 when j(x) and l' (x) are sectionally continuous. The behavior at the singular points
x = ±1 is more complicated. See Ual for details and proofs.
EXAMPLE 3. A Sturm-Liouville Problem with Continuous Spectrum. Consider the
eigenvalue problem

(18) (xx')' + AX-IX = 0, 0< x::: 1,

(19) X(x) is bounded for 0 < x ::: 1,
(20) XO) = O.

It is a Sturm-Liouville problem with p(x) = x, q(x) = 0, w(x) = X-I and (a, b) = (0,1).
In particular, it is singular both because p(O) = 0 and also because w(x) has a singularity
at x = O. In Chapter 9 this problem arises in the analysis of steady-state heat diffusion
in a plate in the form of a circular sector. It is presented here as an example of a Singular
Sturm-Liouville problem that has continuous spectrum and is completely solvable.

Eigenvalues and Eigenfunctions. The differential equation (18) is equivalent to

the Euler equation

(21)

A method for finding the solutions of an Euler equation was given in Section 3.4. It will
be used here to find the eigenvalues and eigenfunctions of (18)-(20). The three cases
A < 0, A = 0 and A > 0 will be treated separately.
Case 1 (A < 0). Write A = _J..L2 with J..L > O. Then the method of Section 3.4 gives the
general solution

This is bounded for 0 < x ::: 1 if and only if (2 = 0 and X(x) = (lXI". The boundary
condition X(1) = 0 then gives (] = 0 and hence X(x) == O. Thus the problem has no
eigenvalues A < O.
Case 2 (A = 0). In this case the general solution is

X(x) = (1 + (2lnx.
This is bounded for 0 < x :::: 1 if and only if (2 = O. Then XO) = (1 = 0 and so
X(x) == O. Thus A = 0 is not an eigenvalue.
554 8 FOURIER ANALYSIS AND STURM-LIOUVILLE THEORY

Case 3 (A > 0). Write A = p.,2 with p., > O. Then the method of Section 3.4 gives the
general solution

X(x) = C1 cos(p., In x) + C2 sin(p., lnx).

This solution is bounded for 0 < x ::::: 1 when p., is any positive number. The boundary
condition X(l) = 0 gives C1 = O. Thus the problem has a continuous spectrum with

(22) Eigenvalues A = p., 2 ,

Eigenfunctions X(x) = sin(p., lnx), p., > O.

Eigenfunction Expansion. It is natural to ask whether a given function f(x) has an

eigenfunction representation

(23) f(x) = 1')0 C(p.,) sin(p., lnx)dp., for 0 < x ::::: 1,

where C(p.,) is an integrable function. An answer to this question can be obtained from
the Fourier sine integral theorem. To derive it, make the change of variables x = e- t (so
lnx = -t) and write

(24) pet) = f (e- t ) = -1 00

C(p.,) sin(p.,t)djL.

This is the Fourier sine integral representation of pet) with solution

-C(p.,) = B(p.,) = -
rr 0
21 00
pet) sin(p.,t)dt.

Returning to the original integration variable x = e- t gives the alternative formula

(25) C(p.,) = -
rr 0
211
f(x) sin(p., Inx)x- 1dx.

These calculations are valid when pet) = f(e- t ) satisfies hypotheses (13)-(15) of the
Fourier sine integral theorem. This implies the following representation theorem.

Theorem. Let f(x) satisfy the conditions

(26) f(x) and f' (x) are sectionally continuous for 0 < x ::::: 1,
(27) f(1- 0) = 0, f(x) = ![j(x + 0) + f(x - 0)] for 0 < x < 1,

(28) 11 [f(x)lx- 1 dx < 00.

Then f(x) has the representation (23), where C(p.,) is defined by (25).
 8.14 SINGULAR STURM-lioUVILLE PROBLEMS 555

Proof. We need only verify that the conditions (26)-(28) for j(x) are equivalent to
conditions (13)-(15) for F(t) = fCe-t). The verification is left for the Exercises.

I Exercises 8.14 1
~"~"""'",",,,,,,~, ~'"'''''m, _" ,',,,,,b,,,,,il+,~,,',,"

1. (A Legendre Series on 0 ~ x ~ 1) Construct the eigenpairs (j"n, Xn) for the singular Sturm-Liouville
problem

((1 - x 2 )X')' + AX = 0 on O:s x :s 1,

X(O) = 0, X(x) and X'(x) bounded for 0 :=; x:=; 1.

2. (Another Legendre Series on 0 ~ X ~ 1) Construct the eigenpairs (.A.n, Xn) for the Singular
Sturm-Liouville problem

CO-x 2)X')'+AX=0 on O:=;x:=;l,

X'(O) = 0, X(x) and X'(x) bounded for 0 :=; x :=; 1.

3. (Orthogonality of Fourier-Bessel Series) Supply the details in the proof of the orthogonality relation

10 1 Xm (x)Xn(x)x dx = 10 1]0 (Smx)]O (Snx)x dx = 0 for m f= n.

CHAPTER 9
 Boundary Value Problems
of Mathematical Physics

The goal of this final chapter is to show how the boundary value problems of mathematical
physics can be solved by the methods of the preceding chapters. This will be done by
solving a variety of specific problems that illustrate the principal types of problems that
were formulated in Chapter 7. Additional applications are developed in the Exercises.
The primary solution method is Fourier's method of separation of variables and the
associated Sturm-Liouville theory of Chapter 8.
Chapter 9 is divided into seven sections, as indicated above. Section 9.1 presents
a final discussion of the one-dimensional heat diffusion problems treated in Chapter 8
(heat equation Ut = uxx} Section 9.2 treats boundary value problems for vibrating strings
(wave equation Utt = c2 u",) Section 9.3 develops Fourier's method for steady-state
temperatures in plates (two-dimensional Laplace equation Uxx + Uyy = 0). Section 9.4
treats transient temperatures in plates (heat equation U t = Uxx +Uyy ). Section 9.5 analyzes
vibrations of circular drums (wave equation un = c2 (uxx +uyy )). Section 9.6 treats steady-
state heat diffusion in solids (Laplace equation Uxx + Uyy + u zz = 0). Finally, Section 9.7
presents an alternative to the Fourier method that is based on the Laplace transform.

9.1 Heat Diffusion in One Dimension

This first section treats problems of transient heat diffusion for which the temperature
is a function of the time t and one spatial coordinate. Eight such problems are solved.
558 9 BOUNDARY VALUE PROBLEMS OF MATHEMATICAL PHYSICS

----{l )~

x= o X=](
FIGURE 1. Homogeneous heat-conducting rod

Some of these problems were already solved, or partially solved, in Chapter 8. In those
cases the results are summarized, and some extensions of the theory are presented.

Problem 9.1.1. Finite Rod, Ends at Zero Degrees. This problem was formulated, and
the uniqueness of the solution was shown, in Section 7.2, see also Figure 1. The problem
is to find a function u(x, t) such that

(1) U(x, t) is defined for 0::: x::: Jr, t ::: 0,

(2) Ut = Uxx for 0::: x ::: Jr, t ::: 0,
(3) U(O, t) = 0 and u(Jr, t) = 0 for t::: 0,
(4) u(x,O) = j(x) for 0::: x ::: Jr,

where j(x) is a prescribed function on the interval 0 ::: x ::: Jr.

The separation of variables method led to the formal solution

(5) u(x, t) =L
00

n=l
bn sin(nx)e- n1t , bn = - 21
]( 0
n
j(x) sin(nx) dx.

This satisfies (1) through (4) and hence defines the solution u(x,t) when u(x,O) = j(x)
and

L b sin(nx),
00

(6) j(x) = n 0::: x ::: Jr.

n=l

It was shown in Section 8.3 that sufficient conditions for this are that j and l' be
sectionally continuous and satisfy the conditions j(O) = j(Jr) = 0 and j(x) =
~ [r(x + 0) + j(x - 0)] for 0 < x < Jr. Clearly, the factors e- n2t in (5) make the series
converge very rapidly when t > 0, even if j(x) does not satisfy the conditions.
Under slightly stronger restrictions on j(x) one has

(7)

and hence the series in (6) converges absolutely and uniformly for allx. This is a corollary
of the following theorem of Churchill; see [Ch-B2J.
Churchill's Theorem. Let j(x) be defined for -Jr ::: x ::: Jr and assume that

(8) j(x) is continuous for - Jr::: x::: Jr andj(-Jr) = j(Jr) ,

 9.1 HEAT DIFFUSION IN ONE DIMENSION 559

(9) f'(x) is sectionally continuous for - 11: ~ X ~ 11:,

(10) ak = -11:117<
_ f(x) cos(kx) dx, bk = -11:117<
_ f(x) sin(kx) dx.

Then

fJa~+b~ < 00
k=l

and the Fourier series for f(x) converges uniformly to f(x) for -11: ~ X ~ 11:.
If f(x) is given for 0 ~ x ~ 11: then one gets similar results for the sine and cosine
series of f(x) by applying Churchill's theorem to the odd and even extensions of f(x) ,
respectively. This gives the following corollaries.
Corollary. (Sine Series). Letf(x) be continuous, letf'(x) be sectionally continuous
on 0 ~ x ~ 11:, and letf(O) = f(11:) = O. Then (7) holds for
bn = -217< f(x) sin(nx) dx.
11: 0

Corollary. (Cosine Series). Let f(x) be continuous and let f'(x) be sectionally
continuous on 0 ~ x ~ 11:. Then

an = - 217< f(x) cos(nx) dx

11: 0

satisfies

(ll)
n=l
Computation of u(x, t). To compute u(x, t) from the series solution (5) we must truncate
the series; that is, replace (5) by a partial sum:

Lb sin(nx)e-
N
(12) UN(X,t) = n
n2t .

n=l

This is equivalent to replacingf(x) by the corresponding partial sum

L b sin(nx).
N
(13) fN(X) = n
n=l

Similarly, the error u(x, t) - UN (X, t) is the solution of the diffusion problem with initial
values f(x) - fN(X) (the truncation error for (6». We wish to show that the truncation
error in u(x, t) does not exceed that for f(x):

(14)
560 9 BOUNDARY VALUE PROBLEMS OF MATHEMATICAL PHYSICS

This is a consequence of the maximum principle for the heat equation; see page 41 in
[Strj. Thus to achieve a satisfactory error bound for u(x, t) at any t it is enough to choose
N so as to achieve the same bound at t = O.

Problem 9.1.2. Finite Rod, Both Ends Insulated. For a rod of length L = Jr and
thermal diffusivity K = 1 this problem was formulated and solved by Fourier's method
in Section 8.4. The problem is to find a function u(x, t) such that

(15) u(x, t) is defined for 0 S x S Jr, t :::: 0,

(16) Ut = Uxx for o S x S Jr, t :::: 0,
(17) uxCO, t) = 0 and ux(Jr, t) = 0 for t :::: 0,
(18) u(x,O) = j(x) for Os x S Jr,

where j(x) is a prescribed function. The Fourier method led to the formal solution
00

(19) u(x, t) = ao + Lan cos(nx)e-n2t ,

n=1

where

(20) ao = - ll
Jr 0
1C
j(x)dx

and

(21) an = -
Jr
2i 0
1C
f(x) cos(nx) dx for n = 1,2,3, ....

In particular, setting t = 0 in (19) gives the Fourier cosine series representation of j(x):
00

(22) f(x) = ao + Lan cos(nx).

n=1

Sufficient conditions for (19)-(22) to give the solution of the boundary value problem
were derived in Section 8.4. Under the stronger conditions of Churchill's theorem,
Corollary 2, one has

n=1

and the series in (19) converges absolutely and uniformly for 0 S x S Jr.

Computation of u(x, t). Denote the Nth partial sums by

N
fN(X) = ao + Lan cos(nx), N = 1,2, ...
n=1
 9.1 HEAT DIFFUSION IN ONE DIMENSION 561

and
N
UN (X, t) = ao + Lan cos(nx)e-n 2t
, N = 1,2, ....
n=1

Then the error in replacing U by UN is

(23) U(X,O - UN (X, t) =

n=N+l

Ifj satisfies the hypotheses of Churchill's theorem, Corollary 2, then (23) and the triangle
inequality give the error estimate

L
00

Iu(x, t) - UN (X, 01 s lanl,

n=N+1

since I cos(nx)e-n 2t lSI. The last sum is an error bound for the Fourier cosine series for
j(x):

L
00

If(x) - jN (x) I S lanl·

n=N+1

°
To see how this may give a numerical error bound for the remainder (23), consider
Example 2 of Section 8.4: j(x) = x for Sx S Jr. Then
-4
a2n = 0, a2n+l = Jr(2n + 1)2' n = 0,1,2, ... ,
and hence, replacing N by 2N, we get

as an error bound for Iu(x, t) - U2N(X, t)l. The integral test for estimating the sum of a
series gives

N = 1,2, ....

Thus
1 1
RN S E if N > -
- 2
+ -.
JrE

As a specific example take E = 0.01. Then

RN S 0.01 if N:::: 33.

562 9 BOUNDARY VALUE PROBLEMS OF MATHEMATICAL PHYSICS

This estimate, which holds for all t ::: 0, is pessimistic. For t > 0 the factors e- n2t make
the series in (19) converge much more rapidly.

Problem 9.1.3. Finite Rod, Mixed Boundary Conditions. This problem illustrates the
Sturm-Liouville theory for a homogeneous rod with different boundary conditions at
the two ends. The boundary value problem is to construct a function u(x, t) such that

(24) u(x, t) is defined for 0 ::::: x ::::: 1, t ::: 0,

(25) Ut = Uxx for 0::::: x ::::: 1, t ::: 0,
(26) ux(O, t) = 0 and ux(l, t) + u(l, t) = 0 for t::: 0,
(27) u(x,O) = j(x) for 0::::: x::::: 1,

where j(x) is a prescribed function. Physically, the boundary conditions (26) mean
that the end of the rod at x = 0 is insulated, while the end at x = 1 radiates
heat into an exterior medium at temperature zero. This problem illustrates the general
Sturm-Liouville theorem in a case where all functions can be calculated expliCitly.

Application of the separation of variables method leads to a series solution of the form

=L
00

(28) u(x, t) cnXn(x)e-Ant,

n=l

where the coefficients Cn are calculated via the orthogonality property as

(29)
Jo1 j(x)Xn (x)dx n=1,2,3, ....
C
n= Jo1 Xn(x)1 2 dx'
I

Here An and Xn(x) are the eigenvalues and corresponding eigenfunctions of the Sturm-
Liouville problem

(30)
X"(x) + AX(X) = 0 for 0::::: x::::: 1,
X/CO) = 0 and X/(l) + X(l) = O.

They will be calculated by the procedure followed in Section 8.13.

The validity of the representation (28) depends on the inclusion of all eigenpairs
(An,Xn(X». The eigenvalues must be real (see Section 8.12). Hence only real AnS need
to be checked. It will be convenient to treat the cases A < 0, A = 0 and A > 0 separately.
Such a A is an eigenvalue if and only if the problem (30) has a nontrivial solution. The
calculations follow the plan of Section 8.13.
 9.1 HEAT DIFFUSION IN ONE DIMENSION 563

Case)... < O. Write)... = _/-L2 with /-L > O. Then the general solution of the differential
equation of (30) is

(31)

with derivative

where hI and hI are arbitrary constants. Thus)... = - /-L I is an eigenvalue if and only if
there is a pair (hI, h2 ) :j:. (0,0) such that

x' (0,)...) = /-Lh2 = 0 (so hI = 0),

X'(l,)...) + X(l,)...) = hl/-L sinh(/-L) + hI cosh(/-L) = o.
But /-L sinh /-L + cosh /-L > 0 for all /-L > O. Thus hI = 0 and so (hI, hI) = (0,0); that is,
X(X,)...) is the trivial solution, and (30) has no negative eigenvalues.
Case)... = O. In this case

so

X'(O,O) = hI = 0, X'(l,O) + X(I, 0) = hI + hI = 0,

so again (h l , h2 ) = (0,0), and)... = 0 is not an eigenvalue.
Case)... > O. Write)... = /-L 2 with /-L > O. Then

(32) X(X,)...) = hI cos(/-Lx) + hI sin(/-Lx),

X'(x,)...) = - hl/-L sin(/-Lx) + hI/-L cos(/-Lx),

and hence

Hence the two boundary conditions give the following linear system for (hI, h2 ):

X'(X,)...)= hl/-L = 0, (so h2 = 0),

X'(l,)...) +X(l,)...) = hI [COS/-L - /-Lsin/-L] = O.

For a nontrivial solution we must have hI :j:. 0, since hI = O. Thus)... is an eigenvalue if

and only if )... = /-L I (with /-L > 0) and

(33) cos /-L = /-L sin /-L.

564 9 BOUNDARY VALUE PROBLEMS OF MATHEMATICAL PHYSICS

y
10

o
o 1f
"2
rr h 2rr
2 FIGURE 2. Graphical solution of tan JL = l/JL

Clearly, cos JL =f. O. Hence, (33) can also be written in the equivalent form
1
(34) tanJL = -.
JL
The corresponding eigenfunction is given by (32) with k2 = 0; that is, nonzero multiples
of

(35) X(X,A) = cos(JLx).

Note that we have proved that all eigenvalues are positive.

It is easy to see that the eigenvalue equation (34) has infinitely many positive solutions
JLn with corresponding eigenvalues An = JL~, n = 1, 2, .... In fact, if we graph the curves
Y = tan JL and y = 11JL in the first quadrant of the (JL,y)-plane, then their intersections
give the values JLn; see Figure 2.
The function tan JL has zeros at JL = 0, n, 2n, ... and vertical asymptotes at JL = n/2,
3n/2, 5n/2, .... Hence if the JLn are numbered in increasing order, then
n 3n 5n
o < JLl < -, n < JL2 < - , 2n < JL3 < - , etc.,
2 2 2

TABLE 1. Roots oftan JL = 11 JL

JLl = 0.274n JL6 = 5.020n

JL2 = 1.090n JL7 = 6.017n
JL3 = 2.049n JLs = 7.014n
JL4 = 3.033n JL9 = 8.013n
JLs = 4.025n JLlO = 9.011n
 9.1 HEAT DIFFUSION IN ONE DIMENSION 565

and

fLn ;:::;; (n - l)rr for n ~ 00.

Accurate values of the eigenvalues can be obtained from equation (34) by standard
numerical methods, as developed in Chapter 1. In Table 1, the first ten crossings of
y = tan x with y = l/x (see Figure 2) were computed by Newton's method.
Eigenfunctions. These have the form (35) with fL = fLn:

(36) Xn (x) = cos (fLnX) , n = 1,2,3, ....

Orthogonality. The Sturm-Liouville theory of Section 8.12 guarantees that the eigen-
functions (36) are orthogonal on the interval °
::s x ::s 1 with weight function
w(x) = 1:

11 COS(fLmX) COS(fLnX) dx = 0, m =J=. n.

This can also be verified by direct integration and use of the eigenvalue equation (34).

Problem 9.1.4. Fourier's Ring Problem. This is the physical problem of predicting the
diffusion of heat in a ring of unit radius when the initial temperature is given; see Figure
3. The problem was already discussed in some detail in Section 8.5. Here we note that
u(x, t), the temperature distribution at time t, is a 2rr-periodic function of the angular

°
coordinate x. Hence u(x, t) is determined by its values on any x-interval of length 2rr.
In Section 8.5 the interval ::s x ::s 2rr was used. It is often convenient to use instead
the symmetric interval -rr ::s x ::s rr.
In this formulation u(x, t) is defined by the follOwing properties.

(37) u(x, t) is defined for -rr ::s x ::s rr, t ~ 0,

(38) Ut = Uxx for -rr::S x ::s rr, t ~ 0,
(39) u( -rr, t) = u(rr, t) and ux ( -rr, t) = ux(rr, t) for t ~ 0,
(40) u(x,O)=j(x) for -rr::sx::Srr,

FIGURE 3. Circular ring and angular coordinate x

566 9 BOUNDARY VALUE PROBLEMS OF MATHEMATICAL PHYSICS

where f(x) is a prescribed function.

The separation of variables procedure goes essentially as in Section 8.5 because the
eigenfunctions cos(nx) and sin(nx) all have period 2n. Thus
00

(41) u(x, t) = C/o + I)a n cos(nx) + bn sin(nx) }e- n2t ,

n=l

where an and bn are the Fourier coefficients off(x), given now by

(42) ao = -1 f71: f(x) dx,

2n -71:

an = -1 f71: f(x) cos(nx) dx, n = 1,2, ... ,

n -71:

bn = -1 f71: f(x) sin(nx) dx, n = 1,2, ....

n -71:

Conditions for the validity of the solution (41), (42) are that f(x) and rex) be sectionally
continuous and satisfy the boundary conditions. A particularly favorable case occurs
when f(x) satisfies the conditions of Churchill's uniform convergence theorem, namely,

(43) f(x) continuous for -n ::: x ::: n,

j'(x) is sectionally continuous for -n :::: x:::: n,

f( -n) = fen).

Then

and hence the series in (41) converges absolutely and uniformly for all x.
Error Estimates. Conditions (43) imply that the partial sums
N
(44) UN (x, t) = ao + L{an cos(nx) + bn sin(nx)}e- n2t
n=l

converge uniformly to u(x, t) as N --+ 00. Moreover, (41), (44) and the triangle inequality
give the error estimate

L
00

(45) lu(x, t) - UN (X, 01::: (Ian I + Ibn!) .

n=N+l

The sum on the right can often be estimated, as in Problem 9.1.2, to give practical error
estimates.
 9.1 HEAT DIFFUSION IN ONE DIMENSION 567

Problem 9.1.5. Heat Diffusion in a Sphere (Symmetric Case). The boundary value
problem of heat diffusion is a solid body Q was formulated in Section 7.6. This problem
will treat the special case where Q is a sphere of unit radius defined by the inequality
(46)

See Figure 4. If the surface of Q is held at zero degrees and if the initial temperature is
a function fer) of the radial coordinate only, then we may expect that the temperature is
a function u(r, t) for all t 2: O. This conjecture will be proved by constructing a solution
of the form u = u(r, t) The uniqueness theorem for the problem (see Section 7.6) then
implies that there can be no other solution.
The heat equation for three-dimensional bodies is

where V2, the Laplace operator, has the form

in rectangular coordinates. If u = u(r, t) depends onx,y and Z only in the combination r,

defined by (46), then a simple calculation based on the chain rule gives (see the Exercises)
2 2
(47) Vu= U rr + -Ur'
r
Hence the boundary value problem for the sphere with symmetric initial values u(r, 0) =
fer) can be formulated as follows
Heat Diffusion in a Sphere (Symmetric Case). Find a function uCr, t) such that

(48) u(r, t) is defined for 0 ::S r ::S 1, t 2: 0,

C49) Ut= U rr + ~Ur for 0 < r ::S 1, t 2: 0,

(50) u(l, t) = 0 for all t 2: 0,
(51) u(r, t) and urCr, t) are bounded for 0 ::S r ::S 1, t 2: 0,
(52) u(r,O) = fer) for O::s x ::S 1,

FIGURE 4. Sphere Q of radius 1

568 9 BOUNDARY VALUE PROBLEMS OF MATHEMATICAL PHYSICS

where fer) is a prescribed function.

We shall see that conditions (51) are needed to eliminate possible singularities at
r = O. From the physical problem, u and its derivatives should be finite at all interior
points of the sphere including the center at r = O.
Separation of Variables. We begin as usual by seeking product solutions of conditions
(48) through (51):

(53) u(r, t) = R(r)T(t).

Substituting this into (49) and separating variables gives
T'(t) R"(r) + (2/r)R'(r)
(54) -= =-A
T(t) R(r) ,

where A is a separation constant. In particular, the time factor T(t) may be taken to be

(55)

For the radial factor R(r) we get the following singular Sturm-liouville problem:
2
(56) R"(r) + -R'(r) + AR(r) = 0 for 0 < r S 1,
r
R(l) = 0,
R(r) and R'(r) are bounded for 0 < r :::: 1.

This may be put into self-adjoint form as

(57) (r2RI)' + ),.r2R = 0 for 0 < r :::: 1,

R(l) = 0, R(r) and R'Cr) bounded for 0 < r :::: 1.

This has the standard form of Section 8.11 with coefficients per) = w(r) = r2, q(r) = O.
The problem is singular because p(O) = O. The eigenfunctions will be orthogonal on
o :::: r :::: 1 with weight w(r)
= r2.
Solution of the Differential Equation. The differential equation of (56) can be
Simplified by the change of dependent variable

vCr) = rR(r), or R(r) = v(r)/r.

Substitution into (56) gives, after a short calculation,

v"(r) + Av(r) = O.

Thus the differential equation for R(r) has the general solution

( )r - l_ k sin.JAr k cos.JAr
(58) R +2 .
r r
 9.1 HEAT DIFFUSION IN ONE DIMENSION 569

Functions R(r) and R'(r) will be bounded on 0 < r ::: 1 if and only if k2 = O. Thus we
may take

sin ,J).r
R( r) = .
r
The boundary condition at r = 1 now gives the eigenvalue equation

R(1) = sin v'A = O.

The roots satisfy,J). = nn en = 1,2, ... ).

Eigenpairs. We have

An = n 2n 2 , n = 1,2, ... ,
sin(nnr)
RnCr) = , n = 1,2, ....
r
Product Solutions. Combining (53) with (55), we have

(59) Un
(r, t ) -_ sin(nnr) e
_n2n2(
, n = 1,2, ....
r
Superposition. The formal solution of the boundary value problem is

~ sin(nnr)
uCr, t ) =
_n2n2(
(60) ~ Cn e.
n=1 r

Initial Values. From (60) we have

~ sinCnnr)
(61) u(r,O)=f(r)=~cn , o ::: r ::: l.
n=1 r

Orthogonality. The eigenfunctions are orthogonal on 0 ::: r ::: 1 with weight w(r) = r2,
so

(62) t
10
(SinCnmr») (SinCnnr») r2 dr =
r r
I ~ if m # n,
- If m = n.
2
This can be checked by direct integration.
Coefficients. The coefficients Cn are given by (61) and (62), namely

~Cm = 11 fer) CinC:mr») r2 dr,

or

(63) Cm = 211 fCr) sinCnmr)r dr, m = 1,2,3, ....

570 9 BOUNDARY VALUE PROBLEMS OF MATHEMATICAL PHYSICS

We see that the coefficients {en} are just the Fourier sine series coefficients of rf(r) on
o :s r :s 1. In particular, if rf(r) satisfies the conditions of Churchill's theorem, then the
series solution (60), (63) converges uniformly.

Problem 9.1.6. Heat Diffusion in a Disk (Symmetric Case). The problem of transient
heat diffusion in a plate Q was formulated in Section 7.3. Here we shall solve it for the
special case that Q is a disk of unit radius defined by

(64) r = -/x 2 + y2 :s 1,

and the initial temperature is a function of r alone; see Figure 5.

We shall construct it as a function u = u(r, t) having the same property. This function
must then be the solution guaranteed by the uniqueness theorem.
Application of the chain rule gives, after a short calculation (see the Exercises),

(65) V 2 u(r, t) = ( a
ax2
2
+ a
ay 2
2
) u(r, t) = Urr + ~UT.
r

Hence the boundary value problem of Section 7.3 specializes to the search for a function
u(r, t) that satisfies

(66) u(r, t) is defined for 0 :s r :s 1, t ~ 0,

(67) Ut = UTT + ~UT for 0 < r :s 1, t ::: 0,
(68) u(l, t) = 0 for t::: 0, u and UT bounded for 0 :s r :s 1, t ::: 0,
(69) u(r,O) = fer) for O:s x :s 1,

where fer) is a given function. The boundedness of u and U T is needed to eliminate

possible singularities at r = O.
The problem defined by (66)-(69) appears to be much like the preceding sphere
problem, differing only in the coefficient of UT • Surprisingly, this small change complicates
the work by causing the appearance of Bessel functions.
Separation of Variables. The search for product solutions

(70) u(r, t) = R(r)T(t)

(o,~o-)---1 (1,0)

FIGURE 5. Disk of radius 1

 9.1 HEAT DIFFUSION IN ONE DIMENSION 571

leads to T(t) = e- At and

1
(71) RI/(r) + -R'(r) + AR(r) = 0,
r
where A is the separation constant. This last equation has selfadjoint form

(72) (rR')' + hR = 0

and thus per) = w(r) = rand q(r) = 0 (see Section 8.11). Note that (72) is singular on
Os r S 1 because p(O) = O.
The eigenpairs (A, R(r)) are the solutions of the singular Sturm-Liouville problem

(73) (rRI)' +hR = 0 for 0 S r S 1,

R(l) = 1,
R(r) and R'(r) bounded on 0 S r S 1.

This problem was solved in Section 8.14.

Eigenpairs. These were found to be (see Section 8.14)

(74) An = s~, Rn(r) =Jo (snr), n = 1,2, ... ,

where S1, S2, ... are the positive zeros ofJo(x), enumerated in increasing order

o< S1 < 52 < ... < 5n < ... ,

and

5n ;:::; (n - ~) JT: for n-+ 00.

Product Solutions. These are the scalar multiples of

(75) n = 1,2, ....

Superposition. The solution of the disk problem will have the form

L enio (snr) e-s~t.

oc
(76) u(r, t) =
n=1

Initial Values. To complete the solution of the boundary value problem, we must have
oc
(77) u(r,O) = fer) = L enio (Snr) ,OS r S 1.
n=1
572 9 BOUNDARY VALUE PROBLEMS OF MATHEMATICAL PHYSICS

This is the Fourier-Bessel series of Section 8.14. The orthogonality property implies that
the coefficients {cn} are given by

10l j(r)10 (snr) r dr

(78) Cn = 1
1015 (Snr) r dr
Note the required weight factor w(r) = r appearing in (78).
Problem 9.1.7. Heat Diffusion in an Infinite Rod. The study of this problem was begun
in Section 8.7. We recall that the temperature function u(x, t) was determined by the
following properties:

(79) u(x, t) is defined for -00 < x < 00, t :::: 0,

(80) Ut = Uxx for -00 <x< 00, t :::: 0,

(81) u(x, t) is bounded for -00 < x < 00, t :::: 0,

(82) U(x, 0) =j(x) for -00 <x < 00,

where j(x) is a prescribed function on -00 <x < 00.

The separation of variables method was shown to lead to the solution

(83) u(x, t) = 100

(A(JL) cOS(JLX) + B(JL) sin(JLx» e- ph dJL,
where A(JL), B(JL) are determined by the Fourier integral theorem as

(84)

Equations (83) and (84) give the solution of the boundary value problem when j(x)
satisfies the conditions of the Fourier integral theorem (see Section 8.9).
To compute the solution u(x, t) by means of (83), (84), it is necessary to carry out
three integrations. We shall now show that this can be reduced to a Single integration.
Equation (84) and the trigonometric identity

imply that
 9.1 HEAT DIFFUSION IN ONE DIMENSION 573

Double Integral Representation. Combining (83) and (85) gives the equation

(86) u(x, t) = ~ 1 (i: j(~)

00
cos IJ./X - ~)d~ ) e-
fl2t
dfL,

which requires only two integrations. We shall reverse the order of integration in (86)
and carry out the inner integration to get a simple integral representation for u(x, t).
The change of order of integration in (86) is justified by a theorem of advanced
calculus if

(87)

This gives
i: If(x) I dx < 00.

(88)

To evaluate the inner integral we recall the Fourier cosine integral

lC<; e- 1L2 I(4a) COS(fLY) dfL = ..j7Uie-ay2 (x real, a > 0),

which was derived in Section 8.9. On taking 1I(4a) = t and Y = x - t we get

(89) roo e-fl2t cos fLeX _ ~) dfL = {iT e- (X~;)2 .

h V~
Finally, combining (88) and (89) gives the desired single integral formula for the solution

1
of the boundary value problem:

1 00
(90) u(x, t) = c;
(x_()2
j(~)e--4-t d~, t> o.
2y nt -00

This formula was derived under the integral assumption (87). However, the formula is
valid for a larger class of initial values. For example, it is enough that j(x) be bounded and
sectionally continuous, in which case it may be shown that if U is defined by (90), then
1
lim u(x, t) = - [[(x + 0) + j(x - 0)] for all x.
t--+O+ 2
The proof, which depends on methods of advanced calculus, will not be given here; see
[Wi2l.
An Alternative Formula. If in (90) we make the change of variables s = (~- x)/(2,.fi.) ,
we get

(91) U(x, t) =
y
1
r.;;.
n
1 00

-00
j(x + 2s,.fi.)e- S2 ds.
574 9 BOUNDARY VALUE PROBLEMS OF MATHEMATICAL PHYSICS

From this formula it is rather easy to show that limHo+ u(x, t) = j(x) whenever j(x) is
continuous and bounded (see the Exercises).

EXAMPLE 1. Consider the special case where the initial temperature is the Heaviside
function: j(x) = H(x). The solution is given by (91). Since x + 2s.Jt ::: 0 if and only if
s ::: -xle2.Jt), we have
(92) u(x, t) = 1
r;;. /00 e-s'ds.
'\j J( -xJ(2..!iJ

The Error Function. The integral in (92) can be expressed in terms of the error
function of probability theory The latter, denoted by erf (x), is defined by the equation

(93) 2
erf(x) = ~ l 0
x
e- s, ds.

The function is tabulated (see [A-S]) and is available via many computer algebra systems.
Numerical evaluation of erf(x) via numerical integration was discussed in Section 1.3.
The complementary error function is also useful. It is defined by

(94) 2
erfc(x) = 1 - erf(x) = ~ 100
x e- s, ds.

Useful properties of erf(x) are

(95) erf( -x) = - erf(x) (erf(x) is odd),

erf(O) = 0, lim erf(x) = 1.
X---++OO

Example 1 (Concluded). The properties of the error function allow us to write the
solution u(x, t) of (92) in the following equivalent forms:

-x ) -x )
(96) u(x, t) = 2:1 erfc ( 2.Jt = 2:1 - 2:1 erf ( 2.Jt
(x)
= 2:1 + 2:1 erf 2.Jt .
From this and the properties of erf(x) we can derive the following results:

limt-Hoo uex, t) = ~ for all x,

u(O, t) = ~ for all t::: 0,
limx -++ oo u(x, t) =1 for all t > 0,
limx -+_ oo u(x, t) = 0 for all t > O.
 9.1 HEAT DIFFUSION IN ONE DIMENSION 575

o 8
FIGURE 6. Graph of ~ + ~ erf(xI(2y't)) with t
-8 -4 4
fixed It = 1)

A graph of u(x, t) as a function of x, with t > 0 fixed, is shown in Figure 6. Recall from
Section 1.3 that erf(x) ::: 0.9999 for all x ::: 3.
The Fundamental Solution K(x, t). The solution (90) of the initial value problem
for the infinite rod is a convolution

(97)

where
U(x, t) = i: K(x - t OJ(/;) d/;,

(98) K(x , t) = _1_e-x2/4t , -00 < x < 00,

t >0.
2JITt
It is easy to verify that K is a solution of the heat equation

Kt = Kxx for - 00 < x < 00, t > O.

See the Exercises. The function K is called the fundamental solution of the heat
equation. To see its physical interpretation we may show that

limHo+ K(x, t) = 0 for all x #- 0,

limHo+ K(O, t) = +00.

°
Thus K(x, t) is the temperature at point x and time t > due to a quantity of heat placed
°
at x = at time t = 0. To determine how much heat is needed to generate K, we define

(99) Q(a, b, t) = lb K(x, t) dx,

where (a, b) is an arbitrary interval. Physically, Q(a, b, t) is the amount of heat in the
segment a :::: x :::: b at time t > 0. We shall calculate its limit as t -+ O+. From the
definition

Q(a, b, t) = -1-
2JITt
lb a
e-x2/4t dx
576 9 BOUNDARY VALUE PROBLEMS OF MATHEMATICAL PHYSICS

and after the change of variables x = (20)s, we have

(100) Q(a, b, t) =
1
'-
l b'(2.ji)
e- 52 ds
v 1T a/(2.ji)

= 1 [erf (2~) - erf (2~)]'

From properties (95) it is seen that for any finite interval (a, b) we have

limQ(a,b,t)=jo ifa<b<OorO<a<b,
[-+0+ 1 if a < 0 < b.

This makes precise the statement that the fundamental solution K(x, t) is the temperature
at point x and time t > 0 due to the release of one calorie of heat at x = 0 at time t = O.

Problem 9.1.8. Heat Diffusion in Semi-Infinite Rods. This problem was solved in
Section 8.10 for cases of zero temperature or heat flux (Dirichlet or Neumann boundary
condition) at the near end of the rod. See Figure 7. Here Single-integral formulas for the
solution are obtained.
Zero Temperature at the Boundary. For this case we found that if u(x, 0) = f(x) and

(101) B(fJ,) =- 21
1T 0
00
j(~) sin(fJ,~) d~

then

(102) u(x, t) = i oo
B(fJ,) sin(fJ,x)e- 1l2t dfJ,.

A single integral formula for u(x, t) in terms of f(x) can be derived by substitution of
(102) into (101), changing the order of integration and evaluation of the inner integral,
as in the case of the infinite rod problem. However, it is easier to apply the result for the
infinite rod to f*(x), the odd extension off(x), as in Section 8.10. Then f*(x) = -fe-x)
for x < 0 and the integral formula (90) of Problem 9.1. 7 gives

(103)

(u)________________~I_ ~
-,,=0 FIGURE 7. Heat diffusion in a semi-infinite rod
 9.1 HEAT DIFFUSION IN ONE DIMENSION 577

Putting ~ = -T] in the last integral gives, after simplification,

(104) u(x, t) = _1_ t>O J(T]) (e-(X-~)2/(4t) _ e-(x+~)2/(4t)) dT],

2Ft 10
or

(105)

where

(106) K (x T] t) = _1_ (e-(X-~)2/(4t) _ e-(x+~)2/(4t))

D " 2Ft .
This function is called the fundamental solution of the semi-infinite rod with Dirichlet
boundary condition.
Zero Flux at the Boundary. In this case the solution is produced by solving the infinite
rod problem with even initial values. The result is

(107) u(x, t) = 1
00
KN(x, T], t)J(T]) dT],

where

(108) KN(X, T], t) = 2~ (e-(X-~)2/(4t) + e-(x+'I)2/(4t)) .

The function KN is called the fundamental solution for the semi-infinite rod with
Neumann boundary conditions. Formally, the fundamental solutions KD and KN differ
by a single sign.

I Exercises 9.1 I
Finite Rod. Solve the follOwing boundary value problems for u(x, t).
1. Dirichlet Problem 9.1.1 ,f(x) = sin 2x - sin 5x.
2. Dirichlet Problem 9. 1.1, f(x) = sin lOx.
3. Neumann Problem 9.1.2,f(x) = 1- cos2x + sin 2 x.
4. Neumann Problem 9.1.2,f(x) = 4 + sin2 2x.
5. Mixed Dirichlet-Robin Problem 9.1.3,f(x) = cos(/-lx), where /-l tan/-l = I and /-l R: 0.27411'.
6. Mixed Dirichlet-Robin Problem 9.1.3,f(x) = 3 cos(/-lx), where /-l tan/-l = 1 and /-l R: 2.04911'.
Fourier's Ring Problem. Solve for u(x, t).
7. Problem 9.1.4,f(x) = sin 2x.
8. Problem 9.1.4,f(x) = cos2x.
578 9 BOUNDARY VALUE PROBLEMS OF MATHEMATICAL PHYSICS

Heat Diffusion in a Sphere. Solve for u(r, t).

9. Problem 9.1.5,f(r) = 1.5 sin(rrr)lr.

10. Problem 9.1.5,f(r) = (3 sin(2rrr) - sin (rrr))lr.

Heat Diffusion in a Disk. Solve for u(r, t).

11. Problem 9.1.6,f(r) = JO(SI r).

12. Problem 9.1.6,f(r) = 2JO(S2r) - JO(Slr).

13. (Radial Heat Equation n = 3) Assume that the 3-dimensional heat equation
Ut = Uxx + Uyy + Uzz
has a solution u = u(r, t) that depends on x, y and z only in the combination r = ,jx2 + y2 + Z2.
Show by the chain rule that the heat equation for u can be written as

2
Ut = Urr + -Ur·
r

14. (Orthogonality) Prove by direct integration, using only calculus methods, that

11 (sin(rrmr)) (sin(rrnr)) 2
---
o r r
- - - r dr=
{O1 ifm i= n,
-
2
if m = n.

