Brauer, Fred - Nohel, John A. - Ordinary Differential Equations - A First Course

a i
—
i
hel
Digitized by the Internet Archive
in 2022 with funding from
Kahle/Austin Foundation
https ://archive.org/details/ordinarydifferenOOOObrau
ORDINARY
DIFFERENTIAL EQUATIONS:
A FIRST COURSE
Second Edition
FRED BRAUER
JOHN A. NOHEL
University of Wisconsin
os
Oe
W. A. BENJAMIN, INC.
Menlo Park, California * Reading, Massachusetts
London « Amsterdam « Don Mills, Ontario * Sydney
UNIVERSITY MATHEMATICS SERIES
Consulting Editors
FRED BRAUER
JOHN A. NOHEL
Copyright © 1973 by W. A. Benjamin, Inc. Philippines copyright 1973 by W. A. Benjamin, Inc.

All rights reserved. No part of this publication may be reproduced, stored in a retrieval system,
or transmitted, in any form or by any means, electronic, mechanical, photocopying, recording,
or otherwise, without the prior written permission of the publisher. Printed in the United States
of America. Published simultaneously in Canada. Library of Congress Catalog Card No.
72-9313.
ISBN 0-8053-1208-0
ABCDEFGHIJ-MA-79876543
Preface
Differential equations are essential for an understanding of many important

physical and mathematical problems. This was first recognized by Sir Isaac
Newton in the seventeenth century, and he used them in his study of the
motion of particles and planets. The development of the subject as a branch
of modern mathematics came in the nineteenth and twentieth centuries
through the pioneering work of a number of mathematicians, notably Birk-
hoff, Cauchy, Lyapunov, Picard, Poincaré, and Riemann. This theoretical
foundation, which has been and is being developed by other mathematicians,
has in turn led to interesting new applications—such as control and guidance
systems—and to a deeper understanding of complex physical systems.
This is a textbook for an introductory course in differential equations
for students who have had a sound calculus course, including functions of
several variables, but with no previous experience in differential equations.
There is an adequate amount of material for courses of various lengths
(quarter, semester, two-quarter, or even a year). By choice of topics and
emphasis on the part of the instructor, one can use this book in either an
honors or a standard course.
In this second edition of our “First Course,” we again use physical
problems to motivate the study of differential equations. First-order equa-
tions, introduced by means of a simple population-growth problem, are
studied in Chapter 1. Second-order equations, motivated by the simple mass—
spring system and the simple pendulum, are studied in Chapter 2. Rather
than proceeding directly to the existence and uniqueness theorems in each
of these chapters (as we did in Chapter | of the first edition), we first study
some simple methods of solution with a large number of worked-out ex-
amples, exercises, and simple applications. We have, as before, avoided an
exhaustive description of such methods, but at the same time we feel we
have not omitted the essential ones. We believe from experience with the
iii
iv Preface
first edition that this procedure paves the way for a better comprehension
of the basic existence and uniqueness theorem presented without proof but
with numerous examples only at the end of each of these chapters. The
proofs of these theorems are given in Chapter 8.
Chapters | and 2 form the core of any course taught from this book.
Students with no previous training in linear algebra, who want a basic”
course, should follow Chapter 2 with Chapter 3 (linear differential equa-
tions). Such a basic course would ordinarily omit the two chapters on linear
systems (Chapters 4 and 5), but should include some of the topics in the
later chapters. We suggest that at least parts of Chapters 6 (series solutions)
and 7 (boundary value problems) are particularly appropriate. Students who
are already familiar with the elements of linear algebra should replace Chap-
ter 3 with Chapters 4 and 5. All the main results of Chapter 3 are included
as special cases in Chapters 4 and 5.
We have taught much of the material in this book in a one semester
course at the sophomore-junior level at the University of Wisconsin in
Madison, meeting three times per week. Approximately three weeks were
spent on each of Chapters | and 2, and four weeks were spent on Chap-
ter 3. The remainder of the course was devoted to a selection of material
from Chapters 6 and 7. For more theoretically oriented students, Chapter 8
might well be covered instead. In modern applications of differential equa-
tions, a significant role is played by high-speed computers. For this reason,
we have included an introduction to numerical methods of solutions in
Chapter 9. This may be of use to students interested primarily in engineering
applications. Such students might also profit from the introduction to
Laplace transforms in Chapter 10.
Chapters | and 2, together with either Chapter 3 or Chapters 4 and 5,
form the basic introduction on which the remainder of the book depends.
Although physical problems are used throughout for motivation and illus-
tration, the treatment is self-contained and does not depend on any knowl-
edge of physics. The last five chapters are almost completely independent,
except that Bessel functions, covered in Chapter 6, enter in Section 7.6. How-
ever, this section could be omitted if necessary in a study of Chapter 7. The
only places in the last five chapters which depend on Chapters 4 of 5 are the
last sections of Chapters 8 and 9, which could be omitted by students who
have not covered Chapters 4 and 5.
Much of the modern theory of differential equations can be explained
properly and efficiently only with the aid of linear algebra. Only a minimal
knowledge of linear algebra is essential for a proper understanding of Chap-
ters 4 and 5. We believe that emphasis on linear algebra is both important
and consistent with current and future trends in the mathematical training
of engineers, physical scientists, and engineers. We heartily endorse the idea
of a curriculum which introduces linear algebra and thus makes it possible to
Preface Vv
study Chapters 4 and 5 of this book in a first course in differential equations.

For students interested in the qualitative theory of nonlinear differential
equations, a knowledge of linear algebra is indispensable. However, we do
provide Chapter 3 as an alternative for the student not acquainted with
linear algebra, and as we have indicated, most of this book is accessible to
such a student.
The principal changes from the first edition are as follows: (a) Chapters
1 and 2 have been completely rewritten, at a considerably more elementary
level, (b) Chapters 4 and 5 and the accompanying appendixes are new (most
of this material was, however, contained in Chapter 6 of the first edition in
a different form). They are based essentially on material from our book on
Linear Mathematics, W. A. Benjamin, Inc., 1970, written jointly with our
colleague Hans Schneider, and we are particularly grateful to him for per-
mitting us to incorporate this material into this book. (c) Sections 5.5 and
8.4, which represent interesting applications, are completely new. (d) The
exercises throughout the book have been rearranged, and new problems
have been added. Rather than placing review exercises at the end of each
chapter, where they remain for the most part untouched, we have incorpo-
rated them into the relevant sections. (e) The appendixes on certain parts of
linear algebra, together with the exposition in the text, make it possible for a
student to follow Chapters 4 and 5 and the last sections of Chapters 8 and 9
with only a minimal background of linear algebra.
The formulation of mathematical models for physical problems is often
misunderstood. In our view, a logically reasonable procedure is the follow-
ing. We start from a concrete physical situation (such as the pendulum),
pass to an “‘idealized” physical model (such as a rod of zero weight, a pivot
of zero friction, Newton’s laws assumed, no air resistance), and then, using
the “‘idealized”’ physical model, construct a mathematical model. The con-
struction of a mathematical model is often difficult and requires much prac-
tice. However, this important aspect of the procedure will not be our main
concern. The mathematical model could be as simple as a quadratic equa-
tion, or so complex that it would take many pages to describe. The next step
is to use mathematical techniques to explore properties of the model, and to
answer specific questions. If such questions can be answered, we are then
able to compare the mathematical model with physical reality as observed
by experiment, thereby testing the validity of the laws assumed in making
the physical approximation. The process may be described by Fig. 0.1.
The choice of physical laws and approximations and the construction
of mathematical models belong more to science than to mathematics. We
shall, therefore, concentrate mainly on the techniques used to pass from the
mathematical model to the prediction.
It is very important not to confuse the model with reality. A model is most
useful when it imitates reality closely, but there will always be aspects of re-
vi Preface
F Comparison
Physical
Prediction
reality
"Laws" Techniques
Physical Mathematical
approximation model
Fig. 0.1
ality which it does not reproduce, and it will always predict events that do
not, in fact, occur. The skill of a scientist lies in knowing how far, and in
what context, to use a particular mathematical model. One simple illustra-
tion may help here. Physicists sometimes speak of light as a wave, and some-
times as a particle. Which is it? The answer is neither, for both are names
of specific mathematical models for light. Both successfully predict (‘‘ex-
plain’) some of the observed behavior of light, but both predict behavior
that light does not exhibit. The moral is clear: Do not expect to have exactly
one correct mathematical model for any aspect fo reality. Newton’s “‘laws
of motion” are not the only “correct” ones (for example, at very high
velocities, Einstein’s laws are closer to reality), nor is Hooke’s law the only
correct law of elasticity.
Because of the way mathematics is often taught, a student is apt to think
of it as merely a collection of techniques, tricks and skills which some pro-
fessor wishes him to learn. As his experience grows, the person who does
research in other physical sciences will learn that for him mathematics is
something entirely different. He will learn to think of mathematics as a tool
which can explain and help him understand various phenomena in the
physical world. At the same time, it is very important for the prospective
mathematician to become acquainted with those mathematical problems
which are closely connected with applications, as many of these problems
are of considerable mathematical interest in their own right. It is our hope
that this book will serve these various needs.
It is a pleasure to acknowledge the help—direct and indirect—from col-
leagues, teaching assistants, and students, particularly at the University of
Wisconsin; their comments aided us in the preparation of this second edition.
We are particularly grateful to Dr. Steve Davis for the checking of exercises,
to Mrs. Phyllis J. Rickli for the typing of the manuscript, and to W. A. Ben-
jamin for preparation of the book. Naturally, any errors that remain are our
responsibility.
Madison, Wisconsin F.B.

September 1972 J.A.N.
Contents
Chapter 1 Introduction to First-Order Differential Equations

Te Some population and closely related growth problems 1
12 First-order equations 5
1.3 Equations with variables separable 9
1.4 Linear equations of the first order 20
ES Direction fields 28
1.6 Existence and uniqueness for first-order equations 30
Chapter 2 Introduction to Second-Order Differential Equations

Dal The mass-spring system 38
2.2 The pendulum problem 43
Le Further examples of motion problems 48
2.4 Second-order equations solvable by first-order methods 49
BS, The existence and uniqueness theorems for second and higher-order equa-
tions 60
Chapter 3 Linear Differential Equations

Spl Introduction 65
32 Linearity 68
333 Linear homogeneous equations 70
3.4 Solution of linear homogeneous differential equations of second order with
constant coefficients 83
3:5 Linear homogeneous equations of arbitrary order with constant co-
efficients 93
3.6 Reduction of order 95
307 Linear nonhomogeneous equations 99
3.8 Resonance 111
Chapter 4 Linear System of Differential Equations

4.1 Introduction 114
Vii
Vili Contents
4.2 The existence and uniqueness theorem 130

4.3 Linear homogeneous systems 134
4.4 Linear nonhomogeneous systems 147
4.5 Nonlinear system of first-order equations 156
Chapter 5 Eigenvalues, Eigenvectors, and Linear Systems with Constant Coefficients

>I The exponential of a matrix 165
32 Eigenvalues and eigenvectors of matrices 169
213) Calculation of a fundamental matrix 174
5.4 Two-dimensional linear systems 180
35) Some population problems 192
5.6 The general case 197
3,7) Solution of Example 9, Section 4.2: An electric circuit 206
Chapter 6 Series Solutions of Linear Differential Equations

6.2 Review of properties of power series 214
6.3 Second-order linear equations with analytic coefficients 217
6.4 Proof of theorem on solutions in power series 223
6.5 Singular points of linear differential equations 228
6.6 Solutions about a regular singular point: Examples 232
6.7 Solutions about a regular singular point: Theorem 239
6.8 Solutions about regular singular point: Exceptional cases 244
6.9 The Bessel equation and some properties of Bessel functions 253
6.10 Singularities at infinity 263
6.11 Irregular singular points, with an introduction to
asymptotic expansions 266
Chapter 7 Boundary-Value Problems

el Introduction 278
Tez Examples of homogeneous boundary-value problems 281
Wea: Examples of nonhomogeneous boundary-value problems — The Green’s
function 288
7.4 Self-adjoint boundary-value problems for the operator L(y)=—y" 293
aS Sturm-Liouville problems 298
7.6 Remarks on a singular boundary-value problem 303
deaf Nonhomogeneous boundary-value problems and Green’s function 306
Chapter 8 Existence Theory

8.1 Existence of solutions 314
8.2 Uniqueness of solutions 329
8.3 Continuation of solutions 333
8.4 The nonlinear simple pendulum 340
8.5 Existence theory for systems of first-order equations and higher-order
equations 346
8.6 Linear systems 349
8.7 Dependence on initial conditions 353
Contents
Chapter 9 Numerical Methods of Solution

Ort The Euler method 357
oF The modified Euler method 368
8) The Milne method 372
9.4 Stability, consistency, and convergence 381
25 Runge-kutta methods 385
9.6 Numerical methods for systems and equations of higher order 389
Chapter 10 The Laplace Transform

10.2 Basic properties of the Laplace transform 393
10.3 The inverse transform 401
10.4 Applications to linear equations with constant coefficients 406
10.5 Applications to linear systems 415
10.6 A table of Laplace transforms 422
Appendix 1 Determinants and Linear Systems 424

Appendix 2 Polynomial Equations 427
Appendix 3 Complex Numbers and Complex-Valued Functions 429
Appendix 4 The Exponential Matrix 433
Appendix 5 Generalized Eigenvectors, Invariant Subspaces, and
Canonical Forms of Matrices 436
Bibliography 449
Answers to Selected Exercises 450
Index 465
é,
—~ ree ‘neal
a, Mie a ects ahiSbeAitbeaee eT 4
1
se AL i bata Hihd gf? od .
> : ; rs >. ny air lacs vit ft ie & y
: a ’ es‘fricna id ay zs
ep i
iP @Miah yi WwebiasMeas aba) abcidtgn basket) ve
i a ey eas es lg aH
pal a ' U , rp as beri: if ’ a
or ern ucstips wit

to cay ee Te :
we . gr
La 18; = ant egVES MT
Ferrer wh PD Ore tenes birt eae ON
‘ aan id Bi asia 2! truce A, SOP
’ te Mya 1)tayta
‘luis je eane bad fran, “a

a i i anieaebitped Val Se igh - pant
ina # sue i ORC aoe ps ox

: ie ¥) rad ea on
Bietan
i tr Re ee din? Br Bee aiy, rf
gr cae eanvaiine
he
7oak
Rein — vw aly es re
= S¢ _ verde ne
‘ i
_ oe
=|
CHAPTER 1
Introduction to
First-Order
Differential Equations
1.1 SOME POPULATION AND

CLOSELY RELATED GROWTH PROBLEMS
Differential equations originated in attempts to give mathematical descrip-
tions of problems in various sciences. The study of growth of some idealized
populations leads one to relatively simple differential equations which can
be solved readily. The solution of these mathematical problems can then be
interpreted in terms of the original population. The study of problems
from other sciences, for example the motion of particles (which goes back
to Isaac Newton in the 17th century), leads to more complicated differential
equations; some of these will be studied in later chapters.
Consider a population of organisms of a single species which has A
members at time t=0. Let y(t) be the number of members of the population
at time ¢. Our goal is to determine this function. Before even stating the
physical problem we shall make the following mathematical idealization.
Evidently the value of y(t) at a particular time ¢ must be a positive
integer or zero (for physical reasons). However, we shall assume that
y(t) isa continuously varying function for t>0. This makes for a reasonable
approximation for large populations and leads to much simpler mathemat-
ical problems.
We will make some other simplifying biological assumptions and try
to use these to determine the function y(t). Thus we shall suppose that the
rate of change of the population, y'(t), (where the prime (') denotes the deriv-
ative with respect to t) depends only on the state of the population at time t and
not at all on what happened before time t. This assumption ignores the age
structure of the population and is quite artificial. Biologists have observed
that this assumption is fairly accurate for micro-organisms reproducing by
splitting (binary fission). However, for species which do not reproduce
themselves until they reach adulthood, ignoring the age structure is unsatis-
2 Introduction to First-Order Differential Equations
factory. For example, if the food supply to a population was suddenly

increased, there might be a rapid increase in the rate of growth in the popula-
tion but then there would be a time lag until the new members started re-
producing. Thus the rate of growth of the population at any point in time
would depend on what happened in the past. Nevertheless we will make
this simplifying assumption. The reader should, however, bear in mind that -
more sophisticated mathematical models are possible.
We next assume that in any time interval of length / the probability that
any member of the population gives birth to a new individual is ah, where
a>0 is a fixed constant for the whole population.
.Now consider the growth of the population from time ¢ to time t+h.
Since there are y(t) members at time ¢ the population increases by ahy(t)
members in this time interval. However, this number must be the same as
y(t+h)—y(t). This simple argument gives the equation
y(t+h)—y(t)=ahy(t) (1.1)
for any ¢>0 and any number />0. Dividing (1.1) by 4 we obtain
y(t+h)—y(t) =ay/(t).
h
Now letting 4-0 and using the definition of the derivative, we obtain the
equation
y'(t)=ay(t). (1.2)
Observe that this limiting process makes use of the assumption that y(r)
is a continuously varying and differentiable function rather than an integer-
valued function.
The mathematical problem is to find the differentiable function y(t)
satisfying (1.2) for t>0 and such that y(0)=A (this number is the initial size
of the population and is given in advance). The equation (1.2) for the un-
known function y is called a differential equation (of first order) and the con-
dition y (0)=A is called the initial condition. We shall now solve this problem.
Since y(¢) represents the size of the population, we expect that y(t)>0.
If y(t) 40, we can divide (1.2) by y(t) and obtain
We recall that d(log |y(t)|)/dt=y'(t)/y(d). Thus the above equation can

be written as
4 log|y(t)| =a
et Population and Related Growth Problems 3
and therefore integration from zero to ¢ yields
log|y(¢)| — log |y(0)| =at.

Using the initial condition y(0)=A, assuming that 440 and using ele-
mentary properties of the logarithm and exponential functions, we obtain
t
log |y(t)| —log|A|=log y(t)= ub
or, since A>0,
y(t)=Ae™. (1.3)
Exercise
1. Show by direct substitution that y(¢) given by (1.3) satisfies Eq. (1.2) on — 00 <t<oo
(even when A =0).
In solving Eq. (1.2) we have assumed that y(t) is never zero. Indeed the
solution (1.3) is never zero. The method of solution cannot be applied if
A=0, because the logarithm of zero is undefined. However, the function
y(t)=0 (which has y’(t)=0) satisfies Eq. (1.2) by inspection. Thus if A =0
the solution is y(t) =0. An interpretation of this result is that if the initial size
of the population is zero, then the population remains zero (i.e. even
mathematically you don’t get something for nothing).
In Eq. (1.2) we can interpret y’ (¢)/y(¢) as a growth rate. It is, in fact, the
time rate of increase of population per member. Thus in the model repre-
sented by Eq. (1.2) we are saying that the growth rate is constant. This
leads to the biological conclusion that the population continues to increase.
In most real-life situations such an assumption is unrealistic. One might
assume that since the food supply decreases as the population increases,
the growth rate should have the form
y'(t)
—“=a—by(t), (1.4)
y(t)
where a and + are positive constants. Thus a more realistic model leads
to the mathematical problem of solving the differential equation
y()=ay(t)—by(), yO)=A. (1.5)

Equation (1.5) is also a differential equation of first order. It cannot be
solved quite as simply as Eq. (1.2), but we will be able to solve it in Section
1.3 (Example 4).
Another population problem closely related to (1.2) arises if we assume,
as before, a constant birth rate a and, in addition, a constant death rate c.
(The assumption that the death rate is constant and independent of age is
unrealistic in most situations; it might, however, apply to an insect pop-
ulation being treated by an insecticide.) Under these assumptions the pop-
ulation y(t), where y(0)=4, satisfies the equation
y()=(4—¢) v0), (1.6).
where a and c are constants. The derivation of this equation is almost
exactly the same as that of Eq. (1.2). Equation (1.6) can be solved exactly
as was Eq. (1.2), provided that a#c. The solution is
y(t)= Aer. (1.7)
If a=c, (1.6) reduces to y’(t)=0 and the solution is y(¢)=A; note that this
result also follows from (1.7) by putting a=c. From (1.7) we see that if
a>c the population grows exponentially — just as for the solution of
Eq. (1.2). If a<c, the solution y(t) decays exponentially, which means that
the population is dying out. Models such as (1.2) and (1.6) have applications
to other areas, and we shall now discuss some of these.
Example 1. Assume that a radioactive substance decays at a rate proportional to

the mass of the substance present. Compute the half-life of the substance.
Let y(t)>0 be the mass of the substance present at time ¢. Let k >0 be the constant
of proportionality. Then the rate of change of the substance per unit mass is y’(t)/y(t)
and in this case
—“=—k or y(t)=—ky(t).
Suppose that y(0)=A. Then, exactly as in (1.2), the solution is
y(t)=Ae™.
The half-life of the substance is defined as the time required for the mass to decrease to
one-half its original value. Thus we wish to find the value of t for which
Ane
D
This is the same as finding the value of ¢for which e“'=2. Taking logarithms of both
sides yields
Ki Ooe
Thus the half-life is log 2/k. Note that both the half-life and k depend on the
particular substance.
Exercises
2. If the half-life of aradioactive substance is 30 days, how long will it take until 99%
of the substance has decayed?
LZ, First-Order Equations 5
J According to Newton's law of cooling, the rate of change of temperature of a

cooling body is proportional to the difference of temperature between the body and
its surroundings.
a) If the temperature of a body is 7(t)°, and the temperature of its surroundings
is A°, show that Tsatisfies an equation of the form 7’ (¢)= —k[T(t)—
A].
b) If a body cools from 100°C to 60°C in surroundings at 20°C in ten minutes,
how long will it take for the body to cool to 25°C?
. Water is heated to the boiling point 100°C. The water is then removed from the
heat and placed in a room which is at a constant temperature of 60°C. After three
minutes, the water temperature is 90°C.
a) Find the water temperature after six minutes.
b) When will the water temperature be 75°C?
. Ifa sum of money is invested at an interest rate of k compounded continuously then
the amount of money S(t) at time ¢ satisfies the equation S’(¢)=kS(t).
a) If the initial investment is 100 and the rate of interest is 4% (so that k =0.04),
show that the amount of money at time ¢ is S(¢)=100e°°*.
b) Show that compounding interest continuously is more favorable to the
investor than compounding interest at fixed intervals, for example annually.
Do this by computing the amount after ten years for interest compounded
continuously, interest compounded annually, and interest compounded semi-
annually. [Hint: If the interest is compounded annually, the amount after ten
years is 100(1.04)'°; if the interest is compounded k times per year, the amount
after ten years is 100(1+0.04/k)'™.]
QO Se What conclusion can you draw by comparing the results of part (b) for interest
compounded annually and semiannually?
. A mold grows at a rate which is proportional to the amount present. Initially the
quantity is two ounces; in two days the quantity has increased to five ounces.
a) Show that if y(¢) is the quantity of the mold at any time ¢, then y(t)=2(3)"”.
b) Find the quantity at the end of ten days.
1.2 FIRST-ORDER EQUATIONS
Having seen how physical problems can give rise to differential equations,
we now begin a systematic study. From the mathematical point of view,
a differential equation is an equation involving a function and some of its
derivatives from which this function is to be determined. Differential
equations involving functions of a single variable are called ordinary dif-
ferential equations, while differential equations involving functions of
several variables are called partial differential equations. We shall be con-
cerned with the former.
Suppose f is a function that is defined and continuous on some region*
* We postpone a precise definition of region to Section 1.6.

Figure 1.1
D in the plane (see Fig. 1.1). Because the independent variable in physical
problems is so often time, we label the axes ¢ and y, rather than x and y. Then
the first-order differential equation associated with the function / is
v=f(ty), (1.8)
where the prime denotes differentiation with respect to ¢. The equations
derived in Section 1.1 are all of this form. The reason for calling (1.8) a first-
order differential equation is that the highest-order derivative appearing
in (1.8) is the first derivative. To so/ve (1.8) means to find an interval J on
the ¢ axis (see Fig. 1.2) and a function ¢ such that:
i) P(t) and @'(t) exist for each ¢ in 1.7
ii) The graph of @ lies in the region D, that is, all points (¢, é(2)), for t
in J, lie in D.
iii) For each ¢ in J we have
(1.9)
Figure 1.2
+ By differentiability of a function at an end point of an interval, we mean the existence

of a one-sided derivative.
12 First-Order Equations yf
Such a function ¢ is called a solution of (1.8) on: the interval /. For

brevity, we say that y=q(¢) satisfies (1.8) on the interval J, or that @ is a
solution of (1.8) on /. This has an immediate geometric interpretation.
The function / assigns some numerical value f(t, y) to each point P in D,
with coordinates (t, y) (see Fig. 1.2). Then (1.9) merely means that the
tangent to the curve y=@(t) has slope f (t, y) at each point P of its graph.
This simple geometrical interpretation will be used to construct solutions
graphically in Section 1.5. We have already found solutions of some simple
differential equations of the form (1.8) in Section 1.1.
In the examples and exercises to follow, we shall employ the notation
{(t, ¥) |a<t<b, c<y<d} to denote the set of points in the plane whose
coordinates (ft, y) satisfy both inequalities a<t<b and c<y<d. Similarly,
{(t, vy) |a<t<b, c<y<d} will denote the set of points whose coordinates
(t, v) satisfy both inequalities a<t<b and c<y<d. Thus for example
{(t, vy) | —l<t<2,5<y< oo} denotes the infinite strip in the plane bounded
by the lines t= —1, t=2, y=5, lying above the line y=5, where the lines
t= —1and y=S are included in the set but the line t=2 1s not. You will have
no difficulty in seeing that {(t, y) | —a<t<0o, —a2<y<oo} is the entire
(t, vy) plane, while {(t, y) | —w<t<a, 0<y<o} is the upper half-plane
excluding the ¢ axis. It should now be clear that {t |a<t<b} denotes the
points on the ¢ axis satisfying the inequality a<t<hb, that is, the open
interval (a,b). For convenience we shall usually write this simply as the
interval a<t<b. We note also that {t | —c<t<oo} is the entire ¢ axis
while, strictly speaking, the inequality — 00 <t<o refers to all real numbers
rather than points on the ¢ axis.
Example 1. y’=2y. Here, f(t, y)=2y, for all t and y. We can therefore take the
region D to be the whole (t, y) plane. The function $(t)=e*' is easily seen to be a
solution of y’=2y on the interval — 1 <t<o, since (i) and (ii) are satisfied for the
choice of / as the interval —co<f<, and (iii) is satisfied since $'(t)=2e*'=2(t).
Note that the function $(t)=ce*' is also a solution for each choice of ¢ on
— 0 <t<oo. For this simple equation we have already learned in Section 1.1 how to find
this solution.
Exercise
1. Sketch these functions for c=0, +3, +1.

The differential equation y’=ay was used in Section 1.1 as a mathematical
model for radioactive decay («<0) and population growth (a> 0).
Example 2. y’=—y?. Here f(t, y)=—y*, and we can again take D to be the whole
(t, y) plane. We can check that $(¢)=1/t is a solution of the equation on either the
interval —oo <t<0, or on the interval 0<t<oo, but not on an interval such as
—2<t<2. To see this, draw a graph of ¢ on —«2<t<0 and 0<t<o, and then
verify that the conditions (i), (ii), and (iii) are satisfied on each of these intervals. We
shall learn in Section 1.3 how to obtain this solution.
Exercises
2. Show that the functions #(t)=1/(¢—c) are solutions of the same equation y’= —y?
for each choice of the constant c on a suitably chosen interval. Draw graphs of
these solutions for c=0, +1, +2.
3. Consider the differential equation y’=2/(t?—1) with f(t, y)=2/(t* —1) defined on
each of the domains D, = {(¢, y) | —o<t<—1, |y|<oo}, D,={(t, y) | —l<r<l,
ly|< oo}, and D,={(t, y) | 1<t<, |y|<oo}. Verify that
t—1
(t)=log t+1
is a solution of this equation on each of the intervals —coo<t<—1, —l<tr<l,

and 1<t<oo. (A graph of ¢ will show why y= @/(t) is not a solution on an
interval such as —2<t<2.)
4. Consider the differential equation y’=(y?—1)/2. Verify that y=(t)=(1+ce’)/
(1 —ce’) is a solution of this equation on an appropriate interval /, for any choice
of the constant c. [Hint: Try this first for specific values of c such as c=0, c=1,
c=—1.] Draw graphs of ¢ for each choice of c.
We have seen in Section 1.1 how physical phenomena may give rise to
mathematical models leading to differential equations. In general, each
differential equation will have many solutions, as is suggested by Example 1.
Thus it is necessary to impose additional restrictions, called initial conditions,
on the equation in order to obtain a unique answer. For first-order
equations of the form y’=f (t, y) a natural additional restriction is to ask
that a solution ¢ satisfy a condition such as ¢ (to)=yo. The differential equa-
tion y’=f (t, y) together with the initial condition y(tg)=yo is called an
initial-value problem. Geometrically, this says that we are only interested
in solutions of the equation whose graphs pass through the point (fo, yo).
Physically, this corresponds to measuring the state of the system described
by the differential equation at a time ¢) and predicting the future behavior
of the system.
Example 3. We have seen in Example | that each function
y=ce!
is a solution of the equation y’=2y. Suppose we wish to find a solution for which
#(1)=5. Imposing this condition on the family of solutions ce’, we are forced to require
that 5=ce? so that c=Se 7. Therefore #(t)= te") is a solution, passing through the
point (1, 5).
Example 4. Among the solutions
Sulsh.cet
~ 1=cet
1.3 Equations with Variables Separable v)
of the differential equation y’=}(y* — 1), let us find a solution @ which obeys the initial
condition $(1)=5. If there is such a solution, it must satisfy
5 1+ oF
l—ce
from which we get 5S—Sce=1+ce, and therefore c=4e1. Thus
1+2e"!
(t) 1—3e r=
is the desired solution.
Exercises
5. Find the solutions of the equation y’=2y (Example 3 above) that pass through the
point (a) (1, 0), (b) (0, 0), (c) (—1, 2).
6. Find solutions ¢ of the equation y’=4(y?—1) (Example 4 above) that obey the
initial condition (a) #(0)=1, (b) 6(0)=0, (c) (2)=0.
* 7. Find an equation satisfied by the coordinates of all points (¢9, yo) with the
property that a solution of y’=/(t, y) through (foyo) has a maximum or minimum
at (to, ¥o). How would you determine whether a maximum or a minimum occurs
at (to, Vo)?
* 8. Apply Exercise 7 to determine the maxima and minima of solutions of

potty
Lae
We shall now study some methods of finding solutions of certain types

of first-order differential equations.
1.3 EQUATIONS WITH VARIABLES SEPARABLE
We shall say that a differential equation of first order
ya) (Ey)
has variables separable if the function f can be written in the form
f(t, y)=9(t) hy). (1.10)

The reason for the name separable is that the differential equation can then
be written as
y'/h(y)=9(2),
provided, of course, that h(y)40. For example, y’=y* and y’=y/(1+?t7)
* An asterisk preceding an exercise indicates that this exercise is rather more difficult
than the others.
are both of this type, while y’=sin ¢—2ty is not. The examples discussed in
Section 1.1 are also of this type. Before explaining the technique in the
general case of variables separable, let us work out one more example.
Example 1. Solve the initial-value problem
y=y?, — v(to)=Vo (1.11)

where fy and yo are arbitrary.
Suppose that the problem has a solution @. According to the definition of solution,
this function @ must satisfy the equation
¢'(N=LO(O]?
on some interval / containing tg. Assuming that p(t)40 for all ¢ in the interval J and in
particular that y)#0, we would have to have
oO)_
[o(.)]?
If we integrate this equation from fg to ¢, we obtain
$'(s) |
| or
to
ds—a\mledse
to
1.12)
where fp 1s given in the initial condition and ¢ is arbitrary. Note that we have used s as
the variable of integration in place of ¢t. We did this in order to avoid confusion
with the upper limit ¢ of integration. To evaluate the integral on the left-hand side of
(1.12), we make the change of variable u= ¢(s). Then, by the change-of-variable theorem
for definite integrals, we obtain
(6)
$(t)
du
|gare (to)
we
This integral is easily evaluated; its value is

$(t)
du 1 > (t) 1 1
u u ¢ (to) 3 bl” blto)

$ (to)
The integral on the right-hand side of (1.12) is of course [{, 1 ds=1—1o. Therefore the
solution @ must satisfy the equation
1 1
#0) Ol)
Since the solution @ is to pass through the point (to, yo), we must have #(to)=yo.
Hence
1.3 Equations with Variables Separable 11
or, solving for (rt),

Yo
p(t) (ese (1.13)
Recall that we have assumed yp) #0. We have now shown that if @ is a solution of Eq.
(1.11).which is never zero, then ¢ has the form (1.13). It remains to be shown that the func-
tion ¢ defined by (1.13) actually satisfies the differential equation (1.11) on some
interval.
The function @ defined by (1.13) is well-behaved for values of ¢ for which the
denominator is different from zero. More precisely, p(t) is differentiable provided
1 —}'9(t—to) #0, and its derivative is given by
r Yo V6
Ws T= yt Ore” ei =Yo(t—to)]?
But
Yo
[o(s)]°=
[1—yo(t—to)]”’
and thus ¢'(t)=[@(c)]? for all ¢ for which 1—yo(t—f9)40, that is, provided that
1Ayo(t—to). Since yo #0, this means that 1/y)>#t—fo, or that tAty+1/Yo. If yo>O,
then 1/y)>>0 and thus the function @ is differentiable on the two intervals
—0 <t<tgt1/yo and tp+1/v9<t<0oo. But remember that the problem is to find a
function ¢ which is a solution of the given differential equation on an interval which
contains to. If y)>0, that interval is —o0<t<tg+1/yo. Therefore, for y)>0, the
function ¢ given by (1.13) is a solution on the interval — co <f<f)+1/yo as shown in
Fig. 1.3. Direct substitution in (1.13) shows that ¢ also satisfies the initial condition
(to)=Yo. Thus we have solved the initial value problem (1.11) if yo >0.
y
(to, Yo) f
Figure 1.3
Exercise
1. Show that if yo <0, the function ¢ given by (1.13) is a solution of the initial-value
problem (1.11) on the interval to + 1/vo <<, and sketch the graph comparable
to Fig. 1.3.
You might well ask if there are any other solutions of the initial-value
problem (1.11). The answer is no, because, having assumed the existence
of a solution on some interval, we found that there was only one possible
candidate for a solution, namely the one given by (1.13).
In the discussion above, we assumed from the start that y)o40 and
in fact that @(¢)40 for all choices of ¢ in the interval /. It is therefore not .
surprising that the solution (1.13) never assumes the value zero for any
choice of ¢. Does this mean that there is no solution of (1.11) if yp=0?
The answer is simple; we have overlooked the fact that ¢(t)=0 is also a
solution of the differential equation y’=y*, and that this indeed does
satisfy the initial condition (tf ))=0. Including the zero solution with the
solutions given by (1.13), we have shown by actual construction that for
every initial point (fo, Yo) there is a solution of the differential equation
through this point. What has happened in this example is quite common;
a “‘general’” method may not succeed in producing all of the solutions of a
differential equation and, for certain exceptional initial conditions, other
methods may have to be used. Note that the function @ as given by (1.13)
is well defined for y)=0. In fact, it happens that the zero solution is ob-
tainable from (1.13) by taking y) =0. For a further discussion of Example |
from a different point of view, we refer you to Example 3, Section 1.6.
Exercise
2. The graphs of the functions ¢ defined by (1.13) above are all hyperbolas. Draw the
solution curves y=(t) passing through the points (0, 4), (0, 1), (0, 2), (0, —1),
(0, —2). Observe how these curves compare with the “special” solution ¢(t)=0.
Let us apply the same technique to the general case of variables separable.
Suppose we have
y'=g(t) h(y)=f(t y) (1.14)
where g is continuous on some interval a<t<b and A is continuous on some
interval c<y<d. The function f(t, y)=g(t)h(y) is continuous on the
7
Figure 1.4
rectangle D=((t, y) |a<t<b, c<y<d} (see Fig. 1.4). Let @ be a solution

of Eq. (1.14) that exists on some interval J contained in (a, b). We must
therefore have ¢’(t)=g(t) h(@ (t)) for every t on J. If we exclude for the mo-
ment the possibility that 4(¢(t))=0 on J, we can write
and therefore
|Rats [oo
t
to
for every ¢ on J. Putting u=¢(s) and using the initial condition ¢ (to)=yo,
we obtain, proceeding as in Example 1,
$ (t) t
|fam |204 ((515)
This equation defines the solution ¢ implicitly. As Example | above and Ex-
amples 2, 3, and 4 below show, it may be possible to solve the resulting
equation for @(t), thus obtaining an explicit solution of (1.14).
Exercise
*3. a) Look up the statement of the implicit function theorem for solving an equation
of the form F(t, y)=0 for y terms of ¢ near a given point (to, Yo) for which
F (to, Vo) =0.
b) Apply the implicit function theorem to Eq. (1.15) and determine conditions which
permit you to solve (1.15) for #(¢) near the point (to, Vo).
What happens if there are values of t where h(¢(t))=0? Let us consider
first the case when h(y,)=0. Consider the function w(t)=yo. Then
w'(t)=0 and A(y(t))=A(y.)=0 and therefore W(t)=yo is a solution of
(1.14). In addition, observe that the left-hand side of Eq. (1.15) may not make
sense, because it is an improper integral which may not converge. If it does
not converge, w(t)=Vo is the solution through the point (to, yo). If the in-
tegral does converge there are at least two solutions through the point
(to, Yo), namely w(t)=yo and the solution #(¢) given implicitly by (1.15).
Exercise 32 below shows that one may have infinitely many solutions in
this case. If hvanishes for some other value of y, say y=y, #Vo, Eq. (1.15)
still defines a solution ¢ implicitly on some interval about ¢= fp. If this solu-
tion @ never reaches the value y,, there is no difficulty. If the solution
“reaches” the value y,, the convergence of the integral on the left-hand side
of (1.15) must be investigated again as above. As before, the constant func-

tion w(t)=y, is another solution.
y=ty, — y(0)=1.
In the notation of (1.14), g(¢)=¢, A(v)=y?, and A(yo)=A(1)=140. Suppose that
there is a solution @. Then on some interval containing the point fg =0, ¢’(t)=r[(t)]°.
If b(t)40 we have ¢'(t)/[(t)]* =2, and therefore, using to =0,
Since ((0)=y,=1, letting v= (s) as in Example 1, we obtain

t
oO
$ (t)
fier g(t) 1
|ay |uen ea:

=— +—,
1 2{p(t)}? 2
10)
Also | s ds=t?/2. Therefore $(1) is defined implicitly by the equation (which corre-
sponds to (1.15) above)
1
alee
{P(t}?
In this example this gives {f(r)}*=(1—¢7)~! and, finally, since ¢(0)>0, we take the
positive square root and obtain $(t)=(1—1*) "1/7, (—1<t<1). We readily verify that
~(0)=1 and that $(t)=(1—1*)~'” satisfies the differential equation y’=ry? on
—1<t<1; note that ¢(t) is never zero for —1<t<1.
Exercises
4. In the solution in the example above, we found that [o(e)]}?=(1 =p7)5%. Why
could ¢ not be given by $(¢)= —(1—1*) 1/7?
5. Can you suggest an initial point (to, yo) for which #(t)= —(1—1?)~1/? would have
been the correct choice?
. What is the rectangle D in this example?
. Why is the solution obtained only valid for —1<t<1?
a
on. Try this method for the initial point (a, 0) and the same equation.
9. Try this method for the initial point (1, 2) and the same equation.
Example 3. Consider again the differential equation y’=ry*, but this time with an
arbitrary initial point (fo, Vo), ¥o >0 (why do we require y)>0?); then, proceeding as
in Example 2, we obtain
t o (t) t
d'(s 1
|Speen
[o(s)3 S =! | Sai
va ae |i xs ds
to yo to
fe
ies! Equations with Variables Separable 15
which yields
l ee
[o@]}’ vo
If we let 1/yo+1t5=c? (c>0) and solve for (¢), this becomes (t)=(c?—12)7!/2,
(\¢|<c). It is easily verified that this is a solution of the differential equation y’=ty?
on the interval —c<t<c for every choice of the constant c>0, that is, for every initial
point (fo, Vo) with yo >0. If yp =0, the corresponding solution is #(t)=0, which is not
obtainable from the expression $(t)=(c?—1t*)~‘/? for any choice of c.
Exercises
10. Sketch the solutions found in the above example corresponding to c=1, c=2,
c=5. Be sure to indicate the appropriate intervals.
11. Solve the above example for an arbitrary initial point (¢9, Yo) with yo <0.
y=ay—by*, y(0)=A, (1.16)

which was formulated as Eq. (1.5), Section 1.1, to describe a population problem. (The
normal biological situation would have a>0, b>0, A>0, and a>bA.)
Let o(¢)40 be a solution on some interval / containing t=0. Then on this interval
d (t)=ap(t)—bo* (1). (1.17)

If ap(t)—bd? (t)=(t) [a—bp(t)] 40, in other words if d(t)#0 and (t)4a/b, then
we may divide (1.17) by af(t)—b?(t) and we obtain
o'(t)
bl [a—b6(0)]
=1.
Integration from 0 to r yields
rf 8) iis
ere lw
Letting vu=¢(s) as in Example | and using the initial condition #(0)=4,
(0)
du
=i. (1.18)
u(a—bu)
A
To evaluate the integral we use partial fractions

1 C D
a
u(a—bu) u a—bu
where C and D are constants to be determined from the above identity. Since
G D _aC—bCu+ Du
ua—bu —- u(a—bu)
we must require
aC +u(D—bC)=1
for all u. Therefore aC=1 and D—bC=0, or C=1/a, D=bC=B/a.
$ (t) $ (t) $(t)

dus eae | du
u(a—bu) a} u aJ a—bu
A A A
1, JO) 1), |abo

Tait pAulast |aShs
p(t) a—bA
=-—lo
A a—b$(t)
Substituting this result in (1.18) yields
—bA
log oe (0) Sh.
A a—bdi(t)
Then the following sequence of easy steps
G—bA oO |
A a—bd(t) — ’
6()=[a—b(]]
—— e*,
w) t 1 ate A Co aA at
Nas eran rey eee

a—bA+bA e™ aA
t = a
(0 ( a—bA a—bA
yields
aAe™ aA
o(t) (1.19)
‘G=bALbAe’ (G=bA)\e “Fb4
The reader can verify by substitution that (1.19) is a solution of(1.16). Note that if b=0
this solution reduces to #(t)= Ae” which agrees with the result obtained for the simpler
model (1.2) in Section 1.1. If 640 the solution behaves quite differently. In particular,
lim,.... &(¢)=a/b. To see this from (1.19), observe that (a—bA)e~ tends to zero as t> 00.
Exercise
12. Sketch the graph of the solution (1.19) assuming a>0, b>0, A>0, a> Ab. (If you
wish, take a=8, b=2, A=2.) Also sketch the solution if a>0, A>0, b=0.
In the simple model discussed in Section 1.1, the solution continues

to increase, which means the population continues to grow indefinitely. In
the present model the population increases from the initial value A to the
limiting value a/b as t>0o. This corresponds more closely to observed be-
havior for many populations.
Exercises
In Exercises 13-27, find all solutions through the given initial point. Decide, if you
can, whether the solution obtained is unique. Be sure to specify the interval in which
the solution is valid.
13. y=t2y (5, 1)

14. y'=t’y (1, 0)
15. y’=t/y (to, Yo); Yo > 9
16. y’=t/y (to, Yo); Yo <O
y-y?
17. veo (to, Yo), to #0, —1;
YoXO, 1
2
is y= y marty
21
_ t+t7 eo)
Ca
Hoy 1
ms 1+e' })
20. y' =t?
y* —4¢? (5, —2)
Dvr
y 4" (te. Yo)
y=i+y (—x/4,0)
23. v=Ty cost) (4;0)
24. y'=tant tany (0, 7/6)
Biy=ty* (0, —2)
Pier
Ak ahi, aed (eile
"log2
27. y'=2t(1—y7)'? (0, 1)
28. By solving the differential equation y’=1+-y’, show that no solution through the
origin exists for —0o0<t<oo.
29. Given a family of curves f(x, y)=c in the (x, y) plane, a family of curves g(x, y)=c
is said to form a set of orthogonal trajectories of the first family if every intersection
of a curve in the first family with a curve in the second family is at right angles.
Find the orthogonal trajectories of each of the following families of curves:
a) x?+y2=c b) y=ce*
C) 7 =4ex d) 2x? +? =c*
[Hint: (for part a)): Show that, for every curve in the given family, dy/dx= —x/y.
Therefore for every orthogonal trajectory dy/dx=y/x; solve this last differential
equation.|
Separable first-order differential equations also arise in certain prob-
lems involving chemical processes. Typically in such problems one has a
tank of a fixed capacity filled with a solution (say brine) containing a
certain amount of a substance (say salt) thoroughly mixed. A solution
(brine) containing a fixed amount of substance (salt) per unit volume enters
the tank at a fixed rate while at the same time the stirred mixture leaves
at a fixed rate (which may differ from the entering rate). If one denotes
by y(t) the amount of substance (salt) at time ¢, then y(t) must satisfy
the ‘‘equation”’:
y’ (¢)=(rate at which substance is being added)
—(rate at which substance is being removed).
Writing this out mathematically leads in certain special cases to a first-
order equation which has variables separable. We now illustrate this type
of problem. For more complicated problems, refer to Exercises 31 and 32,
Section 1.4.
Exercises
30. A tank contains 100 liters of water and 10 kilograms of salt, thoroughly mixed.
Pure water is added at the rate of 5 liters per minute, and the mixture is poured off
at the same rate. How much salt is left in the tank after one hour? Assume
complete and instantaneous mixing.
The following problems illustrate several other important applications
of first-order differential equations.
Exercises
*31. When lying on its side on a smooth horizontal plane, a helical spring is L inches
long (see Fig. 1.5). The spring is a “‘linear spring”’ in the sense that the amount it is
compressed by an axial force is proportional to the force applied. The constant of
proportionality, which is called the “‘spring constant,” is K (pounds per inch), and
the total weight of the spring is m pounds.
L | This much of the spring

i weighs w(x) pounds
Figure 1.5
Equations with Variables Separable 19
When the spring is stood on end, it shortens owing to its own weight. If x
represents the distance (in inches) measured positively downward from the upper
end of the up-ended spring, and if w(x) is the weight of that portion of the spring
above a section x units from the top, then the function w(x) is a solution of the
initial-value problem
dw _ 1E
dx 1—w/KL’
Solve this initial-value problem for the function w(x). Let L’ represent the areal
length of the up-ended spring. Noting that w(L')=m, find L’. How did the spring
shorten when-up-ended? If p(x) is the weight per unit length of the up-ended
spring, then p(x)=dw/dx. For what value of x does p(x) take its maximum? What
is the maximum value of p(x)? What is the ratio of this maximum value to the
average value of p(x)?
32, Find the solution ofthe equation y’ = 3y7'? passing through (¢9, yo). Give a complete
discussion of the various cases. Note that y)>4#0 and yy=0 must be treated
separately and that @(t)=0 is again a solution. However, we do not have uniqueness
of solutions if yg =0. For example, ¢(t)=0 is a solution through (0, 0) and so is
#.0=S 0) —co<t<c
CSU<AOO)
for every c>0. (Verify!) Sketch the graphs of these solutions.

3: Carry out a similar investigation for the equation y’=2y!/?.
*34. Consider the differential equation
v=f(y),
where f(y) is a continuous positive function for y>0 and /(0)=0. By separating
variables, show that if the improper integral {6° [1//(u)]du diverges, then there
is a unique solution through the point (0, yo) which is strictly positive and tends to
zero as t — 00. Show that if |§ [1//(u)]du converges for some a>0, then there is
an infinite number of solutions through the point (0, 0). Give some examples of
functions f(u) and sketch the corresponding solution curves. [Hint: Use the
method of separation of variables and the inverse function theorem.]
“Bey Consider the solution ¢ ofthe differential equation y’=1* + y* passing through the
point (0, 1). Show that the curve y= p(t) has a vertical asymptote for a value t=fo
with 2/4<1)<1. [Hint: To see that fo <1, observe that ¢ satisfies the differential
inequality $’(t)>[@(d)]*. Integrate this inequality, using ¢(0)=1 to obtain
(t)>1/(1—2). To see that t)>7/4, observe that ¢ also satisfies the differential
inequality $'(t)<1+[@(c)]* for 0<r<1 and proceed in a similar fashion.]
SO: Find all continuous (not necessarily differentiable) functions
f(t) such that
tror=| 10 ds, (t>0).

1.4 LINEAR EQUATIONS OF THE FIRST ORDER
We shall now consider another type of first-order equation for which there
is a general procedure for finding all solutions. The equation y’=f (¢, y)
will be called linear if f (¢, y) is a linear function of y. This means that the
equation can be written in the form
y'+p(t) y=q(). (1.20)

Examples are the simplest population model in Section 1.1, and equations
such as y’+ycost=sint, y'+3ty=e', or (1+t) y'+ty=cost. More
generally, a linear equation is written in the form
ao(t) y’'+a, (t) y=b(t). (E21)
If ag (t) #0, this can be reduced to (1.20) by division by a(t). If dp (to)
=0
for some fo, there are difficulties, which will be discussed in Chapter 6.
Here we require that a)(t)40 for all ¢ under consideration.
We shall assume that the functions p and q in (1.20) are continuous on
some interval /; then the function f(t, y)= —p(t) y+4q(d) is continuous in the
vertical strip S= {(t, y) | tin I, —c0o<y<oo} shown in Fig. 1.6.
Df
Figure 1.6
We first work with some examples, which will suggest the approach to the
general problem.
Example 1. y’+2ty=0 with initial point (0, yo).
Here the strip S of Fig. 1.6 is the whole (¢, y) plane. This equation has variables
separable, and we can use the method of the previous section to obtain the solution
do(t)=yoe", for —co<t<o, if yo #0. If yp=0, we obtain the solution ¢(1)=0
which, incidentally, is also included in the formula #9 (t)=yoe".
Another way to express the fact that the solution ¢o(t) is ype" is to write
ebo(t)=Yo-
1.4 Linear Equations of the First Order 21
If we differentiate this relation we obtain

d 2 ; 12 12 12 r
a Let bo(t)] =$(t) e& +.2te bo (th=e [Po (t)+ 2tbo(t)].
Since e #0, this means that $o is a solution of y’+2ty=0. What we learned from this
is that if @p is a solution and if we multiply both sides of the equation 6 (¢)+ 2/9 (t)=0
by e® we obtain on the left side the derivative of e'’(t). Since this derivative is zero, it
follows that e’9(t) is a constant; in fact, the initial condition $o(0)=Yo tells us that
this constant has the value yo. This procedure of multiplying the differential equation
by a suitable function (in this case e'’) gives us another way of finding the solution do.
While these remarks may appear to be a waste of time in this example, they will be
useful for the solution of the next example.
Example 2. y’+2ry=sin/, initial point (0, yo).

Now the variables are not separable. If (¢) is a solution, so that p’(t)+ 2r(t)=sin t,
we can make the left side of the equation the derivative of e (t) by multiplying by
e’, as we have done in Example |. If we multiply the whole equation through by e”,
we obtain d/dt[e'¢(r)]=e" sin ¢. But now we can integrate, obtaining
t t
|Fleroioas=e"is] =e°6()-4(0)= | sins a

0 10)
or
t
o1)=r0+ |e® sins ds

0
and this gives #(t)=yoe +e" [he sin sds for —co<t<oo as a candidate for
a solution. The integral {j e° sins ds cannot be evaluated in terms of elementary
functions, but nevertheless it defines a differentiable function, whose value for t=0
is zero and whose derivative is e’ sint. We now check our result by substituting this
expression for ¢ in the differential equation y’+2ty=sin¢ and we obtain, using the
product rule and the fundamental theorem of calculus,
t
@'(t)= —2te" Yot = e-f Je sins ds

2 d 2 a is
0
t t
d ©
= Ptentyyei” FF|e* sins ds—2te™" |e* sins ds
0 0
t
= ~216-Fyot sine —21e"* | sins ds= —2td(t)+sint.

0
This shows that ¢ is actually a solution of the differential equation. Observe that ¢ is
the sum of two terms, one of which is the solution ¢p in Example 1.
Now let us try to use the method developed in these examples to find
a solution # of Eq. (1.20), satisfying the initial condition $(fo)=Yo for
some fp in J.
We begin by considering the simpler equation
y'+p(t) y=0 (1:22)
which has variables separable and is easily solved. If we define
P= |ns)
the solution ¢y passing through (to, yo) is given by $o(t)=yoe =P *”. We can
write this as e?d,(t)=yo and differentiate, using P’(t)=p(t), to obtain
© [cP Golt)] =e" Po) +P'(1)ef bolt)=e? [45 (0+ (0) bo(t)]=0.
This is just what we need to find the solution ¢ of the full Eq. (1.20)
passing through (to, yo). For, if we multiply both sides of the equation
f' (t)+p(t) d(t)=q(t) by e?®, we obtain d/dt[e’®
¢(t)]=e? q(t). Now
we can integrate from ¢ to t, obtaining
t
eP? (t)—e?©) #(00)= |e? ) g(s) ds

to
or, using P(t o)=0 and $(to)=Vo, we obtain as a candidate for a solution
t
P(th=yoe O+ eel e? ©) q(s) ds (1.23)

to
for fo and ¢ on the interval 7. We must now verify that the function defined
by (1.23) is a solution of (1.20).
Exercise
1. Verify that (1.23) does satisfy both the differential equation for all ¢ in J and the
initial condition. [Hint: Proceed exactly as in the verification that ¢ is a solution, as
in Example 2. |
Observe that the expression for ¢ is the sum of two terms, the solution
of (1.22) through (fo, yo) and the solution of (1.20) through (fo, 0). The
expression is defined and remains finite so long as p(t) and q(t) are con-
tinuous. (This follows from theorems in calculus — which ones?)
‘Combining the result of Exercise 1 with the fact that we found only one
possible candidate for a solution, namely the one given by Eq. (1.23),
we may state the following result.
Theorem 1. If p(t) and q(t) are continuous functions on some interval I and
if to is any point in I, then for each yo the initial-value problem
y+P(y=a(t) — y(to)= Yo
has one and only one solution # given by Eq. (1.23). This solution exists for all
t on the interval I.
Figure 1.7
The situation described in Theorem | may be pictured geometrically

as in Fig. 1.7. Note that the graph of the solution reaches the boundary
of the strip S because, as pointed out above, the functions p, g are assumed
to be continuous on the interval J and the solution ¢ given by Eq. (1.23)
is well defined and satisfies Eq. (1.20) on the entire interval /.
While (1.23) is a formula for the solution, it is easier to use the method of
solution which gave (1.23) than to memorize the formula. This formula is
more complicated than it appears, since
Exercises
In Exercises 2-18, find the solution through the given initial point and determine the
interval on which the solution is valid.
2. y +2y=e' (0, 1)
3. y +py=q (to, Yo), where p and g are constants
45 yy aye 2 i085)
5. y +y/t=1 (1, —2)
6. ty’ +3y=0? (—2, 1)
[Hint: First divide by ¢.]
1 ty
+2a) (to,
Yo) With f9#0
8. y'+y cotf=cost (2/2, 0)
9. 17y'4+ty=2t74+1 — (to, Yo) with tp #0
10. ty’ +y=e' (1, 1)
Hl. ty +2y=0?-1t4+1 (1, 1/2)
Lay =) Oy (0, 0)
13. y'=—2t(y—t?) (0, 0)
14, y'+3y=e* (0, 1)
15. (t+1)(0'+y)=e' (0, a)
162357 y= 2te" (0, a)
17. y’'—y tant=1 (0, a)
18. v(t)=Jo v(s) ds+¢+a. [Hint: Differentiate the equation. ]
19. Discuss the behavior of all solutions of the differential equation
y=Ay
as t> +00 for each of the cases 1>0, A4=0,1<0.
20. Discuss the behavior of all solutions as t> +00 of each of the following differ-
ential equations.
a) y=—2y+e°
b) y'=—2y+e'
c) w=—2yt+1
d) y’=—2y+1/(1+0’). [Hint: Use PH6pital’s rule to evaluate the limit.]
*e) y’=—2y+f (t), where f(t) is a continuous function with lim,.,,,
f(¢)=0.
21. Find the solution of the differential equation y’+ a(t) y=b(t)y* (Bernoulli equa-
tion), through (fo, Vo), where k is a constant, k#1, 0. [Hint: Make the change of
dependent variable z=y'~*.] How could you solve the equation if k=1 or k=0?
22. Find the solution of y’—y=t) passing through the point (1, —1).
23. Find the solution of y’+ y/t=y*/t? passing through the point (—1, 4).
24. Find the solution of the differential equation ty’=2r7y+y logy through the
point (1, 1). [Hint: Make the change of variable log y=v. ]
*25. a) Show (by constructing it) that the equation y’+y=/(r) has a unique solution
bounded for — co <t<oo, where |f'(t)|<M and fis continuous for — 0 <t<oo.
[Hint: Consider the solution ¢4(t) with ¢4(—A)=0, where A>0, and let
A-+0, b(t)=limysybalt).]
b) Given that the function f(t) is periodic with period 2n (that is, f(¢+22)=/f (t)
for —00 <f<oo), show that the solution obtained in part (a) is periodic. [Hint:
Show by a suitable change of the variable of integration that the solution
obtained in part (a) satisfies p(¢+2x)=¢(t) for every ¢.]
*26. Given the differential equation
y +a(t) y=f (0,
where a(t) and f(t) are continuous, where a(t)>c>0, and where lim,.,, f(t)=0,
show that every solution tends to zero as too.
*27. Leta bea positive constant and let lim,_,9 ,f()=5. Show that the equation
ty +ay= f(t)
has a unique solution which is bounded as t>0+ and find the limit of this solution
as t>0+.
Application to Simple Electric Circuits

The simplest electric circuit consists of a generator or a battery acting as
a source of energy called e/ectromotive force e, connected by wire in series
with a resistor r and a switch s. (In practice it is impossible to build a pure
resistance circuit, but we ignore this difficulty.) This circuit is sketched
schematically in Fig. 1.8. When the switch s is closed, the resistor uses energy
s Ife
Figure 1.8
and causes a current i to flow through the circuit. Ohm’s law states that the
current / is proportional to the electromotive force e. This is expressed by the
equation
e=ir (1.24)
where the constant of proportionality r is called the resistance. The con-
sistent units in which these quantities are measured are: e in vo/ts, iin amperes,
and r in ohms. If the quantities e and r are known, the current 7 is immediately
determined from (1.24) (this is not a differential equation). It is assumed that
the resistance of the wire is negligible.
A more complicated circuit places an inductor L in series with a resistor
and an electromotive source e. An inductor causes a change in current. It has
an inertia effect comparable to a mass in mechanics. Our circuit may be
pictured schematically as in Fig. 1.9.
Figure 1.9
When the switch s is closed at t=0, a current i flows in the circuit. The
problem is to determine the current i(¢) as a function of time. To do this we
will set up a differential equation for the current which is derived from the
following empirical laws.
1. Ohm’s law: The voltage drop e, across a resistor of resistance r is
=f.
[This has already been stated in (1.24).]

2. The voltage drop e, across an inductor is proportional to the time
derivative of the current, that is,
er = ie (t) 5
where the constant of proportionality L is called the inductance (measured in

units called henrys).
3. Kirchhoff’s law: The algebraic sum of voltage drops around a circuit (or
a Closed loop in a circuit) is zero. Equivalently, the voltage e supplied to the
circuit must equal the sum of the voltage drops.
Using these laws for the circuit in Fig. 1.9, we obtain the following
equation:
Li’ (t)+ri(t)=e. (1.25)
Assuming that no current flows in the circuit at time t=0 (when the switch
is closed), the initial condition is i(0)=0. If the voltage source e is a battery,
e is a constant; if the voltage source is an alternating current generator,
then ¢ is usually assumed to be of the form
e=ey SINat,
where 2n/q is the frequency and eg is the voltage. Eq. (1.25) is a linear first-
order equation of the form (1.21).
Exercises
28. Solve (1.25) with the initial condition i(0)=0, assuming that L, r, e are positive
constants, and sketch the graph.
29: Show that the solution of Eq. (1.25) with the initial condition i(0)=0, under the
assumption that L, r are positive constants and e=ey sin wt, is
i()= eowL ontriLyt £0

sin(wt —«)
P+tol (r?7 + w*P)'/?
where « is the angle defined by the relation
aol
tan «=—.
r
Sketch the graph ofthis solution and compare with the result of Exercise 28. Note
that the current is the sum of two terms: one, called the transient solution, approach-
es zero, and the other, called the steady-state solution, oscillates.
30. Assume that the resistance r is a linear function of time, r=kt, and that L and e
are positive constants. Show that the solution of Eq. (1.25) with the initial con-
dition i(0)=0 is
t
: C _y2
iGy== ee" esks2/2L aids,
E
0
Sketch the graph of this solution and show that lim,.,, i(¢)=0. Is this result phys-
ically reasonable? Given that e= 100 volts, k=5 ohms per second, and the current
reaches its maximum value when t=2 seconds, determine the maximum current.
An Application to Chemical Mixtures

Linear differential equations arise naturally in chemical mixtures. This is
illustrated by the following exercises.
Exercises
SE. A tank contains 10 liters of water to which is added a salt solution containing 0.3
kilograms of salt per liter. This salt solution is poured in at the rate of 2 liters per
minute and is thoroughly mixed, and then the mixture is drained off at the same rate
(2 liters/min). How much salt is in the tank after five minutes?
[Hint: Let (t) be the
amount ofsalt in the tank at the end of tminutes. Then between ¢ and (t+ /) minutes,
approximately 2 x 0.3 x h=0.6h kilograms of salt enter the tank, and approximately
2x p(t) x zy xh=0.2hG(t) kilograms of salt leave the tank. Thus $(¢+h)—¢(o)
~0.6h—0.2h¢(t); dividing by / and letting h-0, we have #’(t)+0.2(t)=0.6. The
initial condition is ¢(0)=0. }
32. A tank has 10 gal of brine containing 2 Ib of dissolved salt. Brine with 1.5 Ib of salt
per gallon enters the tank at the rate of 3 gal/min. The mixture is kept well stirred
and leaves the tank at,the rate of 4 gal/min.
a) Find the amount of salt in the tank at any time /.

b) Find the concentration of salt (pounds of salt per gallon) at the end of 10 min.
c) Sketch the graphs of the amount of salt and of the concentration of salt against
time and determine the absolute maximum of each quantity.
1.5 DIRECTION FIELDS
The geometric interpretation given in Section 1.2 is the basis for a simple
and effective procedure for finding out something about the nature of the
solutions of the differential equation
y'=f(t, y),
where the function f is defined in a region D in the (¢, y) plane. We can use
a special device to display the values of f in such a fashion that it will be
possible to sketch the solutions directly. In Chapter 9 we shall see that
this crude process is the key to a useful numerical procedure, adaptable to
digital computers, for finding approximations to solutions of a differential
equation. The interested reader should have no difficulty in understanding
Section 9.1 after reading this section.
The function f defines a direction field or tangent field in the region
D as follows. At each point P in D, with coordinates (t, y), we evaluate
f(t, y) and then draw through P a short line segment whose slope is f(t, y).
To do this systematically, it is helpful to construct the curves in the (t, y)
plane on which the function f(t, y) is constant. (These are called the level
curves of the function f:) On each such curve the previously mentioned
line segment has constant slope. Thus choose a value A and construct
the curve f(t, y)=k (such a curve may have more than one branch) and
along this curve draw line segments of slope k. Do this for several values of k.
Figure 1.10
ES Direction Fields 29
For example, in Fig. 1.10, we have constructed part of the direction

field defined by the function / (t, y)=—y. Hence the curves f(t, y)=k are
the horizontal lines — =k. Note that the slope of the direction field is zero
along the ¢ axis; it is —1 along the line y=1; it is +1 along the line y= —1,
and similarly along other horizontal lines. As another illustration, we give
in Fig. I.11 part of the direction field corresponding to the function
J (t,y)=y—t. Note that the function f is constant on each straight line
y=t+k (k constant).
Figure 1.11
How will such a diagram help us find a solution of y’=f(t, y)? The
answer is that a solution curve y=¢(t) must be a curve which “‘flows with”
the direction field; that is, at each point on its graph, the graph of y=¢(t)
must be tangent to the corresponding line segment of the direction field.
This becomes evident if we superimpose on a direction field the graphs of
a number of solutions of the differential equation, as is indicated by the solid
curves in Figs. 1.10 and 1.11. Thus, after the direction field has been
sketched, we connect the line segments by smooth curves. In this way we can
obtain some useful information about the nature of the solutions of a
differential equation by a freehand sketch of curves which obey this simple
geometric condition.
Exercises
Construct direction fields for each of the following differential equations and then
sketch general solution curves:
1. jf =2y 2. y=—y"
3. y =ty 4, y=1-0?—-y?
Seyi aa 6. y=t?+y?
7. y=y-t? 8. y=-t/y
t+y
9. y=y/lt+y) 10. eee
1—t?)y-t
Hye
y
1.6 EXISTENCE AND UNIQUENESS FOR FIRST-ORDER EQUATIONS
It would be desirable to have a theory for differential equations that

guaranteed existence and uniqueness of solutions. Moreover, this theory
should be applicable to a large number of physical problems. In this section,
we shall explain the meaning and application of the basic theorems of this
type, without attempting to prove them. The proofs may be found in
Chapter 8.
In what follows, D will denote a given region in the (t, y) plane. By a
region we shall always mean a set of points with the property that, given any
point (to, Yo) in the region, it is possible to construct a rectangle centered at
(to, Yo) which lies entirely in the region (see Fig. 1.12).
Figure 1.12
For our purposes, rectangles are more convenient, but it would be possible
to use circles instead. A rectangle centered at (fo, vo) is the set
{(t, y) |—a<t—to
<a, —b<y—Yo <b} or, equivalently, {(¢, y) || t—to)<a,
ly —Yo)|<5}, for some a, b>0. The interior of a rectangle or circle, the
whole plane, and an infinite strip
{(t, y) |a<t<b, —a<y<o}
are frequently encountered examples of regions. A less obvious example is

the interior of a circle with the center removed.
1.6 Existence and Uniqueness for First-Order Equations 31
Observe that if (¢9, vo) is on the “boundary” of D, every rectangle cen-

tered at (fo, Yo) contains points which are not in D. (In fact, this statement
can be used to give a precise definition of the boundary of D.) In other words,
boundary points are not included in D. If (to, yo) is a point of D which lies
“close” to the boundary, then the numbers a and b in the definition of a region
must necessarily be chosen small.
In what follows, whenever we say that a certain function f is con-
tinuous on D we are assuming that it is continuous at all points of D, but we
make no assumption concerning its behavior on the boundary. If fis con-
tinuous at some boundary points of D as well as in D, then f is still con-
tinuous on the region D. In many problems which arise, this is precisely the
situation, but the continuity of fat some boundary points is not essential.
Exercises
1. Consider the region D= {(t, y) |(t—1)? +(» +2)?

<4}.
a) Construct at least two rectangles centered at the following points which lie
entirely in D: (1, —2), (0, —1), (2, —3), (1, —7o).
b) Show that the points (1, 0) and (—1, —2) are boundary points.
c) Consider the collection of points {(¢, y) |(t— 1)? +(» +2)?
=4}. Is this a region?
2. Which of the following collections of points are regions?
a) {(t, y) |t>0, —coo<y<oo}
b) {(t, y) | —wo<t<o, —2<y<oo}
c) {(t, y) |O0<t<1, —ao<y<oo}
d) {(t, y)|—w<t<a, —w<y<o}
3. Show that the collection of points {(r, y) | —1<t<1, y=17} contains only boun-
dary points.
4. Consider the region D={(t, y) | t?7+y?<1}. Are the following functions con-
tinuous on D?
eh) eel
1
b) g(t,iy) ») =——
1—-P—y3
c) Discuss the continuity of the functions

fand g of parts (a) and (b) if D is the
region {(¢, y) (t? + y7)<2}.
The main result on existence and uniqueness of solutions of initial-
value problems is the following.
Theorem 1. Let f and Of/0y be continuous functions in a given region D. Let
(to, Yo) be a given point of D. Then there exists an interval containing ty and
exactly one solution , defined on this interval, of the differential equation
y'=f(t, y) which passes through (to, Yo) (that is, the solution @ satisfies the
initial condition (to)=0).
Example 1. Consider the differential equation y’=ay (« constant). Here f(t, y)=ay
and (0f/éy) (¢, y)=«. Both fand df/dy are continuous in the whole (¢, y) plane. Theorem |
shows that there is a unique solution ¢ of y’=«y through every point (¢o, Yo) in the plane. |
We have seen in Section 1.1 that #(t)=y e*“~ is a solution of this initial-value prob-
lem. Note that from the explicit solution we see that this solution exists for —00 <f<oo,
but the theorem does vot give us this information.
Exercises
5. Find a solution of y’=«y through (1, 0). Is this the only solution through (1, 0)?
6. a) Show that —1/r is a solution of y’=y? passing through (—1, 1).
b) Show that $(¢)=—1/t is the only solution of y’=y? passing through (~—1, 1).
Be sure to determine an appropriate region D before applying Theorem 1.
c) What is the largest interval on which #(t)= —1/ris a solution of y’=y? through
the point (—1, 1)? The reader should observe that Exercise 6 shows that a solu-
tion of y’=f(t, y) does not necessarily exist for all ¢ even though f and éf/cy
are continuous in the whole plane.
Example 2. Consider the differential equation y’+p(t) y=q(t), where p and gq are
given functions, continuous on some interval a<t<b. (We may, of course, have p, g
continuous on the whole ¢ axis.) Choose tg with a<t )<b and any yo. Then we can use
Theorem | to show that there is a unique solution @ satisfying the initial condition
(to) =Yo. Since f(t, y)= —p(t) y+q(t) and df/ey(t, y)= —p(t) are continuous in the
region D={(t, y) |a<t<b, —0<y<oo}, Theorem | can be applied for every point
(to, ¥o) in D to show the existence of a unique solution. We have seen in Section 1.4
that the solution @ exists on the whole interval a<t<b.
Example 3. Consider the differential equation y’=3y?/?. Here f(t, y)=3y?/?,

f/6y =2y~*°. Thusf is continuous in the whole plane and @f/éy is continuous except
on the f axis (y=0). Theorem | tells us that there is a unique solution through any point
(to, Yo) With yo #0, but gives no information about solutions through (fo, 0). We can
easily verify that there is actually an infinite number of solutions through (¢, 0), as
follows. The function ¢, defined by
0 if —o<t<c
60-40) if c<t<oo
has a continuous derivative for — oo <t<oo and is a solution passing through (fo, 0)
for every value of c>¢o. This can be verified by direct substitution. In addition to all
these solutions, the identically zero function is also a solution (see Fig. 1.13). These
solutions can be constructed by the method of Section 1.3 (see Exercise 32, Section
1.3).
This example illustrates that in some cases there may be a solution ofthe differential
equation y’=f(t, y) through (¢9,vo) even though the hypothesis of Theorem | is not
(to, 0)
Figure 1.13
satisfied. In this example we have existence, but not uniqueness, of solutions through
(to, 0). Other situations may arise; for example, it can be shown that $(t)=0 is the only
solution of the differential equation y’=|y| passing through the point (0, 0), in spite of
the fact that 0f//Cy, where f(t, v)=|y|, does not exist at (0, 0).
These examples suggest that some or all of the conclusions of Theorem |

may be true even though not all the hypotheses are satisfied. This is
indeed the case; for example, if f is continuous in D, with no assumption
on ¢f/éy, it can be shown that a solution exists but is not necessarily
unique. However, in most problems, Theorem | is applicable and more
refined results are not needed.
Exercises
Use Theorem | to answer each of the following.
7. Does the differential equation y’=1+y? have a unique solution through the
point (0, 0)?
8. a) Does the differential equation y’=(y?—1)'/? have a unique solution through
each of the following points? (i) (0, 2), (ii) (1, —5).
b) What can be said about solutions through the point (0, 1)? Note that f(t)=1 isa
solution.
9. Determine the region or regions in the (, y) plane in which Theorem | can be applied
to obtain existence and uniqueness of real solutions for each of the following
differential equations.
y
a) a
y=—t/y) b
yy /—
7
t y
c) y ie eg rer d
)y
ee ee
l4e+y?
i+
) ya ) WP +ty'—y=0
8) eo ear h) (yP=?—y?
, Y /
k) y=1-t?-y? l) y=1+y?
m) y=t?+y? n) y=y-0?
ae
en =—— , (l-t?)y-t
yy
Oo
ae P) ———— =
F
0 (t<0;—w<y<o)
Vi=< Dy? (t>0; O<y<oo)
ve (t>0;—
0 <y<0).
a) Show that
is a solution on — 00 <f<oo.
b) Is ¢(t) continuous everywhere?
c) Is #’(t) continuous everywhere?
d) Can you apply Theorem | to obtain the existence and uniqueness of solutions
through the point (0, 1)? Explain fully.
We shall now consider another application of Theorem | dealing with

the geometry of first-order equations. While we have not proved it here,
all first-order equations can be solved, at least in theory, in terms of integrals.
However, these are often hard to evaluate and to interpret. Therefore, it is
difficult to draw any conclusions about how solutions behave and this, after
all, is an important matter. We can often get some idea about this,
without solving the equation, by exploiting the notion of direction field
further than we did in Section 1.5 and making use of Theorem 1. We
illustrate by considering an example.
Example 4. Consider the differential equation y’ = y’, and for the present let us ignore
the fact that we have already solved this equation in Example 1, Section 1.3. From
Theorem | we know that through each point (to, Yo) of the (t, y) plane there passes a
unique solution @. However, we do not know on how large an interval, both for t>f) and
for f< lo, this solution exists. We see, using the geometric interpretation ofsolution, that
the slope ofthe solution at the initial point is v3 >0 if yp40. Therefore the values ofthe
function #(t) increase for f> f) and decrease for f < fp (that is, as we move to the left from
fo); see Fig. 1.14. For values of t> fy and near fo, it is clear that, because ¢ (ft)increases as f
does, so does the slope $'(¢)=[(d)]?, and consequently the solution will increase more
rapidly as ¢ increases. It might therefore be expected in the case yg >0 not only that
$(t)> +00 as ¢ increases but also that this might happen as ¢ approaches some finite
value ¢, >fo. [This does in fact happen, as we see by looking at the explicit solution,
Eq. (1.13), Section 1.3.]
(to, Yo)
Figure 1.14
For values of t<f) the solution ¢ decreases as we move to the left and, when ¢(t)
becomes less than 1, the slope [(t)]? becomes smaller. Therefore the solution ¢(¢)
decreases less rapidly as t becomes more negative. However, the solution ¢ can never
cross the ¢ axis; by uniqueness (since the identically zero function is a solution), if it
did ¢ would be identically zero, contradicting the assumption yy > 0. Therefore we would
except the solution ¢ to exist for all ¢<¢ and lim, _,, @(¢)=0. Examination ofthe ex-
plicit solution, Eq. (1.13), Section 1.3, shows that this is indeed the case.
Now consider the case yp <0. Again, the initial slope is y>0, and the solution ¢
increases for t>f,) and decreases for t<fo (that is, as we move to the left from ¢9); see
Fig. 1.15. For values of t>¢o, #(t) will continue to increase, but more slowly as
|\p(t)| becomes smaller. As the solution ¢ cannot cross the ¢ axis (why?), we would
expect @ to exist for all t>fo and that lim,.,,, (t)=0 in the case yy <0. For t<fy
the solution will become more negative and approach — oo; this in fact happens for a
finite value ft; <¢o, as can be seen from the explicit solution (1.13), Section 1.3.
(to, Yo)
Figure 1.15
Exercises
11. Fora fixed yy >0, vary yp and use the argument given above to sketch the solution
curves in the (t, y) plane. Repeat for a fixed yy <0. What happens if y) =0?
12. By a similar analysis sketch the solution curves ofthe differential equation y’= — y’.
13. Sketch the solution curves of the differential equations y’=y and y’= —y.
14. Sketch the solution curves of the differential equation y’=1? +,7.
15. Sketch the solution curves of the differential equation y’=1/(t* +7).
We noted in the above examples and in Section 1.6 that, although -
Theorem | guarantees the existence of a solution on some interval, it gives no
information about the size of the interval of existence. What information
we have obtained about the size of the interval of existence has come from
explicit formulas for solutions. Since explicit formulas for solutions are
seldom available, it would be useful to have a criterion for the size of the
interval of existence. Such a result is beyond the scope of this book. However,
roughly speaking, the following is true: Under the hypotheses of Theorem 1,
the interval J of existence of the solution $(t) through the point (fo, Vo) is
such that the graph of the solution extends to the boundary of D (see Fig.
1.16). A proof of this result may be found in [2], Chapter 3.
y D
Figure 1.16
In particular, if D is the whole plane, or a vertical strip

{(t, y) |a<t<b, —ao<y<oo},
it can be shown, using the above result, that the solution exists on the
entire ¢ interval provided it is bounded on this interval. Note that, if D is
the whole plane, it is not necessarily true that the solution exists for all r;
this is well illustrated by Example 4: y’=y7, y(to)=yo9 analyzed above and
solved in Section 1.3, for which the solution
$(t) Stan
ze
exists only on the interval
1
—o<t<tyt+— if yo>O0.
Yo
CHAPTER 2
Introduction to
Second-Order
Differential equations ofthe type we wish to discuss arise naturally ina variety
of physical problems. Perhaps the simplest class of these problems deals
with the motion of particles. The mass—spring system (considered in Section
2.1) and the pendulum (developed in Section 2.2) lead to differential
equations which are prototypes of mathematical models for other important
physical systems. We shall use these typical examples to motivate our study
of second-order differential equations, just as we used some biological
problems to motivate our study of first-order equations.
Before turning to the construction of these mathematical models, we
recall some aspects of the Newtonian model for the motion of a particle.
In this model, it is assumed that a body, called a particle, can be represented
as a point having mass. (We shall assume knowledge of the rather difficult
concept of mass; for practical purposes, mass can be measured by the weight
of the body.) It is assumed that, in the absence of *‘forces,”’ the motion ofthe
particle is unaccelerated and is therefore straight-line motion with constant,
perhaps zero, velocity (Newton’s first law). The presence of acceleration is
therefore to be interpreted as a sign of the presence of a force. This is a
vector quantity given by Newton’s second law: If F‘ is the force acting ona
particle of mass m moving with a velocity v, then
F=< (my).
The vector quantity mv is called the momentum of the particle. If the mass
is constant, Newton’s second law may be written as
* Vectors will be indicated by boldface type throughout the book.
SH
38 Introduction to Second-Order Differential Equations
dv
F=m—=ma,
dt
where a is the acceleration vector of the particle and f is time.
In the Newtonian model, the gravitational force can be shown (experi-
mentally) to be proportional to mass, so that problems involving gravita-
tional forces on particles near the earth’s surface can be handled con-
veniently by assuming that the acceleration g due to gravity is constant.
2.1 THE MASS-SPRING SYSTEM
A weight of mass m is connected to a rigid wall by means of a spring with

spring constant A>0. An external force F(t)=A cos wt acts on the weight,
and the system slides in a straight line on a frictionless table as shown in
Fig. 2.1. Let y(t) denote the displacement of the weight from equilibrium
of the system, with »>0, whenever the displacement of the weight stretches
the spring and y<0 whenever the displacement r of the weight compresses
the spring. At equilibrium, y=0 and the spring is unstretched. Then we may
pose the following problems:
i) If the system starts from rest at time t=0 with an initial displacement
Vo, determine the motion of the system for r>0.
il— If F(t)=0 and if the system starts from some given initial state
y(0)=a, y’(0)=6, where a, 6 are not both zero, show that the system
executes a simple harmonic motion (here the prime denotes the
derivative with respect to r). Also determine the natural frequency of the
system (that is, the frequency when F(rt)=0).
iii)— Show that if @=(k/m)'/? (that is, if the natural frequency of the system
is equal to the applied frequency), then the solutions of the system
become unbounded at too. (This is the so-called case of resonance.)
F(t)
=A coswt
Uddddddddddae
Figure 2.1
2.1 The Mass-Spring System 39
In order to formulate this problem mathematically, we make the following

assumptions:
i) The spring has zero mass.
li) The weight can be treated as though it were a point mass of mass m.
iii) The spring satisfies Hooke’s law, which in the present context states
that the spring exerts a restoring force toward the (unstretched)
equilibrium position; the magnitude of this force is proportional to
the displacement of the spring from equilibrium, and the constant of
proportionality k>0 is called the spring constant.
iv) Neither the table nor the surrounding medium offer any resistance to
the motion of the system.
v) The system obeys Newton’s second law of motion.
Suppose that at any arbitrary time ¢ the system is in the position shown in
Fig. 2.1 y(t) units from equilibrium. We list a// the forces acting on the
particle of mass m at time f:
i) The restoring force of the spring is —ky(t) by Hooke’s law. (Note that
if y(t)>0 this restoring force acts to decrease the displacement,
whereas if y(t)<0 the restoring force acts to increase the displacement
y(t) of the particle of mass m.)
ii) The external force F(t)=A cos ot.
Thus the sum of all forces acting on the particle at time tis —ky(t)+ A cos at.
Now Newton’s second law of motion applied to the moving particle of mass
m States that
“(mv)=sum of the forces acting on the particle. (2.1)
Here, v is the velocity vector and mv is the momentum. (Newton’s law is

stated in terms of vectors. Since the motion of the present system 1s restricted
to a line, the vectors involved are one-dimensional and vector notation can
be eliminated.) Because the mass m is constant and because in this case
v=y'(t), we obtain, using (2.1),
my" (t)= —ky(t)+A cos at,
or the displacement y(t) satisfies the equation
k A
y(t) tet)y(t)=— cos cos at,
"(t)+—
which we write symbolically as

ek A
Verte eee CON, 22)
Equation (2.2) is a differential equation. It is linear (of first algebraic degree

in y and its derivatives) and of second order (since y” is the highest-order
derivative present). The mathematical problem is to find a function @¢,
called a solution of (2.2) defined in the interval 0<t<oo, such that ¢’ and
ob” exist and such that, for0<t<a,
k A
" (t)+— $(t)=— cos at.
m m
Since the system starts from rest at t=0 with an initial displacement yo,
we shall also require that the function ¢ satisfy
o(0)=yo, $'(0)=0; (2:3)

these are the initial conditions. The problem consisting of the differential
equation (2.2) and the initial conditions (2.3) is called an initial-value
problem. Once such a function has been found, the remaining questions
can be answered. This function ¢ (it can be shown that there is only one
such function) completely determines the motion of the system under as-
sumptions (i}{v).
A different set of assumptions leads to a mathematical model different
from the one described by Eq. (2.2) and the initial conditions (2.3). Consider
now the system described in the preceding example and in Fig. 2.1, but
drop the assumption that the system slides on a frictionless table. Instead,
assume that the table offers a force of friction proportional to the velocity
with constant of proportionality b>0.
i) Assuming that assumptions (i)-{v) hold, determine the motion of the
system with the same initial state as that of the system described above.
il) If the external force F(t)=0, discuss the behavior of the system in each
of the cases b* <4km, b*>4km, and b? =4km.
Here the formulation of the mathematical model is similar to the previous
one. We merely add —by’(t) to the forces considered previously; this
is the force caused by friction. Thus Newton’s second law says in this case
that
my" (t)= —ky(t)—by'(t)+A cosat, (2.4)
or
GD chd Kiera
Vi sian Vise =a COS. COL.
ie 1mm
Equation (2.4) is another differential equation; like (2.2) it is linear and of
the second order. The mathematical problem is now one of determining a
function y, defined for 0<?< oo, such that w’(¢) and w(t) exist for 0<t< oo
and such that
2.1 The Mass-Spring System 41
v'(j4+— yore
OS cosamMt,
m m m
for 0<1<oo. Such a function y is called a solution of Eq. (2.4) on 0<t<oo.
We also require that W should satisfy initial conditions
W()=yo. ww(0)=0. (2.5)

Equations (2.4) and (2.5) constitute another initial-value problem.
Equations (2.2) and (2.4) were obtained by considering the same
physical problem but under different simplifying assumptions. We stress
the fact that a set of physical assumptions different from either of those
made above might lead us to yet another mathematical model. We might,
for example, assume that the restoring force of the spring is proportional
to the cube of the displacement (instead of assuming that Hooke’s law
holds). This, assuming no friction, would lead to the equation
me oie eee
y +— y =— cos at
m m
in place of (2.2). Further, the accuracy of a particular mathematical model

in predicting physical phenomena depends primarily on the degree to
which the physical assumptions are reasonable.
You may now wish to acquire additional facility in the mathematical
formulation of simple physical problems.
Exercises
(In each of the following, assume that the particle is near the earth’s surface and that
gravity is the only force acting on the particle. You will have no difficulty in solving the
differential equations obtained.)
1. A particle is released from a height yo above the earth’s surface (y=0) with initial
velocity (velocity at t=0) in the vertical direction of magnitude vo. Find the
height y(t) of the particle above the earth’s surface and velocity v(t)=dy/dt as
functions of time.
2. A particle is released at an initial position so (distance s is to be considered positive
in the downward direction) on a frictionless inclined plane (as shown in Fig. 2.2)
Figure 2.2
with an initial velocity vg. Find its distance s(t), measured along the inclined plane
from the top, and its velocity v(t) as functions of time.
. A projectile is fired from an initial position (x9, yo) with an initial velocity of
magnitude vg at an angle of inclination 0, 0<@<z/2. Find its horizontal and
vertical coordinates x(¢) and y(¢) as functions of time.
. In each of the above exercises, discuss factors of actual physical situations which \
you feel have been neglected and consider their probable effect on the actual
behavior of the system. What information do you feel can be obtained from the
mathematical models you have constructed? In particular, compare the time it
takes for a body released from height A to reach the ground to that time for a
projectile fired horizontally with initial velocity vg.
. A weight of mass m is suspended from a rigid horizontal support by means of a
very light spring (see Fig. 2.3). The weight is allowed to move only along a vertical
Figure 2.3
line (no lateral motion in any direction is permitted). The spring has a natural
(unstretched) length L when no weight is suspended from it. When the weight
is attached, the system has an equilibrium position at which the spring is stretched
to a length L+a, where a is a positive number. We set the system in motion by
displacing the weight from this equilibrium position by a prescribed amount
and releasing it either from rest or with some prescribed initial velocity. Describe in
mathematical terms the motion of the system. [Remarks and Hints. Since the notion
is restricted to a vertical line, the position of the weight can be described completely
by the displacement y from the equilibrium position (see Fig. 2.3). The mathematical
equivalent of the motion of the mass-spring system is then a function @ such that
y= 6(t) describes the position of the weight for each value of >0, where t=0 repre-
sents the starting time of the motion. In order to determine the motion, that is to
determine the function ¢, we must impose additional restrictions on @. For example,
2.2 The Pendulum Problem 43
if we displace the weight a distance yg and then release it, we would require that
$(0)=Yo.- If we release it from rest at this position, we will also require that #’(0)=0.
Suppose as before that assumptions (i){(v) hold. With reference to Fig. 2.3,
we shall measure the displacement y from equilibrium (y=0), choosing the down-
ward direction as positive. The force ofgravity F, in Fig. 2.3 in mg, and the restoring
force of the spring F, is —k(y+a) by Hooke’s law. Observe that Fig. 2.3 has been
drawn with y>0 so that F, is directed upward [5]. Sketch the analog of Fig. 2.3
with y>0 and compute the forces F, and F, in this case. The total force acting on
the weight is
F,+F,=mg—k(y+a).
The equilibrium position occurs when this total force is zero. Therefore, at equilib-
rium, mg—k(0+a)=0 or a=mg/k. Thus we can rewrite the total force at any
position y of the mass as
mg
F,+F,=mg—k Ve =—ky.
By Newton’s second law,
d 2
eo yp— ky=F imu)=m 2
2.2 THE PENDULUM PROBLEM
A pendulum is made by attaching a weight of mass m to a very light

and rigid rod of length L mounted on a pivot so that the system can swing
in a vertical plane. If we set the weight moving, can we describe the nature
of its motion? For example, can we answer questions such as the following:
(1a) How does the nature of the motion depend upon the mass m and
the length of the rod L? (1b) If we hold the weight out an angle 6,
(see Fig. 2.4) and then release it, how will the speed it attains at the bottom
of its swing depend on m, L, and the angle 0? (Ic) We know from ex-
perience that the weight will swing back approximately to the original
Figure 2.4
position 0). How long will this take? How does this period depend on m,
L, and 0)? (1d) What happens if we immerse the pendulum in some substance
such as water, which exerts a considerable amount of resistance on the mov-
ing weight? How will this affect the nature of the motion? Will the
pendulum eventually come to rest, and, if so, how long will this take? (le) Will
any of the previous answers change if we move the apparatus to the top of a
high mountain or place it in an orbiting satellite?
We have assumed that the rod supporting the weight is perfectly rigid.
This means that the position (state) of the pendulum at time ¢ can be
described completely by the size of the angle 0 (Fig. 2.5). The mathematical
Figure 2.5
equivalent of the “motion” of the pendulum is then a function @ such

that 0= b(t) describes the position of the pendulum for each value of t>0,
where ¢=0 represents the starting time of the motion. To determine the mo-
tion, that is, to determine the function ¢, we must impose additional
restrictions on @. For example, in the question (1b) posed above, we would
require that (0) = 6p. If we released the pendulum from rest at that position,
we would also require ¢’(0)=0.
We must next apply physical principles to obtain additional restrictions
on the function ¢. Again, the basic tool is Newton’s second law. In order
to apply this to our pendulum problem, we shall make a number of
simplifying assumptions:
a) The rod is rigid, of constant length L, and of zero mass.
b) The weight can be treated as though it were a particle of mass m.
c) There is no air resistance; the pivot is without friction, and the only
external force present is a constant vertical gravitational attraction.
22 The Pendulum Problem 45
At any time ¢, the gravitational force F, has magnitude mg and is

directed downward. There is also a force F, of tension in the rod of
magnitude 7 directed along the rod toward the pivot. Let u, denote a unit
vector in the radial direction and let uy denote a unit vector in the tangential
direction, as shown in Fig. 2.5. Then F, can be expressed as
F, =F u,+ Fou,
where F,, Fy are the components of F, in the radial and tangential direc-
tions, respectively, F,=mg cos0, F,= —mg sin@. Also F, can be expressed
as
Eo= a Tu,
where T>0. We recall from calculus that the vectors u,, u, satisfy the
relations
= dy
do je
Let r denote the position vector of the weight, taking the pivot as the origin,
so that
u=Lu.
Differentiation, using the chain rule and the above relations, gives
a a dt
and
d’r do dd\?
= —L 2.6
de) ae (7) ss a
The total force acting on the weight is
F,+F,=(mg cos 0—T) u,—mg sin
0uy. (2.7)
By Newton’s second law (since the mass is constant),
d’r
Thus, equating coefficients of u,, uy in (2.6) and (2.7), we obtain
do 2 d
2 ,
~mL (5) =mg cos @—T, mL—,=—mg sin@.

dt?
Note that the first of these equations contains also the unknown quantity
T. However, if the angle @ can be determined from the second equation,
then the magnitude of the tension 7 can be found from the first equation;
in fact, T=mg cos 0+mL(d6/dt)’. Therefore the motion of the pendulum
is completely determined by the second equation, which may be written
in the form
7A
art 4 sind=0. (2.8)

This equation is the mathematical model for this simplified pendulum
problem. It is a differential equation (of second order). In deriving this
equation we have used only assumptions (a), (b), and (c), and we have not
yet imposed any additional restrictions. If we wish to be more specific
about the questions we ask, we must specify other information. For
example, question (lb) asks us to predict the speed of the weight at the
bottom of its swing, if we hold the weight out at an angle 6, and release
it from rest. The last phrase is equivalent to the pair of initial conditions
P(0)=%, —f'(0)=0. (2.9)

Question (1b) can then be restated as: ‘‘What is Ld'(t) when 0=02”
The mathematical problem is then to find a function ¢, defined for
all t>0, such that 0=¢(t) satisfies the differential equation (2.8), that is,
o"()=—F sinb(t) (t>0)
and such that (2.9) also holds. If we can find such a function, we say that
we have found a solution of the differential equation (2.8) obeying the initial
conditions (2.9). We may then hope that this function ¢ will give a good
approximation to the actual motion of a specific pendulum, and that we
can use ¢ to predict properties of the pendulum which can be measured
experimentally.
We can modify the model in several ways, some motivated by physical
intuition and some by mathematical requirements. For example, we could
replace assumption (c) by:
c’) The pendulum encounters resistance, due to the pivot and the
surrounding air, which is proportional to the velocity vector,
and leave the remaining assumptions unchanged. In this case,
d70 g a dé
= —— sind 2.10
dt? ea 210)
replaces Eq. (2.8) as the appropriate mathematical model. The last term is
the appropriate mathematical translation of the additional resistance force.
Note that Eq. (2.10) reduces to (2.8) if a=0.
22 The Pendulum Problem 47
Exercises
1. Derive Eq. (2.10). |Hint: The resistance force is aL (d0/dt) ug.|

2. What is the magnitude of the tension, assuming that @ has been found from Eq.
(2.10)?
There are many other possible modifications of the model. We could,
for example, attempt to imitate nature more closely by allowing the pendu-
lum rod to be slightly elastic, and let it increase or decrease its length in
accordance with the magnitude of the radial component F,. This ought
to result in a mathematical model whose predictions come much closer
to agreeing with experiment than either (2.8) or (2.10). However, it would
be such a complicated model that it would be extremely difficult, at the
present state of mathematical knowledge, to extract useful information
from it in a simple way.
Until research in mathematics makes it possible to develop easier
techniques to work with complicated mathematical models, we must be
content to work with simpler models which are likely to give less accurate
predictions.
Sometimes, in fact, a complicated model is replaced by a simpler model
that is mathematically related to the original one, even though there may be
no physical justification for this step. Let us reexamine the differential
equation
d*6
oe
g
7 es Pea
7 sind (2.8)
2.8
If @ is small, then sin@ and @ have almost the same value. (Consult any
table of trigonometric functions for radian arguments and also recall that
limy_.9 sin0/@=1.) We might therefore be inclined to replace (2.8) by the
linear equation
d*0 g
rs 0 (2.11)
and hope that this new mathematical model (which, as we shall see, is
much easier to solve) is almost as good for some purposes as the model
(2.8). When we analyze these equations, it turns out that (2.8) and (2.11)
predict quite different qualitative behavior for pendulums, and that some of
the predictions made by (2.11) do not happen at all, but that both agree quite
well with experiment if we restrict the swing of the pendulum to a rather
small arc. (Incidentally, Eqs. (2.8) and (2.11) also predict things that a
‘real’ pendulum does not do at all; for example, they predict that once a
pendulum is set swinging, it will never stop! Equation (2.10) comes closer to
reality by at least predicting that the pendulum will slow down.)
2.3 FURTHER EXAMPLES OF MOTION PROBLEMS
We now formulate several more problems of motion which we shall be able

to solve in the next section.
Example 1. A steel sphere is projected vertically upward from the earth’s surface
with a positive initial velocity vp. Does it take the sphere longer to reach its maximum
height or to fall back to the earth from that height?
Assume that the acceleration of gravity (g) is constant, that the air is still, and that
the retarding force of air friction is proportional to the velocity of the sphere relative
to the air through which it moves. Call the constant of proportionality p.
If you claim that one time is longer than the other, give a mathematical justification
of your claim (not an intuitive physical argument).
Consider Figs. 2.6 and 2.7 for the ascent and descent, respectively, of the ball,
oy
Air resistance
—mg {Py
Ground
Ascent (y’ > 0)

Figure 2.6
te
Air resistance
Ground
Descent (y’ < 0)

Figure 2.7
in which y(t) is the height of the ball above the surface at any time f. We note that in
both the ascent and descent the total force acting on the ball at time ¢ is
—mg—
py’ (t).
2.4 Second-Order Equations Solvable by First-Order Methods 49
Hence by Newton’s second law, the equation of motion is
my” (t)= —mg—py’(t).
We are given that at =0, y is zero and y’ is v9. Hence the mathematical problem is to
solve the initial-value problem
my" +py =—mg, y(0)=0, y' (0)=v. (2.12)
We shall solve this problem in the next section.

Example 2. Rocket Flight. The force of gravity is not constant, but depends on the
distance from the center of mass of the earth (and to some degree on densities of
local rocks). When we are dealing with large distances, it is a reasonable approxima-
tion for some purposes to assume that the force of gravity on an object of mass m out-
side the earth’s atmosphere is given by F=kmy7, where k is a constant and y is the
distance from the object to the center of the earth, which is assumed spherical. During
the initial “powered-flight” period of a rocket, the laws of motion are rather compli-
cated. After “burnout,” things are much simpler, since the rocket is merely coasting,
while its velocity is being decreased by the effects of earth’s gravity. We can therefore
ask the following questions (we assume that the rocket moves directly away from the
earth).
a) Knowing the height and velocity of the rocket at burnout, can we find how high
it will go before beginning to fall back to earth?
b) Can we determine what velocity it must have at burnout in order to continue
rising indefinitely?
We assume that the burnout height is essentially the radius of the earth R. Let y
be the distance from the rocket to the center of the earth. From Newton’s law of
gravitation, we see that the differential equation for the motion is y” = —k/y*, where k
is chosen so that at the surface of the earth (v=R) the acceleration is —g. Therefore,
substituting y=R and y’=—g into the equation y’ = —k/y*, we obtain —g= —k/R?
and thus k=gR?. The initial conditions are that at -=0, y=R and dy/dt= V(velocity at
burnout). Thus the mathematical problem is to solve the initial-value problem
y"=—gR’ly’, yO)=R, yy O)=V. (2.13)

We shall solve this problem in the next section.
Many other physical problems lend themselves to mathematical formu-
lation as second-order differential equations. In the next section we shall
give some exercises on the formulation and solution of such problems.
2.4 SECOND-ORDER EQUATIONS SOLVABLE

BY FIRST-ORDER METHODS
The differential equation in Example 1, Section 2.3, is

my” + py'= —mg. (2.14)
This equation does not contain the unknown function y explicitly. Thus
it can be viewed as a first-order linear differential equation for y’, and it

can be solved by the method of Section 1.4. Equation (2.14) is a special case
of the equation
y'=g(t, y), (2.15)
where g does not depend explicitly on y. Let y’ be a new variable v. Then
y” =v’, and (2.15) becomes the first-order equation
v' =g(t, v) (2.16)
together with the relation v=y’. If we can solve (2.16) by some method
such as those of Section 1.3 or Section 1.4, we obtain v as a function of
t. We then find y by integrating the equation y’(t)=v(c).
If (2.15) is accompanied by the initial conditions y(to)=¥o, y’ (to)=Zo;
then we require the solution of (2.16) to satisfy the initial condition
v(to.)=y'(to)=Zo. When we integrate to find y from y’(t)=v(t), we im-
pose the initial condition y(to)=yvo. Thus, assuming that v(t) has been
found from (2.16), with v(to)=Zo, we obtain
= 0+ |el)
as the solution of (2.15) for which y(to)=yo, y’ (to) =Zo.
Example 1. Solution of Example 1, Section 2.3.
We wish to solve the initial-value problem
my" +py' = —mg, y(0)=0, y(O=25: (2.17)
As above, let y’=v; then y”=v’ and (2.17) becomes

ak:
Veta
+— ==. g v(0)=uo
) —— i)
This is a first-order linear equation in v. By the method ofSection 1.4, its solution is found
to be
m m
o()= (to) exp(—2 ‘|i.
p m p
Thus, integrating y’(¢)=v(t), v(0)=0, we obtain
(or) ha exp( P ln (2.18)

[Ley hay m p
as the (unique) solution of the initial-value problem (2.17).
Exercises
I. Verify that y(1) defined in (2.18) is a solution of the initial-value problem (2.17).
2. Explain why (2.18) is the unique solution of (2.17).
The initial-value problem (2.17) is now solved. To answer the question whether the
sphere takes longer to reach maximum height or to fall back to earth from the maximum
height, we let fo be the time of ascent to maximum height. Thus
, gm gm
v(to)=y t)=(v0+ 2)ep(—? in) =O,
p m p
and therefore, solving for tg, we obtain
tom bog MM), 2.19)

+
Dp mg
To determine whether the time of ascent exceeds the time of descent, we compute y(27o).
Clearly, if y(2t9)>0, the time of descent exceeds the time of ascent; and similarly one
interprets the case of y(2t))<0. One has from (2.18),
mg\m 2 2
v(2ts)=(vo+ M2)*]1-ex(-P rn)|
P/?p m p
= (00+) ™ exp(—2 a)|exp(22) —exp(— 2)|2,

p p m
t
m
ff
m
2D
p
Let exp(pty/m)=x. From (2.19),
pto =log
(“!SEre:
m mg
and hence
mg+ Pvo
x=———_..
mg
Substitution in (2.18) gives
We observe that in our case x>1, because vp >0. The function f defined for x>1 by
1
f (x)=x—-—2 logx
x
has
1 2 /x-1)
Poj=t+a-24( Jeo
x
Therefore f(x) is increasing. Since f(1)=0, one has f(x)>0 for every x>1. Therefore
y(2to)>0, and the ascent is faster than the descent. We remark that this is the result
one would expect on physical grounds as well.
The differential equation in Example 2, Section 2.3, is
wu” gR* (2.20)

a
which does not contain the independent variable ¢ explicitly. Thus it is a
special case of the equation
y’=h(y, y). (2.21)

This equation can be reduced to a first-order equation by the following trick.
Let y’=p and think of p as a function of y. By the chain rule,
, 4p dpdy dp
dt dydt dy
Thus Eq. (2.21) becomes
Ply h(y, Pp) (2.22)

which is a first-order equation for p as a function of y. Assuming that it
can be solved by one of the methods in Chapter 1, we then obtain a solution
y(t) of (2.21) by solving the separable first-order equation
y (t)=p(y). (2.23)
Up to now we have not mentioned initial conditions. If (2.21) is ac-
companied by the initial conditions y(t9)=Vo, y’(to)=Zo, we must solve
(2.22) subject to the initial condition p(yo)=Zo, because Zo is the value of
y’ corresponding to the value yo of y (namely at t=/)). Once p has been
thus determined, we need only solve (2.23) subject to the initial condition
V(to)=Yo-
Because the left-hand side of (2.22) is p(dp/dy), there may be diffi-
culties whenever p=0. You are therefore warned that you should verify
by direct substitution the result of solving any problem by this method.
To clarify this method, we consider three examples.
Example 2. Find the solution of the differential equation
Pn)
y=dy (2.24)
satisfying the initial conditions
y(O)=1, y(0)=-1. (2.25)
Equation (2.24) does not contain the independent variable explicitly, and is a special
case of (2.21). Therefore let y’=p(y); then y’=p(dp/dy), and (2.24) becomes
et (2.26)
Equation (2.26) can be solved by separation of variables. The appropriate initial con-
dition is p(v(0))=p(1)=y’ (0)= —1. Separating variables in (2.26), we have
Ded pen
l+p?dyy’
or
en
ape atity
a ayer
tae |og|yl.
Integrating with respect to y from y=1, we obtain, using p(1)=1,
log(1+p?)—log2=log y—log1!
or
l+p 2 oF
log yo HOBY-
Thus (1 +p?)/2=y, p?=2y—1, and
pi=#Qy—1)!?.
Since p(1)=y'(0)= —1, we must choose the negative square root, and
p(y)=—(Qy—-1)"”.
To find y, we recall that p(y)=y’(t). We must therefore solve the differential equation
y= —(2y—1)?
subject to the initial condition y(0)=1. Separating variables again, we have
y(t)
‘(i Les
Qy()-1)'?
©aie[ev()—)'2]=-1.
Integrating with respect to ¢ from t=0, we obtain, using y(0)=1,
(2y(t)—1)'?-1l=—t.
Thus
2y(t)—1=(-t+1),
y(t)=3[1+(1—-0)7]. (2.27)
The reader may verify by direct substitution that (2.27) is a solution of (2.24) satisfying
the initial conditions (2.25).
Exercise
3. Verify that (2.27) is the solution of (2.24) satisfying the initial conditions (2.25).
Example 3. Solution of Example 2, Section 2.3.

We wish to solve the initial-value problem
—gR2 ‘
y=, yO=R, yO=V. (2.28)
We let y’=p(y). Then y" =p(y) (dp/dy) and (2.28) becomes
d
peeRe a= (2.29)
dy y
This equation is of first order in p with variables separable. It can be solved by the
method of Section 1.3, and its solution is
2gR?
[p(y]? =V7+ —2gR. (2.30)
y
Exercise
4. Separate variables in (2.29), impose the initial condition, and obtain (2.30).
To solve (2.30) for y as a function of ¢ involves solving the separable equation
70m"
¥()=p0)=#( 72-20R+ (2.31)
subject to the initial condition y(0)= R. The choice of sign in (2.31) is determined by the
physics of the problem (the plus sign when the rocket is rising). Unfortunately, the solu-
tion of (2.31) leads to an integral which cannot be evaluated in terms of elementary
functions unless V7 =2gR.
Exercise
5. Solve the initial-value problem
v=)",
2gR*\"/2
y
v=,
(the case of (2.31) V?=2gR).
In spite of the fact that (2.31) cannot be solved in general, we can obtain some
interesting information about the motion of the rocket. For example, we can deter-
mine whether the rocket will reach a maximum height and then return to earth or
whether it will continue to rise and will escape from the earth’s gravitational field.
The rocket will reach its maximum height when y’(¢)=0 or when
Ly (0)? =p?(y)=V2+—
? 2. 2 Re
2g =0.
If V7 <2gR this height is
y 2gR?
: 2aR= Vy)
If V*=2gR the expression for the maximum height is undefined. However, we have
in this case
_ 2gR?*
Ly’ (t)]? y
which was solved in Exercise 5. It is therefore clear that if V? =2gR the rocket ‘‘slows
down to zero” only as yoo. What this means is that the initial velocity V is sufficient
for the rocket to overcome the force due to gravity. Thus V=(2gR)!/” is called the
escape velocity of the rocket.
If V?>2gR, then
]*/?
ly’ (t)| =[e+(2gR?2/y
where e= V?—2gR>0.
With this initial velocity the rocket will certainly overcome the gravitational forces,
and since
9) R2 1/2 _
lim (:+ Z ) =,/e,
yot+o y
the rocket will have a velocity of Je when it leaves the earth’s gravitational field.
If g=32 ft/sec* and R=4000 mi, the escape velocity is given by
43 mi
V=(2gR)'?=
(33)4/? sec.
Exercise
6. In the case V*=2gR, show that the solution y(¢) of Exercise 5 tends to +00 as
too. (This means that the rocket continues to rise.)
Example 4. Solve the equation of motion ofthe linearized simple pendulum
d’0 eB g
ee
ae L (2.32)
subject to the initial conditions 0(0)=6 , 0’(0)=0. This initial-value problem was
derived in Section 2.2, Eq. (2.11), as an approximation of Eq. (2.8).
The equation to be solved is of the form (2.21), which has the independent variable
t missing. Using the method described for this case, we let @’=p(@) and we think of p
as a function of @. By the chain rule we then have
d°@dp dpd@ dp
dt? dt dodt do"
Thus (2.32) becomes
which is a first-order equation in p which has variables separable. The initial con-
dition now becomes p(@(0))=p(@))=0. By the method of separation of variables we
obtain
Since p(6))=0, we obtain
and therefore
This solution is valid for 6? <63 or —|@)|<@< {6p}.

Note that if p(@)>0, the above solution is
p)=(2) (3-8), (2.34)

g 1/2
while if p(@)<0, the solution is
p()= -(2) (63

—62)!g ie 2 2 2
(2.35)
yc
You may find it helpful to draw a sketch of the two situations. To solve the given
problem for @=@(f) we must treat these two situations separately. Taking (2.34) first,
letting k=(g/)'*, p=’, we have
0'()=k[B-9?()}'7, $0)=8.
Separating variables and integrating from 0 to ¢, we obtain
t
$'(s)
—._—_-—_~ ds=kt..
fimo An
0
Making the change of variable v= @(s) in the integral, and using 6(0)=6. we obtain
ou)
du
|(Bw)——__=kt.
80
2.4 Second-Order Equations Solvable by First-Order Methods Sy
The careful reader will now notice that this integral is improper because the
integrand becomes infinite when u*—>
05. Thus by the definition of such integrals (if they
exist), and by use of integral tables or otherwise, we obtain
(0) o()
du , du
ee
2 2\1/2
ein TT
(05—u )/ e> 0+ (03 —u?)!/?
00 Oot+e
yu |e
= lim | arc sin —
e70+ 0 Oo + €.
=arc
= One
sin ——— lim
_ Oo
arc sin ———
0 E> Ot 0
_ p(t)
= arc sin ———arc
:
sin l
0
_ P(t) 7%
=arc sin ————_.
Chy
Thus
t t
are sin Des, or arcsin CA ee
Oy 2 A
and consequently
(t)=9 sin( [2142)=0,cos, (2.36)
Exercise
7. Apply the method just used to solve (2.35) in the case p=’ <0 and obtain
(t)=9 sin(5- [2‘=, cos |is
which is the same result as was obtained for the case p>0.
In the next chapter we shall learn a much simpler way to solve linear
differential equations such as (2.32).
Exercises
8. Use the method of this section to solve the initial-value problem
y"=4y, yO=1, y(O)=0.

9. Solve the differential equation of Exercise 8, subject to the initial condition
y(0)=0, y’(0)=1.
If we apply the method of Example 3 to the nonlinear pendulum problem
0” (t)= =¢ sint, 0(0)=0, 0'(0)=0, (2.37)

which was derived in Section 2.2, Eq. (2.8), we obtain, exactly as in
Example 4,
2g
Ip(A)| =|¢' (H|= (2) [cos p(t)—cos 6 |". (2.38)
Exercise
10. Derive the result (2.38) by applying the method of Example 4 to the initial-value
problem (2.37).
To solve (2.38) for 0= p(t), we must again consider the two cases ¢’(t)>0,
f'(t)<0. In the case #'(t)>0, separation of variables now yields
P(t)
Wyo?
aa (#) t,
(cosu—cos6o)"? \L
(2.39)
90 |;
in which the integral is also improper. Unfortunately, this integral cannot

be evaluated in terms of elementary functions, although it can be shown to
exist and also it can be approximated numerically. The case $’(t)<0 leads
to a similar integral.
Although (2.39) cannot be integrated precisely, it can still provide
useful information. For example, p(@) is the angular speed of the non-
linear pendulum at the angle 0. At the lowest point of the swing 0=0 and
the angular speed at this point is
2g 1/2 i g 1/2 2)
(7) (1—cos0,) 7=2 = sin >.
Exercises
li. Find the angular speed at the lowest point of the linearized pendulum (2.32) and
compare the results if 69 =7/2, if 09=2/3, and if 69=1/10.
*12. Consider the initial-value problem
p—=-— 9, P(0o)=0 (2.40)
which was solved in Example 4. Why is the existence and uniqueness theorem for
first-order equations (Theorem 1, Section 1.6) not applicable to this problem? (Note
that in Example 4 there are actually two solutions: the one we obtained and the
solution p=0.)
alse Repeat Exercise 12 for the initial-value problem
dp 9.
i sind, p(Ao)=0 (2.41)
which arises in Exercise 10.

*14. What can you say about the applicability of the existence and uniqueness
theorem in Exercises 12 and 13 if the initial value is changed to p(0))=1?
1S: Find the solution of the initial-value problem y”=y, y(0)=1, y’(0)=—1.
16. Find the solution of the initial-value problem y”=y'+1, y(1)=0, y’(1I)=—1.
ULE Find the solution of the initial-value problem y’=$ y'y*, y(0)=1, y’(0)=1/3.
yy ty y=
(van der Pol equation), which occurs in the theory of vacuum tube circuits.
a) Show that the change of variable y’=v transforms this to the form
dv
SI Cer
dy v
b) Sketch the direction field and a few solution curves of the equation obtained in
part a).
19: Given that y”= —4/y° and y(2)=4, y’(2)=0, find y(4).
20. Solve each of the following initial-value problems:
a) yy’ +) +1=0, y(1)=1, y'(1)=1
b) 2yy’+(y’)?+1=0, y(l=1,y(D=1
7A. Find the solution of y’=[1+(y’)*]*? such that y(0)=1, y’(0)=0.
elms A small solid sphere of density y (pounds per cubic foot) falls from rest under the
influence of gravity into a large reservoir of liquid whose surface is h feet below
the point at which the sphere is released (see Fig. 2.8). The density of the liquid is
ky, where k <1. The frictional resistance of the liquid to the motion of the sphere
is proportional to the velocity of the sphere, with constant of proportionality p.
Determine the maximum depth to which the sphere sinks.
Figure 2.8
Assume that the reservoir is sufficiently deep that the sphere does not strike
the bottom, and neglect air resistance. Make the approximation that the sphere
is either entirely in air or entirely in the liquid and that the instant of entry
coincides with the time when the center of mass of the sphere passes through the
free surface of the liquid. Recall that a submerged body is buyoed up by a force
equal to the weight of the liquid displaced. If the mass of liquid displaced is w Ib,
then this force is wg. :
*23. A small sphere weighing 0.1 pound is projected vertically upward with an initial
velocity of 1500 ft/sec from a height of 1000 ft above the earth’s surface. It is sub-
sequently acted upon only by gravity and by air friction. The resistance force of
air friction (in pounds) is 10~7 times the square of the velocity in feet per second.
Assuming that the acceleration of gravity has the constant value 32 ft/sec/sec, draw
a graph showing the altitude of the sphere as a function of time from the instant of
release until the instant of contact with the earth.
24. A spherical mass grows at a rate proportional to its instantaneous surface area.
Assuming that the sphere has an initial radius a and that it falls from rest under the
influence of gravity, show that its acceleration at time ¢ is
g y 3a*
4 ie fF
where r is the radius at time ¢. Thus show that the acceleration is constant if and
only if the sphere has zero initial radius.
25. According to Einstein’s special theory of relativity, the mass m of a particle varies
with the velocity v according to the rule
Mo
m=
where c is the velocity of light and my is the mass when v=0 (rest mass). If a particle
falls from rest under a constant gravitational field, show that its velocity at time
t is v=c tanh gt/c, and determine the distance fallen in time f.
2.5 THE EXISTENCE AND UNIQUENESS THEOREMS

FOR SECOND- AND HIGHER-ORDER EQUATIONS
We have seen that equations of the second order arise naturally in many
physical examples. These are of the form y”=g(t, y, y’), where g(t, y, z)
is defined in some region D in the three-dimensional (t, y, z) space (for
convenience we let y’=z). The concept of a region in three-dimensional
space is a natural extension of the two-dimensional situation, with rec-
tangles centered at (fo, ¥o) replaced by rectangular parallelepipeds cen-
tered at (fo, Vo, Zo). Such a parallelepiped is a set of points of the form
{(t, ¥, 2) | t—tol<a, ly—yol<5, |z—Zo|<c} for some a, b, c>0.
We suspect from the physical examples discussed earlier that in order
to obtain a unique solution we should prescribe not only an initial position
ZS Existence and Uniqueness Theorems 61
but also an initial velocity at a time fo. In terms of a solution @¢, this means
that we prescribe #(to) and ¢'(to). Keeping in mind the above differences
from the first-order case, we can state the following analog of Theorem 1,
Section 1.6, for second-order equations.
Theorem 1. Let g, 6g/0y, and 0g/éz be continuous in a given region D. Let
(to, Yo. Zo) be a given point ofD. Then there exists an interval I containing to
and exactly one solution ¢, defined on this interval, of the differential equation
y= g(t, y. y’) which passes through (to, Vo, Zo) (that is, the solution ¢ satisfies
the initial conditions $(to)=Yo, $' (to) =Zo-)
Example 1. What does Theorem | say about the initial-value problem
6” = —k?0, 0(0)=4, 0’(0)=0, (2.42)
which was solved in Example 4, Section 2.4, with k?=g/L?

Here the unknown function is @ in place ofy and the function g in Theorem | is
g(t, 0, 0)=—k20.
This function g is continuous in the entire three-dimensional (¢, 0, z=0') space. The
partial derivatives
(t,0,0')=—k?, —(t,0,0')=0
DIS
Sele
are also continuous in the entire (¢, 0, 0’) space. Thus, by Theorem 1, given any 0p, the
differential equation has one and only one solution 0=¢(t) defined on some interval
I (the theorem doesn’t say how large / is), satisfying the initial condition ¢(0)=p,
g'(0)=0. (We note also that if we impose the more general initial condition 0(to)=a,
0'(to)=b, where to, a, b are any real numbers, the result is the same.) We have found
$(t)=0, coskt as a solution of (2.36) in Example 4, Section 2.4. By Theorem |, this is
the only solution. Note that from the explicit formula for the solution, we can say
that it exists for —co0<t<oo.
Exercises
1. What does Theorem | say about the existence and uniqueness of solutions of the
initial-value problem
O’=—k? sind, 6(0)=%, 0'(0)=0, eat

which was discussed in Section 2.4?
2. Repeat Exercise 1 for the damped nonlinear pendulum equation (see Eq. (2.10),
Section 2.2):
6’=—k2 sind—-- 6, 0(0)=%, 6'(0)=0, k=

m ris
3. Discuss the existence and uniqueness of solutions ¢ of y’+p(t) y’'+q(t) y=/(0),

with initial conditions $(to)=Vo, $'(to)=Zo, where p, g, and

f are given functions
continuous on some interval a<t<b, where a<ty <b.
4. a) Show that #(t)=0 is the only solution of y’+p(t) y’+q(d) y=0 satisfying the
initial condition $(0)='(0)=0, if p and g are continuous on some interval con-
taining 0 in its interior.
b) Show that if y/(f) is a solution of y”+p(t) y’+4(t) y=0 which is tangent to the ‘
t axis at some point (¢,, 0, 0), then y(s)=0.
5. It is easily verified that c, cos 2¢+c, sin 2tisa solution of y’+4y=00n —c <f<oo
for every choice of the constants c, and c).
a) Determine c, and c, so that this solution satisfies the initial conditions
seu «fie
(en a
b) Write the solution @ satisfying the initial conditions
and prove that this is the only solution satisfying these conditions.
Equations of Higher Order

There are problems where higher-order equations are encountered. For
example, in elasticity theory equations of the fourth order arise naturally.
We are interested in equations of order n, of the form y”=h(t, y, y’,...,
y"~)). We observe that the cases when n=1 and n=2 are the ones we have
already discussed. Here / is a function defined in some region D in the (n + 1)-
dimensional (t, y;, ¥2,..., ¥,) space. A “parallelepiped”’ in (n+ 1) dimensions
centered at (fo, 71, %25--+. Mp) is a set of the form
{G Vip Yon Yt (b= toa |e Va Nel one
WREIC Vyu=V Vo) os ea and a, b,,...,b, are positive numbers.

t—)))
We have the following analog of Theorem 1.

Theorem 2. Let h, 0h/éy,,..., Ch/Oy, be continuous in a given region D.
Let (to, M15 ++-5 Mn) be a given point of D. Then there exists an interval con-
taining ty and exactly one solution , defined on this interval, of the dif-
ferential equation y =h(t, y, y',..., v"~) which passes through (to, Ny...
Nn)s (that is, the solution satisfies the initial conditions $(to)=n,, $' (to) =n.
oy PO
(to) =Mn)-
Example 2. What does Theorem 2 say about the initial-value problem
yh = ty” = 3y’ + 2ty?, 9
YO=O,.VO=sla yi Oat yO).