15. (Radial Heat Equation n = 2) Assume that the 2-dimensional heat equation
Ut = Uxx + Uyy
has a solution u = u(r, t) that depends on x and y only in the combination r = ,jx 2 + y2. Show by
the chain rule that the heat equation for U can be written as

1
Ut = Urr + -Ur·
r

16. (Heat Integral) Letf(x) be continuous and bounded on -00 < x < 00. In the integral formula

u(x, t) =
"rr
1
r.; 1 00

-00
f(x + 2s0)e-s2 ds,

show that for each fixed x, limHo+ u(x, t) = f(x) , by using the equality

lim
N->oo
1r.;
"rr
lN-N
e- s2 ds = 1.

17. (Fundamental Solution) Let

K(x t) -_ _
'2v'iii1_ e-x2 /(4t)
, -00 < x < 00
,>.
t °
Verify using only calculus methods that K is a solution of the heat equation Kt = Kxx for -00 < x < 00,
t > 0.
 9.2 VIBRATION OF STRINGS AND TRAVELING WAVES 579

9.2 Vibration of Strings and Traveling Waves

This section treats the vibration of uniform strings under tension. For strings of finite
length, fixed at the ends (Dirichlet boundary condition), the boundary value problem was
formulated and the uniqueness of the solution was shown in Section 7.1. In this section
the solutions are constructed, both for the Dirichlet and for other boundary conditions,
and the theory is extended to infinite and semi-infinite strings.
The governing partial differential equation for vibrations of uniform strings is
d'Alembert's wave equation

(1)

where u(x, t) is the string displacement at point x and time t. In (1), c = .JTlP, Tis
the uniform tension in the string, and p is its mass density. We begin this section with a
discussion of d'Alembert's remarkable discovery that all solutions of (1) represent waves
that travel along the string with constant speed c.
Traveling Waves on Strings. We begin with the observation that if ¢(x) is any
function with a second derivative then

(2) u(x, t) = ¢(x - ct)

is a solution of (1). Indeed, two applications of the chain rule give Uxx = ¢"(X - ct) and
Utt = C2¢"(X - ct), which imply (1).
The graph of (2) in the ex, u)-plane, with t fixed, is identical to the graph of u = ¢(x),
but translated to the right by ct units, as in Figure 8 below.
Equation (2) represents a traveling wave, propagating to the right without change of
form. The propagation speed is clearly c (see the Exercises). In the same way, if u = y,(x)
is another function with second derivative then

(3) uex, t) = y,(x + ct)

is a solution of (1) that represents a traveling wave that moves to the left with constant
speed c. Moreover, since the wave equation (1) is linear and homogeneous, we may add

-- x
:;.
FIGURE 8. Graph of uex, 0) = rf>ex) and
xo x=xo+ct uex, t) = rf>ex - ct)
580 9 BOUNDARY VALUE PROBLEMS OF MATHEMATICAL PHYSICS

(2) and (3) to get another solution:

(4) u(x, t) = ¢(x - ct) + 1/r(x + ct).

It is a surprising fact that every solution of (1) has the form (4) for some functions ¢ and
1/r. More precisely, we have the following theorem.
D'Alembert's Theorem. For each solution u(x, t) of (1) there are functions ¢(x) and
1/r(x) such that (4) holds. Moreover, if ¢ and ~ are a second pair then there is a constant
k such that

¢(x) = ¢(x) - k, ~(x) = 1/r(x) + k.

Proof (Optional). In the (x, t)-plane introduce new coordinates Cr, s) by

(5) r = x - ct, s= x + ct
and write

(6) u(x,t) = vCr,s),

so that for a given u, (5) and (6) define v. Then a simple application of the chain rule
gives (see the Exercises)

Thus (1) holds if and only if in the (r, s)-plane one has

(7)
a2 v
-=0.
aras
Integrating first on r gives
av
as = 8(s),

and hence integrating on s gives

(8) vCr, s) = ¢(r) + f 8(s) ds = ¢(r) + 1/r(s).

Combining this with (5) and (6) gives (4). To prove the last statement of the theorem,
we remark that if

u(x,O = ¢(x - ct) + 1/r(x + cO = ¢(x - ct) + ~(x + ct)

then

¢(r) - ¢(r) = ~(s) - 1/r(s) = k,

 9.2 VIBRATION OF STRINGS AND TRAVELING WAVES 581

where k is independent of both rand s.

Product Solutions-An Example. In Section 7.1 it was shown that the wave
equation (1) has product solutions such as

(9) u(x, t) = sin(jLx) sin(jLct),

where jL is a constant. It is not immediately apparent that this is a sum of traveling waves,
as guaranteed by d'Alembert's theorem. However, the trigonometric identity

sin asin b = ~ [cos(a - b) - cos(a + b)]

implies that (9) is indeed a sum of traveling waves.

The remainder of this section presents solutions of several boundary value problems
for the wave equation (1). D'Alembert's theorem and the separation of variables method
will be shown to provide alternative approaches to these problems.

Problem 9.2.1. Traveling Waves on an Infinite String. If a long taut string is struck
at a point far from its ends then intuition and experience suggest that traveling waves
will propagate out from that point with speed c. To analyze this problem, it is natural to
consider the idealized problem of an infinite string. Thus the displacement u(x, t) of the
string will be characterized by the following properties (see Section 7.1).

(10) u(x, t) is defined for -00 <x< 00, t :::: 0,

(ll) Utt = c2 uxx for -00 < x < 00, t ::: 0,
(12) u(x,O) = j(x) and Ut(x,O) = g(x) for -00 <x< 00,

where j(x) and g(x) are prescribed functions. We have already argued (see Section 7.1)
that to predict the future displacement, we must know the initial displacement and
speed, characterized by j(x) and g(x), respectively. The correctness of this model will
be shown by verifying that there is only one function u(x, t) with properties (10), (11)
and (12) (uniqueness). Then the solution will be constructed by two methods, based
on d'Alemberts' theorem and Fourier's integral theorem, respectively. Finally, the two
solutions will be shown to be the same, as required by the uniqueness theorem.

Uniqueness Theorem. For each choice of the functions j(x) and g(x) there is at
most one function u(x, t) that has properties (10), (11) and (12).
Proof. The linearity of the wave equation (11) implies that the uniqueness property is
equivalent to the statement that if j(x) = g(x) = 0 for -00 < x < 00 then u(x, t) = 0
for -00 < x < 00 and t :::: 0 (see the proof ofthe uniqueness theorem in Section 7.1).
582 9 BOUNDARY VALUE PROBLEMS OF MATHEMATICAL PHYSICS

To prove this we use d'Alemberts theorem to write

(13) U(x, t) = ¢(x - ct) + 1/I(x + ct)

and

Ut(x, t) = c (-¢'(x - ct) + 1/I'(x + ct)).

The initial conditions (12) with j(x) = g(x) = 0 give

(14) U(x,O) = ¢(x) + 1/I(x) = 0 for - 00 < x < 00,

whence

UxCx,O) = ¢' (x) + 1/1' (x) = 0 and

Ut(x,O) = + 1/1' (x») = 0

c (-¢' (x)

for all x. It follows easily that ¢'(x) = 1/I'(x) = 0 for all x, and hence ¢(x) = ki
and 1/I(x) = k2 (k i and k2 constants). Next, ki + k2 = 0 by (14), so ¢(x) = ki and
1/I(x) = - k1. Finally; by (13) we have

U(x, t) = ki - ki = 0 for - 00 < x < 00, t::: 0,

which completes the proof.
Construction of the Solution via d'Alembert's Theorem. A solution u(x, t) of the
boundary value problem (10), (11), (12) has the form (13) for some functions ¢ and 1/1, by
d'Alembert's theorem. To find these functions we will use the initial conditions (12). Thus
u(x,O) = ¢(x)+1/I(x) = j(x) and Ut(x, 0) = c (-¢'(x) + 1/1' (x) ) = g(x)forallx, whence

¢'(x) + 1/I'(x) = j'(x) ,

-¢' (x) + 1/1' (x) = g(x)lc.

Solving for ¢' and 1/1' gives

2¢'(x) = j'(x) - g(x)lc,

21/1'(x) = j'(x) + g(x)/c.

Integrating, a particular solution is

2¢(x) = j(x) - -l1

c
X
g(;) d;,

+ -l1
0
x
21/1(x) = j(x) g(;) d;.
c 0
 9.2 VIBRATION OF STRINGS AND TRAVELING WAVES 583

Substituting this into (13) gives the solution (see the Exercises for details)

(15) u(x, t) = -1 [r(x - ct)

2
+ j(x + ct)] + -1
2c
l x ct

x-ct
+ g(~) d~.

This formula is called d'Alembert's solution of the initial value problem (10), (11), (12).
We have shown that if (10), (11), (12) has a solution then it must be given by this
formula. The converse statement, that (15) defines a function that satisfies (10), (11),
(12), is easily verified by differentiation and setting t = O.
Construction of the Solution via Separation of Variables. The solution (15) can also
be derived by the separation of variables method. To simplify the writing, we shall do
this for the special case g(x) = 0 only: Thus we seek a function u(x, t) that satisfies

(16) u(x,O is defined for -00 <x< 00, t ::: 0,

(17) Utt = c2 uxx for -00 < x< 00, t ::: 0,
(18) Ut(x,O) = 0 for -00 <x< 00,

(19) u(x,O) = j(x) for -00 <x< 00,

where j(x) is a prescribed function. The first step of the separation of variables method
is to seek those product functions

(20) U(x,O = X(x)T(t)

that satisfy all the linear homogeneous conditions of the problem. Here this means
conditions (16), (17) and (18). Thus (20) is a product solution if and only if (see
Section 8.7 and Section 9.1)

X" + AX = 0 for - 00 < x < 00,

X(x) is bounded for - 00 < x < 00,

while

T"+C 2 AT=0 for t:::O,

TI(O) = 0,

where A is a separation constant. The solutions for A, X and Tare

A = j.L2 (j.L > 0),

X(x) = A(j.L) cos(j.Lx) + B(j.L) sin(j.Lx),
T(O = cos(j.Lct).
584 9 BOUNDARY VALUE PROBLEMS OF MATHEMATICAL PHYSICS

Thus the one-parameter family of product solutions is

U(X, t, j.L) = [A(j.L) cos(j.Lx) + B(j.L) sin(j.Lx)] cos(j.Lct),

where A(j.L), B(j.L) are arbitrary functions of j.L.
The second step of the separation of variables method seeks to satisfy the inhomo-
geneous boundary condition (19) by superposition of the product solutions; that is, we
try to solve (16)-(19) by an integral of the form

(21) u(x, t) = 1 00
[A(j.L) cOS(j.Lx) + B(j.L) sin (j.Lx) ] cos(j.Lct) dj.L.

This will satisfy (19) if

u(x,O) = f(x) = 1 00
[A(j.L) cOS(j.Lx) + B(j.L) sin(j.Lx)] dj.L.

This is a Fourier integral. Thus by the Fourier integral theorem we must choose A(j.L),
B(j.L) to be

(22) A(j.L) =;;1 foo

_oof(~) cos(j.L~) d~, B(j.L) = -1 foo f(~) sin(j.L~)~.
J'( -00

Thus (21) and (22) define the solution of the problem. They require three integrations.
However, we shall show that all integrations can be carried out to recover d'Alembert's
solution (15) with g(x) = O. To do this we note that (22) implies

A(j.L) cos(j.Lx) + B(j.L) sin(j.Lx) =;;1 foo

-00 f(~) cos j.L(x - ~) d~

and hence (21) may be written as

u(x, t) = -1
J'(
1 fOO
0
00

-00
f(~) cos j.L(X - ~) cOS(j.Lct) ~dj.L.

Next, the trigonometric identity

2 cos j.L(X - ~) cOS(j.Lct) = cos j.L(x - ~ - ct) + cos j.L(X - ~ + ct)

gives

u(x, t) = -1100 foo f(~) [cos j.L(X - ~ - ct) + cos j.L(x - ~ + ct)] d~dj.L.
J'( 0 -00

Combining this with Fourier's identity

f(x) = -1
J'(
1 foo
0
00

-00
f(~) cos j.L(x - ~) d~dj.L
 9.2 VIBRATION OF STRINGS AND TRAVELING WAVES 585

gives
1
(23) u(x, t) = 2." [f(x - ct) +f(x + ct)].
This is just d'Alembert's solution in the case where g(x) = o.
EXAMPLE 1. (Plucked String). To pluck a taut string, one gives it a (small) displacement
f(x), holds the string motionless, and then releases it. Hence g(x) == 0 and the solution
is given by (23). Thus the initial profile splits into two equal parts that propagate to the
left and right with speed c. See Figure 9 below.

Problem 9.2.2. Traveling Waves on a Semi-Infinite String. This problem is to find the
waves generated on a long taut string with one end fixed. If the string is plucked or struck
at a point near the fixed end (and therefore far from the other end) then it is natural
to model the problem as a semi-infinite string, stretched along the positive x-axis, with
displacement zero atx = 0 (Dirichlet boundary condition). The boundary value problem
is to find the displacement u(x, t) characLerized by the following conditions.

(24) u(x, t) is defined for 0 .::: x < 00, t ::: 0,

(25) Un = c2 u xx for o .::: x < 00, t ::: 0,

(26) u(O, t) = 0 for all t ::: 0,

(27) u(x,O) = f(x) for 0.::: x < 00,

(28) Ut(x,O) = g(x) for 0.::: x < 00,

where f(x) and g(x) are prescribed functions.

Uniqueness. This problem has (at mosL) one solution. To show this, it is again enough
to show that if f(x) = g(x) = 0 for all x ::: 0 then u(x, t) = 0 for all x ::: 0 and t ::: O.
This is most readily done by using d'Alembert's theorem. Thus

(29) u(x, t) = </l(x - ct) + ljr(x + ct) for all x::: 0, t::: o.

-
FIGURE 9. Graph of
u(x,1) = Hr(x - e1) + f(x + ct)]
±f(x + ct) f(x) ±l(x - et) and u(x, 0) = f(x)
586 9 BOUNDARY VALUE PROBLEMS OF MATHEMATICAL PHYSICS

°
Note that (29) requires that 1/r(~) be defined for ~ ::: only, but cJ>(~) must be defined
for all real~. Conditions (26), (27), (28) withf = g = give the equations °
(30) U(O, t) °
= cJ>( -ct) + 1/r(ct) = for all t::: 0,
(31) u(x,O) = cJ>(x) + 1/r(x) = °
for all x::: 0,
(32) ut(x,O) = -c [cJ>'(x) -1/r'(x)] = °
for all x::: 0.

From (30) and (31) we see that

(33) cJ>(x) = -1/r(x) = cJ>( -x) for all x::: 0,

so that cJ> is an even function. Also (31) and (32) give

1/r'(x) = cJ>'(x) = -1/r'(x) for all x::: 0,

°
so 1/r'(x) == and 1/r(x) = k (a constant). Next, cJ>(x) = -k by (31) again. Returning to
(29) gives u(x, t) = k - k = 0, which completes the proof.
Solution of the Problem via d'Alembert's Theorem (g = only). The solution of the °
boundary value problem (24)-(28) can be written in d'Alembert's form (29) again. We

°
shall use the initial and boundary conditions (26), (27), (28) to find the functions cJ> and
1/r. To shorten the writing, only the special case g = is treated. The complementary
°
case f = 0, g of- can be done by the same method. Note that to define u(x, t) for all
° °
x ::: and t ::: we must determine cJ>(~) for all t and 1/r(~) for all ~ ::: 0.
Boundary Condition. By (26) we have

U(O, t) = cJ>( -ct) + 1/r(ct) = ° for all t::: 0,

whence

(34) cJ>( -~) = -1/r(~) for all ~::: 0.

Initial Conditions. By (27) and (28) with g = °we have

(35) cJ>(x) + 1/r(x) = f(x) for all x::: 0,

and Ut(x, 0) = -c [cJ>'(x) -1/r'(x)] = 0. Thus cJ>' and 1/r' are defined by
(36) cJ>' (x) + 1/r' (x) = l' (x),
cJ>' (x) - 1/r' (x) = ° for all x::: 0.

This system has the unique solution

1
cJ>'(x) = 1/r'(x) = '2.1'(x) for all x::: 0,
 9.2 VIBRATION OF STRINGS AND TRAVELING WAVES 587

whence
1
¢(x) = "2f (x) + hI,
1
1jJ(x) = "2f (x) + h2 ,
where hI and h2 are constants. But then (35) implies that hI + h2 = 0, so
1
¢(x) = "2f (x) + hI,
1
1jJ(x) = "2f (x) - hI·

The constant hI may be absorbed into the definition of ¢ and 1jJ, as in Problem 9.2.1, so
1
(37) ¢(x) = 1jJ(x) = "2f (x) for all x 2: o.
Finally, (34) gives
1
¢(x) = -1jJ( -x) = --f( -x) for all x:::: 0,
2
or

~f(x)
1-"2f (
for x 2: 0,
(38) ¢(x) = 21
-x) for x < O.

This may be written as

1
¢(x) = "2/a(x) for all real x,

where fa (x) is the odd extension of f(x) to -00 < X < 00. Finally, the full solution of
(24)-(28) in the case g = 0 is
1 1
(39) u(x, t) = "2fa (x - ct) + "2f (x + ct).
The function ¢ will satisfy
1
¢(O+) = "2f (O+) ,
1
¢(O-) = -"2f (O+).
Thus ¢(x) will be continuous at x = 0 only if f(O+) = O. Physically, this means that
u(x,O) = f(x) must satisfy the boundary condition.
588 9 BOUNDARY VALUE PROBLEMS OF MATHEMATICAL PHYSICS

u
Stage 1 f(x)
_ _ _ _ _ _ _ _. - - ,_ _ _ _ _ _~----~--L-----~ X

Stage 2 -A B-
____________~~--~-L~~---L--~~ X

c-
Stage 3 B-
c-
------------------~ __~--------~~x

1 1
A = 2f (x +ct) B= -f(x - ct)
2

FIGURE 10. Three stages of propagation

Physical Interpretation of the Solution. Consider an initial displacement f(x) that is

zero outside an interval la, b1, where 0 < a < b < 00. There are three stages to the
propagation of the initial displacement, as illustrated in Figure 10.
In the first stage the initial pulse separates into two pulses, each with half the
amplitude of f(x), that propagate to the left and right with speed c. In the second stage
the pulse !fo(x - ct) moves to the right and the pulse !f(x - ct) appears on x ~ O. A
common description is "the left-moving pulse is reflected at x = 0 to produce a reflected
wave." In the final stage the reflection is complete and two pulses propagate to the right.
The right-moving wave !fo(x - ct) consists of a leading pulse !l(x - ct) followed by its
reflected wave.
Solution via Separation of Variables (g = 0). The product solutions that satisfy
(24)-(26) and (28) with g = 0 are readily seen to be

(40) u(x, t, f.L) = sin(f.Lx) cos(f.Lct), f.L ~ O.

Thus superposition gives the solution

(41)

The coefficient B(f.L) is determined by the initial condition

(42) u(x,O) = f(x) = ['0 B(f.L) sin(f.Lx) df.L, x ~ o.

 9.2 VIBRATION OF STRINGS AND TRAVELING WAVES 589

It was shown in Section 8.14 that this has the unique solution

(43) B(jJ,) = - 21
IT 0
00
f(~) sin(jJ,~) d~;

that is, B is the Fourier sine integral of f(x). The solution (41), (43) requires two
integrations. To see the identity of this solution with d'Alembert's solution (39), note that
the Fourier sine integral (42), (43) is meaningful for all x and gives the odd extension of
f(x):

(44) fo(x) = 1 00
B(jJ,) sin(jJ,x) djJ, for all x.

To write the solution (41) in traveling wave form, we make use of (44) and the
trigonometric identity

2 sin(jJ,x) cos(jJ,ct) = sin jJ,(x - ct) + sin jJ,(x + ct).

Combining this with (41) gives

u(x, t) = - 11
2 0
00
B(jJ,) [sin jJ,(x - ct) + sin jJ,(x + ct)] djJ,.

By (44) this gives

1 1
u(x, t) = "2/ (x -
o ct) + "2/o (x + ct).
Finally, sincejo(x) = j(x) for x:::: 0, we see that
1 1
u(x, t) = "2/o(x - et) + "2f (x + et),
as in (39) above.
Neumann Boundary Condition. If u(O, t) = 0, the Dirichlet boundary condition, is
replaced by the Neumann boundary condition uxCO, t) = 0, it is readily seen that the
d'Alembert solution (39) of the semi-infinite string problem is replaced by
1 1
(45) u(x, t) = ie(X - ct) + "2f (x + ct),
wherefe(x) is the even extension off(x) to -00 < x < 00. It follows that

1 I 1 I
ux(x, t) = "2fe (x - ct) + "2f (x + ct),
and hence
1 1
ux(O, t) = -f: (-ct) + -f' (ct) = 0,
2 2
590 9 BOUNDARY VALUE PROBLEMS OF MATHEMATICAL PHYSICS

becausej:(x) is the odd extension ofJ'(x) to -00 <x< 00.

Problem 9.2.3. Vibrating String of Finite Length-Ends Fixed. This problem was
formulated as a boundary value problem in Section 7.1. The string displacement is
characterized by the following five conditions.

(46) u(x, t) is defined for 0 S x S Land t ::: 0,

(47) Ut!= cu 2 xx for o s x S Land t ::: 0,
(48) u(O, t) = 0 and u(L, t) = 0 for t ::: 0,
(49) u(x,O) = j(x) for Os x S L,
(50) Ut(x,O) = g(x) for Os x S L,

where j(x) and g(x) are prescribed initial displacement and speed, respectively. The
uniqueness of the solution was proved by the energy method in Section 7.1.

The solution of this boundary value problem can be constructed by the d'Alembert
traveling wave method, but the details are lengthy. We shall determine this solution by
an indirect method, beginning with the separation of variables method. The latter can
be simplified by noting that if u(f,g) is the solution of (46)-(50) then the linearity of
conditions (46)-(50) implies that

(51) u(f,g) = u(f,O) + u(O,g).

Thus it will be enough to calculate u(f,O) and u(O,g) separately and then add them to get
the general solution.
Construction of u(f,O) by Separation of Variables. We seek first the product functions
that satisfy (46)-(48) and (50) with g = O. These are seen to be

(52) sin
n1lX) cos (mrct)
(L L ' n=I,2,3, ...

and their scalar multiples. Superposition then gives the trial solutions

u(f,O) (x, t) ~ Bn sin (mrx)

= f:t L cos (mrct)
L

where the Bn are to be chosen to satisfy (49):

(53) j(x) = LBn sin

00 (n1lX)
L .
n=l
 9.2 VIBRATION OF STRINGS AND TRAVELING WAVES 591

From Chapter 8 we know that the correct choice is the Fourier sine series coefficients
for j(x):

(54) Bn == bnC}) = ~ lLf(~) sin (n~~) d~, n = 1,2, ....

Thus the solution is given by (54) and

(55) u({,o)(x,t) = ~bnC})sin C~X) cos (n~ct).

Summation of the Series. We shall use the convergence theory of Chapter 8 to sum
the infinite series in (55). First, the identity

. (nrrx)
sm - cos (nrrct)
- - = -1.sm (nrr(x-ct)) + -1.sm (nrrex+ct))
L L 2 L 2 L
implies that

u({,Ol(x, t)-~~b(j)'sm (nrre - )) + ~~bej)'

(56)
x ct (nrr(x+ct)) ,
- ~ n ~ n sm
2 n=] L 2 n=] L
where bnC}) is defined by (54). Note that Dirichlet's convergence theorem for Fourier
sine series gives

L bnC}) sin C~X),

00

(57) f(x) = 0:::: x:::: L.

n=]

Moreover, the series in (57) converges for all real x. Indeed,

sin ( - n~x) = _ sin (n~x) (odd function)

and
nrr(x+2L))
sin ( L = sin (nrrx
L + 2rrn
)
= sin (nrrx)
L ;
i.e., the functions sin(nrrxlL) have a common period of 2L. It follows that

L bnC}) sin C~X) = rex)

oc

(58) for - 00 < x < 00,

n=1

where

(59) rex) = f(x) for 0:::: x:::: L,

r(-x) = -rex) [orall x,
r(x + 2L) = rex) for all x.
592 9 BOUNDARY VALUE PROBLEMS OF MATHEMATICAL PHYSICS

These conditions characterize f*(x) as the (unique) odd periodic extension of f(x) with
period 2L. With this definition of f* (x), equation (56) is equivalent to

1 1
(60) u(f,O) (x, t) = "2J*(x - ct) + "2J*(x + ct).

Conversely, if u(f,O) is defined by (60), where 1* is defined by (59), then it is easy to check
that u(f,O) is the solution of (46)-(51) with g = O.
Construction of u(O,g). In this case we seek the product solutions that satisfy (46)-(49)
with f(x) = O. These are the scalar multiples of

(61) sin (Lmrx) sin (nnct)

L ' n = 1,2, ....

Superposition gives the trial solution

(62) u(O,g)(x,t) = ~Bnsin

00 I (nnx)
L sin (nnct)
L

and

(63) (O,g)
Ut (x,t ) -_ ~
~BnI (nnc)
- . (nnx)
Sin - COS
(nnct)
-- .
n=] L L L

The initial conditions (50) give

(64) (0 )
U t ,g (x, t) = g(x) = L 00
BnI (nnc)
- sin (nnx)
- ,
n=] L L

which implies

nnc) 2 [L (nn~)
(65) B~ ( L = bn(g) = L 10 gmsin L ~.

We may sum the series in (63) via (65) and Dirichlet's convergence theorem to get

(O,g) ) 1~b ( ) . (nn(x - ct))

Ut (x, t = - ~ n g sm
2 n=] L

+ -1 ~b
~ n
(g) sm
. (nn(x+ct))
2 n=] L

= "21 g* ex - ct) +"21 g*(x + ct) ,

 9.2 VIBRATION OF STRINGS AND TRAVELING WAVES 593

where again g*(x) is the odd periodic extension of g with period 2L. To get u(O,g) itself,
we may integrate to get

u(O,g) (x, t) = it u~o,g) (x, r) dr.

Thus

u(O,g) (x, t) = ~ t g*(x - cr) dr + ~ t g*(x + cr) dr

210 210

= -
1 lx+ct g*(~)d~ - -
1 lX-ct g*(~) ~
2c x 2c x

= -2
1 lx+ct g*(~)~.
c x-ct
Finally, adding u(f,O) and u(O,g) gives the general solution

u(x, t) = -21 [[*(x - ct) + j*(x + ct)] + -2c1 lx+ct

x-ct
g*(~)~,

where f* and g* are the odd periodic extensions off and g with period 2L.
Remarks on String Instruments in Music. The vibrating string of finite length
with fixed endpoints is a model for the vibrating strings of musical instruments such as
violins and pianos. The product solution

un(x,t) nnx)
= sin (L [Bn cos (n::nct)
L +Bn, sin (n::nct)]
L '
with parameters Land c = "jf7p, is called the nth characteristic function (or the nth
normal mode) of the string; see Figure 11. It represents a synchronous vibration with
frequency

Vn = C~c) (2~) = ;~ = 2~ If hertz.

They are all integer multiples of the fundamental frequency

v, = ;LIf
In music the higher multiples of VI are called overtones. Thus

V2 = 2 VI = octave above VI,

3
V3 = 3v] = -V2 = fifth above V2,
2
594 9 BOUNDARY VALUE PROBLEMS OF MATHEMATICAL PHYSICS

x
n=l n=2 n=3

FIGURE 11. First three normal modes-snapshots at t = 0

V4 = 4VI = 2V2 = octave above V2,

5
Vs = 5VI = -V4 = major third above V4,
4

The occurrence of these simple fractional ratios is the basis for the harmonious tones of
stringed instruments.
Periodicity of Vibrations of Strings. The solution formula (60) for u(j,O) gives
1 1
u(j,O\x, t + 2Uc) = "'21*(x - ct - 2L) + "'21*(x + ct + 2L)
1 1
= "'2f*(x - ct) + "'2f*(x + ct)
= u(j,O) (x, t),
since1* has period 2L. Thus for any initial displacement u(x, 0) = f(x) the corresponding
motion has a fixed period

2L =2L fE
r=
c VI
and frequency

VI = ~r = ~ fi
2Ly"-P hertz.

Thus every motion of the string is a musical note of frequency VI. The method of tuning
a string is to adjust the tension T until VI has the desired value (for example VI = 440
for middle A).

Problem 9.2.4. Finite Strings with Other Boundary Conditions. Changing the boundary
conditions has a profound effect on the vibrations of a taut string. As an example we
 9.2 VIBRATION OF STRINGS AND TRAVELING WAVES 595

shall discuss the boundary value problem defined by the following conditions.

(66) u(x, t) is defined for 0 ::;: x ::;: 1 and t 2: 0,

(67) Utt= c2uxx for o ::;: x ::;: 1 and t 2: 0,
(68) ux(O, t) = 0 and ux(l, t) + u(l, t) = 0 for t 2: 0,
(69) Ut(x,O) = 0 for O::;:x::;: 1,
(70) u(x,O) = j(x) for O::;:x::;:l,

where j(x) is a prescribed initial displacement. Separation of variables leads to the same
Sturm-Liouville problem as in Problem 9.1.3. Thus the product solutions of (66)-(69)
are the scalar multiples of

(71) cos(f.-Lnx) cos(f.-Lnct), n = 1,2,3, ... ,

where the numbers f.-Ln are the positive solutions of

(72)

Thus superposition leads to solutions of the form

L
00

(73) u(x, t) = Cn cos(f.-Lnx ) cos(f.-Ln ct)

n=1
with

=L
00

(74) j(x) Cn COS(/LnX ).

n=1
Sturm-Liouville theory, as developed in Chapter 8, guarantees the orthogonality of the
eigenfunctions cos(f.-Lnx). Thus the coefficients cn can be computed from

f01 j(x) COS(f.-LnX) dx

(75) Cn = 1 ' n = 1,2, . ...
fo cos 2 (f.-LnX) dx
Of course, d'Alembert's theorem guarantees that u(x, t) is a sum of two traveling waves.
To find them we use the identity

2 cos(a) cos (b) = cos (a - b) + cos(a + b)

to write

2: L Cn cos f.-Ln(x - + 2: L Cn cos f.-Ln(x + ct)

1 00 1 00
(76) u(x, t) = ct)
n=1 n=1
1 1
= 2:f*(x - ct) + i*(X + ct),
596 9 BOUNDARY VALUE PROBLEMS OF MATHEMATICAL PHYSICS

where
00

(77) j*(x) = Len COS(fl,nX), -00 < X < 00.

n=1

This sum converges to j(x) for 0 ~ x ~ 1 by Sturm-Liouville theory However, it is

not clear without further analysis that it will converge for other values of x. A rigorous
analysis of the question is beyond the scope of this text.

Traveling Waves on an Infinite String. Solve for u(x, t) by applying d'Alembert's formula, then graph
three snapshots of the solution on -5 ~ x ~ 5 at t = 0, t = 0.5 and t = 1. Assume c = 1.

1. Plucked string Problem 9.2.1,f(x) = I - Ixl on Ixl ~ l,f(x) = 0 elsewhere, g(x) == O.

2. Plucked string Problem 9.2.1,f(x) = I sin(Jlx)I on Ixl ~ l,f(x) = 0 elsewhere, g(x) == O.
Traveling Waves on a Semi-Infinite String. Find explicit times to, tl and t2 when the solution u(x, t) of
the traveling wave Problem 9.2.2 exhibits the three stages of propagation. Assume c = 1.

3. Problem 9.2.2, g(x) == O,f(x) = sin(rrx) for 2 ~ x ~ 3,f(x) = 0 elsewhere.

4. Problem 9.2.2, g(x) == O,f(x) = sin 2 (rrx) for 2 ~ x ~ 3,f(x) = 0 elsewhere.

Vibrating String of Finite Length. In Problem 9.2.3, let L = rr and assume g(x) == O. Find explicitly the
solution u(x, t).

5. f(x) = sinx.

6. f(x) = sin x - sin 2x.

7. (Speed of a Traveling Wave) Let </>(x) be any function with a second derivative and define u(x, t) =
</>(x - ct), where c > 0 is a constant. Let d(t) be the positive distance that the graph of rp(x) is translated
in order to coincide with the graph of u(x, t). Show that the translation is to the right and d(1) = ct, and
hence the speed of a traveling wave is d' (I) = c.

8. (Change of Variables) Let u(x, t) be a given twice-differentiable function. Introduce new coordinates
(r, s) by

r =x - ct, s = x + ct
and define vCr, s) = u«r + 5)12, (5 - r)l(2c)). Apply the chain rule to establish the identity

9. (D'Alembert's Solution) Given f, I' and g continuous, let rp and 1/1 be continuously differentiable
functions that satisfy the equations

f(x) = rp(x) + 1/I(X),

2rp'm = I'(~) - g(~)/c,
21/1'm = I'm + g(~)/c.
 9.2 VIBRATION OF STRINGS AND TRAVELING WAVES 597

Integrate two of these equations from ~ = 0 to ~ = x to obtain equations for t/> and 1/1, and then substitute
into the equation u(x, t) = t/>(x - ct) + 1/I(x + ct) to obtain d'Alembert's solution

u(x, t) = -1 [r(x - ct)

2
+f(x + ct)] + -1 l
2c x-ct
x+ct
g(~)~.

10. (D'Alembert Solution) Letf,f' and g be continuously differentiable and define

u(x, t) = -1 [r(x - ct)

2
+f(x + ct)] + -1 l
2c x-ct
x +ct
g(~)~.

Show that Utt = c2uxx, u(x,O) = f(x) and Ut(x, 0) = g(x).

The Plucked String. A plucked string lies along the interval 0 ~ x ~ 1. It has wave speed c = 1,
initial velocity zero and initial shape

{
2X 0 ~ x ~ 1/2,
f(x) = 2(I-x) 1/2 ~ x ~ 1.

Then f has an odd periodic extension h of period 2 and the exact solution y can be written in terms of h:

1
Y= "2 [hex - tH hex + t)j.

11. Graph the solution surface yon 0 ~ x ~ 1, 0 ~ t ~ 1 and also the "snapshots" of y at times t = 0,
t = 0.2, t = 0.4, t = 0.6, t = 0.8, t = 1.

12. Determine the first five terms of the series representation of the solution

L Bn sin(mrx) cos(mrt).
00
y(x, t) =
n=l

Let z be the partial sum of the first five terms. Graph the solution surface z on 0 ~ x ::: 1, 0 ::: t ::: 1
and also the "snapshots" of z at times t = 0, t = 0.2, t = 0.4, t = 0.6, t = 0.8, t = 1.
The Semi-Infinite String. Let fex) be a prescribed function with f twice continuously differentiable and
consider the problem

U(x, t) is defined for 0 ::: x < 00, t ~ 0,

Utt = c2uxx for 0::: x < 00, t ~ 0,
u(O,t) = 0 for all t ~ 0,

°
u(x,O) = f(x) for 0 ~ x < 00,
Ut(x,O) = for 0 ~ x < 00.

The exercises below refer to this problem as the semi-infinite string problem, even though it is a special case.

13. Show that the product solutions for the semi-infinite string problem are

U(x, t, JL) = sin(JLx) cos(JLct), JL ~ o.

14. Argue heuristically that superposition of the product solutions of the previous exercise gives the trial
solution
598 9 BOUNDARY VALUE PROBLEMS OF MATHEMATICAL PHYSICS

15. (Even Extension) Assume that f is twice continuously differentiable on x ;:,: O. Let fe(x) be the even
extension off(x) to -00 < X < 00 and define

1 1
u(x, t) = "2fe(x - ct) + "2 f (x + ct).

Show that U is a solution of the wave equation Utt = c2uxx on x ;:,: 0, t ;:,: 0 with wave shape
u(x,O) = f(x) and wave velocity Ut(x, 0) = o.

9.3 Steady-State Diffusion of Heat in Plates

In Section 7.4 the problem of steady-state heat diffusion in plates was modeled as a
boundary value problem for the Laplace equation Uxx + Uyy = 0 in regions Q of the
Cartesian plane, and the uniqueness of the solutions was proved. For most plate shapes
Q this problem can be solved only by numerical methods. However, for some regions of
simple shape the problem can be solved by separation of variables. In this section several
problems of this type are analyzed.

Problem 9.3.1. Rectangular Plates, Dirichlet Boundary Condition. This problem asks
for the temperature distribution u(x,y) that satisfies Laplaces equation in a rectangle and
has prescribed values on its four edges. Before solving the general case, we shall illustrate
the method by the case of a square plate of side Jr with zero boundary conditions on
three of the four edges. Thus u(x, y) must satisfy

(1) u (x, y) is defined for 0 ::: x ::: Jr, 0 ::: y ::: Jr,
(2) Uxx + Uyy = 0 for 0::: x ::: Jr, 0 ::: y ::: Jr,
(3) u(O,y) = 0 and u(Jr,y) = 0 for 0::: y ::: Jr,
(4) u(x, 0) = 0 for 0::: x ::: Jr,
(5) u(x, Jr) = j(x) for 0::: x ::: Jr,

where j(x) is a preSCribed function.

We shall solve this problem by the separation of variables method-determination

of the product solutions of (1)-(4), followed by superposition and determination of
coefficients by an orthogonality condition. Only the main steps of this now familiar
method will be given.
Product Solutions. As usual, we seek the product solutions

(6) U(x,y) = X(x)Y(y)

 9.3 STEADy-STATE DIFFUSION OF HEAT IN PLATES 599

that satisfy all the linear homogeneous conditions of the problem. These are (1)-(4) in
the present case. Substituting (6) into (2) and separating variables gives

Y"(y) XI/(x)
--=---=A
Y(y) X(x) ,

where).. is a separation constant. We satisfy (3) by X(O) = X(Jr) = 0 and (4) by yeo) = 0.
Thus X is a solution of the familiar Sturm-Liouville problem

x" + AX = 0 for O:s x :s Jr,

X(O) = 0, X(Jr) = 0,

with solutions

An = n2 ,
Xn(X) = sin(nx), n = 1,2,3, ....
The corresponding Y -factors, denoted by Yn , satisfy

Y;: - n2Yn = 0 for O:s y :s Jr,

Yn(O) = O.

Finally, the product solutions are precisely the scalar multiples of

un(x,y) = sin(nx) sinh(ny) , n = 1,2,3, ....

Superposition. This gives solutions of the form

I.>n sin(nx) sinh (ny) ,

00

(7) u(x,y) =
n=1

where the coefficients Cn are chosen such that the series converges. We wish to choose
the cn's such that (7) satisfies (5); that is, for 0 :s x :s Jr,

L
oc

u(x, Jr) = j(x) = Cn sinh(nJr) sin(nx).

n=]

This is the Fourier series expansion of j(x) if

Cnsinh(nJr) = bn(j) = -2111: j(~) sin(n~)~,

Jr 0
so
bn(j)
(8) Cn = sinh (nJr) ' n = 1,2,3, . ...
600 9 BOUNDARY VALUE PROBLEMS OF MATHEMATICAL PHYSICS

Finally, substituting (8) into (7) gives the solution

~ . sinh(ny)
(9) u(x,y)=L.,..bn(j)sm(nx). h( ).
n=1 SIn mr

Convergence. Equation (9) is only a formal solution of the boundary value problem
because the convergence of the series has not been shown. To analyze this we note that
the quotient of hyperbolic sine functions satisfies

Sinh(ny ) 1< 1 ~y ~ Jr.