25) Existence and Uniqueness Theorems 63
Here the function 4 of Theorem 2 is
h(t, V1, Yas 3, Ya) = 2ty? —3y, +ty3

which is continuous in the whole five-dimensional space. The partial derivatives
oh ; oh ; oh 5 oh 0
—=4ty,, —=-3, =5). =
Oy, ae Oy, Oy3 OVs
are also continuous in the whole of five-dimensional space. Thus, by Theorem 2, the
initial-value problem (2.43) has a unique solution y=¢(¢) existing on some interval J
containing t=O of unspecified length.
Exercises
6. a) Show that the differential equation y+ 2y”+3y=0 has a unique solution ¢

satisfying the initial conditions @(1)=1, ¢’(1)=0, ¢”(I)=—1, 6” (1)=2.
b) Show that y/(t)=0 is the unique solution of this equation satisfying the initial
conditions (—1)='(—1)=w"(—1)=’"(—1)=0.
7. Show that $(t)=5 is the unique solution of y” +(y—5)?=0 satisfying the initial
conditions $(to)=5, (to)=" (to)=0 for any tp, —00<ty<oo.
As can be seen from the Examples and Exercises following Theorems |

and 2, these existence and uniqueness theorems guarantee that a solution
exists on some interval J, but they give no information about the size of
the intervals /. If one is so lucky as to find an explicit solution, one can
get such information. There are criteria for the size of the interval of existence
even if no explicit solution is known; the situation is similar to that of
the first-order case discussed at the end of Chapter 1. We shall not go into
this problem further.
Exercises
8. Determine the region or regions in (t, y, y’) space in which Theorem | can be applied
to obtain existence and uniqueness of real solutions for each of the following dif-
ferential equations.
gd 2 /\2 ee
a) y= +y+(y) b) y Tee
) ee d) »"=log(y"/)
e) y’+(sint) y'+(logt) y=0 f) y’+y’?+(y)?=0
g) Y=y h) y’=y't+1
i) y"=4y'y° j) y'+0’-1) y'+y=0
k) y’=—4/y° l) yy" +(yP +1=0
m) y’=[1+()7]*” n) y"=[1+(y)?]'”
9. Show that the only solution of
y’+ty'+(1+2t?) y?=0
which is tangent to the ¢-axis at some point (¢g, 0) is the identically zero solution.
w_\y (t20; —o<y<oo)

0 (t<0; —wo<y<o)
a) Is the function
10-5
afd
Sy
(t<0)
a solution on — 0 <t<a?
b) Is #(¢) continuous everywhere?
c) Is #’(t) continuous everywhere?
d) Can one apply Theorem | to obtain the existence of a unique solution wy such
that y(0)=1, w’(0)=1? Explain fully.
CHAPTER 3
Linear Differential Equations
3.1 INTRODUCTION
In this chapter we study linear differential equations. As we have already

seen in Chapters | and 2, such equations arise naturally as mathematical
models for some physical systems. Students familiar with linear algebra can
proceed directly to Chapter 4, where more general results are obtained.
We shall concentrate on second-order equations of the form
Ao(t) y’ +4, (t) +a,(t)y=f(t) (3.1)

where do, 4;, 42, f are given functions continuous on some interval J; the
interval J may be open, closed, or open at one end and closed at the other.
We shall see that all the results concerning (3.1) can readily be extended
to linear equations of order higher than 2.
Equations of the form (3.1) occur in many applications. For example,
the simplest (and least accurate) mathematical model for the simple
pendulum is of the form of Eq. (3.1) with ag(t)=1, a,(t)=0, a,(t)=g/L,
J (t)=0 (see Eq. 2.11, Section 2.2). Many physical problems, such as the
motion of a pendulum, a “‘mass-—spring”’ system, and the oscillations in the
shaft of an electric motor, have equations such as (3.1) as their crudest
mathematical models. By this we mean that in most instances the
mathematical model may well be more complicated ; for example, in the case
of the simple pendulum, the derivation originally led to a nonlinear
differential equation (Eq. 2.8, Section 2.2). In such cases one naturally tries
to see whether the relevant equation may be simplified in such a way that the
new approximating equation can actually be solved. This process usually
involves “‘linearizing”’ the equation. In the case of the pendulum equation,
we accomplish this by replacing sin@ by @ in the equation. Naturally we
hope that for “‘small oscillations” this approximation is good enough to
predict the nature of the motion.
65
66 Linear Differential Equations
It certainly is not obvious at this stage that the linearized equation will
be any simpler to handle than the original one. However, experience will
show that linear equations are relatively easy to handle, while nonlinear ones
usually present serious difficulties.
If one linearizes a problem (for the simple pendulum this means replacing
sin @ by 0 in the equation), the following question arises naturally: How good
an approximation does the linearized equation actually produce? For the
pendulum, we would like to prove that in some sense the motions of the
linear and nonlinear models are close to each other when |@| is “small.”
We can hope to answer such questions only much later (see Section 8.4 for
a treatment of the nonlinear simple pendulum with damping). However, the
material presented here is an essential first step—before we can ask how
good an approximation the linearized equation produces, we must be able
to solve this linearized equation.
Before beginning the study of the general theory of Eq. (3.1), we recall
that we already know something about this equation. Namely, as an applica-
tion of the fundamental existence and uniqueness theorem for second-
order equations (Theorem 1, Section 2.5; see also Exercise 3, Section 2.5),
we can state the following result.
Theorem I. Let do, a;, 42,f be functions continuous on some interval I, and
let ao(t)#0 for all t in I. Then for each to in I, there exists one and only one
solution (t) of the equation (3.1) satisfying arbitrary prescribed initial con-
ditions $(to)=Yo, &' (to) =Zo- This solution $(t) exists on the whole interval I.
The fact that the solution (ft) of the linear equation (3.1) exists on the
entire interval 7 does not follow from Theorem 1, Section 2.5, but can be
proved separately (see, for example, Exercise 1, Section 8.5). In this chapter,
we shall assume the validity of Theorem | as stated. We may formulate this
in another way. For a linear second-order differential equation a solution
with a given initial displacement and slope exists and is unique for as long
as the coefficients are continuous and the coefficient of the leading term (dq (¢)
in (3.1)) is not zero.
Example 1. Consider the differential equation ty” +(cos 7) y’+[1—1/(t+ 1)] y=20.
Discuss existence and uniqueness of solutions.
Here do(t)=t, a, (t)=cos t, a, (t)=1—1/(t+1), f(t)=2¢ are continuous for all f
except a3(f), which is discontinuous at r= —1; also aj(0)=0. Thus we must dis-
tinguish three cases for the initial time fo: Case (i): tg< —1; Case (ii): —1<to<0;
Case (iii): ty>0. We do not take to=0 or fo= —1 (why?). In case (i), by Theorem 1,
given any fg9<—l1, there exists one and only one solution ¢ of the given equation
satisfying the initial conditions $(to)=¥9, $'(to)=Zo, Where Vo, Zo are arbitrary given
real numbers; this solution # exists on the interval —oo<t<—l1 by the last
statement in Theorem |.
ayil Introduction 67
Exercises
1. Discuss in a similar way the existence and uniqueness problem for cases (ii) and (iii)
of the equation in Example 1.
2. Discuss the existence and uniqueness problem for real solutions of the equation
(1+ 1%)y”’+2ty’+(log |t|) y=cost.
3. Do the same for the equation
doy" +a,y +a,y=f (t)
where do, @;, dz are constants and f(t) is continuous on —0<t<o.
Example 2. Consider
Ao(t) y" +a, (t) y +ay(t) y=, (3.2)

where do, @;, 4, are continuous on some interval J and ay(t)40 on J. Show that
$(t)=0 is the only solution satisfying the initial conditions #(to)=0, '(to)=0, where
to isin J.
It is readily verified that the function ¢ defined by ¢(t)=0 for all ¢ in / is a solution
of this initial value problem. Therefore, by Theorem | (here f(¢)=0), f(t)=0 is the only
solution on Eq. (3.2) on J satisfying the initial conditions (to)=0, $'(to)=0 for any
ty in J.
Exercises
4. Show that, if solutions ¢ of Eq. (3.2) are represented as curves in the (¢, y) plane, no
solution of(3.2) except #(t)=0 can be tangent to the ¢ axis at any point of /. [Hint:
Study Example 2. ]
5. For each of the following differential equations, determine the largest intervals on
which a unique solution is certain to exist by application of Theorem 1. In each case,
it is assumed that you are given initial conditions of the form $(to)=Vo, $' (to) =Zo
with f9 arbitrary. Note that the interval to be determined may depend on the
choice of fo.
a) ty"+y=2? b) #7(t—3) y’+y"=0
c) y’+./ty=0. d) (l+0*) y’—y'+ty=cost
e) y’—(sind) yy t+y=0 f) vy”—(log |¢|) y=0

For the linear differential equation of order n
do(t) y¥+a,(1)YP ++? +4,-1) ¥+4,() y= f(d), (3.3)

the following analog of Theorem 1, partly a consequence of Theorem 2,
Section 2.5, is valid.
Theorem 2. Let do, a,,..-, G,, f be continuous functions on some interval I

and suppose ao(t)#0 for all t in I. Then, for each tg in I, there exists one and
only one solution $(t) of Eq. (3.3) which satisfies arbitrary prescribed initial
conditions
os Linear Differential Equations
es ag
O(to)=21. = 3. ---. OP”
=22. H" (to)
— P'(Fo) (Co) =M-
The solution @(t) exists on the entire interval I.
As for Theorem 1, the fact that solutions exist on the whole interval J
does not follow from Theorem 2, Section 2.5 (see Exercise 2, Section 8.5).
e
Exercise
6. Apply Theorem 2 to determine the largest intervals on which the existence of a

unique solution 6(f) is assumed when initial conditions of the form
O(to)=2;. @ (to) =2,---. o° Y(t.)=<,
are given: distinguish different values of fp if necessary.
a) y"+(cos 4) y'+(1—P) y=e' b) »"+(cos 4)» +(1—P) y=tant

c) n+y=e ‘+cos1 d) mn +y=secr
3.2 LINEARITY
To develop the theory of linear differential equations, such as (3.1), it is

convenient to introduce the operator L defined by the relation
L(y) ()=40 (2) ¥ (+41 (9 ¥ (+42(0) 9) (3.4)

which we denote briefly by L(y), where L(y)=agy"+a,y'+a,y. Here we
think of L(y) (*) as the value of the function L(y) at the point 7. Noticing
that L(y) is precisely the left-hand side of Eq. (3.1), we may write the
equation simply as
Loy)=f (3.5)
where it is understood that all functions are functions of r.
An operator is, roughly speaking, a function applied to functions. In the
present case, the operator Z is a rule which assigns to each twice differentiable
function y on some interval / the function L(y), where L(y) (t)=ao(t) (2)
+a, (t) ¥'(t)+a2(2) v(2).
The operator LZ is a particular example of a class of operators called
linear operators: An operator T defined on a collection S of functions is said
to be linear ifand only iffor any two functions y, and y, in the collection S and
for any constants c, and c, one has
T (cy¥y +¢2¥2)=¢,T (¥1)+¢2T (2).
It is easy to verify that our operator Z defined by (3.4) is linear. To see this,
let S be the collection of twice differentiable functions defined on the interval
I. Then if y, and y, are any two functions in S and c, and c, are any
two constants, L(¢,¥y +¢2¥2)=@o(Cy¥; +C2¥2)" +4, (Cy ¥; +e2V2)' +42 (C1);
3.2 Linearity 69
+ ¢22)=¢,L(¥,)+cL(y2) by elementary facts about differentiation (which

ones
?).
Exercises
1. Show that the operator T defined by T(y) (t)={y(s) ds, for any function y con-
tinuous on a</<b, is a linear operator.
2. Give other examples of linear operators.
3. Show that the operator T defined by T(¥)=(y’)’, for any function y differentiable
one some interval /, is not linear.
We shall need some more terminology before proceeding to the theory
of linear differential equations. If the function f#0 on J, we say that
Eq. (3.5) is nonhomogeneous (with nonhomogeneous term f). With every
nonhomogeneous linear differential equation of the form (3.5) we associate
the homogeneous (or reduced) linear differential equation L(y)=0 obtained
from (3.5) by replacingf by the zero function.
We now give two basic properties of solutions of linear differential
equations; these are immediate consequences of the linearity of the operator
j
ot
i) If , and , are any two solutions of the homogeneous linear differential
equation L(y)=0 on some interval I, then for any constants c, and c, the
function c,$1 +2 (called a linear combination of $, and $,) is also a solution
of L(y)=0 on I.
To see this we merely compute: L(c,, +¢.6 )=c,L(¢1)+c,L(¢2), by
the linearity of L. Since ¢, and ¢, are solutions of L(y)=0 on J,
L(¢,)=L(¢2)=0 for every t on J, and therefore L(c,¢,+c,¢,)=0.
Thus c,¢,+0¢2¢; is a solution of L(y)=0. I
Exercise
4. Use mathematical induction and the above result to establish the analog of
property (i) for m solutions ¢,(d),...,¢,(t) of L(yv)=0; that is, show that if
1, D25--+, Pm are m solutions of L(y)=0 on J and if cy, C2,..., Gy are any con-
stants, then cy, +¢o6)+-+:-+Cnm is a solution of L(y)=0 on J.
This result is usually expressed by saying that any linear combination
of solutions of L(y)=0 is again a solution of L(y)=0. It is sometimes called
the principle of superposition of solutions. Our object in the next section
will be to show that the problem of solving the equation L(y)=0 can be
reduced to the problem of finding certain special solutions of L(y)=0 and
obtaining all other solutions as linear combinations of these special so-
lutions.
Another important consequence of the linearity of the operator L is the
following.
ii) If @ and w are any two solutions of the nonhomogeneous linear differ-
ential equation L(y)=f on some interval I, then 6—w is a solution of the
corresponding homogeneous equation L(yv)=0.
To see this, we merely compute L(¢—wW). By the linearity of L we have
L(g@—w)=L(¢)—L(y), for tin J. But ¢ and y are solutions of L(v)=fon J.
Therefore L(@—W)=f—f=0 for ¢ in J, which proves the result. I ‘
This result shows that it is only necessary to find one solution of the
equation L(v)=f, provided that one knows all solutions of L(v)=0. This is
because every other solution of the nonhomogeneous equation (3.5) differs
from the known one by some solution of the homogeneous equation L(v)=0.
Exercises
5. Given that wis a solution of L(v)=0 and v isa solution of L(y)=/fon some interval
T, show that u+v is a solution of L(v)=f on J.
6. Suppose f can be written as the sum of m functions /;,..., f,5 that is, {()=f;(0)
+f3(t)+---+/,(0. for ¢ on some interval 7. Suppose that uv, is a solution of the
linear equation L(v)=/,, v2 is a solution of the linear equation L(v)=/>, and in
general u; is a solution of the linear equation L(y)=/; on / for i=1,..., m. Show
that the function v=u,+u,+---+u,, is a solution of L(y)=/ on J. (This result,
also called the principle of superposition, enables us to decompose the problem of
solving L(v)=/ into simpler problems in certain cases.)
Before closing this section we repeat that the only property of the oper-
ator L is used above is linearity. Therefore our results are much more
general than appears to be the case. In particular, if we define the linear
differential operator L, of order by the relation
L,(y) (t)=ao(t) v (t)+.a,(t) VP") (4+ +4,-1(0) ¥ (O+a,(d) (0
where y is any function which is 7 times differentiable on some interval /,
and the functions a;(j=0, 1,..., 7) are continuous on J, do(t)#0 on J, then
4
all results stated in Section 3.2 hold.
Exercise
7. Formulate and verify the analogs of the linearity properties (i) and (ii) for the
equation L,(y)=/ for n=1, 3,4, and » an arbitrary positive integer.
3.3 LINEAR HOMOGENEOUS EQUATIONS
In this section we go far beyond the result established above, that any
linear combination of solutions of the linear homogeneous differential
equation L(y)=0 is again a solution of L(v)=0. We will show that
every solution of L(y)=0 is a linear combination of certain special solu-
tions. Then in Section 3 3.7 we will show how to use the special solutions
3.3 Linear Homogeneous Equations 71
to find every soluuon of the nonhomogeneous equation L(y)=f, using

the linearity property ii) established in the previous section.
Before we can do this we need the important concept of linear depend-
ence.
Definition. We say that m functions g1, 92... Gm are linearly dependent
on an interval I ifand only if there exist constants b,, b,..., bj,, not all zero,
such that
bigh (t)+b2g2(t)+--: + DinGm(t) =0
for every ton I. We say further that the m functions are linearly independent on
I if they are not linearly dependent on I.
Example 1. Show that the functions sin*7, cos”, 1 are linearly dependent on any
interval.
Since sin’ t+cos*t—1=0 for every t, we merely put g, ()=sin* ¢, g>(t)=cos?t,
g3(t)=1, b; =b,=1, b; = — 1 in the above definition. This proves the linear dependence
of the given functions.
Example 2. Show that the functions e, e’', where r,, r, are real constants, are
linearly independent on any interval / provided that r; ¥rp.
To see this, we suppose that there exist constants b,, b, such that b,e"'+6,e"'=0
for all ¢ in 7. Multiplying by e~™! we obtain 6, +b,e"2~"'=0 for all ¢ in J, and
differentiating both sides of this equation with respect to ¢, we obtain b(r,—r,) e?-"'
=0 for all tin J. Since r; #r, and e”?~"” is never zero, this implies that b, must be zero.
However, then ),e"''+5,e'=0 for all ¢ in /, implies b,e""'=0 for all ¢ in J, and hence
b, must also be zero. Since b, and b, are both zero, e”’ and e”' must be linearly
independent.
Exercises
1. Establish the linear independence of the following sets of functions on the intervals
indicated.
a) sin ¢, cos f on any interval /.
b) e”"", e', e on any interval / if r;, r2, r3 are all different.
c) e”", te’ on any interval /.
d) 1, ¢, 27, 2 on any interval J.
e) 7, ton —1<¢<1 but not on 0<t<Il.
f) The functionsf, (¢), f2(t) on —1<t<1, where
co ioo}
LO0= (1 pant
hO=), ie.
n=0 n=0
2. Prove that the functionsf, g are linearly dependent on / if and only if there exists a
constant c such that either f(t)=cg(t) or g(t)=c/(t) for every ¢ in J.
3. Decide which of the following sets of functions are linearly dependent and which
are linearly independent on the given interval. Justify your answer in each case.
4) Oo,(=e "96; (=e s Va woeice

b) o,(t)=e", ,(t)h=e, —-20<t<o
c) i()= Jt, g2(t)=t, O<t<a
d) ¢,()=1, g2(t)=e, 3(t)=e', —w<t<a
) d=, (j= sint, —=lsrsl
0 0 0
f) oi (t)=1 UE cay 65(=\anco —wo<t<-—l
g) oi (t)=1, f2(t)= Tan $3(t)= cers oo pen
h) $,(t)=07, ej=0, d= "Har. ==tara1
More generally, we have the following result, which will be useful on

several occasions later.
Lemma 1. Then functions
Cte ke—1 wee,pit
ale
COME 2 hrk2-1 eee, t aes
il Vit Tatyana
(EI Sec e Ag?
erst te’s! A ga tks—1 erst

>
where ki +k,+...+k,=n and where r,,1r2,...,1r, are distinct numbers, are

linearly independent on every interval I.
Since the proof of this theorem is technically rather complicated, the

reader is advised not to get involved in the details. He should be sure he under-
stands the statement of the Lemma. Note that Example 2 is a special case
With) =K>=1,n=—2.
Proof. The proof is an extension of the argument used in Example 2 above.

Suppose the 7 functions are linearly dependent on some interval J. Then
there exist nm constants a;;,i=1, 2,...,5,/=0, 1,...,k;—1, not all zero,
such that
of t ; te” t 1 tk 3 lor al e"2 a te’2 +
A109 ; skal ; e Rite A . 20 21 a
ap, Ko
le —1 e pro Lal socelayg) e rs ee] le Vs Se rf ks lo Fst —_
‘_Q
or, more compactly,

Ss
ys (a;oe"" +a;,te™" teks ah P= ee Fert) = 0)

i=1
she Linear Homogeneous Equations 73
for all ¢ in 7. We may define the polynomials
P;(t)=ajo t+ajyt+ ee FOR oe (i= ee s)
to write this condition in the form
P(t) e+ P3(t) ev +--+ Pathe" =0 (3.6)
for all ¢ in 7. Since, by assumption, the constants a;; are not all zero, at
least one of the polynomials P(t) is not identically zero. It is convenient
to assume that P,(t)#0; we can always arrange this by a suitable labeling
of the numbers r,, r3,...,r,. Now we divide Eq. (3.6) by e* and differ-
entiate at most A, times until the first term drops out. Note that all terms
in (3.6) can be differentiated as often as we wish. Then we have an equation
of the form
Q2(t) +. Q3(t) ef +.-- +0,(2)ef)=0 (67)

for every ¢ in J. The term Q,(t) e“'~" in (3.7) is obtained by differentiating
P;(t) e'~' (i=2,..., s), as often as necessary to remove the first term P, (¢).
Note that differentiation of a polynomial multiplied by an exponential gives
a polynomial of the same degree multiplied by the same exponential (think
of the rule for differentiation of products). Thus the polynomial Q, in
(3.7) has the same degree as P,, and does not vanish identically. We continue
this procedure, dividing by the exponential in the first term and then differ-
entiating often enough to remove the first term, until we are left with only
one term. Then we have an equation of the form
R,(t) Css 1) —()
in which the polynomial R, has the same degree as P,, and does not vanish
identically. However, the exponential term in this equation does not vanish,
and we have a contradiction. This shows that all the constants a;; must
be zero, and therefore that the n given functions are linearly independent
on/. |
Exercise
4. To which of the sets of functions in Exercises | and 3 could you apply Lemma | to
deduce either linear dependence or linear independence?
The above discussion of linear dependence and independence of func-

tions has not been, up to this point, related to the differential equation
L(y)=0. Before continuing, review Theorem | and the notion of linearity
as given in Section 3.2. Using these, we now establish one of the key results
of the theory of linear differential equations.
Theorem 1. Let ao, 4,, ay be functions continuous on some interval I and

let ao(t)40 for all t on I. Then the differential equation
L(y)=ao(t) y" +4, (t) Y+.a2(t) y=0

has two linearly independent solutions @,, @, on I. Moreover, if @ is any
solution of L(y)=0 on I, then it is possible to findaunique pair of constants c,, €2
such that for every t on I
P(t)=cr hi (t)+ C22 (2).

Proof. Let tp be any point of the interval 7. By Theorem 1, Section 3.1,
there exists a unique solution ¢, on J of L(y)=0 satisfying the special initial
conditions #,(to)=1, 1 (to) =0. Similarly, there exists on J a unique so-
lution @, of L(y)=0 such that $5 (to)=0, $5(to)=1. We select these par-
ticular solutions because it will be easy to prove that they are linearly in-
dependent on /. You will see later, after studying the proof, that many
other choices are possible (see Exercises 7 and 8).
We claim first that the solutions ¢, and @, are linearly independent
on J. Suppose there exist constants b,, b, such that
bi, (t)+b2$2(t)=0 (3.8)
for every ¢ on J. Since ¢,, @, are solutions of L(y)=0 on J, they are dif-
ferentiable on J and hence from (3.8) we have also
bo) (t)+b265(t)=0 (3.9)

for every ¢ on /. In particular, putting t=¢) in (3.8) and (3.9), we obtain
respectively from the chosen initial conditions
b,-1+b,:0=0, by0+b,-1=—0,
and we therefore conclude that b, =b,=0, which shows that the solutions
&,, 2 cannot be linearly dependent on / and therefore this proves their
linear independence on /.
To complete the proof of the theorem, let ¢ be any solution of L(y)=0
on / and calculate $(to)=a, '(to)=f. (That is, we evaluate p(t) and ¢'(¢)
at f=fg and call the values at fo, « and f, respectively.) If there are to exist
constants c, and c, such that $(t)=c,¢, (t)+c.¢>(¢) for all tin J, this relation
must hold in particular at tp, and we must have
a=(to)=C1P1
(to) +C262 (to) =e1° 1+c2:°0=c,,
B=@'
(to) =C1 9% (to) + C203 (to) = C1 O+C2*1= ep.
Define the function w by the relation w(t)=a¢@, ()+ Bo2(t) for tin J. Clearly
(by the linearity property (i), Section 3.2), y is a solution of L(y)=0 on J;
moreover,
3.3 Linear Homogeneous Equations Th
W (to) =a, (to) + Boba (to) =a 1+ 8-0=a,

W' (to) =2' (to) + BO3 (to) =a: 0+ B 1=B.
Therefore @ and y are both solutions of L(y)=0 on J which satisfy the
same pair of initial conditions at fp. Since, by Theorem 1, Section 3.1,
there is only one such solution, we conclude that $(t)=() =a, (t)+
Bo (t)
on J, which completes the proof.
Exercises
5. Why are the constants c,, c, in the statement of the theorem unique?
6. Carry out the proof of Theorem | by using the solutions y, and w, of (3.5) on J
satisfying the initial conditions Wy (to)=2, Wj (to)= —1 and (to) = —1, W4 (to) =!
in place of the solutions ¢, and @). [Hint: Begin by showing that the solutions
W1, W2 of (3.5) are linearly independent on /.]
7. Let w, and w, be solutions of L(y)=0 on / satisfying the initial conditions
Wi(to)=%, Wi (tCo)=B; Wal(to)=y, Wo (to)=4

respectively. Under what conditions on a, f, y, 6 will the solutions w,, w, be linearly
independent on /?
8. Assuming the condition found in Exercise 7 to be satisfied, use the solutions w,
and w, to complete the proof of Theorem 1.
Example 3. Find that solution ¢ of v’+y=0 such that $(0)=1, ¢’(0)=—1, using
the fact that cost and sint are both solutions.
It is easily shown that cost and sin¢ are linearly independent solutions of
y’+y=0 on any interval / (see Exercise la).) To find the desired solution we apply
Theorem 1, letting ¢,(t)=cos ¢, #2 (¢)=sin ¢, and observing that $,(0)=1, 6, (0)=0,
>(0)=0, $5 (0)=1 as in the above proof. By Theorem | we know that there exist unique
constants ¢c;, Cc, such that $(t)=c, cos t+ c, sin f; as we saw in the proof we may deter-
mine c, and c, by imposing the initial conditions. Thus we obtain
o(0)=1=c,-1+c,°0
¢'(0)= —1=—c,:0+c,°1
Therefore c; =1, c= —1 and the desired solution ¢ is @(t)=cos ¢—sin ft.
Exercise
9. State and prove a theorem analogous to Theorem | for the linear third-order
differential equation
L3(y)=ao(t) ¥" +4, (t) y” =ay(t) y’ +a3(t) y=0,

where do, @;, @>, 43 are continuous on some interval J and a(t)40 on J. [Hint:
For any f 9 on J let ¢, be that solution of L3(y)=0 for which ¢,(to)=1,
$1 (to) =0, Pf (to)=0, let p2 be that solution of L3(y)=0 for which #3 (to)=0,
5(to9)=0, let @; be that solution of L3(y)=0 for which $3(to)=0,

4(t9)=1 and now proceed as in the proof of Theorem 1.]
For the general case of linear differential equations of order n,
L,(y)=4o(t) y +4;(t) YP +--+ +4,-1(0) ¥ +4,(t) y=0,

Theorem | has the following analog:
Theorem 2. Let do, a1,.-., 4, be continuous functions on some interval I,
and suppose ag(t)#0 on I. Then the differential equation L,,(y)=0 has n linearly
independent solutions @,, >,-.-,P, on I. Moreover, if @ is a solution of
L,,(v)=0 on I, then there exist uniquely determined constants C1, C2,..-5 Cy
such that
P(t)=CrG1(t)+Crh2
(+--+ Cnbn(t)
for every t in I.
Exercise
10. Prove Theorem 2.
In practice it is undesirable to restrict ourselves to solutions ¢,, ¢)

which satisfy special initial conditions such as @, (to) = 1, $4 (to) =0, 62 (to) =0,
5(to)=1 at some f, in J. We shall show shortly that instead of the special
solutions @,, ¢2 used, any two linearly independent solutions of L(y)=0
on 7 will serve the purpose just as well. To see this we can use the result of
Exercises 7 and 8. Alternatively, it is convenient to introduce the concept
of the Wronskian, which, as we shall see, also serves another purpose.
Definition. Let f,, f., be any two differentiable functions on some interval I.
Then the determinant
Ws
JiaRites
fr)= ye Oe
is called the Wronskian of f,; and f. Its value at any t in I will be denoted by
W(f,, f2) (t). More generally, iff,,...,f,,aren functions which are n—1 times
differentiable on I, then the nth-order determinant
peer ape:
{0 mekpiene oni,
Whos =[
fe fee feiss pon
is called the Wronskian off,,...,fy.

33 Linear Homogeneous Equations HT.
Exercise
11. Evaluate the Wronskian of the following functions
a) f,(¢)=sin ¢ io) =cost (—20<t<o)

b) fi:(=e, hi=e", (—20<t<o)
ce) A(Q=t? A ()=le| (— 00<t<oo)
d) f,(d)=1 exc | E(d=F (—10<t<o)
The Wronskian of two solutions of L(y)=0 on J provides us with the

following simple test of their linear independence.
Theorem 3. Let ao, a1, ay be given functions continuous on some interval I,
and let ao(t)#0
for all t on I. Then two solutions $,, 62 of
L(y)=ao(t) y" +4, (t) ¥ +ay(t) y=0

are linearly independent on I if and only if W(,, $2) (t)#0 for all t on I.
Before proving this result we give an illustration.
Example 4. Show that cos¢ and sint¢ are linearly independent solutions of y"+y=0
for —c <i<oo.
The functions @, (t)=cos ¢, $2(t)=sin ¢ are solutions of y’+y=0 on —2a0<t<oo.
To test their linear independence we compute their Wronskian
cost sint
W
(cost, sin =| l —wo<t<o.
—sint cost
Therefore, by Theorem 3, ¢, (t)=cos ¢, #2(t)=sin ¢ are linearly independent solutions
of y’+y=00n —0 <t<o. Of course, we already know this result from having applied
the definition of linear independence directly. However, when dealing with solutions of a
linear homogeneous equation L(y)=0, the theorem is often easier to use than the
definition.
Warning. Do not apply Theorem 3 when the functions being tested for linear
independence are not known to be solutions of a linear homogeneous equation
L(y)=0. To see why, consider the functionsf; (t)=0’, f;(t)=t\¢| and take for
I] the interval — 1<1t<1. Then as we saw in Exercise 11 (c), the functions f,, /5
are linearly independent on J and yet W(f,,f2)(t)=0 for every ¢ on
—l<r<l.
Proof of Theorem 3. The proof consists of two parts. Suppose first that
the solutions ¢,(t), 62(t) of L(y)=0 are such that W(¢,, 2) (t)40 for all
ton J and yet ¢,, ¢, are linearly dependent on /. Then by the definition of
linear dependence there exist constants b,, b, not both zero such that
b,o,(t)+b262(t)=0 forallton/ (3.10)

and also
bo (t)+b.65(t)=0 foralltonJ (why?) (310)
For each fixed ¢ on J, Eqs. (3.10) and (3.11) are linear homogeneous algebraic
equations satisfied by b, and b,, and the determinant of their coefficients
is precisely W(d,, $2) (¢). Since, by assumption, W(,, ¢2) (t)40 at any
ton J, it follows from the theory of linear homogeneous systems of algebraic
equations (see Appendix 1) that b, =b,=0, which contradicts the assumed
linear dependence ofthe solutions ¢,, ¢, on J. This shows that if the Wrons-
kian of two solutions of L(y)=0 is different from zero on /, then these so-
lutions are linearly independent on J.
To prove the second part of the theorem, assume that the solutions
;, 62 of L(y)=0 are linearly independent on J and assume that there is
at least one ? on J such that W(¢,, $2) (2)=0. (If there is no such ? there
is nothing to prove!) Now look again at the algebraic system (3.10), (3.11)
for t=?. It follows, again from the theory of linear homogeneous systems
of algebraic equations (see Appendix 1) that, because W(¢,, ¢,) (?)=0, the
system of algebraic equations
bb, (2) +b2$2(f)=0, bi) (2) +b2$45 (7) =0 (3.12)
has at least one solution b,, b,, where b, and 5, are not both zero. To complete
the proof define the function W(t)=6,¢,+5¢.(t), where b,, 6, are taken
as any solution of (3.12). First observe that is a solution of L(y)=0 (why?).
Because of (3.12) the solution w satisfies the initial conditions y(?)=0,
w’(?)=0. Therefore, by Theorem | and Example 2, Section 3.1, w(t)=0 for
every ¢ on J. This means that we have found constants 6,, b, not both zero
such that b, ¢, (t)+52¢, (t)=0 for every t on J. This contradicts the assumed
linear independence of the solutions ¢,, ¢, on J. Therefore the assumption
W (1, $2) (#)=0 is false; that is, no such ? exists and W(d,, ¢) (t)40 for
every ¢ in J. This completes the proof of Theorem 3. I
Exercises
12. Show that e”', e~ *‘ are linearly independent solutions of y’—4y=0 on — 10 <t<oo.
13. Show that e-‘? cos(,/3/2) (ew? sin(./3/2) t are linearly independent solutions
of y’+y’+y=0 on —wo<t<oo.
14. Show that e', te~' are linearly independent solutions of y’+2y’+y=0 on
—0<t<o.
15. Show that sin ?*, cos /? are linearly independent solutions of ty”—y'+4ey=0 on
0<t<o or —«0<t<0. Show that W(sin 77, cos t7) (0)=0. Why does this fact
not contradict Theorem 3?
16. State the analog of Theorem 3 for the nth-order equation L,(y)=ao(t)
+a, (t) yy" +---+a,_,(t)y' +a, (1) y=0.
We can now establish a result which says that for any two solutions ¢,
and @, of a linear homogeneous second-order equation with continuous
coefficients, the Wronskian is either identically zero or never equal to zero.
Theorem 4. Let the hypothesis of Theorem 3 be satisfied on some interval I.
Let (,. 2 be two solutions of L(y)=0 on I. Then either their Wronskian
W(,, 2) (0) is zero for every t in I or it is different from zero for every t in I.
The proof of Theorem 4 is outlined in the following three exercises.
Exercises
17. Let @,, 2 be two solutions on some interval J of L(v)=do(t) vy"+4,(1) y' +.a,(t) y
=0, where do, a,, a, are continuous on J and a,(t)#0 on J. Show that the
Wronskian W(@,, ¢2) (¢) satisfies the first-order linear differential equation
Wn W, (tind). (*)
Pi (t) $2(1)
[Hint :W'(9,, 62) {t)= = (0162-9162)
=9192 — Pia -
Pi(t) $5(t) ,
Now use the fact that ¢,, @, are solutions of L(y)=0 on /to replace ${, 3 by terms
involving $,, $;, $2, 65. Collect terms to obtain (*).]
18. By solving (*) in Exercise 17, derive Abel’s formula
t
W (1. 62) (t)=W(d1, 2) (to) exp(- |a is).

to
19. Use the result of Exercise 18 to prove Theorem 4.

20. State and prove the analog of Theorem 4 for the linear third-order differential
equation
L3(y)=4o(t) y” +4, (y) y’ +4y(t) ¥ +a43(t) y=0.
21. Show that e', cos ¢, sin ¢ are linearly independent solutions of the differential equa-
tion
y"—y"+y—-y=0 on —OK<f<o.
22. Theorem 4, combined with Theorem 3, provides a convenient method for testing
solutions of linear differential equations for linear independence on some interval.
For, according to these results, it is enough to evaluate the Wronskian at some
conveniently chosen point. Thus, for example, show that
em
6-14), 2:3-5-6---(3m—1) (3m)

m 1
{c.9)
p3m+1
p2(t)=t +) 3-4-6-7---(3m) (3m+1)
m=1
are linearly independent solutions of y’—ty=0 on the interval —co<t<oo.

(Here you may assume that it has already been shown that ¢, and @, are
solutions of y’—ty=0, but how could you verify this?)
Recall that the linearity of L implies that any linear combination of

solutions of L(y)=0, is again a solution. We have raised the question:
“Can every solution of L(y)=0 be generated as a linear combination of
some special solution?’? We answered this partially in Theorem | using a
particular pair of linearly independent solutions. With the help of Theorem 3
we can now answer the question completely.
Theorem 5. Let ao, 4,, 4) be functions continuous on some interval I, and
let ag(t)#0 for all tin I. If p, and 2 are any two linearly independent solutions
of
L(y)=do(t) y" +4,(t) y +a2(t) y=0
on I (not necessarily the two special solutions @,, 6 of Theorem 1), then every
solution @ of L(y)=0 on I can be written in the form
b()\=cihi(t)terpr(t) tin (3.13)

for some unique choice of constants ¢c,, C.
From a practical point of view, the theorem tells us that knowledge
(possibly by guessing) of any two linearly independent solutions ¢,, ¢, of
L(y)=0 on J enables us to express every solution by means of Eq. (3.13) by
choosing the constants c,, c, suitably. For this reason, we call the function
defined by (3.13) the general solution of L(y)=0 on J, and we sometimes say
that the linearly independent solutions form a fundamental set.
Proof of Theorem 5. Let @ be any a solution of L(yv)=0 on / and let fy be
any point in J. Compute ¢(t))=«, d'(to)=f. Because ¢, and @, are linearly
independent solutions of L(y)=0, Theorem 4 tells us that W(d,, $3) (t)#0
for all ¢ on /; in particular W(@,, $2) (to) #0. If the representation (3.13)
holds for all ¢ in J, it will have to hold at t=fp. To see if this is possible, we
impose the conditions #(to)=«, ' (to) =f and obtain the system of algebraic
equations
C14 (to) +C2$2(to) =a, C14 (to) +23 (to) =B

with determinant of coefficients W(,, 2) (to) #0. Therefore by the theory
of linear nonhomogeneous systems of algebraic equations (see Appendix 1)
this algebraic system can be solved uniquely for c,, c,, and we obtain
_ 42 (to) — Bb2 (to) 3 _ BG 1 (to) = 24 (to)

(3.14)
; W ($1, 2) (to) ” ; W ($1, 2) (to) -
This choice of c,, c, makes (3.13) hold at t=¢9. To see whether this choice of
C1, C2 does the job for all ¢ in J, we define the function
W(th=c1o (t)t+er¢2(t),
where c;, c, are the numbers given by (3.14). We observe that w(t) (as well
as f(t), 1 (2), $2(t) is a solution of L(y)=0 on J. To complete the proof
we need only show that wy (t)= (1) for every t in /. But using (3.14) we see that
W(to)=a=(to) and W'(to) =B=¢'

(to).
Therefore ¢ and yw are both solutions of L(y)=0 on J and they satisfy the
same initial conditions at t=f9. By uniqueness (Theorem 1, Section 3.1),
@ and w are identical and this establishes Theorem 5. |
Theorem 5 extends easily to higher-order linear differential equations
as follows:
Theorem 6. Let dao, a1,...,a, be functions continuous on some interval I,
and let ag(t)#0 on I. Let $,, b2,.-., 6, be any set ofn linearly independent
solutions of the equation
L,(y)=ao(t) Y +4,(t)YO +---+4,-1(t) ¥ +4,(t) y=,

on I. Then every solution $ of L,(y)=0 on I can be written as a unique linear
combination
P()=Cibi(Qt- +nn(t)
for t on I of the given solutions $,, $2,..-, Py (1-€., there exist unique constants
Gis Coie. C, SUCK that
P(t)=C1
1 (t) +CrG2(t) +> $end, (t)-
Exercise
23. Prove Theorem 6 if n=3.
Remark (for students acquainted with linear algebra). The theory developed
in Sections 3.2 and 3.3 shows that the solutions of a linear homogeneous
differential equation L(y)=0 with continuous coefficients on some interval J
and with nonvanishing leading coefficient on J, form a vector space V
over the
real or complex numbers (see property (i), Section 3.2). Theorem 1 shows
that the dimension of V is 2 if L is a linear differential operator of order 2, by
exhibiting a basis for V consisting of the special linearly independent solutions
@, and ¢, constructed in the theorem. Theorem 5 shows that any two linearly
independent solutions of L(y)=0 also form a basis for V, provided the order
of L is 2. We can derive this more simply using knowledge of linear algebra.
Once we know, by Theorem 1, that V has dimension 2, it follows immediately
that any two linearly independent vectors in V (that is, solutions) span V.
Theorem | for a homogeneous linear differential equation of order n shows
that for such an equation the vector space of solutions has dimension n.
For a more general discussion of this topic we refer the reader to Chapter 4.
Example 5. Use the functions e?‘ and e *' to find the general solution ofthe equation
y”—4y=0; then find that solution ¢ for which $(0)=1, $'(0)=0.
The functions e?', e~*! are solutions of y’ —4y=0 for all ¢ (why?). They are linearly
independent solutions of L(y)=y’—4y=0 on —co<t<oo (why?). Therefore, by
Theorem 5, every solution @ of y’—4y=0 on —c« <t<o can be written in the form
b(t)=cye*+c,e 7
for some unique choice of the constants c,, c>. This is the general solution of y’—4y=0
on —o<tf<oo. To find that solution ¢ of y’—4y=0 on —a<t<oo for which
(0)=1, $'(0)=0, we see from (t)=c,e7"+c,e ~ that c, and c, must satisfy the equa-
tions
Qj tey=—1, 2¢e;—2e,=0
orc; =c, =4. Thus $(t)=(e" +e *')/2=cosh 2¢. The reader should note that Theorem |
alone does not supply enough information to solve this problem because the solutions
e*' and e 7‘ do not satisfy the right initial conditions. However, we could solve the
problem by using Exercises 7 and 8.
Exercises
24. In each of the following, find the solution ¢ of the given differential equations that
satisfies the specified initial conditions. Also find the general solution in each case.
a) y’—4y=0, (0)=2, $'(0)=—1 (see Exercise 12)
b) y’+y’+y=0, ¢(0)=1, G’(0)=3 (See Exercise 13)
c) y’+y'+y=0, $(0)=0, $7(0)=0
d) y’+2y’+y=0, ¢(1)=—-1, @'(1)=1 (see Exercise 14)
e) y’—ty=0, $(0)=0, $'(0)=2 (see Exercise 22)
f) y"—y"+y'-y=0, (0)=2, $'(0)=—-1, P"(0)=1 (see Exercise 21)
25. Show that for any constant «>0, sin wt and cos af are linearly independent solu-
tions on —co<t<oo of y’+a7y=0. What is the general solution? Find all
solutions which pass through the point (z/4a, 5). Which one of these has slope «
at the point (2/4«, 5)? Which has slope —« at the point (z/4«, 5)? Letting ¢, and
#2 represent the solutions of slope « and —« respectively at the point (7/4, 5),
decide whether @, and @, are linearly independent on — co <t<oo.
*26. In each of the following, let @, (1) and @,(¢) be solutions of the differential equation
L(v)=y"
+p(t) y +q(t) y=0
on some interval /.
a) If by (to)=$2(to)=0 for some fy in J, show that ¢, and ¢, cannot form a
fundamental set of solutions on /.
3.4 Linear Homogeneous Differential Equations 83
b) If ¢, and ¢, both have a maximum or a minimum at some point fp in J, show

that @, and @, cannot form a fundamental set of solutions on J.
c) Let @, and @, form a fundamental set of solutions on J which both have an in-
flection point at some point ¢, of /. Show that p(t,)=q(t,)=0.
d) Let @; and ¢, forma fundamental set of solutions on /. Show that wy, =¢,+ 5,
W.=,—2¢, also form a fundamental set of solutions on /.
3.4 SOLUTION OF LINEAR HOMOGENEOUS DIFFERENTIAL

EQUATIONS OF SECOND ORDER WITH CONSTANT COEFFICIENTS
In this section we shall learn, with the aid of the theory just developed,
how to solve the linear second-order equation L(y)=0 in certain special
cases. We shall be guided by the fact that in order to find all solutions of
L(y)=0, we merely need to find two linearly independent solutions and
then apply Theorem 5, Section 3.3. Subsequently, we shall study linear
homogeneous equations of higher order. We shall begin with the simplest
case of constant coefficients and then consider in the following sections
several more complicated cases of variable coefficients. We note that in the
general case of continuous coefficients a solution of L(y)=0 exists but
cannot necessarily be found in terms of elementary functions. Fortunately,
large numbers of interesting physical problems lead to mathematical models
which, when simplified sufficiently, fall into categories which we can handle
easily.
Important examples of models leading to an equation of the form
L(y)=y"+py +qy=0, 10
<4 <<100)
where p and qg are real nonnegative constants, are the mass—spring system and
the pendulum (see Eq. (2.4), Section 2.1 and also Eq. (2.10), (2.11), Section 2.2,
with sin@ replaced by 0). Another important model is a linear electrical
circuit consisting of a capacitance C, a resistance R, and an inductance L
connected in series. It can be shown that the potential difference (voltage)
v(t) across the capacitance can be reasonably described by the equation
pee pe =)
v 7 Lean
Properties of these models will be discussed in future exercises.

We first solve the linear differential equation
+a, y' +4a2y=9, = OO <GE OO)

L(y)=aoy"
where do, @;, 4) are real constants, dj#0. Thus we may as well divide
through by a) and assume that the equation has the form
L(y)=y"+py +qy=0, —ao<t<o0 (3.15)

where p and gq are real constants, not necessarily positive. Our task is to
find two linearly independent solutions of (3.15). Recall that for the first-
order equation y’+ry=0, where r is a constant, e'" is a solution. In
Section 1.1, we found this solution by separation of variables. However,
we could also find it as follows: If for some constant z, e*' is to be a solution
of y’+ry=0, then we must have (e')’
+re*’=0 or (z+1r) e’=0. Since e” 40,
we see that e”' can be a solution of y’+ry=0 only if z= —r which gives e ”
as a candidate for a solution. Direct verification shows that it 1s.
Let us try to find a solution of (3.15) of the form e** on —w2w<t<o.
Then we must have L(e*)=0. But L(e*)=(e7)’
+p(e”)’ + ge" =(z? + pz
+q) e'. Therefore e” can be a solution of L(y)=0 on — 0 <t<oo only if
(2? +pz+q) e*=0
or, since e’ 40, only if z is a root of the quadratic equation
z?>+pz+q=0. (3.16)
Equation (3.16) is called the characteristic equation or auxiliary equation
associated with (3.15), and z7+pz+q is called the characteristic polynomial
associated with (3.15). The quadratic equation (3.16) has the roots
OR eet
of 2 5) iG —
—p=(p?=4q)"”?
a)
Exercise
1. Verify that e*'', e* are solutions of (3.15).
We now have two possibilities:

i) if p?44q the roots z, and z, are distinct,
lieth pt —4gz 2:
In case (i) this means that e*!' and e*‘(z,; #z>) are two distinct solutions
of the differential equation (3.15) on — 00 <t<oo. The only question which
remains is: Are these solutions linearly independent on — 00 <t< oo? They
are, as was shown directly in Example 2, Section 3.3. Alternatively, since
e*"', e*! are solutions of(3.15), we can establish their linear independence by
Theorem 3, Section 3.3, using the Wronskian. We have
e7it e772!
We" 2 )\= =(zZ,— Ze iia.

Bes aoe
Since z; #2, W(e*", e')#0 and Theorem 3, Section 3.3, gives the desired
linear independence. Therefore, in case (i), by Theorem 5, Section 3.3,
every solution ¢ of Eq. (3.15) has the form
o@j=cye! Hee", Sw <t<00. (3.17)

for some unique choice of the constants c;, ¢2; (3.17) is called the general
solution of(3.15), as explained in the remarks following Theorem 5, Section
328"
Example 1. Find the general solution of y’—9v=0.
By the method just given, the characteristic equation (3.16) is z*—9=0, which has
the roots z,; =3, z,= —3. Since z, #23, the solutions e*', e- *' are linearly independent
on —0o <f<co and therefore the general solution of y’—9y=0 is #(t)=c,e*+c,e*,
where c,, c, are arbitrary constants.
Exercises
2. Find the general solution of each of the following equations, and then the
solution ¢ satisfying the given initial conditions:
a) y’—y=0, $(0)=0, ¢’(0)=1.
b) y’—5y’+6yv=0, $(0)=0, ¢’(0)=1.
c) y”—6y"+ Lly’—6y=0, ¢(0)=¢'(0)=0, 6” (0)=1.
*3. In Eq. (3.15) for damped oscillations (that is, assume p, q>0), find values of p and
q such that the roots of the characteristic equation are real and distinct. For such
values, discuss the asymptotic behavior (behavior as t + 00) of the solutions by
computing lim,...., #(t), where ¢(t) is any solution of(3.15).
The careful reader may have noticed that in case (i), the roots z,, Z, will
be real and distinct if p*>4g, and will be complex conjugate (hence dis-
tinct) if p* <4q. For example, the differential equation
y"+y+y=0
has the characteristic equation z?7+z+1=0, and its roots are
—-14
—— 31 — _ aa
-1-J3i
ot 5) oe ae
Therefore
Pa on —1-./3i
exp (a ) and exp (= )
should be, and in fact are, solutions. However, these functions are complex-
valued functions of the real variable t, and up to this point all functions con-
sidered have been real. If you are unfamiliar with complex-valued functions
of a real variable, read Appendix 3 before proceeding.
In view of the theory of complex-valued functions of a real variable
discussed in Appendix 3, every definition and theorem given for real so-
lutions of the real equation
L(y)=do(t) y’ +4, (t) +42 (1) y=b(),

where dp, @;, @>, and b are real functions defined on some interval /, holds for
complex-valued solutions of this equation. This remains true even if do,
a;, a), and b are complex functions. This specifically applies to existence
and uniqueness of such solutions, and linear dependence and independence
of such solutions (including the Wronskian test). There is no change needed
in any of the statements and their proofs; it is only necessary to bear in mind
that the functions which enter each discussion may be complex-valued.
Exercises
4. Show that the functions

js :
Sj ten Shy) —1—/3i
exp (St ) and exp eae )
satisfy the differential equation y’+y’+y=0 for all real ¢.

5. Reprove Theorem 1, Section 3.3, in the case that the coefficients ao(t), a; (t), a2 (0)
are continuous complex-valued function on an interval J and fs real.
We now present a result on complex-valued solutions of rea/ linear

differential equations which is of great importance in applications. Note
that this result is not restricted to equations with constant coefficients.
Suppose that fis any complex-valued function defined in a real interval
JF. Let #f denote the real part of fand let %f denote the imaginary part;
e.g., if f(t)=exp 2it we have
(# f) (t)=cos 2t and (¥ f) (t)=sin 2r.
Theorem 1. Let @ be a complex-valued solution of the differential equation
L(y)=4o(t) y"” +4; (t) ¥ +a, (t) y=0
on some interval I, where ao, ay, ay are given real functions on I. Then the
real functions u=2o, v=IF are themselves (real) solutions of L(yv)=0
onl.
Proof. Since @ 1s a solution of L(y)=0 on J, we have
ao (t) 6" (t) +41(t)$'()+4p(t) o(t)=0

for every ¢ on J. (The fact that @ may be complex valued does not change
anything.) Since @=u+iv, we have, from the definition of derivative
Q (t)=u' (t)+iv' (1), 6” (t)=u" (1)+iv"(1).Therefore
do(t) [u" (t) + iv" (t)] + a, (t) [u'(t)+ iv’ (t)] + a, (0) [u(t) +iv(0)]=0.
Separating the left-hand side into real and imaginary parts, we obtain
(remember that ao, a, a> are real) for all ¢ on J:
Ag (t) u" (t)+a, (t) u(t) +.ay(t) u(t) +ifao(t) v(t) +a, (t) v(t)+a,(t)v(t)] =0
[Note: This also shows that L(¢)=L(u)+iL(v); this is true in general if
L is a linear differential operator with real coefficients.] Since the last
relation holds for every ¢ on J and since a complex number is zero if and
only if both its real and imaginary parts are zero, we have, for all ¢ in J:
L(u)=ao(t) u" (t) +a, (t) u(t) +a, (t) u(t)=0
and
L(v)=ao(t) v" (t) +a, (t) v'(t)+.a,(t) v(t) =0
which shows that u=2 and v=4¢ are both solutions of L(y)=0 on J
and completes the proof. |
Exercise
6. Let ¢ be a solution on some interval / of the differential equation
L(y)=do(t) y” +4, (t) y'+a2(t) y=b(t)

where dp, d;, a are real and 5 is complex. Show that u=&¢@ satisfies the equation
L(y)=Ab and prove an analogous result for v= 4¢.
Example Z. Use the solutions
$4()=exp (at
2
‘), $a(t)=exp cee )
ofthe differential equation y’ + y’+y=0 and Theorem | to find the general solution in
real form on —c<t<o. They are linearly independent on -—c0<t<o, since, by
Theorem 3, Section 3.3, interpreted for complex-valued solutions,
an (Ss ) ae Cae )
W ($1. $2) ()=
ie 3 exp(—4 a ae
2 2
=——a/3ie) £0, =—O<taoO.
Therefore, by Theorem 5, Section 3.3, interpreted for complex-valued solutions, every

solution ¢ (possibly complex valued) of y’+y’+y=0 on —co<t<oo has the form
(t)=c,1(t)+¢2¢2(0) for some unique choice of the (possibly complex) constants
C1, C. By Theorem | (applicable because the coefficients are real) the real functions
u, (t)=2d; (t)=exp(—t/2) cos (,/3/2) t

and
v; (t)= 4; (t)=exp(—t/2) sin(/3/2) t

88 Linear Differential Equations *
are also solutions of y’+y’+y=0 for —co<t<oo. The same statement applies to
Us (t) =A. (t)=exp(—t/2) cos(\/3/2) t

and
v>(t)=.4£o5(t)= —exp(—t/2) sin(,/3/2) t.
You can easily check that W(u,, v,) (t)}#0 on — 0 <t<oo. Therefore, by Theorem 5,
Section 3.3, again, every solution ¢ of y’+y’+y=0 on —0 <t<o has the form
t 5 ane 3
p(t)=a, exp(-5) cos . t+da, exp (-5)sin Se
for some unique choice of the (possibly complex) constants a,, a2. Starting with the
complex form of the solution ¢, we may also arrive at the “real form” as follows.
Using Euler’s Formula (see Appendix 3) and collecting terms, we have
b(t)=c11
(t)+cr$2(t)
=C; exp(—5) (cos? Spy sin )
+C exp(-5) (cos t ee )
Z 2
t 3 t 3
=(c;+C2) exp(-5) cos t+i(c,—c) exp(—5) sin? ihe
2)
If we now define a; =c, + , a, =i(c; —c), we obtain the desired form. It is clear from
this that the solution (t) of the equation y’+y’+y=0 will be real if and only if
C,=C, (the complex conjugate of c,). In this case, of course, a, and a, are both real.
We now return to the general equation L(y)=0, where p and q are

are real constants, and summarize what we have learned up to this point.
Theorem 2. Every solution of the differential equation
y"+py'+qy=0 (3.15)
where p, q are real constants with p?#4q is defined on —0«<t<o and
has the form
b(t)=cye*+c,e, —ao<t<oo (3.17)
The numbers z,, 22 are the distinct roots of the characteristic equation
z*>+pz+q=0 (3.16)
and C,, Cy are constants. If p? >4q, z, and z, are real and distinct. If p* <4q
the roots Z,, Z, are complex conjugates. In this case if z, =«+ if (a, B real) the
solution @ may be expressed in the form
p(t) =e" (a, cos Bt +a, sin Bt) (3.18)
where ay, a, are constants. If @ is real, a, and ay are real.
Proof. We have already proved all of Theorem 2 except for Eq. (3.18).
To prove (3.18), we proceed exactly as in Exercise 5 above; namely, we
know from Theorem | that e%' cos Pr, e* sinft are solutions of y"+py’+
+qy=0, where «+i is a root of z?+pz+q=0. Since these solutions are
linearly independent on — 00 <r< oo, Eq. (3.18) is a direct consequence of
Theorem 5, Section 3.3. 1
Exercises
7. Show that e* cos fr, e“ sin fr are linearly independent solutions on — 00 <t<oo of
(3.15) when p? <4q.
8. Proceeding as in Example 2, show that ay, a in (3.18) are given in terms of c,
and c, by the formulas a; =c,+¢), a, =i(c; —C3), where c,, c, are the constants in
(3.17).
9. Find the solution @ satisfying the initial conditions #(0)=¢'(0)=1 of each of the
following differential equations:
a) y’+y=0 b) y’—4y'4+ 13y=0
c) v"+4r=0 d) y’+2y’+2y=0
*10. In Eq. (3.15) with p, g nonnegative, find conditions on the constants which lead to
complex roots of the characteristic equation and investigate the behavior of the
solutions for various choices of these constants as f> +00.
Theorem 2 enables us to solve Eq. (3.15) completely when the characteris-

tic equation (3.16) has distinct roots.
We now turn to the case of equal roots of the characteristic equation.
This occurs if p?=4g, and the characteristic equation
z*+pz+q=0
then has the double root z= —p/2; therefore exp[(—p/2) ¢] is a solution of
y’+py +qy=0 (3.15)
on — «0 <t<oo if p*=4q. The theory tells us that in all cases, (3.15) should
have two linearly independent solutions. We now employ a useful trick to
find (guess) a second linearly independent solution. Knowing that exp
[(—p/2) t] is a solution of(3.15), we try to determine a nonconstant function
w such that
Hiy=exn( —21)wt (3.19)

will also be a solution of (3.15). Now wy will be a solution of (3.15) on
—0o<t<oo if and only if
Wi"(t) + py'(t)+
ay()=9
or equivalently, using (3.19), if and only if
exp(5 )w"(t)—p exp( )w (+5 exp( )w(t)

Ns
+o| 5 exp(—§ )n(+exn( 5 )w( feaexp(—§ )w(t)=0
or if and only if
exp(—£7)|w+(a-") (0 =o. —0<t<o.
Since g=p?/4, w must satisfy the equation exp[(—p/2) t] w’(t)=0.

But exp[(—p/2) t]40, and so (3.19) will be a solution of (3.15) only if
w is determined such that w’(t)=0 (—coo<t<oo). Thus w(t)=c,+ ct
where c,, C, are constants. We therefore have (t)=(c, +c2¢) exp[(—p/2) ¢]
as a candidate for the solution of (3.15). Direct substitution shows that it is.
Now exp [(—p/2)t] is a solution of y’+py’+qgy=0 if p*=4g, and 1
exp[(—p/2) t] is another solution on — 0 <t<oo (verify!). Since they are
linearly independent solutions on — 00 <t<oo (verify!), Theorem 5, Section
3.3, tells us that we have proved the following result.
Theorem 3. Let p and q be constants such that p? =4q. Then every solution
gon —-oO<t<o of
y’+py'+qy=0 (3.15)
has the form
()=(e.+exthexo( ~Br), —0o<t<o
where c, and Cc, are constants.

There is an alternative and instructive way to establish the fact that
if p?=4g, then t exp[(—p/2) ¢] is also a solution of (3.15). We give this
method also because it is useful for solving higher-order equations. We
know that exp[(—p/2)t] is a solution of L(y)=y’+py'+(p?/4)
y=0.
This means that
p
=e" | 274 pz+— =0.

z=—p/2
a 3 2 2
— (L (e*))=— G (#+pz+ ") = te (#+ pz +) +e (2z7+p)
we see by substituting z= —p/2 that also

Note that 27+p is the derivative of z?+pz+p?/2 and both these vanish at
the double root z= —p/2. (This is a general result about multiple roots—
see Appendix 2.) As you may verify,
Gide \ 0 Ge. Gi) Ones Naata) Oca

éz\ ét) ét\ dz) = az\ at?) at?\ az J’
so that (6/dz) (L(e*)) = L(0/éz) (e”). Therefore
=1(te*) =0.
Z=—
p) 2 c=
— p22
This shows that te‘~?/*™) is also a solution of y”+py’+qy=0 if g=p?/4.
Exercises
11. Find the general solution of each of the following equations. If the equation is
real, express the solution in real form. Note that Theorem 3 is true if p and q are
complex, and thus equations with complex coefficients can be solved.
a) y"+9y=0 b) y’—Sy’+ 6y=0
c) y+ 10y’+ 25y=0 d) p"+2i’+y=0
e) 4y”"—y=0 f) y” + 5y’+ 10y=0
gh) ey’ +2y’+y=0, O<e<l h) 4y"+4y’+y=0
*12. In Eq. (3.15) with p? =4g, investigate the behavior of the solutions as t+ 00 for
various values of the constants.
*13. Recall the Definition: A function f is said to be bounded on some interval I if and
only if there exists a constant M >0 such that |f(t)|<M for all t on I. For example,
sin f, cos f are bounded on any interval, 1/t is bounded on [1, 2] but not on (0, 00),
e~' is bounded on [—5, «) but not on (—co, —5].
a) Determine which differential equations in Exercise 1 have a// their solutions
bounded on [0, 2).
b) Repeat part (a) for the interval (— co, «).
*14. Show that the solutions of the differential equation y”+py’+qy=0, wherep and q
are positive constants, are oscillations with amplitudes which decrease ex-
ponentially when p?<4gq (light damping) and that they decrease exponentially
without oscillating if p?>4qg (overdamping). How do they behave if p?=4q
(critical damping)?
The Phase Plane
Let ¢ be a real solution on 0<t<oo of the linear second-order differential

equation L(y)=0, where L has real constant coefficients. Let y,=o(#),
y=’ (t), where we now think of ¢ as a parameter, ranging over the
D2 Linear Differential Equations
interval 0<t<o. It is of interest to examine the graph of the curve I’,

called a positive semiorbit, given parametrically by the equations y, = (A),
y= (¢) in the (y,;, y2) plane (called the phase plane). For example, for the
equation y” + y=0, we may consider the solution y, = f(t)=c, cos ¢+c, sint,
so that y,=¢'(t)= —c, sint+c, cost, 0<t< oo, for each choice of the real
constants ¢c,, C>. It is clear that y? + y3 =c?+c3, in particular y, =0, y,=0
if ¢,;=¢,;=0, and hence each (positive. semiorbit y,—9(7), y,—©@ (1)
(0 <t< oo), is a circle in the phase plane with center at the origin. The arrow
(Fig. 3.1) represents the direction of motion along each curve in the phase
plane as ¢ increases. Note that through each point in the phase plane there
Vo
Figure 3.1
passes one and only one semiorbit. (This is not hard to prove using the
uniqueness property, Theorem 1, Section 3.1.) We note also that the above
example, though not too difficult to handle, would be even more transparent
if we observed that the solution
(t)=c, cost+c, sint
with c,, ¢, not both zero, could be expressed in still another form, namely,
$(t)=A sin (t+a)
where A=(cj+c¢3)’/ is called the amplitude, and «=arcsin (c,/c? +.c3)"/)
is called the phase shift. It is now obvious that #'(t)= A cos(t+.«) and there-
fore the curve y; = (ft), v2='(t), O<t<oo in the (y,, v2) plane is a circle of
radius A centered at the origin.
15. Establish the above formulas for A, «. Hint: Assuming c,, ¢, are not both zero,
: Cy C2 i
c, cost +c, sint (che)! ( ss2 cost + 2 2\1/2
sin)
(ct +¢3)"/ (Gere
oe)in Linear Homogeneous Equations of Arbitrary Order 93
16. a) Write the general solution of the equation y”+2y=0 in the ‘‘amplitude-phase
shift form.”
b) Determine the amplitude, period, phase shift of that solution @ of y”+9y=0
which satisfies @(0)=1, $’(0)=2.
c) Sketch and identify several typical positive semiorbits (that is, let 0<1<o) of
the equation »”+9y=0 in the phase plane. What happens if we let ¢ range on
the interval —co<f<0 (negative semi-orbit)? Indicate the direction of the
motion along each curve as f increases.
17. Sketch a few typical positive semiorbits in the phase plane for each ofthe following
differential equations. Consider also the negative semiorbits. Indicate the direc-
tion of the motion along each curve as f¢ increases.
a) y’+2y’+2y=0
b) v’—v=0
18. Suppose we had a pendulum for which a crude mathematical model would give
rise either to the equation in Exercise 16a) or 17a) above. Can you give a physical
interpretation of the semiorbits in the phase plane in each case?
19. Consider two solutions ¢(t)=c, cost+c, sint and w(t)=d, cost+d, sint of the
equation y” + y=0, where cj +c3 =dj +d3. Show that these solutions both give rise
to the same positive semiorbit in the phase plane, even though the solution @ need
not be the same as the solution w.
Exercise 19 shows that, although only one orbit passes through each
point of the phase plane, each orbit corresponds to many solutions with
different phase shifts.
3.5 LINEAR HOMOGENEOUS EQUATIONS OF ARBITRARY ORDER

WITH CONSTANT COEFFICIENTS
We can easily generalize the results of Section 3.4 for second-order linear
differential equations with constant coefficients to equations of arbitrary
order. Consider the linear homogeneous equation of order ” with constant
coeflicients a, az, :.., a,
L(y) — y at ay” = 3) se coo de a= 1) ef Any =() (3.20)
and look for a solution of the form e*' as before. Note that Eq. (3.20) reduces
to (3.15) when n=2, with a, =p, a, =q. Since L,(e”)=p,,(z) e, where
Pile aye. boo az,

is a polynomial of degree n, called the characteristic polynomial, we see
that the analogs of Theorems 2 and 3, Section 3.4, (although rather more
involved) may be stated as follows:
Theorem 1. Let 21, 22, -«-5 Zs, Where s<n, be the distinct roots of the char-
acteristic equation (of degree n)
Py(z)=2"+ay2""'+---+a,-12+a,=0
and suppose the root z; has multiplicity m;, i=1,..., 8, (m,+m2+-": +m,=n).
Then the n functions
ete at et
er perm) pe teat
7s, te?" veld, pits 1 pzst
are (i) solutions of L,(y)=0 on —0<t<o, and (ii) linearly independent

on —0o<t<oo. Hence by Theorem 6, Section 3.3, the general solution of
(3.20) is a linear combination of these n functions.
Of course, Theorem 1, Section 3.4, holds without change for higher-order
equations, and if the coefficients a,,..., a, are real each solution of L,(y)=0
can be expressed in real form exactly as before.
We do not prove Theorem | except to remark that if z;, 1<z;<s isa
root of multiplicity m; of the polynomial equation p,(z)=0, then
Pn(2;)=9, p,(Z;)=0, ..., py”? (z;)=0

but p!"(z;)#0 (see Appendix 3). This observation enables us to prove
the result much as was done in Theorem 3, Section 3.4 (alternative approach).
The linear independence of these solutions has been proved in Lemma 1, Sec-
tion 3.3.
Example 1. Find the general solution of the equation y*)+ 1l6y=0.

Since this equation has order 4, is homogeneous, and has constant coefficients,
Theorem | is applicable. The characteristic equation is z*+16=0. To solve this equa-
tion, we write
z= —16=16e""2™ 7 =0, +1, +2...
or, letting z=re'®,
re=t?
16e" +2") n=0, +1, +2.....
Hence r*=16 and 0=2/44+(n/2)n (n=0, +1, +2, ...), and the distinct roots are
Z,=2 exp [i(x/4)]=,/2 (1+7), 22=2 exp [3i(2/4)] =,/2 (—1+1), z3=2 exp [—i(x/4)]
=,/2 (1 —1), z4=2 exp [—3i(x/4)] =,/2 (— 1 —i), corresponding ton=0,n=1,n=—1,
n= —2, respectively. It is clear that the choices n= +2, +3,..., lead us back to one of
the roots 2,, 22, 23, 24 already listed. Since n=4 and since the characteristic equation
has four distinct roots, every solution @ of the equation y+ 16y=0 has by Theorem |
(here n=4, m, =m,=m;=m,4=1) the form
$(t)=c, exp[/2(1 +i) t] +c. exp /2(1—1) f]

+3 exp[,/2(-1+i) t]+¢eq exp[.\/2(-1-3) 4]
3.6 Reduction of Order 95
for some unique choice of the constants c,, ¢>, C3, C4. This may be written in the real
form
(t)=exp[/ 2t] (a, cos,/2t+a, sin ./2t)+exp[— /2t] (a3 cos,/2t+a, sin ,/2t)
for some unique choice of constants a;, @), 3, dq.
Example 2. Find the general solution of the equation y” + 3y"+3y’+ y=0.