Isinh (nJr) -
for 0

Hence the series in (9) converges at least as well as the Fourier sine series for j(x). Criteria
for the convergence of Fourier sine series were given in Section 8.3. Note also that
sinh(ny) e ny e-ny
- -n(1(-y)
----~e for n -+ 00.
sinh(nJr) e n1( - e-nrr

Thus the series (9) converges very rapidly when 0 ~ y < Jr. The slowest convergence
occurs when y = Jr.
The General Case. Now consider the steady-state temperature u(x,y) in a rectangular
plate of arbitrary dimensions Land M units, respectively. If coordinates are chosen as in
Figure 12,
then u(x,y) is characterized by the boundary value problem

FIGURE 12. Steady-state heat

(0,0) u= 11 (L,O) diffusion in a rectangular plate
 9.3 STEADy-STATE DIFFUSION OF HEAT IN PLATES 601

(14) u(O,y) = gl (y) for 0:::: y :::: M,

(15) u(L,y) = g2(y) for 0:::: y :::: M,

where 11 (x) ,h (x), gl (y) and g2 (y) are prescribed functions. Conditions (12)-(15), taken
together, specify u on the complete boundary of the rectangle.
Separation of variables is not directly applicable to the boundary value problem
(10)-(15) when all the functionsl1,h, gL g2 are nonzero. However, if

(16) v(x,y) is the solution for 11,h given and gl = g2 = 0, and

(17) w(x,y) is the solution for gl, g2 given andll = h = 0,

then the solution of the general case (10)-(15) is given by the sum (check this!)

(18) u(x,y) = v(x,y) + w(x,y).

Thus it will be enough to construct v and w. We will show that this can be done by
separation of variables.
Construction of vex, y). The function vex, y) is characterized by the conditions

(19) Vxx + Vyy = 0 for 0:::: x :::: L, 0 :::: y :::: M,

(20) v(O,y) = 0 and v(L,y) = 0 for 0:::: y :::: M,
(21) v(x, 0) =11 (x) and vex, M) = h (x) for 0:::: x :::: L,

where 11, h are given. To solve this by separation of variables we first construct the
product solutions of the homogeneous conditions (19), (20) only. Putting
(22) v(x,y) = X(x)¥(y)

into (19), (20) gives, by the usual argument,

¥I/(y) X"(x)
- - = - - - =A
¥(y) X(x) ,

with separation constant A, which leads to the well-known Sturm-Liouville problem

x" (x) + )"X(x) = 0 for 0:::: x :::: L,

X(O) =0 and X(L) = o.
The eigenpairs of this problem are

An = (mrILi,
Xn(x) = sin(mrxlL), n = 1,2,3, ....
The eigenfunctions Xn are orthogonal on the interval 0 :::: x :::: L. The corresponding
orthogonal expansion is the Fourier sine series for that interval.
602 9 BOUNOARY VALUE PROBLEMS OF MATHEMATICAL PHYSICS

The corresponding product solutions are

Vn(X,y) = sin(nJl'xlL)YnCy),
where Yn is the general solution of

(23)

This solution can be written as

or

YnCy) = al sinh(nJl'yIL) + a2 cosh(nJl'ylL).

However, there are more convenient choices. Let us write
D() sinh (nJl'yIL)
Y y
n
= sinh (nJl'MlL)
.
Then it is easy to check that Y~Cy) and Y~(M - y) are a solution basis (check this by the
Wronskian test). Hence the general solution of (23) can be written

where An and Bn are arbitrary constants. Notice that

y~(o) = 0, Y~(M) = 1,

and so

We will use these coefficients to satisfy the nonhomogeneous boundary conditions (21).
Construction of v (Completed). Superposition gives
00

(24) v(x,y) = LXn(X)YnCy)

n=l

This satisfies Laplace's equation and the two homogeneous boundary conditions. To
satisfy the two nonhomogeneous boundary conditions, we have
00 nJrX
v(x,O) = flex) = LBnsin(T) for 0:::: x:::: L
n=l
 9.3 STEADy-STATE DIFFUSION OF HEAT IN PLATES 603

and
IX) mrx
v(x,M) =f2(X) = "Ansin(-)
~ L for 0<
- x <
- L.
n=1

The orthogonality property of the eigenfunctions sin(mrxlL) gives

Bn = bn(fI) = ~ [L 11 (~) sin(mr~/L)~,

L 10
An = bn(h) = ~ [L h(~) sin(mr~/L)~.
L 10
This completes the construction of v. Note that we can write v as follows.

~ . (nnx) bn(h) sinh(nZY) + bn(/I) sinh(n1r(~-Y))

(25) v(x,y) = ~sm - .
n=1 L sinh(mrM/L)

From this equation, conditions (19)-(21) that characterize v are readily verified.
Construction of w(x,y). The function w(x,y) is characterized by the conditions

Wxx + Wyy = 0 for 0S x S L, 0 S Y S M,

W(x, 0) =0 and w(x,M) =0 for 0 S x S L,
w(O,y)=gl(y) and w(L,y)=g2(y) for OsySM.

These conditions are identical with conditions (19)-(21) for v(x,y) if we exchange x and
y, Land M and (h ,h) with (gl, g2). Thus w(x,y) may be obtained from (25) by making
these exchanges. This gives

~ . (nrry) bn (g2) sinh(n:) + bn (gl) sinh(n1r(~-x))

w(x,y) = ~sm -M . h(!!&) ,
n=l SIn M

where

bn(g) = ~ [M g(l']) sin (nrrI']IM) dl'].

M 10
This completes the construction of vex, y) and w(x, y) and hence also u(x, y) = vex, y) +
w(x,y). For numerical examples see the Exercises.

Problem 9.3.2. Rectangular Plates- Neumann and Mixed Boundary Conditions. These
problems were formulated and the uniqueness of their solutions was discussed in
Section 7.4. Here two examples are solved by the separation of variables method. More
general cases are discussed in the Exercises.
604 9 BOUNDARY VALUE PROBLEMS OF MATHEMATICAL PHYSICS

(O,n)

Ux = a

x
FIGURE 13. Square plate-Neumann
(0,0) Uy =0 (n,a) boundary condition

A Neumann Problem. Consider a square plate of side units, with three sides Jr
insulated (au/an = 0) and a prescribed heat flux (au/an = j) across the fourth. With an
obvious choice of coordinates, see Figure 13,
u(x,y) is characterized by the following boundary value problem.

(26) U(x,y) is defined for 0::::: x::::: Jr, 0::::: y ::::: Jr,

(27) Uxx + Uyy = 0 for 0::::: x ::::: Jr, 0 ::::: Y ::::: Jr,

(28) Ux(O,y) = 0 and ux(Jr,y) = 0 for 0.:::: y .:::: Jr,

(29) uy(x,O) = 0 for 0::::: x::::: Jr,

(30) uy(x,Jr)=j(x) for O:::::x:::::Jr,

where j(x) is a prescribed function. We shall solve this problem by separation of variables.
Product Solutions. The product solutions u(x,y) = X(x)Y(y) must satisfy conditions
(26)-(29). This gives the Sturm-Liouville problem

X" + AX = 0 for 0::::: x ::::: TC,

(31)
X'(O) =0 and X/(TC) =0
for X(x) , and the initial value problem

y" - AY = 0 for 0.:::: y ::::: Jr,

(32)
y'(O) = 0

for Y(y). Solving (31) in the usual way gives the eigenpairs

An = n2 ,
Xn(x) = cos(nx), n = 0,1,2,3, ....
 9.3 STEADy-STATE DIFFUSION OF HEAT IN PLATES 605

The corresponding function Y is

Yn(y) = cosh(ny).

Thus the product solutions are multiples of

(33) Un(x,y) = cos(nx) cosh(ny), n = 0,1,2,3, ....

Superposition. The set (33) generates solutions of the form

+L
00

(34) u(x, y) = Co Cn cos(nx) cosh(ny).

n=l

We know from Section 7.4 that the solution of the Neumann problem is unique only up
to an additive constant. Thus Co is arbitrary We shall use the final boundary condition,
equation (30), to determine the remaining constants Cn. The y-derivative of u is

L nC
00

uy(x,y) = n cos(nx) sinh(ny),

n=l
so

L nC sinh(n:rr) cos(nx).
00

(35) uyCx,:rr) = f(x) = n

n=l

This is a Fourier cosine series on the interval 0 :::: x :::: :rr. The Fourier cosine series
coefficients off(x) are defined by (see Section 8.4)

(36) ao(/) = - 11
:rr
7C
f(~) d~,

21
0

7C
an(/) = - f(~) cos(n~) d~, n= 1,2,3, ....
:rr 0

In equation (35) there is no constant term. Thus a solvability condition for f(x) is that
ao(/) = 0, or

The remaining coefficients Cn are determined by

nCn sinh(n:rr) = an (/), n = 1,2,3, ... ,

or
an(/)
Cn = .
nsinh(n:rr)
606 9 BOUNDARY VALUE PROBLEMS OF MATHEMATICAL PHYSICS

This gives the solution

~ cosh(ny)
(37) u(x,y) = Co + ~ an(/) cos(nx) . ( ).
n=l n smh mr

It is easy to show that if j(x) satisfies the conditions of Churchills theorem (see Section
9.1, Corollary 2), so that

then the series (37) for u, and the derived series for u y , converge absolutely and uniformly
for all x and y in the square 0 :::: x :::: Jr, 0 :::: y :::: Jr. Under these conditions (37) defines
a function u(x,y) that satisfies conditions (26)-(30) and is therefore a solution of the
Neumann problem.
A Mixed Problem. As a simple example of a mixed problem, we shall consider the
function u(x,y) characterized by the follOwing conditions.

(38) U(x,y) is defined for 0 :::: x:::: Jr, 0 :::: y :::: Jr,
(39) Uxx + Uyy = 0 for 0:::: x :::: Jr, °: : y :::: Jr,

(40) u(O,y) = 0 and u(Jr,y) = 0 for 0::::: y ::::: Jr,

(41) uyCx,O) = 0 for 0::::: x::::: Jr,

(42) Uy(x, Jr) = j(x) for 0:::: x ::::: Jr,

where j(x) is prescribed. Physically, u(x,y) may be interpreted as the steady-state

temperature in a square plate of side Jr if the sides at x = 0 and x = Jr are held at zero
temperature, the side at y = 0 is insulated, and the heat flux across the side at y = Jr is
a prescribed function j(x). The solution by separation of variables goes as follows.
Product Solutions. These are multiples of

(43) Un(x,y) = sin(nx) cosh(ny), n = 1,2,3, ....

Superposition. This set (43) generates solutions of the form

= L Cn sin(nx) cosh(ny)
00

(44) u(x,y)
n=l

and

=L
00

(45) uyCx,y) nCn sin(nx) sinh(ny).

n=l
 9.3 STEADy-STATE DIFFUSION OF HEAT IN PLATES 607

In particular, to satisfy condition (42) we have

L nC sinh(nrr) sin(nx).
00

(46) uyCx, rr) = j(x) = n

n=l

Comparing this with the Fourier sine series for j(x) gives the coefficients cn:

nC nsinh(nrr) = bn(j) = - 21"

rr 0
j(~) sin(n~)~.

Thus the solution of the mixed problem (38)-(42) is defined by the series

~ . cosh(ny)
(47) u(x,y) = L.,;bn(j)sm(nx) . h( ).
n=l nsm nrr

If j(x) satisfies the condition of Churchill's theorem then

and the series for u(x, y) and uyCx, y) converge uniformly and satisfy (38)-(42), as in the
preceding example.

Problem 9.3.3. Half-Plane with Dirichlet Boundary Condition. If the temperature

u(x, y) is needed near one edge of a large square plate and far from the other three edges,
then one may model the plate as the half-plane y ~ 0; see Figure 14.
The corresponding boundary value problem is to find a function u(x,y) with the
properties

(48) u(x,y) is defined for -00 < x < oo,y ~ 0,

(49) Uxx + Uyy = 0 for -00 < x < 00, y ~ 0,

x
u(x,O) = j(x) FIGURE 14. Dirichlet problem for a half-plane
608 9 BOUNDARY VALUE PROBLEMS OF MATHEMATICAL PHYSICS

(50) U(x,O) = j(x) for -00 <x< 00,

where j(x) is prescribed. To obtain uniqueness we shall add the conditions

(51) U, Ux and u y are bounded for -00 <x< 00, y ~ 0,

and

(52) f~oo [[(x) I dx < 00.

We shall solve this problem by separation of variables.

Product Solutions. The product solutions u(x,y) = X(x)Y(y) must satisfy (48), (49)
and (51). The functions X(x) and Y(y) are characterized by the properties

X" +)..X=O for - 00 < x < 00,

(53)
X(x) bounded for - 00 < x < 00,

and
y" -)..Y = 0 for 0::: y < 00,
(54)
Y(y) bounded for 0::: y < 00.

The now familiar arguments imply that the eigenpairs for (53) are

A=f.1,2, f.1,~O,
(55)
X(x) = A(f.1,) cos(f.1,x) + B(f.1,) sin (f.1,x) ,

while

(56)

Thus the product solutions have the form

where A(f.1,) and B(f.1,) are arbitrary functions of f.1,.

Superposition. Summing (that is, integrating) over the parameter f.1, ~ 0 gives solutions
of the form

(58) u(x,y) = 1 00
{A(f.1,) cOS(f.1,x) + B(f.1,) sin(f.1,x)} e- ILY df.1,.

Finally, to satisfy the boundary conditions (50) we must have

(59) u(x,O) = j(x) = 1 00

{A(f.1,) cOS(f.1,x) + B(f.1,) sin(f.1,x)} df.1,.
 9.3 STEADy-STATE DIFFUSION OF HEAT IN PLATES 609

This is the Fourier integral representation of j(x) and the Fourier integral theorem of
Section 8.9 gives the solution

(60) A(p.,) =;1 foo

-ooj(~) cos(p.,~)~, B(p.,) =;1 foo
-ooj(~) sin(p.,~) d~.

This completes the solution whenever j(x) satisfies the conditions of the Fourier integral
theorem.
The solution (58), (60) requires three integrations to evaluate u(x,y). We shall show
that two of these can be carried out, leaving a single integration. The analysis parallels
that of Problem 9.1. 7 in Section 9.1. Substituting the identity

A(p.,) cos(p.,x) + B(p.,) sin(p.,x) =;1 foo j(~) cos p.,(x -

-00 ~) d~

into (58) gives

U(x,y) = ~ 1 (i:j(~)
00
cos p.,(x - ~)d~) e- IlY dp.,.
Changing the order of integration (justified when (52) holds) gives

(1
i:
(61) u(x,y) = ~ i:j(~) 00
cosp.,(x - ~)e-llYdp.,) ~
= K(x - ~,y)j(~) dt
where the kernel K is defined by

(62) K(x,y) = - 11
Tr 0
00
cos(p.,x)e- IlY d~.

This integral may be calculated by recalling the Laplace transform (see Chapter 4)

10
[00 e-st cos(cvt) dt = A
s + cv
(s > 0).

The substitution s -+ y, t -+ p." cv -+ x gives

1 y
(63) K(x,y) = --2--2'
TrX +y
Thus the solution of the half-plane problem is delivered by the Single-integral formula

y foo j(~)
(64) u(x,y) =; -00 (x _ ~)2 + y2 dt y > O.

Note that this is valid only for y > O. For Y = 0 the formula yields the meaningless form
Ox 00.
610 9 BOUNDARY VALUE PROBLEMS OF MATHEMATICAL PHYSICS

The integral in (64) is called Poisson's integral. A second form of the solution that
is sometimes useful is obtained by substituting

(65) ~ = x + y tan(t), -JrI2 < t < JrI2,

into (64). The result, after a simple calculation, is

(66) u(x,y) = -1
Jr
1 7C/2

-7C12
j(x + y tan t) dt.

Uniqueness and Boundedness. Hypotheses (50) and (52) imply that

(67) If(x) I : : : M for - 00 < x < 00,

where M is a constant. It follows from (66) that lu(x,y)1 ::::: M for all x and all y :::: O.
Thus Poisson's integral assigns to each bounded boundary function j(x) a solution u(x, y)
with the same bound. We shall not prove the uniqueness of the bounded solution here.
However, note that uniqueness of the solution is lost if we do not require that u(x,y) be
bounded. For in that case we can add to the solution (66) any multiple of y and obtain
another (unbounded) solution.
A Simple Example. Define j(x) by

(68) j(x) = H(x + 1) - H(x - 1),

where H is the Heaviside function. Then the integrand in (66) is 1 when -1 < x +
y tan t < 1 and zero otherwise. Thus

u(x,y) = ~ [tan-1 ~ Cx) - tan- 1(-1y- X) ] '

or, using the fact that tan- 1 x is odd, we may write

(69) u(x,y) = ~ [tan-1(x; 1) _ tan- 1(X ~ 1)].

Throughout this calculation tan- 1 x is assumed to be the principal value of the function
tan-I, defined by -JrI2 < tan- 1 x < Jr12. It is easy to verify that (69) defines a solution
of (48)-(52) when j(x) is defined by (68). For additional examples see the Exercises.

Problem 9.3.4. Half-Plane with Neumann Boundary Condition. This problem is ob-
tained from the preceding one by replacing the Dirichlet boundary condition u(x, 0) =
j(x) by the Neumann boundary condition uyCx,O) = j(x). In this case the separation of
variables method leads to divergent integrals. However, note that the derivative

(70) v(x,y) = uy(x,y)

 9.3 STEADy-STATE DIFFUSION OF HEAT IN PLATES 611

satisfies all the conditions of the Dirichlet problem, whence

(71) v(x,y) = u/x,y) = ~

y 1 00

-00
f(~)
(x _ ~)2 + y2 ~.
Integrating with respect to y gives, after an elementary integration,

(72) u(x,y) = ~
1 1 00
-00 log [(x -~) 2 + l] 112 f(~) ~ + K,

where K is a constant. It can be shown that (72) defines a solution of the Neumann

i:
problem if

~(~) log I~II ~ < 00.

The appearance of the constant K in (72) is not surprising, since the solution of the
Neumann problem is determined only up to an additive constant.

Problem 9.3.5. Infinite Strip-Dirichlet Boundary Condition. This problem models the
case of a long narrow rectangular plate when the steady-state temperature u(x, y) is sought
at points (x,y) far from the two narrow sides. As an example, the strip -00 < x < 00,
o : : y :::: 1, will be discussed; see Figure 15.
We shall consider the case where u(x, 0) = 0 and u(x, 1) =f(x) for -00 < x < 00.
Thus u(x,y) is characterized by the conditions

(73) U(x,y) is defined for -00 < x < 00,0::: y ::: 1,

(74) U xx + Uyy = 0 for -00 < x < 00,0 ::: y ::: 1,
(75) U(x, 0) =0 for -00 <x< 00,

(76) u(x, 1) = f(x) for -00 < x < 00,

u= jex) y=l

FIGURE IS. Infinite strip-Dirichlet boundary

u=o condition
612 9 BOUNDARY VALUE PROBLEMS OF MATHEMATICAL PHYSICS

where j(x) is a prescribed function. To obtain uniqueness we shall add the conditions

(77) U, U x , uy are bounded for -00 < x < 00,0 :::s Y :::s 1,
(78) J~oo [[(x) I dx < 00.

The separation of variables method will be used.

Product Solutions. The product solutions u(x,y) = X(x)Y(y) must satisfy (73)-(75)
and (77). This leads to
X" +).,X = 0 for - 00 <x< 00,
(79)
X(x) bounded,

and
y" -).,y =0 for O:::s y:::s 1,
(80)
yeO) = o.
The product solutions have the form

(81)

where jJ., ::: 0 and sinh jJ., is a convenient normalizing factor.

Superposition. Integrating over jJ., 2: 0 gives solutions

(82) u(x,y) = 1 o
00
{A(jJ.,)cosjJ.,x+B(jJ.,)sinjJ.,x}
sinh(jJ.,Y)
'h
sm jJ.,
djJ.,.

The boundary condition (76) requires that for -00 < x < 00,

u(x,1) = j(x) = 1 00
{A(jJ.,) cos jJ.,X + B(jJ.,) sin jJ.,x} djJ.,.
As before, this is the Fourier integral ofj(x), and we must have

(83)

Equations (82) and (83) define the solution whenever j(x) satisfies the conditions of the
Fourier integral theorem (see Section 8.9). The method of Problem 9.3.3 leads to an
integral formula

(84)

where
u(x,y) = i: K(x - ~,y)j(~) d~,

K(x,y) =-
1 foo cos(jJ.,x).
sinh (jJ.,y)
h djJ.,.
n -00 sm jJ.,
 9.3 STEADy-STATE DIFFUSION OF HEAT IN PLATES 613

However, evaluation of this integral is beyond the scope of this text. Infinite strip problems
with other boundary conditions are left for the Exercises.

Problem 9.3.6. Semi-Infinite Strips with Dirichlet Boundary Condition. This problem
models the case of a long narrow rectangle when the temperature u(x,y) is sought near
one of the narrow sides of the rectangle. With suitable choice of the unit oflength we may
assume that the strip is defined by the inequalities 0 ~ x < 00, 0 ~ y ~ Jr as in Figure 16.
The boundary value problem suggested by the figure asks for a function u(x, y) with
the follOwing properties.

(85) u(x,y) is defined for 0 ~ x < 00, 0 ~ y ~ Jr,

(86) Uxx + Uyy = 0 for o ~ x < 00, 0 ~ y ~ Jr,
(87) u(O,y) = fey) for o ~ y ~ Jr,
(88) U(x, Jr) = g(x) for o ~ x < 00,
(89) u(x,O) = hex) for o ~ x < 00,
where f(Y), g(x) and hex) are prescribed functions. To these we add the conditions

(90) u, Ux , u y are bounded for 0 ~ x < 00, 0~y ~ Jr,

(91) 10 00
Ig(x) I dx < 00, 10 00
Ih(x)1 dx < 00.

The Dirichlet problem (85)-(91) cannot be solved directly by separation of variables.

However, if uCf,g,h) denotes the solution of (85)-(91), then the linearity of the conditions
implies that

(92) uCf,g.h) = uCf,O,O) + uCO,g,O) + UCO,Q,h),

U(X, n) = g(x)
(O,n)

U(O,y) = fey)

(0,0) FIGURE 16. Semi-infinite strip

U(x,O) = hex) with Dirichlet boundary condition
614 9 BOUNDARY VALUE PROBLEMS OF MATHEMATICAL PHYSICS

in analogy with the case of a finite rectangle (see Problem 9.3.1 above). We shall show that
v = u(f,o,O), w = uCO,g,O) and z = uCO,O,h) can all be constructed by separation of variables.
Construction of v = u(f,o,O). The function v is characterized by the conditions

(93) v(x,y) is defined for 0 :::: x < 00,0:::: y :::: Jr,

(94) Vxx + Vyy = 0 for 0:::: x < 00,0:::: y :::: Jr,

(95) v(x,O) =0 and vex, Jr) =0 for 0:::: x < 00,

(96) v(O,y) = I(y) for 0:::: y :::: Jr,

(97) v, VX , Vy are bounded for 0:::: x < 00,0:::: y :::: Jr.

Product Solutions. The product solutions of (93)-(95) and (97) are the multiples of

vn(x,y) = e- nx sin(ny) for n = 1,2,3, ....

Superposition. This leads to solutions of the form

00

(98) v(x,y) = L cne- nx sin(ny).

n=1

The coefficients are to be chosen so as to satisfy the final boundary condition (96). Thus
00

v(O,y) = I(Y) = L Cn sin(ny) for 0:::: y :::: Jr.

n=1

This is the Fourier sine series representation of I(y) and is given by the sine series
coefficients

(99) n = 1,2,3, ....

The function v defined by

00

(100) v(x,y) = Lbn(f)e-nx sin(ny)

n=1

is the desired solution if I(y) satisfies the conditions of Churchill's theorem (see Section
9.1).
Construction of w = u CO,g,O) • The defining conditions for ware

(101) w(x,y) is defined for 0 :::: x < 00,0:::: y :::: Jr,

(102) Wxx + Wyy = 0 for 0:::: x < 00,0 :::: y :::: Jr,

(103) w(O,y) = 0 for 0:::: y :::: Jr,

(104) w(x,O) = 0 for 0:::: x < 00,

 9.3 STEADy-STATE DIFFUSION OF HEAT IN PLATES 615

(lOS) w(x, n) = g(x) for 0 ~ x< 00,

(l06) W, Wx , Wy are bounded for 0 ~ x < 00, 0~Y~ n,

(107) fo
oo Ig(x)ldx < 00.

This can be solved by separation of variables, following the plan of Problem 5 above.
However, the results for that problem can be used much more directly. To see this, let
go(x) be the odd extension of g(x) and let u(x,y) be the solution of the infinite strip
problem (see Problem 5 above) in the strip 0 ~ x< 00, 0 ~ y ~ n. The result is

(l08) u(x,y) = 1 00

o
{A(JL) cOSJLX +B(JL)sinJLx}
sinh (JLY)
. h
sm JLn
dJL,

where

However, cos(JL~) and sin(JL~) are even and odd, respectively, and hence A(JL) = 0 and

(1l0)

It follows that

(Ill) W(X,y) = 1
o
00 sinh(JLY)
B(JL) sin(JLx) . h
sm JLK
dJL,

with B(JL) defined by (110), is the desired solution. Properties (101), (102) and (04)-
(06) hold because (108), (l09) solve the infinite strip problem. The final condition,
equation (03), is obvious from equation (Ill).
Construction of z = u(O,O,h). The case parallels that of wand is left for the Exercises.

Problem 9.3.7. Circular Disk-Dirichlet Boundary Condition. This problem asks for
the steady-state temperature distribution u(x, y) that satisfies the Laplace equation in a
circular disk and has prescribed values on the circumference. Thus if the disk has radius
a then u(x,y) must satisfy

(1l2) uxx + uyy = 0 for x 2 + y2 _< a2 ,

(1l3) u(x,y) = g(x,y) for x 2 + y2 = a2 ,
where g is a prescribed function on the circumference.

Use of Polar Coordinates. It is natural to introduce polar coordinates (r, e), defined by

(1l4) x = rcose, y = rsine,

616 9 BOUNDARY VALUE PROBLEMS OF MATHEMATICAL PHYSICS

to take advantage of the circular symmetry of the disk. We shall write

(llS) v(r,8) = u(x,y),

where x and yare given by (114). A straightforward, but lengthy, application of the chain
rule of calculus then gives

(1l6)

Hence if we write

(1l7) f(8) = g(a cos 8, a sin 8)

then the boundary value problem (12), (13) becomes

1 1
(llS) Vrr+ -Vr
r
+ 2vee
r
= 0 for 0 < r S a and all 8,
(1l9) v(a,8) = fee) for all e,

e.
where f(8) is a prescribed function of There are other, implicit, conditions that are
needed to calculate vCr, 8). First, by (117) ,f(8) must have period 27l', as must vCr, e):

(120) v(r,e + 27l') = v(r, e) for all e,

(121) f(8 + 27l') = fee) for all e,

because of the geometric interpretation of e; see Figure 17.

Finally,

(122) v(O, e) must be finite and independent of e,

since v(o, e) = u(O,O) is a constant.

y
___--,--~u(a, e) =fee)

(X,y) = (rease, r sin e)

FIGURE 17. Circular disk-Dirichlet

problem
 9.3 STEADy-STATE DIFFUSION OF HEAT IN PLATES 617

Product Solutions. We look for product solutions of (ll8),

(123) v(r,O) = R(r)8(e),

that satisfy (120)-(122). Substituting (123) into (ll8) gives

1 1
R"8 + -R'8
r
+ -R8"
r2
= 0,

or

2R"(r) R'(r) 8"(e)

(124) r --+r--=---="A
R(r) R(r) 8(0) ,

where "A is a separation constant. For R(r) we get the conditions

(125)
r 2R"(r) + rR'(r) - "AR(r) = 0 for 0::::: r < a,
R(r) finite for 0::::: r ::::: a.

For 8(0), equation (124) and the periodicity requirement (120) imply

(126)
8"(0) + A8(0) = 0 for 0::::: 0 ::::: 2ll",
8(2ll") = 8(0) and 8 ' (2ll") = 8'(0).

This is just the Sturm-Liouville problem of Section 8.5 with periodic boundary
conditions. The eigenpairs are

(127)
Ao = 0, 8 0 (0) = Ao,
An = n 2 , 8 nCO) = An cos(ne) + Bn sin(nO), n = 1,2,3, ....

Turning to the radial factor R(r) defined by (125) with "A = n2 , we have an Euler equation

(128)

It is solved by Eulers substitution R(r) = ra. The indicial equation for a is a 2 - n2 = 0,

618 9 BOUNDARY VALUE PROBLEMS OF MATHEMATICAL PHYSICS

or Ol = ±n. Thus

n = 0 gives Ro(r) = Co + Do log r,

n = 1,2, ... gives Rn (r) = Cnrn + Dnr-n.
The finiteness condition of (125) requires that Dn = 0, n = 0,1,2, .. " Thus the product
solutions of (l18) that satisfy (120)-(122) are

(129) vo(r, f)) = Ao,

vn(r, f)) = rn (An cos nf) + Bn sin nf)) , n = 1,2,3, ... ,

where the An and Bn are arbitrary constants.

Superposition. A solution of the boundary value problem (118)-(122) is sought as a
sum

L rn (An cos nf) + Bn sin nf)) .

00

(130) vCr, f)) = Ao +

n=1

This formally satisfies all the conditions except the boundary condition (119). For (19)
we have

= f(f)) = Ao + L (anAn cos nf) + anBn sin nf)) .

00

yea, f))
n=1

The last equation must be the Fourier series for f(f)), and hence the coefficients An, Bn
must be determined as follows.

(131) n= 0, Ao = ao(j) = ~ (2n f(41) d41,

2n 10
n:::: 1, anAn = an(j) = ~ (2n: f(41) cos n41 d41,
IT 10
n:::: 1, anBn = bn(j) = ~ tn: f(41) sin n41 d41.
IT 10
Combining this with (130) gives the solution

(132) vCr, f)) = ao(j) + f

n=1
(~) n (an(j) cos nf) + bn(j) sin nf)) ,
where an(j) and bn(j) are given by (31).
Convergence. The coefficients of an(j) and bn(j) in (132) are bounded by 1 in the disk
o S r S a. Hence if f(f)) satisfies the conditions of Churchill's theorem (see Section
9.1) then the series in (132) converges uniformly in the closed disk and the boundary
conditions yea, f)) = f(e) can be verified.
 9.3 STEADy-STATE DIFFUSION OF HEAT IN PLATES 619

The Poisson Integral for the Disk. It is of both theoretical and practical interest that
the series in (132) can be summed to provide a simple integral representation for vCr, e).
The calculation is as follows. First, replace an (j) and bn (j) in (132) by their integral
representations from (131) to get

(133) vCr,e) = -
1 121t jC¢)d¢
2n 0

00 Crlat (1t
+ I:
n=1
--
n
Jo jC¢) Ccos n¢ cos ne + sin n¢ sin ne) d¢
0

= -
1 121t jC¢)d¢+ I00 :Crlat
- 121t jC¢)cosnce-¢)d¢
2n 0 n=1 n 0

= hJo
1(1t j(¢) (00
1+ ~2(~)rn cosnce-¢) ) d¢.

The second step is to sum the series in parentheses. It is essentially a geometric series.
The calculation goes as follows Cwhere R{z} means the real part of z):

Thus if
r .,/,
Z = _el~
a
then

= Z+Z + z 3 + ... =
2 Z
-- (if ria < 1)
l-z

~ei'" (1 - ~e-i"')

11 - ~ei"'12

+ (D
2 2 .
(1 - ~ cos 0/) sin20/
Hence
00 r)n r
- cos ,/.
'f' -
(r)2
-
( _ cos no/ =
~
a a
a 1- 2 (D cos 0/ + (D 2
620 9 BOUNDARY VALUE PROBLEMS OF MATHEMATICAL PHYSICS

and

(134)

a2 + r2 - 2ar cos 1/r .

Combining (133) and (134) gives the integral formula

(l35) v(r, e) = -1
2n
1
0
2JT

a2 + r2
a2 - r2
- 2ar cos(B - ¢)
f(¢) d¢,

°
valid for :s r < a. This result is called Poisson's integral for the disk.
Caution. Poisson's integral does not make sense when r = a!

Problem 9.3.B. Circular Disk-Neumann Boundary Condition. This problem differs

from the preceding one only in replacing the Dirichlet boundary condition v(a, B) = f(B)
by the Neumann boundary condition av(a, B)!an = f(B). Since the normal derivative, on
the disk is the radial derivative the boundary condition is equivalent to vr(a, B) = f(B).
Thus vCr, B) is characterized by the following conditions.

(l36) vrr + lV
r r
+ l2r Vee = ° for 0 < r -< a, -00 < B< 00,

(l37) vr(a, B)= f(B) for -00 < B < 00,

(l38) vCr, B + 2n) = vCr, e) for -00 < B < 00,

(l39) f(B + 2n) = f(B) for -00 < B < 00,

(140) v(o, B) is finite and independent of B.

Product Solutions and Superposition. These are the same as for the preceding problem
and again lead to solutions of the form

L rn (An cos nB + Bn sin nB)

00

(141) vCr, B) = Ao +
n=l

with arbitrary coefficients An and Bn. To calculate them we determine the radial derivative
Vr by formal differentiation of (141) to get

L nrn-
00

(142) vr(r, B) = f(B) = 1 (An cos nB + Bn sin nB) .

n=l
 9.3 STEADy-STATE DIFFUSION OF HEAT IN PLATES 621

Use of the Neumann Boundary Condition. The boundary condition (137) and
equation (142) give

= fce) = L
00

(143) vrCa, e) (nan-IAn) cos ne + (nan-IBn) sin ne.

n=1

This gives the Fourier series for fce). In particular, there is no constant term, and hence
a solution exists if and only if fce) satisfies

(144) ao(j) = -
1 12n fee) de = O.
2n 0

Since fce) = avCa, e)!an, we see that (144) means that there is no net flux of heat into or
out of the disk. This condition clearly is necessary for thermal equilibrium. Also, (143)
places no restrictions on Ao in (141). This agrees with the fact, proved in Chapter 7, that
the solution of the Neumann problem is unique only up to an additive constant.
The coefficients in (143) must coincide with the Fourier coefficients off(8). This
uniquely determines the constants An, Bn Cn 2: 1) by

(145) nan-IAn = an(j),

nan-IBn = bn(j), n = 1,2,3, ....

Combining (141) and (145) gives the solution

(146) vCr, e) = Ao + a L --
00Crla)n
n=1 n
{an(j) cos ne + bn(j) sin ne},

where Ao is an arbitrary constant. The function vCr, e) defined by (146) satisfies each of
conditions (136)-(140). Condition (137) is verified by differentiating (146) with respect
to r and setting r = a to get (recall that ao(j) = 0)

L {an(j)cosn8 + bn(j)sinne} =fce).

00

vrCa,e) =
n=1

An Integral Representation for the Neumann Problem. The series solution (146) can
be summed by an indirect method to get an analogue of Poisson's integral for the disk.
To do this, note that from (146) one has

(147) L -
vr(r,e) = -a 00 (r)n {an(j) cosne + bn(j) sinne}.
r n=1 a
622 9 BOUNDARY VALUE PROBLEMS OF MATHEMATICAL PHYSICS

Now, the series in (147) is exactly the series solution of the Dirichlet problem with
boundary values 1(0). Thus it coincides with Poisson's integral, and we can write

(148) vr(r,B)
a
= --
r 2n
1 1 0
2n

a2 + r2 -
a2 - r2
2arcos(B - ¢)
I(¢)d¢.

Integrating with respect to r gives the solution of the Neumann problem as the integral

(149) vCr, B) =Ao + ~ [2n log [a 2 + r2 - 2ar cos(B - ¢)

n 10
r l/2I(¢) d¢.

This formula is more easily derived by complex variable methods - see page 318 in
[Ch-BJ for a proof.

Problem 9.3.9. Circular Sector-Dirichlet Boundary Condition. This problem asks for
the steady-state temperature distribution vCr, B) in a circular sector of radius 1 and angular
opening a (0 < a < 2n); see Figure 18.
The function vCr, B) is characterized by the follOwing conditions.

(150) v(r,B) is defined for 0 :s r :s 1,0 :s B :s a,

(151) Vrr + ~Vr + ';'VIiO = 0 for 0 < r :s 1,0 :s e :s a,
(152) v(l, e) = I(B) for O:s B :s a,
(153) v(r,O) = g(r) for O:s r :s 1,
(154) v(r,a) = her) for O:s r:S 1,
(155) v(O, e) is finite and independent of e.

Here I, g and h are prescribed functions that specify the boundary values of v on the
complete boundary of the sector.

v = h(r)
v = j(8)

x
FIGURE 18. Dirichlet's problem for a
o v = g(r) circular sector
 9.3 STEADy-STATE DIFFUSION OF HEAT IN PLATES 623

Reduction to Three Simpler Problems. If v = v(j,g,h) denotes the solution of

(150)-(155) then the linearity of the conditions implies that

(156) V(j,g,h) = v(j,o,O) + veo,g,O) + v(O,O,h) .

We shall show that each of the last three functions can be constructed by the separation
of variables method. When this has been done, the solution of the general problem will
be given by (156).
Construction of w = v(j,o,O). This function is characterized by the boundary value
problem

(157) w(r,e) is defined for 0::::: r::::: 1,0::::: e::::: a,

(158) Wrr + ~Wr + ~wee = 0 for 0 < r ::::: 1,0 ::::: e : : : a,

(159) w(r,O) = 0 and w(r,a) = 0 for 0::::: r::::: 1,
(160) w(l,e) = fee) for 0::::: e ::::: a,
(161) wen, e) is finite and independent of e,
where fee) is a prescribed function. We proceed to construct w(r, e) by separation of
variables.
Product Solutions. The product functions

(162) w(r, e) = R(r)8(e)

satisfy (157)-(159) and (161) if and only if

(163) r2R" + rR' - AR = ° for 0::::: r ::::: 1,

R(O) is finite,

and

(164)
8" + A8 = °
for 0::::: e : : : a,
8(0) = 0, 8(a) = 0,

where A is a separation constant. The Sturm-Liouville problem (164) gives the eigenpairs

An = C:f
8 n (e) = sin(mrela), n = 1,2,3, ....

The corresponding radial factors R(r) = Rn(r) satisfy

2" + rRn, - (nrr)2
r Rn -;:;- Rn = 0.
624 9 BOUNDARY VALUE PROBLEMS OF MATHEMATICAL PHYSICS

Euler's method gives the general solution

The value Rn (0) is finite only if h2 = O. Thus the product solutions are scalar multiples of

wn(r, e) = rnJr/(x sinCnnB/a), n = 1,2,3, ....

Superposition. This gives solutions of the form

00

wCr, e) = L AnrnJr/(x sinCnnBla) ,

n=1
where the An's are constants. To satisfy the final condition (160) we have
00

wCl,e) = fce) = LAnsinCmre/a).

n=1
This is a Fourier sine series for fce) on the interval 0 S e s a. Thus the proper choice
of the coefficients An is

(165) An = bn(j) = -21(X fCc/» sinCmrc/>/a) dc/>, n = 1,2,3, ....

a 0

The solution of the boundary value problem is

00

(166) w(r, e) = L bn(j)rnJr/(x sinCmre/a).

n=1
Example. Take a = n (so the sector is a half-disk) and
fCe) = 5 sinCe) - 11 sin(3e).

Then (166) becomes

00

wCr,e) = Lbn(j)rnsinCne),
n=1
and
00

fce) = L bn(j) sinCne).

n=1
Hence, by inspection, bl (j) = 5, b3 (j) = -11 and all other bn (j) = O. Combining these
results gives the solution

wCr, e) = 5r sinCe) - 11r3 sinC3e).

 9.3 STEADy-STATE DIFFUSION OF HEAT IN PLATES 625

Construction of z = veo,g,O), This function is characterized by the conditions

(167) z(r,e) is defined for 0:::: r:::: 1,0:::: e:::: a,

(168) Zrr + ~Zr + ~Zee = 0 for 0 < r:::: 1,0:::: e : : a,
(169) v(r,a)=O for O::::r::::l,
(170) z(l, e) = 0 for 0:::: e :::: a,
(17l) z(r,O)=g(r) for O::::r::::l,
(172) z(O, e) is finite and independent of e,
where g(r) is given.
Product Solutions, The product z(r, e) = R(r)8(e) satisfies (167)-(170) and (172) if
and only if

(173)
r2R" + rR' + AR = 0 for 0:::: r :::: 1,
R(l) = 0, R(r) bounded for 0:::: r:::: 1,

and

(174)
8" - A8 = 0 for 0:::: e : : a,
8(a) = 0,

where A is a separation constant. Now, (173) is the singular Sturm-Liouville problem of

Section 8.14, Example 3. It was shown there to have a pure continuous spectrum with
eigenpairs

A = fL2 (fL 2: 0),

(175)
R(r) = sin(fL log r).

The corresponding function 8 is the solution of (174) with A = fL2 A convenient choice
is
8(e) = sinh fLea - e)
sinh(fL e)
with 8(0) = 1. Thus the product solutions are the multiples of
sin(fL log r) sinh fL(a - e)
(176) z" (r, e) =
sinh(fLa)
, fL 2: 0,

Superposition. Multiplying (176) by an arbitrary function C(fL) and integrating over

o :::: fL < 00 gives
(177) zer, e) = 1 00

o
CefL) sin(fL log r)
sinh fLea - e)
. h(
sm fLa
) dfL·
626 9 BOUNDARY VALUE PROBLEMS OF MATHEMATICAL PHYSICS

We wish to choose C(jL) so as to satisfy the final boundary condition (171). By (177)
and (171) we must have

(178) z(r,O) = g(r) = lCC C(jL) sin(jL log r) djL.

It was shown in Section 8.14 that the proper choice of C(jL) is

(179) () 211 ()
C jL = - gr
nor
sin(jL log r) d
r.

Thus the solution is given by (177) with C(jL) given by (179). Sufficient conditions for
the validity of (178), (179) were given at the end of Section 8.14.
Construction of v(O,O,h). This may be obtained from (177) by the substitutions () ~ ex-()
andg ~ h.

Problem 9.3.10. Other Plates Describable by Polar Coordinates. The preceding ex-
amples do not exhaust the regions for which the Dirichlet problem can be solved by
separation of variables in polar coordinates. Further examples are rings defined by 0 <
a :::: r :::: b, 0 :::: () :::: 2n, the region exterior to a disk, defined by 0 < a :::: r < 00,
o :::: () :::: 2n, and truncated sectors, defined by 0 < a:::: r :::: b, 0 :::: () :::: ex. For further
discussions see the Exercises.

IExerc!!!~!.1
Square Plate. Compute the solution u(x,y) in series form.
1. Problem 9.3.1 with L = M = rr,fl (x) = O,h(x) = 1, g1 Cy) = g2(y) = o.
2. Problem 9.3.1 withL = M = rr,!1(x) =h(x) = 0,g1Cy) =g2(y) =y/rr.
Numerical Methods for the Square Plate. Compute the solution u(x,y) at x = y = rr/2 correct to 3
decimal places using Leibnitz's theorem for alternating series. The exact answer in each case is 1/4.

3. Problem 9.3.1 with L = M = rr,j1 (x) = O,h(x) = 1, g1 Cy) = g2(y) = o.

4. Problem 9.3.1 with L = M = rr,j1 (x) = hex) = 0, g1 Cy) = g2(y) = y/rr.
Superposition for the Square Plate. Compute the solution u(x, y) by applying the superposition principle
to appropriately chosen square plate problems.
5. Problem 9.3.1 with L = M = rr,j1 (x) = O,h(x) = 1, g1 (y) = g2(y) = y/rr.
6. Problem 9.3.1 with L = M = rr,j1(x) = hex) = xirr, glCy) = 0, g2Cy) = l.
Uniqueness. Verify by the uniqueness theorem that the given solution is the unique solution of the
boundary value problem.
7. Problem 9.3.1 with L = M = rr, fl (x) = 0, hex) = 1, g1 Cy) = g2(y) = y/rr. Unique solution
u(x,y) = y/rr.
 9.3 STEADy-STATE DIFFUSION OF HEAT IN PLATES 627

8. Problem 9.3.1 with L = M= n,fl(x) = hex) = xln, glCy) = 0, g2(y) = 1. Unique solution
u(x,y) =xln.
Square Plate with Dirichlet Boundary Condition. Find the Fourier sine series solution u(x,y).
9. Problem 9.3.1 with L = M = n,fj (x) = 0,12 (x) = f(x), gl (y) = g2 (y) = 0, where

f(x) = {x, 0 :s: x :s: nl2,

n - x,nl2 :s: x :s: n.

10. Problem 9.3.1 withL = M = n,fl(x) = 0,f2(X) =f(x), gl(y) = g2Cy) = 0, wheref(x) = x(n -x)
for 0 :s: x :s: n.
Square Plate with Neumann Boundary Condition. Find the Fourier cosine series solution u(x,y) for
the given Neumann boundary condition.
11. Problem 9.3.2 with f(x) = x - nl2 for 0 :s: x :s: n.

12. Problem 9.3.2 withf(x) = sin 2x for 0 :s: x :s: n.

Square Plate with Mixed Boundary Condition. Find the series solution u(x,y) for the given mixed
boundary condition.
l3. Mixed Problem 9.3.2 withf(x) = 1 for 0 :s: x :s: n.
14. Mixed Problem 9.3.2 withf(x) = x(n - x) for 0 :s: x :s: n.
Half-Plane with Dirichlet Boundary Condition. Find the Poisson Integral solution u(x,y) for the given
half-plane problem and verify by direct calculation that the given answer satisfies the differential equation
and boundary conditions.
15. Problem 9.3.3 withf(x) = H(x + 1) - H(x - 1) for -00 < x < 00 andfO) = f(-1) = 1/2. The
solution is

u(x, y) = ~ [tan -I (X; 1) _ -I C~ 1) ]. tan

16. Problem 9.3.3 withf(x) = H(x - a) - H(x - b) for -00 < x < oo,f(a) = feb) = 1/2, where a and b
are constants such that -00 < a < b < 00. The solution is

u(x,y) = ~ [tan- I (x ~ a) _tan- I (x ~ b)].

Half-Plane with Dirichlet Boundary Condition. Miscellaneous results for the Poisson integral solution
of Problem 9.3.3.
17. Show by direct integration that for all-oo < x < 00 and ally> 0,

18. Assume that a constant M exists such that If(x) I :s: M for all -00 < x < 00. Verify that the Poisson
integral solution of Problem 9.3.3,

u (x, y) = -
y
n
1 00

-00
f(~) d~
(~)2
x - 5 +y
2'

satisfies for y > 0 the inequality lu(x,y)1 :s: M.

Infinite Strip with Dirichlet Boundary Condition. Let u (j,g) (x,y) denote the solution of the infinite strip
problem 9.3.5 with boundary conditions u(x, 0) = f(x), u(x, 1) = g(x). Verify the follOWing results.
628 9 BOUNDARY VALUE PROBLEMS OF MATHEMATICAL PHYSICS

19. u(j,g) = u(O,g) + u(j,O).

L:
20. The solutions u(j,O) and u(O,g) are given by the formulas

= ~,y)g(~)~,
L:
u(O,g) (x,y) K) (x -

u(j,O)(x,y) = K2(X - ~,y)f(~) d~,

where

rrK) (x,y) = 1 00
cOS(JLX).
sinh(JL(l -
smhJL
y» dJL,

=1
-00

00 sinh(/lY)
rrK2(X,y) coS(/lX)-.- - d/l.
-00 smh/l

Semi-Infinite Strip with Dirichlet Boundary Condition. Let u(j,g,h)(x,y) denote the solution of the
semi-infinite strip problem 9.3.6. Verify the results below.
21. u(j,g,h) = u(j,o,O) + u(O,g,O) + u(O,O,h).
22. The solution u (O,O,h) is given by the formula

(0
u "
°h)
(x,y) = 1 00

-00
A(/l) cosC/lX).
sinh(/l(l -
smh/l
y» d/l,

where

A(/l) = 3:.
rr
roo h(~) cOS(JL~) d~.
10
Half-Plane Numerical Results. . Evaluate u(O, 1/2) by numerical integration, accurate to five digits.
2
23. Problem 9.3.3 withfCx) = e- x ,using the Poisson integral formula

( )- 1
u x,y - rr
~
-00
00
j(~) ~
(x _ ~)2 + y2 .
2
24. Problem 9.3.4 withf(x) = e- x ,using the integral formula

uCx,y) = - 11
rr -00
00
log[(x - ~)2 + ilf(~) d~.
Circular Disk. Determine explicitly the series solution of the given problem. The first two have Dirichlet
boundary conditions and the last two have Neumann boundary conditions.
25. Problem 9.3.7withfCO) = 1 + 5cosO - 3cos30 and a = 1.
26. Problem 9.3.7 withf(O) = cosO - 5 sin 30 and a = 2.
27. Problem 9.3.8 withfCO) = 4cosO - cos20 + cos 30 and a = 1, Ao = o.
28. Problem 9.3.8 withfCO) = cos20 and a = 2, Ao = o.
Circular Disk Numerical Methods. Determine numerically to five digits the solution u(r, 0) at r = ro,
o = 00, using the stated approximation method, given
fee) = {Io 0<< 00 << rr,
rr 2rr.
 9.3 STEADy-STATE DIFFUSION OF HEAT IN PLATES 629

29. Problem 9.3.7 with a = 1. Use a series representation for u and solve for ro = 0.1, eo = Jr/2.

30. Problem 9.3.7 with a = 1. Use Poisson's integral representation for u and solve for ro = 0.1, eo = Jr/2.
31. Problem 9.3.7 with a = 1. Use the exact solution representation below and solve for ro = 0.1, eo = Jr/2.

u(r, e) = -
1 + -1 tan -I (2rSine)
--.
2 Jr 1 - r2

This formula should not be verified (it's difficult!). The arctangent function above is defined by -Jr/2 <
tan- I x < Jr/2.

32. Problem 9.3.8 with a = 1, AO = O. Use a series representation for u and solve for ro = 0.1, eo = Jr/2.
Change Ion Jr < e < 2Jr from its defined value of 0 to the new value -1 (needed for ao(f) = 0).
Circular Sector. Derive the results below for the quarter disk.
33. The solution w = uCf,O,O) of Problem 9.3.9 for the quarter disk a = Jr/2 is given by

L bn(f)r2n sin(2ne),
00
w(r, e) =
n=1

34. The solution w = uCf,O,O) of Problem 9.3.9 for the quarter disk a = Jr/2 andiCe) = 2 sin 6i1 is given by
w(r, e) = 2r 6 sin(6i1).

Circular Hole in an Infinite Plate, Consider a large plate with a circular hole that is far from the plate
edges. We shall model this as the idealized problem of finding the steady-state temperature near a circular
hole of radius a in an infinite plate.
If rand e are polar coordinates with origin at the center of the hole then the steady-state temperature
vCr, e) satisfies

1 1
(180) Vrr + -;:vr + ;:2vee = 0 for a ::'0 r < 00 and all e,
(181) v(a, e) = I(e) for all e,
(182) vCr, e + 2Jr) = vCr, e) for all a ::'0 r < 00 and all e,
(183) vCr, i1) is bounded for a ::'0 r < 00 and all e,

where I(e) is the prescribed temperature at the hole.

35, Show that vCr, i1) = In r - In a is an unbounded solution of (180)-(182) with I ;; O.
36. Show that the solution to (180)-(182) is not unique Conclude that boundedness hypothesis (183) is
necessary
37, Show that the product solutions of (180), (182), (183) are vo(r, e) = Ao and for n = 1,2,3,
Vn (r, e) = r -n (An cos ne + En sin ne) ,
38, Apply the superposition principle to obtain the series solution

+L
00

vCr, e) = Ao r- n (An cos ne + En sin ne).

n=1

39. Calculate explicit formulas for An and En using orthogonality methods.

630 9 BOUNDARY VALUE PROBLEMS OF MATHEMATICAL PHYSICS

9.4 Transient Diffusion of Heat in Plates

In the preceding sections we have treated only boundary value problems whose unknown
functions depend on two independent variables. In this and the next two sections
problems whose solutions depend on three variable are solved. An interesting feature of
such problems is the necessity of two applications of the separation of variables procedure.
The general boundary value problem of transient heat flow in flat plates was
formulated in Section 7.3 and the uniqueness of the solution was shown. Here the special
cases of rectangular and circular plates are solved.

Problem 9.4.1. Rectangular Plate-Zero Boundary Condition. For simplicity, the case
of a square plate of side Jr will be treated. The case of a rectangle of arbitrary size can be
dealt with by the same method. We seek a function u(x,y, t) that satisfies

(1) u(x,y, t) is defined for °: : x ::: °: :Jr, y ::: Jr, t :::: 0,

(2) Ut = Uxx + Uyy for 0::: x ::: °: :
Jr, y ::: Jr, t :::: 0,
(3) u(O,y, t) = ° and u(Jr,y, t) = ° for 0::: y ::: Jr, t :::: 0,
(4) u(x,O,t)=O and u(x,Jr,t)=O for O:::x:::Jr,t::::O,

(5) u(x,y, O) = j(x,y) for 0::: X::: Jr, °: : y ::: Jr,

where j(x, y) is a prescribed function.

Product Solutions. First we shall separate the variable t from x and y by constructing
product solutions of (1)-(4) of the form

(6) U(X,y, t) = <I>(x,y)T(t).

Substituting this into (2) and dividing by <I>T gives

<I>xx + <I>yy T'

(7) --<I>---''-'- = T = -)..,

where).. is a separation constant. In particular, T is a multiple of the function

(8)

Eigenvalue Problem for <I>(x,y). Equations (1)-(4) and (7) imply that <I> must satisfy

<I>xx + <I>yy + )"<I> = ° for 0::: x ::: Jr, y ::: Jr, °: :

(9) <I>(O,y) = ° and <I>(Jr,y) = 0 for 0::: y ::: Jr,
<I>(x,O) = ° and <I> (x, Jr) = ° for 0::: x::: Jr.
 9.4 TRANSIENT DIFFUSION OF HEAT IN PLATES 631

This problem is our first example of an eigenvalue problem for a partial differential
operator. Each nontrivial solution of (9) defines an eigenpair for the linear operator
L<t> = -<t>xx - <t>yy.
Program. Our plan for solving the heat diffusion problem (1)-(5) is to carry out the
following steps. First, find the eigenpairs O"n, <t>n (x, y)) of problem (9). Second, use
superposition to build a solution of (1)-(5) of the form
00

(10) u(x,y,t) = Z:::Cn<t>n(X,y)e- Ant .

n=]

Third, prove the orthogonality property

(ll) 1][ 1][ <t>m(X,y) <t>n (x,y) dxdy = 0 if m #: n.

Finally, calculate the coefficients Cn in (10) as

C = an(/) = Jo][ J; j(x,y)<t>n(x,y) dx dy

(12)
KI<t>n(x,y)12 dx dy
n
Jo][

by applying (11) to the initial condition

00

(13) j(x,y) = .2::>n<t>n(X,y) for 0::: x::: Jr, 0::: y ::: Jr.
n=]

Solution of the Eigenvalue Problem. We seek eigenfunctions of (9) as product

functions

<t>(x,y) = X(x)Y(y).
This second application of separation of variables gives
X" (x) Y"(y)
(14) X(x) = - Y(y) - A = -/L,

where /L is a separation constant. Equation (14) and the boundary conditions of (9) lead
to the following familiar Sturm-Liouville problems for X and Y:

X" + /LX = 0 for 0::: x ::: Jr,

X(O) = X(Jr) = 0,

and
y" + vY = 0 for 0::: y ::: Jr,
yeo) = Y(Jr) = 0,
632 9 BOUNDARY VALUE PROBLEMS OF MATHEMATICAL PHYSICS

where

]) = A - fl.

The eigenpairs of the two problems are

fl = flm = m 2 , X(x) = Xm(x) = sin(mx), m = 1,2,3, ... ,

and

]) = ])n = n2 , YCy) = YnCy) = sin(ny), n = 1,2,3, ....

The corresponding eigenpairs of (9) and their time factors are

(15)

<l>mn(X,y) = sin(mx) sin(ny) ,

Tmn(t) = e-Cm'+n')t.
Note that the eigenpairs can be enumerated (in various ways) using a single index.
However, here the double index notation is more natural.
Product Solutions. We have constructed the product solutions

(16) umn(x,y, t) = <l>mn(x,y)Tmn(t) = sin(mx) sin(ny)e-Cm'+n')t,

wherem,n = 1,2,3, ....
Superposition. From (16) we shall build additional solutions of (1)-(4) of the form

LL
00 00

(17) u(x,y, t) = Cmn sin(mx) sin(ny)e- Cm '+n 2 )t,

m=l n=l

where the Cmn are constants. We wish to choose them in such a way that the initial
condition (5) is satisfied.
Initial Condition. Equation (15) gives, for 0 S x S Jr, 0 S Y S Jr,

= j(x,y) = L L
00 00

(18) u(x,y,O) Cmn sin(mx) sin(ny).

m=l n=l

Orthogonality. The orthogonality of the eigenfunctions (15) will be verified and used
to calculate the coefficients Cmn . The following calculation is elementary.

(19) lIT lIT <l>mn(X,y) <l>pq(x,y) dxdy

= lIT lIT sin(mx) sin(ny) sin(px) sin(qy) dx dy

 9.4 TRANSIENT DIFFUSION OF HEAT IN PLATES 633

= (11'( sin(mx) sin(px) dx) (11'( sin(ny) sin(qy) dY)

= jJr2/4 if (m, n) = (p, q),
o if (m, n) =j:. (p, q).
Coefficients. The orthogonality property (19) allows us to calculate the Cmn in (18) as

(20) Cmn = amn(j) == 2' 411'(11'( j(x,y)sin(mx)sin(ny)dxdy.

Jr 0 0

Solution. Combining our results gives the solution of (1)-(5):

L L amn(j) sin(mx) sin(ny)e-Cm2+n2)t.

00 00

(21) u(x,y, t) =
m=ln=l

Convergence. The proposed solution (21) is only formal, because we have not shown
that the series converges. Equation (20) implies that
411'(11'( If(x,y)ldxdy
lamn (j)I.:::2' forall m,n=I,2,3, ....
Jr 0 0

It follows that if t > 0 then the series in (21) converges absolutely and uniformly in x
andy.
Double Fourier Series. For t = 0, (21) gives the formal expansion
00 00

(22) f(x,y) = LLamn(j)sin(mx)sin(ny),

m=1 n=1

where amn(j) is given by (20). This series is called the double Fourier sine series for
f(x,y). The convergence theory for such double series is an advanced topic that cannot
be developed here. For a careful treatment see [B-McGl.

Problem 9.4.2. Circular Disk-Zero Boundary Temperature. To simplify the notation,

the case of a disk of unit radius will be treated. The general case is left for the Exercises.
Also, to exploit the circular geometry of the disk, we shall use polar coordinates, with
the origin at the center of the disk. Thus the temperature will be given by a function
u(r, e, t), and the problem has the following form

Boundary Value Problem. Find the function u(r, e, t) that satisfies the conditions

(23) u(r,e, t) is defined for 0 .::: r .::: 1, e real, t ::: 0,

(24) Ut = Un + ~Ur + ~Uef! for 0 < r'::: 1, e real, t ::: 0,
(25) u(l, e, t) = 0 for e real, t ::: 0,
(26) u(r, e, 0) = fer, e) for 0.::: r .::: 1, e real,
634 9 BOUNDARY VALUE PROBLEMS OF MATHEMATICAL PHYSICS

where fer, e) is a prescribed function.

Implicit Conditions. The geometry and physics of the problem necessitate the
following additional conditions.

(27) u(r, e + lrr, 0) = u(r, e, t) for 0:::: r :::: 1 and e real,

(28) u(O, e, t) must be finite and independent of e for all t :::: O.
Product Solutions. We seek functions of the form

(29) u(r, e, t) = <P(r, e)T(t)

that satisfy (23)-(25), (27) and (28). The usual separation argument gives

(30)

where A is a separation constant, while (A, <P(r, e)) is a solution of the following
eigenvalue problem.
Eigenvalue Problem. Find all eigenpairs (A, <p(r, e)) such that <P =f. 0 and

(31) <Prr + ~<Pr + ~<PIJ!) + A<P = 0 for 0:::: r:::: 1, e real,

(32) <P(1,8) = 0 for all real e,
(33) <Per, (J + 2n) = <Per, (J) for 0::: r ::: 1, (J real,
(34) <P(r,(J) is bounded for 0:::: r:::: 1,8 real,
(35) <p(0, e) is finite and independent of 8.

Solution of the Eigenvalue Problem. Again we seek product solutions

(36) <P(r,8) = R(r)8(8).

Substituting this into (31) gives, by the usual argument,
r2RI/(r) + rR'(r) + )..r2R(r) 81/(e)
(37)
R(r) = - 8(e) = JJ..,

where JJ.. is a separation constant. For the 8-factor, (31) and (33) imply that we have the
periodic Sturm-Liouville problem

81/ + JJ..8 = 0 for 0:::: (j :::: 2n,

(38)
8(lrr) = 8(0), 8 ' (2n) = 8 '(0).

The eigenpairs are

JJ..=JJ..m=m2,
(39)
8(e) = 8 m W) = Am cos(me) + Bm sin(me), m = 1,2,3, ... ,
 9.4 TRANSIENT DIFFUSION OF HEAT IN PLATES 635

where Am and Bm are arbitrary constants. Next, the R-factor is a solution of the singular
Sturm-Liouville problem

(40) R" + ~R'

r
+ (A - m2) R = 0,
r2
R(l) = 0, R(r) bounded for 0:::: r :::: 1.

Here m2 is fixed as one of the eigenvalues of the 8-factor, and A is an eigenvalue of

problem (40). The special case m = 0 of (40) was done in Section 8.14. Proceeding in
the same way leads to the eigenpairs

(41)
A = Amn = s~n'
R(r) = Rmn(r) = ]m(smnr),

where Smn is the nth positive zero of the Bessel function]m(x), numbered in increasing
order: 0 < Sml < Sm2 < .. '.
Combining (39) and (41) gives the eigenpairs of the <I>-problem:

A = Amn = s~n'
(42) <I>(r, e) = ]m(smnr) cos(me), m = 0,1,2, ... ,
<I>(r, e) = ]m(smnr) sin(me) , m = 1,2,3, ... .

Note that for m = 0 the eigenspace has dimension 1, while for integers m ~ 1 it has
dimension 2. The eigenfunctions listed in (42) are linearly independent. This is most
easily proved by showing that they are orthogonal when integrated over the disk. For
example, we have

(43) 11 1" Om (Smnr) cos(me» (Jp(spqr) cos(pe») r de dr

= (1 1
1m (Smnr)lp(spqr)r dr) (1" cos(me) cos(pe) de)

=0

if (m, n) i=- (p, q). Indeed, if m i=- p, then the cosine integral is zero, while if m = p but
n i=- q, then the Bessel function integral vanishes by the orthogonality theorem for the
singular Sturm-Liouville problem (40). A similar calculation holds for the product of
any two of the eigenfunctions (42).
Normalizing Constants. The eigenfunctions of problem (40), given by (41), are not
normalized. Instead, it is known that

(44) 1o
1 2 1 2
lm(smnr)rdr = - lm+l(smn).
2
636 9 BOUNDARY VALUE PROBLEMS OF MATHEMATICAL PHYSICS

For a proof see, for example, page 252 in [Krey].

Superposition. The product solutions (42) give the formal solution of the heat diffusion
problem for the disk as a sum
00

(45) u(r,(J,t) = LAonJo(sonr)e-s6nt

n=l
00 00

+L L {Amn cos(m(J) + Bmn sin(m(J)}Jm(smnr)e-s~nt,

m=l n=l

where the coefficients Amn and Bmn are arbitrary constants, to be chosen such that the
initial condition (26) holds. This is true if
00

(46) u(r, (J, 0) = fer, (J) = L AonJo (SOn r)

n=l
00 00

+L L {Amn cos(m(J) + Bmn sin(m(J)}Jm(smn r).

m=l n=l

The orthogonality and normalizing constants give for n :::: 1 and m :::: 1,

Convergence. The convergence theory of the Fourier-Bessel series (46), (47) requires
a deeper study of the properties of Bessel functions. It will not be presented here.
The Symmetric Case. If u(r, (J, 0) = fer) is independent of (J then Amn = Bmn = 0 for
all m, n :::: 1 by (47). It follows from (46) that u = u(r, t) is independent of (J and
00

(48) u(r, t) = L AonJo(sonr)e-s6nt,

n=l

where

(49) Aon = aon(j) == ~t

]1 (SOn) Jo
f(rHo(sonr)r dr.

Apart from notation this result coincides with the solution of Problem 9.1.6, Section 9.1.
 9.4 TRANSIENT DIFFUSION OF HEAT IN PLATES 637

Other Boundary Conditions. With minor changes the results of this section apply to
plates that satisfy the Neumann and Robin boundary conditions. See the Exercises.

Exercises 9.4

Square Plate. Solve the transient heat flow problem 9.4.1 for the prescribed functionJ and obtain the
formula for uCx,y, t) without the use of integration tables.
1. f(x,y) = sin 3x sin Sy - sin 4x sin 6y,
u = sin 3xsin Sye- 34t - sin 4xsin 6y e- 52t .
2. f(x,y) = lI'sin3xsiny + sinxsin2y,
u = lI'sin3xsinye- lOt + sin x sin 2ye- 5t .
Square Plate. Solve the transient heat flow problem 9.4.1 for the following special choices of j(x, y).
Express the answer as an infinite series with all coefficients evaluated through integration tables.

3. f(x,y) = 1. 6. f(x,y) = x 2y.

4. f(x,y) = y. 7. f(x,y) = x(lI' - x).

S. f(x,y) = xy. 8. f(x,y) = x(lI' - x)y(lI' - y).

Rectangular Plate. Complete the details for the following generalization of Problem 9.4.1 to an arbitrary
rectangle of sides 0 ::: x ::: L, 0 ::: y ::: M.
9. Formulate the transient heat flow Problem 9.4.1 for the rectangle case.
10. Show that separation of variables leads to product solutions X(x)YCy) in which the Sturm-Uouville
problems have solutions Xm(x) = sin(mll'xlL) and Yn(y) = sin(nll'yIM).
11. Show that

m1l')2
Amn= ( L + (nll')2
M '

12. In the series solution u = L~=l L~l cmnXmYnTmn, verify the formula

Cmn = ~ {M (L f(x,y)Xm(X)Yn(t)dxdy.
LM 10 10
Neumann Conditions for the Square Plate. Complete the steps below to solve Problem 9.4.1 with the
Dirichlet conditions replaced by Neumann boundary conditions Ux = 0 on the vertical boundary and Uy = 0
on the horizontal boundary.
13. Apply separation of variables to obtain Sturm-Liouville problems with solutions Xm (x) = cosCmx) and
YnCy) = cos(ny).
14. Show that the time factors Tmn(t) and eigenvalues Amn use the same formulas as Problem 9.4.1.

15. Show that u = L~=o L~O cmnXmYnTmn·

16. Determine the constants Cmn from the initial condition u(x,y,O) = f(x,y). In particular, 1I'2coo =
J; g J(x,y)dxdy.
Mixed Conditions for the Square Plate. Complete the steps below to solve Problem 9.4.1 with the
Dirichlet condition on the horizontal boundary replaced by the Neumann condition uy = O.
638 9 BOUNDARY VALUE PROBLEMS OF MATHEMATICAL PHYSICS

17. Apply separation of variables to obtain Sturm-Liouville problems with solutions Xm (x) = cos(mx) and
Yn(y) = sin(ny).

18. Show that the time factors TmnW and eigenvalues Amn use the same formulas as Problem 9.4.1.

19. Verify u = 2::;;;'=0 2::~1 (mnXmYn Tmn.

20. Determine the constants (mn from the initial condition u(x,y, 0) = f(x,y).
Circular Disk. Solve the transient heat flow Problem 9.4.2 for the special cases below without using
integration tables.

21. f(r,8) = JO(501 r) - 2JO(503r).

22. f(r,8) = 3cos8Jl(S12r) + 5 sin 3(113 (S37r).
Circular Disk GraphicS. Graph the 3-D solutions of the above transient heat flow problems using the
formula supplied and a computer graphics program. Supply a 3-D snapshot for each of the time values
t = 0, t = 0.5, t = 1.

24. «(r, 8, t) = 3 cos (Ill (s12r)e-s12t

2
+ 5 sin 3(113 (s37r)e-s37t
2
.
Other Circular Disk Problems. Establish the following generalization of Problem 9.4.2.
25. (Plate of Radius a) Formulate and solve by separation of variables the transient heat diffusion problem
for a circular plate of radius a > o.

9.5 Vibrations of Drums

In Section 7.5 the physical problem of predicting the motion of a taut drumhead
was modeled as a boundary value problem for the two-dimensional wave equation.
Specifically, if the resting drumhead occupies a domain Q in the Cartesian (x,y)-plane,
then u(x, y, t), the displacement of the drumhead out of this plane, is uniquely determined
by the following conditions. See Figure 19.

(1) u(x,y, t) is defined for all (x,y) in Q and t ~ 0,

(2) Utt = Uxx + Uyy for all (x,y) in Q and t ~ 0,

x
FIGURE 19. Drumhead over n
 9.5 VIBRATIONS OF DRUMS 639

(3) U(x,y, t) = 0 for all (x,y) in an and t ::: 0,

(4) u(x,y,O) = f(x,y) and ut(x,y,O) = g(x,y) for all (x,y) in n,
where f and g are prescribed functions, namely the initial displacement and velocity
of the drumhead, respectively. The uniqueness of a function u(x,y, t) with properties
(1)-(4) was proved in Section 7.5 by the energy method.
Rayleigh's Principle. An observation due to Rayleigh 1 will be used to simplify the
boundary value problem (1)-(4). Rayleigh noted that ifthe solution of (1)-(4) is denoted
by U (f,g) then

(5)

To verify (5), we note that the two sides of equation (5) satisfy conditions (1)-(4) (check
this!). Hence they are equal, by the uniqueness theorem. Equation (5) is Rayleigh's
principle. It allows us to concentrate on the special case of (1)-(4) for which f = O.
The remainder of this section treats the vibrations of a circular drumhead of unit
radius. It is natural to use polar coordinates with origin at the center of the drumhead,
so that the displacement is a function u(r, e, t). We shall consider first the symmetric
case where g = g(r) and u = u(r, t) are independent of e. Then the general case where
g = g(r, e) will be analyzed.

Problem 9.5.1. Circular Drumhead-Symmetric Case. We seek a function u(r, t) =

uCO,gl(r, t) that satisfies

(6) u(r, t) is defined for 0 S r S 1, t ::: 0,

(l) Utt= U rr + ~ur
1
for o < r S 1, t ::: 0,
(8) u(l, t) = 0 for t ::: 0,
(9) u(r,O) = 0 for o S r < 1,

(10) utCr,O) = g(r) for OSrs1,

where g(r) is a prescribed function. In addition, there is the regularity condition

(11) u(r, t) is finite for 0 S r S 1, t ::: O.

Of course, we expect this condition to hold for the physical problem of the drum.
However, we shall see that it must be used expliCitly to obtain a unique solution.

I Lord Rayleigh (john William Strutt), 1842-1919, English physicist and mathematician.
640 9 BOUNDARY VALUE PROBLEMS OF MATHEMATICAL PHYSICS

Product Solutions. The product solutions of (6)-(9) and (11) are the scalar multiples of

(12) un(r, t) = JO(Snr) sin(snt), n = 1,2,3, ... ,

where a < S1 < S2 < ... are the positive roots S of the equation Jo(s) = O. The
Sturm-Liouville problem leading to the radial factorsJo(snr) was solved in Example 1 of
Section 8.4 and in Problem 9.1.6, Section 9.1.
Superposition. The solution of problem (6)-(11) should have the form
00

(13) u(r, t) = L cnJO(sn r) sin(snt).

n=1
We wish to choose the coefficients en so as to satisfy condition (10). Now, formally,
00

(14) ut(r, t) = L cnsnJO(sn r) cos(snt).

n=1
Thus
00

ut(r, 0) = g(r) = L cnsnJO(sn r).

n=1
By the orthogonality of the eigenfunctions JO(snr), we must take

(15) CnS n = an(g) == -2-

2
J1 (Sn)
11° g(r)JO(snr)r dr;

see Section 9.4 above.

Combining (13) and (15) gives

u(O,g)(r, t) = f
n=1
an (g) JO(snr) sin(snt).
Sn
This result and the Rayleigh principle give
00

uCf,O) (r, t) = u;oj) (r, t) = Lan(/) JO(Snr) COS(Snt),

n=1
where an(/) and an(g) ae defined by (15). Finally,

(16) uCf,g) Cr, t) = uCf,O) (r, t) + u(O,g) (r, t)

~ { sin(snt) }
= f=tJo(Snr) an(/)coS(Snt)+an(g)-S-n- .

The function uCf,g) defined by (16) is easily seen to satisfy (6)-(11), provided that the
series involved are convergent. The convergence question involves a deeper study of
Fourier-Bessel series. It will not be pursued here.
 9.5 VIBRATIONS OF DRUMS 641

Problem 9.5.2. Circular Drumhead-General Case. In the general case, the drum
displacement u(r, e, t) = u(o,~) (r, e, t) will be determined by the following conditions.

(17) u(r, e, t) is defined for °: : r ::: 1, e real, t :::: 0,

(18) = + ~Ur + ~uee for ° < r ::: 1, e real, t :::: 0,
°
Ut Urr

(19) u(l, e, t) = for e real, t :::: 0,

(20) u(r, e, 0) = ° for 0::: r < 1, e real,
(21) ut(r, e, 0) = g(r, e) for 0::: r ::: 1, e real,

where g(r, e) is a prescribed function. In addition, the physical problem requires the
following.

Implicit Conditions.

(22) u(r, e + 2n, t) = u(r, e, t) for °: : r ::: 1, e real, t :::: 0,

(23) u(O, e, t) is finite and independent of e for all t :::: 0.
Much of the separation of variables method for the problem was carried out for the heat
equation in Section 9.4 above. Here we shall only summarize the results.
Product Solutions. These satisfy all of conditions (17)-(23) except condition (21).
They have the form

(24)

where m = 0,1,2, ... and n = 1,2,3, ....

Superposition. Summing over the product solutions gives as a candidate for the
solution of the full problem,

L Aon]o(son r) sin(sont)
00

u(r, e, t) =
n=1

+L L {Amn cos(me) + Bmn sin(me)}Jm(smn r) sin(smnt),

00 00

m=1 n=1
and hence

L Aonson]o(sonr) cos(SOnt)
00

ut(r, e, t) =

+L L {Amnsmn cos(me) + Bmnsmn sin(me)}Jm(smnr) cos(smnt).

00 00

m=1 n=1
642 9 BOUNDARY VALUE PROBLEMS OF MATHEMATICAL PHYSICS

FIGURE 20. Symmetric drumhead

u(r, t) = sin(s2t)Jo(s2r)/s2

On setting t = 0 and using the orthogonality of the eigenfunctions Csee Section 9.4
above), we get

Amnsmn = amnCg),
Bmnsmn = bmnCg), m = 0,1,2, ... , n = 1,2,3, ... ,

where amn Cg), bmn Cg) are as in Section 9.4. Combining these results gives the solution

u(O,g) Cr, e, t) = fn=1

aon Cg) ]oCSOnr)
SOn
sinCsont)

Similarly,

u(f,O) (r, e, t) = u~oj) Cr, e, t)

L aOn (g)]O(SOnr) cos(SOnt)

00

u(f,O) (r, e, t) =
n=1

+L L {amn(g) cos (me) + bmn(g) sin (me) } ]m(smn r) cos(smnt).

00 00

m=1 n=1
The symmetric case and the general case produce Significantly different 3-D plots,
as shown in Figures 20 and 21.

Rayleigh's Principle. Let v = u(OJ) and w = u(O,g). Assume additionally that the partials Vttt. Vtxx and
Vtyyare continuous. The follOwing steps verify Rayleigh's principle.
1. Verify that u = Vt + w is a solution of the wave equation Utt = Uxx + Uyy .

FIGURE 21. General drumhead

u(r, e, t) = 0.2J1 (S13r) cas (e) COS(S13t)
 9.6 STEADy-STATE DIFFUSION OF HEAT IN SOLIDS 643

2. Verify that U = Vt + w satisfies u(x,y, t) = 0 for t "= 0, (x,y) in aQ, and u(x,y, 0) = j(x,y) , Ut(x,y) =
g(x,y) for (x,y) in Q and t "= O.
Circular Drum. Verify the following results for the symmetric case in Problem 9.5.1.

3. The circular drumhead solution for the initial conditions fCr) = 0 and g(r) = JO(S2 r) is solved without
integration of Bessel functions

4. The circular drumhead solution for the initial conditions j(r) = 0 and g(r) = 1 is given by

u(r, t) = L
DO 2
-2--JO(snr) sin(snt).
n=1 s~h (sn)

The integrations use a change of variables x = Snr, the identity xJO(x) = (xl! (x)), and the identity from
Section 9.4, J~ J6(Snr)r dr = ±H(Sn).
Rectangular Drum. A membrane stretched over a rectangular frame is called a rectangular drumhead.
Establish the following extensions of the circular drumhead theory by using the rectangular methods of
Problem 9.4.1.

5. Formulate and solve the rectangular drumhead problem for shape 0:5 x :5 L, 0 :5 Y :5 M.

6. Formulate and solve the square drumhead problem for shape 0:5 x :5 IT, 0 :5 Y :5 IT.

Circular Drum. Formulate and solve the follOwing extensions for the circular drumhead of radius a > 0
and propagation speed c = y'Tlp, using the method of separation of variables.

7. The symmetric case as in Problem 9.5.1.

8. The general case as in Problem 9.5.2.

Nodal Lines. Forthe drum of radius 1 and speed 1 treated in Problem 9.4.1, the normal mode functions
are

un(r,t) =JO(snr)sin(snt)·

The nodal lines of mode Un are the roots of lo(snr) = 0 in the domain 0 :5 r < 1. The nodal lines are
stationary relative to the time oscillations of Un.
9. Compute and graph in polar coordinates the nodal lines of mode U3.

10. Show that the nodal lines of Un are exactly the n - 1 concentric circles r = sm/sn, m = 1,2, ... ,n - 1.

9.6 Steady-State Diffusion of Heat in Solids

This section is a continuation of Section 7.7. There the boundary value problems
governing steady-state heat diffusion in three-dimensional solid bodies were formulated
and the uniqueness of the solutions was discussed. Here solutions are constructed by
separation of variables for bodies of two simple shapes and various boundary conditions.
First rectangular parallelepipeds are considered. Then spherical bodies are treated. These
644 9 BOUNDARY VALUE PROBLEMS OF MATHEMATICAL PHYSICS

choices of shapes and boundary conditions are only illustrative. Many similar problems
with rectangular or spherical shapes can be solved by the same method.

Problem 9.6.1. Rectangular Parallelepiped with Dirichlet Boundary Conditions. Let

° °
the parallelepiped be defined by the inequalities :s x :s L, :s y :s M, :s z :s N, °
wherex,y and z are Cartesian coordinates. It will suffice to treat the case where five of the
six faces of the parallelepiped are held at zero temperature (see Problem 9.3.1, Section
9.3). If the nonzero boundary condition is in the plane z = N then the temperature
u(x,y, z) is determined by the conditions

(1) u(x,y, z) is defined for °:s x :s L, °:s °:s z :s N,

y :s M,

(2) Uxx + Uyy + U zz = ° for o:s x :s L, °:s y :s M, °:s :s N, z

(3) u(O,y, z) = ° and u(L,y, z) = ° for O:s :s M, °:s :s N,

y Z

(4) u(x,O,z)=O and u(x,M,z)=O for O:sx:sL,O:sz:sN,

(5) u(x,y,O) = ° for O:s x:s L, O:s y:s M,
(6) u(x,y,N) = f(x,y) for O:s x :s L, °:s y :s M,

where f(x, y) is a prescribed function.

Product Solutions. We shall seek solutions of the linear homogeneous conditions

(1)-(5) of the product form

(7) U(x,y, z) = X(x)¥(y)2(z).