Again Theorem | is applicable and the characteristic equation is 2*+3z7+3z+1
=(z+1)*=0. Thus z= —1 isa triple root and e~', te~', ?e~' are by Theorem 1 linearly
independent solutions on — 00 <¢<oo. Hence every solution # has the form
P(t)=e ‘(cy +ct+c3t7)
for some unique choice of the constants c,, ¢3, c3.
Exercises
1. Find the general solutions of the following differential equations.

a) py” —27v=0 b) vy? —16v=0
c) YO 2y"+y=0 = d) y\ +5y"+4y=0
e) YO+y=0
2. Find that solution @ of y+ 16y=0 for which $(0)=1, $/(0)=0, 6”(0)=0, 6” (0)
=0. (See Example | above.)
3. Given equation y“+/y=0, where 4 is a constant, find the general solution (in
real form) in each case: (a) 2=0, (b) 2>0, (c) 2<0.
* 4, Which of the equations in Exercise | have the property that (a) all their solutions
tend to zero as t> + ©, (b) all their solutions are bounded on 0<t< ©, (c) all their
solutions are bounded on —2 <1t<a@?
In closing this section we emphasize that the methods of solution which
we have developed are applicable only when the coefficients are constant
and the equation is linear! The analog of the phase plane (n=2) is n-
dimensional phase space and again solutions @ of Eq. (3.20) can be
pictured as curves in this space. However, we shall not pursue this topic
further at this point.
3.6 REDUCTION OF ORDER
The methods of Sections 3.4 and 3.5 do not apply to linear equations
with variable coefficients. Thus even though our theory tells us that there
are two linearly independent solutions of asecond-order linear homogeneous
equation, it may not be possible to find them. Sometimes it is possible to
guess or by some other means find one solution ¢, of the linear equation
L(y)=4o(t) y" +4, (t) y’ +42(t)p=O

on some interval J where do, a;, 22 are continuous on J/and a(t) 40 on /. Then
the same trick which led to Theorem 3, Section 3.4 (see Eq. 3.19) in the case
of constant coefficients with equal roots of the auxiliary equation will enable
us to find a second, linearly independent solution of L(v)=0 on / by reducing
the problem to one of solving a first-order equation.
Assuming that we know a solution ¢,, we let ¢,(t)=w/(s) ¢, (t) and try
to find a nonconstant function w so that L(@,)=0 for every ¢ on J. (Why
should w be nonconstant?) Since
bh=wo.+wd,, $h=w'd +2w'd, + wi,

we see that
L(b2) =o w" + (2doG, +4191) Ww +wLl(¢;).
But L(¢,)=0, since @, is a solution of L(y)=0. Therefore L(¢,)=0 for all

t on Jif and only if w satisfies the equation
do (t) y(t) w” +[2a9(t) $4 (t) +41 (2) $1 (t)] w =0 (3.21)

for all t on J. Note that (3.21) is a first-order linear equation in w', and is
readily solved as follows. Let w’=v and assume that @#0; then (3.21)
becomes
Separating variables and using properties of the logarithm and ex-

ponential function, we obtain the solution
el apa eat ey
rigmer] [esataa)*|
=exp(tosL 6.0 ~? logo (to)] a.vo(—| zAg(s

9i) to
t
“Taare | lexcta
Bava xp| — ay
0
where c is the constant [, (fo)*. Then, for fg and fin J, using w’ =v, we have
ee | c ta { a,(s) b
u exc exp( lee is)
to
3.6 Reduction of Order 97
and therefore
t o
5(t)=cd, (t) lz ar exp(-|% 7ts)do. (3.22)

to to
The reader should verify by direct substitution that L(d,)=0. This leads us
to the following result.
Theorem 1. If , is a solution of L(y)=0 on I, where do, a,, a, are
continuous on I and ay(t)#0 on I, and if b,(t)#9, then the function dy given
by (3.21) is also a solution of L(y)=0 on I. Moreover, the solutions d),, d4 are
linearly independent on I; hence every solution p of L(y)=0 on I has the form
b=c,6,+C22 for some unique choice of ¢5, Cy.
We have only to prove the linear independence ofthe solutions ,, @, on
I. This is done by computing W(d,, @,) and using Theorem 3, Section 3.3.
Exercise
1. Carry out the proof of linear independence of #,, h2 on J.
Example 1. One solution of y’+ty’—y=0 is ,(t)=t. Find a second linearly

independent solution @).
We could, of course, apply Eq. (3.21). Rather than try to remember such a com-
plicated formula, we proceed directly by putting @,=¢tw and forming L(,)= L(tw).
Since $5 =tw’+w and $3 =tw"+2w’ we have L(fp,)=tw" + 2w'+t?w’. Thus L(p,)=0
if and only if v=w’ satisfies
tr’ +(2+¢7) v=0.
Separating variables, we obtain
Thus, for ¢#0, v(t)=c(e~"!7/r?), w()=cfio(e~”!7/s”) ds, (to%0), and ¢,(#)

=ct fig(e *!?/s”) ds, (to #0, 140), where c is a constant, which we may take to be 1.
To establish the linear independence of #, and #,, we form their Wronskian
t
— 2
e $*/2
t | cor ds
W (1, b2) (t= =e? ¢£0,
to
Since this Wronskian is different from zero, , and ¢, are linearly independent on any
interval not containing the origin.
Exercise
2. Given one solution ¢#,, in each case find a second linearly independent solution
> on the interval indicated.
D
desea, Pi(J=", — 0<t<oo
b) y’—4ty’
+(4t? —2) y=0 $,()=e7, —0<t<0
c) (1—t?) y—2ty’'+2y=0 $i (t)=t, 0<t<1
d) ty”—(t+1) y+y=0 $, (t)=e', t>0.
In general if #,,..., d,, where k <n, are linearly independent solutions

on some interval / of the linear equation of nth order
L,(y)=4o(t) y +4,(t) y"" 9 +---+4,-1() y +4,(0 y=0.

where d,,..., d, are continuous on J and a,(t)#0 on J, one can reduce the
problem successively to a linear equation of order n—k. We illustrate this
with the following exercises.
Exercises
3. Suppose ¢,, #2 are linearly independent solutions on an interval / of the

L3(y)=y" +4, (t) y” +a9(t) y¥+a3(t) y=0

(we are taking ao (t)= 1).
a) Let 6=w@, and compute the linear equation of order two which must be
satisfied by w’ in order that L,()=0.
b) Show that (#3/¢,)’ is a solution of the equation of order two found in part a).
c) Use the result of part b) to reduce the second-order equation to one which is
linear and of first order.
4. Two solutions of
ty” —3ty'+3y=0, t>0
are f,(t)=1, 62(t)=0°. Use this and Exercise 3 to find the general solution of
the given equation for +>0.
5. a) One solution of the equation
L(y) =ao(t) YX+4, (t) ¥" Y+---+a, (td) y'+a,(d y=0.

is ¢ '/? sin ¢. Find the general solution for r>0.
b) Repeat part a) for the equation
2ty” +(1—4t) y'+(2t—1) y=0,

given that e' is one solution of the homogeneous equation.
Sui Linear Nonhomogeneous Equations 99
6. Make the change of variably y=w(¢) v(t) in the equation
y"+p(t) y +q(t) y=0

and choose the function v(t) to make the coefficient of wu’ in the resulting equation
for uw equal to zero. Show that the equation for u then becomes
u" —4f[p(t)]* + 2p (t)—4q(t)} u=0.

7. Apply the change of variable suggested in Exercise 6 to the equation
t?y" +ty' +(t?—n’) y=0

and find the resulting equation.
8. a) Show that the change of variably y=w'/q(t) u reduces the nonlinear first-order
equation
y¥ +p(t) y+q(t) y?=r(t),

known as the Riccati equation, to the second-order linear equation
b) Apply this procedure to solve the equation
ty +ty+t?y?=1.
3.7 LINEAR NONHOMOGENEOUS EQUATIONS
We now turn to the nonhomogeneous second-order linear equation
L(y)=4o(t) y" +a, (t) ¥ +42(t)y=S(0) (3.23)

and, more generally, the mth-order linear equation
L,(y)=do(t) ¥ +4, (1) YP +--+ +4,(t) y=S( (3.24)

where throughout do, d;,.... d,, and fare given functions continuous on some
interval J, and a,(t)40 on J. In physical problems having (3.23) or (3.24) for
a mathematical model, the nonhomogeneous term f(t) represents an external
force acting on the system. For example, if the damped linear mass—spring
system considered in Section 2.1 is subjected to a given periodic external
force A coswt, then the equation of motion is
k A
y’ +by' +— y=— cosat. (3.25)
m m
(see Eq. 2.4, Section 2.1). Initial conditions are imposed as before. We
remark that the equation for the current in an electrical circuit having
resistance, inductance and capacitance in series and a periodic impressed
voltage also has the form (3.25) under the appropriate physical assumptions.
The entire development is based on the following fundamental result:

Theorem 1. Suppose , is some particular solution of L(y)=f on I, and
suppose that $1, b» are two linearly independent solutions of L(y)=0 on I.
Then every solution w of L(y)=f on I has the form
W=C1—, +C2b2+W, (3.26)
where C1, C, are constants which can be determined uniquely.

Since every solution of (3.23) has the form (3.26), we refer to (3.26) as
the general solution of (3.23). According to Theorem 1, to find any solution
of (3.23), we need only find two linearly independent solutions ¢,, ¢, of
L(y)=0 and some particular solution w, of L(y)=f, and then use the given
initial conditions to determine c, and c).
Proof of Theorem 1. Since y, is a solution of (3.23) on J, we have L(w,)=f
for all t on J. Since yw is also to be a solution of(3.23) on J, we have, using the
linearity of L,
LWy—,)=LW)-LW))=f—-f=9.
This shows that y—y, is a solution of the homogeneous equation L(y)=0
on /. (Recall that this much of the proof was already established in Section
3.2 property ii). Therefore, by Theorem 5, Section 3.3, there exist unique
constants c,, c, such that
W—W,=C19, + Cro? for all ton J

which completes the proof. I
The nth-order linear nonhomogeneous equation (3.24) can be treated
in the same way, and Theorem | has the following analog.
Theorem 2. Suppose i, is some particular solution of L,(y)=f on I, and
suppose that p,, 2,..., b, aren linearly independent solutions of L,,(y)=0 on I.
Then every solution W of L,(y)=f on I has the form
W=C1P, 4+Coh2t? + 6,9, t+Wp
where Cy, Cz,.-., C, are constants which can be determined uniquely.
Exercises
1. Prove Theorem 2.
2. Compare Theorem 2 in the case n=I with the results of Section 1.4, in
particular with Theorem 1, Section 1.4.
We shall now study some methods for finding a particular solution ofthe
equation L(y)=for L,(y)=f.
37 Linear Nonhomogeneous Equations 101
The Method of Variation of Constants
This general method for finding a particular solution is applicable whenever

one knows the general solution of the associated homogeneous equation
L(y)=0 or of L,(y)=0 on J. We begin with the second-order case.
Let @;, (2 be two linearly independent solutions of L(y)=0 on J. (These
may either be given to us or in some instances we can find them by one of
the methods already studied.) The method consists offinding functions u,, u,
such that the function
Wp=Urh; +Urpy (3.27)

will be forced to satisfy the equation L(y)=/ for all ¢ on J. It is remark-
able that such a simple device works, because when uw, and wv, are constants
the function (3.27) satisfies L(y) =0 and thus cannot be a solution of L(y)=f
for f40. It is of course not obvious that such functions u,, uv, can be found.
We first argue in reverse; suppose we have found functions w,, uw, such that
(3.27) satisfies L(y)=fon I. Then we have, for all ¢ on J,
(uid, +Urh2)
=u, $b)+25 +P, +uyd2,
(ub, +Urh2)" =U,5 +urh5 + 2u)P+ 2u,$5 + uh, +uyo2
and using L(¢,)=L(¢,)=0 we obtain
L(uyP; +urb2) =U, L(G) + UL (b2) +40 (Piru + G2u3)

+ 2d (Puy + Pps)
+ay(Pru + bur)
= do (Put + haus) +2(Piui + hour)
+a, (Pu + 2u5)=f
for all ¢ on J. We would now like to obtain two relations from which to
determine the two functions u,, u,. We note that if 6,u;+,u5=0 for all f
on J then also (¢,u,+@,u,) =0 for all ¢ on J. But (¢,u,+¢,u5) =¢,u{
+o,u;+\u;,+5u5. Therefore, if we assume
piu, + pou, =0 (3.28)
for all ¢ on J, then the requirement
Ag (Pru + G23) +2(hiuis + hour) |+ay (Givi + b2u,)=f
implies, on using (3.28) and the equation obtained by differentiating (3.28),
that we must also have, since (3.28) implies that @,uj{+.uj+o{u;
+ hou, =9,
(3.29)
Nor tl ‘ Leen 4 if
Pu, + ou, =—
ag
for all t on J. Thus the assumption of the existence of a solution of the
form (3.27) of the equation L(y)=/f has led us to the two equations (3.28),
(3.29) from which we hope to determine wu, uv and then the functions
U,, Uz. But now reversing the argument we see that if we can find twofunctions
u', u to satisfy equations (3.28), (3.29), then indeed w,=u, 6; +u2h2 will
satisfy L(y)=fon J.
To find a particular solution of the equation L(y)=/, we may therefore
concentrate on equations (3.28), (3.29). These are linear algebraic equations
for the quantities w',, uw, and the determinant oftheir coefficients is W(,, o>).
Since the solutions $,, ¢, of L(y)=0 are by hypothesis linearly independent
on J, it follows that W(¢,, $2) ()40 for all t on / (Theorem 2, Section 3.3)
and the system ((3.28), (3.29)) of equations can therefore be always solved
(in fact uniquely) for the quantities wv, w,. By Cramer’s rule (Appendix 1),
the solution of the algebraic equations (3.28), (3.29) is
=) Or gfe weedSEM Exoni=

agW (1, b2)’ . dyW (,, b2)’
Thus a possible choice for u,, uz 1s
£(3) 269) OO
t
ON =|PROMS CMO L is(0=| ao(s) W(b,. 62) 6)”

to to
for any fo, ¢ in J, where we have taken the constant of integration to be

zero. Substituting in (3.27), we find that
[L620 6) 0bo),
Ze to
NOC oy
is a solution of L(y)=f on J, as may be verified by direct substitution. We
have thus sketched the derivation of the following important result.
Theorem 3. Let $1, $y be any two linearly independent solutions of the equa-
tion
L(y)=4o(t) y" +4)(t) y +ay(t) y=0,

where do, 1, Az are continuous functions on some interval I and ao(t)40 on I.
Then a particular solution , of L(y)=f, where f iscontinuous on I, is given by
Eq. (3.30).
Equation (3.30) is usually called the variation-of-constants formula.
The reason for this name is clear from the method. Although the condition
expressed by Eq. (3.28) is artificial, the fatt that we can solve the problem
using it justifies it, and this is actually the essence of the method. This
method of finding a particular solution of L(y)=fcan be used whenever the
coefficients a, a,, a7 in L and the function f are continuous on J and ay #0
on J, and whenever one knows the general solution of the associated homo-
geneous equation. It is not restricted to equations with constant coefficients.
3.7 Linear Nonhomogeneous Equations 103
Exercise
3. Prove Theorem 3 by direct substitution of the function w, given by (3.30) into

L(y)=fon I. (Hint: Write w, in the form
(£6) dr(s) i £(9) bls)

=O
==)
|camecere’
to
1s—
| awe caTG
to
ds
before beginning the differentiation. ]
Example 1. Find the general solution of the equation
seas t us t ml
y +y=tant, SSD eas
es 2 2
Since, by Section 3.4, @,(t)=cos t, @,(t)=sin ¢ are linearly independent solutions of
y’+y=0 on any interval, they are linearly independent on —2/2<t<z/2; in fact
cost sint
W (61, $2) (t)= =1.
—sint cost
Instead of memorizing Theorem 3, it is simpler to remember the key steps of the

method. By what we have just seen, W,=u,; cost+u, sint will be a solution of
y’+y=tant on —2/2<t<n/2 if and only if the functions uw, and uw, are such that
u,, U5 satisfy Eqs. (3.28), (3.29), that is, if and only if
ui, cost+u sint=0 1 1

ar 5 —x~<t<~
—u', sint+u, cost=tant 2 v
Thus
. sin*t 1—cos’t
u; (t)= —tant sint= — = =cost—sect
cost cost
' T T
u3(t)=cost tant=sint, aint
and we may take

t t
us(= |(cost see dt, vo(= [sined

0) 0
or
Tl Tl
u, (t)=sint—log |sect+tant|, — u2(t)= —cost, Ary
Therefore y, =(sin—log(|sect+tan¢|) cost+(—cost) sint=—cost log(|sec t+ tan Z|).

Therefore, by Theorem 1, every solution of y’+y=tant on —2/2<t<z/2 has the form
W(t)=c, cost+c, sint—cost log |sect

+tan ¢|
for some unique choice of the constants c, and c).

Exercises
4. Find the general solution of each of the following differential equations.

a) y’+y=sect, —1/2<t<n/2
b) y"+4y'+4y=cos 2
c) y’+4y=f(t), where fis any continuous function on some interval / :
d) y’bsg + 4y=3e°'+2r°+sin ¢
€) y’+(1/4?) y=/(d), (t> 0), fcontinuous, given that o, (¢)=./7 is a solution ofthe
homogeneous equation.
5. Given that ¢ is a solution of the equation y’+k*y=f(t), where k is a real con-
stant different from zero and fis continuous for 0<t<oo, show that c, and c, can
be chosen so that
t
c I
o(t)=c, coskt+—* sin ket |sink(t—s) f(s) ds
0
for 0<t<oo. (Use cos kt and sin kt/k as a fundamental set of solutions of the
homogeneous equation.) Find an analogous formula in the case k=0.
6. Given the equation
y’+5y'+4y= fl),
use the variation-of-constants formula and Theorem | to prove that:
a) If fis bounded on 0<t< oo (that is, there exists a constant M >0 such that |/(s)|
<M on 0<t<oo), then every solution of y’+5y’+4y=/(t) is bounded on
VEST <66),
b) If also f(t)>0 as too, then every solution # of y’+5y'+4v=f(t) satisfies
(t)>0 as t> 00.
* 7, Can you formulate Exercise 6 for the general equation
y"+ayy' +azy= f (t), d,@> constant
with a,, a, suitably restricted?
The method of variation of constants and Theorem 2 are applicable

to the nth-order equation L,,(v)=/f with coefficients do, a;,... d,, fcontinuous
and ay(t)#0 on some interval J provided one knows n linearly independent
solutions (1, p2,..-, Py, Of the homogeneous equation L,(y)=0 on J. Using
the second-order case for motivation, we try to find finer: High, banat then
not all constant, so that
Wr=U191 + Urb. + ae UO
will be a solution of L,(y)=f on L. If (see the second-order case)

uyo,+ush,+...+u,g,=0 on J, then wi=u,,+---+u,¢), on J and if
up, t+---+u,?,=90 on J, then Wi=u,¢{+---+u,¢%. Continuing in this
Sy Linear Nonhomogeneous Equations 105
manner we find that if wv, w5,..., uj, are chosen to satisfy the system oflinear
algebraic equations on /
+:-+ud, =
uid, tuo,
ud +u,¢, +:-+uid,, =0
(3.31)
.
+u, gf 9+... 4g" =0
ud"
purged t
VP uy GO- D4.
ao
then the function
Wp=U1h, +h. +-- +u,d,
will satisfy L,,(v)=fon J.

Note that the determinant of coefficients of the system of equations (3.31)
is W(d,, b2,..., @,), which is different from zero for every ¢ in J since
1, P2,--., @, are linearly independent solutions of L,(y)=0 on J.
Exercise
* 8. Verify that the function ¥,=u,, +uU2p7)+u3¢3 is a solution of L3(y)=/f on 1.

[Hint: Solve (3.31) by Cramer’s rule and integrate to find uw, v3, u3.]
Thus the entire problem is reduced to solving the algebraic system

(3.31). Since its determinant of coefficients is W(¢,, ¢2,..., @,), and since
the solutions ¢,,..., ¢, of L,(v)=0 on J are linearly independent on /,
W(,,---, Pn) (Q)40 for ¢ on J and (3.31) always has a unique solution for
the quantities wu/,..., uj), on J. In fact, letting W;(1) be the n by n determinant
having the same elements as W(¢,,..., ¢,) (t) except with (0, 0,..., 1) as its
jth column, we see that Cramer’s rule (Appendix 1) gives
a) a)
u y(t) = W (1... Pn) Ao(t) hao Orr,
The uw; are obtained by integration, so that
ENG 4 (eine OL abaaly

Him), se |Ws. O)) dol) 2
where ft, and ¢ are any two points of /.
We have therefore sketched the derivation of the following result, which
generalizes Theorem 3.
Theorem 4. Let @,, bo, .«., P, ben linearly independent solutions of the equa-
tion
La(y)=4o(t) y +4,(1)y+ +4,-1(t) ¥ +4,( y=f (9),

where do, Q4.---, 4, f are continuous functions on some interval I, and
dy (t)#0 on I. Then a particular solution , of L,(y)=f is given by Eq. (3.32).
Application of Theorem 2 now yields the general solution of L,,(y)=/fon J.
Exercise
9. Find the general solutions of the differential equation

a) y” —8y=e!
b) yx?+ l6y=/(t), fcontinuous on — 0 <t< co
The Method of Judicious Guessing (also known as the method of undeter-

mined coefficients or the annihilator method)
It is clear from the above discussion that the method of variation of
constants always yields a solution of the nonhomogeneous linear equation
if one knows the general solution of the homogeneous equation, regardless
of whether the equation has constant or variable coefficients. However,
there is a quicker way to solve the equation L,(y)=/ provided that L,, has
constant coefficients and provided that f has a rather special form. We shall
study the general equation of order 7 here because it is no more difficult than
the second-order equation. We shall assume that fis defined on — 00 <t< oo
and has at most a finite number of linearly independent derivatives. This 1s
equivalent to assuming that f must satisfy some homogeneous linear
differential equation with constant coefficients. (Why? Think of the defini-
tion of linear dependence of the functions f, f’,...,f’” for some m>1.)
Therefore (by Theorem I, Section 3.5) it must in fact be assumed that f(t)
is a linear combination of terms of the form
tkemt
where k>0 is an integer and m is a real or complex number including

possibly zero. By the principle of superposition, Section 3.2 (see Exercise 12),
we may as well assume that f(t) consists of only one term of the form
Ki=cle™
where c is a constant. (Why?) Before we proceed we stress again that the
method we are about to explore further only works under the special con-
ditions stated above. But if these conditions are not satisfied, the method of
variation of constants is applicable.
For simplicity of exposition let us first assume that f(t)=ce™’, where m
is a real or complex number, possibly zero, and let us consider L,(y)=ce™,
where L,(y)=y"
+a,y"" +---+4,_-1y'+a,y. Our task is to find a par-
3.7 Linear Nonhomogeneous Equations 107
ticular solution w, of L,()=ce™ and then to apply Theorem 2 to find all

solutions. By an elementary calculation (see Section 3.5)
L,,(Ae*) = Ap,(z) e*,

where p,,(z) is the characteristic polynomial
Py(Z)=2"+a,2""'4++++-+a,_,Z+a,.
Thus Ae” can satisfy L,(y)=ce™ if and only if Ae” p,(z)=ce™ by all t.
This will be true if we choose z=m, and A =c/p,(m) provided that p, (m)#40.
We therefore try as a solution of the equation L,(y)=ce™ the function
W,(t)=(c/p,(m)) e™ provided that m is not a root of p,(z)=0 (that is, provided
that p,,(m) #0). Indeed, if p,,(m)40, we have by the linearity of L,
Cc é c
L,(W
Ea afon) J pylon)
)= L, Ge e)Se Lite") See
ny Pe EN CE = Cet
Note that the assumption that m is not a root of p,(z)=0 is equivalent to

the assumption that e” is not a solution of L,(y)=0. Having a particular
solution, we then apply Theorem | or 2 to find the general solution.
Example 2. Find the general solution of the equation y”—9y=e'.
The general solution of the homogeneous equation is c,e*’+c,e*'. Since e' is not
a solution of the homogeneous equation we try y,=Ae'. Then w/ —9y, = Ae'—9Ae'
= —84Ae'. Then )—9y, =e' if and only if —84=1 or A= —¥. Thus y,= — ge’ and by
Theorem | every solution has the form w=c,e"+c,e*— $e".
Example 3. Find the general solution of the equation y”—9y=sin 21.

Since sin 2t =(e"), we consider first y’ —9y =e" and try p, = Ae*". Then i, = 2iAe*",
w= —4Ae?". Thus W—9,=(—4A—9A) ec?" =e" if and only if A= —75. Therefore
W,p=(—7s) e7". By Theorem 1, Section 3.4, (—7s) sin2¢ is a particular solution of
y”—9y=sin2t and the general solution has the form W(t)=c,e*!+c,e *'—(7s) sin 20.
Returning to the equation L,(y)=ce™, let us now suppose that e” is a

solution of L,,(y)=0 but te” is not; this is equivalent to assuming that m is a
simple root of the algebraic equation p,(z)=0. This means that p,(m)=0, but
p,,(m) 40 (see Appendix 2). We have
L, (e%") =p,(z)
Since (0/dz) L,,(e**) =p,,(z) e +p, (z) e*, and since (0/0z) L, (e”) = L, (de*/0z) =
L,(te‘) we see that L,,(te**)=p,,(z) e* + te*'p,(z). (Compare with the case of
equal roots in Section 3.4 (immediately following the statement of Theorem
3).) Therefore L,,(te’”’)
=p;,(m) e”, which suggests that if m is a simple root of
P,(Z)=0, the equation L,(y)=ce™ hasa particular solution of the form
W,= Ate™
for some constant A. Indeed, we see that ,=[c/p,(m)] te™. Of course, in
practice we would simply assume a solution of the form Ate”, substitute and
determine A.
Example 4. Find a particular solution of y”—9y=e™.
We observe that e*' is a solution of y” —9y=0, but e*" is not and therefore we guess
a particular solution of the form w,(¢)= Ate*’. Then
Wi,(t)=3Ate* + Ae*, ”(t)=9Ate*
+ 6Ae™.
Therefore ’,(t)—9W,(t)=6Ae*=e* if and only if A=. Thus w,(t)=(¢/6) e*' is a
particular solution of y’—9y=e*.
Similarly, if m is a double root of p,(z)=0, then p,(m)=p;(m)=0, but
p,, (m) 0. This implies that e”", te” are solutions of L,,(y)=0, but t7e”" is not,
and this in turn suggests assuming a solution of L,(yv)=e”™ of the form
W,=Ate™ for some constant A. As before, one determines A by direct
substitution.
In general, if m is a k-fold root of p,(z)=O(k<n), then p,(m)=p’,(m)
=...=p"-(m)=0, but p™(m)40. This implies that e™, te™,...,1%
te”
are solutions of L,(y)=0 but t*e” is not. Thus in this case one would “ju-
diciously guess” a solution y, of L,(y)=e™ of the form y, = At*e™ and sub-
stitute to determine the constant A.
We turn now to the more general equation L,(y)=ct*e”, k a positive
integer, c a constant. A straightforward but tedious calculation shows that
L(t*e™)
= p,,(m) t*e™ + pj,(m) kt*~*e™ +k(k—1) t*~7p%(m) e™+ <
where there are at most a total of k+1 nonzero terms. Therefore when
guessing the form of a particular solution it is clear that we cannot now
merely try w,(t)=Ar‘e™, but must include terms te” (j=0, 1,..., k) which
will serve to cancel out the terms we get in the above calculation. Thus we
try w,(t)=Aye™ + Ante™ +--+ A, ,t*e”™ and find the A,’s by substitution.
Example 5. Consider the equation y’ —9y=Pe'.
Here k=3, m=1 (and m=1 is not a root of m*—9=0). Determine the form of a
particular solution. In this case we try ,(t)=A,te'+ A,17e'+ A3te'+ Axe’, where
the constants A,, A, A3, A, are determined by direct substitution.
Exercise
10. Find the general solution of y’—9y=Pre'.
As in the case k=0 considered previously, additional complications
arise if now one or more of the functions e”", te”",..., “e™ is a solution of
the homogeneous equation L,(y)=0. The ‘‘judicious guess’? based on the
above considerations is to multiply each one of these functions by the lowest
power of t, say t’, such that none of the resulting functions is a solution of the
homogeneous equation L,(y)=0; then assume a solution of the form ,(t)
=A,te™+Azt* e™+---+A,4,0%*4e™", and determine the constants A,,
Ay,..., Ay, by direct substitution in the equation
Si Linear Nonhomogeneous Equations 109
Lap ecttent:
Example 6. Find the form of a particular solution of the equation y”—9y=re*.
Here e *‘is a solution of y”—9y=0, but te> is not. Therefore, to find Wp, we try
W,(t)=Ayte *+A,07?e *+A4,8e* 4 Aytte*. (Note that there is no point in includ-
ing also a term Aje*' since e * is a solution of the homogeneous equation.)
Exercises
11. Find the general solution of the equation y”—9y=re~*.

12. Examine the problems in Exercises 4 and 9 which were done by variation of
constants and decide which ones could be done by the method ofjudicious guessing
and which ones could not. Carry out the details of those that could.
13. Find the general solution of each of the following.
a) y"+4y=sin 21 b) y’—4y=2e-
Cy aay oy ty=— =e" d) y"—4y=te'
e) vy’ —4y=te”! f) y’—4y=r'e#
*14. Consider the equation y’+k?y=2k sinkt, and show that all solutions are un-
bounded as t«. This phenomenon always occurs if the nonhomogeneous term is
a sine or cosine function which is a solution of the homogeneous equation (physical-
ly speaking if the “‘applied” frequency is a “natural” frequency of the system).
a) y’+y=cosectcott, O<t<n
fox) Py" +0'+4vssin log |t|
) py” —3r?y" + 6ty’—6y=0
d) y’—6y' +9y=e'
) y’—6y'+ 9y=te**
ia’)
f) ty" —31y'+4y=log ||
g) y+ 8y + l6y=0
h) ?y’+ty’—2iy=0
i) VO —2yp"+y=e'+sin ¢
jy ty stant, O<t<n/2
k) ¥?+y=g(t) g continuous
l) y’+y=h/(t), where
h(t)=t, Ox<t<x, h(t)=ncos(xn—#t), n<t<2n,
and h is periodic with period 2z.

16. A 50-gram mass can stretch a spring 2 cm under its own weight. The spring
satisfies Hooke’s law and there is no friction or external force. Determine the fre-
quency with which the mass will oscillate (Example 1, Section 2.1). [Hint: The
spring constant is determined from the information in the first sentence. Note
also that g =980. |
. Suppose that the mass-spring system in Exercise 16 is oscillating with an
amplitude of 5 cm. Find the maximum velocity. [Hint: Consider the initial
conditions (0)=5, #'(0)=0.]
. Suppose that the mass-spring system in Exercise 16 is at rest in an equilibrium
position at time r=0, and a force 500 cos 2¢ is applied. For what value of ¢ will the
displacement first equal one centimeter?
19. A spring is stretched 15 cm by an 8-kg weight. Suppose a 4-kg weight is
attached to the spring and released 30 cm below the point of equilibrium with
an initial velocity of 180 cm/sec directed downward. Determine the motion of the
system (Example 1, Section 2.1).
20. Determine the motion of the system in Exercise 19, with air resistance 1000 dynes
when the velocity is one centimeter per second. (Note that g=980.)
Mal Consider the mass-spring system of Exercise 19, but with an additional external
force 5 cos 2t. Determine the motion.
De A body of mass m falls from rest from a height 4 above the surface of the earth.
Assume that the only forces acting on the body are the force of gravity and a force
of air resistance which is ¢ times the velocity of the body. Find the motion of the
body, and show that its velocity approaches the limit mg/c as t>00.
ZS: The bob of a simple pendulum of length 2 feet is displaced so that the
pendulum makes an angle of 5° with the vertical and then is released. Assume
that the motion is determined by Eq. (2.11), Section, 2.2.
a) Find the angle 0 which the pendulum makes with the vertical as a function of
time.
b) Determine the frequency of the vibration.
c) Calculate the distance traveled by the pendulum bob during one period.
d) Find the velocity and acceleration of the bob at the centre of its path.
24. A simple pendulum of unit mass vibrates in a medium in which the damping is
proportional to velocity. If the pendulum bob passes through the equilibrium
position 8=0 at t=0 with velocity vo, show the angle @ is given by
OG) e “/? sinwt

0)
where w=(g/L—k?’*)'/*, k is the damping constant, and L is the length of the

pendulum. Find k if the distance traveled during one complete vibration is half
the distance traveled during the previous vibration. Use the linearized version of
Eq: (2.10); Section 22:
. The current /in amperes in an electrical circuit with resistance R ohms, inductance
L henrys, and capacitance C farads in series is governed by the equation
1
EU ERE SIE
3.8 Resonance 111
where E(t) is the applied voltage. Suppose the applied voltage is a constant Eo.
a) Show that the current decreases exponentially if CR? >4L
b) Find the current if CR?=4L and 1(0)=0, 1'(0)= E/L
c) Find the current if CR*<4Z and /(0)=0, I'(0)=Eo/L
26. Consider the circuit of Exercise 25. Find the current if
0 t<0
E(t)= 7
Ey sinat t>0
assuming /(0)=0, 7’(0)=0.

27. Find the current in an electrical circuit with inductance and capacitance, but no
resistance, and an applied voltage E(t) given by
St
Ee O<r<l
1 t=
with /(0)=0, 7’(0)=0.

28. A sphere of radius R floating half submerged in a liquid is set into vibration. Given
that y is the vertical displacement of the diametral plane of the sphere from the
equilibrium position, show that
pg yay:
P= ee
Show that for small vibrations (neglecting the nonlinear term) the sphere vibrates
with frequency 1/272(3g/2R)'/?.
3.8 RESONANCE
Consider the undamped mass-spring system with a given periodic external

force. The equation of motion is (2.2), Section 2.1. For convenience, we
write this equation in the form
y’ +kéy=A coskt. (3.33)
We call k, the natural frequency of the system and k the applied frequency.
We can solve Eq. (3.33) by the method of judicious guessing. If k#ko, a
particular solution is given by
A
Vol)= ap COS kt. (3.34)
Exercise
1. Verify the particular solution (3.34) of Eq. (3.33) with k# kp.
Observe that if the applied frequency k is close to the natural frequency

ko, then the particular solution given by (3.34) represents an oscillation

with large amplitude. If we think of kp as fixed and let, in (3.34), the
applied frequency k approach ko, then the amplitude of the oscillation
becomes unbounded. This phenomenon is called resonance. If k=ko, a
particular solution cannot be obtained from (3.34). However, returning to
(3.33) with k=ko, we obtain a particular solution
A
¥(t) = eo t sin kot
of the equation y”+k$y=Acoskot.
Exercise
2. Find the above particular solution.
Thus we see that in the case of resonance, the amplitude of the particular
solution (often called a forced oscillation) is not constant, but is an un-
bounded function of t. This phenomenon predicted by the model (3.33) does
not occur in a real physical system because of the presence of friction, which
causes damped oscillations rather than periodic solutions in the homo-
geneous case. In the presence of friction Eq. (3.33) is replaced by
y"+ay+k2y=Acoskt, a>0. (3.35)
Exercises
3. Find a particular solution of Eq. (3.35).

4. Show that as k approaches ko, the amplitude of the particular solution found in
Exercise 3 increases, but remains bounded if a>0.
The situation described in Exercise 4 is of practical importance in tuning

radio circuits, where the object is to adjust the applied frequency to resonance,
thereby maximizing the amplitude. In other situations, the object is to avoid
resonance.
5. Determine a particular solution of the differential equation
y"+py +qy=A coskt
of the form $(t)=B cos (kt—a). Show that the amplitude of the oscillation B is a
maximum if k=(q—p7/2)'’? (called the resonant frequency) provided p?<2g.
What happens in the case p*>2q? Show that at resonance the amplitude of the
oscillation is inversely proportional to the damping p.
6. Consider the differential equation
my” +ky= f
(t)
where
3.8 Resonance 113
iA O<t<e
0, t>é
Find the solution #(t) such that #(0)=0, '(0)=0. Is @ a solution in the sense of
Theorem |, Section 2.5? Give a physical interpretation of this problem and its
solution, thinking of ¢ as a small positive constant. Discuss the behavior of the
solution as ¢0; does lim, .9, ’(f) exist for all 1?
CHAPTER 4
Linear Systems of
In this chapter we will study the theory of systems of linear differential

equations, together with an outline of the basic theory of nonlinear systems
at the end of the chapter. As we shall see presently, mathematical models
of physical systems somewhat more complicated than those investigated
in Chapters 1, 2, 3 (motion of several interacting particles, electrical net-
works, population problems involving more than one species, etc.) often
lead to systems of more than one differential equation. Such systems can
be reduced in most instances to /inear systems of first-order differential
equations with the aid of certain simplifying assumptions. To study these
linear systems, we shall make use of linear algebra (vector spaces and
matrix algebra). As a very special case of every result obtained in this chapter
we will obtain a corresponding result for a scalar linear differential equation
of second (or higher) order, such as those studied in Chapter 3. Thus students
familiar with linear algebra can study this development directly and omit
Chapter 3.
4.1 INTRODUCTION
We shall consider systems of first-order linear differential equations of the

form
V4 = 441(0)¥y +442(t) 2 ++ +44,(1) Ve +91(0)

V2 = 43; (t) ¥y +422(t) v2 +--+ +4,(1)¥,+.92(t) (4.1)
Vn = ant (7) vy + Qn? (1) v2 6 + Ann (2) Vat Gu (r).
where the given functions a;,(t), where i, 7=1,...,m, and g;(t), where
i=1,...,, are continuous on some fixed interval .. Unless mentioned spe-
114
cifically otherwise the interval . can be open, closed, half open, finite or
infinite. If n=1, we have the important special case of a scalar first-order
equation, which we write in the form
y'=p(t) y+q(t), (4.2)

where p(t) and q(t) are given functions.
System (4.1) is linear in y 1, y2,...,V,, and yj, y3,...,¥,. The scalar
equation y’=2y? is an example of a nonlinear differential equation. These
much more complicated equations will not be considered in this chapter;
of course some simple cases of nonlinear equations have been studied in
Chapter 1.
Example 1. Consider the system
Vu Vi tye +e
Yo=UYi—Ys (4.3)
¥3=Yity2—y3t+2e “,
where .¥ is the real line, {t|— 00 <t< oo}. Here, n=3, and in the notation of(6.1)
ay,(t)=1 ay2,(t)=—t a,3(t)=0 gi(t)=e

az,(t)=t? a,(t)=0 a3 (t)=—1 g2(t)=0
a3,(t)=1 a3,(t)=1 a33(t)= —1 g3(t)=2e'.
Consider now the array
1 -=t 0
ao=|r 0 1] (4.4)
1 1 -1
A(t) is a matrix whose entries are functions. The properties of matrix addition,
multiplication by scalars, and matrix multiplication with constant entries also hold
for matrices whose entries are functions defined on a common interval %. Let y and y’
be the column vectors
a yi
Y=] 2 |, v-|4),
A ee y3
and let g(t) be the vector
x-| 0 | (4.5)
Dem
Then, observing that matrix vector multiplication of A(t) and y gives
Vary
A(t) y= |tyi—Ys |
Vitae) a)
we see that system (4.3) may be represented conveniently in the matrix vector form
y=A(t)y+e(),
where A(t) and g(¢) are given respectively by (4.4) and (4.5).
116 Linear Systems of Differential Equations
Returning to the general case of system (4.1), we define the n x n matrix
Ayi(t) @y2(t) ++ 44,

(t)
A(t)= a(t) 422(t) te; 4n(1) (4.6)
Gyi(t) prt) -* Oyn(t

whose entries are the n? functions a;;(t), where i, j=1,..., 7. Next, define the
vectors g(t), y, y’ by the relations
“91 (2) Vy vy
g(=| 7), yap], y= A (4.7)
an (0) Vr y,
Then the system (4.1) can be written in the form
y'=A(t)y+e(0). (4.8)
Exercise
1. Given the system
é =y,+cost
yo=V1
Define the matrix A(t) and the vectors y, y’, g(t), and write this system in the
form (4.8).
Before proceeding with the definition of a Solution and a discussion of
the system (4.8), we need the following definitions.
Definition 1. A matrix (such as A(t)) or a vector (such as g(t)) is continuous
on an interval ¥ if and only ifeach of its entries is a continuous function at each
point of F.
Definition 2. An nxn matrix B(t) or a vector u(t) with n components, de-
fined on an interval ¥% and given respectively by
byi(t) by2(t) ++ Binl(t) u(t)

aya] ral) ba) ~ bal] gif al -)
Dnt (t) bn2 (t) seas ke (t) Uy (t)
is differentiable on # ifand only if each of its entries is differentiable at every

point of ¥%. Their derivatives are given by
1) ial) bial) u(t)

B(t)= Pas (t) may fas: 2n(t) >
u'(t)= ua
bis(@) bia(O\e seein ui,(t)

respectively. Similarly, the matrix B(t) or the vector u(t) is integrable on an

interval (c, d) if and only ifeach of its entries is integrable on the interval (c, d).
Their integrals are given by
foun fain - fou

faca [bata feo . feaina
Rent ere
ate
feta fina
Exercises
2. Evaluate the derivatives of each of the following vectors or matrices:
t Ca em
a) B(t)=|sint 0 cost for —o<t<o.
i? t 1
b) B(t)= ae nh] for —co<t<0oo.

= SiliE a GOST
2t
) BO= |es te™

92 (241)& for
ae —a0<t<a
logt
d) u-| tlogt | for 0<t<oo (and where logt is the natural
t? logt logarithm of f).
ia} ~ ll
—} —_ for —l<t<2.
eee ae
ee
BwWN
3. Evaluate {{ B(t) dt or {{ u(t) dt for each of the matrices B(¢) or vectors u(t) in Exercise
1. [Hint: In parts (c) and (d), integrate by parts.]
4. Is the vector
w(=||
continuous on the interval 1<r<2?
Is it continuous on the interval —1<t<1? Explain.
We are now ready to say what is meant by a solution of system (4.8).
Definition 3. Let A(t) be a continuous nxn matrix on an interval ¥. Let

g(t) be a continuous vector with n components on the same interval ¥. A
solution of the system
y =A(t) y+g(¢) (4.8)

on some interval J (where Y is contained in ¥) is a vector u(t) whose derivative
u’(t) is continuous on the interval J and such that
u’(t)= A(t) u(t)

+8(¢)
for every ton J. (Note at this point that the interval Y is not necessarily the
same as ¥.)
Example 2. Consider the scalar (n= 1) differential equation y’=—y+1. Then u(t)=
=e ‘+1 is a solution on the interval —oo<t<oo. For u(t) is continuous on
—0<t<oo and u’(t)=—e'. Thus u’(t)= —e '=—u(t)+1.
Example 3. Show that
Clearly (t) is differentiable for — oo <t<oo (because e’ is) and
o'()= |
On the other hand, by matrix vector multiplication,
Thus
o'(t)=A(t) O(t), —cao<t<o.
In Chapters | and 2, we saw that with a differential equation, one usually

associates a particular initial condition. For example, the solution u(t)
=e ‘+1 of Example 2 satisfies at t=0 the initial condition u(0)=2. More
generally, suppose we consider the system (4.8) together with the initial con-
dition y(t))=Yo, where fy is a given number in the interval . and where
Yo IS a given vector in n-dimensional Euclidean space.
Definition 4._ By a solution of the initial value problem,
y=A(t)y+g(t), y(to)=Yo. (4.9)

we mean a solution u(t) of the system y'=A(t) y+g(t) on an interval Y con-
taining the point to, such that u(to)=Yo.
Our object will be to learn as much as possible about such initial value
problems. As a matter of fact, when n=1 the initial value problem (4.9) can
always be solved, and we have already obtained a formula for the solution.
(See Section 1.4, Theorem 1.) Unfortunately, for n>2 the situation is much
more complicated.
Example 4. Show that the vector
is a solution of the system

Ol Vi
y ‘— & ale where
h y— B 5
on —0<!t<oo, satisfying the initial condition
Obviously,
Since cos f and sin ¢ have continuous derivatives everywhere, we have
Vivo P= sint || us (6) | eae u,(t)}_| 9 1 ;

¥ =| eee |=| a Li Gi ucrenlenenimo
Exercises
5. Show that
ial ea
is a solution of the system of Example 4 on —«<t<oo satisfying the initial
condition
6. Show that
w(t)=c,u(t)+c,v(t),
where u(t), v(t) are given in Example 5 and Exercise 5, respectively, and where
C1, Cy are any constants, is a solution of the initial-value problem
y’ -|NS 4y, v(0)=| &]

for —co <t<oo.
7. Show that
vid=|_<|
is a solution of the initial-value problem
on the interval — co <t<o.
If the linear system (4.8) has a very special form, it can be solved com-
pletely. We illustrate this with the following examples and exercises.
Example 5. 1) Solve the initial-value problem (4.9) with n=2,
a=|6 n g(t)=0
where d,, d, are constants.
Here the system (4.9) has the very simple form
Yi=dy1, Vi(to)=Yor
; 4.10
Y2=adzy2, V2(to) Be ( )
in which the differential equations are not linked to one another and each can be solved
separately. By separating variables (see, for example, Example | and Exercise 1, Section
1.1) we have that ¢,(t)=e""~" yo, is the solution of the first equation for
—o<t<oo and ¢,(t)=e"""™ yo, is the solution of the second equation for
—0o<t<oo. Thus
_| P:(t) |_| You expd, (t—to)

= [Fale expd, = 7s
is the solution of (4.10).
ii) Solve the initial-value problem (4.9) where A(t) is the n xn diagonal matrix
d, 0
A(t)= and g(t)=0,
and where d,, d), ..., d, are constants. It is clear that the jth equation of this
system is simply
yj=diy;, Yj(to)=Yoj-
Its solution (by the same method as in part (i)) is given by
$j (t)=Yo; expd;(t—to), —0o<t<o.
Thus
o(t)=| . =|". =o=<t<0.

Pn (t) Yon XP d,(t —to)
iii) Solve the initial-value problem (4.9) if A(t) is the matrix in part (ii) and g(¢) is any
continuous vector function g(t) on —0<t<oo. Here the jth equation is
y,;=djy;+9,(t), Y;(to)=Yoj
which is a linear first-order scalar differential equation. Thus by Theorem 1,

Section 1.4, especially Eq. (1.23), with p(t)=d; and q(t)=g,(t), we have
y(t) Yo; expdj(t -193[eats)] gj(s) ds

and
rs
You €xpd, (t—fo)+ |[expd,(t—s)] g(s) ds

to
Yoo €xpd,(t—to)+ | [exp d2(t—s)] g(s) ds
Yon exp d,,(t—to) + |[exp d,(t—s)] g,,(s) ds

to
Exercises
8. Find a solution @ of the initial-value problem
yi=-)1, 2
; j= J;
Wy Vir vas ¥(0) fil
[Hint: Solve the first equation and substitute in the second equation. Solve this
equation by using Theorem 1, Section 14. What is the interval of validity?]
9. Find a solution @ of the initial-value problem
W=—-N>
; )) = 1
Vos Vurty2: ¥(0) H
10. Describe a method for solving the “‘triangular system”
Vr = 11 FA 2y27 + inYn
y2= A22y2 1°": + Q2nYn
a= = Oeil. ile
Vie GnnYn>
where q;;, with j>i, are constants; note that a;;, with j<i, are zero.
11. Find a solution of the ¢ of the initial-value problem
Y¥i=yityotf(d,t 0
0)=] , |;
Vo= Vi Vas (0) 0
where f(t) is a continuous function. [Hint: Define v(t)=y, (t)+y2(d.]
Exercises 8, 9, and 10 show that “triangular” systems of first-order

differential equations can be solved by successive solution of scalar first-order
equations. However, a system that is not triangular cannot, in general, be

solved by putting it in triangular form as one might first guess, because the
elementary row operations that triangulate the right-hand side (that is, the
coefficient matrix A) destroy the isolation of the derivatives on the left-hand
side. Note that this difficulty does not arise in the case of linear algebraic
equations.
The examples in Section 2.1, particularly Equation (2.11), lead us to a
specific linear second-order scalar differential equation with initial con-
ditions. These can be reduced to an initial-value problem for a linear system
oftwo first-order equations ofthe form (4.9) by the following method.
Example 6. Show that the scalar second-order linear initial value
y'+p()y
+a) y=r(),
¥ — vto)=m, — ¥'(to)= M2, (4.11)
where p, g, r are given functions continuous on an interval ¥, fp is in .%, and y,, n, are
given constants, can be reduced to a system of the form (4.9).
In agreement with Definition 4, by a solution of (4.11) on an interval Y contained
in #% we mean a function w(t) such that w’(t), w(t) exist and are continuous at each
point of .¥, such that w”(t)+ p(t) w’(t)+ q(t) w(t)=r(t) for every ¢ in ¥, and such that
w(to)=1, W’ (to) =N2- The idea is to introduce new unknowns y, and y, by means of
the definitions y,;=y, y,=y’. Then
ys=y"= —p(t) y'—4(t) y+ r()= —p(t) y2—-4()v1+ (0).

This suggests that the given initial-value problem (4.11) can be described by the initial-
value problem
ve
|| it |
we Ee
0 I 0
=| 10 -n0|"*[x0}
_
[3]
where
|| ahi
Note that (4.12) is a special case of (4.9) with n=2 and A(t), g(t) displayed in (4.12).
We will now show the initial-value problems (4.11) and (4.12) are equivalent in the
sense that given a solution of either one, we can construct a solution of the other one.
More precisely, let y(t) be a solution of (4.11) on some interval Y containing fy. Define
the functions ¢, and ¢, on Y by the relations
di(N=V(), G2 (=w'(0).
Define the vector by the relations
v-|
ooh
We claim that (ft) is a solution of (6.13) on Y. Clearly,
#0 [tol Lyal Ll
moreover,
for every t on Y, which shows that (¢) is a solution of (4.12) on Y. Conversely, let u(¢)
be a solution of (4.12) on some interval Y containing fy. Let
and define the function w by w(t)=u,(¢). We claim that w is a solution of (4.11) on J.

From the first equation of (4.12), we have w’(t)=u;(t)=u,(t), and from the second
equation in (4.12), and the definition of w, we have
w"(t)=u3(t)= —q(t) us ()—P(t) u(t) +r(0)
= —9(t) w(t)—p(t) w(t) +r(2).
Hence, w"(t)+p(t) w'(t)+¢(¢) w(t}=r(t) on Y. We also have w(to)=u,(to)=7,, and
w' (to) =U (to)=n2. Hence, w is a solution of (4.11).
The reader will note that the initial-value problem in Example 4 is

equivalent to the scalar second-order differential equation y’’+y=0 with
initial conditions y(0)=1, y’(0)=0. (Compare with (4.11) and (4.12).)
Example 7. More generally, the scalar nth-order linear initial-value problem
y+py(t)yO +---+p,-1(t) ¥ +P,(t)y=r(t), 4.13

; ‘
Y(to)= Y(to)N1,
=N2,---. YX (to) =, oy)
where /,, P2,.--, P, are given functions continuous on an interval ¥, fy is in %, and
nee H, are given constants is equivalent to the system
0 1 0 0
0 Om 1 0 : p
y= Roy eyes ee
0 0 0 1 (.)
Pal’) —Pr-ilt) - —prlt) —pi(t) ;
ny
vo)=| | (4.14)
where
Vy yi
Wi==|| 8 lhe y= || 2 |.
Vn Vn
Put
Way = J), y2 = y’, v3 = Wee teey Yn ae y" =2) .
Then
\ i y — V2
Y2= a V3
Ya-1 =" Y=,

y= Y= —P,(1)Vi Pn () Yo P(t) yy,ed)
and
V1 (to) =Y(to)=M1, Y2(to)=Y

(to) =N25--+5 Vato) ="? (to) =Mn-
This suggests that the given initial-value problem (4.13) should be equivalent to the
system
yi y2
y= Y2 = J3
Yn —Pn(t) ¥1 —Pn-1(t) Y2—**-Pi(t)

Yn+1(C)
0 it © 0 0
0 OP iG 0 0
0 0 0 1 0
— p,(t) —p2(t) —pr(t) r(t)
11
y(to)=n= |.”
"n
Indeed, similar to Example 6, let /(t) be any solution of (4.13) on some interval Y con-
taining fo. By this we mean that Wy, Wy), .... Wi) exist in Y, are continuous and satisfy
the differential equation (4.13) while W(to)=11, W'(to)=N2,.... VW"(to)=n,. Define
the functions ¢,, $2,..., @, on Y by the relations
bi(th=W(t), d2(=W'(d),-- P(Q=W" (0):

Define
Clearly @(to)=1. Moreover, exactly as in Example 5, we have
4 (t) w(t) 2(t)

p(t) y(t) 3(t)
O()=| : : =
dni} fw" P(t) ,(t)
o,(t) y(t) —pil) ¥ *()—--—p,(t) YO +r)
a(t)
p3(t)
,(t)
—Prlt) d1()—- — Pit) $n ()+r(d)
0 16 0 $1 (t) 0
0 0 1 0 0 $2(t) 0
= : s +| :
0 ig ose 0 1 hy —1(t) 0
=p; (t) tee —p2(t) —p,(t) $n (t) r(t)
which shows that this particular @ is a solution of (4.14). Conversely let u(r) be any
solution of (4.14) on ¥. Define w(t) =u, (¢) (i.e., the first component of u). We claim that
this function w is a solution of (4.13) on Y. This proof is also very similar to the
special case n=2 carried out in Example 6 and we shall leave it to the reader as an
exercise.
We remark that whereas every nth-order scalar equation is equivalent toa
system of first-order equations (as shown in Example 7), the converse is not
ik Le]
true. For example, the system
es | 1 0 > Wa
cannot be made equivalent to a second-order scalar equation, because the

coefficient matrix has the wrong form.
Exercises
12. For each of the following initial-value problems, write an equivalent initial-value
problem for a first-order system:
a) y’+2y'+7ty=e', y(I=7, y(=—2

b) 2y”—5St?y’ +(cost) y=logt, y(2)=1, y’(2)=0
c) y”—6y"+3y'
+e ‘y=sint, y(0)=0, y’'(0)=0, y’(0)=0
d) y+ 16y=te', y(0)=1, y(0)= —1, y’(0)=2, y’”(0)=0
13. Reduce each of the following initial-value problems to an equivalent initial-value

Pk A g
a) y’+— y=— cost b) 6”+— 6=0
m m ib
y(0)=yo, (0) =0 0(0)=0o, 0'(0)=0

14. Reduce each of the following initial-value problems to an equivalent initial-value
a) y’+5z’-7y+6z=e'
z” —2z+13z'—15y=cost,
where
[Hint: Let w,=y, w.=y', W3=Z, Wa=Zz".]
b) y+5z+2y=0?
z” +6y' + 11z’—3y—z=t,
where
In Exercise 13 above we have seen how the equations for the simple
mass-spring system (Section 2.1) and for the linearized pendulum can be
reduced to initial-value problems for linear systems of first-order equations
of the form (4.9). To close this introductory section we shall consider two
more examples of more complicated physical systems and show how they
also lead to an initial-value problem of the form (4.9).
Example 8. A weight of mass m, is connected to a rigid wall by a spring having
spring constant k, >0. A second weight of mass m, is connected to the weight of mass
m, by means of a spring having spring constant k,>0. An external face F(t) is applied
to the second weight. The whole system slides in a straight line on a frictionless
table, as shown in Fig. 4.1. Let y,(t) denote the displacement of the first weight from
its rest position (equilibrium) and y(t) the displacement of the second weight from
equilibrium. At equilibrium y,=y,=0 and both springs are unstretched.
Ye
-
NN
m, =
ETT
m; / F(t)
7
yi: =0, y,(t) y2=

0, y2(t)
Figure 4.1
a) If at time t=0 the system starts from rest with initial displacements y,(0)= 1,
¥2(0)=y29, determine the motion of the system.
b) If m,=m,=m, ki =k,=k, and F(t)=0, show that the motion ofthe system is
a superposition of two simple harmonic motions with natural frequencies
Lessr(y” aw Le EN"
In order to formulate this problem mathematically, we assume (i)-(v) of Section 2.1
and apply Newton’s second law of motion to each moving weight which we regard as
a point mass. Suppose that at time ¢ the system is in the position shown in Fig. 1.5;
then the only forces acting on the particle of mass m, are the restoring force of the first
spring —k,y,(¢) (by Hooke’s law), and the restoring force of the second spring whose
net extension is y,—y, (because it is stretched y, units by the second weight and
compressed y, units by the first weight). Thus, the restoring force of the second spring
is k,[y2(t)—y,(t)]. Applying Newton’s second law to the particle of mass m,, we there-
fore obtain
mi (t= —kyy1 ()+k2(y2(t)—y1 (0). (4.15)
The only forces acting on the particle of mass m, are the restoring force
—k,[y2(t)—y,(t)] of the second spring whose net extension is y,—y, units, and the
external force F(t). Thus, Newton’s second law applied to the second particle yields
may2(t)= —k2Ly2()—yi (J+ F()- (4.16)

From the statement of the problem we also have the initial conditions y,(0)=y4o,
y1(0)=0, y3(0)=y20, = ¥2(0)=0. Equations (4.15) and (4.16) govern the motion and the
mathematical problem is to find a pair of functions ¢,, $2 defined for t>0, such that
1, 1, 5, 05 exist for t>0, such that
ie aa at RAR i $,(t)]
m23(t)= —k2,[b2()—- bi (0 + F(d,
for every t>0 and such that
$:(0)=Y1o, $1 0)=0, 62(0)=y20, (0) =0.
If such functions have been found (it will be seen later (Section 4.2) that there is only
one such pair of functions), we say that we have found a solution of the initial-value
problem consisting of the system of linear differential equations of second order
myi=—kiyy +k2(yv2—-y1)
oe ~ka(v2—Ya) + F (0)
a
i)
4.17
and the initial conditions
Yi)=Yio, —y2(0)=Y20, ¥i(0)=9, —_y2(0). (4.18)

In order to determine the motion and answer question (b), we will have to learn
how to solve the initial-value problems (4.17) and (4.18). This will be done in Chapter 5
(see Exercise 6, Section 5.7).
Exercise
15. Find a system of first order equations of the form (4.9) equivalent to (4.17), (4.18).
(Hint: Put w;=y1, W2=Yi, W3=Y2, We =); then w, =Wy, etc.)
Closely related to mechanical systems are electrical systems that also

lead to linear systems of ordinary differential equations.
Example 9. Consider the electrical circuit shown in Fig. 4.2, in which the known
time-varying source current i,(¢) is connected to nodes A and C; in the circuit v, (¢) is
the variable voltage (unknown) across the 5/3-farad condenser, i, (¢) is the variable cur-
rent through the 3/5-henry inductor. The polarities are indicated in Fig. 4.2. When
v,(t)>0, the potential of node A is larger than that of node C (measured with respect
to a common reference). When i, (t)>0, the current flows from node A to node B.
Suppose that at t=0 we are given
v,(0)=0.6 volt, i,(0)=1.0 ampere, v>(0)=1.2 volts.
A 3/5 henry Borge
54 farads
Figure 4.2
It is desired to determine the voltages v,(t), v2(¢), and the current i,(t) as functions of
time (in terms of the given source current i,(f)).
To set this problem up mathematically, we shall use Ohm’s law (v=iR, relating
voltage, current, and resistance) and the formulas i=Cv’'(t) and Li'(t) (relating the
“current through” to the ‘‘voltage across” condensers and inductors; here C is the
capacitance, L is the inductance, and '=d/dt). We shall also need Kirchhoff’s law of
currents: The sum of the currents entering and leaving a given node is zero.
Suppose that at a time f the source current i,(t) leaves node C and enters node A
as shown in Fig. 4.2. At the same time the current i, leaves node A (through the
inductance of 3/5 henry) and a current (5/3) vi(t) leaves node A through the
capacitance of 5/3 farads. Thus, by Kirchhoff’s law applied to node A we have
i,(t)—i, (t)—3v;, ()=0. (4.19)

Similarly, a current i, enters node B, while currents §v}(t) and v,(¢)/1 leave node B
through the 4-farad condenser and 1l-ohm resistor, respectively. Thus, Kirchhoff’s
law applied to node B gives
iy (t)—§v2(t)—
0, (t)=0. (4.20)
Then in the middle loop of the circuit shown in Fig. 4.2, since the sum of the voltage
drops must also be zero (another one of Kirchhoffs laws), we have
Solving each Eqs. (4.19), (4.20), and (4.21) for the quantities vj, 7;, 02, respectively, we
see that the mathematical problem is to solve the initial-value problem

v= 31, +3i,(1), v0) =0.6 volt
i, =30, —20, i, (0)=1 ampere (4.22)
v2 =6i, —6v2, v(0)= 1.2 volts
for the unknown functions 0, i;, v2, where i,(t) is a given source current, that is, a
given function defined for 0<1<oo. This initial-value problem consists of a system of -
three linear first-order differential equations and it is clearly of the form (4.9). It will be
solved completely in Section 5.7.
4.2 THE EXISTENCE AND UNIQUENESS THEOREM

In discussing the scalar first-order differential equation (Theorem 1, Section
1.4), we were able to give an explicit expression for the solution of the initial-
value problem. However, for second- or higher-order scalar differential
equations and more generally for first-order systems, it is frequently not
possible to give an explicit expression for the solution. For example, the
scalar differential equation
t?y" +ty’+(t?
—p?) y=0,
where p is a constant, is called the Bessel equation (of index p), and arises in
many problems of mathematical physics. This equation and its solutions
have been studied extensively. Except for special values of p, such as p= or
p=3, these solutions cannot be expressed in terms of a finite number of
elementary functions. Nevertheless, the Bessel equation does have solutions
for every initial-value problem with t, #0. This fact is a special case of the
following genera: theorem.
Theorem 1. Let A(t) be a continuous nxn matrix on some interval %. Let
g(t) be a vector with n components continuous on the same interval %. Then
for every to in ¥% and every constant vector n, the initial-value problem
y=A(t)y+g(t), y(to)=n (4.9)
has a unique solution existing on the same interval F¥.
You must refer to Section 8.6 for the proof. Our objective here will be
to learn to apply the theorem. In Theorem 1, the matrix A(t) and the vector
g(t) may have real- or complex-valued entries.
Example 1. Let n=3 and consider the initial-value problem (4.9) with
1 —t 0
zak 7 1 e I
A(t)=|- t?—-1 _ (t)=
g(t)=| cost =0,
to=0, N=n=| 0
1 —é —1
t?+1
Determine whether this initial-value problem has a unique solution and find the largest
interval ¥ of existence of this solution in accordance with the theorem.
4.2 The Existence and Uniqueness Theorem 131
The entries of A(¢) and g(t), with the exception of 1/(t?—1), are continuous on
—0<t<oo. However, 1/(t? — 1) fails to be continuous at = +1. Since tp=0, Theorem
1, therefore, tells us that the given initial-value problem has a unique solution
((0)=n), and the solution @ exists on the interval —1</<1. It is worth pointing
out that if we choose a different ¢), for example tg=10, the new initial-value
problem will also have a unique solution W(W(10)=y) and the solution W will exist for
<=60;
We now apply Theorem | to the important special case of the initial-

value problem for a linear second-order scalar equation
v' tp) y' +a var, v(to)=m1, vy"(to)=N2- (4.11)
where p, g, and r are functions continuous on some interval .% and fy is a
point of .¥. In Example 6, Section 4.1, we have seen that the initial-value
problem (4.11) is equivalent to the initial-value problem
v= [40 alt al w=[R] en

to which Theorem ! can be applied directly. The requirement that A (¢) and
g(t) be continuous on | translates into the following result.
Corollary I to Theorem 1. Let p, q, r be given functions continuous on an

interval % and let ty be in ¥. Then the initial-value problem (4.11) has a unique
solution w(w(to)="1, W'(to)=N2) that exists on the same interval ¥.
Readers who studied Chapter 3 should note that this is precisely Theo-
rem 1, Section 3.1.
Example 2. Consider the initial-value problem

(7 +4) y’+ty'+(sint)y=1, y(I)=2, y(1)=0.
Determine the existence and uniqueness of the solution as well as the interval of
existence.
To apply Corollary 1, we must first reduce the given differential equation to the exact
form of(4.11). This is accomplished by dividing by (¢? +4). Thus, in the notation of the
corollar sint l
‘ pi=s- 5, 10 apearr Et) Saecenaee
Since these functions are continuous for — 0 <t< oo, the given initial-value problem has
a unique solution existing on — 0 <f<oo.
Similar to Corollary 1, we may consider the more general initial-value

problem
y+ pi(t) YY +--+ Par(8)¥' + alt) Y=") (4.13)
Y(to)=n1, (to) =N29-- YP (to) =n |
for the scalar nth-order linear equation. From Example 7, Section 4.1, and
from Theorem | we obtain immediately the following.
Corollary 2 to Theorem 1. Let p,,P2,---,Pn be given functions continuous

on an interval ¥ and let ty be in ¥. Then the initial-value problem (4.13) has
a unique solution w (w(to)=M1, W' (to) =N2,---. W"” (to) =N,) that exists on
the same interval #.
Note that, unlike the study of linear systems of algebraic equations, we
consider only systems in which the number of equations and the number of”
unknown functions is the same. In Chapters | and 2, we observed that phys-
ical problems should lead to mathematical formulations of differential
equations with exactly one solution. For a system of first-order linear
differential equations, if the number of equations is not the same as the
number of unknown functions, the initial-value problem does not in general
have a unique solution. For example, consider the initial-value problem
¥i=Y2 Vi)=1, y(0)=1,

that is, one equation in two unknown functions. Then
is a solution, but so is
vo=["7"]
In fact, there are infinitely many solutions, for we can add an arbitrary
differential equation for y,,
yp =a(t) yi + b(t) v2 +9(0).
where a, 6, and g are arbitrary continuous finctions. With this additional
differential equation, the initial-value problem will have a unique solution
by Theorem 1.
From this example, you can see that, similar to the situation for linear
algebraic systems, an initial-value problem for a system of linear differential
equations with fewer equations than unknowns may have infinitely many
solutions. Unlike the algebraic case, however, such a problem may have no
solution, as is shown by the example y,+y3+y5=1, yi +y5+y3=0,
y,(0)=y2(0)=y3(0)=0. If there are more equations than unknowns, the
system may be inconsistent. For example,
yi —
val yi ve (0) 1
is obviously inconsistent.
Exercises
1, What does Theorem | tell you about each of the following initial-value problems for
systems of the form (4.9)?
The Existence and Uniqueness Theorem 133
3 t logt e' |
4=| | Ta 2(=[5 | n=|_4|
b) Same as part (a), except tg = —1
c) Same as part (a), except fp =0
d) Same as part (a), except
g(t)=| 1
t?—9
e) 2=3, fh=5
1 cost —sint 0 1
AG)=| err ete g(t)=|0 q=|-1
0 e t 0
f) Same as part (e), except
0
n=| 0
0
Can you guess the solution?
. What does Corollary 1 to Theorem | tell you about each of the following initial-
value problems for linear, scalar second-order equations of the form (4.11)?
a) y’+ry’—y=0,
y(I)=1, y'(1)=1.
b) Same as part (a), except y(1)=1, y’(1) unspecified.
c) Same as part (a), except y(1)=0, y’(1)=0. (Can you guess the solution?)
d) ty’ +ty'+(t?-4)
y=0, y(—1)=0, y'(—1)=1.
e) Same as part (d), except y(0)=0, y’(0)=1.
. What does Theorem | tell you about each of the initial-value problems of Exercise
14, Section 4.1?
. Discuss, using Corollary 2 to Theorem 1, each of the following initial-value
problems:
a) y"+ry'+(tan t) y=2t, y(0)=0, y'(0)=9, y"(O)=1.
b) x—ey=1, y(0)=0, v'(0)=0,
c) yp” +(t?—1
y=0,)'?
y(—1=1,
d) y"+(P=1)!? y=0, y(2)=0, y'(2)=0, y’(2)=1
. Prove Corollary 2 of Theorem 1.
6. For each of the following differential equations, determine the largest intervals on
which a unique solution is certain to exist as an application of Corollary | of
Theorem. In each case it is assumed that you are given initial conditions of the form
P(to)=Yo (to) =Zo- ;
a) ty" +y=0?
t?(t—3) y"+y'=0
Veet y=0
(1+27) y’—y’+ty=cos t
e) e'y”—(sin t) yy’t+y=0
f) y"—(log |e) y=0
4.3 LINEAR HOMOGENEOUS SYSTEMS
We shall study the structure of solutions of
y =A(t) y+g(t). (4.23)
If g(t)#0, the system (4.23) is said to be nonhomogeneous. With every

nonhomogeneous system (4.23) we associate the corresponding homogeneous
linear system
y =A(t)y (4.24)
This section is devoted to the study of the algebraic structure of the set of all
solutions of (4.24). Here we are concerned not with a specific initial-value
problem for (4.24), but rather with the algebraic structure of the set of all
solutions. We assume, throughout, that the matrix A(t) is continuous on
some fixed interval %, which may be finite or infinite. The entries q;;(t) of the
matrix A(t) can be real or complex-valued.
By Definition 3, Section 4.1, a solution of the system (4.24) is a vector
u(¢) whose derivative u’(¢) is continuous on .¥. In the language of linear
algebra, this means that a solution of the system (4.24) is an element of the
vector space of functions with n components, real- or complex-valued, and
having continuous first derivatives on the interval %. We shall call this vector
space C, (4). (Here C stands for continuity, ’ stands for the derivative, the
subscript v is there because each vector has n components, and .¥ represents
the interval under consideration.)
We remind the reader of the following important facts about vector
spaces (see for example [3], Chapter 5). Let ¥ represent the real or complex
numbers.
Definition 1. A vector space V over F¥ isa collection of elements called vectors
for which two operations, called addition and scalar multiplication, are defined:
For every pair of vectors u, VE V there is a unique vector u+veV called
the sum of wand vy. For every vector ue V and every scalar we F, there is a
unique vector awe V called the product of « and u.
Addition and scalar multiplication satisfy the following properties:
(A,) (u+v)+w=u+(v+w)
for all vectors u, v, we V.
4.3 Linear Homogeneous Systems 135
(A,) There exists a unique vector in V denoted by 0 such that u+0=u

for allue V.
(A3) For every ueV there is a unique vector v such that u+v=0: v is
denoted by —u and is called the additive inverse.
(A,) u+v=v-+u
for all vectors u, ve V.
(S,) For all xeF and u, ve V, «(u+v)=ou+av.
(S.) For all «, BEF and ueV, (4+f) u=au- pu.
(S3) For alla, BEF and ue V, (4B) u=a(Bu).
(S4) For allueV, lu=u.
Definition 2. Let V be a vector space over ¥. A subset W of V(WSV) is
called a subspace of V ifand only if W is itselfa vector space over F with the
same operations of addition and multiplication by scalars as in V.
A simple criterion for testing whether a subset of V is a subspace is the
following. A subset W of a vector space V is a subspace ifand only if ou+ Bye W
for every u,veW and for every a, BEF (i.e. every linear combination of
vectors in W is again in W).
Let u(t) and v(t) be any two solutions of (4.24) on the interval %. Let «
and f be arbitrary real or complex constants. We claim that au(t)+ Bv(t) is
also a solution of (4.24) on ¥. For, by the rules of vector calculus discussed
in Section 4.1 and the fact that u(t) and v(t) are solutions of (4.24),
[ou (1) + By (t)}’ =u’ (7)+ Bv'(2)

=A (t) u(t)+ BA(t) v(t)
= A(t)[ou(t)
+Bv(s)].
Since every linear combination of solutions of (4.24) is again a solution, the
set V of all solutions of (4.24) is a subspace of the vector space C,(/¥).
It is easily verified that for every integer k>0 the following vectors
(with n components) in the space C, (4)
1 t te f*
0 0 0 0
OVE 4:0: I, E
0 0 0 0
are linearly independent. Since k is an arbitrary positive integer, it follows

easily that the dimension of C,(%) (i.e. the number of linearly independent
vectors in C,/(.4%) which span C,(¥4)) cannot be finite. This fact emphasizes
the importance of the problem of finding the dimension of the vector space V
of solutions of (4.24) which is a subspace of C,(.%). We have the answer in
the following basic results.
Theorem 1. If the complex nxn matrix A(t) is continuous on an interval

SF, then the solutions of the system
y =A(t)y (4.24)
on .£ form a vector space V of dimension n over the complex numbers.
In view ofthe remarks preceding the statement ofthe theorem, it is signifi-
cant that, according to Theorem |, to find any solution of (4.24) it suffices
to find a finite number of solutions, namely, a set that forms a basis for the
vector space V.
Proof of Theorem |. We have already established that the solutions form

a vector space V over the complex numbers. To establish that the dimension
of V is n, we need to construct a basis for V consisting of7 linearly indepen-
dent vectors in V, that is, of m linearly independent solutions of (4.24) on ¥.
We proceed as follows. Let tp be any point of¥ and let 6,, 65,..., 6, be any
n linearly independent points (vectors) in complex Euclidean n-space, (for
example, e€,, €2,..., @,5
are obviously n such vectors.) By Theorem 1, Section 4.2, the system (4.24)
possesses n solutions @,, @5,..., @,, each of which exists on the entire interval
#, and each solution @; satisfies the initial condition
(to) =9;, p=, Daven (4.25)
We first show that the solutions ,, @5,..., ,, are linearly independent on

SF. Recall that this involves examination of linear combinations of vector
functions, but with scalar (constant) coefficients. Suppose there exist complex
constants A,, A3,..., a,, Such that
a1, (t)+.a,2(t)+---+a,6,(t)=0 for every t on ¥.