The Laplace equation (2) holds for (7) if

X" (x) ¥"(y) 2"(z)
(8) --+--+--=0
X(x) fey) 2(z)

° °
for :s x :s L, :s y :s M, °:s z :s N. It follows that each of the three quotients in (8)
is a constant, say
X" (x) yll(y) 2"(z)
(9) X(x) = -A, y(y) = -p., --=-v
2(z) ,

and

(10) A + p. + v = 0.
Use of the boundary conditions (3)-(5) gives

(ll)
X" + AX = ° for O:s x :s L,
X(O) = 0, X(L) = 0,
 9.6 STEADy-STATE DIFFUSION OF HEAT IN SOLIDS 645

and

yl/+jJ.,Y=O for O:::y:::M,

(12)
yeO) = 0, Y(M) = o.

These are simple Sturm-Liouville problems with the familiar eigenpairs

fhr 2
)... =)..., = - -
(13) < L2'
X(x) = Xe(x) = sin
(cnx)
L ' C = 1,2,3, ... ,

and

m2 n 2
jJ.,=jJ.,m=M2'
(14)
YCy) = YmCy) = sin (m~y) , m=1,2,3, ....

Finally, by (10) and (13)-(14),

and Z(z) = Zem (z) is determined, up to scalar multiples, by

(15)

It will be convenient to choose the solution

sinh (
Zem(Z) =
sinh
(
(2
iJ
m'
+ M2 nN )'

so that Zem (N) = 1. Then the product solutions (7) are the nonzero multiples of

c ) sinh ( G+ :~ nz)
(16) uemCx,y, z) =
(
sin ~x sin (":Y) -.--:(~=e2===m=2::-\) ,
smh iJ + M2 nN
whereC,m= 1,2,3, ....
646 9 BOUNDARY VALUE PROBLEMS OF MATHEMATICAL PHYSICS

Superposition. Linear combinations of the product solutions (16) all satisfy (1)-(5).
To satisfy the final boundary condition (6), we try

L L ClmUem(X,y, z)
00 00

U(x,y, Z)=
e=1 m=l

&+ ~~ nz)
C:
(17)

8 ~ cem
00 00 (£) sinh (
= sin ~X sin
y
) -sl'-n-hl~e=2===m2=-""':}
L2 + M2 nN

Coefficients. We wish to choose the coefficients cem to make (6) hold:

oc
L L cemuem(x,y,N)
00

j(x,y) =
(18) e=1 m=1

=
~~
f:r. ~ Cl m sin (£nx)
L sin (mny)
M
for 0 ::: x ::: L, 0 ::: y ::: M. This can be done by using the orthogonality of the
eigenfunctions sin(£nxlL) and sin(mnyIM) on 0 ::: x ::: Land 0 ::: y ::: M, respectively
The latter can be written

1\in CZX) sin (PZX) dx = 8ep~,

(19)
1M sin (e;:) C;:) dy
sin = 8eq ~,
where 8u = 1 and 8mn = 0 if m f. n. Applying (19) to (18) gives

1o
M
j(x,y) sin (qny)
- dy = -M ~
M
~ Ceq sin (£nx)
2
-
l=1
,
L

and

(20)

These calculations are, of course, only formal. Questions of convergence of the series
and interchange of summation and integration are left for more advanced courses. For
details see [B-McGl.
An Example. Consider a cube of side n so L = M = N = n and take j(x, y) = xy. Then

cpq = :2 (1" x sin (px) dx) (1" y sin(qy) dY) .

 9.6 STEADy-STATE DIFFUSION OF HEAT IN SOLIDS 647

Now,

1" x sin (px) dx = [;2 sin(px) - ; COS(PX)] t: JT(-1)p+1

p
so
4 (_l)p+1 (-l)q+l 4( -l)p+q
Cpq = 11"2 P q pq
Hence the formal solution is

(-1 )p+q sinh ( JP2+q2 z)

LL
00 00

u(x,y, z) = 4 sin(px) sin(qy) ( ).

p=1 q=1 pq sinh J
p2 + q2 11"

Problem 9.6.2. Rectangular Parallelepiped with Mixed Boundary Conditions. We shall

consider the case defined by conditions (1)-(5), as before, but with (6) replaced by

(6') uz(z,y,N) = j(x,y) for 0.:::: x.:::: L, 0.:::: y .:::: M.

This defines a mixed problem with the Neumann condition (6') on one face and Dirichlet
boundary conditions u = 0 on the remaining five faces.

Product Solutions. As before, the product solutions of (1)-(5) are the nonzero scalar
multiples of

(21) UCm(x,y, z) = sin (€~x) sin C:y) sinh (J~: + :: rrz).

For this problem there is no special advantage to the normalization (16).
Superposition. We shall take

u(x,y, z) = ~~
L.., L.., Cem sin (€rrx)
- sin (mrry)
- - sinh 2 + 2m rrz ) .
(g;g2 2
e=1 m=1 L M L M

Then formally,

~ L..,
uz(x,y, z) = L.." ~ cern
, sin (€rrx)
- sin (mrry)
- - cosh (g;g2 2
2 + m rrz ) , -2
f=1 m=1 L M L M

where c~m = Cfmrr &+ ~~ , whence boundary condition uz(x,y, N) = j(x,y) gives
(€rrx) (mrry)
j(x,y) = 8~Cfmsin
00 00"
L sin M '
648 9 BOUNDARY VALUE PROBLEMS OF MATHEMATICAL PHYSICS

where c~m = CCmn &+ ~~ cosh ( &+ ~~ nN). The coefficients are now calculated
by means of the orthogonality (9), as before. The result is

1L 1Mj(x,y) sin (P~x) sin (q;) dx dy

LM ~
=4nYV+Wcos h
(~)
YV+WnN cpq .
Problem 9.6.3. Sphere with Dirichlet Boundary Conditions, Axisymmetric Case. We
shall construct the steady-state temperature distribution u(x,y, Z) in a sphere of unit
radius, defined by the inequality x 2 + y2 + Z2 S 1, with prescribed values on the surface
of the sphere. For this problem it is natural to use spherical coordinates

x = rsin4>cos8,
(22) y = r sin 4> sin 8,
Z = rcos4>,

where r ~ 0, 0 s 4> s n, and 0 S 8 S 2n, and to seek u as a function of r, 4>, and

8: u(x,y, z) = vCr, 4>, 8). Moreover, we shall treat only the axisymmetric case where v is
independent of 8:

(23) u(x,y,Z) = v(r,4».

Substituting (23) into the Laplace equation, we find after a lengthy but straightforward
calculation, based on (22)-(23) and the chain rule, that

(24)

With these assumptions, the boundary value problem asks for a function vCr, 4» that
satisfies

(25) v(r,4» is defined for 0 S r S 1,0 S 4> S n,

(26) Vrr + ~vr + r2s~n4> (sin4>v4»4> = 0 for 0 < r S 1,0 S 4> S n,
(27) vO,4» = j(4)) for 0 S 4> S n,

where 1(4)) is a prescribed function, defined for 0 S 4> S n. We note that the partial
differential equation (26) is Singular at r = 0 and at 4> = 0 and 4> = n. The physical
interpretation of the model requires that vCr, 4» be finite and single-valued at these points.
 9.6 STEADy-STATE DIFFUSION OF HEAT IN SOLIDS 649

This suggests the additional conditions

(28) v(O, ¢) is finite and independent of ¢ for 0 :::: ¢ :::: n,

(29) v(r,O) and vCr, n) are finite for 0 :::: r :::: 1.
We shall see that these additional conditions are needed to determine vCr, ¢) completely.

The boundary value problem (25)-(29) is more difficult than those considered
above because the partial differential equation (26) has variable coefficients that have
singularities on the z-axis. Nevertheless, we shall find that it can be solved by Fourier's
separation of variables method.
Product Solutions. We search for product solutions

(30) vCr, ¢) = R(r)<P(¢)

that satisfy (25), (26), (28) and (29). As usual, the boundary condition (27) is left for the
superposition phase. Substituting (30) into the Laplace equation (26) and multiplying
by r2 gives
R
(r2R" + 2rR) <P + -.- (sin¢<P')' = O.
sm¢
Separating the variables then gives, by the usual argument,
_1_ (sin ¢ <p')'
sm</>
(31) - ' - - - - - = A,
<P
where A is the separation constant. For the radial function R(r), equation (31) is an Euler
differential equation

(32) r2R" + 2rR - AR = O.

It is solved below by the method of Chapter 3, Section 3.4. First the pairs (A, <P(¢)) will
be determined as the eigenpairs of a singular Sturm-Liouville problem. Conditions (29)
and (31) define the problem: Find all eigenpairs of

(33)
_1_~
sin¢d¢
(sinA.~)
'f'd¢
+ A<P = 0 for 0 <
-
A. < n
'f' - ,

<P(O) and <P(n) finite.

This is singular at both ¢ = 0 and ¢ = n because the Sturm-Liouville coefficients
w(¢) = p(¢) = sin ¢ vanish at both points (see Section 8.14).
Substitution. To recognize (33) as a previously solved problem, it is helpful to change
the independent variable ¢ to x = cos ¢ and define

X(x) = <P(¢).
650 9 BOUNDARY VALUE PROBLEMS OF MATHEMATICAL PHYSICS

The chain rule then gives

<p'=- sin¢X',
dx
<P"=(-sin¢)X" - - (cos¢)X'
d¢
=(-sin¢)2X" - (cos¢)X'
=(1 - x 2 )X" - xX'.

Substituting this into the differential equation of (33), written as

<P" + c~s¢ <P' + A<P = 0,

sm¢
gives Legendre's differential equation

(1 - x 2 )X" - 2xX' + AX = 0,

or, in standard Sturm-Liouville form,

d ( (1 - x 2 ) dx
dx dX) + AX = O.
Since x = cos ¢, the boundary conditions of (33) are transformed into
Xc -1) and X(l) finite.

Thus the transformed problem (33) is the well-known Singular Sturm-Liouville problem

((1-x 2 )X')'+AX=0 for -lsxS1,

XC -1) and X(l) finite,
which was solved in Section 8.14, Example 2. The results are

A = An = n(n + 1),
(34)
X(x) = Xn(x) = PnCx), n = 0, 1,2,3, . .. ,

where Pn (x) is the nth Legendre polynomial. In particular, since x = cos ¢, the eigenpairs
of (33) are

The corresponding radial functions R = Rn satisfy (32) with A = nCn + 1), so

r2R~ + 2rR~ - nCn + l)Rn = 0,

Rn(O) finite.
Solving the Euler differential equation gives the general solution

Rn(r) = kIr n + k2r- n- l ,

 9.6 STEADy-STATE DIFFUSION OF HEAT IN SOLIDS 651

where hI and h2 are arbitrary constants. For Rn(O) to be finite, we clearly must have
h2 = 0. This gives the following product solutions.
Product Solutions. The product solutions of (25), (26), (28) and (29) are the nonzero
multiples of

(35) vn(r,¢)=rnpn(Cos¢), n=0,1,2, ....

Superposition. We seek the solution of the boundary value problem (25)-(29) as a

sum of product solutions (35):
00

(36) v(r,¢) = l:>nrnPn(COS¢).

n=O

This is a formal solution, provided that

00

vO,¢)=j(¢) = LCnPn(COS¢) for 0<S¢<S7i.

n=O

The eigenfunctions <'Pn(¢) = Pn(cos¢) are orthogonal on °<s ¢ <s 7i with weight
w(¢) = sin¢:

Jo
r Pm(cos¢)Pn(cos¢) sin¢d¢ = /1
-1
Pm (x)P n (x) dx = ° for m:j:. n.

Moreover,

Jo
r P~(COS¢) sin¢d¢ = 11 P~(x) dx = _2_,
-1 2m + 1
m = 0, 1,2, ....

Thus (36) is the solution if

(37) 2n -
Cn = - + 110" j(¢)Pn(cos¢)sin¢d¢, n = 0,1,2, ....
2 0

Of course, we have only shown that (36), (37) define a formal solution, since we have not
discussed the convergence of the series and their derivatives. For rigorous results see Ual.

EXAMPLE 1. Let j(¢) = sin 2 ¢ = 1 - cos 2 ¢ = 1 - XI Since

2 2 2
j(¢) = 1 - x = -Po(x) - -P 2 (x).
3 3
652 9 BOUNDARY VALUE PROBLEMS OF MATHEMATICAL PHYSICS

Thus the corresponding solution is

2 0 2 2
v(r,¢) = -r PO(X) - -r P2 (x).
3 3

Problem 9.6.4. Sphere with Neumann Boundary Conditions, Axisymmetric Case. This
problem differs from the preceding one only in replacing the Dirichlet boundary
condition v( 1, ¢) = j(¢) by the Neumann boundary condition

(38) vr(l,¢)=j(¢) for O~¢~rr.

The product solutions remain the same and hence vCr, ¢) again has the form
00

(39) v(r,¢) = I::>nrnPn(COS¢).

n=O

Differentiating with respect to r and setting r = 1 gives

00

vr(l,¢) =j(¢) = LncnPn(COS¢).

n=O

The orthogonality of the Legendre functions implies that

(40) ~Cn = f"j(¢)Pn(cos¢) sin¢d¢ for n = 0,1,2, ....

2n +1 10
In particular, (40) with n = 0 implies that a necessary condition for the solvability of the
Neumann problem is

(41) l rrj(¢)sin¢d¢ = O.

Moreover, if (41) holds then Co may be any constant. This is consistent with the
nonuniqueness of the solutions of the Neumann problem, as stated in the Exercises of
Section 7.7.

Parallelepiped Problems. Solve the steady-state diffusion problem in the unit cube 0 =s x =s 1,0 =s Y =s 1,
o =s Z =s1 under the given boundary condition with zero Dirichlet conditions on the remaining five faces.
1. u(x,Y, 1) = sin(rrx)sin(2rry).

2. uz(x,y,1) = sin(rrx) sin(2rry).

3. uz(x,y,O) = sin(4rrx) sin(5rry).
4. uy(x, 1, z) = sin(rrx) sin(4rrz).
 9.7 THE LAPLACE TRANSFORM METHOD 653

Sphere Problems. Solve the steady-state diffusion Problem 9.6.3 in the unit sphere under the given
=
boundary condition u( 1, ¢) j(¢) without use of integration tables.

5. j(¢) = sin2 ¢.
6. j(¢) = cos2¢.
Sphere Problem Formulas. Verify the given solution to the steady-state diffusion Problem 9.6.3 in the
unit sphere under the supplied boundary condition u(l, ¢) = j(¢). Assume the identity

r 1 (_l)k(2k)1
Jo P2k+l (x)dx = 22k(kl)2(2k + 2)'
7. j(¢) = 1 for 0:::: ¢ :::: lf/2,j(¢) = -1 for lfl2 < ¢ :::: If, and for x = cos¢,

3 73 115
u(r, ¢) = '2rP1 (x) - Sr P3(X) + 16 r P5(X) - ....

8. j(¢) = 1 for 0:::: ¢ :::: lf/2,f(¢) = 0 for lf/2 < ¢ :::: Jr, and

~ + 3) (2n)1
n(4n2n+2 2n+l
u(r,¢) = f;;o(-l) 22n+l(n!)2 r P2n+l(COS¢).

9.7 The Laplace Transform Method

The major theme of Chapters 7, 8 and 9 has been applications of Fourier's method
of separation of variables to the solution of boundary value problems of mathematical
physics. The goal of this final section of Chapter 9 is to present an alternative approach,
the Laplace transform method, which can provide solutions of some boundary value
problems that are not solvable by separation of variables. These include problems of heat
diffusion and wave propagation with nonhomogeneous boundary conditions.
The Nature of the Laplace Transform Method. In Chapter 4 the Laplace
transform was applied to linear ordinary differential equations. The resulting algebraic
equations were then solved and their inverse transforms were found to solve the original
differential equations. If a solution u(x, t) of the heat or wave equation is transformed to
form

(1) u(x, s) = .c(u(x, t» = 1 00

e-stu(x, t) dt

then U will satisfy an ordinary differential equation with respect to x. The Laplace
transform method is to solve this equation, by the methods of Chapter 2 and Chapter 3,
and then find the Laplace inverse u(x, t). Here we shall present three examples of the
method. The first example, heat diffusion in an infinite rod, was solved by Fourier's
method in Problem 9.1. 7 of Section 9.1. Here the results of Problem 9.1.7 will be obtained
by the Laplace transform method. The second and third examples treat boundary value
654 9 BOUNDARY VALUE PROBLEMS OF MATHEMATICAL PHYSICS

problems for the heat and wave equations with nonhomogeneous boundary conditions-
problems not hitherto studied in this text. Many additional examples can be found in
classical texts; see, for example, [Churl.

Problem 9.7.1. Heat Diffusion in an Infinite Rod via Laplace Transform. In Section 9.1
the problem was formulated as the search for a function u(x, t) that satisfies the following
four conditions.

(2) U(x, t) is defined for -00 <x< 00 and t ::: 0,

(3) Ut = Uxx for -00 <x< 00 and t ::: 0,
(4) u(x, t) is bounded for -00 < x < 00 and t ::: 0,
(5) u(x,O) = j(x) for -00 < x< 00,

where j(x) is a prescribed bounded function.

The Transformed Problem. . The boundedness condition

(6) Iu(x, 01 ~ M for -00 <x< 00, t ::: 0,

implies that the Laplace transform U(x, s) exists for all s > a and
(7) IU(x, s)1 ~ 1 o
00
e-stlu(x, 01 dt ~ -.
M
s
The derivative rule for the Laplace transform implies that

(8) ..c (ut(x, t)) = SU(X, s) - j(x).

Moreover, we shall assume that

(9) ..c (Uxx(x, t)) = roo e-stuxx(x, 0 dt = a22 roo e-stu(x, t) dt.
10 ax 10
Hence the transform of the heat equation (3) gives

(10) sU(x, s) - j(x) = Uxx(x, s) for - 00 < x< 00.

We shall show that (10) and the boundedness condition determine U(x, s) and u(x, t)
completely
Solution of the Transformed Problem. The general solution of the ordinary differential
equation (10) is

(ll)
 9.7 THE LAPLACE TRANSFORM METHOD 655

where VI and V 2 are linearly independent solutions of the homogeneous equation

Vxx - sV = 0, and Vp is any particular solution of (10). We may take

(12)

For Vp we shall use the method of variation of parameters and the basis VI and V 2 (see
Section 3.2). The Wronskian of VI and V 2 is

(13)

Thus, from Section 3.2,

vex, s) = (CI - r e-yS~f(~) d~)

_1_
2010
eySx

r eyS~f(~) d~)
(15)
+ (C2 + _1_ e-ySx.
2010
For vex, s) to be bounded when x ~ ±oo, we must have

CI = 1Ie
2-ys
1 00
e-yS~f(~) d~,

1
0

0
C2 = 1Ie eyS~f(~) d~.
2-y s -00

Substituting these into (15) gives the bounded solution

(16)
656 9 BOUNDARY VALUE PROBLEMS OF MATHEMATICAL PHYSICS

We wish to find the inverse Laplace transform of the function

e-v'slx-~I
(17) K(s) = ..;s'
2 s
Now, standard tables of Laplace transforms give the pair

e- alt e- 2"f<iS
get) = 1-;
y rrt
G(s) = £(g) = ..;s .
Taking 2Ja = Ix - ~I, or a = (x - ~)2/4, gives
_(X_$)2
-1 e 4<
k(t) == £ (K(s)) = 1-; ,
2yrrt
which is the heat kernel of Section 9.1 (see Problem 9.1.7). Thus

u(x, t) = £-1 (U(x, s)) = £-1 ( [ : K(s)J(~) d~) ,

and if £-1 may be brought inside the integral sign, then

(18) u(x, t) =
/00J(;)e
1 _(x_,)2
~,
2yJrt -00
1-; 4<

which is the solution formula of Section 9.1, Problem 9.1.7. Of course, once equation
(18) is obtained, it may be verified by direct calculation.

Problem 9.7.2. The Semi-Infinite Rod with End Temperature Given. Consider a laterally
insulated semi-infinite rod that is at constant temperature zero at time t = 0, as in Figure
22.
If the end of the rod is controlled to have temperature J(t) at each time t ~ 0, then the
internal temperature u(x, t) will be a solution of the following boundary value problem.

(19) u(x, t) is defined for x ~ 0, t ?: 0,

(20) Ut = Uxx for x?: 0, t ?: 0,
(21) u(x,O) = 0 for x?: 0,
(22) u(O, t) = J(t) for t ~ 0,

U -_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ ~

x=o FIGURE 22. Heat diffusion in a semi-infinite rod

 9.7 THE LAPLACE TRANSFORM METHOD 657

where f(t) is a prescribed function. For simplicity, we shall add the following
supplementary conditions.

(23) u(x, t) is bounded for x ?: 0, t ?: 0,

(24) u(x, t) ---+ 0 as x ---+ 00 with t fixed.

Note that (22) and (23) imply that f(t) is bounded.

This boundary value problem is not directly solvable by the separation of variables
method. Here we shall solve it by the Laplace transform method.
The Transformed Problem. We shall write

(25) vex, s) = £(U(x, t»,

as in the preceding problem. Then, since u(x, 0) = 0, we find that V satisfies the problem
(26) V"", - sV = 0 for x?: 0, s > 0,
(27) vex, s) is bounded for s > 0,
(28) V(O,s) = £(j(t») = F(s).

Solution of the Transformed Problem. The general solution of (26) is

Vex,s) = C1e0x + C2e-0X (s> 0),

where C1 and C2 are arbitrary constants. For V to be bounded when x ---+ 00, we must
take (1 = O. Then (28) implies that (2 = F(s), so

(29) vex, s) = F(s)e- 0X .

The Inverse Transform. The convolution theorem for the Laplace transform implies
that (29) has the solution

(30) u(x, t) = £-1 (FCs)e- 0X ) = ltf(r)g(x, t - r) dr,

where

(31) g(x, t) = £-1 (e-0X) = _x_e-X2/4t.

2~
The last equation can be found in most tables of Laplace transforms; see, for example,
[ERJ. Combining (30) and (31) gives the integral formula

(32) u(x, t) = -X-

2.Jli
it
0
f(r)(t - r)-3/2 e-x 214(t-r) dr.
658 9 BOUNDARY VALUE PROBLEMS OF MATHEMATICAL PHYSICS

An alternative solution formula may be derived by changing the variable of integration

in (32) from r to z, where

so that
x2
r=t--.
4z 2
Moreover, r -+ t as Z -+ +00, and r -+ 0 as Z -+ 2~' Thus the substitution gives

(33) u(x, t) =
2
,.fii /00
2-J (t - X2) e
4z 2
-Z2
dz.
20

The boundary conditions can be verified directly from (33). In particular, U(x,O) = 0
and u(O, t) = J(t).

EXAMPLE 1. The special case that J(t) = Ua (a constant) gives the solution

This is usually written as

U(x, t) = Ua erfc (2~) ,

where erfc is the complementary error function of probability theory, defined by
2 [00 z2
erfc (w) = ,.fii lw e- dz.

Problem 9.7.3. The Semi-infinite String with End Displacement Given. As a final
example of the Laplace transform method we shall calculate the motion of a semi-infinite
string, initially at rest, when the end of the string is moved in a prescribed way The
corresponding boundary value problem asks for a string displacement function u(x, t)
that satisfies

(34) U(x, t) is defined for x ~ 0, t ~ 0,

(35) Utt = (2Uxx for x ~ 0, t ~ 0,
(36) U(x,O) = 0 and Ut(x,O) = 0 for x ~ 0,
(37) u(O, t) = J(t) for t ~ 0,
 9.7 THE LAPLACE TRANSFORM METHOD 659

where J(t) is the prescribed displacement. Of course, we must assume that u and Ux
remain small enough for the linear wave equation (35) to hold. In particular, we require
the following supplementary conditions.

(38) u(x, t) is bounded for x 2: 0, t 2: 0,

(39) u(x,t) -+ °
asx -+ 00, t 2: 0.

Condition (38) implies that the Laplace transform

vex, s) = £(u(x, t))

exists for all x 2: °and s > 0. To find V, we have

by the derivative rule of Laplace transform theory and (36). Also, we assume as before that

£ (uxx(x, t)) = Vxx(x, s).

It follows, on transforming (35) and (37), that Vex, s) satisfies

(40) Vxx - (S2/c 2)V = ° for x 2: 0,

(41) vex, s) is bounded for x 2: 0, s > 0,
(42) V(O, s) = F(s),

where F(s) = £(j). The general solution of (40) is

(43)

where Cl and C2 are arbitrary constants. Moreover, (41), (42) and (43) imply that Cl = °
and C2 = F(s), so

(44) vex, s) = F(s)e-sx!c.

Finally, (44) is easily inverted by using the second shifting theorem of Laplace transform
theory The result is

(45) u(x, t) = £-1 (F(s)e-sx!C) = J (t - ~) H (t - ~) ,

!°(
where H is the Heaviside function. Equivalently, (45) can be written as

J X)
t -- for 0::: x < ct,
u(x, t) = C
for x> ct.
660 9 BOUNDARY VALUE PROBLEMS OF MATHEMATICAL PHYSICS

Thus the prescribed motion of the end of the string generates a wave that propagates out
along the string with speed c.

IExercises 9.71
Infinite Diffusion. Consider the boundary value problem

U(x, t) is defined for -00 < x < 00 and t ~ 0,

Ut = Uxx for - 00 < x < 00, t ~ 0,
u(x, t) is bounded for -00 < x < 00 and t ~ 0,
u(x,O) = j(x) for - 00 < x < 00.

Solve by Laplace transform methods the following instances of this problem for the solution u(x, t). The
answers involve the error function

erf(x) =
2 r e-
-./7i 10 X
2
dx.

1. j(x) = x for Ixl :::: 1,f(x) = 0 otherwise.

2. j(x) = e-x for x ~ O,f(x) = 0 otherwise.

Semi-infinite Diffusion with End Temperature Given. Consider the boundary value problem

U(x, t) is defined for 0 :::: x < 00, t ~ 0,

Ut = Uxx for 0:::: x < 00, t ~ 0,
U(O, t) = j(t) , x~~ u(x, t) = 0 for t ~ 0,
u(x,O) = 0 for 0:::: x < 00.

Solve by Laplace transform methods the following instances of this problem for the solution u(x, t). The
answers will be in unevaluated integral form.

3. j(t) = sin(at), for a > o.

4. jet) = l~t.
Semi-infinite String with End Displacement Given. Consider the boundary value problem

U(x, t) is defined for 0 :::: x < 00, t ~ 0,

Ut[ = c2uxx for 0:::: x < 00, t ~ 0,
u(O, t) = j(t), lim u(x, t) = 0 for t ~ 0,
x-+oo
u(x, t) is bounded for 0 :::: x < 00 and t ~ 0,
u(x,O) = 0 and Ut(x,O) = 0 for 0:::: x < 00.

Solve by Laplace transform methods the following instances of this problem for the solution u(x, t). Graph
snapshots of the solution for c = 1,0:::: x:::: 1, t = 0, t = 0.2, t = 0.6.

5. j(t) = sin(m) for 0 :::: t :::: 1,f(t) = 0 otherwise.

6. jet) = sin2 (2m) for 0:::: t :::: 1,f(t) = 0 otherwise.

 9.7 THE LAPLACE TRANSFORM METHOD 661

7. (Semi-infinite String with Distributed Load) Use Laplace transform methods to find u(x, t) for the
modified semi-infinite string problem below. Then verify that the answer is correct.

U(x,t) is defined for 0 :::: x < 00, t ~ 0,

Utt = c2 uX)( + 1 for 0:::: x < 00, t ~ 0,
u(O,t) = 0 for t ~ o.
u(x,O) = 0 and Ut(x,O) = 0 for 0:::: x < 00.
APPENDIX
 Answers and Hints to
Selected Exercises

Section 1.2
I. Y = 0.209 + 0.52x,y = 0.443 + 0.26x
3. Y = -0.7 + 2.2x, y = 6 - 4.5x
5. ao = 0.469, al = 0.52, a2 = -0.3714285713
7. ao = 0.4, al = 2.2, a2 = -6.7.
9. Y = -15000x + 149000, y = 7000(x - 1) + 134000, y = 23000(x - 2) + 141000.
II. Check that the quadratics interpolate the data points

(0,149000), 0,134000), (2,141000), (3,164000).

Then show that the polynomials can be pasted together to make a continuously differentiable function
f withj'(O) = o.
13. Choose Xo = 0, then (x - xo)(x - Xl)(X - X2)(X - X3) = h4 t(t - l)(t - 2)(t - 3) has its value of
largest magnitude on 0 :s t :s 3 at some point t (t = xlh). The possible values of t are ~, ~ + 1,
~ - 1. The largest magnitude is I - 11·
21. Item (b) follows by changing variable in the integral, u = -v, du = -dv.
27. Wild oscillation at the ends of the interval.
29. The proof is by induction. Given Zo and the relations mk = C(fk+l-ikl), Zk+l
Xk+l-Xk
= 2mk - Zk, for
k = 0, 1, ... , n - 1, it suffices to prove the relation Zk = F(k) where F(k) = 2 L~~~ (_l)k-l-i mi +
(-l)k zo .

31. P2 (x) = 0.2229166666x 2 - 0.1266666665x + 0.5454374999 and P2 (l) = 0.6416875.

33. Approximate 0.9876471951, exact 0.9876569764. At least 5 digits.
664 ANSWERS AND HINTS TO SELECTED EXERCISES

35. Let j(x) = (x - xo)(x - Xl). The largest value of If(x) I is If((XQ + xl)l2)1 = (Xl - xo)2/4.
Section 1.3
1. 0.170825, exact 0.1666666667.
3. 0.3858779367, exact 2In(2) - 1 = 0.386294361.
5. 0.1667000000, exact 1/6 = 0.1666666667.
7. 0.9095424082, exact 3 In(3) - 1 - 2 In(2) = 0.909542506.
9. n = 17.
11. 1.999835504, exact answer 2.
l3. 25 divisions. Table of erf(O.I) to erfO):

0.112463 0.880204 0.997020

0.222702 0.9103l3 0.998137
0.328626 0.934007 0.998857
0.428391 0.952285 0.999311
0.520499 0.966105 0.999593
0.603855 0.976348 0.999764
0.677800 0.983790 0.999866
0.742100 0.989090 0.999925
0.796907 0.992790 0.999959
0.842700 0.995322 0.999978

15. 6 divisions or m = 3.
8x 2 - 4 2
19. j"(x) = -/ii e- x , n = 20 in trapeZOidal rule. The estimate is 0.2226989755 for exact erf(0.2) =
0.2227025892, error 0.0000036137.
23. Definej(x) = ]0 (x). Let C(h) = (f0 + h) - jO - h»/(2h), R = (4C(hl2) - C(h»/3 (Richardson
extrapolation). The computations, using a hand calculator:

Co = C(0.4) = jO.4) - j(0.6) = -0.4314371789,

0.8
Cl = C(0.2) = j(1.2) - j(0.8) = -0.4378865213,
0.4
4Cl - Co
Rl = 3 = -0.4400363019,

C2 = C(O.I) = j(Ll) - j(0.9) = -0.4395088980,

0.2
4C2 - Cl
R2 = 3 = -0.4400496901,

C3 = C(0.05) = j(1.05) - j(0.95) = -0.4399151220,

0.1
4C3 - C2
R3 = 3 = -0.4400505299.

Section 1.4
1. See Figure 1.
 ANSWERS AND HINTS TO SELECTED EXERCISES 665

-0.5 4 FIGURE 1. Graphical solution to X2 + 1 = 3x

3. See Figure 2.

5. 0.3819671633 in 15 iterations on [0.3,0.4]. Exact 0.3819660113.

7. -0.6823310855 in 17 iterations on [-1, -0.5]. Exact -0.6823278038.

9. 0.2679466751 in 5 iterations on [0.2,0.3]. Exact 0.2679491924.

11. 0.1984380936 in 3 iterations on [0.1,0.3]. Exact 0.1984372145.

l3. 0.198437215 in 2 iterations on [0.1,0.3]. Exact 0.1984372145.

15. 0.6180357143 in 2 iterations on [0.6,0.7]. Exact 0.6180339887.

19. 0.5109649660; exact for 1 - sin x = x is 0.5109734294.

21. Use the definition oflogarithm.

25. The derivative off is positive, because l' (x) is the sum of two positive terms (except at x = a). Apply
the previous exercise.

o FIGURE 2. Graphical solution to 1 - x = x3

666 ANSWERS AND HINTS TO SELECTED EXERCISES

29.

Xo 1.000000 100.000000 1000000000 0.00250000001

Xl 0.694400 0.693160 0.693147 19.336971

X2 0.694952 0.694955 0.694955 0.693212

X3 0.694951 0.694951 0.694951 0.694955

X4 0.694951 0.694951 0.694951 0.694951

Section 1.5

1.

Y X Y x

1.00 0.0 1.36 0.6

1.01 0.1 1.49 0.7
1.04 0.2 1.64 0.8
1.09 0.3 1.81 0.9
1.16 0.4 2.00 1.0
1.25 0.5

3.

Y x Y x

0.50000000 1.0 0.71428571 0.4

0.52631579 0.9 0.76923077 0.3
0.55555556 0.8 0.83333333 0.2
0.58823529 0.7 0.90909091 0.1
0.62500000 0.6 1.00000000 0.0
0.66666667 0.5

5.

Y x Y x

0.00000000 0.0 0.77459667 0.6

0.31622777 0.1 0.83666003 0.7
0.44721360 0.2 0.89442719 0.8
0.54772256 0.3 0.94868330 0.9
0.63245553 0.4 1.00000000 1.0
0.70710678 0.5
 ANSWERS AND HINTS TO SELECTED EXERCISES 667

7. No, the inverse table is not a function.

x y x y

0.0 0.00 1.2 0.24

0.2 -0.16 1.4 0.56
0.4 -0.24 1.6 0.96
0.6 -0.24 1.8 1.44
0.8 -0.16 2.0 0.20
1.0 0.00

11.

y = 0.00 xo = 0.00000000 x = 0.00000000

y = 7f/4 xo = 0.00000000 x = 0.95299367

Y = 37f/4 xo = 0.95299367 x = 2.4821593

y=7f xo = 2.4821593 x = 3.1415927

Y = 57f/4 xo = 3.1415927 x = 3.801026

Y = 37f/2 xo = 3.8010260 x = 4.5109464

Y = 77f/4 xo = 4.5109464 x = 5.3301916

Y = 27f xo = 5.3301916 x = 6.2831853

668 ANSWERS AND HINTS TO SELECTED EXERCISES

13.

Y = 0.00 xo = 0.00000000 x = 0.00000000

Y = rr/4 xo = 0.00000000 x = 0.99374338
Y = 3rr/4 xo = 0.99374338 x = 2.5041474

y=rr xo = 2.5041474 x = 3.1415927

Y = 5rr/4 xo = 3.1415927 x = 3.7790379
Y = 3rr12 xo = 3.7790379 x = 4.4709969
Y = 7rr/4 xo = 4.4709969 x = 5.2894419
Y = 2rr xo = 5.2894419 x = 6.2831853

15. Use a2 = [2 + b2 and a ~ [ (by geometry).

Section 1.6
1. Explicit as written.
3. Implicit. Can be converted to explicit.
5. No, because Fx = Fy = 0 at x = 0, y = 1.
7. Initial x = 1.00, YO = 0.6 obtained graphically.

x = 1.00 Yo = 0.60000000 Y = 0.65963921

x = 0.75 Yo = 0.65963921 Y = 0.86042689

x = 0.50 YO = 0.86042689 y=l

x = 0.25 YO = 1.00000000 Y = 1.079327

x = 0.00 Yo = 1.079327 Y = 1.1049876

9. Substitute the polar coordinate equations.

dy -FxCx,y)
11. dx Fy (x, y)

13. FyCO,O) = 1.
15. FyCO,O) = 2.
 ANSWERS AND HINTS TO SELECTED EXERCISES 669

17. Gy(x,O) = -x/4 violates Gy ::: L < 1 when x> 4.

19. A most interesting proof can be found in the classical analysis text of Dieudonne.
21. To = Qlh = lOQ/(lOh) = IOQlx.

Section 1.7

1. N = 66.

3. N=4.

5. N = 26.

Section 2.1
[24 24
1. Typically, 10 (1 + 20 dt = (t + (2 )lo = 600.
3.
1
i] = - (e- I +sint-cost),
2

1
i2 = - (11e- 1 + sm t - cos t); see Figure 3.
2
5. The cooling equation has formal solution u(O = 20 + SOc-hI because in this setting to = 0, T = u(t),
TO = 100. The additional condition, eg., u(10) = 60, determines h uniquely as h = In(2)/l0 The
cooling time t is the solution of 21 = 20 + SOe-ht for the given value of h.

7. To check the initial condition, set x = 0 to obtain yeO) = eO +0 2 = 1. To check the differential equation,
differentiate y to obtain y' = - 3e- 3x + 2x.

9. x = 0 and x = 2.
11. See Figure 4.

13. See Figure 5.

15. See Figure 6.

17. Isoclines are horizontal lines x = 3/2 ± V0.25 + c, when c takes on values c 2: -1/4.
Section 2.2

1. x 3/3 = 1/3 + t212, x = VI3]I + ~t2

3. In(x + 3) = InO) + In(t + 5) - In(5), x = 3tl5.
S. ~ arctan((x - 3)12) = t, x = 3 + 2 tan(2t)

\
x
FIGURE 3. Plot of hand i2
610 ANSWERS AND HINTS TO SELECTED EXERCISES

I / /
/ /

-1

o FIGURE 4. Direction field for x' = t + x

1.0

,
\

0.1
o 1 FIGURE 5. Direction field for XXi +t = 0

7. -In 1 - eX = t - In 1 - e, x(t) = In(e - l)e + 1

9. x(t) = ~(t + 2et )

11. xCt) = et

13. x(t) = In(l + tiS)

1
1 /

-1

o FIGURE 6. Direction field for x' = x3

 ANSWERS AND HINTS TO SELECTED EXERCISES 671

15. x(t) = 3 + Insect

17. x(t) = 1 - t + (1 + t) InO + t)
19. x(t) = fd (1 + t 3 )1/2 dt

21. The plotted data points are (tk, Vk), k = 0 to k = 24. The recursion is Vv+1 - Vk = (%+1 - %)12
(k = 0 to 23), where the tk's and qk'S are given in the flow meter table.
23. 0.0008842 inches.
25. 135.12125 minutes.
27. Because a half-full tank empties at lower velocity, it takes longer to empty a half-full tank. Il
523.94332 seconds, I2 = 1788.8544 seconds.
29. t = 19.4334 by the trapeZOidal rule, t = 19.4334 by Simpson's rule.
31. 69.44 days.
Section 2.3
1. x(t) = xoe-Aot.
3. x(t) = e-1+C05(t)
5. x(t) = e-1+e- t

7. x(t) = 5/.
9. x(t) = t2 + 2.

t5 1
11. x(t) = "9 - 9t 4 '

1
13. x(t) = sin t + - (cos t - cos 1 - sin 1).
t
15. x(t) = t - 2 tanet) sec(t)

17. pet) = p(0)(l/2)U17500

19. 1600 years.

21. 30In(0.0l)lln(0.5) = 199.32 minutes or 3.32 hours.
23. q(t) = 5e- 10t

25. vet) = ~ (e- lOt + 10sint - cost).

101
27. The amount from cominuous interest is 100e4 = 5459.82 to the nearest cent. The amount from
compound interest is 100(1 + 0.08/n)50n The value of n is much greater than 365 (for n = 400, the
compound interest is 5457.63)
Section 2.4
1. xy=c.
3. xy + eX = c.
5. tan-l(xly) = c.

7. x 3y + 4x 2y2 + 4y 3 = c.
672 ANSWERS AND HINTS TO SELECTED EXERCISES

9. x sin(y) = c.

2
11. )1-x2 +L = c.
2
y4
13. xy3 + y3 - - = c.
4
15. (y+x)e'Y=c.

17. x 2y+y= 1.

19. xy sin (x) +y = 3(rr/2 + 1).

23. /L(t) = e- Ja(t)dt.
25. Since Uy 01 0, we can solve for y in terms of x, yO) = 1. Since Ux 01 0, we can solve for x in terms of
y, x(l) = 2.

Section 2.5

1. xCt) = Cjt + CI
3. x(t) = Cjt + (2 + t2 /2 + t3 /6.

5. x(t) = Cje- 2t + (2.