In particular, putting t=fo, and using the initial conditions (4.25), we have
a,6,+a,0,+-:-+a,6,=0.
But this implies that a,, a,..., a, are all zero because of the assumed linear
independence of the given vectors ¢,, ¢2,...,0,. Thus, 1, @5,..., @, are
linearly independent on .¥.
To complete the proof we must show that these n linearly independent
solutions of (4.24) span V; that is, they have the property that any solution
W(t) of (4.24) can be expressed as a linear combination of the solutions
;. D2,..., @,. We proceed as follows. Compute the value of the solution
W at fy and let (to) =o. Since the constant vectors 6,, 6>,..., 6, form a basis
for complex Euclidean n-space, there exist unique constants c,, C,..., € n
such that the constant vector o can be represented as
O6=C,6, +C,0,+-::+C,6,,.
Now, consider the vector
o(t)=c1o, (t) + c22(t)+ ‘ee +c, (t) ,
Clearly, (ft) is a solution of (4.24) on %. (Why? Prove this.) Moreover, the

initial value of @ is (using 4.25)
O(to)=c,6, +€,6,+---+¢,6,=6.
Therefore, (t) and W(s) are both solutions of (4.24) on ¥% with (to)
=W (to)
=o. Thus, by the uniqueness part of Theorem 1, Section 4.2, @(¢)=w(d) for
every ton ¥, and the solution w(t) is expressed as the unique linear combina-
tion
Wl) =r; (+ e262 (1) +- + ey6,(t) for everyron 4. (4.26)

Exercise
1. Show that this expression of(¢) as a linear combination of , (4), ..., @,(¢) is unique.
[Hint: Assume (t)=d, $,(t)+---+d, ,(¢ in addition to (4.26) and show that
d,=c,, where j=1,..., n.]
Thus, we have shown that the solutions @,, @,..., o, of (4.24) span the
vector space V. Since they are also linearly independent, they form a basis
for the solution space V, and the dimension of V is n. This completes the
proof of Theorem 1. I
We often say that the linearly independent solutions @,,..., @, form a
fundamental set of solutions. There are clearly infinitely many different funda-
mental sets of solutions of (4.24), namely, one corresponding to every basis
6,,..., 6, of Euclidean n-space.
Exercise
2. Prove the following analog of Theorem | for systems with real coefficients. If the
real nxn matrix A(t) is continuous on an interval Y, then the real solutions of
(4.16) on .% form a vector space of dimension n over the real numbers. [Hint: This
is not a trick question; just check that the proof of Theorem | applies here.]
When we wish to apply Theorem |, it is useful to restate the result in the

following manner: The system (4.24) possesses n linearly independent solutions
on the interval %. Moreover, every solution @(t) of (4.24) on ¥ can be ex-

pressed as a unique linear combination of those n solutions. In practice, this
means that it suffices to find, in any manner, n solutions of (4.24) and show
that they are linearly independent. We shall devote considerable attention to
this in special cases later in this chapter. Unfortunately even in special cases
this is not a trivial problem, because there does not exist a procedure for-
finding a basis for the solution space V in the completely general case.
We now apply Theorem | to the scalar linear homogeneous second-order
y"+p(t) y+q(t) y=0, (4.27)

where p and qg are continuous functions on the interval #.
Corollary 1to Theorem 1. Let p and q be continuous on the interval %.
Then equation (4.27) possesses two linearly independent solutions w, (t), W(t) on
the interval %. Moreover, if w(t) is any solution of (4.27) on ¥, then there
exist unique constants C,, C, such that
W(th=c,W,(t)+c.W,(t) foreverytin ¥.
Proof. By the method of Example 6, Section 4.1, the scalar equation (4.27)
is equivalent to the linear system
rl a0 vo] Ln ae
which is a special case of (4.24). By Theorem 1, there exist two linearly
independent (vector) solutions @, (t), @2(¢) of (4.23) such that every solution
o(¢) of (4.28) has the form (t)=c,,(t)+c,6,(t). By the equivalence of
(4.27) and (4.28),
wo-[20} e0-[20]
and w,(t), w2(t) are solutions of (4.27) on .%. We know that the vector
solutions @,(t) and ¢,(t) are linearly independent on .¥. We wish to show
that w,(¢) and (2) are linearly independent on ¥%. Suppose that c,W;, (t)
+ ¢,/2(t)=0 for every t in %, then cy} (t)+c.W5(t)=0 for every ¢ in ¥.
Thus, c,, (¢)+c22(t)=0 on ¥. Since @, (t), 2 (¢) are linearly independent
on ¥%, c,=0, c,=0. Therefore w,(/ and w,(t) are linearly independent
on ¥. Also, every solution w(t) of (4.27) is the first component of the corres-
ponding vector solution o(t) of the system (4.28). Since (t) has the form
(1)=c1, (1)+cr> (1),W(t) has the form W()=c,W(t) +c2Wo(d). |
The same reasoning applied to the scalar linear homogeneous differential

equation of order n gives the following result.
Corollary 2 to Theorem 1. Let p,,...,~, be continuous on the interval ¥.

Then the differential equation
y +p, (t) y"" ++ +p,(t)y=0 (4.29)

possesses n linearly independent solutions W,(t),..., W,(t) on the interval JF.
Moreover, if y(t) is any solution of (4.29) on ¥, then there exist unique constants
SES c, such that (th=cyW, (t)+ coo (+--+: +¢,W, (0).
Exercise
3. Prove Corollary 2 to Theorem 1.
We can interpret Theorem | ina different and useful way. A matrix ofn
rows whose columns are solutions of (4.24) is called a solution matrix. Now,
if we form anv X n matrix using n linearly independent solutions as columns,
we will have a solution matrix on ¥ and also its columns will be linearly
independent on .¥. A solution matrix whose columns are linearly indepen-
dent on &¥ is called a fundamental matrix for (4.24) on ¥.... Let us denote
the fundamental matrix formed from the solutions @,, 5, ..., @,, as columns
by ®. Then the statement that every solution w is the linear combination
(4.26) for some unique choice of the constants c,,..., c, is simply that
W(t)= P(t) c, (4.30)
where @ is the fundamental matrix constructed above and ¢ is the column
vector with components ¢,,..., c,. (The vector &(t) ¢ is obtained by forming
the linear combination of columns of ®(t) with c,,...,c, as coefficients.) It
is clear that if ® is any other fundamental matrix of (4.24) in Y, then the above
solution W can be expressed as
W(t)=P(t)€ for every ton ¥
for a suitably chosen constant vector é. Clearly, every solution of (4.24) on
can be expressed in this form by using any fundamental matrix.
We see from the discussion above that to find any solution of (4.24) we
need to find a fundamental matrix. A natural question, then, is the following.
Suppose we have found a solution matrix of (4.24) on some interval ¥; can
we test in some simple way whether this solution matrix is a fundamental
matrix? The answer is contained in the following result.
Theorem 2. A solution matrix ®(t) of
y =A(t)y (4.24)
is a fundamental matrix if and only if det &(t)40 for every t in ¥. Further,
ifdet P(t) #0 for some to in F, then det &(t)40 for all t in %. (By det H(t)
we mean the determinant of the matrix ®(t).)
Proof. If det &(t)40 for every t in ¥, then the columns of the solution
matrix ®(t) are linearly independent on ¥. For suppose there exist constants
Cisse C. Such that
C1, (t) +c. (t)+---+¢,0,(t)=0 foreverytin ,
where @,(0),..., @,(¢) are the columns of ®(t). This can be written in the
form -
-{]
@(t)e=0 foreverytin J,
where
Fix ¢ at t=fo in %. Then O(tyo) c=0 is a system of n algebraic equations

for the m unknowns ¢),..., c,- Since det O(tp)40, c,=0, c,=0,..., c,=0
by Cramer’s rule. This proves that the columns
; (1),-.-, On(t)
are linearly independent; hence ®(¢) is a fundamental matrix on ¥.
Conversely, suppose ®(t) is a fundamental matrix of (4.24) on ¥. Let
o(¢) be a solution of (4.24) on ¥. By Eq. (4.30), there exists a unique vector
ec such that o(t)=@(t)¢ for every ¢ in ¥. Fix fy in ¥; then, in fact, the
constant vector ¢ is uniquely determined by solving the algebraic system
(to) Xx=(to). Since this algebraic system has a unique solution for each
right-hand side (to), the coefficient matrix ®(t)) has rank n. Hence ®(tg) is
nonsingular, and therefore, det ®(t)) #0. This is true for each fixed fy in ¥,
and therefore det ®(r)#0 for each t in Y. It may appear that the vector c
depends on the choice of to. However, it does not for the following reason.
Since @(¢)=@(t) c for every ¢ in Y, if 4;Ato in FY, then o(t,)=(¢,) c. Thus,
the unique solution of the algebraic system ®(t,) x=@(t,) is the same vector
¢ obtained as the unique solution of the algebraic system ®(to) x= (to).
Finally, if det ®(to)40 for some fy in Y, let 6, =, (to), ..., 6,=, (Co).
The vectors 6,,..., 6, are linearly independent, and therefore form a basis for
Euclidean n-space. We claim that the solutions @,(¢),..., o,(t) are linearly
independent on ¥; for if not, there exist scalars c,, C,..., ¢, not all zero
such that
C1, (t) +022

(t) ++ +¢,,(t)=0 on F.
Putting t=, we obtain
C10, +0C,0,+-:-+c,6,=0
which contradicts the linear independence of 6;,...,6,. Hence, ®(¢) is a
fundamental matrix of (4.24). Therefore, by the second part of the proof,
det ®(t)40 for every rin ¥. |
The reader is warned that a matrix may have its determinant identically
zero On some interval, although its columns are linearly independent. Indeed,
let
ae tS
P()=)0. 2 ¢
Oe0 an
Then clearly det ®(t)=0, —00<t<oo, and yet the columns are linearly
independent. This, according to Theorem 2, cannot happen for solutions of
(4.24).
Example 1. Show that
We first show that ®(t) is a solution matrix. Let @, (¢) denote the first column of (7);
sn-(SHLo AISHG ts
then
eof 2H TEE se
for —co <t<oo. Similarly, if @,(t) denotes the second column of ®(t), we have
for —co<t<oo. Therefore, &(t)=[, (t), 2 (2)] is a solution matrix for —00<t<oo.
By Theorem 2, since det ®(t)=e7'40, O(¢) is a fundamental matrix for —0o<t<oo.
By Theorem 2 also, it is enough to compute det ®(f) at one point, for instance t=0.
Since &(0)=/, this gives det 6(0)=1+#0.
Exercises
4. Show, with the aid of Theorem 2, that

cost sint
—sint cost
is a fundamental matrix for the system y’= Ay, where
Om
= |—1 ‘l
5. Show, with the aid of Theorem 2, that
| exp(r,t) cae
r, exp(r,t) rz exp(rot)
is a fundamental matrix for the system y’= Ay, where
A=
kere
a ea,
,
and r,, rz are the distinct roots of the quadratic equation z” + a,z+a,=0. (We shall
learn in Section 5.3, Exercise 1, how to construct this fundamental matrix.)
Corollary 1 to Theorem 2. If ®(t) is a fundamental matrix of y'=A(t)y on

an interval ¥ and if CE F mn is anonsingular constant matrix, then ®(t) C is also
a fundamental matrix of y'=A(t) y on J.
Proof. Let &(t)=[, (0), 02(0),...,0,(0)]. Then the columns of @(t)C
are linear combinations of the columns of ®(t) (by matrix multiplication).
Since the columns of ®(t) are solutions, ®(¢) C is a solution matrix on ¥. But
det &(t) C=det &(t) det C. By Theorem 2, det &(t)40 on YY, and since C is
nonsingular, det C40. Thus, det &(t) C¥ 0 on -¥ and, again by Theorem 2,
(t) C is a fundamental matrix on %. |
Exercise
6. Show that C®(t), where C is a constant matrix and @(t) is a fundamental matrix,
need not be a solution matrix y’= A(t) y.
The converse of Corollary | is also true.
Corollary 2 to Theorem 2. If ®(t) and (t) are two fundamental matrices of

y'=A(t) y on J, then there exists a nonsingular constant matrix C such that
Y (t)=@(t) Con ¥.
Proof. Letting w; be the jth column of Y, we see from (4.30) that y,;=e,,
j=1,...,n, where c; are suitable constant vectors. Therefore, if we define
C as the constant matrix whose columns are the vectors c;, J=1,...,, we
have at once that ¥(t)=@(t) C for every t on .¥. Since
det ¥ (t)=det H(t) det C
and since det ® and det ¥ are both different from zero on ¥% (why?), we
also have det C40 so that C is a nonsingular constant matrix. |
Exercises
7. a) Show that
is a fundamental matrix for the system y’= A(t) y, where
A(t)=]} —
Se
on any interval ¥ not including the origin.

b) Does the fact that det (0)=0 contradict Theorem 2?
8. Show that if a real homogeneous system of two first-order equations has a
fundamental matrix
then
cost sint
—sint cost
is also a fundamental matrix. Can you find another real fundamental matrix?
[Hint: Let &(t)=[,
(2), d2(t).] Show that Zp,(t) and Zh,(2) are solutions of
y =Ay, A real, where & is the real part. By the real part of a vector we mean,
of course, the real part of each component. A similar result holds for the
imaginary parts of o, (7) and o,(¢). ]
We shall now apply Theorem 2 to the scalar linear homogeneous second-

order equation
y’+p(t) y +q(t) y=0, (4.27)

where p and q are continuous functions on a given interval ¥. As we have
seen in the proof of Corollary 1 to Theorem 1, (4.27) is equivalent to the
system
saeOmeOa os
If (2) is a solution matrix of (4.28) on ¥, then (t)=[, (2), b2(d], where
6, (0 be
t
W(t)
= / b) 62(0=| 200
W(t)
t = /
with w,(t), W(t) solutions of the scalar equation (4.27). By Theorem 2, ®(t)
is a fundamental matrix of (4.28) on % if and only if
i(t) i Wolt
det &(t) = det \¥ ;
vin Fo ford ine?
This determinant is called the Wronskian of w,(t) and w(t). Thus, by the
proof of Corollary 1 of Theorem 1, if det &(t)40, then the solutions yw, (2),
W(t) of the scalar equation (4.27) are linearly independent on ¥, and every
solution of (4.27) can be written as a linear combination of ,(¢) and (2).
This is one-half of the following result.
Corollary 3 to Theorem 2. Two solutions W,, Wz of (4.27) on F are linearly
independent on # if and only if their Wronskian
= Wilt) Wo(t)
W (w(t), W2(t)] =det lye(t) |
is different from zero for all t in #.

Proof. We must still prove that if the solutions W,, w2 are linearly inde-
pendent on .¥, then their Wronskian is different from zero for every f in ¥.
Suppose there is at least one ? on % such that W[yW, (2), W2(2)]=0. (If no
such ? exists, then there is nothing to prove.) Consider the algebraic system
a,W, (€) + a22(#)=0
a, Wy(2) +a2W2(#)=0,
for the unknowns a,, a,. By the theory of linear homogeneous algebraic
equations, this system has a nontrivial solution @,, a, where a,, a, are not
both zero. Consider the function w(t)=4,W, (t)+4,W2(t). Since (4.27) is
linear, is a solution of (4.27) on ¥, and w(?)=0, w’(7)=0. By Corollary 1
to Theorem 1, Section 4.2, there is only one solution to the initial value
problem consisting of (4.27) together with the initial conditions y(?)=0,
y’(?)=0. Since the identically zero function is a solution of this initial value
problem, we conclude that w(t)=0 on ¥. Therefore, 4,W,(t)+a,2(t)=0
for every ¢t in 4%. Since a,, a, are not both zero, w, and w, are linearly
dependent on ¥. Thus, if Ww, and w, are linearly independent on -Y, there
can be no such 7, and W[wW, (2), ¥2(t)]
40 for every t in ¥. |
By Corollary | to Theorem 1 every solution of equation (4.27) on ¥ has
the form c,W, +c, for some unique choice of the constants c,, c,. For this
reason a pair of linearly independent solutions, such as Wj, Wz, of Eq. (4.27)
are said to form a fundamental set of solutions.
Exercises
9. Show that e?', e 7‘ are linearly independent solutions of y” —4y=0 on — co <t<oo.

10. Show that e~'? cos ./3t/2, e~"? sin J3t/2 are linearly independent solutions of
y +y +y=00n —O<t<o0,
11. Show that e~', te~' are linearly independent solutions of y’+2y’+y=0 on
—0o<t<o.
12. Show that sin 1’, cos ¢? are linearly independent solutions of ty”—y’+4r3y=0 on
0<t<a@ or —o<t<0. Show that W(sin 7, cos t*) (0)=0. Why does this fact
not contradict Corollary 3 of Theorem 2?
13. a) Let 1, ¢2 be any two solutions on some interval 4%, of L(y)=do(t) y’ +a, (t) y’
+a,(t) y=0, where do, ay, a) are continuous on ¥ and ag(¢)#0 on .¥. Show that
the Wronskian W(¢,, ¢2)(¢) satisfies the first-order linear differential equation
t
wa 20 ton ¥. (s)
ao (t)
[Hint:
bi(t) P2(t)
W'(o1, $2) (= =(6162—
$162) =9163—-Fib2-
P(t) $5(t)
Now, use the fact that ¢,, ¢, are solutions of L(y)=0 on ¥ to replace
”
, and $3 by terms involving ¢,, 6}, $2, 4. If you then collect terms you
should get Eq. (*).]
b) By solving (*), derive Abel’s formula:
t
W(b1.62)()=W (61.62) to) exo(—| “as,

to
for to, t on .¥. This gives another way of seeing that if the Wronskian is
different from zero at one point, then it is never zero.
14. State the analog of Corollary 3 to Theorem 2 for the linear third-order
L3(y)=4(t) ¥" +4, (t) y’ +ay(t) y’+.43(t) y=0.

15. Show that e’, cos f, sin ¢ are linearly independent solutions on — co <t<oo of the
differential equation y” —y’+y’—y=0.
16. Show that
(9)
tm
m=1
co
pmr
ba()=t+) 3-4-6-7---(3m) (3m+ 1)’

m=1
are linearly independent solutions of y”—ty=0 on the interval — 00 <t<oo. (Here

you may assume that it has already been shown that ¢, and @, are solutions of
y”—ty=0, but how could you verify this?)
By a similar argument we can establish the following analog of Corollary

3 to Theorem 2 for the scalar equation of order n.
Corollary 4to Theorem 2. A set ofnsolutions W,, W2,..., W, on F of
y +p, (t) yD +--+,(t)y=0, (4.29)

where D1, P25 +++5 Py are continuous on F, is linearly independent on # if and

only if the Wronskian
Wil) Walt) alt

Wilt)... va(}=der| M9 ¥200 val
YO D(t) WED) Yet)
is different from zero for every t on F.
A set of n linearly independent solutions of Eq. (4.29) is said to form a
fundamental set of solutions (see the remarks following Corollary 3 to
Theorem 2 in the second-order case).
Exercises
17. Prove Corollary 4 to Theorem 2. [Hint: Imitate the proof of Corollary 3 to

Theorem 2. ]
18. In each of the following, let @, (¢) and #(t) be solutions of the differential equation
L(y)=y"+p(t) y'+q(t) y=0

on some interval 4%, where p and g are continuous on .¥.
a) If @; (to)=2(to)=0 for some fo in %, show that the solutions ¢, and @, cannot
form a fundamental set of solutions on ¥.
b) If the solutions ¢, and ¢, both have a maximum or minimum at some point
t; in ¥%, show that d, and @, cannot torm a runaamental set of solutions on -¥.
c) Let @, and ¢, form a fundamental set of solutions on ¥% which both have
an inflection point at some point f, in 4. Show that p(t,)=q(t,)=0.
d) Let ¢, and #, form a fundamental set of solutions on .%. Show that
W,=¢,4+¢2, W2=62—2¢, also form a fundamental set of solutions on ¥.
19. a) Let ¢, and ¢, be solutions of L(y)=y" —4ty’+(42? —2) y=0 on the interval
—0<f<oo, satisfying the initial conditions $,(1)=1, ${(I)=3, ¢2(1)=3,
$5 (1)=1. Are these solutions linearly independent on — 00 <t< 0? Justify your
answer.
b) Show that W, (t)=exp (t’) is a solution of the equation L(y)=0 and find a second
linearly independent solution on — 00 <t<oo. [Hint: Look for a solution w,
of the form w,(t)=u(t) ¥,(d); substitute and find u(¢) to make w,, > linearly
independent solutions.|
c) Find the solutions ¢, and @, in part (a).
20. Given that the equation
ty” —(2t+1) y+2y=0, t>0
has a solution of the form e* for some c, find the general solution. [Hint: First
find what c must be; then find a second linearly independent solution as in
Exercise 19(b).|
4.4 Linear Nonhomogeneous Systems 147
21. a) One solution of the equation
1
Luj=ry ty +(2-3) y=0, t>0
is t'/? sin ¢. Find the general solution of the equation L(y)=37"/? sin t, where
t>0. [Hint: Use the method suggested in Exercise 19(b) previously. |
b) Repeat part (a) for the equation
2ty” +(1—4t) y’+(2t—1) y=e.
To provide the reader with some easy, concrete examples, we advise that
he study the solution of linear scalar differential equations with constant
coefficients as carried out in Sections 3.4 and 3.5. While this material is an
easy special case of linear systems with constant coefficients, to be studied in
Chapter 5, it is, nevertheless, helpful to see the special case first.
4.4 LINEAR NONHOMOGENEOUS SYSTEMS
We now use the theory developed in Sections 4.2 and 4.3 to discuss the
form of solutions of the nonhomogeneous system
y=A()yta(d), (4.31)
where A(t) is a given continuous matrix and g(t) is a given continuous vector
on an interval .%. The entire development rests on the assumption that we can
find a fundamental matrix of the corresponding homogeneous system y' = A(t) y.
The vector g(t) is usually referred to as a forcing term because if (4.31)
describes a physical system, g(t) represents an external force. By Theorem 1,
Section 4.2, we know that given any point (fo, ), fo in ¥%, there is a
unique solution @ of (4.31) existing in all of ¥%such that (to) =n.
To construct solutions of (4.31), we let ®(t) be a fundamental matrix of
the homogeneous system y’=A(t)y on ¥; @ exists as a consequence of
Theorem 1, Section 4.3 (see also remarks immediately following its proof).
Suppose , and @, are any solutions of (4.31) on 4. Then @, — @,j is a solu-
tion of the homogeneous system on -¥.
Exercise
1. Verify this fact.
By Theorem 1, Section 4.3, and the remarks immediately following its proof
(in particular, see Eq. 4.30), there exists a contant vector ¢ such that
o,—o,=@e. (4.32)
Formula (4.32) tells us that to find any solution of (4.31), we need only
know one solution of (4.31). (Every other solution differs from the known
one by some solution of the homogeneous system.) There is a simple method,

known as variation of constants, to determine a solution of (4.31) provided
we know a fundamental matrix for the homogeneous system y'=(At) y. Let ®
be such a fundamental matrix on .%. We attempt to find a solution of
(4.31) of the form
W(t) = (t) v(t), (4.33)
where v is a vector to be determined. (Note that if v is a constant vector, then
W satisfies the homogeneous system and thus for the present purpose v(t)=c
is ruled out.) Suppose such a solution exists. Then substituting (4.33) into
(4.31), we find for all ¢ on #
W (= F(t) V+ P() V(Q=A() OD) V(t)

+(0).
Since @ is a fundamental matrix of the homogeneous system, ®’ (t) = A (t) &(2),
and the terms involving A(t) ®(t) v(¢) cancel. Therefore, if w(t) =®(¢) v(t) isa
solution of (4.31), we must determine v(t) from the relation
P(t) v(t)=g(0).
Since @(t) is nonsingular on ¥ we can premultiply by ® *(t) and we have, on
integrating,
t
v= |P~'(s) g(s) ds, to, ton F
and, therefore, (4.33) becomes

t
vio=o09 |@~(s) g(s) ds, to, Con. (4.34)

to
Thus, if (4.31) has a solution of the form (4.33), then W is given by (4.34).
Conversely, define w by (4.34), where © is a fundamental matrix of the
homogeneous system on ¥. Then, differentiating (4.34) and using the
fundamental theorem of calculus, we have
t
VI=# |ORO.H+90) O80

t
= 4()(0)[O*6) 26)ds+20
and using (4.34) again,
for every ¢ on 4. Obviously, W(to)=0. Thus, we have proved the variation

of constants formula:
Theorem 1. If ® is a fundamental matrix of y'= A(t) y on ¥, then the func-
tion t
is the (unique) solution of (4.31) satisfying the initial condition
W (to) =0
and valid on F.
Combining Theorem | with the remarks made at the beginning of this

section, we see that every solution @ of (4.31) on ¥ has the form
(1) =,() + (0) (4.35)

where yp is the solution of Eq. (4.31) satisfying the initial condition (t))=0,
and @, is the solution of the homogeneous system satisfying the same initial
condition at to as , for example ,(to)=1.
leh Le} of
Example 1. Find the solution of the initial-value problem
We have seen in Example | Section 4.3 that
is a fundamental matrix of the associated homogeneous system on —0co<f<oo.

Taking the inverse of the matrix ®(t), we obtain
e —se*
is t
Since ©(0)=/, the solution of the corresponding homogeneous system satisfying the
initial condition
Exercises
2. Consider the system y’= Ay+g(t), where
Verify that
‘lo 2} rn} e0-[
sm
is a fundamental matrix of y’=Ay. Find that solution @ of the nonhomo-
geneous system for which
1
=| |}
3. Find the solution @ of the system y’ = Ay +g/(t) with A the same as in Exercise 2 and
with
satisfying the initial condition
=|||
4. Consider the system y’= A(t) y+g(¢), where
01
7]
A(t)=|—2 2
eo-[i]
alg
Find the solution @ satisfying the initial condition
and determine the interval of validity of this solution. [Hint: Use the fundamental
matrix given in Exercise 7, Section 4.3.]
We now consider the form of the variation of constants formula for the
scalar second-order linear nonhomogeneous differential equation
y’+p(t) y'+q(t) y=r(t), (4.36)

where p, g, and r are continuous on an interval 4%. We have seen in Corollary
1 to Theorem 1, Section 4.3, that the corresponding homogeneous equation
y"+p(t) y +q(t) y=0 (4.27)

has two linearly independent solutions ,, Y, on .%. These solutions are the
first components of the (vector) solutions,
r=[-a0 -s0) In Lh ee
We apply Theorem | to the system
er nto|?*[to} s-{>| ore

which is equivalent to (4.36). Since w,(t), W(t) are linearly independent
solutions of (4.27), by the equivalence of the equations (4.27) and (4.28), the
matrix
P(t)= W(t) Wr (t)
Wilt) W2(t)
is a fundamental matrix of the system (4.28) on ¥. Let
= u, (t)
oY bs4
be the solution of (4.37) satisfying the initial condition u(t )=0. By Theorem
I,
u(t) =9(0) i© (5) hal ay,

to
From
21 W2(s) —W2(s)
s) meal ave r 0)
I W/(s) seal
Wi (s),Wo(s\] |—Wils) Was)
weobtain
ca ). W2(5)] ae } v0 |\eige 3 S
- |e roc
W(t) W1(s)—W1(t) W2(s)
ove! ep
The solution of (4.36) satisfying the initial conditions y(to)=0, y’ (to) =0
is, by the equivalence of (4.36) and (4.37), the first component aeicf u(t).
Therefore this solution is
nig [as val) 4,

CARA)
Thus we have proved ie variation of constants formula for the scalar second-
order linear equation:
Corollary to Theorem 1. Let W,(t), z(t) be linearly independent solutions of

y”+p(t) y+q(t) y=0 (4.25)
on an interval %. Then the function
E: [Wo(t) Wi(s)—Wilt
‘ 0- | Walsh W is
s)]
bs detk9) (4.38)
is the unique solution of
y’+p(t) y+4(t) y=r(0) (4.36)

on ¥& satisfying the initial conditions y(to)=0, y'(to)=0. Moreover, every

solution of (4.36) on ¥% has the form
w(t)=cyWy(t)+c2h2(t)+u, (t) (4.39)

for some unique choice of the constants c,, C).
This last expression (4.39) is called the general solution of (4.36).

Example 2. Find a particular solution of the differential equation
TT Tl
y’+y=tant, 3
We apply the corollary to Theorem | directly using the linearly independent solu-
tions /, (t)=cos ft, 2 (t)=sin ¢ of the homogeneous equation y” + y=0. We have
cost sint
W(Wi(t), W2(0)= ; =1.
—sint cost
Hence formula (4.38) yields (using tp =0) the particular solution
Hi ()= ((in cos s—cost sins) tans ds

(0)
t t
=sint |sin s is—cost| sins tans ds

(0) oO
=sine(1 cost)+051 |(coss—sees) ds

0
= sint(1—cost)+cost(sint—log |sect
+tant).
=sint—cost log |sect+tant].
We note that, since sin ¢ is a solution of the homogeneous equation, the function
u(t)= —cost log |sect

+ tan¢|
is also a particular solution. We also remark that we could apply Theorem | directly
by first converting the given differential equation to an equivalent system of first-order
equations as was done for Eq. (4.36); however, for second-order scalar equations, it is
more efficient to employ the Corollary.
Exercises
5. Verify by direct substitution that w,(¢) is a solution of (6.36) on ¥ satisfying the

initial conditions uw, (to) =0, vu; (to) =9.
a) y"+y=sect —aj2<t<n/2.
b) y’+4y'+4y=cos 2¢.
c) y’+4y=f(t), where f is any continuous function on some interval ¥.
d) y’—4y'+ 4y=3e '+20? +sin ¢.
. If @ is a solution of the equation y’+k*y=f(t), where k is a real constant _
different from zero and
f is continuous for 0<t<oo, show that c, and c, can be
chosen so that
t
1
o(th=cy cos kt +" sinkt +7 [sinke—s f (s) ds
0
for 0<t<oo. (Use cos kt and sin kt/k as a fundamental set of solutions of the
homogeneous equation.) Find an analogous formula in the case k=0 and show
that it can also be obtained by computing lim,_.5 o(¢).
. Given the equation
y'+5y'+4y=
f(0).
Use the variation of constants formula to prove that:
a) If f is bounded on 0<t<oo (that is, there exists a constant M>0 such that
[f(t)| <M on 0<t<oo), then every solution of y’+5y’+4y=/f(t) is bounded
on 0<f<oo.
b) If also f(t)>0 as too, then every solution ¢ of y’+Sy’+4y=/(t) satisfies
(t)>0 as t>00.
. Can you formulate Exercise 8 for the general equation
y"+a,yy+a,y=f(t), a,, a2 constant

with a,, a, suitably restricted?
10. Formulate the analog of the corollary to Theorem | for the third-order equation
y"+pi(t) y+pay(t) y'+ps(t) y=f (0d.
[Hint: Let w, (t), W2 (0), W3(¢) be linearly independent solutions of the corresponding
homogeneous equation. Then proceed as in the corollary.]
a) y”—8y=e".
b) y+ l6y=f(d, fcontinuous on — co <t<oo.
12. Find the general solution of each of the following differential equations:
a) y’a cosectcott, O<t<z
b) y”—6y’+ 9y=e'
c) y’—6y’+9y=te*!
d) Ono. l6oy=0
e)My YY —2y"+y=e'+sint
f) y”+y’=tan ¢, 0<t<n/2
g) w+ y=g(t); g(t) continuous
h) y’+y=h(t), where
h(t)=t(0<t<nz), h(t)=ncos(x—-t), nm<t<2n
and A is periodic with period 27.
13. The current / in amperes in an electrical circuit with resistance R ohms, inductance
L henrys, and capacitance C farads in series (see Fig. 4.3), is governed by the
equation
R
E(t)
Figure 4.3
Lt Ri4ci
i) — E'(t) 5
C
where E(t) is the applied voltage. Suppose the applied voltage is a constant Ep.
a) Show that the current decreases exponentially if CR? >4L.
b) Find the current of CR?=4L and i(0)=0, i’(0)=E£)/L.
c) Find the current if CR* <4Z and i(0)=0, i’(0)=£,/L.
14. Consider the circuit of Exercise 13. Find the current if
0 t<0
BO=46, sinat, t>0O
assuming i(0)=0, i’(0)=0.
15. Find the current in an electrical circuit with inductance and capacitance in series,
but no resistance, and an applied voltage E(t) given by
BE
poral 0<t<l
2
le ifsIh
with i(0)=0, i’(0)=0.
16. Determine a particular solution of the differential equation
y" +py' +qy=A coskt,
where p and q are positive constants, of the form $(t)=B cos (kt—«). Show that
the amplitude B of the oscillation is a maximum if k=(q—p7/2)'/”, called the

resonant frequency, provided p?<2g. What happens in the case p*>2q? Show
that at resonance the amplitude of the oscillation is inversely proportional to the
damping p. [Hint: For the corresponding homogeneous equation, see Exercise 14,
Section 3.4.]
4.5 NONLINEAR SYSTEMS OF FIRST-ORDER EQUATIONS
Nonlinear systems of first-order differential equations can be treated in a

way which has many points of similarity with the treatment given in
Section 4.1 and 4.2 for linear systems. While the algebraic structure of
solution sets does not carry over to nonlinear systems, much of the notation,
basic existence theory, and reduction of scalar equations of order n to systems
is much the same as in the linear case. In this section we shall study systems
of first-order differential equations of the form
Y,=filt Vio Va. +e Yn)

Yr~=hlt, V1, Ya2> 2i? Yn) (4.40)
Yn = Sn(t, Vi> V2. +++ Yn)
where f;, f>,...,f, are n given functions defined in some region D of

(n+ 1)-dimensional Euclidean space and y,, y2,..., y, are the n unknown
functions. To solve (4.40) means to find an interval J on the ¢ axis and n
functions ¢,,..., ¢, defined on J such that
i) $i (0), 65(0), ..., o,(0) exist for each ¢ in J.
ii) the point (¢, o,(t),..., @,(t)) remains in D for each ¢ in J.
iii) Oj)()=f,(t, O1 (1), $2(0),---. On (0) foreach tin) = (j=1,..., n).
Thus (4.40) is the analog of the single equation y’=f(t, y) studied in
Chapter 1. Naturally, the functions f; may be real or complex-valued. We
shall assume the real case unless otherwise stated. While the geometric inter-
pretation is no longer so immediate as in the case n=1, a solution of (4.40)
(that is, a set of m functions ¢,,..., ¢, on an interval /) can be visualized as a
curve in the (7+ 1)-dimensional region D, with each point p on the curve
given by the coordinates (¢, 6, (¢),..., ¢,(t)) and with ¢/() being the com-
ponent of the tangent vector to the curve in the direction y;. This inter-
pretation reduces to the one already given when n=1 and the curve in D
defined by any solution of (4.40) can therefore again be called a solution
curve. The initial-value problem associated with a system such as (4.40) is the
problem of finding a solution (in the sense defined above) passing through a
given point Po: (fo, 11, N25---» Mn) (we do not write (to, ¥10,---, Yao) to avoid
double subscripts) of D. In general, we cannot expect to be able to solve (4.40)
except in very special cases. Nevertheless, it is desired to obtain as much
4.5 Nonlinear Systems of First-Order Equations 157
information as possible about the behavior of solutions of systems. For this

reason we shall develop a considerable amount of theory for systems of
Example 1. Consider the second-order equation
y"=g(t, y, y) (4.41)
where g is a given function. Put y=),, y’=y2; then one has y,=y, and from (4.41)
y’=y5= g(t, ¥;, ¥2). Thus (4.41) is apparently equivalent to the system of two first-
order equations
Yi=Y2
4.42
y2=9(t. V1, V2) Rae)
which is a special case of (4.40) with n=2, f(t, ¥1, ¥2)=V2, fo(t, V1, V2) =9G(t, V1, V2):
To see this equivalence let ¢ be a solution of (4.41) on some interval /; then y, =(0),
yv2=¢ (t)is a solution of (4.42) on J. Conversely, let b,, 2 be a solution of (4.42) on J,
then y= @,(t) (that is, the first component) is a solution of (4.41) on J.
Exercise
1. Write a system of two first-order differential equations equivalent to the second-

order equation
6” += sind=0
:
with initial conditions 6(0)=6, 6’(0)=0, which describes the motion of a simple
pendulum (Section 2.2).
Example 2. The scalar equation of nth order
yM=g(t,y, Ys. y¥” ”) (4.43)

can be reduced to a system of v first-order equations by the change of variable
Vi=Vs V2= V's-00s Vn ="). Then (4.43) is seen to be equivalent to the system
Yi=Y2
Y2=Ys
: (4.44)
Ya-1=Yn
Yn=9(t V1; V2.0 Yn)
another special case of (4.40).
Exercises
2. Establish the equivalence of (4.43) and (4.44).

3. Reduce the system oe
Virey2Vt the
2y1 +3y2=2y1y2
to the form (4.40). [Hint: Solve for y{ and y}.]
To study systems of first-order equations such as (4.40) systematically it

is convenient to introduce vectors. We define y to be a point in n-
dimensional Euclidean space, E,,, with coordinates (y,,..., y,). Unless other-
wise indicated, E,, will represent rea/ n-dimensional Euclidean space, that is,
the coordinates (y,,..., ¥,) of the vector y are real numbers. However, the
entire theory developed here carries over to the complex case with only minor
changes, which will be indicated where necessary. We next define functions
AGVN=Slby8I)s GStegn
and thus the system (4.40) can be written in the form
yi =f, (t, y)
Y2= Salty) (4.45)
Vy = fultsy).
Proceeding heuristically (we will be more precise below), we next observe
that /,,...,/,can be regarded as n components of the vector-valued function
f defined by
f(t, y)=col(f, (t, 4 ees A y)),
where col means column vector. We also define
y’=col(yj,..., Yn):
Thus the system of n first-order equations (4.40) (and all the systems which
arose earlier in this section (see also (4.45))) can be written in the very compact
form
¥ =f(¢ y) (4.46)
Equation (4.46) resembles the familiar single first-order equation y’=f(t, y),
with y, f replaced by the vectors y, f, respectively.
Example 3. We may write the system (4.42) above
so that
f(t, y) =col(y2, g(t, ¥1, Y2))-
The Euclidean length of the vector y is defined by the relation
n 1/2
yl =(il? +--+ parya( > vi?) |

Notice that |y;| is well defined for y; complex and thus ||y|| is also defined for
a complex vector y. We need the notion of length in order to measure dis-
tances between solutions of systems. However, for the purpose of dealing with
systems such as (4.46) it turns out to be more convenient to define a different
quantity for the length (or norm) of a vector y than the familiar Euclidean
length, namely,
Iyl=[il+Iyal +--+ bol=20 Lil.

Again, |y| is well defined for either real or complex vectors y. No confusion
need arise from using the absolute value sign for different purposes; on the
left-hand side |y| is the notation for length of the vector y; on the right-hand
side we sum the absolute values of the components of y. Observe, for ex-
ample, if y=(3 +i, 3—i), then ||y|| =(/3 +47 +|3 —@?)'/? =(10 + 10)!/? =(20)!/?
and |y|=|3 + i|+|3 —i]=(10)'/* +(10)!/? = 2(10)'/*; clearly |y| > ||y|| in this case
and in fact, |y|=./2|/y||. In general, the quantities ||y|| and |y| are related, as
follows.
Exercise
4. If y is an n-dimensional vector, E,,, show that
lyl'siyisJ/n lyl
[Hint: Use the inequality 2|uv| < |u|? + |v}? and show |ly||? <|y|? <nlly|7.]
The important point about this inequality is that |y| is small if and only if
lly||is small.
The length function ly| has the following important properties:
i) jy/=0 and |y|=0 if and only if y=0.
ii) if c is any complex number, |cy| = |c| ly].
ili) for all y and z, y+z|<|y|+|z].
The proofs are immediate from well-known properties of complex numbers.
For example, to prove (ii) we have
leyl= > leyd= Y tel bl=lel YYWl=lel yl

n
j=1 A =1 i=
Similarly for (iii) we use the inequality |w+v| < |u|+ |v| valid for any complex
numbers u and v.
Exercise
5. Show that the Euclidean length ||y|| of a vector y also satisfies the properties
(i), (ii), (iii) above. [Hint: To prove (iii) you will need to apply the Schwarz inequality
for sums, that is,
to the term 2&( 9" y,Z;) which arises in the expansion of

i=1
ly+z\|7= S ly; +2,|7 = (yi+2,) (vit+2;)

= i
zoe
-> lyp+2a(S yizZ

i+ X ele |
i=1
Using the length function we define the distance between two vectors y
and 2, d(y, z), by the relation
d(y, z)=|y—2|.
The distance function d(y, z) has the following important properties:
i) d(y, z)>0 and d(y, z)=0 if and only if y=z.
il) d(y, z)=d(z, y).
iii) d(y, z)<d(y, v)+d(v, z) (triangle inequality).
The proofs of these properties follow immediately from the corresponding
properties (iwe(ii), (iii) of the length function. For example, to prove (ili) we
have d(y, z)=|y—z|=|(y—v)+(v—z)|<|y—v|+ |v—z|=d(y, v)+d\v, z).
Any function satisfying the properties (i), (ii), (iii) is called a distance
function. For example, p(y, z)=||y—z|| for any vectors y, z is such a func-
tion, and represents the Euclidean distance between the points y and z in E£,,.
Exercise
6. Show that p(y, z)=||y—z|| also satisfies the properties of a distance function.
[Note: The proof of (iii) is harder than for the distance function d. You will need
to use the Schwarz inequality as in Exercise 5 above.]
To define continuity, differentiability, and integrability of vector func-

tions, we need the notion of limit for vectors. We use the distance func-
tion d to define convergence. A sequence {y“} of vectors in E,, is said to
converge to the vector y if and only if d(y™, y)=|y“ —y|+0 as k->oo and
in this case we write lim y“ =y or fy}. Since
ly —y|=ly — yl + ly? —yalt+--- +1 —y,

where y™ =(y, . ., yY), y=(1,---, Vn), the above definition says that the
sequence of vectors {y“)}—>y as k>oo, if and only if each component of
y”) (these form a sequence of real or complex numbers) tends to the
corresponding component of the vector y. It is clear that all properties of
limits of sequences of complex numbers may now be assumed to hold for
sequences of vectors without further explanation.
If we use the Euclidean distance function p(y, z)=||y—z||, we say that
the sequence {y“)} converges to the vector y if and only if p(y, y)= ||y — y||
+0 as k-—oo. It seems clear that the concept of convergence should not de-
pend on the particular distance function used. We establish this for the
distance functions d(y, z) and p(y, z) in Exercise 11 below.
Exercise
7. Let {y} be a sequence of vectors. Show that |y“ —y|0 as ko if and only if
ly“ —y|| 0 as koo. [Hint: Use Exercise 4.]
A vector-valued function g=g(t) is a correspondence which assigns to

each number t in an interval I one and only one vector g(t); we write g(t) =(g, (2),
.-5 G,(t)) and we call g, the kth component (real or complex-valued scalar
function) of the vector function g. Because of our definition of convergence
we can now define g to be continuous, differentiable, or integrable on / if
and only if each component of g has this property. As in Section 4.1, if g is
differentiable on J we denote its derivative by g’ and we define
@ =collg.. dose gee
Similarly, if Eis continuous on J we denote its integral from a to b (a and b

on J) by f° g(s) ds and we define
a
a(s) o-(|9,(s) ds, fy jest is)
a a a
We take note of the very important inequality

b
|g(s) ds = |Ig(s)| ds a<b. (4.47)

a
To prove (4.47) we have
[nef
b
se fo )ds
<[int
(s)| ds+-- +for a= i)| ds
Exercise
8. Justify each step in the proof of inequality (4.47). Note that in the middle steps you
have ordinary absolute values.
It is also true that
uve < |ig( ds
for any continuous vector g, but the proof is more difficult than the one for
(4.47).
We can now return to the system
y =f(t, y) (4.46)
where the vector-valued function f is defined in some (n+ 1)-dimensional
region D in (t, 1, 2,---, ¥n) space. To find a solution of (4.46) (compare
Section 1.2) means to find a real interval J and a vector function @ defined on
Tsuch that
i) '(¢) exists for each t on J.
ii) The point (t, @(t)) lies in D for each ¢ on J.
ili) @'(t)=f(¢, b(t) for every ¢ on J.
Thus the analogy between (4.46) and a single scalar equation of first order
is complete. Just as for the scalar equation, to solve an initial-value problem
for the system (4.46) with the initial condition @(¢o)=M, (fo, n) a point of D,
means to find a solution @ of (4.46) in the above sense passing through the
point (to, n) of D, that is, satisfying @(to) =. While it is not in general possible
to solve (4.46) explicitly, we can illustrate the concepts with some simple
problems.
Example 4. The system
is of the form (4.46) with y=(y;, ¥2), f(t, vy)=(V2, ¥1). Clearly, D is all of (t, y,, v2) space
and (¢)=(e', e’) is a solution valid for —0o <t<o, since (i), (ii), (ili) of the definition
are satisfied. Note that cd, c a constant, is also a solution.
Exercise
9. Can you find (guess) another solution w(t) of the system in Example 4 on — 00 <t<00
which is not of the form cd(t)?
There are few systems of the form (4.46) of any real interest, other than
those which are linear in the components of y with constant coefficients,
which can actually be solved explicitly. We therefore refrain from making up
special “‘textbook” problems for this purpose. On the other hand, one wishes
to analyze the behavior of solutions of systems of the form (4.46) that cannot
be solved explicitly. The first question that comes to mind in this analysis
is: Does the given system have a unique solution satisfying the given initial
condition? The following result provides an answer in the affirmative in
most problems which arise in practice. We note that because of the
equivalence between nth-order equations and systems of first-order equations
already established, this result includes all results of this type discussed in
Chapters | and 2 as very special cases. For the proof of this result we refer to
Section 8.5.
In what follows we let D represent a region in (n + 1)-dimensional space
with the property that (compare Section 1.6) given any point (¢o, y) in D,
the interior of the (7 + 1)-dimensional “‘box”
B={(t, y) ||t—tol<a, ly—nl<b}
will, for a, b>0 and sufficiently small, lie entirely in D. (We note that if we
use the Euclidean norm |/y—1|| <4 then the set
C={(t, y) |t—tol<a, lly—nl
| <b}
would specify a “‘cylinder’’ whose cross section by a hyperplane f=constant
would be an n-dimensional sphere.) The most important special cases: the
whole space, a half space {(t, y) |0<t<oo,|y|>0}, and “‘infinite strips” (for
example, {(t, y) ||t—to|<00, |v|<2}) have the above property.
For systems of first-order differential equations one has the following
existence and uniqueness theorem (compare Theorem 1, Section 1.6).
Theorem 1. Let f be a vector function (with n components) defined in a
domain D of (n+1)-dimensional Euclidean space. Let the vectors f, 0f/0y,
(k=1,..., n) be continuous in D. Then given any point (to, 9) in D there exists
an interval containing ty and exactly one solution , defined on this interval,
of the system
y =f(t, y) (4.46)
satisfying the initial condition (to) =M.
The reader is advised to refer to the discussion and examples following
Theorem 1, Section 1.6. All remarks and examples about the scalar equa-
tion, of course, apply to systems of differential equations.
Example 5. Discuss the problem of existence and uniqueness of solutions of the
initial-value problem for the system
Yo ys
y2=(cost) yy +0?ys
Vea via.
This system is of the form y’=f(¢, y) with y=(¥1, Y2, ¥3), f(t, y)=(2 +3, (Cost) yi +
+17y3, y;—Y2); hence f(¢, y) is continuous for |t|<0o, |y|<0o. Moreover, f/éy, =
(0, cost, 1), éf/éy,=(t,0, —1), 6f/Ay3=(1, 17,0), which are also continuous for
|t|<oo, ly|<oo. Thus D is all of four-dimensional (t, y,, ¥2, 3) space and by
Theorem 1, through any point (¢o, n) there passes a unique solution @ existing on some
interval containing ft). It can be shown, that the solution @ actually exists on the
interval — co <t<oo.
Exercises g
10. Discuss the existence and uniqueness of solutions of the system
Vi=)i
Yo=Vityo.
11. Find a solution @=(@,, $2) of the system in Exercise 10 which satisfies the initial
condition @,(—1)=1, ¢2(—1)=0. Discuss the interval J on which the solution
exists.
CHAPTER 5
Eigenvalues, Eigenvectors,
and Linear Systems
with Constant Coefficients
We have seen in Chapter | how to solve the scalar equation y’=ay, and we
know that every solution is of the form ec, where c is a constant. In this
chapter we will learn how to find a fundamental matrix of the system y’ = Ay,
where A is a constant n x n matrix. The explicit calculation of a fundamental
matrix will lead us naturally to the study of eigenvalues and eigenvectors of
matrices. As in Chapter 4, some knowledge of linear algebra is essential,
but for students with this knowledge, this chapter, which contains the results
of Sections 3.4 and 3.5 as very special cases, can be studied instead of those
sections. We emphasize that the techniques discussed in this chapter are not
applicable to systems for which the coefficient matrix is not constant.
5.1 THE EXPONENTIAL OF A MATRIX
In order to find a fundamental matrix of the system

y =Ay, (55)
we first need to define the exponential of a matrix. If M isa nxn matrix, we
define the matrix exp M (or e™) to be the sum of the series
M2 M? Mé . Mé
eM aM ce tee
3 k! ie (52)
S i °o
where / is the n xn identity matrix. (Note that M° =/ and 0!=1.) To justify

this definition, we must show that the right-hand side of (5.2) makes sense.
It is not difficult to define a suitable notion of convergence of a series of
matrices and to show, using this definition, that exp M is well defined for
every matrix Me F,,,. This is done in Appendix 4.
165
166 Eigenvalues, Eigenvectors, and Linear Systems
An important property of the exponential matrix is that if P, Me F,,, and

if P and M commute (MP= PM), then
exp(M+ P)=exp M exp P. (5.3)
To prove this, we apply the definition (5.2) to the left-hand side of (5.3).
We obtain ;
INGER:
exp(M + P)= »; on (5.4)
k=0
By the binomial theorem and MP= PM,

k
k!
(M+Ph=) k __
MP
; i pk-l
[=0
(If x, y are real or complex numbers and k>0 is an integer, the binomial
theorem states that (x+y)*=Yieo K!/[/'(K—-D!] xy". If x and y are
matrices which commute, the same result holds.) Therefore, canceling k!,
we obtain _ :
M! Pk!
Oae=0
On the other hand,
M' . Pi
expM-expP= 1 aT:
Iie j=0 ,
By multiplication of absolutely convergent series, we have
exp Mien PC.

k=0
: (5.6)
M'! pk-1
Gr= Sores .
I! (k—D)!
1=0
Comparison of (5.6) with (5.5) proves (5.3).

A useful property is that if T is a nonsingular n xn matrix,
TL (exp.M) P= exp( han). (5.7)
Exercises
1. Verify (5.7). [Hint: Use (5.2).]

2. If MeF,,, show that:
a) exp (c,\M+c,M)=exp c,M exp c,M for any ¢, c.6F.
5.1 The Exponential of aMatrix 167
b) (e%)"* =e" ™,
c) (e“)K =e, where k is any integer.
d) e°=J, where 0 is the nxn zero matrix.
We are now ready to establish the basic result for linear systems with
constant coefficients
y =Ay. (5:1)
Theorem 1. The matrix
P(t)=exp At (5.8)
is the fundamental matrix of (5.1) with ®(0)=I on —20<t<o.
Proof. That @(0)=T/ is obvious from (5.2). Using (5.2) with M= At (well
defined for — 1 <t<o and every nxn matrix A), we have by differentia-
tion’
; fh GRO wi ke Bed saa
(exp At) Seger mye earn at exp At,
—a<t<o. Therefore, exp Af is a solution matrix of (5.1) (its columns

are solutions of (5.1)). Since det &(0)=det/=1, Theorem 2, Section 4.3,
shows that ®(t) is a fundamental matrix of (5.1). I
It follows from Theorem | and Eq. 4.30 (Section 4.3) that every solution
of the system (5.1) has the form
o(t)=(expAt)he (—aw<t<oo) (5.9)
for a suitably chosen constant vector c.
Exercises
3. Show that if ¢ is that solution of (5.1)satisfying (to) =n, then (t)=[exp A(t—to)|
—00 <f<00.
4. Show that if &(t)=e*", then @-'(t)=e“.
We now proceed to find some fundamental matrices in certain special
cases; that is, we evaluate exp At for certain matrices A.
Example 1. Find a fundamental matrix of the system y’=Ay if A is a diagonal

matrix, dj 0
0 au;
t It is easy to prove that the familiar theorems on differentiation of power series
(Section 6.2) with real or complex coefficients hold essentially without change for power
series having n xn matrices as coefficients.
From (5.2),
d, Oe d? 2
exp At=I+ cr, —-+ ob: =a
; ata a2| 2!
dj 0) -
+ +
0 dk |
exp d,t 0
as exp dt
0 exp d,t
and by Theorem | this is a fundamental matrix. This result is, of course, obvious, since
in the present case each equation of the system is y,=d,y, (k=1,...,) and can be
integrated separately.
Example 2. Find a fundamental matrix of y’=

Ay if
maleic
BIG 3
Since
and since these two matrices commute, we have
exp At=ex fy.0 tmex 0

PALES HONS Plo oSF
Pier et! Pr cs aN ia pe.

hOn te OFF0 Qe (O\, ait ‘
But
OP 1512 hao
Ol Quiles baa
and the infinite series terminates after two terms. Therefore,
exp At=e* lo i
and by Theorem | this is a fundamental matrix.

2.2 Eigenvalues and Eigenvectors of Matrices 169
Exercises
5. Find a fundamental matrix of the system y’=

Ay if
—2 I 0
A= 0 -2 |
0 0 -2
and check your answer by direct integration of the given system.

6. Find a fundamental matrix of the system y’= Ay if
Myth 2 0ek0
A= ee
0 0
where 4 is ann Xn matrix.
7. Find a fundamental matrix of the system y’= Ay, where A is the n xn matrix
8. What is wrong with the following calculation for |an arbitrary continuous matrix
A(t)?
=exp(|4(s) is)|= A(t) exp (|40) is),
so that exp (jf, A(s) ds) is a fundamental matrix of y’= A(t) y for any continuous
matrix A(t).
9. Consider the system
LY Ave
where A is a constant matrix. Show that |t\4=e4!°*!"! is a fundamental matrix for 140
in two ways: (i) by direct substitution, (ii) by making the change of variable |t|=e’.
5.2 EIGENVALUES AND EIGENVECTORS OF MATRICES
You will have noticed that the examples and exercises presented so far,
all of which involve the calculation of e4‘, are of a rather special form.
In order to be able to handle more complicated problems and in order to
obtain a general representation of solutions of (5.1) (that is, if we want to
evaluate explicitly the entries of the matrix exp(Az), we will need to intro-
duce the notions of eigenvalue and eigenvector of a matrix.
To motivate these concepts, consider the system y’= Ay, and look for a
solution of the form
b(j=clcrmeZ0:
where the constant 4 and the vector c are to be determined. Such a form
is suggested by the above examples. Substitution shows that ec is a solution

if and only if
hee = Ae“e.
Since e*“ £0, this condition becomes
(AI— A) c=0
which can be regarded as a linear homogeneous algebraic system for the
vector c. By elementary linear algebra (see for example [3], Theorem 1,
Section 3.8 and Theorem 2, Section 4.5) this system has a nontrivial solution
if and only if A is chosen in such a way that
det (AI — A)=0.
This suggests the following definitions. Unless otherwise stated we shall
assume that our field of scalars is the complex numbers.
Definition 1. Let A be a real or complex nxn matrix. An eigenvalue of A is
a scalar 4 such that the algebraic system
(AI— A) x=0 (5.10)

has a nontrivial solution. Any such nontrivial solution of (5.10) is called an
eigenvector of A corresponding to the eigenvalue 4.*
Definition 2._ The polynomial of degree n,
p(A)=det (AI — A)
is called the characteristic polynomial of A.**
Therefore, the calculation preceding Definition 1 shows that e*‘c is a solu-
tion of the linear system y’ = Ay if and only if A is an eigenvalue of A and ¢ is
a corresponding eigenvector. We will return to a discussion of the system
y’ = Ay in Section 5.3 after we have become familiar with properties of eigen-
values and eigenvectors.
In view of the remarks immediately preceding Definition 1, the eigen-
values of A are the roots of the polynomial equation p(A)=0. As p(A) is a
polynomial of degree n, there are exactly n eigenvalues, not necessarily
distinct. In particular, there is at least one eigenvalue and one eigenvector for
every matrix A. If A=Ap is a simple root of the equation p(A)=0, then A, is
called a simple eigenvalue. If A=A is a k-fold root of the equation p(A)=0
* Even though the entries of A are real, the scalar 4 may be complex (see Example |
following).
** The function p defined by the expression p(A)=det(A/— A) is a polynominal of
degree n. We shall tacitly assume that such determinantal polynomials obey the rules
of determinants.
5.2 Eigenvalues and Eigenvectors of Matrices 171
(that is, p(A) has (A—A )*, but not (A—A )**1, as a factor), then Ay is an eigen-
value of multiplicity k. Since the constant term in p(A) is p(0)=det (— A), if
A=0 is not an eigenvalue ofA, then p(0) #0, and in this case A is nonsingular.
Example 1. Find the eigenvalues and corresponding eigenvectors of the matrix
[3 J
The eigenvalues of A are roots of the equation
j=)
det(4—A1)= det |-#-02+34=0.
=) 3=/)
Thus, 4,;,,=3+5i. The eigenvector
[i
corresponding to the eigenvalue 2,=3+5/ must satisfy the linear homogeneous
algebraic system
= —S5i 5 uy aa
~Anu=| 3 5||u Fo
Thus, u,, u> satisfy the system of equations
and, therefore,
corresponding to the eigenvalue 2, =3—Si is found to be
ral
1
for any constant [.
Example 2. Find the eigenvalues of the matrix
oe pe At
Consider the equation det (A/— A)=0.
det [477 iy [R@-2 A441 = 7261490.

Thus, 4=3 is an eigenvalue of A of multiplicity two. To find a corresponding

eigenvector we consider the system
(31—
A) c=0
or
1 —| Cy 0 or Cy —C, =0
i —1 C2 0 ¢,—¢,=0.
Any vector c with components c;=c, is an eigenvector. Thus, the vector
ila
where « is any scalar, is an eigenvector corresponding to the eigenvalue A= 3.
Exercises
1. Compute the eigenvalues and corresponding eigenvectors of each of the following

matrices.
—-3 1 7
a) |e: ni b)} 0 4 -1 c) f |
0 0 2
a4 1 0 3 ers |
d) bs | ¢) 8 tee fy) 72 Ot
5 1 —-1l 1 -1 2
—-1 0 0 —1 LO
Be On = 0 hol aioe t eo
On S015 OF Ses
2) 13" 3
i) |4 —5 3] (eigenvalues are —1, 2, —2)
4 -4 2
1 21
j) el el (eigenvalues are —1, —1, 3)
2 1
0 1 0
k) 0 0 1] (eigenvalues are —1, —2, —3)
=o ith 6
0 1 0
l) 0 0 1 (eigenvalues are —1, —1, —2)
=O, os toed
A Om Onn
QO 4 i © @
m)|0 0 4 0 0
OR OMOR ASO
O205-07010
Sy Eigenvalues and Eigenvectors of Matrices 173
ad ee
n)
phvnasirig
eee
nea5 mi a !
[Hint: Characteristic polynomial is (A— 1)? (A+ 1)?.]
;
1 2 —-l +3
2. Show that if A is a triangular matrix of the form

CoRR Ole oe aye
pee 0 az2
: OR e:
0 Bo a en
the eigenvalues of A are A=a,;, where i=1,..., 7.
You will note that in Example | preceding, the two eigenvectors u and
v are linearly independent if «40 and B40, since
detfu, v]=del|on% 6
[2840
Therefore, the vectors u and v form a basis of (complex) two-dimensional
Euclidean space. However, in Example 2, the eigenvectors form only a one-
dimensional subspace. In applications to differential equations as well as
in matrix theory it is important to know whether the set of all eigenvectors
(corresponding to the various eigenvalues) of a given matrix A form a basis.
As Example | shows, even if the matrix A is real, the eigenvectors may
have complex components. Thus, we consider the eigenvectors as vectors
with complex components. If the xm matrix A has n distinct eigenvalues,
the corresponding eigenvectors form a basis for complex n-dimensional
Euclidean space.
Theorem 1. A set of k eigenvectors corresponding to any k distinct eigenvalues
is linearly independent.
Proof. We shall prove the theorem by induction on the number k of eigen-
vectors. For k=1, the result is trivial. Now, assume that every set of (p— 1)
eigenvectors corresponding to (p—1) distinct eigenvalues of a given matrix
A is linearly independent. Let v,,..., v, be eigenvectors of A corresponding
to the eigenvalues /,,...,A,, respectively, with 1;4A,; for i¢j. Suppose
that there exist constants c,, C2,..., C,, not all zero, such that
C1Vp + C2V, + #e,v,=0. (5.11)

We may assume c,#0. Applying A—A, J to both sides of this equation,
and using (A — 4,1) v,=(A;—A,) v,(j=L..., n), we obtain
C2 (Az — Ay) V2 +03(A3—A1) V3 +++ +0,(A,—A4) Vp =0. (5.12)
But y>, V3,..., ¥, are linearly independent by the inductive hypothesis, and
therefore c,(A;—A,)=0, where j=2, 3,...,p. Since 1;4A,, where j=2,
3,..., p, we have c;=0, where j=2, 3,..., p, and (5.11) becomes c,v, =0. Since
¥, £0, c, =0as well, which shows that v,,..., v, are linearly independent. This
proves the theorem by induction. I
We remind you that since the characteristic polynomial of an n x n matrix
is a polynomial of degree n there are at most n distinct eigenvalues. Since
the eigenvectors span a subspace of n-dimensional space, there are at most
n linearly independent eigenvectors. Of course, in any case, there exists at
least one eigenvector, since there is at least one (distinct) eigenvalue.
Example 3. Determine the subspace spanned by the eigenvectors of the matrix A
in Example 2.
As we saw in Example 2
is an eigenvector corresponding to the eigenvalue 2 = 3 of A of multiplicity 2 for any value

a0. Since 1=3 is the only eigenvalue of A, every eigenvector of A is of this form for
some «#0. Thus, the set of all eigenvectors of A is the subspace of two-dimensional
space spanned by the vector
ol
1
Clearly, this subspace is the line passing through the point (1, 1) and the origin.
Exercises
3. Determine the subspace, and its dimension, spanned by the eigenvectors of each
matrix in Exercise 1.
4. In the matrix A of Exercise 2 assume the diagonal elements a;;, where i=1,..., n,
are all distinct. Find the dimension of the subspace spanned by the eigenvectors
of A.
5.3 CALCULATION OF A FUNDAMENTAL MATRIX
We have seen in Theorem 1, Section 5.1, that exp tA is a fundamental

matrix of the linear system with constant coefficients, y’= Ay. We have also
seen in Examples | and 2, Section 5.1, how to compute exp (¢4 in certain
special cases; in particular, we have seen how to compute exp tA when A
is diagonal. We will now show how to compute a fundamental matrix ® of
the system y’= Ay when A has » linearly independent eigenvectors. This is,
in particular, true if all the eigenvalues of A are distinct. We postpone to
Section 5.5 consideration of the completely general case of an arbitrary
matrix A.
Suppose the matrix A has n linearly independent eigenvectors v,, V>,..., V n
corresponding to the (not necessarily distinct) eigenvalues /,, A,...,A n°
5.3 Calculation of aFundamental Matrix 175
Motivated by the discussion at the beginning of Section 5.2, we claim that

each vector function
o,(t)=exp(Ajt)v; j=1,....n
is a solution of y’= Ay, on — 00 <t<oo. For,
5 (t)=exp(A,t) A,v;
=exp(Ajt) Av;
=A exp(A,t) v;
= Ag, (t) Ha ee
where we have used the fact that Av;=A,v,, 7=1,..., n. Define
P(t) =[3 (0), b2(t),..-, On (0).

Since each column of @ is a solution of y’= Ay, @ is a solution matrix of
y = Ay on —o<t<oo. We have
det (0) =det ['v,,..., v,,] 40,
because the vectors V,, V>,..., ¥, are linearly independent. It now follows from
Theorem 2, Section 4.3, that det &(t)40 for —0<t<oo and that ®(t) isa
fundamental matrix of y’=Ay on —0<t<oo. We have therefore proved
the following result.
Theorem 1, Let A bea constant matrix (real or complex). Suppose V;, V>,.--,
v, are n linearly independent eigenvectors corresponding respectively to the
eigenvalues 11, 4,..., An. Then
@(t)=[exp(A,t) v1, exp(A2t) vo,..., exp(A,t) V,]
is a fundamental matrix of the linear system with constant coefficients y'= Ay
on —0 <t<oo. In particular this is the case if the eigenvalues A,, Ax, ..., An
are distinct.
Example 1. Find a fundamental matrix of the system y’=
Ay if
cle
LL
By Example 1, Section 5.2, 2; =3+5i and 2, =3-—Si are eigenvalues of A and
are (linearly independent) eigenvectors corresponding to 2,, 2, respectively. By

Theorem |
3+ sit jeS= Snr
2(0)=| fgoesn sas |
is a fundamental matrix on —00<f<oo.

In general, Theorem 1 does not yield exp tA, even though it does yield a
fundamental matrix &(t) of y’= Ay. By Corollary 2 to Theorem 2, Section 4.3, since exp
tA and @(t) are both fundamental matrices of y’= Ay on — 0 <t<oo, there exists a
nonsingular matrix C such that
exptaA=@(t) C. (5.13)
Setting =0 in (5.13), we obtain C=®~'(0). Thus, :

expta=@(t) & '(0). (5.14)
Example 2. Find exp #4 if A is the matrix in Example 1.

By (5.14), Example | and Example 3, Section 5.2, we have successively
eStsit je svt 1 i a
exp p tA=| . + Sitd
je(3 go desit ‘
hole
re. eft sit jes St 1 a
D jeSt5at eS SHt = 1
LP e3t5dt4 oB-Sit Sie ae

73 Hee veer souk) eet Yt GG sit |
sit cos St sin St

si —sin 5¢ cos 5t |’
If A is real, exp fA is real from the definition (5.2). Thus, Eq. 5.14 gives at the
same time a way of constructing a real fundamental matrix, whenever 4 is real.
Example 2 is a special case of this remark.
Exercises
1. Find a fundamental matrix of the system y’= Ay; also find exp tA for each of the
following coefficient matrices.
a) A |; Al (see Exercise 1(c), Section 5.2)
2 —3 3
d) A=| 4 —5 3 (see Exercise 1 (i), Section 5.2)
4 —4 2
Oat 0
e) A=} 0 0 1 (see Exercise 1(k), Section 5.2)
5.3 Calculation of aFundamental Matrix 177
2. Show that the scalar second-order differential equation u" + pu’ + qu=0 is equivalent
to the system y’= Ay with
0 1
A=
a a
and compute the eigenvalues 2,, 2, of A.

3. Compute a fundamental matrix for the system in Exercise 2 if 2, 42>, that is, if
p? #4gq, and construct the general solution of the scalar second-order equation.
An alternative way of producing a real fundamental matrix if A is a real

2 x 2 matrix is contained in the following exercises.
Exercises
eae
4. Given the matrix
0 1
show that 4*?=—J, 4=—A, A*=I and compute A”, where m is an arbitrary
positive integer.
5. Use the result of Exercise 4 and the definition (5.2) to show that
a, cos t sin ¢
e= F 5
—sin t cos t
2 9h tBass
[Hint: cos t=1-> +7 ,+-, sii¢=t———$4 |
6. Compute e'4, if
Zea
[Hint: Use Exercises 4 and 5.]
We close this section with the solution of the nonhomogeneous system

y'=Ay+g(t), (5.15)
where 4 is a constant matrix and g is a given continuous function on
— oo <t<oo. The variation of constants formula (Theorem 1, Section 4.4)
with ®(t)=exp tA as a fundamental matrix of the homogeneous system now
becomes particularly simple in appearance. We have ®~'(s)=exp(—s4A),
&(t) ®'(s)=exp[(t—s) A]; if the initial condition is (to)=n, >,(t)=
exp[(t—fo) A] n and the solution of (5.15) is
t
o(t)=exp[(t—to) A] nt {exp[(t—s) A] g(s)ds —-w<t<o, (5.16)

to
where e'“ is the fundamental matrix of the homogeneous system that we

can construct by the method shown in this section. Note how easy it is to
compute the inverse of ® and also &(t) & '(s) in this case. However, it may
not be possible to evaluate the integral in (5.16) explicitly except in special

cases.
Example 3. Find the solution ¢ of the system y’= Ay+g(o) satisfying the initial con-
dition
60-1 | if 4-|_3 | and 2=| ‘5|:

From Example 2 preceding, we have
ip ra=e| cos St sin Bf
—sin 5t cos St
Substituting in (5.16), we obtain (using t, =0)

yall COS O06 .SINLST 0
A Eo 5t cos i H
fana| —cOss(t—s) —sind(t—s) wes
+[e pee tees kb Jo

(0)
t
jz sin St 3¢-9 =3| cS (t=)

a beak © “| sin 5(t—s) os:
0
In this case, we can evaluate the integrals as follows
(=e E abe |=| cos St cos 5s+sin 5¢t sin | a

cos 5t —sin S5t cos 5s+cos 5t sin Ss
Using the formulas (these can be found by integration by parts)

t
—=<As
|e-4 cos 5s ds=——__ (—4 cos 5s+5 sin 5s)

16425 rae
te)
ri
e 4s s=t
Je~45 sin
sin 5s
Ss ds=———_
ds 16405 |(—4 sin
sin 5s—5
5s—5 cos
cos 5 5s) lee
0
we obtain
Cane 4
cos St i (—4—4 cos St+5 sini si+a}

—
e 4 5
in
+sin 5t \(
4 (—4 sinsin St 5t— 500s 5+}
a
o()=e EB St ae
COS St| ert 4
—sinsin St }a (—4— cos St+5 sini s+a}
=
e+ 5
+ cos St }a — cos ssa

(—4—4sIsin 5t—5 =
5.3 Calculation of a Fundamental Matrix 179
Further simplification seems pointless. You will note that even such a simple example
as above leads to a rather complicated answer.
Exercises
7. Find the solution @ of the system
y = Ay+g(t)
in each of the following cases:
a) oo=| |} 4-| sl at)=["| (see Exercise 1a),

0 1 = 0
b) 6(0)=0, A=] 0 0 it Si) 0
=o il —6 exe
(see Exercise le).
1 2 =3 3)
c) o(1)=] 0], A=|4 —5 3], g(t) arbitrary
0 pe Ce
(see Exercise 1d).
. By converting to an equivalent system, find the general solution of the scalar

equation
y’—-y=f (0),
where f is continuous, by using the theory of this section.

. Use the results of this section and Exercise 5 to find the general solution of the
scalar equation
Very — ple):
where f is continuous.
10. Suppose m is not an eigenvalue of the matrix A. Show that the nonhomogeneous
system
y =Ay+ce™
has a solution of the form
and calculate the vector p in terms of A and c.

11. Suppose m is not an eigenvalue of the matrix A. Show that the nonhomogeneous
system
k
y=Ayt+ j=0
) oem
has a solution of the form
[Hint: Show that the vector p, satisfies the algebraic system
(A—ml) pp= —&%

(A—ml) p;=((+ 1) pj41 -&, F220: ooo B=!
and that these systems can be solved recursively.]

12. Find the general solution of the system
ty’ = Ay+b(t).
[Hint: Use the result of Exercise 9, Section 5.1.]
5.4 TWO-DIMENSIONAL LINEAR SYSTEMS
In this section we will solve completely two-dimensional systems of the form

y =Ay, (5.17)
where A is a 2x2 matrix. More precisely, if we choose a nonsingular
constant 2 x 2 matrix 7, then the change of variable
Y= 77 (5.18)
transform (5.17) to the system
y=T7—AT2),
so that
2’ =(T~1AT)z. (5.19)
We now show that it is always possible to choose the matrix Tin such a way
that the resulting system (5.19) can be solved easily. If W(t) is a solution of
(5.19), then (t)=7(t) is a solution of (5.17), using (5.18). Conversely,
since 7 is nonsingular, every solution of (5.17) gives a solution of (5.19).
In other words, to solve (5.17) completely, it suffices to construct the matrix
T in such a way that (5.19) can be solved.
Our first result concerns the case where the coefficient matrix A is complex.
Theorem 1. Let A be a complex 2 x 2 matrix. There then exists a nonsingular
complex matrix T such that T~' AT has one of the following forms:
i) [ , Aen
Proof.
of Case (i) arises when A has two distinct eigenvalues A, py. In this
5.4 Two-Dimensional Linear Systems 181
case the result is an immediate consequence of Theorem | and of the follow-

ing calculation. Let v, and y, be the (linearly independent — see Theorem 1,
Section 4.6) eigenvectors corresponding to the distinct eigenvalues A, p,
respectively. Define the matrix
T=([v,, ¥,]-
Then
Tr*AT=T “Aly, ¥,j=T) -(ety, Ay)

=T “(dv nv, J=T“ver y= fi ab
where by matrix multiplication, using T~'T=J, since vy, is the first column
of T, T~'y, is the first column of J, and similarly, since v, is the second
column of 7, T~'v, is the second column of J.
Case (ii) arises when 4 has an eigenvalue 4 of multiplicity two for which
there are two linearly independent eigenvectors. In this case, the result is
found by the same calculation used in (i), with yw replaced by 4.
Case (iii) arises when A has an eigenvalue / of multiplicity two but the
subspace consisting of all eigenvectors of A has dimension one; that is,
any two eigenvectors of A are linearly dependent. ‘hen there exist nonzero
vectors which are not eigenvectors of A; at the same time we recall that
there exists at least one eigenvector of A. Let v be a nonzero vector which
is not an eigenvector of A and let
u=(A—Al)v. (5.20)
Since v is not an eigenvector, and A—AJ is not the zero matrix, u#0; we
will show that u is an eigenvector.
We first assert that the vectors u and y are linearly independent. Suppose
then there exist constants c,, c, such that
cyu+c,v=0. (5.21)
Since u and v are both different from zero, either c, and c, are both zero (in
which case there is nothing to prove) or they are both different from zero.
Using the definition of u and the fact that c, #0, we may rewrite (5.21) as
¢,(A—Al) v+-c,v=0
(re )opo[-b-2)
or
This says that A—c,/c, is an eigenvalue of A. Since A is the only eigenvalue,

c,=0, which implies c,=0. Thus, u and v are linearly independent and
span two-dimensional Euclidean space.
Let x be any eigenvector of A; we may therefore write

x =au-+ by.
If a=0, then x is a multiple of v, and therefore v is an eigenvector (but by
hypothesis it is not); thus a#0. Now,
0=(A—AI) x=(A—AI) (au+ bv)
“fot
=a(A—Al) u+b(A—Al) v=a(A—Al) ut bu
f(t
This says that A—b/a is an eigenvalue, and since 4 is the only eigenvalue,
b=0. Therefore, x is a nonzero multiple of u, and u must be an eigenvector.
Now, we define
T =[u, v] [i a (5.22)
V2
As in case (i),
po diee I aes
U;,V2—Uz2V, —UuUz uj;
We have, using (A — AJ) v= Av—Av and (5.20), (5.22)

T-'AT=T~'[Au, Av]
=T~'[Au,u+dv]=T~* {[0, u] + [Au, Av}
ST -o(yaEny |r 4 fer
Sah ape cel fini Mais

U,V2—UnV, | —U2 Ur TO rs
ie | bei
=41+]9 allo |
With reference to the above proof, you should note that in showing
that u=(A —2/) vis an eigenvector, we have actually shown that (A —1/)? w=
0 for every vector w. For, we may write w=au-+bv because u, v form a
basis. Since u is an eigenvector, we obtain
(A—Al) w=a(A—Al) u+b(A—Al)

=b(A-Al)v.
Hence,
(A —Al)’w=(A—Al) b(A—Al) v=b(A—Al) u=0
because of (5.20) and the fact that u is an eigenvector of A.

We remark that if 4 is a real matrix and ifits eigenvalues are real, then the
matrix T constructed in each of the three cases above is real. However, if A
is real but has complex eigenvalues (necessarily complex conjugates), the
matrix JT will not be real and it is of interest to learn the simplest form of
T~' AT which can be achieved with a real matrix T. The answer lies in the
following result.
Theorem 2. Let A be areal 2 x2 matrix with complex conjugate eigenvalues
a+ifp. There then exists a real constant nonsingular matrix T such that
i -1 ar=|
= a B
il
Proof. Let u+iv be an eigenvector corresponding to the eigenvalue

a+if, where u and vy are real. If v=0, so that this eigenvector is real,
Au=(«+if) u
the left side of which is real and the right side of which is not real. Thus,
v0, and a similar argument shows that u#0.
We define the matrix 7 with columns u and vy,
= _f[ ur YY
T =[u, n-[i al
In order to show that 7 is nonsingular, we must show that u and vare linearly
independent. Suppose not; then there exist real constants c,, c, both different
from zero such that
cyut+c,v=0.
Since c, 0, and A(u+iv)=(«+if) (u+ iv), we have
Ati)=A( —2 vw) =(04)( 241) Nie

1
thus,
Av=(a-2 5)v.
Cy
Then v is a real eigenvector. But, as remarked above, u40 and therefore
this is impossible. Thus, u and v are linearly independent and T is non-
singular.
Taking real and imaginary parts in the equation
A(u+
iv) =(«+ if) (u+iv),
we obtain
Au=ou—fvy, Av=fu+ayv.
Therefore,
AT =[Au, Av] =[ou— fv, But av]

and
Tare 1 V2 —v, || au,—fv, Pu,t+av,|}_}| a B -

Uy au, — pv, Bu,+av, —B al
U,V2—Ugv, | —U2
Consider now the system (5.19), where 7” 'AT=B has one of the forms
in Theorems | and 2. We can write down the fundamental matrix exp Bt in
each case. We have the following possibilities,
CASE i) If
alo ih
A 0
then
Eu 0
exp Br=|j a
CASE ii) If
alo a
A O
then
CASE ill) If
then
At At
exp B= of |
Further, if
ai)
B=|_§2284 with
: « and f real,
then
whos le COS Be IsimiBE
oo ae [ sinBt cos a
Exercise
1. Show that exp Bz has the form indicated in each of these cases. [Hint: For special
cases of the last one, see Examples 1 and 2, Section 5.3, or Exercises 4 and 5,
Section 5.3.]
In order to describe the behavior of many physical systems, it is con-

venient to picture all solutions of the two-dimensional systems geometrically.
For the system
/ Vy
y, = A
y,; =| |)
ia ,
this is conveniently done by regarding ¢ as a parameter and then picturing

the solutions in the (y,;, v2) plane. This plane is called the phase plane, and
the graphical representation of all solutions is called the phase portrait of the
system. Each curve in the phase portrait is called an orbit.
Definition 1, A point p with coordinates (p,, p2) is called a critical point of
the system y' = Ay if and only ifAp=0.
The origin is always a critical point. If A is nonsingular, the origin is clearly
the only critical point since the system Ax=0 has only the trivial solution.
Example 1. Construct the phase portrait for the system y’= Ay, where
“lcm eek
Se 0 [M1
Since det A 40, the origin is the only critical point. A fundamental matrix is
Cae)
exp at={) ea
Let (rt, n) be that solution of y’=Ay for which (0, n)=y. Then
eu nn
OE MS lexP a NS aA:
Here we have arbitrarily chosen fp =0. Notice that @(t—‘o, 9) is that solution passing
through the point n at =f). Let y=(7;, 12) be any point in the (y,, y2) plane. Then the
solution (t, n) for t>0 is represented by the parametric equations y, =, (=e “ni,
¥2=(t)=e *2 for t>0 and this represents the portion of the curve shown in Fig. 5.1
between n and the origin, as is verified by elementary calculus; the arrow indicates the
direction of increasing ft. Notice that the slope of the tangent to this curve, dy,/dy,,
also tends to zero as t> + ©, because
3t
dy. VI es 312 =F,
2t
= Ca
dy, yy —2mye- 2m
Similarly, t<0 represents the portion of the curve in Fig. 5.1 above the point y. It should
be noted that lim ,., , .(¢, n)=90; that is, both the solution and the orbit approach the
origin as t— + 00. Proceeding in this way by choosing various points of the phase plane
as initial points, we obtain the phase portrait of the system, shown in Fig. 5.2. Notice
that every orbit approaches the origin (as f— + 00).
Figure 5.1
Ye
Figure 5.2
Exercises
“f 3
2. Obtain the phase portrait of the system y’= Ay, where
sale ow
lee 2 0
5. Obtain the phase portrait for the scalar equation y’ +4y=0. [Hint: Use the system
¥;=2, ¥2= —4),. By a phase portrait of a scalar second-order equation we
mean the phase portrait of the equivalent first-order system.]
We shall now analyze the general rea/ two-dimensional linear system

y = Ay with constant coefficients. As at the beginning of this section, let us
make the change of variable y = 7z, where Tis a nonsingular constant matrix
(to be determined), and substitute, obtaining the system
ZT CAT 2 (5.19)
whose coefficient matrix T7~ ' AT is similar* to A. For simplicity, and because
this is the case that arises in applications most frequently, we consider only
the case det A#0. This means that zero is not an eigenvalue of A, and that the
origin is the only critical point. For the case when det A =0, the reader is
referred to Exercises 19 and 20.
As is done in Theorems | and 2, we may show that there is a real non-
singular matrix T such that T~'AT is equal to one of the following six
matrices:
i) E At where p<A<O0 or O<ypK<aA.
I whereA>0O or A<O.
: At pw<0<A.
: 1
iv) E | whereA>0O or A<O.
|2 I where v#0 and o>0, or o<0O.

6:
vi) le o where
v#0.
The Cases (v) and (vi) correspond to complex conjugate eigenvalues of 4,

o+tivand +iy, respectively. In the four remaining cases the eigenvalues A, yu
are real. We obtain the possible phase portraits of (5.19) by assuming that
T~!AT is one of the forms (i){vi).
* For a discussion of similarity, see Appendix 5.
CASE (i) (This is essentially Example | and Exercise 2 above.) The

solution of (5.19) through the point (7;, 72)4(0, 0) at t=0 is
en,
(=| Sor|
y2
Figure 5.3
If u<A<0, we have (t)>0 as t>+00 and we obtain the phase portrait

in Fig. 5.3 with every orbit tending to the origin as t>+oo. If 0<p<A,
we obtain the phase portrait in Fig. 5.4 with every orbit tending away from
the origin as t>-+00. Arrows indicate the section of increasing ¢. The
origin in Figs. 5.3 and 5.4 corresponding to Case (i) is called an improper
node.
2
" Figure 5.4

Exercise
6. Justify the phase portrait for Case (i) with 0<p</.
CASE (ii) Here the solution of (5.19) through (11, 12)4(0, 0) at t=0 is
* en,
(t) =|
and if A>0, we obtain the phase portrait in Fig. 5.5, whereas the case
2<0 corresponds to Fig. 5.6. Note that all orbits are straight lines tending
away from the origin if 2>0 and toward the origin if 1<0.
The ratio $(t)/P, (¢) if 7, 40 is constant, as is d, (t)/b2(t) if yn.=0. The
origin in Case (ii) is called a proper node.
y2 y2
Figure 5.5 Figure 5.6
CASE (ili) Here

en:
(=| 2
with w<0 and j>0, is the solution through (7,, 72) at t=0. Now, as
t>+0, o,(t)> + according as 7,>0 or 4,<0 and ¢,(t)>0 as t> + 0.
It is easy to see that if |A|=|y|, the orbits would be rectangular hyperbolas;
for arbitrary 1>0, u<0O they resemble these curves as shown in Fig. 5.7.
Quite naturally, the origin in Case (iii) is called a saddle point.
y2
Figure 5.7
Exercise
7. Construct the phase portrait in Case (iii) if 2<0 and p>0.
CASE (iv) Here
$(1)= i ea
N2
@
is that solution passing through (7, 72) at t=0 and if A <0 the phase portrait
is easily characterized by the fact that every orbit tends to the origin as
t>+0oo and has the same limiting direction at (0,0). For dy,/dy,=
5/61 =[Ab2/(Ad, + G2) ] 0 as t> + 00 (see Fig. 5.8). The origin in Case (iv)
is called (as in Case (i)) an improper node.
¥2
aya
Figure 5.8
Exercise
8. Construct the phase portrait in Case (iv) with 2>0.
CASE (v) Here the solution, for the case ¢>0, passing through the point
(11, 2) at t=0 is
|
Ny COS vt+n sin vt
o(t)= el—n, Sinvt+n, Cos vt
Let p=(n7 +n3)"7, cos a=n,/p, sin a=n>/p. Then

b()=e"| p cos(vt—a) I
—p sin(vt—«)
Letting r, 0 be the polar coordinates, y, =r cos, y,=r sin@, we may write

the solution in polar form r(t)=pe%, 0(t)=—(vt—a). Eliminating the
parameter t, we have r=C exp[(—a/v) 0], where C=p exp[(a/v) «]. Thus,
the phase portrait is a family of spirals, as shown in Fig. 5.9, for the case
5.4 Two-Dimensional Linears Systems 191
o>0, v>0 and the origin is called a spiral point. In this case, the orbits
tend away from zero as t— + 00 (or, equivalently, approach zero as t— — 00).
¥2
Nv
Figure 5.9
Exercise
9. Sketch the phase portrait for the Case (v) in case <0, v<0.
¥2
Figure 5.10
CASE (vi) This is a special case of Case (v) with o=0. From the above
formulas we see that the orbits are concentric circles of radius p oriented as
shown for v>0 in Fig. 5.10. The origin is called a center.
10. Sketch the phase portrait Case (vi) when v <0.
We observe from the possible cases considered above that all solutions of
(5.19) and also their orbits tend to the origin as t— + co if and only if both
eigenvalues of A have negative real parts; in this case we say that the origin
is an attractor of the linear system. (5.19).
Exercises
Sketch the phase portrait of each of the following scalar equations by converting to
an equivalent system. Identify the origin and decide whether it is an attractor.
eee x0 12. x”—3x'+x=0
13. x"+3x'+x=0 14. x"+3x'-—x=0
15. x”—3x'+2x=0 16. x” +3x'+2x=0
17. x” —2x'+x=0 18. x” —x'—6x=0
19. To illustrate the complexity of the case when the origin is not the only critical
point of a linear system consider the system
Vi= Va v2
yz=2y—2y2.
a) Show that there is a line of critical points.
b) Sketch the phase portrait.
[Hint: y,=2y;, and the eigenvalues of the coefficient matrix are 0 and —1.]
20. Repeat as much as you can of Exercise 19 for the system
Vi =4y1V1 +4422,
V2 =41V1 + 422y2,
where 4, ,@22—@,24,,=0, but not all of a,,, a5, @21, a>» are zero.
21. For each of the following systems y’= Ay with A given below, sketch the phase
portrait, identify the origin as a node, saddle point, spiral point or center and decide
whether or not the origin is an attractor.
A a-[; i b) a=] if
5.5 SOME POPULATION PROBLEMS
In Section 1.1, we formulated some models to describe the population of a

single species. Here we shall discuss some problems involving populations of
two interacting species. First we consider a predator-prey problem. Let x(t)
Ses) Some Population Problems 193
be the number of members of the first species, known as the prey, at time f,
and let y(t) be the number of members of the second species, known as the
predator, at time ¢. If there were no predators, then according to one of the
models formulated in Section 1.1, the growth rate x’(t)/x(t) of the prey
population satisfies an equation of the form
x'(t)/x(t)=A—ax(t),
where / and a are positive constants. Now, however, we assume in addition
a negative growth rate (that is, a death rate) proportional to the size of the
predator population at time t. Thus the growth rate of the prey population
satisfies an equation of the form
x'(t)/x(t)=2—ax(t)—by
(0),
where 4 and a are positive constants. Now, however, we assume in addition
a negative growth rate (that is, a death rate) proportional to the size of the
predator population at time ¢. Thus the growth rate of the prey population
satisfies an equation of the form
y (O/y(t)="+ ex(t)—dy(t),
where py, c, d are positive constants. We have now obtained a model for the
predator-prey problem, namely a system of two first-order differential
equations x’ =x(A—ax—by)
y'=y(u+cex—dy). ee)
We cannot solve this system explicitly, but we can examine the possibility
of an equilibrium solution. To do this, we find the critical points of the
system (5.23), the points (xo, Vo) in the phase plane where the right-hand
sides of both equations in (5.23) are zero. Thus we must solve
Xo (A—ax 9 —byo)=0,
VYo(U+cxXo —dyo)=0.
It is easy to see that there are four critical points, shown in Fig. 5.11.
iy;
MetaCD yea)
\— ar — by =0
Figure 5.11
One is at the origin, a second is at the intersection of the line A—ax —by=0
with the x-axis, a third at the intersection of the line 4+cx —dy=0 with the
y-axis, and a fourth at the intersection of the lines A—ax—by=0 and
u+cx—dy=0. The reader will note that the line 1—ax —by=0 has negative
slope and positive x- and y-intercepts and that the line w+cx—dy=0 has
positive slope, positive y-intercept and negative x-intercept because all the
constants A, a, b, u, c, d are positive. This means that the fourth critical point
is in the first quadrant of the (x, y) plane, as shown in Fig. 5.11. We denote
this critical point by (xo, yo). Since this is the only equilibrium position which
does not involve extinction of either predators or prey or both, we examine
it more closely.
We would like to show that solutions whose initial values are sufficiently
near the equilibrium position tend to the equilibrium position as too. The
reader should note that the other three equilibrium positions are biologically
plausible and have been observed experimentally. If a solution tends to
the origin, this may be interpreted as the situation where the predators first
kill all the prey and then die themselves because of a food shortage. We
concentrate on the equilibrium position (xo, yo) because the other three
equilibrium positions are mathematically trivial, not because of any lack of
biological significance.
To study the behavior of solutions near the critical point (x9, Yo), we
make the change of variable which transforms (x9, Vo) to the origin
u=X—Xo, V=y—Y)o-
Substitution in x’ =x(A—ax — by) gives
u! =x’ =(xo +u) [A—a(xp +u)—b(yo + v)]
=X (A—aXxo — by) + Xo(—au—bv)+u(A—axy
—byo) +u(—au—bv).
Since (Xo, Yo) satisfy the algebraic system
ca ers eo)
This reduces to
u’ = —axou—bx ov —au? —buv.
In a similar way, substitution in y’=y(u+cx—dy) gives
v’ =Cyou—dyou+cuv—dv?.
Thus we obtain the system
Gu [= 12 2a
Br epeteine aes ce (ees)

It is possible to show that the solutions of (5.25) with u(0) and v(0) small
behave in the same way for large ¢ as the solutions of the linear system
5.5 Some Population Problems 195
obtained from (5.25) by neglecting the quadratic terms in (5.25), namely

u’ = —axogu—bxov,
v’=cyou—dyov.
(5.26)
For a proof of this, see, for example [2, Chapter 4], or [7, Chapter 15].
Our problem is now reduced to the analysis of the phase portrait near the
critical point uw=0, v=0 of the linear system (5.26). The values of x9 and yo
may be found explicitly by solving (5.24). The reader may easily determine
the eigenvalues of the coefficient matrix
[teh]
ex SOX
529
to be
ae (axo + dyo) + [(axo saedy) —4bexoyo] ia

: 2
5.28
= —(aX9 + dyo)+[(axo +dyo)?
5 —4x9¥o(be + ad) ]"/? : ra
Exercise
1. Find the eigenvalues of the matrix A given by (5.27).
Since X9>0, vo>0, a>0, b>0, c>0, d>0, it is clear that
(axo + dyo)” —4xoyo(be + ad) <(axo + dyo)*

and thus we see from (5.28) that the eigenvalues of A both have negative
real parts.
If (aX9
—dy)? —4bcxoVo
=0, the eigenvalues of A are real and negative.
Thus we are in one of the Cases (i), (ii), or (iv) of Section 5.4, and the critical
point w=0, v=0 is a node. The critical point v=0, v=0 of (5.26) or (5.25)
corresponds to the critical point x = xo, y=o of (5.23), and we see that solu-
tions of (5.23) with initial values near (x9, Yo) tend to (xo, Yo) with a definite
limiting direction.
If (axX9 —dyo)? —4bcxo yo <0, the eigenvalues of A are complex, and we are
in Case (v) of Section 5.4. Thus the critical point u=0, v=0 is a spiral point.
In this case, solutions of (5.23) with initial values near (xo, Yo) spiral about
(x, Yo) as they tend to (xo, Yo).
Another closely related biological problem is the question of popula-
tions of two species competing for the same food supply. The governing
equations are very similar to the predator-prey equations (5.23). In the com-
peting species problem, with x(t) and y(t) representing the respective
populations of the two species at time ¢, we have the system
x'=x(A—ax—by),
y’=y(u—cx—dy). (5.29)
The terms —bdy in x’/x and —cx in y’/y correspond to the assumption of
a decrease in the growth rate proportional to the population of the other
species caused by the decrease in the food supply. Mathematically the
only difference between (5.29) and (5.23) is that (5.29) has a term —cx
where (5.23) has a term +cx. The lines A—ax—by=0 and up—cx—dy=0
now both have negative slope and may not have an intersection (Xo, Yo)
in the first quadrant of the phase plane. (Of course, negative populations
have no biological meaning, and only the first quadrant of the phase plane
is of interest.) The two possibilities are illustrated in Fig. 5.12, Fig. 5.12(a)
being the case in which there is an equilibrium in the first quadrant and
Fig. 5.12(b) being the case where there is not.
My of
\— ar — by = 0
wu — cu —dy=0 [Cl
— Ay—()
Nar — by =" 0
(a) (b)
Figure 5.12
Exercise
2. Show that the solution (xo, Vo) of the algebraic system 2 —ax—by=0, u—cx—dy=0
IS X9=(Ad—b)/(ad—bc), yo=(wa—Ac)/(ad—bc) and deduce that there is an
equilibrium position (Xo, Yo) with x9 >0, yo>0 if and only if the ratio 7/y lies be-
tween the ratios a/c and b/d.
If there is an equilibrium position (x9, yo) with x9>0, yo>0, then the
analysis may be carried on exactly as in the predator-prey problem. The
coefficient matrix of the linearized system is now
A= —aXo —bxo
—CYo —dyo
5.6 The General Case 197
with eigenvalues
—(axX9 +dyo)+[(ax9 —dyo)? + 4bcxoyo |"?
2
A=
—(axo +dyo)+[(axo +dyo)? + 4X9Vo (be — ad) ]"/?

5 :
Exercises
3. Show that if be<ad, both eigenvalues are real and negative and deduce that the
critical point (Xo, Vo) is a node to which solutions tend.
4. Show that if bc> ad, one eigenvalue is real and negative while the other is real and
positive. Deduce that in this case the critical point is a saddle point, which
means that one of the competing species eventually dies out.
5.6 THE GENERAL CASE
In Section 5.3, we learned how to find a fundamental matrix of the system

y =Ay in the case when the matrix A possesses n linearly independent
eigenvectors. (In particular, this covers the case when the eigenvalues of A
are all distinct.) We then learned how to determine the matrix exp ¢A in this
case, and we found that each element of exp ¢A is a linear combination of
exponential functions. In Section 5.4 we treated the general 2 x 2 case, and
we observed that if the matrix A has fewer than two linearly independent
eigenvectors, then exp tA contains at least one term of the form te” (see Case
(iii) of Theorem 1, Section 5.4). This shows that the solution may have a
more complicated form than that suggested in Section 5.3.
In this section we will determine the form of exp tA when A 1s an arbitrary
nxn matrix. We note that the following method, which leads to the formula
(5.34), is always applicable and contains the results of Sections 5.3 and 5.4
as special cases. However, we warn you that this section is somewhat more
difficult.
We will need the following result from linear algebra; its proof may be
found in Appendix 5.
Let A be a (complex) nxn matrix. Compute 4,, A2,..., A,, the distinct
eigenvalues of A with respective multiplicities n,,,...,,, where n, +n,
+-+-+n,=n. Corresponding to each eigenvalue A; of multiplicity n; consider
the system of linear equations
(A-A,I"x=0 j=1,2,...,k. (5.30)
The solutions of each such linear system obviously span a subspace which
we call X,, where j=1, 2,..., k. The result from linear algebra needed tells
us that for every x in Euclidean n space there exist unique vectors X,, X2,.--, Xx»
where x,;€X ;(j=1,..., k), such that
X=X, +X, +--+. (5.31)

It is important to know that the linear algebraic system (5.30) has n;
linearly independent solutions so that the dimension of the subspace _X;; is ;.
We note that if all the eigenvalues of A are distinct, that is, if each n;=1,
where j=1,...,k and k=n, then the vectors x;,.%>..... X, ate suitable
multiples of fixed eigenvectors that are linearly independent and span
Euclidean n space. Thus, if v,,..., v, is a fixed set of linearly independent
eigenvectors of A and if x is an arbitrary vector, the vectors x, are given by
x,;=cj,V; for some scalars c,, where J lucas. 1.
It may, in fact, happen that in (5.30), (4 —A,/)*x=0 for all x in X; and
q<n,, as the following example demonstrates.
Example 1. Consider the matrix
4 1 0 0 0
0 + 1 0 0
A=]|0 0 4 0 0
0 0 0 4 0
0 0 0 0 4
Here n=5, and 2=4 is the only eigenvalue, with multiplicity 5; that is, k=1. Since
there is only one eigenvalue, no decomposition into subspaces is necessary. According
to the theorem from algebra n,=5, so that certainly (A —4/)° =0. However, as you
can easily verify (A —4/)?=0 but (A —4/)? 40. Therefore, g=3 <n, =S.
To apply this theory to the linear system y’= Ay, we look for that solu-
tion (t) satisfying the initial condition @(0)=y. By Theorem 1, Section 5.1,
(t)=e'4y and our object is to evaluate e'n explicitly, that is, to see exactly
what the components of @ are. From the definition of exp(tA) (see (5.2),
Section 5.1) we know that in general the components of exp(tA) y will be
infinite series and thus hard to evaluate. The algebraic theory we have just
discussed enables us to decompose the initial vector yn in such a way that
the components exp(tA) n can be written as finite linear combinations of
exponentials and powers of t. We compute A,, A,,..., A,, the distinct eigen-
values of A of multiplicities n,,,,..., n,, respectively. We apply the theorem
to the initial vector n and in accordance with (5.31), we have
WS Maa es (5.32)
where vy; is some suitable vector in the subspace X;, with j=1,..., k. Since the
subspace X; is generated by the system (5.30), v,; must be some solution of
(5.30). Now, by (5.32), e4n=)%_, e'4v,, and we may write (and here is the
trick!)
e'4y =exp(A,t) exp[(A—A,I) t] v;
2
A I)? +e + (n,;—1)!
pri}
=exp(A) [I+t(A—A) +—(A—

m1 (4A ye 9;
for —0<t<+, where the series in parentheses terminates because Vv; 1s

a solution of (5.30); (thus the term (4 —/,/)"v,=0 and all subsequent terms
in the expansion of the matrix exp[(4—A,/)t] are zero). Observe that the
vectors w,=(A—A,/Pv;, for p=0, 1,...,n;—1, belong to the subspace X;
because
(A—A,I)!w,=(A—A,I™ [(A—Ajl)?v,=(A—AL)
Pv,Y*
=0.
Thus, the vector e'v, remains in X; for each t, — 00 <t<oo. Applying the
above calculations to the solution o(t)=e'4y of y’= Ay, we have
o(t)=en=e'4 y v= 3 ey,
j=1 j=1
k
sy exp(/,t) [i+a—a)+ Re
or, finally, the solution @ satisfying @(0)=n is
nj-1
o=) exp(A,t) 1 faa Vue 00 =O 9 (533)

j=1
We point out again that if (A —,/)” =0, where g;<n;, then the sum on / in
Eq. (4.65) will contain only g;, rather than n,, terms. This formula also tells
us precisely how the components of the solution behave as functions of ¢ for
any given coefficient matrix A.
If A has only one distinct eigenvalue, there is no need to decompose as
in (5.32). In this case we know that (A —AJ)" x=0 for every x; that is, that
(A—AI)" is the zero matrix. Therefore, from the series definition of exp 7A,
we have
n—1}
t' é
exptA=e"24 "=e" > = (A—Aaly’; (5.34)
i!
i=0
that is, the series terminates after at most terms and exp/A 1s easily com-
puted.
Example 2. Solve the initial problem y’= Ay, y(0)=9, if A is the matrix in Example 2,
Section 5.2. Also, find exp (tA).
From Example 2, Section 5.2, we know that 2, =3 is an eigenvalue of multiplicity 2.
In the above notation, n, =2. Therefore, only the subspace X, is relevant. We readily
calculate
a-si=|
Sil i and we also see that (A—3])
baw e020
-|5 Al
200 Eigenvalues, Eigenyectors, and Linear Systems
so that (5.30) is satisfied for every vector in x. Substituting in (5.33) with
ie n=[
we find o(t)=e*[1+t(A—31)] y
and therefore 11 u] ni ht +t(—n, +n2)

(t)th=e"21+t
} E& ae meer N2+t(—n: +n2)
5B5
Wee
is the solution with @(0)=y. To construct exp (tA) we may use the formula (5.34) and
we obtain (since (A —3/)? =0),
exptA =e* [I+(A—3]) ¢]
Saale enue pene eal
Alternatively, to construct exptA, we may use (5.35) as follows. We recall that exp tA
is that fundamental matrix which reduces to the identify at =0. Thus
1 0 1 0
exptA=exptA E | |=
[exer lo}exptA Hit
where the solution vectors
exptA | and exptA ny
are found from (5.35) by substituting first
and then
respectively. This gives the matrix already found above.
Xs =3x,—-X,+x;
Xp =2% + Xs
X3=X,—X2+2x;
which has coefficient matrix
Seals al
Ae Ome!
i Sk 2
Find that solution @ satisfying the initial condition
Ny
(0)=} 12] =n,
3
and also find exp tA.
The characteristic polynomial of A is det (AJ—A)=(2—1) (2-2), and therefore
the eigenvalues are 24, =1, 24, =2 with multiplicities n, =1, n,=2, respectively. In the
notation of (5.30), we consider the systems of algebraic equations
(A—I)x=0 and (A—2I)?x=0
in order to determine the subspaces X, and X, of 3-dimensional Euclidean space.
Taking these in succession, we have first
2 —1 1 2x1—X2+x3=0
(A—I)x=]2 =—1 I)x=0 or 2x,—x,+x,=0

ly e188) X,—xX,+x;=0.
Thus X, is the subspace spanned by the vectors
xX}
25;
X3
with x, =0, x,= x3, and clearly dim X, =1. Next,
with x, =x, and x; arbitrary; clearly, dim ¥,=2. You are advised to picture these
subspaces. Observe that the rank of the matrix A—/ is 2. Thus, by a well-known result
in algebra, [see, for example, [3], Theorem 2, Section 5.6], dim X, =3 —2=1. Similarly,
the rank of the matrix (4—2/)? is clearly | and dim X¥,=3—1=2.
We now wish to find vectors v,EX,, ¥,€X, such that we can write the initial vector
n as in (5.32):
n=YV; +YV>.
Since
v, €Xj,
0
vi a
R
for some scalar «, and since v,€ X,,
for some scalars f, y. Therefore,
so that B=n,, «+ B=n2, ~+y=N3. Solving these equations for a, 6, y, we find that
a=N2—M, B=, Y=N3—N2 +m, and
0 Ny
Vite ope ta Vvo= Ny
N2—- Na—-m +m
Thus, by the formula (5.33), we find that the solution such that @(0)= 1 is given by
o(t)=e'v, +e* (1+ t(A—2I)) vy,

0 1 -1 1 Ni
=e |y,—n,| te* | I4+t]2 -—2 1 Nn
N2—N L a0 n—notm
0 tt =f ¢ Ny o
=e |n.—n,| +e | 2t 1-—2t ¢ nN
N2—M t =F fli tia stacttia
To find exp tA, we put n successively equal to
1 0 0
0], Ile 0
0 0 1
in (5.36). We obtain the three linearly independent solutions that we use as columns of the
matrix
(1+1) e?! —te** te!
4 —e'+(1+t) et e' — te" te?!
ape ners e'—e2! e2t
Example 4. Find exp 74 if
ae il OY Oe @
ORAS tena O
Ale |i) @) 4. a) @
OOF O40
OOM OM One
Here n=5. Using the results of Example 1, we have (4 — 4)? =0, so that (A —4/)°x
=0 for any vector x and the initial vector y remains arbitrary. Since there is only
one eigenvalue (A =4), only the subspace X, is relevant and we have, from (5.34) (since
(A —4/)? =0),
t?
exptd=e"[I+t(A—4])+ 74 —4I)’].
Therefore
O10: O20 O40 pile05.6

OD 1 OTe a 10 F020 Oro
exptdA=e"/1+r]0 0 0 0 0 oT 0} ..0' 30°. 05.0
OwOraD® OVO £ 1cOLeOe 0 O20
0 0 0 0°90 OF Ov 20910
t?
l1t= 00
2!
ae OO) al ie OW
OOM OmO
Od O04) O
CFOs 0,70. 71
y =Ay+g(?),
where A is the constant matrix in Example 2 and where
with the initial condition o(0)=n.

From Example 2 we have
1—-t t
wiih atonal a
tA et
DOR in |-t a
1—(t—s)+e7 **(t—s)
exp[(t—s) A] eae ora ne
Note that it is not necessary to compute ® '(s) separately. Therefore, using (5.16)
(Section 5.3),
3t 1—t t 3t | b= s)he (Es)

o(t)=e |a) i fiee (plein eet Gard
There seems to be little point in evaluating the integrals.

204 Eigenvalues, Eigenyectors, and Linear Systems
Exercises
Find the fundamental matrix exp tA for each of the following systems y’ = Ay having
the coefficient. matrix given. Also find a particular solution satisfying the given initial
condition.
1. a-|, Al n=|3| (See Exercise 1c, Section 5.2.)
aft a} wf] 1 © 3 0
3. A=]8 1 -1]); n=|—2]. (See Exercise le, Section 5.2.)
Ss itt —7
3 -1 —4 2) 1
2 3 —2 -4 : ;
AA ea eee 5 n 1 (See Exercise In, Section 5.2.)
1 2 -1 -—3 0
[Note: The characteristic polynomial is (A—1)? (A+1)?.]
5. Find that solution of the system
Yi=Vityo2tsint, y,=2y,+cost
such that y,(0)=1, y.(0)=1. [Hint: Find a fundamental matrix of the homo-
geneous system; then use the variation of constants formula (5.16) in Section 5.3.]
Consider the scalar linear differential equation of second order
y’+py +qy=0, (5.37)

where p and q are constants. We can solve this equation as a special case of
the theory developed here as outlined in the following exercises.
Exercises
6. Show that Eq. (5.37) is equivalent to the system y’=

Ay with
Arad
and compute the eigenvalues 2,, 2, of A.
7. Compute a fundamental matrix for the system in Exercise 6 if 2, 42,; that is, if
p’ #4q, and construct the general solution of Eq. 5.37 in this case.
8. Compute a fundamental matrix for the system in Exercise 6 in the case
2,=2,=/; that is, p?>=4g, and construct the general solution of (5.37) in the
case p*=4q. Note that A—AJ/ is never zero in this case, so that the fundamental
matrix, as well as the general solution of (5.37) must necessarily contain a term
in te”.
9. Generalize the results of Exercise 6, 7, and 8 to the scalar equation
y"+Piy’ + poy +p3y=9,

where p,, P2, P3 are constants. (Needless to say, you are not expected actually
to solve a cubic equation.)
10. Find a fundamental matrix for the system y’= Ay, where A is the matrix.
a) 4-[3 | b) a=] |
1
e) 4=| 3raeSi Tak ty hl
11. Find the general solution of the system y’=Ay+b(r) in each of the following
raf!
Cases.
fe) ao o-||
b) 4-|—) | =| 4|
‘ |
ah
e) 4-[) il Bil= coe

12. Consider the system of differential equations
Via rt Va —
Via 2Zy1+V2t y2=9.
a) Show that this system is equivalent to the system of first-order equations
u’ = Au, where
Uy ui 0 | 0
u=]Uu,| = V4 . A= —4 4 y
U3 V2 2
b) Find a fundamental matrix for the system in part (a).

c) Find the general solution of the original system.
d) Find the solution of the original system satisfying the initial conditions
y1(0)=0, yi (0)=1, — y2(0)=0.

13. Repeat the procedure of Exercise 12 for the system

Yity2—y2+y2=0,
Vatyitye+y2=0.
In part (d) find that solution satisfying the initial conditions y,(0)=0, y;(0)=1,
Y2(0)=0, y2(0)=2.
14. Consider the matrix differential equation
Y=AY-- YB;
where A, B, and Y aren xn matrices.
a) Show that the solution satisfying the initial condition Y(0)=C, where C is a
given n Xn matrix, is given by
Y@Q=e"Ce™.
b) Show that
Z=- |CGE ahi

0
is the unique solution of the matrix equation
AX +XB=C
whenever the integral exists.
c) Show that the integral for Z in part (b) exists if all eigenvalues of both 4
and B have negative real parts.
5.7 SOLUTION OF EXAMPLE 9, SECTION 4.1: AN ELECTRIC CIRCUIT
The particular circuit in question led us to the initial-value problem (4.22)

which we write in the form
v', = —3i, +3i,(t), v, (0)=0.6
i) =30, — 302 i,(0)=1 (5.38)

v, = 6i, —6v2, o> (= 1.2.
Define
y=|i}, 4=|$0 06 -§], g@=| 00

vy 0 =3 0 31,(t)
V2 —6
Then (5.31) has the form
y=Ay+g(t), y(O)=] 1 |]. (5.39)

5.7 Solution of Example 9, Section 4.1 207
To solve this initial-value problem we first solve the homogeneous system

y’ = Ay in (5.39). By Theorem 1, Section 5.1, exp tA is a fundamental matrix
of the homogeneous system and to find it we first find the eigenvalues of A.
The eigenvalues are the roots of the characteristic polynomial,
A 2 0
det (AI — A)= Li A 2 |a2s02 i246
0 -6 A+6
=(A+ 1) (A+2) (A+3).
Hence the eigenvalues are A, = —1, 4, = —2, A,=—3.
Since the eigenvalues of A are distinct, we may proceed by the method of
Section 5.3, and we next compute an eigenvector corresponding to each
eigenvalue. Corresponding to the eigenvalue 1, = — 1, we consider the system
—1 2 0
0 -6 5
By elementary row operations we have
—1 2 0
R
Qn
0 -6 5 lee
|
(co
Oe
- MN
uln
© (ee
|
ra) |
eooor VN
OR
4h
D
yw ©
Oaey
——
Thus the vector
is a solution of (—/—A) x=0. Notice that (miraculously) v,; =y(0) (it was
intentionally given this way). Similarly,
3 =
1%
10 10
o-['] [fF
32 1
are eigenvectors corresponding to the eigenvalues 4,=—2 and 1,=—3

respectively. By Theorem 1, Section 5.3,
D(j=len'vay eaten 795]

is a fundamental matrix of the system y’ = Ay. By formula (5.14), Section 5.3,

exptd=@(t) (0).
By a familiar formula from linear algebra [see, for example, [3], Theorem 1,
Section 4.7],
2820
~ det &(0)’
and we find, after an elementary but tedious calculation, that
25. 2895 S)
6 6
fb +*(O= 4 8 | (5.40)
5 -9 5
You can also verify that (0) 6~'(0)=@~'(0) &(0)=/. Hence

Ze7t ye? he * 25 2.5 2
exptA=| e! e 7 te3] |20 gs —12 (5.41)
Sem eat area 5 -9 5
(3e7'—2e7 *+5e7 *) (—3e7'+12e7 7* ee **) ($e '—e **+4e~ 7)
= Gee '— Be * se *) (—2.5e-'+8e **—3e **) (2e~'—42e- 7" + Se 7)
(Se~'—10e7
"+. 5e7*) (—3e7'+12e° *—9e- 3") (e'—Se
7+ Se **)
(Notice that at =0 this matrix reduces to the identity matrix as it should.)
Thus, the solution @, of the homogeneous system y’ = Ay satisfying the given
initial condition at t=0 is (by matrix vector multiplication)
0.6 0.6e°'
®,(t}=exptA | 1 |=exp(tA)vi=| ex ie
12 Fee.
By the variation of constants formula applied to the system (5.39), we obtain
as the solution @ to the given initial-value problem
t
(t)=exp(tA) v, + [exp(e—s) Ag(s) ds.
In view of the special form ofg(s) we obtain from (5.40) (replacing t by :—t):
0.6 f [ Se 792-40 te 3G)
oem] 121 |etsg ||Beer aera ege-s-o]

Se7 =) — 10-24-94 Se 3-9
ide, (549)
0
Thus, if the source current i,(¢) is given, the solution (5.42) can be simplified
further and if 7,(¢) is sufficiently simple, the integrals may be evaluated explic-
Shi! Solution of Example 9, Section 4.1 209
itly (for example, if the source current is 3 sin wt) and then the desired voltages
Vy, V2 and the current 7, can be determined explicitly.
Exercises
1. Find the voltages v, (¢), v2(¢), and the current i, (¢), if i,(t)=3 sin wr.
2. Discuss the behavior as too of the voltages v, (1), v2(t), and of the current i, (1),
in the case that 7,(t)=3 sinat.
3. Repeat Exercise 2 if i,(t)=3e~™ sin wt, «>0.
We close this section with several remarks. Define the matrix

TP =O(0)=[vao¥o.¥o|,
and make the change of variable
v= 17:
Then y’=7z’, and the system (5.36) becomes

Tz =ATz+
g(t)
or
Z=T *ATz47T ‘git), zO)=1 *y(0). (5.43)
The matrix T~'AT is, of course, similar to the matrix A for any matrix T
(see Appendix 5), but with the present choice of the matrix 7 and the vector
g(t), the system (5.43) becomes
—1 0 0 1
Za Oo) Ol zr (td Oe (5.44)
0 0 =3 0
where z(0)=7~ 'y(0) and where T~' is given by (5.40). Of course, (5.44) is
readily solved (almost by inspection) because it is uncoupled. If we denote by
W(t) the solution of (5.44) satisfying the initial condition y(0)=z(0)=T™ 'y(0),
then the solution @(t) of the original problem (given by (5.42)) is @(t)= Tip(2).
Exercises
4. Find the solution w(t) of the system (5.44) satisfying the original condition
W(0)=7~ 'y(0) with y(0) given in (5.39).
5. Use the result of Exercise 4 to find the solution of the original initial value
problem. (Compare your result with (5.42); they should be identical.)
The procedure outlined here, which involves the diagonalization of the

coefficient matrix A, in general, has the following interpretation. According
to Theorem 1, Appendix 5 the coefficient matrix A is similar to a diagonal
matrix if and only if A possesses n linearly independent eigenvectors. (This
is always the case if the eigenvalues of A are distinct.) Then the motion of any
system such as (5.39) driven by any input g(t) and starting from any set of
initial conditions can be thought of as the superposition of the motions of
the uncoupled first-order systems (5.44); for if w is a solution of the un-
coupled system then d¢=7¥ is a solution of the original coupled system
(5.39).
Another remark of some physical importance is the following. You may
prefer to think in terms of a mechanical rather than an electrical system.
An electrical system such as the one in Fig. 4.2, Section 4.1, which we have
studied has an equivalent mechanical analog. This~is obtained as follows.
Dash _pol
Figure 5.13
Consider the mechanical system sliding on a frictionless table as shown in

Fig. 5.13. Two masses m, and m, are connected by a spring (obeying Hooke’s
law), having spring constant k. A force f is applied to the mass m, and the
mass m, is connected to a rigid wall by means of a dash pot which offers a
resistance to motion proportional to the velocity, with proportionality
constant p. The system moves in a straight line on a frictionless table. Let
X1, X2 denote the displacements from equilibrium of the masses m, and m,,
respectively. At equilibrium f=0 and the system is at rest. Let v;=x;, with
i=1, 2, be the velocities of the masses and let /=x,—x, be the increase in
length of the spring with respect to its unstretched length (length under no
tension). By Newton’s Second Law of Motion, we have
i) for the mass m,:
rit |
Le oe:
ii) for the mass m3:

3
UL Rp Hie eel U6 Toes
l .
k >
and, from the definition of / and p,,

l=v,—0,.
Sal Solution of Example 9, Section 4.1 211
Putting this together, we obtain
= mk
that
mm,
'=—v,+0v, (5.45)
v, = ———
ee
|-— v
- mk m, 7
If we now identify
aes 3 1 3
f =i,, Sa m,=s, M2=6> k=3,
the mechanical system (5.45) and the electrical system (5.38) already solved
become identical.
There is a procedure for establishing mechanical analogs from electrical
ones and vice versa. For our purposes it suffices to see one example, such as
the one above. For more details, the interested reader is referred to one of the
standard references such as [8].
Exercise
6. Solve the problein (4.17), (4.18) for the mechanical system posed in Example 8,
Section 4.2. [Hint: Reduce the linear system (4.17) of two second-order equations
to an equivalent linear system of four first-order linear equations and apply the
methods of this chapter.]
CHAPTER 6
Series Solutions
of Linear
6.1 INTRODUCTION
In many mathematical and physical problems, we are led to differential

equations which cannot be solved in closed form. (By a solution in closed
form we mean a solution expressed in terms of elementary functions, that
is, polynomials, rational functions, exponentials, logarithms, trigonometric
functions, etc.) For many such differential equations it is possible to prove
that the solution can be expanded in a power series, and to calculate the co-
efficients of this power series. The purpose of this chapter is to give an ex-
position of these methods of solution. While most of the chapter is devoted
to linear differential equations, we start with a very simple method which is
applicable to some nonlinear equations. It is, however, usually cumbersome
to apply unless one needs only the first few terms of the series solution.
Example 1. We begin by returning to the differential equation
y’=-siny—y (6.1)
which represents a possible model for a damped pendulum (see Section 2.2), and
which cannot be solved in terms of elementary functions. Let us try to find the
solution ¢ of (6.1) which satisfies the initial conditions
o0)=Z ¥(0)=0. (6.2)

From Theorem 1, Section 2.5, we know that Eq. (6.1) has a unique solution satis-
fying the initial conditions (6.2). Moreover, it can be shown that this solution can
be expanded in a convergent power series in powers of t about t=0.* Assuming this
fact, how can we find the coefficients of this power series? Recall first that if @ can be
* This follows from an extension of Theorem 1, Section 6.3 below, to nonlinear equa-
tions, but will not be studied here.
212
expanded in a power series about t=0, then @ must have derivatives of all orders at
t=0 and @ has the Taylor series expansion
¢"(9) , o” (0)
P(t)=
(0) +4’ (0) t+ y PL Gls aries a oc
1
where the series on the right side converges to ¢(t) in some interval |t|<r. Thus we
need only evaluate the solution and its derivatives at t=0 in order to find the expansion
(assuming that there is one).
From (6.2) we already know that the function ¢ must satisfy #(0)=7/4, ’(0)=0.
In terms of the graph of ¢, this means that we know that y=¢(t) passes through the
point (0, 2/4) with slope 0. Since ¢ is a solution of (6.1), we know that
o¢"(t)=—sind(t)—¢'(t) t>0. (6.3)
In particular, we can set t=0 in (6.3), and find $”(0)=—sin (0)— ¢’(0)= —sin (2/4)
a =P Since $”(0) is negative, this tells us that the graph of ¢ is concave
downward near t=0, and the value 256/53) would enable us to compute its curvature
there (see Fig. 6.1). How can we find the values of the higher derivatives of ¢ at the origin?
i
(0, 1/4)
Figure 6.1
Differentiating the identity (6.3), we obtain
$()=—[cos $()] ¢'()-4"() 120. (6.4)

Therefore $°)(0)= —[cos (7/4)] (0)—( —./2/2)=,/2/2. Here it should be noticed that
we need the value of $”(0) that we found just above. In exactly the same way, we can
differentiate (6.4), and obtain
o ()=[sin (1)] Ld’ (Q]? — [cos o(1)] $"()- 4 (t)
and therefore $')(0)=(,/2/2)(0)? —(./2/2) (—./2/2)—./2/2=(1 —/2)/2. Continuing
this procedure, we could determine the values of a// the derivatives of ¢ at the point
t=0. Therefore, the expansion of the solution @ begins with the terms
Oo 2 wv2t 3 be ~ et 4 i
DD) eee 2a)! 2 4!
214 Series Solutions of Linear Differential Equations
Exercises
1. Continuing with Example 1, find the numerical values of $'°(0) and $ (0).
2. Let w be the (unique) solution of (6.1) that obeys the initial conditions
w(0)=0, w’(0)=1. Find the values of the first four derivatives of y at =0 and
write down the first few terms of the power series expansion of iy about t=0, assum- _
ing that it has one.
3. Consider the first-order differential equation y’=(?+y°. If @ is the solution
satisfying the initial condition $(1)=0, find the values of ¢’(1), ¢”(1), ¢° (1),
and $(1). Also, write down the first few terms of the power series expansion of
o about t=1, assuming that it has one.
4. Consider the solution ¢ of the first-order equation y’+y=0 such that $(0)=1.
Use the above method to find the power series expansion, if there is one, of ¢.
Can you sum this series and verify that the sum represents a solution?
5. Repeat Exercise 4 for the solution ¢ of the differential equation y” + y=0 such that
$(0)=1, ¢'(0)=1.
While the above procedure can sometimes be used, it presents several

difficulties. As Example 1 above shows, each successive derivative is usually
more difficult to evaluate. Unless one can compute a formula for ¢” (0)
for every n, it is impossible to verify the convergence of the resulting series
and thereby justify the method. This also is demonstrated by Example 1.
Another difficulty that can arise is illustrated by the following example.
ene
Example 2. Consider the differential equation y’=/(t), where f is defined by
== jl 2
The function f has derivatives of all orders at the origin and f(0)=0 (n=0, 1, 2,...).
If @ is the solution of y’=/(t) such that ¢(0)=0, and if @ can be expanded in a
convergent power series about t=0, then every coefficient in this power series must be
zero, and ¢ must be the zero function. However, the zero function does not satisfy the
differential equation.
We will study systematically methods of obtaining power series solu-

tions which avoid the difficulties illustrated by these examples. We will
limit ourselves to linear equations. We propose to concentrate on second-
order equations since they are of the greatest interest for physical applica-
tions and since they exhibit most of the important phenomena.
6.2 REVIEW OF PROPERTIES OF POWER SERIES
Every power series )'”- 9c,t" has a radius of convergence R>0. (ick=0)
the facts stated below are meaningless.) The numbers c,, as well as ¢ can be
real or complex. The following properties hold:
6.2 Review of Properties of Power Series 215
i) The series )°* 9¢,t" converges absolutely for |t|<R.

li) If we let f(Q)=) °°oc,t”"for |t|<R, then f is a continuous function and
J has derivatives of all order for |t|<R. Its derivativef’ is the sum of the
power series )°” nc,t"1, and the differentiated series also has radius
of convergence R. More generally, for any k>1,
f(t)= ¥ n(n-1)---(n—k+1
¢,t"*
) for |t]<R.
n=k
ili) If a and b are in the interval of convergence, then

b oe) b
[roae) c, |sa
In particular
fr
t=)’ feue)a
which is a power series also having radius of convergence R.

a
~~
Identity theorem for power series: If °° od,t” is another power series
with radius of convergence R>0, and if )* od,t"=
>) 9c,t" on some
interval (more generally, some disc in the complex plane) in which both
series converge, then d,=c,, for all n. In particular, if the sum of a power
series is zero for all t in some interval, then every coefficient in the power
series must be zero.
Ss Convergent power series can be added a adding corresponding terms.
oo precisely, if f()=)°~ oc, t" and g(t)=)2 od,t”, then f()+g()=
2 o(c, +d,) t", wherever the power oe forfand g both converge.
Vi — Convergent power series can be multiplied by multiplying individual
terms and eos terms ee the same power of t. Specifically,
if f()=> ~~ 0¢,0" and g(t)=) ~-o4,t", then f(t) g()=) c= oPat", where
ie bee Ween +++++Cod,,. Thé series )\*op,t" converges for all values
of ¢ for which the power series for f and g both converge.
We remark that properties (i){vi) hold equally well for power series
of the form )'- 9c, (t—a)", with the region of convergence |f—a|< R.
To simplify the notation we introduce the following terminology.
Definition. A function f defined on an interval containing the point t=a is
said to be analytic at t=a iff can be expanded in a Taylor series f(t)= > -°-o¢n
(t—a)" with a positive radius of convergence.
Note that iffis expanded in such a power series, the coefficients are given
by c,=f“(a)/k!, as can be verified by using property (ii) and then setting

t=a.
Exercise
1. Determine the real values of f, if any, at which the following functions fail to be ;
analytic: sin ¢, 1/t, 1/(t—7), 1/(t? +1), (t—2)/(St+4) (t—3).
The ratio test is a useful and simple method of determining the radius of
convergence of a power series.
Ratio Test
Consider the power series )°°- o¢,(¢—a)". If lim, len +1/Cn| exists and has
the value p, then the power series has radius of convergence R= 1/p if p40,
and R= co if p=0, and the power series converges absolutely for |t—a|<R.
We remind the reader that the end points of the interval of convergence
must be investigated separately.
Another useful method of estimating the radius of convergence of the
power series )'-°_ 9c, (t—a)" is the following.
Comparison Test
If es cS eons tet C-= 0' such, thatiie |<G forn—Onl. Dy sand at,
yo " converges for |t—a|<R, then )'* oc,(t—a)" also converges
at palesa ee <R.
We will make frequent use of the following fact about analytic functions
whose power series begins with different powers of (t—a). We state the result
for two functions, but the extension to any number of functions is obvious.
Lemma 1. Let f and g be functions analytic at the point a, given by the power
series oo
fO=Y alta, g=¥ alta
respectively. Suppose p<q and c,#0, d,#0. (This ensures that the series
for f begins with a term in (t—a)? and the series for g begins with a term in
(t—a)*.) Then the functions f and g are linearly independent on every interval
I on which both their series expansions converge.
Proof. Suppose f and g are linearly dependent on J. Then there exist
constants A and B, not both zero, such that
Af (t)+ Bg(t) =A alt +B ay d,,(t —a)"=0
for every ¢ in J. Using property (v), we may add these power series term-
by-term. The resulting power series has the lowest-order term Ac, (t—a)?
6.3 Second-Order Linear Equations 217
v),
(since p<q). Using property (iv), we conclude that 4=0. But then
for every ¢ in J, which mee itl again by property (iv). Thus 4 = B=0,
which contradicts the hypothesis of linear dependence, and completes the
proof. |
6.3 SECOND-ORDER LINEAR EQUATIONS WITH

ANALYTIC COEFFICIENTS
Before presenting a general theorem, we study a specific problem.