7. x(t) = Cjet + CI - t - t2 /2.

9. xCt) = Cj InCt) + C2.

dv
11. Vdy =0,v=Cj,xCt)=Cjt+C2·

13. v :; = y, v 2 = i + c, xCt) = Cje t + CIe- t

15. v:; = -4y, v2 = -4i + c, x(t) = Cj cos(2t) + (2 sin(2t).

~
17. v dy = -2v, v + 2y = C, or v = 0, xCt) = Cj + CIe- li .

19. 2yv dv = 1 +v 2 , 1 +v 2 = kyo

dy

21. ex =
~
yk 2 +4f.g, vo
-k - ex
= - - , VM =
-k +ex
- - 0- , v= VOVM
exte"t-1 ,J
-1
(t) = va +
VO(VO - VM)
e"t '

i
U h ~ -~ ~ -~
t -1 ( )d
J t t = vat + voCvo - VM) In VM - voe-at . I I
a voex VM - vo

gt 2 b
23. x(t) = Re + bt - - + -(M - at) In 11 - aUMI·
2 a
1
25. y(x) = 1 + .
eX 2- x (I; _tet2 - t dt + C)
 ANSWERS AND HINTS TO SELECTED EXERCISES 673

5 ( 2g - 0.23 )
27. t) = l(v)) = - ( ) = 0.1568155886. For the others, employ Simpson's rule, for
2 g g - 0.23

example, t2 = l(v2) = ~ (312 + 32_40.23 + 32~0.4) = 0.314667519, t3 = l(v3) = t) +

5( 1
3
4
32 - 0.23 + 32 - 0.4 + 32 _ 0.64
1) = 0.4733927036.

Section 2.6

L Define D to be the region -00 < t < 00, X < L The solution sin t exits D at t = ±nI2, at which time
x' (0 ::: 0, so x(o = 1 for t > n12.

3. The issue is the continuity of the derivative x' (1) at the points t = a and t = b.

Section 2.7

L The slope field is F(t,x) =x.Usegrid(x,y),x = 0.2Sn,y = 0.2Sm+l,nandmfrom-2t02instepsofi.

3. The slope field is F(t,x) = x(l - x) Use grid (-2h,y), (-h,y), (O,y), (h,y), (2h,y) for y = 0.5 + 2h,
O.S + h, 0.5, 0.5 - h, 0.5 - 2h and h = 0.25.

S. The exact solution is x(t) = et .

x Exact

0
0.1 1.1 1.10S17
0.2 1.21 1.2214
0.3 1.331 1.34986
0.4 1.4641 1.49182

The error is about 2 percent.

7. The exact solution is x(t) = 1/(1 + e- t ).

x Exact

0 0.5 0.5
0.1 0.525 0.52498
0.2 0.5499 0.54983
0.3 0.57469 0.57444
0.4 0.59913 0.59869

The error is about 1 percent.

9. The exact solution is xU) = -1 - t + 2et .

x Exact

0 1.0 1.0
0.1 1.11 1.1103
0.2 1.24205 1.2428
0.3 1.39847 1.3997
0.4 1.58180 1.5836

The error is about 0.1 percent.

674 ANSWERS AND HINTS TO SELECTED EXERCISES

11. The exact solution is x(t) = 1/(1 + e- t )

x Exact

0 0.5 0.5
0.1 0.525 0.52498
0.2 0.54988 0.54983
0.3 0.57450 0.57444
0.4 0.59877 0.59869

The error is about 0.01 percent.

13. The exact solution is x(t) = -1 - t + 2e t .

x Exact

o l. 00000 l. 0
0.2 1.2428 1.2428
0.4 l.58364 l.5836

The error is about 0.002 percent.

15. The exact solution is x(t) = 11(1 + e- r).

x Exact

0 0.5 0.5
0.2 0.54983 0.54983
0.4 0.598686 0.59869

The error is about 0.0005 percent.

17. The exact solution is x(t) = et .

x Exact

0
0.2 1.22 1.2214
0.4 1.4884 1.49182

The error is about 0.2 percent.

19. The exact solution is x(t) = 1/(1 + e- t )

x Exact

0 0.5 0.5
0.2 0.54975 0.54983
0.4 0.59852 0.59869

The error is about 0.03 percent.

 ANSWERS AND HINTS TO SELECTED EXERCISES 675

21. The exact solution is x(t) = e1.

x Exact

o
0.4 1.49173 1.49182

The error is about 0.0006 percent.

23. The exact solution is x(t) = 1/0 + e- t ).
x Exact

o 0.5 0.5
0.4 0.598686 0.59869

The error is about 0.0006 percent.

25. The exact value of xCt) = 1/0 + e- 1) at t = 1 is xCl) = 0.7310585787. The results are as follows,
where the last column of the table is xO) - XN, and XN is in column 2.

N Euler x(l) Error

4 0.7368351728 -0.0057765941
64 0.7314271978 -00003686183
256 0.7311508052 -0.0000922263
4096 0.7310643445 -0.0000057653

27. The exact value of x(t) = 1/(1 + e- t ) at t = I is xO) = 0.7310585787. The results are as follows,
where the last column of the table IS xO) - XN, and XN is in column 2.

N Taylor4 x(l) Error

0.7291666667 0.0018919127
2 0.7309788031 7.97757e-05
4 0.7310545273 4.0517e-06
8 0.7310583515 2.277e-07

29. The exact value of x(1) = 1/0 + e- 1) at t = 1 is xO) = 0.7310585787. The results are as follows,
where the last column of the table is xO) - XN, and XN is in column 2.

N Rk4 x(l) Error

0.7309103404 0.0001482387
2 0.7310474077 1.1 1717e-05
4 0.7310578607 7.187e-07

31. The exact answer is f;/2 cos(tJdt = 1. The results for the four methods appear in the table below.

Exact 1. 000000000
Simpson 1.000003392
Euler 1.076482803
Heun 0.9979429868
Rk4 1000000212
676 ANSWERS AND HINTS TO SELECTED EXERCISES

Section 3.1

1. q" + 100q' = 5, q(O) = q'(O) = o.

3. q" + 100q' + 10q = 5, q(O) = 0, q'(O) = o.

5. q" + 10q = 5, q(O) = 0.05, q'(O) = o.

7. lOx" + 2x = 0, x(O) = 0, x'(O) = 1.

9. 0" + 10- 2 0 = 0,0(0) = 1,0'(0) = 1.

11. 0" + 30' + 20 = lOsin(wt), 0(0) = 0, 0'(0) = O.

32
13. O.Olx" + lOx = 0, x(O) = 7(rrI180), x'(O) = O.
15. (";xu')' - x 2 sin(x)u = O.
17. The derivation depends on formulas dyldx = R' (r)lw and d2yldx 2 = R" (r)/w 2 .
19. Because of the product rule applied to the first term.
21. It suffices by linearity to check that e- t , e- 2t , and e- 3t are solutions.

23. By Kirchhoff's laws, ij + i2 + i3 = i. Drop formulas give ij = CLi~ = Cu", i2 = u'IR, i3 = U/L
1 1
25. Ljq~ = --qj + R(i2 - ij), L2q~ = --q2 + R(ij - i2). The isomorphism is
Cj C2
(mj,m2,kj,k2,C) --+ (Lj,L2, 1/Cj, 1/C2,R).

Section 3.2

1. W = -2.
3. W = cosh2(t - to) - sinh 2(t - to) = 1.
5. W = _2e- 3tJ2

7. W = _3et2 /4 .

9. X2 (t) = sin t.
11. x2(t)=tlnt.

13. X2(t) = 1.
15. X2(t) = -1 + -tin
1 11-+-t I.
2 1- t
4
17. xp(t) = _t 7l2 et
35
1
19. xp(t) = '2t2(ln t)2 - t 2 ln t.

1
21. xp(t) = -'2tcost.

300 + 30.jl49
23. t = 49 = 13.59585.
 ANSWERS AND HINTS TO SELECTED ExERCISES 677

Section 3.3
1. et , e2t .
3. e3tJ2 cos(J9lU2), e3t12 sin(J9lU2).
5. e3t11O cos(UlO), e3tJlO sin(UIO).
7. e- 2t15 , te- 2t15 .
3 4 1
9. xp(t) = - - cos(2t) +- sin(2t) + -.
50 50 2
1 t
11. xpW = -"2e- .

1 2t
13. Xp(t) = -te + (19
- - 3-t - -t
1 2) e.t
5 32 8 4
15. Q = CI cos(50t) + C2 sin(50t).
17. Q = CI cosCU20) + C2 sin(U20).
19. i(t) = -lOOe- t sin(20t).
kAo(k - mw2 ) -kAocw
21. Al (w) = ,A2(W) = .
(k - mw 2 )2 +
c2 w 2 (k - mw2 )2 c2w 2 +
k 2A2
23. A(w)2 =
(k - mw)
2 20
+c
2 2' A' = 0 when -[2(k - mw2 )( -2mw)
w
+ 2c2wl = 0, so at a rest point
4m 2 k 2A2
A(w)2 = 2 O2 ,but otherwise the maximum is at w = O. ThenAmax = Ao for 2km - c2 ::'S 0;
c (4km - c )
2mk
otherwise Amax = Ao . Resonance occurs at those frequencies w that maximize the
cJ4km - c2
amplitude A(w).
25. See Figure 7.
27. Let the roots be)..2 = -all + ../B12, )..2 = -al2 - ../B12, where D = a2 - 4b =discriminant. Then
(D> 0) W = be-at, (D = 0) W = e- at , (D < 0) W = Jb - a 2/4e- at .

29. See Figure 8.

31. The characteristic equations)..2 + 16)" + 64 = 0 and)..2 + 15)" + 64 = 0 have roots -8, -8, and
-1512 ± i.J3I/2, respectively.

33. LetfI(t) = eC- 33 +v'6s)t14 + /-33-v'6s)t14,h(t) = eC- I7+J33)t12 + /-17-J33)t12. They cross at
t = 0.1230872708. See Figure 9.

0.35

0.00
o FIGURE 7. Voltage Vet)
678 ANSWERS AND HINTS TO SELECTED EXERCISES

60

° -2 o 2 FIGURE 8. Plot of e- I + e- 21

35. T = In(100) = 4.61. It crosses once,

Section 3.4

1. co = 1, Cn = ° for n > 0, x(t) = 1.

3. ck+2 = -Ck+l/(k + 2), Co = 0, Cj = 1. Then Cn = (_l)n-l Inl for n ~ 1 and x(t) = 1 - e- t ,
5. ck+2 = q/[(k + l)(k + 2)], cO = 1, Cj = 0, Then Cn = lin! for n ~ 1 and n even, zero otherwise, and
x(t) = (e t + e- t )/2,
7. ck+2 = q/[(k + l)(k + 2)], Co = 0, Cj = 1. Then Cn = lin! for n ~ 1 and n odd, zero otherwise, and
x(t) = sinh(1) = (el - e- t )l2,
9. q+2 = -Ckl(k + 2), Co = 0, Cj = 1, and x(t) = t - t 3 /3 + t 5/lS - t 7/lOS + ... ,
11. ck+2 = -Ck-2/[(k + l)(k + 2)], Co = 1, CI = 0, C2 = 0, C3 = 0, and x(t) = 1 - t4 /12 + .. "

13. Ck+2 = ck-l/[(k + l)(k + 2)], CO = 1, Cj = 0, C2 = 0, and x(t) = 1 + t 3 /6 + t 6 /180 + ".,

15. Xl (t) = 1, x2 CO = 1.
17. XI (t) = 1, X2 CO = 1- e-1. For x2, co = 0, Cj = 1, ck+2 = -Ck+l/(k + 2), Then Cn = (-l)n+l/n!
for n ~ 1.
19. XI (t) = 1- t 3/6+ t 6/180 - " ' , X2(t) = t - t 4 /l2 + t 7/504 - ... , Use ck+2 = -ck-l/[(k + l)(k+ 2)]
°
and Co = 1, Cj = C2 = for Xl, similar for X2, Exact solutions involve Bessel functions, too complicated
to cite,

21. xI (t) = 1, X2(t) = t.

23. XI (t) = 1, X2 (I) = t2

°° FIGURE 9. Plots off1 andf2

 ANSWERS AND HINTS TO SELECTED EXERCISES 619

25. Xl (t) = t l/2 , x2 (t) = t l/2 In(t).

27. Xl(t) = t 3/2 ,X2(t) = t 1l2

29. Roots 1 and O. Xl (t) = cO, X2 (t) = qt.

31. Roots 0 and 2. Xl (t) = cO, X2 (t) = qt 2. Recursion relation is k(k - 2)ck = 0 for the smaller root.
33. Roots -3 and -2.Xl (t) = co L~o t 2k - 2/(2k + 1)! = cot- 3 sinh(t).x2(t) = Co L~O t 2n - 3/(2n)! =
cot- 3 cosh(t).

35. Roots 0 and 2. Xl (t) = cot2 L~o( _1) n t 4n +2/(2n + 1)!. Recursion Cn = -4cn-4/[n(n + 2»), Co =I 0,
q = c2 = C3 = O. X2(t) = cOt2 L~0(_1)nt4nl(2n)!. Recursion Cn = -4cn-4/[n(n - 2)], Co =I 0,
q = Q = c3 = 0 (Q =I 0 reproduces the solution Xl with C2 replacing co).

37. The integrand is smooth; differentiate under the integral sign. Then apply integration by parts with
u = sin(t sin 0), du = t cos(t sin 0) cos 0 dO.

39. The integrands upon addition collect with factor [1 - cos 2 0 - sin 2 0] = o.
41. Expand both sides into a series, using term-by-term differentiation on the left. The two series are equal,
by making a change of variable k = m + 1 in the series on the right.

43. The terms (nlx)]n(X) and -(nlx)]n(x) cancel upon addition.

45. Truncate to L~~o( _1)k(xl2) 2k l(k!)2 for plotting, using Taylor estimates on 0.1 ~ X ~ 10. Plot with
a hand calculator or a computer algebra system.

47. When 11- = 0, then 0 - x2)y" - 2xy' = O. The roots are 0 and O. The recursion cn+! = -n(n +
1)cnl[2(n + 1)2] leads to y(x) = Co. Choose Co = 1 to make yO) = 1. When 11- = 1, then
(1- x2)y" - 2xy' + 2y = O. Theny(x) = Co + q (x - l);yO) = 1 gives q = Co = 1 andy = x.

Section 3.5

1. x(t) = t 3/6, 0 ~ t ~ 1, x(t) = t 2/2 - t 4/12 - t/6 - 1112, 1 ~ t ~ 3.

3. x(1) =t- sint, 0 ~ t ~ n:,x(t) = t 2/2 +sint+ (3 -n:)t+n: 2/2 - 2n:, n: ~ t ~ 2n:.

5. x(2) = 1.5 + f12(2 - s)j(s)ds = 2.124 by the trapezoidal rule.

7. x(2) = 1.5 + f12(2 - s)j(s)ds = 1.9086 by the trapezoidal rule.
9. Exact is X = cos 4t.

k tk Xh Yk Ek E'
k Exact

0 0 0 0 0
0.1 -1.6 -0.0789 0.0423 0.92106099
2 0.2 0.84 -3.2 -0.1433 0.3306 0.69670671
3 0.3 0.52 -4.914 -0.1576 0.8158 0.36235775
4 0.4 0.0656 -5.376 -0.0948 1.3777 -0.029199522
680 ANSWERS AND HINTS TO SELECTED EXERCISES

11. Exact is x = t 2 e- tn.

h tk xh Yh Eh E'
k Exact

0 0 0 0 0 0 0
0.1 0 0.1 0.00452 -0.01404 0.004524
2 0.2 0.01 0.17048 0.00637 -0.02311 0.016375
3 0.3 0.027048 0.21726 0.006288 -0.02835 0.033337
4 0.4 0.04877 0.245185 0.00485 -0.03068 0.053626

13. Exact isx = sin(4t)/4.

h tk xk Yk Ek E'
k Exact

0 0 0 0 0 0
0.1 0.1 -0.0265 -0.07894 0.09735
2 0.2 0.2 0.84 -0.02066 -0.14329 0.17934
3 0.3 0.284 0.52 -0.05099 -0.15764 0.23300
4 0.4 0.336 0.0656 -0.08611 -0.09480 0.24989

15. Exact is x = sinh t - t.

h tk xk Yk Ek E'
k Exact

0 0 0 0 0 0 0
1 0.1 0 0 0.000167 0.005 0.000167
2 0.2 0 0.010017 0.001336 0.01005 0.001336
3 0.3 0.001 0.03015 0.003520 0.01519 0.004520
4 0.4 0.004017 0.060602 0.006735 0.62047 0.010752

17. Exact is x = t - sin t.

h tk xk Yk Mean error

0 0 0 0 0
0.2 0.00133 0.019933 0
2 0.4 0.01058 0.078939 0

19. Exact is x = 10 - 10 cos t.

h tk xk Yk Mean error

0 0 0 0 0
0.2 0.19933 1.98669 0.0000115
2 0.4 0.789389 3.894186 0.000025
 ANSWERS AND HINTS TO SELECTED EXERCISES 681

21. Exact is x = t4/12.

k tk xk Yk Mean error

0 0 0 0 0
0.2 0.000133 0.0026667 0
2 0.4 0.002133 0.0213333 0

23. Exactx = cos 4t. Data points (0, 0), (0.1, -0.0789), (0.2, -0.14329), (0.3, -0.15764), (0.4, -0.0948)
25. Exact x = t2 e- t /2. Data points (0,0), (0.1,0.0045), (0.2,0.0064), (0.3,0.0063), (0.4,0.0049).
27. q(t) = 10-10costonO ~ t ~ 10,q(t) = 100-coslO)cos(t-lOHlOsinlOsin(t-lO)ont > 10.
29. x(t) = 0 for 0 ~ t ~ to, xU) = (t - to)4/108 for t > to defines infinitely many solutions.
31. Use to = 0, Xo = 0, Yo = 1, Fl = y, F2 = -x, h = 77:/5 for 5 steps. The recursion starts with
Xl = h - h 3 /6, Yl = 1 - h2 /2 + h4/24

33. Xl = 0.004424998264, exact is 0.0044167009; Yl = 0.08321660418, exact is 0.0832582469; Zl =

0.6942544095, exact is 0.69940463519.
Section 3.6
1. X = qe- t + (2e-2t + (3e-3t
3. X = q e- t + e- t ((2 cos t + (3 sin I).

5. x = qeU2 + Qe- t+,/6t + (3e-t-,/6t.

7. x = qe67t/50 + qe t cos(../3o + (3et sin(../30.

9. x = C] e7U3 + (2te7U3 + (3e,,)3t + (4e-)3t.

11. x = ~eU2 _ v'6 + 2 e-t+vI6t + v'6 - 2 e-t-vI6t

3 12 12
13. x = 6e- t - 8e- 2t + 3e- 3t
15. k(t) = ~e-t - e- 2t + ~e-)t
2 2
17. k(t) = e-tO - cos t).
67U50 t h t h 2500 -2500 -850J3
19. x=qe +Qe cos(v3t)+(3e sin(v3t),q = --,C2 = - - , C 3 = - - - .
7789 7789 23367
r> r> -189 9 189 57
21. x = cle7U3 + C2t/t13 + (3e",3t + (4e-",3t, Cl C C
= 242' 2 = 22' 3 = 484
+../3
242 '
189 57
q= - - -../3.
484 242
23. x=t 3/6
1
25. x = cos t + 2t2 - 1.

t5
27. x = 6000'

t6
29. x = 720'
682 ANSWERS AND HINTS TO SELECTED EXERCISES

FIGURE 10. Plot of J(t) = H(t - ]f)

1 1 3 1
31. x = _e- t + -tet - _et + -(cost + sint).
8 4 8 4
t3
33. x = q + C2 t + C3 t2 + 6.
2 t5
35. x = q +C2 t + Clt + 6000.
37. x=qet+Cle-t+c3cost+qsint-1.
Section 4.1
1. See Figure 10.
3. See Figure U.

5. Use Jooo = Jt +J::'.

7. Use Jooo = J~ + Jaoo . The integrand is KePt with K = e- a and fJ = 1 - s.
9. J(t) = e- 7t

ll. J(t) = H(t) + H(t - 2).

13. J(t) = et- 2 H(t - 2).

Section 4.2
1. 2e-s~ (a cos at + s sin at) from t = 0 to t = 00.
s +a
-st . -at -2as -st a2 _5 2
3. e (A cos at + B Sin at), A = s2+a2 + (s2+a2)2' B = s2+a2 + (s2+a2)2 from t = 0 to t = 00.

-Xt
5. ;;f+b 2 (b cos bt + X sin bt), X = s - a, from t = 0 to t = 00.

7. Use sinh at = ieat - ie-at.

2 FIGURE 11. Plot of J(t) = H(t - 1) - H(t - 2)

 ANSWERS AND HINTS TO SELECTED EXERCISES 683

9. After substitution, t~ = A + (O)B, which gives A = -1.

11. A = -1/6, B = -1/2, C = 213.
13. j(t) = ~ cos 2t - s'j sin 2t - ~e-7t
15. j(t) = _27e- 3t - 27te- 3t + 28e- 2t - 8te- 2t

17. j(t) = e2t cos 2t.

19. jet) = A60c7UlO + 12{oe-8U10

21. jet) = et sin 2t + tet cos 2t.
n. j(t) = (iet - ie-t) sin t = sinh t sin 1.
25. The steps use linearity, formula (11), series rules, and the sum formula for geometric series.

27. Use fooo = fg + faoo and integration by parts.

29. Use repeatedly the relation that precedes the one to be proved.
Section 4.3
1. i(t) = 5 - 4e- 2U3
3. i(t) = H(t - 3)i¥ (e- t - e- lOt +27 ).
5. i(t) = lOe- UIO - H(t - 11110) lOe-UIO+1lI100

7. y=25(1-cos2t).

9. y = i - i cos( ht) - H(t - 1) 0- i cos h(t - 1)), Y55 = i (cos h(t - 1) - cos h t)
1
11. (5-5)2'

6 I
13. (5-1)4 - (s+ 112)2 .

,,10 (-I)"nl
15. L...n=O (5-1)"+1 .

17. j(t) = 2te- 2t .

19. jet) = et sin 2t.

21. jet) = A sin 2t + b cos 2t.
23. jet) = et cos ~t.

25. jet) = e- t (~ cos(t! vis) + 3f sin(t! vis»)

31.
e- 2s
2
(L +
5-1
e- 2 )
5+1 .

33. jet) = H(t - 2)(t - 2)

35. jet) = H(t - l)e t - 1

37. jCt) = iet sin(2t - 2)H(t - 1)

684 ANSWERS AND HINTS TO SELECTED EXERCISES

39. jet) = 9 - ge- t .

41. j(l) = It + $i~~4t.

43. jet) = 1 - cost + sint - t.
45. The equation transforms to s(1 - s)C(y) + (l - s + N)C(y) O. Partial fraction theory implies
l-s+N = N+I +.J'L hence the result.
s(I-s) s 1-5'

47. fC t) -- e- 2'_e- 3'

t .
4
49. fi'

51. e~'" C2~w2)'

53. h(t) = rt,e4t - rt, - ~.
55. Y = jSint(l-cos3 t) - jcostsin 3 tory = ~sint- ~sin2t.

57. y = is sin t - to sin 4t for 0 S t S n, y = -fa sin 4t otherwise.

59. y = 2et.
61. The extra term arises because of boundary term evaluation in integration by parts, applied to the
improper integrals over [0, to) and (to, 00).
63. Fubini's iterated integral interchange theorem.

k ke- PS
67. ps2 - s(l-e-PS)'

69. Write OJ) = i (L~l a" ("~p + j(O») and compare with C(j').
+1 t 00 f(n) (0) ,
71. Since C(eatt") = n!/(s - a)n ,the transform of ea jet) is the series Ln=O -n!- (s_:jn+l . Compare
this series to the series for C(j).

Section 4.4
1. x(t) = 200e U20

3. x(t) = 2e- U2 cos t.

5. x(t) = 2e- lOt sin lOt.
7. x(t) = -ie-t + ~e-2t + tV.
9. x(t) = tet +2 - 3e t + e2t.
II. xCt) = ie- 3t - e- 2t + ie-t.
13. xCt) = e- t - e- t cos t.

15. xCt) = -Jre-7U2 - ~tet + Met

17. x(1) = 2500 e67USO - ~etsinCJ3t) - 2S00etcoscJ3t).
7789 7789v'3 7789

19 . xCt) -- -Lte7U3
22
189 elU3
_ ill + ....2...
242
(63+38v'3 ev'3t + 63-38v'3 cv'3t)
6 6 .
 ANSWERS AND HINTS TO SELECTED EXERCISES 685

FIGURE 12. LRC circuit transient current

21. Q(t) = ~ sin SOt + Q(O) cos SOt

23. i(t) = -WOe-lOt sin(20t), plot on [0, 2nI20[. See Figure 12.

25. V = 10 - (Wcost+ 10sint)e- t onO:S: t:s: 1, V = 10e-t(ecos(t -1) +esin(t-l) -cost-sint)

1O(! _ e- s )
for t > 1. With Laplace transforms, .c(Q) = (s~ 1)2~ 1 .

27. x(t) = J6 e- t+x sin2~t-x)f(x)dx.

29. See Figure 13.

Section 4.5

1. x(t) = e3t , yet) = - ~ + ¥e4t

3. x(t) = ~ (l - cos 2t), yet) = sin t.

5. x(t) = i (e5t + et ), yet) = eSt - et .

7. x(t) = e2t (cos 3t - j sin 3t), yet) = e2t (cos 3t + 3 sin 3t).

9. x(t) = -~ + ~e2t + ~e-2t, yet) = t _ *e2t + ~e-2t.

11. x(r) = j- ~ cos t - -b cos 3t, yet) = i sin t - ~ sin 31.

13. il(t) =¥- te-4000t - ~e-lOOOt, i2(t) = ~ + te-4000t _ ~e-lOOOt

15. f! = --65 + m613, r2 = _2 -

6
m6 ' il (t) =
'1m
120 (1 + 4T]+1 efjt - 4"2+1 e r2t )
r2m '
h(r)
llQ. (er2t _ e'l t).
m
Section 5.1

o 2n FIGURE 13. Oscillator solution

686 ANSWERS AND HINTS TO SElECTED EXERCISES

3.
7xI - 3X2 = 9,
-3xI + 10X2 - 2x3 = 0,
- 2x2 + lOX3 - X4 = 0,
- X3 + 12'4 - 2xs = 0,
- 2'4 + 6xs = O.

5. XI = ~,X2 = ~,X3 = l&,X4 = ~,Xs =~.

7. P = 112 for one day ahead, p = 0.48 forthree days, and p = 0.481 for five days.
Section 5.2
1. Row 2+ = (-3)(row 1), Row 2* = -1/2, Row 1+ = (-2/3)(row 2), Row h = 3.
3. XI = 2/3, X2 = 1/3.
5.

7. Xl = -2,X2 = 3,X3 = 1,X4 = O.

9. Xl = 0, X2 = -1, x3 = 2.
11. Case B, X = -11l0,y = 1/10 - z/5, Z = z.
13. Case D, all planes equaL
15. Case C, meet pairwise in skew lines.
17. The first three equations represent planes that meet in a unique point (x,y, z) that does not satisfy the
fourth equation. Inconsistent.
19. For a unique solution, there are no independent unknowns, and the number of dependent unknowns
is three.

21. X = ~K - G, y = ~G - ~K.
n+l-k
23 . xk = Ii+l'

Section 5.3
1. Null space span{O}, range R2.

5. VI and V2 are independent, but {VI, V2, V3} is dependent.

7. Independent.
9. Proper means not all ofRn. Show 5 = {O} is closed under addition and scalar multiplication.
11. Show V is closed under addition and scalar multiplication.
 ANSWERS AND HINTS TO SELECTED EXERCISES 687

13. Verify v3 = Vj + 2V2. Then solve Cj Vj + C2V2 = dj V2 + d2V3 for Cj, Q in terms of dj, d2.
15. Any n + 1 vectors in V are dependent. Argue that any v in V is a linear combination of a basis Vj, ... ,
Vn of the subspace V.

17. There are ten separate proofs. For example, items (I) and (VI) hold by the definition of function addition
and function scalar multiplication.
Section 5.4

3. Yes.

7. Yes, they are conformable

9. G~).
11. AB=BA=diag(-1,-4,-2,-2)

13. D = cdiag(l, 1, 1) for some c.

17. Show rref (A) = I to prove that A is nonsingular.

19. The diag(aj, a2, ... , an) is Singular if and only if aj a2 ... an = O.

21. Yes, its rref is the identity

23. The solutions of rref(A)x = 0 and Ax = 0 are the same, so Ax = 0 implies 0 = rref(A)x = x.
25. Find the rref of the augmented matrix (A I 0.

27. (11010)
001
0 ,(1
0010)
101
0 ,(1
010 0)
011
0 ,(1
01 0 0)
001
1,0(1
0
o1) , 0 1 0)
0 0 1
0 .
0 0 1
C
(
-9/5 2215
-115 )
29. -1 2 o .
4/5 -7/5 115

( )
-11 5 -2
31. 7 -3
5 -2

33.
(
315
-2/5
-1
-3/5
215
2
115
115
-1
).
( )
0 0 217 -117
-1 -1 1617 617
" -2
0
-1
0
1517
317
317
217
688 ANSWERS AND HINTS TO SELECTED EXERCISES

-1/2 113 -2
1/3 -1 713 )
4/3
37. ( -1/~ -1/3 3 -10/3 .
-113 -1/3

39. Entries on the left have the form Lk=l L~=l aikbkfcfj-
41. Entries on the left have the form Lk=l (aik + bik)ckj.

Section 5.5

1. Choose a, b to span the plane; then a x b completes a basis in R3 , and V-1 = span{a x b}.

3. Yl = XL Y2 = x2 - (X2 , Yl)Yl/IIYl f, Y3 = x3 - (X3, Yl)Yl/IIYlI1 2 - (X3 , Yl)Yl / IIYlI1 2

5. Yl = Xl, Y2 = x2 - (X2, Yl)Yl/IIYlI1 2

7. A = (~ ~ ~). b= G)'
9. 4.
11. 1.

13. The null space is spanned by (- ~) , the nullspace of A* is the zero vector, the range of A is spanned
by G)' (~). and the range of A * is spanned by G) CD and

15. A b~idcnhe mog"! Ai> mmm Th, nnff;poceo! A M; b~" n} A "'i;lnnh,

mnge o! A' i; m(j) (-1) Thenn!br occo! A' M; b"i, ( = l)

213 2/3 1/3)
17. ( 1/3 -2/3 213 .
-2/3 113 213

19. By the previous exercise, X = (~).

21. Here, m = -27170, b = 2164/3.

23. Each side expands to q L~=l xkuk + Q Lk=l XkVk·

25. Both sides expand to Lk=l xkYk
27. Expand Ilx + yI1 2 using inner product rules and apply the CBS inequality to obtain Ilx + yl12 :::
(11xll + Ilyll)2
 ANSWERS AND HINTS TO SELECTED EXERCISES 689

Section 5.6
1. -1.
3. -1.
5. c.
7. b2
9. -10.
11. 0.

13. 0.

15. 1.
17. -1.
19. abed.
21. -abed.
23. 21.
25. 4.
27. 1.
29. Xl =x2 =-1.
31. Xl = 0, x2 = 1, X3 = 1.
33. Xl = -1,X2 = 0,X3 = 0,X4 = 1.

35. °
37. Experiment with dimension 4. Subtract row I from the otherthree rows to reduce it to a 3 x 3 determinant
of the same form. Repeat.
39. Use the area equal to one half the area of the corresponding parallelogram, which is i la x hi.
41. The equation after expansion is linear in X, y, Z, hence it represents a plane. The points are in the plane
because of determinant rules.
43. 3.
45. 4x - Y + 2Z = 4.
47. 1.
49. 18.
51. Yes. Since det(A) = det(B) det(CB), then det(CB) =f. 0, and CB is invertible.
Section 5.7
1. 5 - 4)" + )., 2 , 2 + i, 2 - i.

3. ).,2-1,1,-1.

5. ).,2 + 2)", 0, -2.

7. 2)..2 - ).,3, 0, 0, 2.
690 ANSWERS AND HINTS TO SELECTED EXERCISES

9. (9 - A)(4 - A)(l - A), 9, 4, l.

11. _A 3 , 0, 0, O.
13. Eigenvalue 0 produces only one eigenvector (1,0).

15. The eigenpairs are -/6, (-/6 - 2, 1); --/6, (-../6 - 2, 1)

17. The eigenpairs are 1, (-2,1,0); -1, (1,-3,3);2, (1,0,0).

19. The eigenpairs are 1, (-1, -7,4); 2, (0, -1, 1); 0, (0, -3, 1).

21. Eigenvalue A = 0 produces only one eigenvector (1,0,0). Geometric multiplicity 1, algebraic
multiplicity 3.
23. Eigenpairs are 2, (0,1,0,0); 2, (1,0,0,0); 1, (0,0,1,0); -1, (0,0,0,1).

l
25. Let P be the matrix of eigenvectors of A; then p- 1AP = AI. Multiply by P and p- 1 to obtain A = AI.
1 0

27. SeIK, N, N', N', N', N', wh", N = (

o
o 0 1 0 ~ ~)
o 0 o 1
o 0 o 0

29. Yes, (~ ~) and (~ ~)-

31. It suffices to show that the columns of <I> are independent, i.e., that the kernel of <I> is zero.

33. Eigenpairs are 0, (1,1,0); 0, (-1,0,1); 1, (-3, -1, 1).

35. Divide by mT(t)q in the first equation and by mT(t)c2 in the second equation.

37. It is equivalent to solving for q and C2 in the equation ( 2 -=- ~ 2 -=- ~ ) G~) = (~). Apply rref
methods.
39. A2 - 4A +5= 0, roots 2 + i, 2 - i.
41. A2(2 -A).

43. (2 - A)(1- A)(-l- A).

45. Directly expand the determinants, or else use the determinant expansion formulas for 2 x 2 and 3 x 3
characteristic polynomials given in earlier exercises.

47. (~ ~)-
49. Use the subspacecriterion: "If x and y are in V and q, C2 are constants, then qx + C2Y is in V."
51. Let x = (1,1,1). Show Ax = 7x.
53. Multiply Ax = AX by the inverse matrix.
55. Let y = Akx. Show Ay = Ay.

Section 6.1

1. i5P = (b + d)12 = (a + c)/2.

3. The vector equations a - b = 0 and a = b are equivalent. The first is equivalent to a - b = O.
 ANSWERS AND HINTS TO SELECTED EXERCISES 691

5. Two vectors are parallel if and only if a = cb for some constant c.

7. Verify that r = (a + b)/2 satisfies a - r = r - b.

9. Apply the triangle inequality to (a - b) and b.

11. 41!'.

13. 1.33,2.23,2.14.

15. The angles are 35.6 and 144.4 degrees (their sum must be 180°).

17. 1.25.

19. 2.145.

21. e= 113.5°, case = -0.3987.

23. x = 1.

25. 2a· (b x c)

27. r = 0 + t, -1 + 2t, 1 - t).

29. (2,3,-3)

31. 0.256x - 0.246y - 0.0476z = 0 or x - 0.961y - 0.186z = O.

33. 64.9 degrees.

35. Apply the parallelogram rule.

37. If A = GP, then let B = PO.

39. Apply the scaling rules.

41. Factor 1 implies no scaling.

43. Write out both sides explicitly in terms of components and compare terms.

Section 6.2

1. r(t)=0,-1,3)+t(2,5,4)

I(
5. ret) = 0 + 7 cos t, 2 + 7 sin t, 5).
1,0,0 Ht( 0,1,0), (l,0,0) --+ 0,1,0),
7. ret) = (1,1,0) + t( 1,2,1 ), (1,1,0) --+ (2,3,1),
( 2, 3, 1 ) + t( -1, - 3, -1 ), (2,3,1) --+ (1,0,0).

9. T = (l!J30)(i + 2j + 5k)
11. r'(t) =- sin(t)(i - j) + cos(t)(i + j) + k, r'(t) = .f3
13. rCt) = 2 cos e(t)i + 2 sin eWj, r' (I) = 20.
692 ANSWERS AND HINTS TO SELECTED EXERCISES

15.
r' (t) = 4 sin O(cos Oi + sin OJ)
+4(1 - cos 0) (- sin Oi + cos OJ),
rl/(t) = 16 cos O(cos Oi + sin OJ)
+ 16 sin O( - sin Iii + cos OJ)
+ 16 sin O( - sin Oi + cos OJ)
+ 16(1 - cos 0)( - cos Oi - sin OJ).

17. The product formula from calculus implies (aj(t)bj(t))' = aj(t)bjCt) + aj(t)bj(t),j = 1,2,3. Add.

19. 2.,f5n.
21. 3n.
23. Start with the result of the previous exercise; then insert KN = dT/ds, r' = r'T.
25. Use K = [r' (t) x rl/ (t)[l[r' (t)3[.
27. The circular helix is a space curve on the cylinder x 2 + l
= 1 that wraps around the cylinder like a
coil spring. The tangent vector points in the direction of the curve. The normal points to the center of
the circle of curvature, which is on the axis of the circular cylinder x 2 + y2 = 1, i.e., the z-axis.
29. Use rl/(t) = (d2s/dt2 )T + (ds/dt)2N and N = B x T.
31. Apply the previous exercise.
33. The triple product is -2. Then h3 =i xj = k; h2 = (i - j)/2; hi = (i + j - 2k)/2.
35. 5v'IO
37. O.
Section 6.3
1. ru = 2ui + vk, rv = 2vj + uk
3. r(u, v) = ui + vj + !C7 - 3u + 2v)k.
5. r(u, v) = (u + v)i + (1 - 2u)j + (u - l)k.
7. For y'3 :::: v :::: 3y'3, 0 :::: u :::: 2n, let r(u, v) = vcosui + v sin uj + (v/y'3) k.
9. r(u, v) = 4 sin u cos v i + 4 sin u sin v j + 4 cos u k for 0 :::: v :::: 2n, 0 :::: u :::: n/3.
11. (0,0,0).
13. No. The trouble is theconditionru xrv i= 0, whichfailsatu = v = 0, becauseru = rv = Oat u = v = 01
15. n = -2i - 4j + k.
17. -6(x - 1) - 2(y - 1) + z - 4 = O.
19. Show that r is in the plane of PQR. Show that (x,y, z) is inside triangle PQR using plane geometry.

*'
21. r(u, v) = sinucosvi + sinucosvj + cosuk, 0 = {(u, v) : 0 :::: u :::: n/2,0 :::: v :::: n/3}, ru
cos u COSy i + cos u cosvj - sin u k, rv = - sin usin vi - sin u sin vj, ru x rv = sin2 usin vU - i),

Area =
23. r(u, v) = u 2 i + uv j + 5:v2 k, 0 = {(u, v) : 0 :::: u :::: 1,0 :::: v :::: 3}, ru = 2u i + uj, rv = vj + v k,
ru x r; = (2u 2 + v2)2, Area = 11.
 ANSWERS AND HINTS TO SELECTED EXERCISES 693

FIGURE 14. Surface and contours of z = Jx 2 + 4y2

25. There are five faces to parameterize. Because of symmetry, only three need to be detailed. The base:
r(u, v) = ui + vj, 0 = leu, v) 0:::: u :::: 1,0 :::: v :::: 2}, ru = i, Iv = j, Iu X Iv = 1,
Area of base = f5 fOI
dudv = 2. The side in the (x, z)-plane: I(U, v) = ui + vk, 0 = {(u,v) : 0 ::::
u :::: 1,0 :::: v :::: 4}, Iu = i, Iv = k, Iu X Iv = 1, Area of (x, z)-side = f; f5
dudv = 4. The
side in the (y,z)-plane: I(U, v) = uj + vk, 0 = {(u,v) : 0 :::: u :::: 2,0 :::: v :::: 4}, Iu = j, Iv =
k, Iu X Iv = 1, Area of (y, z)-side = f04 f5 dudv = 8. The total surface area of the open box (no top)
is 2 + 4 + 4 + 8 + 8 = 26.
27. I(U,V) = vsinacosui + vsinasinuj + vcosak, 0 = {(u,v) : 0 :::: u :::: 2n,0 :::: v :::: h},
Iu = -vsinasinui + vsinacosuj, Iv = sinacosui + sinasinuj + cosak, Iu X Iv = vsina,
Area = nh 2 sin a.
29. I(U, v) = a cos u i + a sin uj + v k, 0 = leu, v) : 0 :::: u :::: 2n, hI :::: v :::: h2}, Iu = -asin u i + acos uj,
Iv = k, Iu X Iv = a, Area = 2an(h2 - hI)'
31. I(U,V) = ui+vj+ (u 2 +v 2 )k, 0 = {(u,v): 0:::: u:::: 1,0:::: v:::: I}, Iu =i+2uk,I v =j+2vk,
Iu X Iv = "II + 4u 2 + 4v 2 , Area = fOI f5 ,,11 + 4u 2 + 4v 2 dudv ="=' 1.86.
33. x= 1/3, Y = 1, Z = -213.
Section 6.4
1. See Figure 14.
3. See Figure 15.
5. See Figure 16.
7. Lift each contour to the height specified, making one curve in each of four planes. Shade the figure.
9. 36n.