Example 1. Suppose we wish to find the solution @ of the differential equation
y’—ty=0 (6.5)
satisfying the initial conditions $(0)=a, ¢’(0)=6. The existence and uniqueness
theorem for linear equations (Theorem 1, Section 3.1) tells us that this problem has a
unique solution ¢ which exists for all t. However, no method of solution that we have
studied will yield this solution in closed form. (We must ask the reader to accept this
fact.)
The power series method begins with the assumption that the solution ¢ is analytic
at t=0. (If the initial conditions were given at t=to, we would begin by assuming that
@ is analytic at t=f,.) This assumption means that olf) can be expanded in a power
series b(t)=cotcytt-+et*+---=-> Ct" (6.6)
which converges in some interval |t|<A, where the een A is positive and to be
determined. Our object is to determine the coefficients c, in this power series so that
¢@ satisfies the equation (6.5) and the initial conditions (0)=a, $'(0)=
If the power series (6.6) represents a solution of (6.5), we must be able to
differentiate it twice. Proceeding formally (that is, without worrying about convergence)
for the present, we obtain
f (t)=cy +2egt+--- +keth +---= Keir
eee 0 = (6.7)
f" (t)=2c,+3-2c3t
+++» +k(k—1) qth 7+--- = Y k(k—-l) et" 7.
Also a ag
th(t)=cotteyt?
+. +ctt t+--= PY oth*? (6.8)
k=0
Using (6.7) and (6.8), we obtain
+3+2c3t+ +k(k—1) et
" (t)—th(t)=[2c, 2
h++]
+c,t? Abo Gt
—[cot dake |
=2c,+3-2c3—Co) t+-::
Veeseps |2s
Lite
=2c,+ ¥) [kk-1l)q—q_3] f°’,
k=3
where we obtain the general term by adding coefficients of the same power of f.
Before proceeding, we note that it is not necessary to write out the calculations as
explicitly as we have done above. From (6.7) and (6.8) again, we have
2 ES
$" (t)—tb(t= ¥ k(k—-1) eth 7- Y gyth*?. (6.9)
k=2 k=0
To combine these series, we observe that the first one begins with a constant term while J
the second one begins with the term cot. Therefore, we separate the constant term from
the first series. We must also rewrite one of the two series in such a way that the general
terms of both series contain the same power of t. To do this with the second series, we
let kK+1=n—2, or k=n—3. Then the second series is )29¢,f"' =) 36,230"7.
We also observe that the particular letter used for the index of summation is of no
importance, and we may use k in place of n again if we wish. We rewrite (6.9) as
p"(t)—th(t)=2c2+osk(k—1) -¥ C, 30k?
(6.10)
yes Y [k(kK—1) q,—cy-3] 8?
which is the same result as that obtained before.
Assuming that the operations which led to (6.10) are justified, we see that ¢, repre-
sented by the series (6.6), is a solution of the differential equation (6.5) if and only if
the c, satisfy the relation
)) [k(k—-1) ,—G-3]
2e,+ K=3 7 =0. (6.11)
By the identity theorem for power series (property (iv), Section 6.2), the coefficient of
each power of ¢ in (6.11) must vanish. Therefore
Dor) k(k—1) c,—c,-3=0 (k=: ya (6.12)
Relations such as (6.12) are said to determine the coefficients c, recursively. Solving
these in succession, we find
6 =0 Cs
1
mance C4=>—
1 Cy O55
1 2
a _ —_ — are
56 es 6 Oe gh eo
Ce— = Co; Co = = Boas
In general, we can describe the solution by the following formulas:
1 2
Cc, =0 Clea =F Cy
1) (4 2) (5
c5=0 peo RANE
(1) (4)---Gm—2) (2) (5) (8)---@Gm—1)

Cam42=9 C3_,=———— Cy) C341= (3m +1)! Cy
Thus all coefficients can be expressed in terms of cy and c,; these in turn can be
determined from the initial conditions. If the solution ¢ of (6.5) is to satisfy the initial
conditions $(0)=a, ¢'(0)=5, we find that cy=a, c, =b, and, therefore, a candidate for
the solution is
: iG Gne eee
Had i+) v=
Gm! |
psfer YQ EMaD aes (6.13)

m=1
provided that the above procedure can be justified.
Exercises
1. Use mathematical induction to establish the formulas for 3m, C3m+1> ANd C3m 42
above.
2. Show how (6.13) is obtained from the formulas of Exercise 1.
3. Use the ratio test to prove that each infinite series in (6.13) converges for |t|<0oo.
Using the convergence of the series established in Exercise 3, we now observe that
all the formal calculations beginning with (6.6) and leading to the solution (6.13) are
fully justified for — co <t< oo by applying the properties of power series given in Section
6.2. Moreover, the solutions ¢,, ¢, represented by the power series
2) (5)---(3m—
.{)ar+ ate ) eee
) OCOD
m=1
are linearly independent by Lemma 1, Section 6.2. Therefore, if a and 6 are considered
as arbitrary constants, the general solution of Eq. (6.5) is given by ad, (t)+5¢,(t) for
—0O<t<o.
While the series in (6.13) converge for all t, only a small number of terms is needed
to give a good approximation to the solution for small values of ¢. This is of practical
importance in obtaining numerical approximations of solutions. For large values of
t (say, t= 10) the series converges too slowly to be of practical value. In this case, however,
the method of asymptotic expansions and numerical techniques are available which
make use of the values computed from the power series for small t. Some of these
techniques will be discussed in Section 6.11 and Chapter 9.
Exercises
Employing the method used in solving the equation (6.5), find the solution ¢ of each .
of the following initial-value problems.
4. y’+y=0, ¢(0)=a, $'(0)=6. (Write the solution in closed form if you recognize
the resulting series.)
5. y’—ty’'+y=0, ¢(0)=1, o/(0)=0.
6. y” —2ty’+2ny=0 (n an even integer, n=2m), (0)=(— 1)” (2m)!/m!, o'(0)=0.
7. y’—2ty’+2ny=0 (n an odd integer, n=2m+1), $(0)=0, $'(0)=2(—1)" (2m
+1)!/m!.
The solutions of Exercises 6 and 7 are called the Hermite polynomials.
Exercise
8. For which values of « does the equation y”—2ty’+ay=0 have solutions which
are polynomials in ¢? [Hint: Assume a power series solution and determine a
condition on « which causes the series to terminate. ]
The above examples suggest that second-order linear equations with

analytic coefficients can be solved by power series. The reader should observe
that in equation (6.5) as well as in all the equations considered in the
exercises the coefficients are polynomials in f, hence analytic at every point,
in particular at t=0 where the initial conditions are imposed. In fact,
the following theorem, which will be proved in Section 6.4, is true.
Theorem 1. Consider the differential equation
y"+p(t) ¥+a() y=f (0.

If p, q, and f are analytic at to, then there is a unique solution @ satisfying
the initial conditions $(to)=a, $'(to)=b. This solution is analytic at t=to
and its expansion
[o.e)
P()= kK=0
DL ex(t—to)®
converges for at least those values of t for which the power series expansions
of p,q, and fin powers of (t—to) converge. The coefficients c, may be determined
recursively by direct substitution.
We remark that Theorem 1, as well as all other considerations in this

chapter, holds without change for the case in which the independent variable
t is complex; the proofs of the results are more meaningful in the complex
domain.
You are warned that differential equations must be put in the form given
above (with leading coefficient 1) before the theorem can be applied.
Example 2. (Legendre equation) Consider the equation (1—1*) y’—2ty’+a(a+1) y
=0, where « is a given constant. Determine whether this differential equation has a
series solution about t=0.
To apply the theorem, we must divide the equation by 1 —?7?, and rewrite it as
2t a(a
+ 1)
lie po eee
which is certainly justified if +# +1. Then
t=0 pi=-—52t a=
a(x
+1)
f)=0
and p and q are analytic at t=0; in fact they can be expanded in power series valid for
|t|<1 as follows.
hs a
p(t)= ——5= —2¢(1+0?+t*+---)=—2 }) t?*
1-t k=0
a(a+1) oa ae Se hoy
LG pera
2 =a(a+1)(14+07+t*4+---)=a(a+1) ¥ 17.
we k=0
Theorem | tells us that the Legendre equation has a unique analytic solution ¢
satisfying any pair of initial conditions #(0)=a, ¢’(0)=6, and the power series ex-
pansion of ¢ converges at least for |t|<1.
Exercises
9. Compute the first 5 terms of the series expansion of ¢ if the initial conditions are
~(0)=1, ¢’(0)=0. Can you guess the general term? Show that if a is an even
nonnegative integer, <=2m, then ¢ is a polynomial of degree 2m. Compute this
polynomial for m=0, 1, 2, 3.
10. Repeat Exercise 9 for initial conditions ¢(0)=0, ¢’(0)=1. Show that if « is an
odd positive integer «= 2m+ |, then ¢ is a polynomial of degree (2m+ 1). Compute
this polynomial for m=0, 1, 2.
Note that although Theorem | tells us that the series expansion of ¢

converges at least for |t|< 1, it may actually converge ona larger interval. This
is illustrated in Exercises 6-10 where the solution turns out to be a
polynomial.
Exercises 9 and 10 above show that for each nonnegative integer n, the
Legendre equation (1—17) y’—2ty’+n(n+1)y=0 has a solution which
is a polynomial of degree n. If this polynomial is multiplied by a suit-
able constant to make its value for t=1 equal to 1, the resulting solution
(why is it a solution?) is called the Legendre polynomial of degree n,

written P,(t).
Exercises
11. Compute the first 5 Legendre polynomials.
13. Compute [1,[Po(é)]? dt, f=, [Pi (0)? at, [£1 [P2(O]? @t.
It can be shown that in general |*,P,,(t) P,(t)dt=0 if m#n, and
1 [P,.(0]? dt=2/(2n+1) (n=0, 1, 2,...). The Legendre polynomials and
their properties play an important role in many physical problems having
spherical symmetry, including problems of potential theory and heat transfer.
Exercises
14. Consider the Legendre equation with « an even integer, «=2m, namely,
(1-2?) y’—2ty’+2m(2m+1) y=0. In Exercise 9 above, the solution ¢@ with
$(0)=1, '(0)=0 was found to be a polynomial of degree 2m. Find a second
linearly independent solution valid in a neighborhood of t=0.
15. Show that the second solution found in Exercise 14 in the form of a power series
converges for |t|<1 but diverges for ‘= +1.
16. Consider the Legendre equation with « an odd integer, ~=2m+1, namely,
(1 —27) y’ —2ty’+(2m+ 1) (2m+2) y=0. In Exercise 10 above, the solution ¢ with
(0)=0, $'(0)=1 was found to be a polynomial of degree 2m+1. Find a second
linearly independent solution valid in a neighborhood of t=0.
17. Show that the second solution found in Exercise 16 in the form of a power series
converges for |t|<1 but diverges for t= +1.
These exercises show that the Legendre polynomials are the only bounded
solutions of the Legendre equation on —1<f<1 when «@ is an integer.
In fact, if «is not an integer, it is easily shown that no solution of the Legendre
equation is bounded on —1<tr<1.
In applying Theorem | to a specific equation, we divide through by
the coefficient of y” before applying the theorem. However, to solve the
equation it is usually easier to return to the original form to calculate the
series solution. For example, to find a series solution of (1 + ¢*) y’+y=0, we
substitute (zt) =) 2 oc,¢* into this equation to give
(L+07) ¥ k(k-1) qt"? + Y et*=0,

k=2 k=0
carry out the multiplication, and proceed as usual.
Exercises
18. Apply Theorem | to each of the following differential equations and initial
6.4 Proof of Theorem on Solutions in Power Series 223
conditions when applicable, but do not solve the equation. In each case give the
interval of convergence of the series solution @ guaranteed by Theorem 1.
a) (+1?) y"+y=0 P()=1 $'(0)=2
b) (1-0?) y’+y=0 P(1)=1 '(1)=0
c) (sint) y’+(cost) yy +y=0 6(3)=« #(5)=6
d) (sint) y’+(cost)y’+y=0 o(x)=a_~ ¢'(n)=b

~~ 19. Solve the initial-value problem given in Exercise 18(a).
20. Solve the initial-value problem given in 18(c).
We remark that for the equation of order n
y+ (t)y++ Pn s(t) Y' + Pal)
V=S(0)
where the functions p,,..., p, and f are analytic at t=¢p, there is an exact
analog to Theorem | which asserts the existence of a unique solution ¢
analytic at t=f, satisfying prescribed initial conditions.
Exercises
21. State the analog of Theorem | for the equation of order n.

22. Find the solution ¢ of the differential equation y°)+ ty’—y=0 such that 4(0)=0,
'(0)=1, $”(0)=0.
6.4 PROOF OF THEOREM ON SOLUTIONS IN POWER SERIES
In this section we shall prove Theorem 1, Section 6.3, in the special case
J (t)=0. This is the case which commonly arises; we shall make some remarks
on the nonhomogeneous problem after the completion of the proof. The
proof is constructive, in the sense that it shows us how to proceed to find
a power series solution for any given second-order linear equation with
analytic coefficients, and is based on the useful method of majorants.
Thus we consider the differential equation
y"+p(t) y'+4(t) y=0 (6.14)
and we seek the solution ¢ satisfying the initial conditions
We assume that p and q are analytic at fy. This means that we can write
p=5ralt—to® al)Yaxlt—to}
with both series converging in some interval |t—f9|<A, where A4>0. The
object is to see whether the solution ¢ of (6.14), (6.15), whose existence is
guaranteed by Theorem 1, Section 3.1, is analytic at ft); that is, whether
it is possible to write $()=)7 oc,(t—to)*, with the series converging in

some interval |t—t)|<B, where B>0.
Assuming for the moment that the solution ¢ is analytic at t=), we
can write, for |t—to|<B,
00
=P ke(t—to)' o" (= Y k(kK—-1)

Ge(t-to)*?.
k=1 k=2
Substituting the series for p, g, ¢, ¢’, and ¢” into Eq. (6.14), we obtain
go"(t)+ ea
$'(t)+4(t) P(t)
Se k—1) ¢(t—to)k~7 + bnD,(t—to)*- 2h ke,(t—to)k~?
(oe) foe)
+ g(t—to)*: 2a; c(t —to)*=0.

k=0 k=
Shifting indices in the first two terms, we obtain
O°()+P(0) 6'()+4(0) B= Y (k+2) (+1) exealt—to}

foe)
fo Py(t—to)X: 9 (kK+1) Cea i(t—to)

iT
M48
8
a y Ox (t—to)*: » c(t —to)*=0.
k=0 k=0
Carrying out the multiplication and addition of these power series (proper-
ties (v) and (vi), Section 6.2 and grouping terms in the same power of (t— fo),
we obtain
i [(k+2) (K+1) Cera +(K+1) Docess + kp ice to + PKC

+ ole + 41Cn—1 + 7+ +44C0] (t—to)*=0. (6.16)
We remark that if the power series for p and q converge for |t—to|<A
and if the power series for @ converges for |t—t)|<B, then the power
series in (6.16) actually converges for |t—to9|<C where C=min (A, B)>0.
By the identity theorem for analytic functions, the coefficient of every power
of (t—fo) in (6.16) must be zero; this gives the following recursive system
of equations for the coefficients c,.
2° 1¢2 + Poli + Golo =9
3-2¢3
+ 2poC2 + P10; +Go€1 +41Co =O0
4-364
+ 3poc3 +2p1C2 + P2C1 + J0C2 +4161 +9209 =0
(6.17)
(k+2) (k+ 1) Ch+2 a 2s [(m+ 1) Pxr—m©m+1 Gye Cl =0
We can solve these in succession. The first equation gives c, as a linear

combination of cy and c,. The second equation gives c3 as a linear com-
bination of cg, c,, and cj; since c, has already been expressed as a linear
combination of cg and c,, this means that c; is a linear combination of co
and c;, and so on. By induction, the equation expressing the fact that the
coefficient of (¢—‘)* is zero gives c,, as a linear combination of co, ¢1,...,
Cx+1, With the coefficients in this linear combination depending on the
given coefficients Po, P1,---5 Ps Jos V1>+++> J. Our method of successive
solution shows then that c,, 1s actually a linear combination of cp and c,.
We can write ¢,42=Gy42Co t+e+ 2€, (kK=0,-1, 2,...), where dy45, &+> can
be calculated successively from (6.17) and depend on po, P1,---s Ps Yoo 19+
g,. Thus we have found the coefficients in the power series for ¢, and we
can write
co
(t)=Cco
+c, (t—to)+ Dat(di.Co
+ &,C1) (t—to)*
eo] 1+ 3 ltt} |+e (t++ Yedee (6.18)
Now cy and c, can be determined from the initial conditions. In fact,

Co=(to)=a4, Cy='(to)=b. Thus we have shown that if the solution
¢ is analytic at tp, then it must have the form (6.18), where the coefficients
are computed successively with the aid of the system of equations (6.17).
The reader should notice that it is not at all obvious that the series (6.18)
which represents the solution @ converges, or what its radius of conver-
gence is. To complete the proof of Theorem 1, Section 6.3, we will show that
the series (6.18) has a positive radius of convergence B, and in fact that
B>A. Once this has been done, all the operations which led from the
differential equation (6.14) to the formal solution (6.18) are readily justified
by using properties of power series just as in the examples in the
previous section. Incidentally, once we have established the convergence
of the solution (6.18), we should note that we can write $(t)=cod,(‘)+
C(t), where $,()\=1+)224,(t—to), b2(t)=(t—-to) + Dice 2€u(t— Lo).
Since the series for ¢, and ¢, begin with different powers of (t— fo), the func-
tions ¢, and @, are linearly independent by Lemma |, Section 6.2, and
we have expressed the solution ¢ of (6.14) as a linear combination of the
linearly independent solutions ¢, and ¢,. Thus ¢(t) in (6.18) is the general
solution of (6.14). We shall show that the series ) i 9¢,(¢—{o)* converges for
|t—to|<.A by constructing a majorant series which converges for |t— to) <A.
To establish the convergence of the series solution (6.18), we shall
need to make use of the following fact about power series, known as
Cauchy’s inequality: If 1? 94,(t—to)* converges in |t—to|< R,then for every
r, 0<r<R, there is a constant M>0 such that |a,| r’<M (k=0, 1, 2,...).
This follows immediately from the fact that if ))% oar converges, then
lim,.... |a,“*|=0, and therefore |a,|r* is bounded; that is, there exists a
constant M>0, independent of k, such that |a,| “<M (k=0, 1, 2,...). (If
t is complex, this argument must be modified somewhat.)
The coefficients c, are determined by the equations (6.17) which may
be written as
k
(kK+2)(k+1) G+2=— 2 [ee Dp meme
ee eee i (6.19)
Since the series ) 2op,(t—to), ogj(t—to converge in |t—tol<A,
corresponding to every Lee number r<A there is a constant M>0
such that
pir<M |gjrP=oM j=0,1,2..2: (6.20)
Using (6.20) in (6.19), we obtain
M k
(k+2)(k+1)lene alse DL [(m+1)lems

sl+emi]
k
=<
[(m+ 1) ee il+|cml] r"+M lex +41 i (6.21)
Se 2
where the positive term M |c,,,|r added on the right side of (6.21) only
increases this side of the inequality and is needed in what follows. Now
we define Co =|co|, C; =|c,|, and for k=0, 1, 2,..., we define C, recursively
by
M
(k+2)(k+1) Cita =a ys[iat Cait Call + MC aa. (6.22)
=0
Comparison of(6.22) with (6.21) shows that 0< |c,|<C, (k=0, 1, 2,...).
Exercise
1. Use induction to show that 0<|c,|<C,.
Replacing k by (k—1) in (6.22), we obtain
M
(k+1) C41 =Saq DE[in el) Ge on b Colira mi Graas (6.23)
Using (6.22),
k
M
7 [(m+1) Ca+1+C,,] 7°=(k+2) (k+1) C,42.—MrC,,,.
m=0
On the other hand, from (6.23),

k k= 1
M 1 M
a > [(m+1) Casi tC] Nise er a iO Oe a i ae
m=0 m=0
+M(k+1) Cy4,+MC,
1
el KCy44 -MC,+M(k+1) Cy+1+MC,.
Combining these, we obtain

(k+1)k
(42) 41) yea] M(k+1)+Mr |C
and therefore
Cero (kt k+M(k+ 1)r+Mr? _k+M 1 Mr

Cae r(k +2) (k+1) — k+2 r (k+1)(k+2)
Since lim,..,, C,+2/C,+1;=1/r, the ratio test shows that the majorant series
Vireo C,(t—to)* converges in |t—t9|<r. This implies by the comparison
test, that )9 ¢(¢—to)* also converges in |t—to|<r. Since this is true for
every r <A, the series )'7-9¢,(t—to)* converges in |t—to|<A, and the proof
is complete. |
In the nonhomogeneous case,
y"+p(t)
y +a(t) y=f()
let
N= Y Alero} ‘
It is easy to see that the coefficients c, in the series solution )'?.9 c;(to)*
are now determined by the equations
2-1¢.+ Pol: +4olCo=fo
3-2¢3+2poC2+ Pier +9001 + U1C0=/t
k
(k +2) (k+ 1) Cy42t 2 [(m+ 1) Py—mm+1 + Gx—mCm] = Si
in place of Eq. (6.17).

Exercises
2. Derive the above set of equations for the coefficients c, in the nonhomogeneous
case. [Hint: Give a development which parallels (6.14){6.17) in this case.]
3. Show that the c, in Exercise 2 can be determined recursively from the above system.
It is possible to establish the convergence of the formal series solution -

obtained in this manner either by an extension of the methods for the homo-
geneous case, or by use of the variation of constants formula (3.34), Section
3b
6.5 SINGULAR POINTS OF LINEAR DIFFERENTIAL EQUATIONS
Many differential equations which arise in applications, the so-called equa-

tions of mathematical physics, fail to satisfy the hypotheses of Theorem 1,
Section 6.3. For such equations the power series method developed in Sec-
tions 6.3 and 6.4 can be modified. We propose to study these modifications
systematically for a broad class of such equations. As has been our
custom, we shall concentrate on equations of the second order. To see what
we may expect, we consider first the very special second-order equation,
called the Euler equation,
(t—to)’y" +(t—to) ay’ taay=0, (6.24)
where a, and a, are constants. The Euler equation serves as a simple but
useful prototype of the general situation which will be explained in detail
below. We notice immediately that Theorem 1, Section 6.3, cannot be
applied to the equation (6.24) in a neighborhood of the point t= fp) because
the functions p(t)=a,/(t—to), q(t)=a>/(t—to)? are not analytic at t=%p.
However, in a neighborhood of any point f,#?to, Theorem 1, Section 6.3,
is applicable.
Exercise
1. Deduce the existence of a unique analytic solution ¢ of (6.24) which satisfies the
initial conditions $(t,;)=Yo, ’(t,)=Zo, Where t; # to. [Hint: Show that Theorem 1,
Section 6.3, can be applied; note that
1 1 1 tr,
kalo else listo ti >to
1 U— Ty
= ic space |,
ty —lo t)—lo
and similarly for 1/(t—f9)?.]

We now take advantage of the special form of the Euler equation, and
we solve it explicitly by the following trick. Make the change of independent
variable |t—t)|=exp s. Then for t>t) we have
6.5 Singular Points of Linear Differential Equations 229
w(s)=y(to+exp s)=y/(t),
and by the chain rule (using s=log(t—tfp))
dy _dwds_ 1 dw
dt as dt (i=, ds
d*y d*w (=) dw d*s
di? ds* \at) * ds de
d*w a vdw: dale
~ ds? (=) Pals |-(=-)
Pe \2 ew dw
-(=-] (Ga-%).
Thus Eq. (6.24) becomes
d?w dw
ie, a (a, — 1) Fe +a,w=0 (6.25)
and this equation has constant coefficients. We leave it as an easy exercise

for the reader to show that if t<¢) the change of variable |t—t)|=exp s
also transforms (6.24) to (6.25). Since Eq. (6.25) has constant coefficients, its
general solution is given by
w(s)=c, exp(z,s)+c, exp(z,5)
where c, and c, are arbitrary constants and where z,, z, are roots of the
characteristic equation
z?+(a,—1)z+a,=0 cee
if z, #Z, and it is given by
w(s)=(c; +c5) exp zs
if z, is a double root of the characteristic equation. Thus, using the change
of variable |t—t,)!=exp s, again we obtain the following explicit solution of
the Euler equation (6.24)
P(th=cy |t—tol?t +c |t—tol?? or = (t)= (Cy +2 log |t— tol) |t— tol”!
according as z,, Z, are distinct or equal roots of the characteristic equation
z?+(z,—1)z+z,=0. Incidentally, we remind the reader that if z is complex
|t|? means exp(z log |t|); thus if z=a+
if, «, B real, then
|t|’ =exp [(« +iB) log |t|] =|t|*(cos B log |t|+i sin B log t).
Exercise
2. Find the general solution of each of the following differential equations valid near
f=0)
a) t?7y"+2ty’—6y=0
==bry Fy +4y—0
c) 2ry"+ty’—y=0
d) ?y’—(2+i)ty’+3iy=0
er me "—ty’+y=0
We notice also that although Theorem 1, Section 6.3 cannot be applied

to Eq. (6.24) to deduce the existence of a solution analytic at f= fo, it may hap-
pen that a solution analytic at t=¢,) does exist (for example, if z, or z2 or
both are distinct nonnegative integers, either the solution (t—f)*! or (t—o)”?
or both are analytic at t=fo).
Exercise
3. Make up an example of a differential equation of the form (6.24) (with ft) =0 if you
wish) which has at least one solution analytic at fg, and another example for which
every solution is analytic at fo.
For the Euler equation, we see that the behavior of solutions when / is
near fy and when f¢ is near ¢, #/g is quite different. Every solution is analytic
at t; #t,) but not necessarily at fy. This suggests that the point f=f) plays a
special role here. The point t=¢,) will be called a singular point of the
equation
do(t) y" +a, (t) y’+a2(t) y=0 (6.26)

where do, 41, a2 are given functions analytic at fo, if dg(to)=0 but a, and a,
are not both zero at fo. (If ao(to) =a; (to) =a2(to)=0, then we can divide
the equation (6.26) by a suitable power of (t—f) so that at least one of the
coefficients is different from zero at to.) More generally, t=f, is a singular
point of the equation
do(t) y +a,(t) vy" +--+» +4,_1(t) y +4,(t)y=0 (6.27)
if dp, @;,..., @, are analytic at t) and ap(tp)=0 but aj, a2,..., a, are not all
ZELO at hay
We shall see that the behavior of solutions of Eq. (4.26) or (6.27) near
the singular point tg depends very much on how rapidly a(t) approaches
zero as {fy. For this reason, we distinguish two different types of singular
points and we make the following classification.
Definition. The point to is called a regular singular point of the equation
(6.26) if it is a singular point and if
6.5 Singular Points of Linear Differential Equations 231
have the property that (t—to) p(t) and (t—to)* q(t) are both analytic at to.
More generally, the point to is called a regular singular point of the equation
(6.27) if it is a singular point and if
Px(t) a
~ do(t)
a do(t)’ eeAo (t)
i=) N as
——
=) = Cas
have the property that (t—to) p,(t), (t—to)? p2(t),..., (t—to)" Pa(t) are all
analytic at to.
Examples
1, The Euler equation (6.24) is perhaps the simplest example of an equation which
has a regular singular point at t=f9 because ty is a singular point and
P(t)=4,/(t= to), q(t) = 42/(t— to)” have the property that (¢— to) p(t)=41,(t—to)” (1)
=a, are analytic everywhere, in particular at fo.
2. The equation fy” +(3) ty’+ty=0 has t=0 as a regular singular point because
p(t)=3/2t, q(t)=1/t have the property that tp(t)=3, t?q(t)=t are both analytic
everywhere, in particular at t=0.
3. The equation (t—1)? y’+2(t—1)? y'—7ty=0 does not have a regular singular
point at r=1 because p(t)=2/(t—1), g(t)=—7t/(t—1)° do not have the property
that (t—1) p(t)=2, (t—1)? g(t)= —7t/(t—1) are both analytic at t=1.
Ifa singular point t=, for Eq. (6.26) or (6.27) (such as t= 1 in Example 3
above) is not regular, then we say that f, is an irregular singular point. On
the other hand, points t=¢, which are not singular points of an equation
(6.26) or (6.27) with coefficients analytic at tg are called ordinary points (or
regular points). Thus in Examples 1, 2, 3 above the points f=?o, t=0, t=1
respectively are singular points, and all other finite values of ¢ are ordinary
points. The points t=¢,) in Example | and t=0 in Example 2 are regular sin-
gular points, while the point t= 1 in Example 3 is an irregular singular point.
Theorem 1, Section 6.3, describes completely the behavior of solutions in a
neighborhood of an ordinary point; our next task is to discuss the behavior
of solutions in a neighborhood of a singular point. As we shall see, it is
considerably easier to do this for regular singular points than for irregular
singular points, and we shall begin with the study of the former. Naturally it
is essential that, given a differential equation of the form (6.26) or (6.27), we
first locate and classify its singular points.
Exercises
Locate and classify all the singular points for finite values of t of each of the following
4. t?y"+y'=0
5, (1—t?) y"—2ty’+a(a+1)y=0 (Legendre equation)
Dy nm Series Solutions of Linear Differential Equations
_ Cy" +ty'+(x?—27) y=0 (Bessel equation)

Oo
1
. t(1—2?) y"+2(t—-1) yy —Sty=0
8. y"+5r sinty +7tv=0
A
9. y+ y+2y=0
1+t
10. r(1—t) y°+[e—(a+b+1)t] y—aby=0 (hypergeometric equation, with a, b, c
constants)
11. ryP+ _ y’ —(sint) yp=0

‘ (2+1t)-~ semis
12. (1-27 By 4 5e5(1+2) y"—277(1-7) vy+y=0
6.6 SOLUTIONS ABOUT A REGULAR SINGULAR POINT: EXAMPLES
It follows immediately from the definition given in Section 6.5 that the
equation
do(t) y” +a, (t) y¥ +a,(t) y=0 (6.26)
has a regular singular point at t=/, if and only if (6.26) can be written in
the form
(t—to)? vy +(t—to) x(t) ¥+B(t) y=0 (6.28)
where x(t)=(t—19) a; (t)/ao(t) and B(t)=(t—to)? a, (t)/ao(t) are analytic at
fo, With at least one of the three numbers x(t), B (to), 6’ (to) different from zero.
(If all three of these numbers are zero, then (6.28) has (t— fg)? as a factor and
f=fp is only apparently a singular point. If we divide (6.28) by (t—19)?, the
resulting equation will have an ordinary point at t=/,.) Notice that the Euler
equation (6.24), is of the form (6.28) with «(7) and f(r) constant functions.
Exercise
1, Show that the equation (6.27), Section 6.5, has a regular singular point at t=fp if
and only if the equation can be written in the form
(t—to fy +(t—to) tay (t) YP" +--+ H(t) & 1 (1) +4,(0)y=0 (6.29)
with a, &,..., &, analytic at fp.
In order to simplify the discussion, it is convenient to make a preliminary

transformation. The change of independent variable t=x+ 1, enables us to
transfer the singular point f) in (6.28) (or (6.29)) to the origin without
changing the form of the equation in any essential way. Namely, if 1540,
we let &(x)=x(f9
+x), B(x)=B(to+x). Then % and B are analytic at x=0
since x and f are analytic at f=). We also let }(x)=y(t
9+). The chain rule
gives
6.6 Solutions about a Regular Singular Point 233
cy eee did3
Te) totX) «= a= V"(to +x)
and therefore (6.28) becomes
Cy tie ede wl ny.
= a2 + XE) xt BO) y=0. (6.30)
This is of the same form as (6.28), but with x=0 a regular singular point.
Conversely, if }(x) is a solution of (6.30), the function y(t)=$(t—to) is a
solution of (6.28). Thus (6.28) with a regular singular point at t=t) #0 and
(6.30) with a regular singular point at x=0 are equivalent. The same
transformation t=x+f may, of course, also be applied to (6.29). We will
therefore assume that such a preliminary simplification has already been
made, and we will consider the equation
t?y" +ta(t) y+ f(t) y=0 (6.31)

where « and f are given functions analytic at t=0 and having power series
expansions n Rs
a()= Loti B= Bat!
which converge in some interval |t}|<r(r>0), and such that the numbers
%9, Bo, B, are not all zero. Thus r=0 is a regular singular point for (6.31).
Every statement about the equation (6.31) can, of course, readily be changed
to apply to (6.28) by means of the above change of variable.
Exercises
Make the appropriate change of independent variable in each of the following

equations to transfer each singular point to the origin, and find the transformed
equation with singular point at the origin.
2. (1-2) y+ 2ty’+3y=0
3. (1-2?) y’—2ty’+a(a+1) y=0 (Legendre equation)
4. t(1—2) y’+[c—(a+b+1) ¢] y'—aby=0 (hypergeometric equation)
Before proceeding with a general theory, let us consider three specific

equations.
Example 1. The equation 2ty’+y’+ty=0 has a regular singular point at +=0. Find
an expression for the general solution valid in an excluded neighborhood of t=0.*
You should recall that it suffices to find two linearly independent solutions
(Theorem 5, Section 3.3). This equation may be written as
* By an excluded neighborhood of t=0, we mean a set given by the inequality 0<|t|<a

for some a>0.
which is of the form (6.31) with «(t)=4, B(t)=377. Since « and f are both analytic at
1=0, =0 is a regular singular point. If « and B were both constants, then (6.32) would
be an Euler equation and would have at least one solution of the form |r|’. To take
into account the fact el in this example f is not a constant, we try to find a solution of
the form |t\? 32 oct“ (co#0), where the constants z, c, are determined by sub-
stitution into the differential equation (6.32), and where the series )% 9¢,t* converges .
on some interval about t=0. Note that we cannot deduce the existence or uniqueness
of any solution with initial conditions prescribed at r=0 from Theorem 1, Section 3.1
(why not?), let alone a solution of the above form. Since t=0 is a singular point for
(6.32), we separate the cases t>0 and t<0. We consider first the case >0 and try asa
solution the function $(t)=f V2 o¢t*=V 2 oct? **, which for 1>0 has
-> eletk Et? b= Y oleh) (e+k—1) e?.

We postpone any justification of the differentiation for the moment Substitution into
(6.32) gives, for those >0 for which these series converge absolutely (see Section 6.3,
for a similar argument):
th" (t) + 5to' (t) +37 (t)

- y c,(2 +k) (z+k—1) #**+4 ¥ alz+k) t7tk+4 osON si
+=) c, ((z +k) (gz+k—1)4+3(2+k)] 77*+4 fy Cans
= ¢€oz(z—}) +e,(2+1) (2+3Jette [(z+k) (z+k—3) q+3e-2] 7.
Writing f(z)=z(z—4), we can write this last relation (for r>0) as
17h” (+o +4PO=] cof

e)teut et 1) +> {f (z+k) q%+4¢,—2} |
(6.33)
Therefore the equation (6.32) can be satisfied by the function ¢(1)= VP oc,t* for t>0
only if the coefficient of every power of ¢ in the right side of (6.33) vanishes. Since we
assumed cg #0, we must therefore have
f(2)=0, ecf(ztl1)=0, f(ztk)qtiq_,=0 k=2,3,4....

The function f(z) is a polynomial, called the indicial polynomial. For a second-order
differential equation with a regular singular point at t=0 this polynomial is always
quadratic, as we shall see in Section 6.7. In the present example, f(z)=z (—4) and thus
the relation f(z)=0 is satisfied if z=} or if z=0. Taking first the case z=4, we must next
choose c, =0 to satisfy the relation c, f(z +1)=0, since /(3)40. Then we must calculate
the coefficients c, (k>2) from the relations f(¢+k) c,+3¢,-2=0 (k=2, 3, 4,...). Since
* You are reminded about shifting indices in series (Section 6.3, Example 1).
fG+k)=(k +4) k #0 for k=2, 3, 4,..., we can write these relations as

—1
C= Ch k=2, 3, 4,....
E Dk(eee) 2
Now we can express all the coefficients in terms of cg, which remains arbitrary. We obtain
Gi OF (aU roa Coy sh 0) one
Exercise
5. Verify the formula for c,,, by induction.
Substituting these quantities into the assumed form of the solution @(t)
=? P~ ¢,t*, we obtain as one candidate for a solution of the equation (6.32) for
t>0 the function
1/2 . (=1)" 2m
(=e! rE) Did asenn (On) 5195-0 inten): |
We have chosen the arbitrary constant cy as 1, since it is merely a factor multiplying the
whole series. Similarly, taking the root z=0 of the indicial equation f(z)=0 we find that
Co is arbitrary, c,=0, and f(k) c,+5c¢,-2=0 for k=2, 3,.... Since f(k)=k(k—3)40
for k=2, 3,..., we can write these relations as
1
a ee
and
we find
ee ae
Co
Com-1 = 0" Com=(—

ye 2-4 =
eee SS (2m):3-7---+-(4m—1)
Again taking cy =1 we obtain as a second candidate for a solution of (6.32) for t>0 the
function
(—1)” 2m
b.()=14 )Acer ve On\aaoe Been
The approach now is to prove that the candidates ¢, and ¢, are in fact solutions of
(6.32) in some interval 0<t<a and that ¢, and @, are linearly independent on this
interval. Before doing this we observe that, assuming the convergence of the relevant
236 Series Solutions of Linear Differential Equations —
series in some interval, the above calculations are all valid for r<0 if f is replaced by
|r|F =e eel,
Exercise
6. Verify this last statement by calculating @, (1) for 7<0.
We thus have, for those +0 for which the series converge absolutely,
1/2 (—1)" 2m
$1 (=Iel ES) 2-4-5: Qamy:S-9->->> (4m+1)- |
(6.34)
(—1)" 2m
b.()=1+) Fy heer (2m)-3-7-- >> (4m+1)-
m=1
as candidates for solutions of (6.32). Next, we apply the ratio test to the series for @, (¢).
We let
m4 :
Um(t)= (2m)-—
2:4--+-- 5:9< <= (4m + 1) :
Then
Us. 1 (t) = t? = 0) m—> oO

U,,(t) | (2m+2) (4m-+5) ;
and thus by the ratio test the series
=f tm
EAL = iy
2:4---:- (2m):5-9+++-- (4m + 1)
0
converges absolutely for —00<t<oo.
Exercise
7. Show that the series for @3(¢) also converges for —co<1<oo.
Now, because of the relevant properties of power series (Section 6.2), it is clear that
all of the calculations which lead from the assumption of the solution of the form
It? 0 cyt* to the two candidates for solutions ¢,, @ given by (6.34) are fully
justified for — oo <r<0 and for 0<r<oo. The value r=0 must be omitted because the
differential equation (6.32) has no meaning at the singular point r=0.
Exercise
8. Show that the solutions ¢, and @, given by (6.34) are linearly independent on
—0o<t<0 and 0<t<oo, and hence on any interval not containing t=0.
[Hint: Modify the argument of Lemma 1, Section 6.2, to fit the present situation.|
Using the result of Exercise 8, we see that the general solution of(6.32) on any inter-
val which does not contain the origin is a, (t)+a, (t), where a, and a, are arbitrary
constants.
Exercise
9. Use the technique of Example | to find the general solution of the equation
ty” +3y'+y=0, and determine the interval of validity of this solution.
In Example 1, the assumption that the given equation has a solution of

the form |t\? }°79 ¢,t* leads to a quadratic equation in z, called the indicial
equation. Each root of the indicial equation leads to a solution of the
differential equation, and the two solutions obtained are linearly independent.
As the next example shows, the indicial equation may have equal roots,
making the search for two linearly independent solutions considerably more
difficult.
Example 2. Find the general solution of the differential equation ty”+ y’ + y=0, valid
in an excluded neighborhood of t=0. The given equation can be written as
t?y"+ty+ty=0. (6.35)
Clearly t=O is a regular singular point and (6.35) is of the form (6.31) with
a(t)=1, Saget Assuming again the existence of a solution of the form ¢(t)
=|t\’ YZ ocxt* (co#0) on some interval (excluding the origin), we consider the case
t>0. Proceeding as in Example 1, completely formally at first, we have
(k+z)(k+z—1) c,t**? + eK(k-+2) ctht?+ 2 cythtet!

(t)+th’(t)+ h(t) => (
t*p"
=27Cot7 + 7 eteee |e
k=1
=f? |e) Cot ys {f (kK+z) G.+C,-1} |
where f(z)=z*. Thus, since (6.35) can have a solution of the assumed form only if the
coefficient of every power of ¢ in this expression vanishes, we must have f(z)=0 and
f(k+2e+¢-:=0 k=1,2....
The indicial polynomial in this example is f(z)=z*, quadratic as in Example 1, but the
indicial equation has a double root z=0. Since /(k+z)40 for k=1, 2,..., we can solve
for the coefficients c, from the recursive equations c,= —c,_,/k? (k=1, 2,...).
Exercises
10. Determine the coefficients c, and complete the derivation of the solution of the
assumed form as in Example 1.
11. Determine the interval of validity of this solution. (Notice that we can see
directly from the recursion formulas, without finding the coefficients, that
ley 440°* !/e,t*|=t/(k
+1)? 0 as k00.)
Note that although the differential equation (6.35) makes no sense at the
singular point r=0, the function defined by the series })~9 c,t* in this case
is well defined at the singular point r=0. This remark is important in ”
applications (see especially Section 6.9).
We see in this example that, because the indicial equation has a double
root, there is only one solution of the form |t|? ))@o c,t*. Since the differential
equation (6.35) is of the second order, it must have two linearly independent
solutions, though not necessarily both of the same form. We postpone to
Section 6.8 the finding of a second, linearly independent solution.
From Examples | and 2, we might suspect that whenever the indicial
equation corresponding to a regular singular point at the origin has distinct
roots, the differential equation has two linearly independent solutions of
the form |t? )o ¢,t*, one corresponding to each root of the indicial
equation. However, the following example shows that this is not always
the case.
Example 3. The equation ty” +ty’—y=0, which may be written as
t?y"+t?y'—ty=0 (6.36)
has t=0 as a regular singular point. Find the general solution, valid in an excluded
neighborhood of t=0.
In seeking a solution of the form $(¢) = |t|? V2 o¢xt* (co #0), we first restrict ourselves
to the case t>0, and we find exactly as before
Pp" ()+07¢'(t)—th(t)
+) (k+2z) cy pete+1 ay c,tkte+!
=<) (k+z)(k+z—1) an
k=0 k=0
ee
=) (k+z)(k+z—-1) otk? Peale.
ech k=0
2(z—1) cot? + on z—1) G+(k+2—2) c_1] pete
marBe (f(k+z) + (k+z—2) c-1} «|

where f(z)=z(z—1). As in Examples | and 2, ¢(t) can be a solution of (6.36) only if z
is a root of the indicial equation f(z)=0, that is, only if z=0 or z=1, and
f(k+z)q,4+(k+z—2)cq,-,=0 k=1,2,.... (6.37)
Taking z=1, we have f(k+1)=k(k+1)40 (k=1, 2,...), and so c,=—(k—-1) c_4/
k(k+1) (k=1, 2,...). This gives c,=0 (k=1, 2,...), and, taking cy=1, we obtain the
solution @, (t)=|t| (t#0).* There is obviously no difficulty about convergence.
* We could have guessed this solution from a careful inspection of (6.36).

To see if there is a second solution of the assumed form, we consider the root
z=0 of the indicial equation. Now the recursion formulas (6.37) become
k(k—-l)q,+(k—2)cq,-,=0 k=1,2....
Taking k=1, we see that c, must be determined from the relation 0-c,—cy=0.
Since cyo#0, this is impossible, and there can be no solution of the assumed form
corresponding to the root z=0. As in Example 2, a second, linearly independent,
solution of a different form may be found by a method to be studied in Section 6.8.
Exercises
12. Find a formula for a second, linearly independent, solution of the equation (6.36)
by the method of Section 3.6.
Using the methods of the examples studied in this section, find as many linearly
independent solutions of the form |¢|? )%c,t* (co#0) as possible for each of the
following differential equations. Also find the interval of validity of each such
solution.
13. 27?y’+4t(1+2) y'+2v=0
=> 14. Py" + 4ty'+2(1+22) y=0
15. ?y’+2(1+2) »’-y=0
= 16. Py +(0? —35 )y=0
17. Py’ +ty’=y=0
6.7 SOLUTIONS ABOUT A REGULAR SINGULAR POINT: THEOREM
The examples of Section 6.6 suggest that for any linear second-order differ-
ential equation with a regular singular point at the origin there is at least
one solution of the form ||? u(t), where u(t) is analytic at t=0. Indeed, let
the equation be
L(y)=t?y" + ta(t) y'+B(t) y=0 (6.31)
where «(t)=) 720 %E*, B(t)= 2-0 Byt* for |t|<r and not all the numbers
9, Bo, By are zero (see Section 6.6), We consider first the case ¢>0, and
we will show by the same formal procedure used in the examples of
Section 6.6 that (6.31) has at least one solution of the form $(t)= )'79ct"
(co #0), which might be called a generalized power series, whose coefficients c,
may be computed recursively. It is also true (see Theorem | below) that this
series expansion is a valid representation of a solution on the “punctured”
interval 0<|t|<r (that is, the interval —r<t<r with the center r=0
removed). The proof of this last statement parallels the proof of the
corresponding statement in Theorem 1, Section 6.3, for an ordinary point,
and is carried out below.
Assuming the existence of a solution of the desired form on some punc-
tured interval 0<|t|<A(A <r), we have, for 0<1<A,

(j= > (ktz)qt* * Oo()= > (k-2) (k+2z—1) c,t?**~?
k=0 k=0
and therefore, using the rule for multiplying power series (property (vi),
Section 6.2),
wee: pu (2 ¢ rs) = Dy(3,pss) yer

(k+z)c
Jott)
re"()= > (k-+2) (k-+-2—1) ct?

Therefore ¢(t) can be a solution of 0<?t<A only if on this interval
Lo ()=# S| d+a)h+e—Net YU+2)em-+¥, fies] =0
ge2(z—1)+a9z+ Bo) cot 5,ar (k+z—-1) c¢,
asia» ((i+z) %—j+ Bj) “|=o.

But this relation holds on the interval 0<t<A if and only if the coefficient
of every power of t is zero. Writing this condition for k=0, 1,..., we find
[(z—1)+a92+
Bo] co=0 ae
[(k+z) (k+z—1)+09(k+z)+Bo] at 2d (j+.2) 0% jC;
ck)
an s C;B,-j;=9 k=1 Ds pith:
jJ=0
Using the hypothesis cy #0, and letting z(z—1)+49z+fB)=f(z), the indicial

polynomial, we may write this set of equations as
f(z)=0, f(k+z) at > (G2. pth le =O kh— te

(6.38)
Notice that the sum in (6.38) depends on z, on the functions «(¢) and A(t),
and on the coefficients Co, ¢,,...,¢,-1, but not on c,. Now we proceed
exactly as in the examples. The indicial equation f(z)=0 determines two
roots z, and z,, and we must see whether the recursion formulas (6.38)
can be solved for the c, when z takes the values z, or z,. But the system
(6.38) corresponding to the root z,; can obviously be solved uniquely for
c, im terms of Co, Cy,..., Cj—1 (kK=1, 2,...) if f(z, +k)
#0 for every positive
integer k. Similarly, the system (6.38) corresponding to the root z, can be
solved uniquely for c, in terms of Co, ¢1,..., Cy-1 (kK=1, 2,...) iff(z, eae
for every positive integer k. The resulting functions $, (=f! Vo c,t* and
f2(t=t? Veo qt, with coefficients determined by this procedure are
candidates for solutions of (6.31) on the interval 0<t<A. We shall call such
candidates for solutions formal solutions of (6.31). However, as we have seen
in Examples 2 and 3 of Section 6.6, we cannot always find formal solutions
for both the indices z,, z,.
We now label the roots z,; and z, so that 2z,>z,, and we will show
that there is always a formal solution of (6.31) of the desired form correspond-
ing to the root z, with larger real part. Indeed, since z, and z, are the roots
of the indicial equation, the indicial polynomial f(z) can be written
I (z)=(z—2Z,) (z—2Z2). Thus
Sf(2, +k)=K(k+z,—z,)40 k=1,2....
since both factors k and (k+z,—Z,) have positive real parts. Therefore
the recursive formulas (6.38) give c, uniquely in terms of Co, C4,..., Cy—1
(co#0) for k=1, 2,... when z=z,, and can be solved recursively for c, in
terms of cy (k=1, 2,...). This procedure gives a formal solution ¢, (¢) cor-
responding to the root z, in some interval with t>0. Exactly as in the ex-
amples of Section 6.6, we verify that the above calculations are valid for
t<0 if f7' is replaced by |t|?'.
We now examine the problem of finding a second, linearly independent,
solution corresponding to the index z,. (Obviously, if z; =z , there is only
one solution of the desired form.) By the argument used to find the formal
solution corresponding to the index z,, we need only check whether
f(Z2+k)#0 for every positive integer k. But f(z, +k)=k(k —(z, —2)), and
it is clear that f(z, +k)=0 for k=k if and only if z, —z, =k. Thus f(z, +k) #40
for every positive integer k if and only if the difference between the two
roots of the indicial equation is not a positive integer. [fz, —2Z is not a positive
integer, we may solve the ne (6.38) corresponding to z=Z», and we obtain
a second formal solution $(t)=|t\? V0 Cxt* of (6.31).
If z; —Z> is a positive integer m, then the following situation can occur.
The recursive formulas (6.38) corresponding to z=z, can certainly be
solved for C1, C2,-+-, Cn—1 because f(z, +k)=k(k—m)+0 for k=1, 2,...,
m—1. Clearly, f(z,+m)=0 but if it should happen that )"""o' [(7+ 2) %m—j+
+B,—-j] ¢;=0, then the equation (6.38) with z=z,, k=m becomes 0°c,,=0,
and is satisfied by an arbitrary constant c,,. We can continue the successive
calculation Of Cm+i5 Cm+2.--. because f(z,+k)=k(k—m)#40 for k=m+1,
m+2,....Wewill refer.to this situation again in Section 6.8.
We may now summarize what our findings up to this point suggest.

t?y"+ta(t) y+ B(t) y=0 (6.31)
where «(t) and B(t) are analytic at t=0 and have expansions «(t)= Y= o %xt*,
th= So Byt* which converge for |t|<r for some r>0. Let z, and z be the
roots of the indicial equation
f(2)=2(z—1)+0%92+
Bo=0 ==
with Rz,>Rz>. Then there is a solution of the form
(j= |e?! DY Gt Co=1
in the punctured interval 0<|t\|<r, whose coefficients c, can be determined

recursively from the equations
fat+hoq= =i) [Gi+z) oy —j+By_j]o) k=1,2..... (6.39)

If z, —Z, is not zero or a positive integer, there is a second, linearly independent
solution of the form
also in the punctured interval 0<|t|<r. The coefficients ¢, are also determined
recursively from the equations (6.39), with z; replaced by z and c, replaced by ¢,.
It should be stressed that it is simpler in practice to substitute the
assumed form of the solution into the differential equation than to use the
recursive formulas (6.39) to solve for the coefficients.
We have not yet completed the proof of Theorem |. It remains to be
shown that the series for ¢,(t) and #,(t) converge for 0<|t|<r. Once this
has been done, it follows from the properties of power series that all the
calculations which lead from the assumption of the form of the solution to
the expressions for ¢,(t) and #3(¢) are justified.
To complete the proof of Theorem 1, let us prove the convergence of
the series for #,(¢) for |t|<r. The proof parallels the convergence proof
in Theorem 1, Section 6.3. Since f(z; +k)=k(k+(z,—z.)), it is easy to see
that
LF (21+ A> K(k—| 21-29). (6.40)
Since the series }\9 a;t/, 29 Bt? convergence for |t|<r, by Cauchy’s
inequality corresponding to every positive number p <r there is a constant
M>0 such that
la) p'<M. |Blpi'<M j=0,1,2.... (6.41)

Using (6.41) and (6.40) in (6.39), we obtain
lice!
K(k—|21—zal) lalsM ) G+lzi1+1) led Ka 12h ee (O.A2)
‘6
Let N be the integer such that N—1<|z,—z,|<N, define Cy=|col,
Cay ancenc Cy—1=|cy— |, and then define C, recursively for k>N by
peel
k(k—|z,—22|) Ch=M >) (i+ lz,141)
pi *C; k=N,N+1,.... (6.43)
j=0
Comparison of (6.43) with (6.42) shows by induction (see Exercise 1,
Section 6.4) that 0<|c,|<C, (k=0, 1, 2,...). Replacing k by (k — 1) in (6.43),
we obtain
Kae
(k=1)(kK-1-[z1~2Zal) Cx-s=M Y Gtlel+1)
pi**1C;. (6.44)
=
Combining (6.43) with (6.44), we see that, since
kak. |
ae P)
pM 2d G+lzil+1) p“Cj=M DY G+lzs|+1)

pi 4 4C,4+M(k+ 24) C.-:
= 1=9
we have
pk(k—|z; —Z2|) C,=(k—1) (k-—1—|z, —22|) C,-. +M(k+|z2,|) C,_1.-
ce _(k=1) (kK=1 2, = 221) + M(k +211)

Cy-1 pk(k—|z, —Zal)
and lim,.,, C,/C,-1=1/p. Thus, the ratio test shows that the series
Veo C,t* converges for |t|}<p. This implies, by the comparison test, that
Vo Ct* also converges for |t|<p. Since this is true for every p<r, the
series ) 72 9 ¢,t* converges for |t|<r. Exactly the same argument can be used
to prove the convergence of the series for @,(¢), and the proof of Theorem 1
is now complete. The reader should observe that the convergence proof is
similar to the convergence proof in Theorem 1, Section 6.3 (and in fact
contains the earlier convergence proof as a special case). |
Example 1. Consider the equation fy’+3ty’+ty=0. Use Theorem 1 to discuss
the nature of solutions valid in a punctured neighborhood of t=0.
Here t=0 is a regular singular point, with «(¢)=3, B(f)=¢; these functions are
obviously analytic at t=0, and their power series expansions, being the functions
themselves, converge for |t|<0o. The indicial equation is z(z—1)+(3) z=z?+z/2=0,
and thus z, =0, z,= —}4. Since z, —zZ, is not a positive integer, Theorem | tells us that
this differential equation has two linearly independent solutions ¢,(t)=1+)
2 1¢,t*
and #3(¢)=|t|1/7(1+ 02 14,¢") valid for 0<|t|<0o. Observe also that in spite of the
fact that the differential equation is undefined at the singular point t=0 one of the
solutions, namely @,, is analytic at -=0.
Exercises
1. Write out the statement of Theorem | for the case of a regular singular point at
t=t. [Hint: Recall the discussion at the beginning of Section 6.6].
. Prove that if z;—z, is not a positive integer, the solutions d, and @, are linearly
independent.
. Use Theorem | to determine the number of solutions of the form |t\? V2oct"
(co =1) for each of the following, without solving the differential equation. Also, use
Theorem | to determine the region of validity of each solution, and whether each
solution is analytic at the singular point. Note that in parts (c), (e), (f) the nature of
the solution may depend on the values of constants, and discuss the various
possibilities.
a) ?y"+tv'+(t?—-4) yp=0
ay b) 3¢7y"+Sty’+ 3ty=0
c) ty’+(1—1) y’'+yy=0 — (y constant)
d) ?y"+t' +(?—4) y=0
e) Py" +ty'+(t??—y?) y=0 (Bessel equation)
f) r(l—t) y"+[e—(a+b4+1) t] y'—aby=0 (hypergeometric equation)
g) ?y’+ty'+(1—1) y=0
-_ h) t?y”+te'y’+y=0
4. For each equation in Exercise 3, find the solutions of the form |t? V2 o¢t*
—— (co =1). For the equation in Exercise 3(c) show that this solution is a polynomial of
degree n, a constant multiple of which is called the Laguerre polynomial of degree
n, if the constant y is the non-negative integer n.
. Use the result of Exercise | to determine the number of solutions of the form
[t+ 17 PRoce(t+1) (co=1), and of the form |t—1|? V2oe,(t—1)* (co =1), of the
Legendre equation (1 — 7?) y” —2ty’+a(«+1) y=0. Do not calculate the coefficients;
see Exercise 6 below.
. a) Find a solution of the equation (1—77) »’—2t)’+a(~+1) y=0 of the form
|t— 17 VP oc (t— Ik (Co = 1). [Hint: The algebra is easier if you do not multiply
by (1—2)/(1+12) to put the equation in the form (6.28); since t=1+(t—1) it is
easy to expand all the coefficients in powers of (¢—1) and then substitute the
assumed form of the solution into the equation.|
b) Show that for certain values of « there exists a polynomial solution.
6.8 SOLUTIONS ABOUT REGULAR SINGULAR POINT:

EXCEPTIONAL CASES
For the differential equation

L(y)=t?y" +ta(t) ¥+B(t) y=0 (6.31)
having a regular singular point at =0, Theorem 1, Section 6.7 guarantees
the existence of at least one solution of the form ¢,(t)=|¢/! Vo ct!
(co=1), valid in a punctured neighborhood of t=0, where z, is that root

of the indicial equation which has the larger real part. We have also seen
that if the second root z, of the indicial equation is either equal to z, or
differs from z, by a positive integer, then equation (6.31) may fail to have a
second, linearly independent, solution of the same form as ¢,. Our object
is to discover the form of a second, linearly independent, solution in these
exceptional cases. Our method is based on the device employed in Section 3.4
to find a second solution of a linear differential equation with constant co-
efficients when its characteristic equation has equal roots. We will discuss
both types of exceptional cases, and will summarize our findings in
Theorem | below. Then we will make a few remarks about the nth-order
equation.
CASE |. (z;=Z ) We shall work on the interval 0<¢<r in a formal (non-
rigorous) way. In order to proceed, we must first refer to our previous
calculations at the beginning of Section 6.7 preceding equation (6.38).
There we found that if d()=/ Vo ct" for 0<t<r where co #0, then
LE W)=F {ele 1)+492+Bo) cot Ge (k+2-1) 0,

+ ao (k +2) atBoat Sy,
((j+2) %—j;+B,—,;) «||
=f? Jost, [f(z+k) o,.+9,(z)] “| (6.45)
where pas)
f(z)=2(z-M+a2+Bo =9(Z)= », [G+z) %%-j+ Bj] ¢;.
We now proceed a little differently than in Section 6.7. Since f(z) is a

quadratic polynomial, numbers z such that f(z+k)40 for k=1,2.,...
certainly exist (there are only two zeros of f(z)). For any such z we may
therefore define c, recursively by the relations
Coz
co 70 f (2+h) p= 9, (2) ksh Jee: (6.46)
where Cy is any constant. In other words, we define c,(z) as a solution of
the recursive system (6.46), but for an unspecified z. The equation (6.45)
then becomes
LiP(, 2)\\=cor £2) 6560: (6.47)
Here we have written ®(r, z) in place of ¢(t) to emphasize the dependence

of the function $(t)=f },=0 ¢¢* on both ¢ and z. It is obvious, incidentally,
that the function ®(f, z,) is a solution of (6.31) where z, is the double root
of the indicial equation (z) =0; this is the solution which we previously called
$,(t). However, (6.47) gives more information than this.
Exercise
1. Show by induction that c,(z), defined by (6.46), is a rational function of z (that is,
a quotient of two polynomials in z).
Since z, =z, is a double root of f(z)=0, we have f(z,)=f'(z,)=0 (see

Appendix 2). If we differentiate (6.47) with respect to z we obtain é
é
rs L[®(t, z)]=coLf' (2)+ f(z) logz] t? cg #0
and therefore, inverting the operations 0/dz and L and the setting z=z,
(just as in Section 3.4), we have
£1 0(, Jenn=H] lt) |=0.