11. 2.
13. 6.
15. I(t) = (2 + 2t)i + (4 + 3t)j.

FIGURE 15. Surface and contours of z = x2 - i

694 ANSWERS AND HINTS TO SELECTED EXERCISES

FIGURE 16. Surface and contours of z = 4 + cos(x - U2)

17. (1,1), (1.2,1.4), (1.592, 1.88), (2.29888,2.5168), 0.565736448,3.436352), (5.927439462,

4.862646579).
19. 4x - 6y + 25z = 15.
21. O.
23. (xlr)i + (ylr)j is a unit vector directed radially away from the origin.
25. At the origin. Yes.
27. As the fluid in the stream goes around the corner, it moves with higher velocity on the outer radius.
The water moving faster on one side of the boat causes the bow to rotate into the new direction.

29. O.
31. 0
33. 4xi - 6yj + 2zk.
35. -zj - (sin(x) cos(y»k.
37. Apply the calculus theorem on equality of mixed partials.
39. Apply the calculus theorem on equality of mixed partials.
41. 4.
43. -512n.

45. n.

Section 6.5

1. J~ Ct + 2t2 )dt = 7/6.

3. J; sin(t)( - sin(t»dt = -nI2.

5. J~Jf (- sin t + cos t)dt = O.

7. J; sin(t) cos(t)dt = o.
9. J;/2 sin(t)dUO + sin(t» = (t + 2/(tan(U2) + 1)1 0/2 = nl2 - 1.
11. O.
13. 2.
15. Near point (2/3,5/3, -4/3) x = 1- U3, Y =2- U3, z = -1 - U3, 0 :s t :s 1.

17. x = 4cost,y = 1, z = 4sint, O:s t:s 2n.

 ANSWERS AND HINTS TO SELECTED EXERCISES 695

19. x = 2, Y = acos t, z = b sin t, 0 ::::: t ::::: 2n where a = v's and b = "fi are the semi-axes of the ellipse.
21. The answer is e = e) + e2 + e3, where
e): x = 1 - t, Y = t, Z = 0 0::::: t ::::: 1,
e2: x = 0, y = 1 - t, Z = to::::: t ::::: 1,
e3: x = t, Y = 0, z = 1 - to::::: t ::::: 1.

23. The set is described by 0 ::::: x ::::: 1,0 ::::: y ::::: 1. This is both an x-domain and a y-domain. The outer
normals: segment x = 0: n = -i, segment x = 1: n = i, segment y = 0: n = -j, segment y = 1:
n =j.

25. The domain can be described as 0 = {(x, y) : -1 ::::: x ::::: 1, -1 ::::: y ::::: "II - x2} with normals:
upper circle: n = xi + yj; segment x = -1: n = -i; segment x = 1: n = i; segment y = -1: n = -j.
27. To describe 0 we use three regions obtained by splitting the given region along lines y = 2 and x = 2.
Define y-domains D) = {(x,y) : 0 ::::: x::::: 2,2 ::::: y ::::: 3), 02 = {(x,y) : 0 ::::: x ::::: 2,x ::::: y ::::: 2), 0 3 =
{(x,y): 2::::: x::::: 3,0:::::y::::: 3). The normals are D)-segment x = 0: n = -i, D)-segment x = 2: n =
i, O)-segmenty = 2: n = -j, O)-segmenty = 3:n =j, 02-segmenty = 2: n =j, 02-segmentx = 0:
n = -i, 02-curvey = x: n = (i - j)/"fi, 03-segmentx = 2: n = -i, 03-segmentx = 3: n = i,
03-segment y = 0: n = -j, 03-segmenty = 3: n = j.

29. The domain is described by region R = {(x,y) : 0 ::::: x::::: 1, -1 ::::: y ::::: I} and the inequality -2 :::::
z ::::: 2. The latter has the form g(x,y) ::::: z ::::: hex, y), and therefore the given set 0 is already in the form
of a z-domain. The normals: face x = 0: n = -i, face x = 1: n = i, face y = -1: n = -j, face y = 1:
n = j, face z = -2: n = -k, face z = 2: n = k.
31. The domain is described by region

and the inequality 0 ;:0 z ;:oj4x 2 + 4y2. The latter has the form gex,y) ::::: z ::::: hex, y), and therefore the
given set 0 is in theform of a z-domain. The normals: face z = 0: n = -k, cylinder ex - 1)2 + y2 = 1:
8 2'+8 2'_k
n = ex - I)i + yj, cone Z = j4x 2 + 4y2: n = Xl YJ
v')6z 2 +!
33. n.
35. 0
37. -2.
39. 4.
41. -4.
43. 2.
45. 32JT/3.
47. 4.
49. 32n.
51. 4.
53. -2n.

55. -2n.
696 ANSWERS AND HINTS TO SELECTED EXERCISES

57. -4Jr.

59. 2Jr.

61. Let F = xi. If the divergence is constant, then one side of Greens theorem is a multiple of the area of D.

63. Apply a previous exercise to obtain the ordinary integral J5" cos 3 t(3 sin 2 t cos t)dt.
65. A shared boundary arc C* between two subdomains DI and D2 of D has to be oriented in opposite
ways, in order for D to be regular.

Section 6.6

1. 4.41 x 10 22
1 1 1 2 -I 1 3 -I
3. a =2 a2 xa 3=L: a l,a =Ta3xal=L:a2,a =T a l xa 2=a3·
5. The plane equation is 12nl + 4n2 + 3n3 = 12, which applies to triples (nl, n2, n3) corresponding to
+ n2a2 + n3a3.
vectors nl al
7. The family 12nl + 4n2 + 3n3 = L defines a family of parallel crystal planes.

9. k' = (2 + 12Jr)al + (l + 4Jr)a2 + 6Jra3.

Section 7.1

1. Draw a right triangle with sides 1, ux, )1 + u~ to obtain formulas for sin/! and cos/! in ter.ms of ux.
Apply the binomial series formula to expand (l + u~)-1/2 in a Taylor series.
3. The period of UI or u2 in t is 2Jr divided by nJrclL Hence they also have period P.

5. Let u be the difference of two solutions. Define E(l) = ~ f~(pu~ + Tu;)dx. Show that E/(t)
-ka(u(O, t)2/2))lot and hence E(t) ::: E(O). Then E(O) = 0 implies E(t) ::: 0 or E(t) = 0 for all t. So the
integrand must vanish, which implies u(x, I)=constant. Since u(x, 0) = 0, then u == 0 as required.

Section 7.2
1. The cosine and sine terms satisfy a differential equationy" + JL2 y = O. Use it to find second derivatives.

3. Show that the difference u of two solutions is zero. Define I(t) = ~ f~ u 2(x, t)dx. Then show that
I' = -K J~ u~ dx ::: 0; hence I(t) ::: 1(0) = 0, and then 1(1) == O. Integrand zero implies u == O.
5. The details are the same as Exercise 3, with the exception of the treatment of the boundary terms in I' (I).
Here, the term KAu(x, t)ux(x, I)1~~6 appears, which is zero, so again I'et) ::: O.

Section 7.3

-*
1. Again Jar? u(V'(u) . n)dS = O. The rest ofthe proof is unchanged from the text.

3. In this case, Jar? u(V'(u) . n)dS = Jar? u 2dS ::: O. The rest of the proof is unchanged from the text
version.

5. A solution u = 0 if and only if one of the sine factors is zero at x = 1 or y = 1, which happens exactly
when A or JL is an integer multiple of Jr.

Section 7.4
1. Expand the dot product V' . Cu V' (u)). Compare with the right side, V' Cu)2 = u~ + u;.
3. A constant satisfies the differential equation, so the only question is the boundary condition. This is
settled quickly by observing that V' (u) = 0 when u is constant.
 ANSWERS AND HINTS TO SELECTED EXERCISES 697

5. Let U be the difference of two solutions. Then u satisfies a Neumann problem with mixed boundary
conditions. By the proof of the uniqueness theorem, V(u) = 0 on n, because u(V(u) . n) = 0 on an.
Hence for connected n, u is a constant on n. Since q '" 0, then this constant is zero, and hence u is zero.
Section 7.5
1. The sine and cosine terms are solutions of an ordinary differential equation wI! + a2 w = 0, where
depending on the context, a = j.i, a = A, or a = we. Therefore, Utt = -w 2c2u, Uxx = -A 2U,
Uyy = _j.i 2u. It remains to compute both sides of the differential equation and compare terms.

3. Define E(t) = ~ f fa (PUf + To(u~ + u;))dxdy, where [2 = To/p. Then E'(t)

for all t :::: 0, which implies that u is constant. Since u = Oat t = 0, then u == 0.
= 0, E(t) = E(O) = °
Section 7.6

° ° °
1. It suffices to show that the difference u of two solutions is zero. Let let) = ~ f f fa u 2 dx dydz. Then
1'(1) = gives I(t) = 1(0) = and finally u ==

-f f faa u2 dS :s: 0; hence I'(t) :s: °as before, to give 0 :s: let) :s: 1(0) = 0.
3. The method parallels that of the first exercise. The only new detail is that the boundary integral of
u(V(u) . n) evaluates to

5. A solution u vanishes on the boundary if and only ifthe product of the sine factors vanishes atx = y = z =
1, and this is exactly the condition sin Asin j.i sin v = 0. Therefore, A, j.i, and v are integer multiples of Jr.
Section 7.7
1. The proof uses the same integral method as in Section 7.4. Here, V . (uV(u)) = V(u)2, and by the
divergence theorem the integralofVu 2 over n is zero (u is zero on the boundary). Hence Vu = Oin n, and
then by connectness of n it follows that u is constant. The constant is zero because u = 0 on the boundary.

3. The difference u = U2 - ul satisfies L'.u = 0 and zero boundary conditions. Methods of the first exercise
produce Vu = 0, and finally u is some constant c on n by connectness. Then u = c gives u2 = ul + e.

5. The proof parallels that of the first exercise. The difference in detail is that V(u)2 = V· (uV(u)) = -hu 2
on the boundary; therefore, by the divergence theorem 0 :s: f f fa V(u)2dx dydz :s: 0. As before, u = c
°
on n. Then V(u) . n + hu = 0 on the boundary gives c = and finally u == 0.
Section 8.1
1. See Figure 17.
3. See Figure 18.
5. u = sin(2x)e- 4t + 2 sin(3x)e- 9t .
7. u = sin(2x)e- 4t + sin(4x)e- 16t
9. Integrate cos(m - n)x - cos(m + n)x over [0, Jr] and use a trigonometric identity similar to the one of
the previous exercise.

FIGURE 17. Surface u = sin(2x)e- 4t

698 ANSWERS AND HINTS TO SELECTED EXERCISES

FIGURE 18. Surface u = ~ sin(4x)e- 16t

11. Use 2 cos 2 (u) = 1 + cos(2u) and integrate using u = nx.

13. Both partials equal L~=l -n 2 bn sin(nx)e- n2t

15. Let t = 0 in the summation. Then the exponential term is 1. The sum is exactly j(x).
17. Each term Un satisfies homogeneous boundary conditions. Add these N equations to show that the sum
also satisfies the conditions.

Section 8.2

1. b2 = 2, b3 = 1, b6 = 1/2, all others zero.

3. bn = Jr(n~~ 1)
Section 8.3
1 u(x t) = II ,",00 sin(2k+l)x e-(2k+1)2 t
. , Jr L..k=O 2k+1 .

3. u(x, t) = L~l .& [1 + (l + n)( _1)n+11 sin(nx)e- n2t

5. u(x, t) = L~l 4(2+~31)n) sin(nx)e- n2t

7. u ( x, t) = ,",00 8 . (2k
L..k=O Jr(2k+l) sm + 1)x e-(2k+1)2 t . See Figure 20.
9.

U= L00
- 2
k=l n(2k + 1)
sine 4k + 2)x e-(4k+ 2)
2
t +L
00

k=O n(2k
6
+ 1)
sin(2k + l)x e-(2k+l)
2
t.

See Figure 21.

11. u = L~l 4(2(-1~tl_1) sin(nx)e- n2t See Figure 22.

13 . f( x ) -_,",00 sin(2kx)
L..k=l -2-k- + ,",00 (1
L..k=O 2k+l + Jr(2k+l)2
2(_l)k ) . (2k
sm + 1)x. Ii0 ven·fy tel
hd . k -
ennty ta ex - n
12
.

Section 8.4
1. u(x, t) = -1.

FIGURE 19. Surface u = * sin(x)e-t + -/,;- sin(3x)e-9t

 ANSWERS AND HINTS TO SELECTED EXERCISES 699

FIGURE 20. Surface

u = ~ (6 sin(x)e- t- 2 sin(2x)e- 4t + 2 sin(3x)e- 9t )

t = 0.0 t = 0.05 t = 0.01

FIGURE 21. Snapshots of u = 4 sin(x)e- t - ~ sin(2x)e- 4t -

~ sin(3x)e- 9t

3 u(x I) = 2+rr - i ,",co __1_ cos(2k + l)x e-(2k+l)2 t

. , 2 rr Lk=O (2k+ 1)2 .

b ( ) -n b rr3 b -rr b
5. u (x, t)
24 2rr
= b0 + ,",CO
2t
Ln=l n cos nx e ,0 = 12' 2k = 2k2' 2k+1 = (2kH)4 rr - (2k+l)2'

7. u(x, t) = 3 represents a plane parallel to the (x, I)-plane.

9 . u(x, t) = irr ,",CO (_l)k cos(2k + l)x e-(2k+1)2 t See Figure 23

Lk=O 2k+1 . .

11. U(x, t) = h3 + Ln=l

CO 21(6-n 2JT2)(-1)"-61
n4 JT
2t .
cos(nx)e-n . See Figure 24.

13. U(x, t) = ~ + L~o ~k~;' cos(2k + l)x e-(2k+1)2r. Select x = a to obtain the identity
15. For)" = fJ.,2, X = A cos fJ.,X + B sin fJ.,X and the boundary conditions lead to B = a and sin fJ.,JT = O.
Hence X = cos fJ.,X and fJ., = n for some integer n ::: 1. For)" < a there is no nontrivial solution X. For
)" = 0, X = A + Bx and B = O.
Section 8.5
1. Substitutef(x) into Euler's formulas to obtain bo = JT/4, bn = ((_1)n -l)/(JTn2 ), an = (_l)n+1/n.
3. Apply the theorem for x = 0 to Exercise 1.
5. Apply the theorem with x = 0 to Exercise 2.

FIGURE 22. Surface u(x, t) = * cos(x)e- t

700 ANSWERS AND HINTS TO SELECTED EXERCISES

//"--'~"-

I ///

~ -----"

t = 0.0 t = 0.05 t = 0.01

FIGURE 23. Snapshots of u(x, t) = ~ + 2rr~-24 cos(x)e- t
~ cos(2x)e- 4t

7. Let U be the difference of two possible solutions. Then u satisfies the ring problem withf = O. Define
I(t) = i f~;r u 2 (x, t)dx. Show I'(t) == 0, hence I(t) = 1(0) = 0, and then u == O.
Section 8.6

1. Apply induction on n. The case n = 1 duplicates the periodicity hypotheSiS. Handle negative n similarly.

3. Assumef(-x) = -f(x), g(-x) = -g(x). Let h = qf + c2g. Show he-x) = -hex)

5. Assume f( -x) = f(x) , g( -x) = g(x). Let h = fg. Show he -x) = hex)

7. Write the integral as the sum of two integrals and change variables x = -t in one of them.

9. Usef(-x) = f(x) and the chain rule to show J'(-x) = -J'(x).

11. Make a change of variables x = x( v) and apply the results in the text for other periods to g(v) = f(x( v».
For example, xCv) = LviJr maps -n ::5 v ::5 n onto -L ::5 x ::5 L. This is outlined in the text, but
the details have been omitted. In particular, show that g and g' are sectionally continuous and relate
g(x + 0), g(x - 0) to f(x + O),f(x - 0). ExpliCitly carry out the change of variables in the integrals in
order to obtain limits of integration -L ::5 x ::5 L.

13. To evaluate the integral f"!.;r cos(ku)du, use orthogonality of f(x) = cos(Ox) = 1 and g(x) = cos(kx)
on [-n,nJ.

15. Start with the second e'}:tation in the theorem. Isolate F(x) in this formula and replace the integral by

f:
its eqUivalent fraction 2n+l ,where M is (-1) times the integral whose integrand contains Gl Estimates
of the size of M are obtained from the calculus inequalities cose ::5 1 and If: f(x)dxl ::5 j(x)dx.

Section 8.9

1. (1) Both f and J' are continuous (1' (0) has to be computed from the Newton quotient definition). (2)
Nothing to do for continuous f. (3) The integral off is twice fooo xe- X dx.

3. (1) Classic definition of a piecewise infinitely differentiable function. (2) Nothing to do unless x = -1
or x = 1. (3) The integral off is 2 f~ x 2 dx.

5. A(/h) = 0, B(/h) = ;r(1'1~4)2'

7. A(/h) = 0, B(/h) = ~ sinl'-icosl'.
I'

9. The integration 4 fooo Si~1' cos(/hx)d/h is done from Example 1 in the text.

11. The indefinite integrals for A(/h) and B(/h) can be found in most integral tables.
 ANSWERS AND HINTS TO SELECTED EXERCISES 701

13. Use the trigonometric identity 2 sin a cos b = sin(a+b) +sin(a-b) to integrate A(JL). Use 2 sin asin b =
cos(a - b) - cos(a + b) to integrate B(JL). The addition theorem for the cosine is used on the integrand
in the Fourier integral theorem to simplify
Section 8.10
2
1. The product solutions are B(JL) sin(JLx)e- 1l t. Superposition produces the solution u(x, t), and B(JL) is
determined by the Fourier sine integral theorem.
3. The odd extension of f is defined by f*(x) = f(x) for x 2: 0 and f-(x) = -f( -x) for x < O. To be
i
shown: (1) f* and its derivative are sectionally continuous; (2) f*(x) = ([*(x + 0) + f*(x - 0)]; (3)
The integral of f* over (-00, (0) is finite. The integral relation for f results from applying the Fourier
integral theorem of Section 8.9 to f*, which agrees with f on 0 ~ x < 00. By oddness of f*, A(JL) O. =
The equation for B can be Simplified to involve only f on the interval 0 ~ x < 00, because f* is odd.
Section 8.12
1. w(x)=x 4 ,[0,2].
3. w(x) = (l + xle"x, [0, 1]
5. Compute X~, Xn(O) and Xn(rr). They are shown to equal -AnXn , 0, 0, respectively. To verify that all
solutions are represented, solve the differential equation for the three cases A = -JL 2 , A = 0, and A = JL 2 ,
with JL > O. The third case produces eigenpairs A = n 2 , X = sin(nx) for n = 1,2,3, ....
7. Change variables by x = et , UCO = X(e t ). Then U" + AU = 0, U(O) = U(l) = O. Let t = s/rr,
yes) = U(s/rr). Then yll + rr- 2AY = 0, yeO) = Y(rr) = O. Apply a previous exercise to solve for Y and
hence U; then change variables back to x.
9. Apply the trigonometric identity sin nx sin mx = i [cos(n - m)x - cos(n + m)x].
11. The products X1X2W and X2X3W are each odd functions on [-1, 11; hence their integrals are zero. The
integral of X1X3W = (5x 4 - 3x 2 )/2 is twice its value over [0,1], because it is even. Direct integration
gives zero.
13. Substitute X(x)TCt) into the differential equation and assume X(x) TCt) i= 0 to obtain the two differential
equations for X and T. One issue is TCt) i= 0 for all t, which is settled by solving the differential
equation for T. This allows cancellation of TCt) in the boundary conditions on u, hence producing
alX(a) + a2X'(a) = 0, blX(b) + b2X'(b) = O.
Section 8.13
1. Eigenpairs are A = (2k - 1) 2/4, X = cos((2k - l)xl2) and A = k 2 , X = sin(h) for k = 1,2,3, ....
Orthogonality uses weight w(x) = 1, and the relation is I~" XpXqdx = 0, where Xp and Xq are two
different eigenfunctions from the above list.
3. A = (rrnlln 2)2, X = sin(nrr In(x)lln(2)), n = 1,2,3, ... , w(x) = x-I,
112 x-I sin(nrr In(x)lln(2)) sin(mrr In(x)lln(2))dx = 0 for n i= m.
5. A = n 2/4, X = sin(nx), w(x) = 4, g sin(nx) sin(mx) (4)dx = 0 for n i= m.
Section 8.14

1. Eigenpairs are A = (2k + 1)(2k + 2), X = P2k+l (x), k = 0,1,2, ... , where Pn is the nth Legendre
polynomial.
3. A direct approach can be carried out following the general method of Section 8.12. The functions Xn
and Xm satisfy differential equations; hence we multiply one differential equation by Xn and the other
by X m , and then subtract to give

Xm(xX~)' - Xn(xX~)' = (Am - An)xXnXm.

702 ANSWERS AND HINTS TO SELECTED EXERCISES

Integrate over [0, 1]; then integrate by parts on the left. Finally, show that the left side is zero.
Section 9.1
1. U(x, t) = sin(2x)e-4t - sin(5x)e- 25t .

3. u(x, t) = ~ - ~ cos(2x)e- 4t .

5. u(x, t) = cos(f.lX)e-/Jh , J.t ~ 0.274Jr.

7. u(x, t) = sin(2x)e- 4t .
9. u(r,t) = ~r-l sin(Jrr)e-;rr2 t .
2
11. u(r, t) = JO(SI r)e-S1 t.
13. One second partial is Uxx = (l/r)ur - xr- 2(xlr)ur + (xlr)2 Urr . Results for y and z are similar. Add to
obtain a formula for Ut.
IS. Verify Uxx = (llr)ur - xr- 2(xlr)ur + (xlr)2 urr . The result for y is similar. Add to find a formula for Ut.

17. Verify Kt = K (tr - t) and Kx = -2Kx1(4t), Kxx = K (tr - -t); then Kt = Kxx·
Section 9.2

1. u(x, t) = ! (j(x - t) +f(x + t)). See Figure 24.

3. to = 1, tl = 3, t2 = 4.
S. u(x, t) = ! (sin(x - ct) + sin(x + ct)).
7. The origins of t/J(x) and t/J(x - ct) are at x = °and x = ct > 0, respectively. Translation is to the right by
amount d(t) = ct.
9.
2l/J(x) f(x) - t It g(~)~ + 2l/J(0) - f(O),

21/F(x) f(x) + t It g(~)~ + N(O) - f(O).

11. See Figures 25-27.

13. Let J.t = ..;;:., where A is a separation constant. Boundedness of the product solution at infinity implies
A 2: O. Hence the product solutions are products of sines and cosines.
IS. The formula for u can be written as u(x, t) = !Je(X - ct) + !fe(x + ct). It follows from the d'Alembert
theorem that U satisfies the wave equation and the boundary conditions u(x, 0) = fe(x) , Ut(x, 0) = O.
Since fe (x) = f(x) for x 2: 0, the proof is complete.

I
I~ i
\

t =0.0 t = 0.5 t = 1.0

FIGURE 24. Snapshots at t = 0, t = 0.5, and t = 1

 ANSWERS AND HINTS TO SELECTED EXERCISES 703

t = 0.0 t = 0.2 t = 0.4

FIGURE 25. Snapshots at t = 0, t = 0.2, t = 0.4

t = 0.6 t = 0.8 t = 1.0

FIGURE 26. Snapshots at t = 0.6, t = 0.8, t = 1

FIGURE 27. Solution surface

Section 9.3

1.

u= ~ t
rr k=O
sin«2k + l)x) sinh((2k + 1)y).
(2k + 1) sinh((2k + 1)rr)

3. 0.2499040972, n = 2 terms.

5. Add the answers from Exercise 1 and Exercise 2: U = Uj + U2,

Uj = ~ t
rr k=O
sin((2k + l)x) sinh((2k + l)y)
(2k+ 1)sinh((2k+ l)rr) ,
704 ANSWERS AND HINTS TO SELECTED EXERCISES

U2 = ~ I : (_1)n+1 sin(ny)(sinh(ny) + sinhn(rr -x».

rr n=l n sinh (nrr)

7. Show that u = y/rr satisfies Laplaces equation and the boundary conditions. By uniqueness, it is the only
solution.

9.

u= ~ I:(_1)n+1 sin(nx) sinh(ny ).

rr n=l n2 sinh(nrr)

11.

u = Co _ ~ I : cos«2k + l)x) cosh«2k + l)y).

rr k=O (2k + 1)3 sinh«2k + 1)rr)

13.

2 ~ [( )n+ I 1sin(nx) cosh(ny)

u=-L.., -1 +1 .
rr n=l n 2 sinh(nrr)

15. Even though u is bounded, it fails to be true that Ux and uy are bounded. However, by direct calculation,
u satisfies the differential equation and boundary conditions. The boundary evaluation uses the formula
limy~o+ tan-leafy) = ~~.

17. Let u = (x - ~)Iy, du = -~/y to obtain an arctangent integral that evaluates to l.

19. Let w be the difference of the left and right sides of the identity. Then w satisfies the differential equation
with zero boundary conditions. By uniqueness (Chapter 7), w == O.

21. Let w be the difference of the left and right sides of the identity. Then w satisfies the differential equation
with zero boundary conditions. By uniqueness (Chapter 7), w == O.
I N _.2 ~
23. u(O, 0.5) ~ IT fo /+0.25 ~ = 0.6156903104 for N = v Sin 10 = 3.393l. The exact answer is
0.6156903442.

25. vCr, e) = 1 + 5r cos e - 3r 3 cos 3e.

27. Consistency requires ao(f) = 0 Cit is). Then vCr, e) = 4rcose - tr2 cos2e + ~r3 cos3e.

29. The alternating series error is controlled for k = 2, giving vCO.l,rr/2) = 112 + 1495031750000rr =
0.5634510439.

31. vCO.l,rr/2) = ~ + ~ tan- l (~) = 0.5634510349.

33. Apply the formula of Problem 9.3.9 to w = u(f,o,O) with 0: = rr/2.

35. The first two equations are verified by direct calculation of the left side. The last equation uses v
independent of e. Unboundedness of the logarithm is a basic property developed in calculus.

37. Product solutions v = R(r)e(e) satisfy for separation constant Je the equations r2R" + rR' - JeR = 0,
e" + Jee = 0, where e is 2rr-periodic and R is bounded at r = 00. Then R = qrn and e =
aI cos ne + bl sin ne, Je = n2 These formulas also apply for Je = O.
 ANSWERS AND HINTS TO SELECTED EXERCISES 705

39.

Ao = - 1
2rr
1 0
2,,-
f(O) dO

a -nAn =-11
rr 0
2 ,,-
f(O) cos (nO) dO

a-nB n = -1 1211: feo) sin (nO) dO

rr 0

Section 9.4
1.
U = sin 3x sin 5y e-(9+25)t - sin 4x sin 6y e-(l6+36lt.

3.
00 00 41+(-l)m+l1+(-l)n+l 2 2
U= " ,,- sinmxsinnye-(m +n )t.
L L rr2 m n
m=l n=l
5.

U = 4 LL
00 00 (_l)m+n
sin mx sin ny e-(m +n
2 2
)t.
m=l n=l mn

7.
_ 8 ~ ~ [1 + (-l)m+l][1 + (_l)n+1]. . _(m2+n2)t
U - 2 L L 3 Sill mx Sill ny e .
rr m=ln=l m n

9. Let u(x,y, t) be defined on domain D: 0 :::: x :::: L, 0 :::: y :::: M, t ::: O. Assume Ut = Uxx + U yy on D,
U = 0 on the boundary of D, and u(x,y, 0) = f(x,y) on 0 :::: x :::: L, 0 :::: y :::: M.

11. The details for I mn duplicate Problem 9.4.1. The other relation follows from the previous exercise and
A = It + v.
13. The problems are X" + ItX = 0, X'(O) = X'(rr) = 0, y" + vY = 0, y'(O) = yl(rr) = 0, with
A = It + v. Then It = m 2 , v = n2
15. Apply superposition to the product solutions of the previous exercises.
l7. The problems are X" + ItX = 0, X' (0) = X' (rr) = 0, yll + vY = 0, yeO) = y(rr) = 0, with A = It + v,
m = 01,2, ... and n = 1,2,3, ... Then It = m2 , v = n 2 In the t-variable, I' + AI = 0, rcO) -10.
19. Apply superposition to the product solutions of the preceding two exercises.

21. U = JO(SOl r)e-S61 t - 2Jo(s03r)e-s63t

23. 501 = 2.404825558, 503 = 8.653727913. See Figure 28.

25. The product solutions are T = e- At , R = Jm(smnr/a), e = Am cos(mO) + Bm sin(mO), f.l = m2 ,
A = s~n' The formula for U is unchanged from the case a = 1, except that r is replaced on the right by
706 ANSWERS AND HINTS TO SELECTED EXERCISES

1=0.0 1=0.5 1=1.0

FIGURE 28. 3-D Snapshots at t = 0, t = 0.5, and t = 1

= 1, the coefficients in the series are obtained by orthogonality methods. See

ria. As in the case for a
the solution manual for the very complex formulas that result.
Section 9.5
1. Argue that Vt satisfies the wave equation by differentiation on t of the wave equation.
3. The series answer in Problem 9.5.1 reduces to a Single term, for n = 2. So a2(g) must be determined.
However, a2 (g) = 1 from the representation of g, without further integrations (it's already a Bessel series).

5. Let U be defined over the rectangle D: 0 :::s x :::s L, 0 :::s Y :::s M. Require Utt =
Uxx + Uyy on D for t ~ 0,
U = 0 on the boundary of D for t ~ 0, u(x,y, 0) = J(x,y), Ut(x,y, 0) = g(x,y) on D. Product solutions
X(x)Y(y)T(t) satisfy the problems X" + /LX = 0, X(O) = X(L) = 0, y" + vY = 0, yeO) = Y(M) = 0,
T" + AT = 0, nO) f. 0, A = /L + v. Then /L = (mrr/L) 2, v = (nrr/M) 2, Amn = (mrr/L)2 + (nrr/M) 2,
Xm = sin(mrrxlL), Yn = sin(nrryIM), Tmn = Amn cos(~t) + Bmn sin(~t), and
00 00
U= L LTmn(t)Xm(X)Yn(y),
m=l n=l

where

Amn = - 41L1M Bmn = 4~ lL1M

°°
°°
fXmXndxdxy, gXmXndxdxy.
LM LM-VAmn

7. Let Q be the region 0 :::s r :::s a in polar coordinates. Assume that Utt = c2 (u rr + ~ur) on Q for t ~ 0,
u(a, t) = 0 for t ~ 0, u(r,O) = 0 for 0 :::s r < a, ut(r, 0) = g(r) on Q for t ~ 0, and u(r, t) is finite.
Product solutions are R = ]o(5 nrla) and T = sin(5nct). Then u = L~1 Cn.!O(5nrla) sin(5nct) with
Cn = (2I]r(5 n)) JJ g(ax)]o(5 nx)xdx. The solution u(f,Ol is obtained by differentiation as ui°J).
9. The roots in 0 :::s r < 1 of]O(53r) = 0 are exactly at 53r = 51 and 53r = 52.
Section 9.6
1.

sinh ( -v'srrz)
u = sin(rrx) sin(2rry) -v's'
sinh ( 5rr)

3.
sinh(v'4Irrz)
u = sin(4rrx)sin(5rry) r;;, .
-v41rr
 ANSWERS AND HINTS TO SELECTED EXERCISES 707

5.

7. Verify it from

Cn 2n+-1
= -
2
(11 -1
Pn(x)dx - 2 1°)
-1
Pn(x)dx,

which gives c2k = 0 and c2k+! = (4k + 3)fd P2k+l(X)dx. In particular, Cj = 3/2, C3 = -7/8,
C5 = 11/16. Bonnet's theorem can be used to show that Pn+l (0) = -(n/(n + l))Pn-l (0), and then
f01 Pn(x)dx = Pn-l (O)l(n + 1) results in
4k + 3 (_1)k(2k)!
+
c2k+l = 2k 2 4k(kl)2 .

Section 9.7
1.

3.

5. u(x, t) = sin net - x) for 0::'0 x < t, u(x, t) = 0 otherwise. See Figure 29.
7. u(x, t) = -it2 - -i (t - xld H(t - xlc)

t = 0.0 t = 0.2 t = 0.6

FIGURE 29. Snapshots at t = 0, t = 0.2, and t = 0.6
 References

[A-S] M. Abramowitz and I.A. Stegun, Handbook of Mathematical Functions, New York, Dover
Publications, 1965.
[B-M] G. Birkhoff and S. Mac Lane, A Survey of Modem Algebra, 4th ed., New York, Macmillan, 1977.
[B-McG] PW Berg and].L. McGregor, Elementary Partial Differential Equations, Holden-Day, 1966.
[B-R] G. Birkhoff and G.C Rota, Ordinary Differential Equations, 3rd ed., New York, Wiley Publishing
Co., 1978.
[C-L] E.A. Coddington and N. levinson, Theory of Ordinary Differential Equations, New York, McGraw-
Hill, 1955.
[Ch-B2] R.V Churchill andj.W Brown, Complex Variables and Applications, 5th ed., New York, McGraw-
Hill, 1990.
[Ch-B] R.v Churchill and].W Brown, Fourier Series and Boundary Value Problems, 3rd ed., New York,
McGraw-Hill, 1978.
[Ch-K] W Cheney and D. Kincaid, Numerical Mathematics and Computing, 2nd ed., Monterey,
Brooks/Cole Publishing Co., 1985.
[Churl R.v Churchill, Operational Mathematics, 3rd ed., New York, McGraw-Hill, 1972.
[D-S] H.F. Davis and A.D. Snider, Introduction to Vector Analysis, 6th ed., Wm. C Brown Publishers,
1991.
[DenH] JP Den Hartog, Mechanical Vibrations, 4th ed., New York, Dover Publications, 1985.
[ER] A. Erdelyi et aI., Tables of Integral Transforms, Volumes I and II, McGraw-Hill, 1954.
[Fr] A. Friedman, Advanced Calculus, Holt, Rinehart and Winston, 1971.
[Ga] PR. Garabedian, Partial Differential Equations, New York, John Wiley and Sons, 1964.
[He] P Henrici, Elements of Numerical Analysis, New York, Wiley Publishing Co., 1965.
Ua] D. Jackson, Fourier Series and Othogonal Polynomials, Mathematical Association of America, 1941.
IKe] ].P Keener, Principles of Applied Mathematics, Addison-Wesley, 1988.
IKrey] E. Kreyszig, Advanced Engineering Mathematics, 7th ed., John Wiley and Sons, 1993.
[Mcl] N ,W Mclachlan, Bessel Functions for Engineers, 2nd ed., Oxford, Clarendon Press, 1961.
[No] B. Noble, Applied Linear Algebra, 3rd ed., Prentice Hall, 1988.
[Prj WHo Press et aI., Numerical Recipes, London, Cambridge University Press, 1986.
710 REFERENCES

[Sch] E.R. Scheinerrnan, Introduction to Dynamical Systems, Prentice Hall, 1996.

[V-P] D. Varberg and E.]. Purcell, Calculus, 7th ed., Prentice Hall, 1997.
tWa] G.N. Watson, A Treatise on the Theory of Bessel Functions, 2nd ed., London, Cambridge University
Press, 1944.
[Wi2] D.V Widder, The Heat Equation, New York, Academic Press, 1975.
[Wi] DV Widder, The Laplace Transform, Princeton University Press, 1941.
[Za] WHo Zachariasen, Theory of X-Ray Diffraction in Crystals, New York, Dover Publishing Co., 1967.
 Index

vEev-L, 318 Arc length, 385

V-L,318 Arc length parameter, 385
K thermal diffusivity, 474 Arccosine
V' notation, 412 principal value, 57
«-Curves, 396 Arctangent identities, 20
v-Curves, 396 Area of a triangle, 344
x-Domain, 418 Astronomical unit A. U., 114
y- Domain, 419 Asymptotic estimates, 198
z-Domain,422 Atmospheric pressure, 96
Augmented matrix, 284
Abel's fonnula, 157 Axis or revolution, 85
Abel, N.H., 157
Abstract vector space, 302 Back substitution, 288
Adaptive numerical method, 136 Balancing a chemical reaction, 279
Addition of geometric vectors, 367 Basis, 297, 299
Adjoint matrix, 319, 327 Gram-Schmidt, 316
Admittance, 117 Basis of solutions, 158
Algebraic and geometric multiplicity, 349 Beam direction, 435
Algorithm for A-I, 311 Bending, 463
Anagram, 77 Bernoulli, D. and Euler, L, 473
Analytic, 182 Bessel
Angle between vectors, 372 equation of order n, 147
Anomaly, 61 Friedrich Wilhelm, 3
Approximations of]o(x), 203 numerical differentiation, 34
712 INDEX

table, 3 general applied voltage, 176

Bessel equation, 154, 189 LRC, 139
order zero, 549, 550 numerical integration, 176
Bessel function, 3,20, 194 periodic square wave input, 251
behavior, 197 Riccati method, 11 7
finiteness property, 196 sinusoidal input, 250
integral, 198 switch, 179
numeric, 199 transform method, 234, 249
recurrence, 198 transform method for square wave, 253
second kind, 196 two-body analogue, 270
Bessel inequality, 505 two-loop, transform method, 263
Binormal vector, 386 undetermined coefficients applications,
Bisection 171
advantages, 43 Wheatstone bridge, 277
algorithm, 43 Circular cylinder parameterization, 403
disadvantages, 43 Circular disk
flask cooling application, 43 graphics, 638
method of, 42 heat diffusion, 615
Boundary conditions for the wave numeric, 628
equation, 445 Circular drumhead, 467
Bragg condition, 436 Circular flat plate transient diffusion
Bragg, Sir William and Sir Lawrence, 436 Bessel functions, 635
Brahe, Tycho, 58 boundary value problem, 633
Bravais, A., 431 convergence and Fourier-Bessel series,
636
Calculation of curvature and torsion, 387 eigenvalues, 634
Calculus of scalar and vector fields, 403 implicit conditions, 634
Cantilever, 122 normalizing constants, 635
Cartesian coordinates, 369, 371 other boundary conditions, 637
Cauchy kernel, 166 periodic Sturm-Liouville problem, 634
Causality principle, 176 product solutions, 634
Center of mass, 403 Singular Sturm-Liouville problem, 635
Central difference approximation, 32 symmetric solution, 636
Change of basis, 351 zero boundary temperature, 633
Characteristic equation, 165 Circular hole in an infinite plate, 629
Characteristic polynomial, 348, 361 Circular sector heat diffusion, 622
Chemical balance, 301 Circulation, 410
Churchill's theorem, 558-561, 566 Clock pendulum, 141
Circle parameterization, 390 Coefficients of a DE, 156
Circuit Column space, 321
multiloop, 149 Columns of a matrix, 306
Circuits Companion matrix, 361
LR, steady state response, 253 Complementary error function, 658
constant input, 250 Complete Fourier series, 492
DC applied voltage, 172 Complex field and matrices, 323
finite duration applied voltage, 177 Complex Fourier series, 499
 INDEX 713

Complex matrix, 283 DAlembert's theorem, 580

Complex vector space, 303 D'Alembert's wave equation, 445, 579
Components of a vector, 370 D'Alembert, Jean Ie Rond, 442
Composition of linear operators, 306 Del notation, 412
Compressing flow, 409 Dependent vectors, 296
Cone, 392,401 Derivative, see Vector, 382
Conformable products, 312 Derivative rule, 231
Conical column, 86, 91 Determinants, 326
Connected, 100,456 basic properties, 334
Conservation of energy, 448 cofactor and adjOint, 338
Constant coefficients, 155, 164,211 column vector representation, 330
Continuous, 156,382 combinatorial formula impractical, 339
Continuous interest, 97 computation rules, 339
Convergence, see Fourier, 480 Cramers rule, 338
Convergence of Fourier series, 482 efficient, 339
Conversion of systems to matrix form, 283 evaluation, 339
Conversion to systems, 211 function, 326
Convolution, 167,237,247,657 Gauss reduction, 340
Cooling law and the finite rod, 533 minors and cofactors, 336
Coupled oscillators parallelepiped, 344
eigenvalues, 347 uniqueness, 331
eigenvectors, 347 Diagonal matrix, 307
equation of motion, 346 Diagonalization of matrices, 352
normal mode, 346 Differentiability of solutions, 127
normal mode equations, 347 Differentiability theorem, 126
normal modes, 346 Differential equation
synchronous vibration, 346 dxldt = FCtl, 85
Coupled systems, 273 circuit, 78
Coverup method, 229 circuit example, LR, 94
Cramer's rule, 326, 328 direction field, 80
Critically damped spring, 181 domain of definition, 83
Cross product coordinate representation, Euler's method, 128
377 exact, 98
Cross product determinant formula, 377 definition, 99
Cross product of vectors, 376 flow past a cylinder, 104
Cross-sectional area, 85 fluid flow, 103
Crystal integrating factor, 102
components relative to a basis, 433 solution method, 10 1
dual basis, 433 exactness criterion, 100
Crystal plane calculation, 434 examples from partial DE, 144
Crystal structure, 431 flask coo ling, 78
Cubic interpolation, 20 higher-order, 139, 149
Curl of a field, 409 fourth-order, 152
Curl operator, 409 general first-order, 92
Current, 96 geometry of solutions, 80
Curvature, 386, 391 homogeneous linear equation, 93
714 INDEX

initial value problem, 123 reservoir, 85

linear analytic, 182 reservoir draining, 87
linear equation of order n, 211 Differential equation of order two
linear first-order, 92 applications, 106
linear second-order, 155 autonomous, 110
damping, 120 constant, 164
nature, 77 free fall, 107-109
nonhomogeneous linear equation, 93 independent of x, 106
numerical method, 127 launch to Jupiter, 114
Runge-Kutta, 135 linear, ll5
order 3, 151 nonhomogeneous, 166
ordinary, 80 Riccati equations, 115
oscillation examples, 149 Riccati method, 117, 118
partial, 80 space probe, 112
power series, 182 Diffusion in a rod, 452
reservoir, 77 Diffusion in a semi-infinite strip, 613
second-order Dimension of a vector space, 298
numerical method, 204 Direct Laplace transform, 223, 228
singular point, 189 Directed line segment, 366
solution, 79 Direction field, 80, 82, 127
solution definition, 156 Direction field of a vector field, 404
solution structure, 155 Directional derivative, 406
sources, 79 Dirichlet kernel, 503
Taylor method, 132 Dirichlet problem, 461
transfonn method for systems, 263 Dirichlets integral fonnula for Sn, 504
variable coefficients, 142 Dirichlets limit theorem, 515
variation of parameters, 93 Discontinuous input, 156
Differential equation of order n Discriminant, 165
basis, 212 Disk heat diffusion, 570
basis recipe, 215 Disk heat diffusion radially symmetric case
coefficients, 212 Bessel, 571
constant, 214 eigenpairs, 571
different bases, 213 equation, 570
fundamental theorem, 212 initial values, 571
homogeneous, 212 product solutions, 571
independence, 212 separation, 570
nonhomogeneous, 214 superposition, 571
nontrivial, 212 uniqueness, 570
solution definition, 211 Distance in space, 371
trivial solution, 212 Divergence, 407, 409, 421
uniqueness, 213 Domain of definition, 83
Differential equation of order one Door stopper, 140
cylindrical reservoir, 88 Dot product, 372, 373
elastic column, 85 Double arrow notation, 285
flask cooling, 87 Double Fourier series, 633
natural reservoir, 88 Drum, 463
 INDEX 715

circular, 467 Energy conservation theorem, 448

Newtons laws, 465 Energy of a vibrating string, 447
perfectly flexible, 463 Equation
problem, 466, 638 eqUivalent, 41
Rayleighs principle, 639 linear system, 277
rectangular and other, 467 solution, 41
tension vector, 464 Wheatstone bridge, 277
thin uniform membrane, 463 Equation of a line using determinants, 344
uniqueness, 466 Equivalent
Drum, circular geometric vectors, 366
uniqueness, 639 linear systems, 285
Drum, circular nonsymmetric parameterizations, 383
general problem, 641 Error function, 40, 574
implicit conditions, 641 complementary, 574
product solutions, 641 erf,26
superposition, 641 Escape velOcity, 113
Drum, circular symmetric Euclidean plane, 366
Bessel function of order zero, 640 Euclidean space, 370
figure, 642 Euler and d'Alembert, 462
formulas, 640 Euler equation, 189
problem, 639 Euler formulas, see Fourier, 496
product solutions, 640 Euler method
superposition, 640 error, 207
Drumhead, see Drum, 638 Euler method for systems, 207
Drums and membranes, 462 Eulers method, 128
Dual basis, 433 error, 130
Even and odd, 507
Eccentric anomaly, 61 Even extension, 532, 589
Eigenvalue, 345 Exact differential equation, 98
complex numbers, 348 Existence of independent solutions, 158
existence, 348 Existence-uniqueness, 124
matrix products, 362 Extended derivative rule for transforms, 247
square matrices, 348
vibration frequency, 345 Faraday, M. and Maxwell, ].e., 403
Eigenvector, 345 Finite difference method, 128, 206
Eigenvector bases, 350 First shifting rule, 236
Eigenvector linear independence, 349 Fixed point method, 44
Electrical oscillator, 255 advantages, 51
Elementary row operations, 285 convergence, 45
Ellipse error, 50
area of a sector, 60 fixed point, 44
Kepler's equation, 59 flask cooling, 45
parameters, 59 nonlinear resistor, 47
Ellipse parameterization, 384 Flask cooling, 91
Ellipsoid parameterization, 403 fixed points, 54
Embedded systems, 20 implicit functions, 66
716 INDEX

problem, xvi Fourier ring problem, 565

stirring effects, xix Fourier series
Flow lines, 405 convergence, 480, 482, 488
Flow lines of vector fields, 415 cosine, 486
Flow meter, 19 Dirichlet problem, 482
sample data, 3 Euler formula, 480
Flow out of a parallelepiped, 408 orthogonality, 479, 514
Fluid flow, 67, 416 sectional continuity and derivatives, 480
Flux of a constant field, 414 Fourier transform, 528
Forward difference, 13 Fourier's heat conservation law, 453, 457,
approximation, 31 468
Four steps in mathematical analysis, xv Fourier's heat equation, 451, 453, 469
Fourier Fourier's heat flux law, 452, 457, 468
Bessel inequality, 505 Fourier's laws of heat conduction, 452
coefficients, 496 Fourier's method, 557
complex series, 500 Fourier, ].B., 451, 473
convergence, 496, 501, 506, 508, 510, Fourier-Bessel series, 551, 572
511 Fredholm alternative, 320
Dirichlet integral formula, 504 Frenet, E and Serret, ].A., 387
Euler formulas, 502, 508 Frenet-Serret formulas, 387
even and odd functions, 507 Frobenius method, 191
integral, 529 Frobenius theorem, 193
integral formula for partial sums, 503 Frobenius, EG., 191
integral theorem, 531, 532, 553 Fubini theorem, 573
other periods, 508 Full wave rectifier, 248
partial sums, 503 Fundamental theorem of calculus, 77, 417
period 2L Euler formulas, 509 Fundamental theorem of linear algebra,
ring problem, 492 314,320
sine coefficients, 479 Fundamental theorem on existence, 124
sine series, 474 Fundamental theorem on
sum of cosines, 503 existence-uniqueness, 156
Fourier cosine series, 486
Fourier integral, 516 Galileo's law, 108
complex form, 527 Gauss algorithm, 287
convergence, 524 Gauss's theorem, 422
definition of, 516 Gauss, Karl Friedrich, 61
directly verified, 517 Gauss-Jordan algorithm, 290
Euler coefficient formulas, 516 Gaussian elimination, 284
even damped oscillation, 521 Gaussian quadrature, 26
even exponential, 520 Geodesy, Handbook of, 290
Gibbs phenomenon, 518 Geometric and algebraic multiplicity, 349
odd extension, 521 Geometric classification of linear systems,
Riemann-Lebesgue, 526 293
single-pulse square wave, 517 Geometric series, 71, 74
theorem, 516 Geometric vectors are a vector space, 367
Fourier integral theorem, 572 Geometry of space curves, 385
 INDEX 717

Gibbs phenomenon, 518 Helix, 384

17.9 percent, 519 Heun's method, 134
example, 519 Higher-order equations, 139, 149
Gibbs,Josiah Willard, 365 Hill's equation, 143
Gradient, 406 Hill, George William, 143
interpretation, 406 Homogeneous, 93,155,211
maximum increase, 406 Homogeneous constant coefficient
Gradient theorem, 417 equation, 164
Gram-Schmidt basis construction, 316 Homogeneous equation, 156
Green's theorem, 417, 420 Homogeneous linear system, 281
Green, George, 417 Homogeneous rod, 452
Gregory-Newton, 13 Hooke's law, 141, 278
Growth and decay, 96 Hyperbolic basis and exponential basis, 602

Half-life, 97 Identities for vector operations, 378

Half-plane heat diffusion with Dirichlet Identity matrix, 308
conditions, 607 Impedance, 117
Half-plane numerical results, 628 Implicit function, 64
Halley's comet, 62 application, 98
Harmonic function, 415 applied to first order, 84
Heat-conducting body, 468 flask cooling, 66
Heat conduction in a rod, 452 fluid flow, circular cylinder, 67
Heat diffusion theorem, 65
disk, 570, 615 uniqueness, 66
infinite rod, 512, 572 Improper integral, 220
nonhomogeneous finite rod, 533 Increasing and decreasing functions, 55
one dimension, 557 Independent
plates, 456, 598 vector test, 299
rectangular plate, 598 vectors, 296
semi-infinite rod, 531 Indicial equation, 190
semi-infinite rod, 576 Indicial polynomial, 193
solids, 468 Induction, 184
sphere, 567 Infinite strip, 611
Heat equation, 451 Infinite sums, 242
Heat equation in 3D, 567 Initial value problem, 123
Heat equation in polar coordinates, 567 fundamental theorem, 123
Heat flow in a semi-infinite rod, 529 Inner product, 315, 372
Heat flow in plates, 456 Input voltage, 156
Heat flow in rectangular and circular Integral operator, 220
plates, 630 Integral rule, 235
Heat flow in solids, 468 Integral rule for transforms, 240
Heat kernel, 656 Integral test, 73
Heat transfer, see Flask, 41 Integrating factor, 102
Heaviside function, 221 Interpolation
Heaviside's coverup method, 229 error control, 10
Heaviside, Oliver, 219 flow meter data, 5
718 INDEX

higher-order polynomial, 11 solvability, 61

error, 14 Kepler, Johannes, 58
uniformly spaced nodes, 13 Kinetic energy, 447
interpolate, 2 Kirchhoff
inverse, 56 current law, 149
linear, 4 laws, 117
uniformly spaced nodes, 5 voltage drop formulas, 139
linear algorithm, 5 voltage law, 150,278
linear error, 6
nonlinear resistor, 19 Lagrange, Joseph Louis, 161
polynomial, 2 Laguerre polynomial, 246
quadratic, 7 Laplace equation
flow meter data, 8 polar coordinates, 147, 191
uniformly spaced nodes, 9 space, 147,470
quadratic algorithm, 8 spherical coordinates, 148
quadratic error, 9 Laplace reduction formulas, 230
quadratic spline, 15 Laplace transform
spline, 14 s-derivative rule, 239
Inverse alternative second shifting rule, 236
definition of, 56 applications, 249
function theorem, 55 circuit, 234
functions, 55 convolution, 237
Inverse table, 63 derivative rule, 231
Invertible matrix, 309 differentiation, 239
Invertible matrix test, 310 direct, 220
Irrotational field, 415 elementary functions, 224
Isobaric map, 403 existence, 221
Isobars, 98 exponential and trigonometric, 22 7
Isomorphic physical problems, 177 exponential rule, 225
Isomorphic system table, 178 first shifting rule, 236
Isomorphic systems, 141 Heaviside function, 221
Isothermal map, 403 infinite rod, 654, 656
Isotherms, 98 integral rule, 235
Iteration, 45 integration, 239
Lerch's theorem, 222
Jordan, Wilhelm, 290 linearity, 221
Jupiter launch, 114 nature, 220
Jupiter rendezvous, 122 operational rules, 231
partial differential equation, 653
Kepler's equation, 58 partial fractions, 227
derivation, 61 periodic, 243
eccentricities, 62 power series, 231
Halley's comet, 62, 64 second shifting rule, 236
laws, 58, 60 sectionally defined, 240
Mercury, 62, 63 semi-infinite rod, 656, 657
Pluto, 64 semi-infinite string, 658, 659
 INDEX 719

sine integral, 235 critically damped, 120

systems of differential equations, 263 damping, 120
three properties, 224 general solution, 118
transfer function, 257 overdamped, 120
trigonometric powers, 226 physical interpretation, 120
trigonometric rules, 226 Riccati method, 118
undetertnined coefficients, 228 underdamped, 120
uniqueness, 222 Linear second-order equations with
Lateral thermal insulation, 452 analytic coefficients, 182
Lateral vibrations, 442 Linear systems
Law of cosines, 373 abstract vector space, 302
Leading pulse, 588 algebra of, 282
Leaf motion, 415 augmented matrix, 284
Left-moving pulse, 588 back substitution, 288
Legendre equation, 149, 154, 189, 199 basis for n-space, 299
Legendre functions, 199 basis theorem, 297
finiteness, 201 CAS examples, 290
Legendre polynomial recurrence, 204 chemical balance, 279
Legendre polynomials, 189,551 chemical balance example, 301
Legendre series, 551 complex entries, 283
boundedness condition, 551 complex field, 323
heat diffusion in a circular sector, 552 complex numbers, 292
orthogonality, 551 complex vector spaces, 303
polynomials, 551 composition, 306
representation, 551 determinants, 326
Legendre, A.M., 149 differential equation, 302
Length of a curve, 385 dimension, 298
Level curves, 403 dimension theorem, 298
Level set equations, 105 double arrow notation, 285
Level surfaces, 404 eigenvalues and eigenvectors, 345
Limit, 382, 395 elementary row operations, 285
Line equation in space, 373 empty solution set, 287
Line integral, 388, 389, 391 equivalence, 285
Line parameterization, 384 Fredholm alternative, 320
Linear, 92, 155 fundamental theorem, 314
Linear m x n systems, 280 fundamental theorem of linear algebra,
Linear algebra, 277 320
Linear combination of solutions, 156 Gauss algorithm, 287
Linear combination of vectors, 295 Gaussian elimination, 284
Linear dependence geometry, 286
differential equations, 156 homogeneous and non-homogeneous,
Wronskian, 157 281
Linear dependence test, 157 independence and dependence, 296
Linear equations of order n, 211 independence test, 299
Linear operator, 221 large, 281
Linear second -order linear combination, 295
720 INDEX

linear operator, 305 Mass inside a surface, 401

linearity property, 282 Mass of a wire, 388
matrix, 281 Mathematical analysis of a physical
matrix algebra, 305 problem, xv
matrix product, 306 Mathematical physics
mechanical equilibrium, 278 transform methods, 653
nonsquare, 280 Mathematical physics applications, 557
null space, 294 Mathieu!; equation, 143
other vector spaces, 302 Mathieu, Emile-Leonard, 143
polynomials, 303 Matrices
range space, 295 adjOint, 319, 327
rank and nullity theorem, 323 and linear algebra, 305
rank theorem, 321 associative and distributive laws, 307
reduced row echelon matrix rref, 289 change of basis, 351
row echelon matrix, 288 characteristic polynomial, 348
rows and columns, 306 column space, 321
scalar product rule, 306 complex, 323
solution structure, 323 composition, 306
solvability theorem, 320 determinant, 326
spaceC[a,b],303 diagonal, 307
space of all functions, 303 diagonalization, 352, 353, 355
structure of solutions, 294 eigenvalues and eigenvectors, 345,348,
subspace, 295 358
subspace construction, 296 identity, 308
vector matrix notation, 281 invertible, 309-311
vector scalar multiple, 282 linear operator, 305
vector space basis, 297 matrix product, 306
vector spaces, 294 non-Singular, 308
vector sum, 282 noncommutative, 307
Unear systems and Sturm-Uouville orthogonal, 315
problems, 539 row operations, 310
Unearly independent vectors, 297 row space, 321
Unes in a crystal lattice, 432 rows and columns, 306
Upschitz scalar product rule, 306
condition, 46 Similarity, 351
constant, 46 singular, 308
Upschitz, Rudolph, 46 square, 307
Loan payments on a car, 416 transpose, 319
Logistic equation, 129 Matrix notation, 281
Lorenz attractor, 405 Maximal interval of existence, 127
Lorenz system, 405 Maximum principle, 560
Llltzen, ]esper, 473 Mean anomaly, 61
Mechanical equilibrium, 278
Magnitude of a vector, 368 Mechanical oscillator, 140, 151,255
Main Frobenius theorem, 193 Membrane, 463, see Drum, 638
Mass denSity, 512 Mercury, 62
 INDEX 721

Method of Frobenius, 191 Nontrivial solution, 156

Method of isoclines, 82 Nonuniform rod heat diffusion, 154
Method ofreduction of order, 159, 163 Norm of a vector, 368
Method of successive approximation, 45 Normal line, 397
Method of undetermined coefficients, 168 Normal modes of vibration, 346
Method of variation of parameters, 161 Normal plane, 386
Miller indices, 434 Normal to a level surface, 407
Moisture loss, 97 Null space, 294
Moment of inertia, 141 Nullity of a matrix, 323
Monotone function, 55, 56 Number theory and determinants, 344
Multiloop circuit, 149 Numerical algorithm, 1
Multiplicity, 349 Numerical analysis, 1
Music, 442, 593, 594 Numerical differentiation, 30, 31, 34
Numerical integration, 23
Natural oscillations, 172 error control, 26
Nature of the Laplace transform, 220 other methods, 26
Near point, 380 Numerical method
Net flow rate, 408 adaptive, 136
Neumann problem, 462 second-order, 204
Newton Numerical methods
algorithm, 52 direction field, 127
inverse square law, 112 Euler method for systems, 207
law for rotating bodies, 141 finite difference method, 128
law of cooling, xvii, 41 finite differences, 206
laws applied to strings, 444 first order system, 205
method, 51, 53 other methods, 208
second law, 141 Runge-Kutta 4, 207
Newton, Isaac, 77 initial value problem, 127
Nodal lines, 643 second order
Nodal values, 3 initial value problem, 204, 205
Nodes, 3 unique continuation, 204
Nonhomogeneous linear system, 281 uniform mesh finite difference, 206
Nonsingular matrix, 308 Numerical series identities, 492
Nonuniqueness of initial value problems, n-vectors, 281
126
Nonorthogonal bases and their duals, 432 Odd extension, 530, 587, 592, 593
Nonuniform string, 450 Ohms law, 278
Nonhomogeneous, 93,155,211 Open box parameterization, 402
Nonhomogeneous equation, 158 Operational rules for the transform, 231
Nonlinear, 92 Orbital point, 61
Nonlinear resistor, 16 Order of a Bessel equation, 147
fixed-point method, 47 Order of a numerical method, 132
fixed-point table, 49 Orientable surface, 407
inverse interpolation, 56 Orthogonal
Newton's method, 52 complement, 318
sample data, 18 sets, 316
722 INDEX

subspaces, 317 other boundary conditions, 459

vectors, 315 problem, 458, 461, 462
Orthogonal bases, 316 radiation and cooling, 460
Orthogonal sums, 317 steady-state heat, 460
Orthogonality uniqueness, 459, 461
Fourier coefficients, 479 Plates
Fourier integral coefficients, 514 circular disk diffusion Neumann
relation, 479 product solutions and superposition,
sines and cosines, 495 620
Orthonormal,316 solution formulas, 621
Oscillator solvability, 621
electrical and mechanical, 255 heat diffusion, 456, 598
transform method, 265, 268 limitations, 598
Oscillators other and polar coordinates, 626
damped with periodic square wave rectangular Dirichlet, 598
input, 258 rectangular with transient diffusion, 630
damped with sinusoidal input, 257 Plates circular disk diffusion
homogeneous equation, 255 angle interpretation, 616
nonhomogeneous equation, 256 caution for Poisson's integral, 620
Other vector spaces, 302 convergence, 618
Outer product, 376 finiteness condition, 617
Overtones and octaves, 593 periodic Sturm-Liouville problem, 617
Poisson integral, 619
Parallelogram rule, 366 polar coordinates, 616
Parameterization nonuniqueness, 393 problem, 615
Paratrooper, 210 product solutions, 616
Partial fraction coverup method, 229 superposition, 618
Partial fractions, 22 7 Plates circular disk diffusion Neumann
Pendulum of variable length, 143 problem, 620
Perfectly flexible, 442, 463 Plates circular sector diffusion Dirichlet
Perihelion, 61 v(O,O,h) 626
Period,493 yCo,g,O) , 625
planet, 60 vCf,O,O): 623, 624
Periodic boundary conditions, 493 decomposition into three problems, 623
Periodic function, 242, 493, 501 problem, 622
extensions, 502 Plates half-plane Dirichlet
properties, 501 Fourier integral, 609
transform rule, 243 Heaviside, 610
Permutation matrix, 313 Poisson integral, 610
Phase lag, 95 problem, 607
Pitt, H.R., 527 product solutions, 608
Plane superposition integral, 608
equation in space, 374 uniqueness and boundedness, 610
Planes in space, 286, 392 Plates half-plane Neumann
Plate integral solution, 611
applications, 462 problem, 610
 INDEX 723

Plates infinite strip Dirichlet Power series method, 182

diffusion problem, 611 Principal normal vector, 386
Fourier integral, 612 Principal value, 57
separation and product solutions, 612 Probability vector, 283
superposition integral, 612 Product solution, 475
Plates rectangular Product solutions of the wave equation, 445
construction of v, 601 Pulses, 588
construction of w, 603 Put the problem on the computer, xvi
convergence, 600 Puzzle, 77
Fourier series, 599
general, 600 Radial heat equation, 578
Neumann and mixed, 603 Radial heat equation on a disk, 570
product solutions, 598 Radiocarbon dating, 97
separation, 598 Radius of convergence, 188
superposition, 599 Radius of the earth, 112
superposition to obtain v, 602 Range space, 295
use of Churchill's theorem, 606 Rank of a matrix, 321
Plates rectangular mixed Ratio test, 188
problem, 606 Rayleigh's principle, 639, 642
product solutions, 606 Rayleigh, Lord, 639
sine series solution, 607 Recipe for constant equations, 164
Plates rectangular Neumann Recipe for Euler equations, 190
Fourier cosine series solution, 605 Reciprocal basis, 391
problem, 604 Rectangular drum, 643
product solutions, 604 Rectangular plate
solvability, 605 double Fourier series, 633
superposition, 605 transient diffusion
uniqueness up to a constant, 605 coefficients, 633
Plates semi-infinite strip Dirichlet convergence, 633
u(O,O,h) 615 eigenvalues, 630
u(O,g,o)' 615 orthogonality, 632
u(j,o,o): 614 product solutions, 630, 632
Churchill's theorem, 615 Sturm-Liouville problem, 631
decomposition u(j,g,h), 614 superposition, 632
problem, 613 transient diffusion with zero boundary
Plucked infinite string, 585 conditions, 630
Plucked string, 597 Recurrence relations, 198, 203
Plucking a string, 446 Recursion formula, 193
Poisson integral, 610 Recursion relation, 184
Position vector, 373 Reduced row echelon form, 289
Potential energy, 447 Reduction of order, 159
Power series Refinery, 294
Cauchy product, 183 Reflected wave, 588
differentiation, 183 Reflected X-ray beam, 438
rules, 182 Reflections of beams by a crystal plane, 435
vanishing rule, 183 Regression line, 325
724 INDEX

Regular domain decomposition, 420 heat diffusion and Fourier series, 474
Regular domain in 3-space, 425 insulated at the ends, 455
Regular singular point, 189 nonuniform, 455
Reservoir product solutions, 474
consistency of volume data, 38 radiation at the ends, 455
depth data, 27 separation constant, 475
draining times, 90 series truncation, 559
filling of, 2 sine series, 559
numerical differentiation, 35 superposition, 477
sample depth and volume data, 30 truncation error, 559
water stored, 29 uniform convergence, 559
Resistive network, 283 uniqueness, 454, 486
Resonance, 174 zero heat flux, 486
Riccati, 140 Rod finite
equation, 115 ends insulated, 560
method, 164 heat content, 512
Richardson extrapolation, 33, 34, 40 mixed, 562, 565
Riemann's theorem, 505 mixed tangent equation, 564
Riemann-Lebesgue theorem, 526 nonhomogeneous, 533
Right-hand and left-hand limits, 481 numerical bounds and error, 561
Right hand rule, 376 separation, 534
Right-moving pulse, 588 Sturm-Liouville series, 537
Ring Rod infinite
eigenpairs, 566 alternate fonnula, 573
error estimates, 566 cosine integral, 573
orthogonality, 495 double integral, 573
periodic boundary conditions, 493 eigenvalue, 512
periodic conditions, 492 error function, 574
periodiC eigenvalue problem, 493 heat diffusion, 512, 572
problem, 492, 565 no explicit boundary conditions, 512
solution of Fourier's problem, 496 separation, 512, 572
superposition, 495 superposition, 513
uniqueness, 500 Rod Neumann
Rk4, see Runge-Kutta, 135 constant initial gradient, 489
Robin problem, 462 constant initial temperature, 488
Rocket, 122, 163 problem, 486-488
Rocket launch, 112 sawtooth, 490
Rod sectionally constant, 489
boundary value problem, 454, 474 superposition, 487
cosine series, 559 Rod semi-infinite
cosine series and orthogonality, 487 boundary at zero temperature, 576
cross-sectional area, 512 Dirichlet diffusion, 531
Dirichlet eigenvalue problem, 475 integral formula, 576
ends at zero, 558 Neumann problem, 531
even and odd extensions, 559 Rod semi-infinite, 529
fundamental solution K, 575 Dirichlet diffusion, 529
 INDEX 725

Dirichlet fundamental solution, 577 semi-infinite string, 659

heat diffusion, 576 Sectional continuity on the line, 516
Neumann fundamental solution, 577 Sectionally
zero flux, 577 continuous, 156
Rod sine series defined functions, 240
constant gradient, 484 Sectionally continuous functions and
constant temperature, 483 derivatives, 480
parabolic profile, 485 Selfadjoint matrix diagonalization, 355
sawtooth, 484 Semi-infinite rod, 531
solution, 483 Semi-infinite string by Laplace methods,
Roentgen, We., 431 658
Rotating fluid example, 411 Semi-infinite string with distributed load,
Row echelon matrix, 288 661
Row operations on a matrix, 310 Semi-infinite strip diffusion with Dirichlet
Row space, 321 condition, 613
Rows of a matrix, 306 Separation constant, 475
rref,289 Separation into pulses, 588
history; 290 Series, 70-73
Runge's example, 21 Shear, 463
Runge, K. and Kutta, W, 135 Shock absorber, 140
Runge-Kutta 4 method, 135,207 Similarity of matrices, 351
Runge-Kutta-Fehlberg numerical method, Simply connected, 100
136 Simpson
error estimate, 25
Salt crystals, 438 reservoir application, 89
Sarruss rule, 342 Thomas, 25
Sarruss rule warning, 342 Simpsons rule, 25
Sawtooth wave, 156,248 Sine integral, 235
Scalar field and L. Euler, 403 Singular matrix, 308
Scalar multiple of a matrix, 305 Singular pOint, 189
Scalar multiple of a vector, 282 Singular Sturm-Liouville problems, 548
Scalar product, 372 Singularity of Legendres equation, 551
Scalar product in n dimensions, 315 Smooth space curves, 383
Scalar product rule, 306 Smooth space surfaces, 395
Scalar quantity, 367 Snapshots of heat diffusion solutions, 486,
Scalar triple product, 377 491
Schwartz inequality, 315 Solids
Second-order boundary value problem, 469
constant coefficient, 164 cube, 646
differential equation, 106 divergence theorem, 469
linear differential equation, 155 heat conservation law, 468
nonhomogeneous, 166 heat diffusion, 468, 643
Second shift rule heat flow, 468
alternate, 236 heat flux laws, 468
Second shift theorem homogeneous material, 468
rule, 236 Laplace equation, 470
726 INDEX

other heat diffusion problems, 469 Sturm-Liouville, 568

sourceless body, 468 superposition, 569
steady-state diffusion, 470 Spherical cap, 401
uniqueness, 469 Spherical coordinates, 147
Solids parallelepiped Spring-mass structure, 278
coefficients, 646 Square matrices, 307
mixed problem, 647 Square plate Neumann diffusion problem,
product solutions, 644 604
steady-state heat, 644 Square wave, 156,244
superposition, 646 Standard basis series calculations, 202
Solids sphere Static sources F(x), 145
Legendre polynomials, 650 Steady-state current, 96
Neumann conditions, 652 Steady-state diffusion in plates, 598
product solution formulas, 651 Steady-state heat diffusion in
product solutions, 649 parallelepipeds and spheres, 643
special substitution, 649 Steady-state heat diffusion in solids, 470
standard Sturm-Liouville problem, 650 Steady-state heat in a plate, 460
steady-state diffusion, 648 Steady-state temperature in a disk, 147
superposition solution, 651 Steady-state temperature in a sphere, 147
Solution, 156 Stereographic projection, 394, 401
Solution basis, 158 Stokes' theorem in 3-space, 425
Solution set, 41 Stokes, George Gabriel, 425
Solution structure, 158 Streamlines, 67, 69, 98, 106,405
Sourceless LRC circuit, 178 Striking a string, 446
Space CIa, bl, 303 String
Space curve, 383 d'Alembert, 582
Space curve examples, 384 d'Alembert solution from the integral
Space curves, 381 from, 584
Space curves in three dimensions, 381 even extension, 589
Space of n-vectors, 281 finite, 590
Space of all functions, 303 fixed at the right, 449
Space probe, 112 infinite taut, 581
Space triangle parameterization, 402 integral form, 584
Special relativity, 92 musical instruments, 593
Specific heat, 512 plucked, 585
Sphere problem, 442
diffusion, 567 separation, 583
heat equation, 567 uniqueness, 581
polar equation, 567 String finite
steady-state heat diffusion, 648 uCO,g) 592
Sphere symmetric u(j,O) , 590

diffusion, 567 general solution, 593

Euler coefficients, 569 odd periodic extension, 592
orthogonality, 569 other boundary conditions, 594
product solutions, 569 periodic vibrations, 594
separation, 568 String instruments, 593
 INDEX 727

String oscillations, 143 Sums of matrices, 305

String semi-infinite Superconducting circuit, 172
boundary condition, 586 Superposition for diffusion in a rod, 477
d'Alembert, 586 Superposition of product solutions, 477
initial condition, 586 Surface area, 397
Neumann problem, 589 Surface area increment notation, 400
odd extension, 587 Surface geometry, 395
physical interpretation, 588 Surface integral, 401
separation, 588 Surface integral interpretation, 407
sine integral solution, 589 Surface of revolution, 399
three stages of propagation, 588 Surfaces in space, 392
traveling waves, 585 Symmetric interval, 502
uniqueness, 585 Synchronous oscillations, 266
Structure of solutions, 158 Synchronous vibrations, 145, 346, 593
Structure theorem for second order, 159 Systems of linear equations, 277
Stunn, c.F. and Liouville, j., 473
Sturm-Liouville Tangent crossing, 543
basic theory, 534 Tangent plane, 396
Bessel equation of order zero, 549, 550 Tangent vector, 384, 396
boundary condition, 533 Tangential and nonnal components of
Dirichlet and Neumann, 533 acceleration, 390
Dirichlet example, 540 Taylor
eigenvalue problem, 533 series expansions, 182
Euler differential equation, 544 Taylor 2 algorithm, 132
existence of eigenvalues and Taylor 4 algorithm, 133
eigenfunctions, 534 Taylor's method, 132
Fourier-Bessel series, 549 error, 134
general boundary conditions, 537 Telescoping sum, 504
graphical solution of eigenvalues, 543 Temperature in a nonuniform rod, 144
heat diffusion, 537 Tension, 463
Legendre series, 551 Tension vector, 443, 464
mixed example, 541 Tenninal speed, 108
Neumann example, 541 Tetrahedron parameterization, 402
orthogonality, 535, 550 Tetrahedron volume, 345
problem, 534 Thennal diffusivity, 474
series expansion, 536 Thennally insulated, 456
singular, 548, 549 Thin elastic membrane, see Drum, 463
substitution method, 545 Thin flat plate, 456
theory, 533, 539 Thin string, 442
variable coefficients, 544 Thin unifonn membrane, 463
Stunn-Liouville continuous spectrum, 552 Third-order system, 211
eigenfunction expansion, 553 Three stages of propagation, 588
eigenvalues, 552 Topographical map, 403, 414
Subspace construction, 296 Torricelli's law, 87, 91
Summation Torricelli, Evangelista, 87
series, numerical, 70 Torsion of a space curve, 387
728 INDEX

Torsional pendulum, 141 Unit sphere parameterization, 393

Total compression, 85 Unsolvable numerical problem, 210
Transfer function, 257
Transform method for partial differential Variable coefficients, 142, 155
equations, 653 Variable tension, 143
Transform rules, 231 Variation of parameters, 93, 161
Transient current, 96 example of method, 162
Transient diffusion in plates, 630 Vector
Transition probability, 283 addition, 367
Transpose matrix, 319 algebraiC, 366
Trapezoidal rule, 23 angle, 372
error, 24 binormal, 386
reservoir application, 89 calculus, 395
Traveling wave, 579 calculus of curves, 381
infinite taut string, 581 calculus or fields, 403
semi-infinite string, 585 Cartesian coordinates, 369
strings, 579 circulation, 410
Trial solution Frobenius method, 192 components, 370, 371
Triangle parameterization, 390 continuity, 382
Triangular wave, 244 cross product, 376
Tridiagonal systems, 292 curl,409
Trigonometric series estimates, 75 curvature, 386, 387
Triple product, 377 derivative, 382
Trivial solution, 156 direct line segment, 366
True anomaly, 61 direction field, 404
Tuning and notes, 594 directional derivative, 406
Twist angle e, 141 displacement, 366
Two-body damped oscillator, 268 divergence, 407
Two-body undamped oscillator, 265 dot product, 372, 373
Two-loop circuit, 149 equivalence, 366
equivalent parameterizations, 383
Uncoupled systems of differential Frenet-Serret formulas, 387
equations, 273 functions, 381
Undetermined coefficients geometrical, 366
addition theorem, 170 geometrical scalar multiplication, 367
examples, 168 geometry of curves, 385
Laplace transform, 228 gradient, 406
method,168 identities, 378, 413
rules, 170 length of a curve, 385
Uniform convergence, 72 limits, 382
Uniform string, 442 line equation, 373
Unique continuation, 204 Lorenz attract or, 405
Uniqueness norm, 368
wave equation, 447 normal and gradient, 407
Uniqueness of surface area, 400 normal line, 397
Uniqueness of the inverse matrix, 310 normal plane, 386
 INDEX 729

notation, 367, 378 Vibrations of drums and membranes, 462

outer product, 376 Vibrations perpendicular to a plane, 463
parallelogram rule, 366 Voltage drop formulas, 139
plane equation, 374
position, 373 Wave equation, 445-447
principal normal, 386 Wave function, 443
right-hand rule, 376 Wave number, 435
smooth curve, 383 Wave vector, 435
smooth surface, 395 Waves
space curve, 383 D'Alembert's theorem, 580
surface area, 397 infinite taut string, 581
surface geometry, 395 product solutions, 581
surface integral, 401 propagation speed, 579
surface of revolution, 399 sums of traveling waves, 581
surfaces, 392 traveling, 579
tangent, 384, 396 uniqueness, 581
tangent plane, 396 Weight function calculations, 538
torsion, 387 Wheatstone bridge, 277
triple scalar product, 377 Wronski, Count, 157
unit sphere, 393 Wronskian function, 157
Vector algebra, 282 Wronskian test
Vector analysis, 365 use in plates, 602
Vector linear combination, 295 Wylie, c.R., v
Vector notation, 281
Vector space Mm,n, 312 X-ray beams in physical optics, 435
Vector spaces, 294 X-ray diffraction, 431
Vector subspace, 295 X-ray diffraction experiment, 437
Vector sum, 282
Vectors, 281 Yeast culture, 97
Vibrating circular membrane, 145
Vibrating string, 442, 579, 590 Zero vector, 281
Vibration absorber, 271 Zeros of 10 (x) and the radial heat equation,
Vibrations of drums, 638 571
Sequence formulas.
1- UN+!
-I-u
- - = I+u+···+UN
N-I
aN - al = ~)a"+1 - an)

t (N)a·~-
n=1

(a+b)N = •
• =0 k

(P)q '" __ p!_ = pep - 1) ... (p - q + 1)

q!(p - q)! q(q - I)···I

Series formulas.
1 00

-=LU"'
I-u "=0
lul<I

(1 + u)P = ~ (:)x·, lui < 1

00 x"
e'=L-
n=O n!

• 00 (-I)"x2"+!
smx = "
;0 -'-;::-'-:--:-:-.,-
(2n + I)!

00 (-I)"x2"
cosx=L---
n=O (2n)!

In(l +x) = L (-I)"x"+1

00

"=0 n+I
' -1 < x ~ 1

tan
-I
x= L00

"=0
(-I)"x2 "+!
2n+I
' Ixl ~ 1

Derivative formulas.
(u")' = nu·-1u'
(uv)' = U'y + Uy'
( ~)'
Y
= U'y - Uy'
y2

(eo)' = u'e"
(In u)' = u'/u
(sin u)' = u' cos u
(cos u)' = -u' sin u
(tan u)' = u' sec 2 u
(cot u)' = -u' csc 2 u
(sinh u)' = u' cosh u
(cosh u)' = u' sinh u
(arcsin u)' = u'(l- U 2 )-112
(arctan u)' = u'/(I + u2 )
Trigonometric formulas.
sin(-O) = -sine, cos(-9) = cos9
sin'9+cos'e = 1
sin (a + b) = sin a cos b + sinbcosa
cos (a + b) = cos a cos b - sinasinb
tan (a + b) = tan a + tanb
1- tanatanb

sin' u = 21 (1 - cos(2u»
1
cos' u = 2(1 + cos(2u»
u I-cosu
tan- = - - -
2 sinu
e"'=cos9+isin9 ( i = H )

Hyperbolic functions.

Integral formulas.

J tanxdx = -In I cosxl + C

J cotxdx=lnlsinxl+C

J I
secxdx = In secx + tanxl + C

J cscx dx = In I cscx - cotxl + C

Jx'+a'
~ ~arctan~+C
=
a a
J.... x'+a'
~=sinh-l~+c
a

J.Ix' - a' Jx' -a')

_ _ = In (x +
_dx +C

J'/a'-x' = a +
_dx
__ arcsin ~ C

J Inxdx=xlnx-x+C

J e""sin(bx)dx= a2:b2(aSinbx-bcosbx)+C

f t!"'·
e"" cos(bx)dx = a' + b' (acosbx + bsmbx) + C