This shows that (@®/0z) (t, z;) is also a solution of (6.31); we will denote this
solution by ¢,(t). From the definition of ®(t, z) we see that
rs) oo ea) 00
b= 5] on C,(z) dh.
at logit) ¢(2,) +0 ), eZ) t
k=0 Z=Z1 k=0 k=0
=, (t) logt+¢”! ¥ a
where ¢, (f)is the solution already found in Section 6.7. Since co (z) is constant,
Co(z1)=0, and the series 7-9 c,(z,) “ actually begins with a term in 1.
Thus we may write
$2(t)=¢,(t) logt+t! 2, ¢ (6.48)
For ¢ negative, we must replace f*' by |t|?! and log ¢ by log |¢].
Exercise
2. Show that c,(z,) exists for k=1, 2,.... [Hint: Use the result of Exercise 1 and the
fact that f(z,+k)#0 for k=1, 2,...]
The result of Exercise 2 shows that the series (6.48) is well defined.
Our work suggests that in the case z;=z,, (6.31) has a second, linearly
independent, solution of the form
bx(D=Mi TYby+
4,(0)logle (6.49)
with the coefficients 6, determined by substitution of (6.49) into (6.31).
In an actual problem (see Example | below) we first find the solution @, as in
Theorem I, Section 6.7, and then if z)=z, we assume a second solution of
the form (6.49), substitute it into the differential equation, and solve for the
coefficients b,.
To justify (6.49) as a second solution, valid for 0<|t|<r (where the
power series expansions of a(t), B(t) in (6.31) converge for |t|<r), we

substitute (6.49) into (6.31) and we find that the coefficients b, can be
determined recursively in terms of the already known coefficients c,
(A=0, 1, 2,...). The operations required in this substitution are, of course,
justified on any interval 0 <|t|<B on which both series in (6.49) converge; as
before, the point s=0 must be omitted. It can then be shown, as in the
proof of Theorem 1, Section 6.7, that the series in (6.49) converge in
0 <|t|<r. However, we shall omit the proof.
Exercise
3. Prove that the solutions ¢,(¢) and #,(t) are linearly independent in 0<|t|<r.
Example 1. Find two linearly independent solutions ofthe equation fy" + ty’ + ty=0
valid near t=0.
As we saw in Example 2 and Exercise 10, Section 6.6, the indicial equation is z? =0,
which has z=0 as a double root, and one solution is
valid for 0<|t|<oo. We now try to find a second solution of the form
p2(t)=It| rs,b,t*
+ 4 (t) log|¢| (6.50)
k=
suggested by (6.49). Working on the interval t>0, we find
bl)=Y (k+1)
byt+$40)logr+—
oi(0)
2 1
O3()= Y (k+1) kb! +$4 (0) loge += $4 (0-36,(0.
k=0
Substituting into the differential equation and using the fact that @, is a solution, which
causes the coefficient of log t to vanish, we find
¥. [(k+1) kbp +(k+ 1) by t+by] 8°? +2t$5 (=.

bot+k=1
Substituting the series for ¢{(t), combining into a single series, and equating the
coefficient of each power of ¢ to zero, we find b)=2 and
(—1)**!2(k+1) k=1,2,....
(k+1)°b, = = bea
17272
2(ke 1)?
This is a recursive relation from which the coefficients b, can be determined, though
not without some difficulty. According to the remarks following the equation (6.49),
and also Theorem | below, the solution @, is also valid for 0<|t|<oo.
Exercise
4. Compute at least four nonzero terms of the solution ¢, in Example 1.

CASE 2. (z;—Z, =m (positive integer)) From Theorem 1, Section 6.7 we al-

ready know that there exists a solution ¢, of the form , (t)=|t\"! Vo Cxt*
(co=1), where Zz, > Az. As in Case 1, we proceed without regard for rigor
in order to determine the form of a second solution. We refer to the remarks
already made about the case z; —z, =m immediately preceding the statement
of Theorem 1, Section 6.7. From these it follows that we can use the.
previously employed recursion formulas
f(k+2) ¢.= —9,(2) (6.46)

with z=z,, to determine the numbers c,(Z>),..., Cm—1(Z2) just as before,
but since f(z, +m)=f(z,)=0 we cannot define c,,(z2). As we mentioned in
Section 6.7, if in a particular problem it happens that g,,(z2) is also zero
(that is, if g,,(z) has a factor (z—z,)), then we could define c,,(z2) arbitrarily.
We have no difficulty in proceeding to find Cy,+1(Z2), Cm+2(Z2),--. from (6.46),
and we will obtain a second, linearly independent, solution ¢,(t)
=|t\?? Vo c;(Z2) &. However, if g,,(z2)#0 (and this is the usual case in
practice) the above method does not work.
To proceed with the case g,,(z2)#0, we observe that because of its
definition in (6.45), g,,(z) is a linear homogeneous function of cg, ¢1,.--, Cy—1-
This remark enables us to introduce (z—z,) as a factor of g,(z) by a proper
choice of co. Namely, we define co(z)=z—Z2, and c,(z), c2(z),.-., Cn—1(Z)
by (6.46). Then co(z), ¢;(Z),.--, Cm—1(Z) all have a factor (z—z,), and co(z2)
=C,(Z2)=--- =Cp—1(Z2)=0, but (6.46) can be solved for c,,(z) for all z. This
means that if we define
WiEZ=C cz) eC Bcolz)=2—25 (6.51)
k=0
for t>0, and if we substitute (6.51) into the differential equation (6.31),
then we find, as in Case | of this section,
L[w(t, 2)]=co(z) £(2) F=(2—-22) (2) &. (6.52)

This implies that y(t, z2) is a formal solution of (6.31). However, if we
examine w(t, z2), we find from (6.51) and c¢o(z2)=---=Cm—1(Z2)=0 that
W(t, Z2)=0? VP m Cy(Z2) “= 2?*™ u(t)=#" u(t), where u(2) is analytic at the
origin and u(0)#0. In fact, it follows easily that w(t, z2) is a multiple of the
solution @, (¢) already known, and we have apparently accomplished nothing.
Exercise
5. Show that —(¢, z.)=a¢, (¢) for some constant a.

However, if we differentiate (6.52) with respect to z, we find
=6) Ele i=1| HO Jrf+e-2) [FF +e / (lose

so that L[(éw/dz)(t, z,)]=0 for each t>0. Thus, at least formally, we

have a solution ¢ 2(t)=(0w/0z) (t, z2) of (6.31). From (6.51),
ow 7 io) fe)
a (t, 22)=t? Yc, (z2) th+t logt Y ¢,(z2) t Co(z)=z—2Z,
Z k=0 k=0
and using the information that co (z2)=c, (Z2)=-+:=Cm—1(Z2)=0, and that
¢,(Z2) is a finite constant, we have
7)
br(=Z ( (EZ) oC (22) +07 logt a G,(Z5) it.
From Exercise 5 and the calculations preceding it, we see finally that for
t>0, our second formal solution has the form $,(t)=f? Vo cj (z2) “+
ag, (t) logt, where a is a constant and ¢, is the solution corresponding to
the index z,.
These findings suggest (see also Theorem | below) that (6.31) has, in the
case Z, —Z,=m>0 an integer, a second, linearly independent, solution of
the form
b2(0)=[t Y)butt +aq (logit (6.53)
valid for 0<t<r, where a is a constant (possibly zero), and where @, is the
solution corresponding to the index z, given by Theorem 1, Section 6.7.
It is the form (6.53) which we use for actual calculations in practice and we
determine the constants a, b, by direct substitution into the given differential
equation. The procedure justifying (6.53) as a second solution valid for
0<|t|<r would be the same as in Case | of this section. If the constant a
turns out to be zero, then the solution (6.53) reduces to the special case
mentioned earlier in which c,, may be chosen arbitrarily. To illustrate the
idea we consider a rather special problem.
Example 2. The equation /?y’+?*y’—ty=0, discussed in Example 3, Section 6.6,
has z,; =1, z,=0 as roots of the indicial equation corresponding to the regular singular
point t=0, and the solution corresponding to the index z,=1 is ¢, (t)=|t| (t40). Since
Z,—2Z,=1, the expression (6.53) suggests that we should assume a second solution
$2 (t) of the form
=> bt*+a |t\ loge. (6.54)
k=0
Taking t>0, we find
$5(t)=atalogt+ > kb,tk~? s()=—+ Y k(k—1) byt*?.

k=1 k=1
Thus the requirement that
175 (t) +t?) (t)—t2(t)=at+ ) k(k—1) byt*+at? +at? log t

k=0
+ ¥ kbyt*t! —at? logt— }° b,t*** =0

k=1 k=0
implies :
by] A+ =0
at tat? —bot+ ¥ [(k+1) kby41+(k—1)
k=1
a—b)=0
ete al Ore
k(k+1) by41= —(k-1) by $=, 3haose
We may choose by =1, and then a=1, b,=—4, b;=1(2°3), ... We note that 5, is left —
undetermined; this is because ¢ is a solution of the differential equation and therefore
any multiple of ¢ is a solution. In particular, we may take b, =0. Substitution of these
results into (6.54) gives the second solution for r>0.
Exercises
6. Show that formally we obtain the same solution for t<0 from (6.54).
7. For the differential equation in Example 2, obtain a general formula for b, and
show that the solution @, is valid for 0< |t|< oo.
8. In Example 2 find the solution ¢, by the method of Section 3.6.
We summarize our findings as follows:

t?y"
+ ta(t) y+ B(t) y=0 (6.31)
where « and f are analytic at t=0 and have power series expansions valid
for |t|<r, with some r>0. Let z,, 2, (2z, =z») be the roots of the indicial
equation
f (z)=z(z—1)+«(0) z+ B(0)=0.
Then if Zz;=2Z, there are two linearly independent solutions ,, @2 of the form
bi ()=le7 o ct* Co=l
f2(t)=|t/*? na b,t*
+ ,(t) log|e|
valid for 0<|t|<r, whose coefficients c,, b, may be determined by direct sub-
stitution in the equation (6.31).
If z; —Z» is a positive integer m, there are two linearly independent solu-
tions &,, b> of the form
$1 (t)=(t/? ce Gt “co
b2(t)=|t|? on byt*+ad,(t) log|t| by #0
valid for 0<\|t|<r, where ais a constant (possibly zero) and the coefficients
Cy, 5, may be determined recursively by direct substitution into the equation
(6.31).
Exercises
2 Prove the linear independence of ¢, and #, in both cases covered by Theorem 1.

10. Write out the statement of Theorem | for a regular singular point at some point
t— fo
Me Obtain two linearly independent solutions valid near t=0 for each of the fol-
lowing differential equations.
a) 4ty”+2y’+y=0 b) ty’+2y'+y=0
c) Py"+r7y’—2y=0 d) ?y"+Sty’+(3—27) y=0
12. Consider the Bessel equation
t?y"+ty’ +(t?7-«?)
y=0
PRN?
where «>0 is a constant.

a) Find the roots of the indicial equation.
b) Without computing the solutions, discuss (using Theorems 1, Section 6.7 and 1)
the forms of solutions valid near the regular singular point t=0. Be sure to
consider all possibilities for different values of «.
13: Find the general solution of the hypergeometric equation t(1 —1) y’+[c—(a+b+
+1) t] y’—aby=0, where a, b, c¢ are positive constants, valid near the regular
singular point ¢=0. [Hint: Solve the indical equation and then consider five cases —
(i) c not zero or an integer, (ii) c=1, (iii) c=0, (iv) c a positive integer >2, (v) ca
negative integer. The answer to part (i) is ©
ab a(a+1)b(b+1)_,
TAOS Sie eer EAI
a(a+1)---(a+k—1) b(b+1)---(b+k-1)
igi
kie(c+1)---(c+k-1)
= Fila, be th {W<))
$2 (t)=|t| “F(a—c4+1,b-—c4+1,2-c,t) 0<|t\<1.
(For further details concerning the hypergeometric equation see [12] or [17].)]
. Repeat Exercise 13 relative to the singular point =1.
. For each of the following differential equations, determine the roots of the indicial
equation at t=0, the form of the general solution, and the region of validity of the
general solution as given by the appropriate theorem.
a) 2ty”’+y’—y=0
—
sb) ty’ +3y'—-t’y=0
—>c) (317 +1") y’—ty+y=0
d) #(1—1) y"—2y/ +2y=0
e) ty’+(1—t)y+qy=0 (q constant)
f) ty’+(1—t) yy+my=0 (ma positive integer)
g) ty" +2ty’ +ty=0
i)) (7 +07) y’-(t7 +1) y+ y=0

j) ty ty” —ty—y=0
k) t?y’+ty'+(4—1t) y=0
l) (t? —t3) y’—3ty +5y=0
m ) t y’ +ty’'+(t? —4) y=0
Patel
16. Find the general solution valid in a neighborhood of t=0 of each of the dif-
ferential equations in Exercise 2.
17. Find the general solution valid in some neighborhood of the indicated singular
point of each of the following differential equations, and give the interval on which
it is valid.
a) (1-2?) y’—2ty’+a(«+ 1) y=0 (Legendre equation); t= — 1.
b) c(1—2) y’+[c—(a+5+1) t] y’—aby=0 (hypergeometric equation); t=1.
18. a) Find the general solution of the confluent hypergeometric equation
ty” +(c—t) y—ay=0
valid near t=0, assuming that c is not an integer.
b) Define M(a, c; t) to be that solution of the equation in part (a) which is analytic
at ¢=0 and had the value | at t=0. Show that the general solution found in part
(a) is
(a, c; t)+c,|t\*
$(t)=c,M “M(1+a—c, 2—c; t)
if c is not an integer.
c) Obtain the general solution of the equation on part (a) when a=1, c= 1.
d) Obtain the general solution of the equation in part (a) when a=1, c=0.
e) Obtain the form of the general solution of the equation in part (a) valid for
large tf.
We conclude this section with a few remarks about the case that t=Oisa
regular point of a general nth-order equation. According to the definition in
Section 6.5, we may write such an equation in the form
L(y) = t"y +0" *04(t) yO +--+ to,-1() ¥+0,(t) y=0 (6.55)
where «,, %2,..., %, are analytic at r=0. It is apparent that the methods of
Section 6.7 and this section are applicable to (6.55). Naturally, the situation
can now be considerably more complicated. We again assume a solution of
the form #(t)=|t\? Yo ext (co = 1). Formal substitution into (6.55) leads to
L,(@)=# |wo Se ice Gs a

where f,,(z) is the indicial polynomial,
6.9 Bessel Equation; Some Properties of Bessel Functions 253
Sn(2) = 2(2—1)---(—z—n+1)+a,
(0) z(z—1)---(z-—n +2) +--+» +.0,(0)
of degree n, and g,(z) is a linear homogeneous expression in C9, Cy,.--5 Cy—1
as in the second-order case. Obviously ¢ is a formal solution of (6.55) if the
coefficients c, are determined recursively from the relation f,(z+k) c,=g;,(z)
(which can certainly be done if f,(z+k)#40 for k=1, 2,...), and if z is a root
of the indicial equation /,(z)=0. There are n roots of the indicial equation;
if they are all distinct and no two of them differ by an integer, there will be n
linearly independent formal solutions of the assumed form. This leads to an
extension of Theorem 1, Section 6.7 to the case of the equation (6.55). The
devices used to obtain Theorem | for the exceptional case of equal roots and
roots differing by a positive integer in the second-order case can also be
adapted to this more general case. The interested reader is referred to more
advanced books such as [7], pp. 132-135, or [18]. To examine one very
trivial case of (6.55) when n=1, it is suggested that the reader find a series
solution valid near t=0 of the equation ty’+«(t) y=0, where «a is analytic
at t=0 and has a series expansion valid for |t]}<r, with some r>0.
6.9 THE BESSEL EQUATION AND SOME PROPERTIES

OF BESSEL FUNCTIONS
As an application of Theorems 1, Section 6.7 and 1, Section 6.8, we discuss

one of the important equations of mathematical physics. The Bessel equation
arises in a natural way in many problems having axial (cylindrical) symmetry,
and may be written in the form
L(y)=t?y’ +ty +(t? —p’) y=0 (6.56)
where p is a constant, possibly complex, with 2p=0. The point r=0 is a
regular singular point, and in the notation of Theorems 1, Section 6.7 and
1, Section 6.8, «(¢)=1 and B()= —p? +¢’. Both these functions are analytic at
t=0 and their power series expansions converge for |t|<0o. The indicial
polynomial is f(z)=z?—p’, and its zeros are z,; =p, z,= —p. According to
Theorem 1, Section 6.7, if p40 and if z; —z,=2p is not a positive integer
(that is, if p is not zero, an integer, or half an integer), there exist two
linearly independent solutions ¢,, $2 of (6.56), valid for 0<|t|<0o, of the
form
1 (t)=Ie/? ee ct" co #0 (6.57)
ba(t)=ldl? S ett lo #0 (6.58)
where the coefficients c, and ¢, are determined recursively by substitution

in (6.56).
We first compute the solution ¢, and as usual we assume ¢f>0. Then we

find
On= Spek te Oe) oS ¢,(p +k) (p+ k—1) t?**-?

k=0
so that
L[¢,())J=0? |r) cotf (p+1) ott {f (p+k) e+ Cy—2} |=0
from which we conclude c,=0, f(p+k) cq,+¢-2=0 (k=2, 3,...). Since

f(pt+k)=(p+k)
—p? =k (2p +k) 40, we have
Cp = Ca)
Co Co
"pli = 22ers
ce han
Com >=
2?"m!(p +1) (p+2)---(p+m)’
Exercise
1. Establish the formula for c,,, by induction.
We may therefore write the solution as
_ = (- ieee
$1 (t)=colt| 1-) 2?"m!(p+ 1) eoypeat

oe)
To define the Bessel functions in the usual way, we must make a particular
choice of co, and for this purpose we need to define the gamma function,
which generalizes the notion of the factorial. This function, denoted by I,
is given by the relation
ii Ga |Ca hi ax (6.60)
and elementary tests for improper integrals show that this function is well
defined and continuous for 2z>0. We observe that ['(1)= | e-* dx=1,
and I'(g)=|¢ e7* x71? dx=2[2 exp (— y?) dy=\/n* Integration by parts
gives i recursion formula I’(z)=(z—1) hee 1) for 2z> 1, since
A A
F(z)= lim ieee dx= lim {Loe tee) me ax}

A> oo A~> oo
0 0
=(z -|¢ xt =o tdn= (2 1) (gi):
In particular, if z is a positive integer n, we see by repeated application

of this formula that ['(n)=(n—1)! We also remark that it is obvious from
the defining relation (6.60) that it does not define I (z) for @z <0. However, for
a positive integer k, ['(z+k) is well defined for 2z>—k. The recursion
formula tells us that [(z+k)=(z+k-—1)(z+k—2)...(z+1) zI'(z). Thus we
may define ['(z) for —k<#z<0 (z#0, —1, —2,..., —k) by
['(z+k)
NMGay
z(z+1)---(z+k—1)
Doing this for k=1, 2,... we may define ['(z) for all complex z except
z=0, —1, —2,.... With the aid of (6.60) for z>0 and using this definition
for z<0, the reader will easily see that for real z the graph of I’(z) is as given
m Fig. 6.2.
We now define cy in the solution ¢, given by (6.59) as
1
C=:
2°T (p+1)
(6.61)
The resulting function, denoted by J, and called the Bessel function of the
first kind of index p is given by
J,(t)=
t SOE Dene
2 eri) , Bed
* To evaluate the integral =|? exp(—y7) dy one uses the following trick.
r-| exp(—y”) ay]=|exp(—y’) dy |exp(—x*) dx

,oe m/2 oO
-|[ootx? + y?)] dx dy= ||exo«

(-r’ )rdr d=".
, 0 0
Thus /=,/7/2, Fie itesteps can be justified by methods of advanced calculus [6, p. 149].
Figure 6.2
This function is well defined for all ¢ and satisfies the differential equation
(6.56), for 0<|t|< oo.
Exercise
2. Show that all calculations which led to the solution ¢, and then to J,(¢) carry over
to z,=—p with no change, provided 2p is not an integer, so that
(6.63)
J_,(t)= 2
: )are i 1) (5)"
m=
is a second, linearly independent, solution of (6.56) for 0<|t|<0oo. Observe that

this requires the definition of I'(z) for 2z<0, z#0, —1, —2.,...
We have excluded the possibly exceptional cases where p is zero, an

integer, or where 2p is an integer but p is not an integer. In each of these
cases we still obtain the solution ¢,(t)=J,(t) just as before. If 2p is an
integer but p is not an integer, and we attempt to find a second solution
po(t)=t ? > eo ct*, we obtain the system of recursion relations (see
6.9 Bessel Equation ; Some Properties of Bessel Functions 257
Exercise 2 above) c, =0, f(—p+k) +c. =0(k=2, 3,...). Since f(—p+k)

=(—p+k)?—p*=k(k—2p), the method may break down for k=2p, that
is, when we try to calculate c,,. Since p is not an integer, 2p is an odd integer
and since clearly (note that c;=0) c; =c3=---=C,-.=0, the equation for
C2p IS O-cy,+0=0; this may be satisfied with an arbitrary choice of Cop.
Thus we are in the special case mentioned in Section 6.8, Theorem | where
a=0, that is, where no logarithmic term enters in the solution @,, and we
may take as a second, linearly independent, solution the function J_,(t)
given by (6.63). This might have been predicted because '(m—p +1), which
appears in the definition of J_,, is well defined when 2p is an integer but p
is not an integer.
Up to this point, we have proved the following result:
Theorem 1. If p is not zero or a positive integer, then the functions J,(t) and
J_,(t) given by (6.62) and (6.63) respectively are two linearly independent
solutions of the Bessel equation (6.56) on any interval which does not include
t=0.
Exercises
3. Let ¢ be a solution of the Bessel equation (6.56). Show that the function y defined
— by W(t)=(t|'/7 b(t) satisfies the equation
hee
vfs = [v=o. (6.64)
4. Show that
for 0<t<oo. [Hint: Use Exercise 3.]

5. Use the definition (6.62) to show that
amt a) Jy-1(t)+Jp+1(Q=— J,(t) t#0
6. Show that
= 2n\/? /sint 2h \ie a cosias

J32(t)= a2 ee J_3/2(t)= 7a ————sint
for 0<t<oo. [Hint: Use Exercises 4, 5, rather than solving the Bessel equation
of index 3 directly.]
With the aid of Exercises 4 and 5 above, one can, as is done in Exercise
6, obtain representations of J,,.(t) and J_,,2(t), where n is a positive integer,
in terms of sin ¢ and cos f¢.
The cases p=0 and p a positive integer in equation (6.56) still remain
to be treated. In the case p =0, the indicial equation z* =0 has zero as a double
root, and Theorems 1, Section 6.7 and 1, Section 6.8 give the existence of
two linearly independent solutions of the form
pi (t)= Ct eneg HO
= «M18
i]
b2()=le Yuh +4 (0) logit

As we have already seen, we may take ¢, (t)=J,(t), where, from (6.62),
IOs y‘ai (5) (6.65)
Notice that the function Jo(t) is analytic at t=0 even though the differential
equation (6.56) makes no sense there. To find a second solution, we
take 0<t<oo and let
Uy, b,t**!+Jo(t) logt.
We then compute
Be 1
,(t)= ¥ (k+1)
b,t* +Jo(t) logt +— Jo(t)
k=0
5(t)= ))
” =
(k+1) kb,t*"
k-1
++ JG
”
(0) logt +&yJo()—s
, 1
Jo(t)
k=1
and substitute into (6.56) (with p=0). Thus

fo.@)
L[o2()]= Y LS aie bh(k+1) bt***+ Y (k+1) kbyt**?

k=1
+ [176 (t)+tJo(t)+t7Jo(t)] logt+ 2tJo(t) =0

and therefore, using the fact that Jo satisfies (6.56), with p=0,
pee \\i2 2m
bot 2b +) [(k+1)? b,+b,_.] 0**=—2 beat

d 2°" (m!)
k=2 m=
Observing that the right side of this equation contains only even powers of
t, we obtain
by=0, b,=4, 32b,4+b)=0,...
and finally
= et 1
nes ) (Ht 4 2)...
2°" (mn)? m
Thus we may define as the second solution of (6.56), with p=0, the function
>, usually denoted by Ko, given by
#20=Ko
EM(sadam
0=- l)(SPesertons.
Y (666
Clearly, (6.66) is also a solution of (6.56) with p=0 for t<0 if we replace
log t by log |r|. The solution Kg is called the Bessel function of the second
kind of index zero. Thus we have two linearly independent solutions
Jo(t) and Ko(t) of (6.56) with p=0 on any interval not including r=0.
If p is a positive integer n, then it is easily verified from (6.62) and
(6.63) that the functions J, and J_, are linearly dependent, in fact, the
following relation holds between them.
Exercise
7. Show that J_,(t)=(—1)"J,(t) when n is a positive integer.
We also know that in this case the indicial equation has roots n and —n,
and that the solution @, (t)=J,(¢) of (6.56) with p=n corresponding to the
root n exists and is given by (6.62). However, the recursion formulas
corresponding to the root —n are c, =0, f(—n+k) c, +c, 2.=0 (k =2, 3,...)
(see Exercise 2 and the paragraph which follows it). Since f(—n+k
=(—n+k)?—n?=k(k—2n), the calculation of the coefficients c, breaks
down for k=2n. We must therefore use Theorem 1, Section 6.8, to find a
second, linearly independent, solution. Accordingly, we substitute the func-
tion given by
b2()=l-" Y bt +aJ,(t) login

k=
into the differential equation (6.56) with p=n. As usual, it suffices to

consider the case t>0. Rather than go through the calculations, we
formulate a sequence of exercises leading to the desired result.
Exercises
8. Let L be the operator defined by (6.56). By considering L[,(t)]=0, show that the
coefficients b,, b2,...$b2,-, are given by b,=0, k(k—2n) b,+b,-.=0 (k=2,
3,..., 2n—1); and therefore
bo m=1,2,...,.n—1.
b, =b3=---=b2,-1=9, bom
~ 2?™(m!)(n—1)---(n—m)’
9. By examining the coefficient of 1°" in L[,(t)]=0, show that
bo
oy Peteyt:
10. By examining the coefficients of 2"*1, 72"*3,... in L[@,(t)]=0, show that
bon+1= 2n43='*:=0, by, is undetermined, and 2m(2m + 2n) bon +2mt+92n+2m-2=
—2a(n+ 2m) dom (m=1, 2,...), where d3,, is the coefficient of 17" in the expansion
Ola) =e modale ane
11. Show that the choice
d i
b>,= Me (14h++2) by= —2n7'(n—1)!
2 n
(so that a=1), leads to a second solution of the form

ie
b2()=Ky()=—4 55 Verio t
lerics coils
Cy ls |(1H+--+3)
k=0
+(14i+-+ 2) (5)oe log |¢| (6.67)

called the Bessel function of the second kind of index n.
We can now summarize our results as follows:
Theorem 2. If p=0, Jo(t) and Ko(t), given by (6.65) and (6.66) respect-
ively, are two linearly independent solutions of the Bessel equation (6.56) on
any interval not containing the origin. If p is a positive integer n, then J,(t)
and K,,(t), given by (6.62) and (6.67) respectively, are two linearly independent
solutions of the Bessel equation (6.56) on any interval not containing the
origin.
Exercises
12. Show that Jo(t)= —J, (¢).

13. Show that |?’J,_,(¢) dt=0"J,(¢). [Hint: Compute {7’/,(¢)}’ and apply Exercise 5.]
*14. Prove that between two positive zeros of Jo(t) there is a zero of J, (t). [Hint: Use
Exercise 12.]
*15. Show that between two positive zeros of J,(t) there is a zero of J,,,(¢). [Hint:
Show that {¢7"J,(a)}'=—1-"J,,41(0.]
*16. Show that between two positive zeros of J,,,(t) there is a zero of J,(t).
*17. Let ¢ be a solution of y”+p(t) y=0 which is not identically zero on a<t<b, and
let y be a solution of y” + q(t) y=0 which is not identically zero on a<t<b. Suppose
that p and g are both continuous on a<1<hb, and that q(t)>p(t) for a<t<b. Prove
that if ¢;, ft. are successive points in a<tO for t; <<t<t, and that y(t)>0 for t; <t<t,. From the differential equations,
(vo — ov’) =06"— dv" =[4()—p()] (0) (0. Integrate from 1, to f,. Since
by hypothesis, JiLq()—p()] (2) (2) dt>0, we obtain W(t) ¢'(t.)—W(4)
¢'(t,)>0, from which we draw a contradiction.]
*18. Show that for every p>0, J,(t) has an infinite number of zeros on 0<t<oo.
[Hint: Combine the results of Exercises 3 and 17, where for t> fo the equation (6.64)
satisfied by ¢’/?J, (t) may be compared with the equation w” +4w=0 if fy is sufficient-
ly large, and every solution of w’+4w=0 has infinitely many zeros on 0<t<oo.
Then apply Exercise 17 with p(t)=4, g(t)=14+(4—p?)/?? for t>to.]
Lo: a) Show that if p>0, A>0, then ¢,(0) =,/t J,(At) satisfies the equation y’+
+[(G-p?)/t?7] y= Ay. a Use the equation (6.64).]
t) Prove that (22-1) fbal0) dl?)dt=[0) 65(0)—d4(0) A6(01E. CHine: Form
(6,6, = $03) =$101— Pu and use the difierendal equation.|
20. Prove that |6rJ,(Az) J,(ut) dt=0 if A#u whenever / and p are positive zeros of
J,. [Hint: Use Brercise 19.]
AE Prove that ie (At) dt=4[J;(A)]*, where p>0, A>0, and J,(A)=0. [Hint: (i) Show
9
that y(t)=J,(At) satisfies the equation of Exercise 23 below. (ii) Show that this
equivalent to the equation (ty’)’+(A?1t—p?/t) y=0, or, on multiplication by ty’,
to the equation (d/dt) (ty’)’
+(A720? —p’) (d/dt) (y?)=0. (iii) Integrate from 0 to 1,
using y(0)=y(1)=0, y’ (t)=AJ,(Az), and integration by parts.]
Exercises 20 and 21, with p=0, and Exercise 12 give the useful formula
1
_f0 AF
[ol Jota d= Ppa a
0
where /, are positive zeros of Jo.

The function K,(t) given by (6.67), and known as the Neumann form
of the second solution of the Bessel equation (6.56), is not the most con-
venient one for all applications. In particular, as we shall see in Section
6.11, another form is useful in studying the behavior of solutions for large
t. We define the Weber form of the solution
¥,()== [K,(t)=(log2—y) J,(t)] n=O, 1,... (6.68)
where the constant y, known as Euler’s constant, is defined by
1
y= lim (:
thto-+]—logk).
k> oo
A proof of the fact that this limit exists may be found in [6, p. 189].
Exercise
22. Show that J, and Y, are linearly independent solutions of (6.56) on any interval
excluding the origin.
In applications, the Bessel equation often arises in a form different

from (6.56). For example, we have already noted that the equation (6.64)-
can be reduced to (6.56) by a change of dependent variable. We give some
exercises to indicate other equations whose solutions can be expressed in
terms of Bessel functions. In each case, the form of the given solution
suggests the appropriate change of independent or dependent variable
(or both).
Exercises
23. Show that one solution of (y"+1t)'+(k?t?—p’) y=0 is J,(kt), where k is a

constant different from zero. What is the general solution?
24. Show that one solution of sy" + ty’—(f?—at?’) y=0 is Jy/(/at'/s), where a, B,
are constants and as40. What is the general solution ?
25. Show that one solution of y’+k?r"y=0 1s
tian
oe he ine 2)|
where mk 4 —2k; what is the general solution?
26. Show that one solution of (/"y’)'+kt"y=0 is 1°~"7J,(./k ft/s), where s=(m—n+
+ 2)/2, p=(1—n)/2s, n#Am+2. What is the general solution?
27. Show that one solution of ty’+y’+ay=0 is Jo(2(at)'?). What is the general
solution?
The most general equation whose solution can be expressed in terms of

Bessel functions is
ty" +¢(a+2bt’) y+ [ce+dt?>—b(1—a—r) t' +671") y=0
one of whose solutions is giz erry VadtS/s), where p={[(1—a)/
2]7\"/?/s—c (see [11], pp. 155-156). The equations given in Exercises
23-27 are special cases of this equation; they are more easily applied and
perhaps less cumbersome.
If we consider solutions of the Bessel equation for values of ¢near r=0,
Lor PST nen

we observe that, for example,
Since )F_, [(—1)"*1/(m!)?] (¢/2)?"-? converges uniformly and absolutely

for |t|<oo, there exists a constant K such that |)?_,[(—1)"*!/(m!)?]
(t/2)?"~?|<K for |t|<1. Then |Jo(t)—1|< Kr?/4 for |t|<1, or equivalently,
6.10 Singularities at Infinity 263
1 —Kt?/4<Jo(t)<14+ Kr?/4 for |t|}<1. We may express this by saying

that Jo(¢) behaves, for small 7, like 1 up to terms of order #7, and we write
Jo(t)=1+4+O(t?) as t0. (In general, we write f(t) =O (g(d) as toaif |f(d/g(d|
is bounded as ta.) A similar argument using the definitions of the Bessel
functions and simple properties of power series gives the following estimates.
Theorem 3. As t>0
2
Jo(t}=1+0(t?) K,(t)=log|t|-+0(t?) Yo(t)=— log|t| + O(t’)
= TG)
iAa ea x
j=
00) KEE
(n—1)!]t|-"
O(t"t2
ofr = be —n+2
n—1)!|t}""
y(t) =! = 5 +O(t~"**), napositive integer
1
J,(t)= Deie O(RF2). ad 6)1 ihe’
en (een
p>O, p not an integer
Exercise
28. For each of the following differential equations, obtain the general solution in
terms of Bessel functions.
a) ty”—y’—ty=0 b) ty”—3y’+ty=0
ce) ty” +ty’—(t? +4) y=0 d) ty"—y'+4ry=0
6.10 SINGULARITIES AT INFINITY
In many problems it is important to study the behavior of solutions of the

equation
y”+p(t) y+q(t) y=0 (6.69)
as |t|>0o. The simplest way to treat this problem is to make the change
of variable r=1/x and to study the behavior of solutions of the resulting
equation as x0. Thus, we let ¢ be a solution of (6.69) for |t|>R. Let
w(x)=(1/x), p(x)=p(1/x), 9(x)= 9(1/x); these functions are well defined
for |x|<1/R and by the chain rule we have
$(Q=-BVX)=- VO), b'M=2V'W ta V0)

1 A ur I ” 2 /
= xt" (x) + 2x? (x).

Therefore y satisfies the equation
x42" + [2x3 —x7p(x)] 2'+4(x) z=0 (6.70)
in which x is the independent variable and z is the dependent variable.
Conversely, if w satisfies (6.70) and if ¢())=wW(1/d, then @ satisfies (6.69).

We may, of course, make a similar transformation for an equation of any
order.
Definition. We say that co is an ordinary point, a regular singular point,
or an irregular singular point for the equation (6.69) if and only if zero is
respectively an ordinary point, a regular singular point, or an irregular singular
point for the equation (6.70).
We remark that a function f(t) is analytic at oo if and only if g(x)=
f(1/x) is analytic at x=0, or equivalently, if and only if f(t)=)VRa,/4,
where the series converges for |r|>r, with some r>0.
Example 1. Investigate the point at infinity for the equation y”+ay’+by=0, where
a and 6 are constants.
The change of variable r=1/x transforms this equation to x*z”+(2x*—ax?) z’+
z=0, which obviously has an irregular singular point at x=0. Thus the given equation
has an irregular singular point at f= 00.
Exercises
1. Show that the Euler equation 7" + aty’+by=0, where a and 6 are constants, has
a regular singular point at oo.
2. Show that the hypergeometric equation
t(1—t) y’"+[ce—(a+b+1)t] y —aby=0

has regular singular points at 0, 1, and co.
3. Show that the change of variable t=t/b transforms the hypergeometric equation
(Exercise 2) to an equation having regular singular points at t=0, t=b, t=00.
Let boo in the transformed equation, and show that the resulting equation,
called the confluent hypergeometric equation has a regular singular point at 0 and
an irregular singular point at oo.
4. Show that the Bessel equation (6.56), Section 6.9 has an irregular singular point
at oO.
. Prove that co is a regular singular point for the equation (6.69) if and only if(6.69)
mn
can be written in the form

t?y"+ta(t) y+ B(t) p=0
where «(r) and f(r) are analytic at oo, that is, «()=VPo%/A, B(Q=VE of,/t*,
with these series converging for |t|>r with some r>0.
6. Obtain a criterion for co to be an ordinary point for the equation (6.69).
Naturally, Theorems | (Section 6.3), 1 (Section 6.7), and 1 (Section 6.8)

can be applied to Eq. (6.69) at the point t= 0o without change in principle;
we must only replace every series in powers of ftby a series in powers of
l/t. For example, if co is an ordinary point, we expand p(t) and q(¢) in
powers of 1/t and we look for two linearly independent solutions, valid
6.10 Singularities at Infinity 265
near t= 00, of the form ))2 9c,t~* and )'7- od,t~*. We substitute into the
differential equation to obtain recursion formulas for the coefficients c, and
d,. For a regular singular point at t=oo, the obvious modifications of
Theorems | of Section 6.7 and 1 of Section 6.8 hold.
Exercises
7. State the analog of Theorems | (Section 6.3), 1 (Section 6.7), and 1 (Section 6.8)
relative to the point at infinity.
8. a) Compute the roots of the indicial equation relative to the point at infinity of the
hypergeometric equation t(1—?) y’+[ce—(a+b+ 1) t] y’—aby=0.
b) Find two linearly independent solutions valid for large t when a#b. What is
the range of validity of these solutions?
9. Show that the change of variable t=(1—1)/t transforms the Legendre equation
(1-1?) »’—2ty’+a(a+1) y=0 into a hypergeometric equation, and calculate
a, b, c. [Hint: Put z(t)=y(t(t)), and use the chain rule.|
10. Locate and classify the singular points of each of the following differential
equations, including t= 00
a) tA iaike aa +(2t—1) y=0

b) (t—1)? y’+(t? +1) y=0
c) (sint) y-—y=0
d) 2t°y"+3ty’ eo
e) (2t7 +17) y” —y=0
f) (cost)y”ee
1
Nore Meh oa
h) els hae —ty=0
i)(t?-1)? (t—1) (t+2) y’+y'=0
je ae ere:—y=0
k) y’+ey=0
l) (140?) y’+t?y'—ty=0
m) t?(1—t?)? y’+2¢(1—1)
y+ y=0
n) y” +(sect) y=0.
11. Find the general solution valid in some neighborhood of the point = 00 of each of
the following differential equations, and give the interval on which it is valid.
a) ¢?y”+aty’+by=0 (Euler equation)

b) t t(1—2) y’+[c—(a+6b+ 1) ¢] y’—aby=0 (hypergeometric equation)
c) (1—2*) y’ —2ty’ + a(a+ 1) y=0 (Legendre equation)
d) (1- t?) y’—-ty’+y=0
e) y”’+2(P—-2t) y’+y=0
f)<Py"+(?—)y'+2-) y=0
g) 2ey’+ry'—(t4+1) y=0
12. Show that t= oo is an irregular singular point and determine (formally) the form
of the general solution valid for large |¢| for each of the following differential
equation. Find whether the series converges on any interval.
a) aoe
1 ees eA
b) y’—6y'+5y=0
c)
d) y’—ty=0
e) ty’+(c—t) ¥—ay=0 (confluent hypergeometric equation)
6.11 IRREGULAR SINGULAR POINTS, WITH AN INTRODUCTION

TO ASYMPTOTIC EXPANSIONS
We have encountered equations of importance for applications which

have irregular singular points. For example, the Bessel equation has an
irregular singular point at oo, as does every linear equation with constant
coefficients. To illustrate the type of behavior we may expect at an irregular
singular point, we consider the simple first-order equation
t?7y'+y=0.
This equation has an irregular singular point at t=0. However, we can
solve it explicitly by separation of variables and we see that every solution
has the form ¢$(t)=ce'/'=c)
7 o(1/k!) t-* (t#0) for some constant c.
Thus ¢ is not of the form |t|? yee , aS It would be if t=0 were a regular
singular point (see Theorem 1, Section 6.7, and the discussion at the end of
Section 6.8).
Exercise
a) Show that r=0 is an irregular singular point for the equation fy” +ry’—y=0.
b) Find two linearly independent solutions near t=0. [Hint: $,(t)=¢ is one solu-
tion, and another solution can be found by the method of Section 3.6.]
The above example shows that the behavior of solutions near an irregu-
lar singular point may differ sharply from the behavior of solutions near a
regular singular point. To present further difficulties which can arise, we
consider the equation 17y”+(3¢—1) y’+y=0, having an irregular singular
point at t=0. If we try to find a solution of the form |¢|? V2 9¢,¢*, we obtain
z=0, c,=k! (k=0, 1, 2,...), giving a formal solution )'% ok!¢*. Since this
series fails to converge on any ¢ interval, it cannot represent a solution.
6.11 Irregular Singular Points 267
It is our purpose to give an introduction to the study of solutions in a

neighborhood of an irregular singular point. A systematic and rigorous
treatment would require the tools of complex analysis, and must be deferred
to a more advanced course.
In most applications which the reader is likely to encounter, the irregu-
lar singular point is at oo. For this reason, we shall consider the equation
y"+a(t) y'+b(t) y=0 (6.71)

where a(t) and b(t) are analytic at oo, with expansions
= » Gb." b(t)= » ee (6.72)

k=0 k=0
converging in some region |t|>r, with t>0, and we shall assume that
oo is an irregular singular point. If we make the transformation x =1/t and
let z(x)=y(1/t), as in Section 6.10, the equation (6.71) becomes
x42" +[20°—x
4 xX eo f+ > bt |=0 (6.73)
k=0 k=0
By inspection of Eq. (6.73) at x=0 we see that Eq. (6.71) has an irregular
singular point at t= 00 if and only if at least one of the numbers do, bo, b, is
different from zero. Since we wish to assume that oo is an irregular singular
point for (6.71), we shall assume throughout that this is the case.
Motivated by the simple examples at the beginning of this section,
where the irregular singular point is at the origin, and by our study of
regular singular points, we try to see whether (6.71) can be satisfied formally
by a series of the form
Pij=ete Gla sty

k=0
where A, z, and c, (k=0, 1, 2,...) are constants to be determined. We consider

first the case t>0, so that f(t)=e")2 o¢,t **7, and
’ (t)=Ae** y et Me eM y c,(—k +z) f ka ltz
k=0 k=0
Oi =ne- ss Cpe ee hee yy c,(—k +z) tk -itz
k=0 k=0
+e" ¥ c,(z—k)(z—k—-1) "777.

k=0
Using (6.72) and property (vi) (Section 6.2) for multiplication of power series
(here in powers of 1/t), we also have
b(t) P(t)=e"t? bsie Beat«)

a(t) d’ (t)=Ae*t? 3 6s Ay jCa)rtee 3s a,—;(z—j) ct*~*.

k=0 k=0 j=
Substituting in (6.71), after elementary manipulations with series we

obtain (formally)
1
evr at)Cot {g(A) cy +C9(2A +a) 2+AayCo +5400}
+ ¥ <g(A) co,+24(z—k +1) G1 +(2-kK +1) (2—k +2) c_2

=2
k-—1
+ ys [b,-;+Aa,-j;+4-1-,(z—J)] a lao (6.74)

j=
where
g(A)=A7 +A +bo. (6.75)
To have a solution of the assumed form the coefficients of all powers of

1/t in the series (6.74) must vanish. Since we assumed c)#0, we must
choose / to be a root of the equation g(i)=0. To make the coefficient of t~?
vanish, we must have g(2) c, +(2A +4) z+4a;, +6, =0. Since g(A)=0, this
allows us to determine z from the relation
(24 +45) z= —Aa,—b, (6.76)
provided 2A+a,)#0. Note that by assumption dg, bo, b; are not all zero.
If the roots 2,, 4, of g(A)=0 are distinct, since 4, +A,=—dp, we see that
2A+a )#0, when 4 is either 4, or /,. If, however, 4 is a double root of the
equation g(A)=0, the present method breaks down; see Example 3 below.
More precisely, if g(A4)=0 has a double root, then the assumed form of the
solution is incorrect. Ignoring this problem for the moment, we consider
the case that / is a simple root of the equation g(A)=0, so that we may deter-
mine z uniquely from (6.76). We now proceed as follows. Setting the coeffi-
cient of t-* in (6.74) equal to zero and noting that g(A) c,=0 for A=A, or A,
we find that we can determine c,_, from the relation
[by +Aay +(ao + 2A) (z—k+1)] qG-1 = —(2-—k +2) (z-k +1) G_»
je
=o [b,-;+Aa,_j;+4,~;-1(z—J)] Cj, Kon Sh Anae (6.77)
‘=
Exercise
2. Verify with the aid of (6.76) that the coefficient of c,_, in (6.77) is different from zero
fork=2, 3:...
These considerations lead to the following result.

Theorem 1. Consider the equation
y’+a(t) y+b(t) y=0 (6.71)
with a(t), b(t) analytic at co, and having an irregular singular point at o.
Then if the equation g(A)=/? +4 9A+b)=0 has distinct roots 4,, 4, and
if Zm is defined by (2)_,+ 49) Zm= — Ama, — 6, (m=1, 2), then equation (6.71)
is formally satisfied by the two series p_(t)=e?™ |t\?" V9 cht *with ch £0
(m=1, 2), whose coefficients ct”(m) are determined recursively from the equations
(6.77).
Exercise
1. Show that if 4, #A,, then the two series ¢,,, determined by Theorem 1 are
(formally) linearly independent.
To see that Theorem | by no means finishes the problem even if 2, 42,

(unless it should happen that the series terminate after a finite number of
terms), consider the following example.
Example 1. Consider the equation
vs(142) =o 9
where « is a constant, which is of the form (6.71) with a(t)=0, b(t)=1+4/t?. Thus
t=oo is an irregular singular point; g(A)=A?+1, and the roots of g(A)=0 are A= +i.
By Theorem 1, the series $(t)=e"|t/? V2 o¢,t“ (co #0), satisfies (6.78) formally. In the
notation of (6.72), we have a,=0. (k=0, 1, 2,...), and O)=1, b;=0) 5, =a, b,=0
(k=3, 4,...). Then (6.76) gives z=0, and (6.77) gives the recursion formula
2i(—k+1) c,-, = —(k+2) (—k+1) 2 +0c,-2 koe Bere
or
i (k—2) (k—-1) +0
SS
2 k-1
Cy—9 —o7 3 See (6.79)
We observe that if «= —(n—2) (n—1) for some integer n> 2, then the series terminates
after a finite number of terms. However, if « does not have this form, then
oe (k—2) (k— Ito

ee |
co, ko.
Cy-2 k-1
Now the ratio test applied to the formal solution e“ )@9c,t~* of (6.78) whose
coefficients are given by (6.79) shows that the series diverges for all t. Therefore it cannot
be called a solution of (6.78) in the usual sense. We will still find it convenient to call a
series which satisfies a differential equation formally a formal solution even if the series
fails to converge on any interval.
Exercise
4. Find the formal solution of (6.78) corresponding to the root A= —i, and show
that it diverges everywhere.
In view of Example 1, it appears that we have reached an impasse.

This is particularly serious since the Bessel equation, when written in the
form (6.64) (Section 6.10), is of the form (6.78). However, it turns out that
the formal series solution, although divergent everywhere, can still be
given a meaning of great practical significance. We can only give an
introduction to this subject, and for this purpose we consider another
example before returning to the equation (6.78) considered in Example 1.
Example 2. Consider the equation y’+y=1/t. It is easy to verify that the series
p(t)=)
ok !t-*~! diverges for all ¢ but satisfies the differential equation formally.
What meaning can we attach to this series? Instead of working with the series, let us
begin differently. Since the equation is linear and of first order, it follows from Theorem
1, Section 1.4 that every solution has the form
t
e
e =i ‘c+e = — ds
Ss
— oO
for some constant c, where the integral converges for r<0. In particular, the function
is well defined for every t<0 and satisfies the differential equation for t<0. Repeated
integration by parts gives
t t
1 x e ile al 7 es mee
AO era ; |Sdsata 542k ‘ |Sas=) kim * 1+R,
ies! = 66 k=0
where
t
Ss
R,=(n+1)!e* |ae
s
We observe that the integrated terms in this expression for the solution w (well defined
for t<0) coincide with the first (n+ 1)-terms of the (divergent) formal series solution.
Now, let us examine the remainder term of R,. We see that for 1<0
or
PHO), Me 2. sev
(n+1)!
feet
W()— Yee & >
I¢|
or
lim |¢|"*?
te
wi Skeet == (|) 72s()alee
To put this in another way, we may say that even though the series ) 9k!¢7-*~!
diverges for all r, the error made in approximating the solution w(t) by the first n terms
of the series is less in magnitude than the (n+ 1)st term, for every integer n and for
t<0. We note, however, that for a particular value of 7, the approximation may not be
improved by taking more terms of the series (for example, t= — 1).
The above considerations suggest the following definition, due to the
French mathematician H. Poincaré.
Definition. Theformal series ) 7° 9a,t~* is said to be an asymptotic expansion
of a functionf(t) defined for t>r as t> ifand only if for each n=0, 1,...,
r|£00 y ato, t 00.
In this case it is customary to write f(t)~ V2. oa,t~* (t 00).

Slightly more generally, the formal series e” )'? 9a,t* is said to be an
asymptotic expansion of a function f(¢) as t> 00, written f(t)~ eo"Vat",
if and only if for each n=0, 1,...,
fens Dy at Loo, to.

k=0
A similar definition may be given for f(t) defined for t<0 as t>—o.
In the above example we have shown that the (divergent) series ) 2 ok !¢~*~
is an asymptotic expansion of the solution w(t) as t> — oo, and its partial
sums may be used to evaluate the integral e~'|‘_,, (e*/s) ds. Incidentally, the
integral |‘,,(e*/s) ds is called the exponential-integral function, denoted by
Ei(t), and its values have been tabulated by precisely this method [13, 14].
Of course, every convergent series is also an asymptotic series, but as
we have seen, the converse is false. If a formal series =)’ 9a,t~* is an
asymptotic expansion of a function f as too, we write f~ ® (too). The
coefficients a, in this asymptotic expansion are uniquely determined by the
function /, since
@j=him fit), ag—lim tf ()—ag |,

t>o
a, = lim | £()-a0- 2 ae ot].

t=
However, different functions may have the same asymptotic expansion.

For example, e‘ and 0 are both represented asymptotically by the zero
series. Thus if f~@® (too), then f+e ‘~® (too). More generally, if g
is any function asymptotic to zero, then f+g~@®. Although we shall not
go into these matters, it is possible to operate with asymptotic series much
as with power series. The interested reader is referred to other books such as
[545 (9nd [.184:
To illustrate how some of these ideas may be applied in the case of an
irregular singular point, we return to Example 1, equation (6.78), in the form
y’+y= = y. (6.80)
We have already seen that (6.80) has a formal solution e” )'7- 9¢,t~“,where
Cyo=1 and c, (k>1) is determined recursively from (6.79), and that this
series diverges for every t>0. We shall now show that the series represents
a solution of the equation in the sense that (6.80) has a solution for which
the formal series constructed above is an asymptotic expansion. For large t,
the equation (6.80) “resembles” the equation y” + y=0, and this suggests that
(6.80) may have a solution $(t) which behaves like e” as too. If (6.80) has
such a solution, we can use the method of variation of constants (Section 3;
see in particular Exercise 5,) to see that ¢ must satisfy the equation
t
b()=e'e—a | 969 ds (6.81)

s
to
for some constant c. The reader may verify that if @ satisfies (6.81), then
it also satisfies (6.80). The form (6.81), however, is not suitable for our
purpose, for even if the integral on the right side exists as too, this
solution @ would not behave like e'* unless the integral approaches zero.
To obtain a more suitable form, we write
isng) $0am |ints) 2 a| sin(e—3) 29 as,
If the first integral on the right side exists, it is a solution of the homogene-
ous equation y”+y=0 (verify this fact), and may be thrown into the term
e“c in (6.81). Then, we take c=1 to obtain the new equation
[e)
(yea |sin(e—s) Pas, (6.82)

t
It is easy to verify by direct substitution that if @ satisfies (6.82), then ¢

satisfies (6.80), and ifit can be shown that a solution $ of (6.82) is bounded for
t>a>0, then the integral in (6.82) approaches zero as too, and thus
p(t)—e"0 as too. It will be shown in Section 8.1, Exercise 13, that (6.82)
has a solution @ on 1 <t<oo and that
Ip(t)i<el", 1<t<oo. (6.83)

Assuming this result we now obtain
co
16(t)—e| <a |sin (e— 3) OO as ja ol (2

Breil tes aniuephes
t t
(6.84)
and thus lim,.,,,(¢(t)—e")=0, or $(t)~e'(t-00). We also observe that we
have obtained the first term of the formal series solution of (6.78). But we
can do more; we write (6.82) as
ice) 00
(t)=e"+a |sin(t—s) _ ds+a |sin(t—s) ee ds. (6.85)
We now use the estimate (6.84) to obtain

foo)
is
(yea |sin(t—s) z ds
t
|p(s)—e*| : (ds |a|7e!*!

<a ||sin(t — s)|ee
——,—— ds <|q|e! | 3 nore
We can now evaluate the integral

oO
; e's ets) pe t@-s) e's
[sint—9)S as= | 5; ds
i
eit eit e2's A 1 e2's

= ii
ree. (= 2i alee s d t)
—
elt
=5, +9!)
where |g(t)|<k/t? for > 1 and some constant k. We may now write
ieee2it |eBone
eae
Or
——-+|g(t)|, tl.
Thus t[e~"#(t)—1—«a/2it]+0 as t>00, or $(t)~e"(1+4/2it) as too.

Thus we have obtained the first two terms of the formal series solution of
(6.78). The above calculations show that for very large values oft,we obtain a
better approximation to the solution by using two terms of the formal series
than by using one term.
We can repeat the procedure by using #(¢)—e“(1+a/2it)=A(t), where
h(t) t? is bounded as too, in (6.82) to obtain the first three terms of the
formal series. We would also see that this gives a still better approximation
to the solution for large ¢. Continuing, we could show that for the solution
(0) of (6.78) which behaves like e'’ as t> 00,
b()=e" Y qt *+h,(), 121 (6.86)

for every integer n>0, where c)=! and c, is determined from the equa-
tion (6.79) for k>1, and where |h,(t)| °*! is bounded as t>00. Thus
p(t)~e" Ve oct “ (too). This is the sense in which the divergent series
found by formal operations is useful. The expression (6.86) may be used to
approximate the solution ¢(f) for large t.
Exercises
5. Show that the Bessel equation

ty" +ty +(t?—n’) y=0
has a solution #(t) such that
e i: Aam_Q-n) =) =0?) G0) @—?)
2t aa Nie 2 ae
as too. [Hint: Use Exercise 3, Section 6.9, and the results of this section. You
may assume the validity of (6.86).}
6. Using Theorem |.and arguments similar to the above, show that there is also a
solution y(¢) of the Bessel equation such that y(t)~e~"/t- '/? YR odyt~* as t00
(do =1), and compute the recursion formula for the coefficients d, and the first four
nonzero terms of the asymptotic expansion.
7. By suitably combining the series ¢ and w in Exercise 5 and 6, show that there are
solutions y, and y, such that
1
y1(t)~— LU, (t) cost+ U,,2(t) sint]
t
=70O
1
y2(t)~—= [U,1(t) sint —U,2(t) cost]
t
where
It can be proved that the Bessel function J,,(t) satisfies
J)~(=) o.(t) cos} ("

+5)he U2 (t) sin} (n+5)h
(6.87
as t>oo. To illustrate the use of the asymptotic series in actual calculations,
we remark that if we use the series
pt ao
SG a SS ree
with t=6, we obtain 0.15067 as an approximation to Jo(6), correct to
four decimal places, provided that we use terms of the series up to the one in
1?°. On the other hand, the asymptotic series (6.87) gives 0.15064, correct to
five decimal places, with the use of only four terms of each of the zeries U9 (¢)
and U,,(t).
From this and similar analysis for Y,,(t), defined by (6.68), we can obtain
the following result which complements Theorem 3, Section 6.9.
Theorem 2
1)~(2) cos| Ot
wd~(Z) “4 jee |
One important consequence of Theorem 2 is that J,(f) has infinitely
many zeros 0<?, </,<...<#,<..., and for large k, t,,, —#, is close to x.
This fact is very useful in the study of boundary-value problems involving
the Bessel equation.
The above considerations should suggest to the reader that the general
theory for the second-order equation (6.71) with an irregular singular
point at infinity is quite complicated. By a suitable generalization of the
above techniques it may be shown that if g(A)=A*
+a A+bp has distinct
zeros, then corresponding to the formal series obtained in Theorem 1, the
equation (6.71), has solutions @ and W having these formal series respectively
as asymptotic expansions,as t > ©.
Finally, we make a few remarks about the case of a double zero of g(A),
for which the above method breaks down completely. We consider an

example which indicates the possibilities in the general case.
Example 3. Consider the equation
ty” —2ty’+(1+2) y=0. (6.88)
This has the form of the equation (6.71) with a(#)=—2, b(t)=1+1/t, and g(a)=
A? —2A+1. Thus A=1 is a double root of the equation g(4)=0. Put y(t)=e"u(t). Then
y (t)=ceu(t)+ eu (t)
y" (t)=c?eu(t) + 2ce“w (t)+ eu" (t).
Thus
ty” —2ty +(1+t) y=e"[tu” +2(c—1) tu + {tc* —2tc+(1+8) c} ul].

If we choose c=1, we see that u satisfies
tu -+-u=0.
(This equation actually has A =0 as a double root of the corresponding indicial equation.)
We now make the stretching transformation t=x?, and we let v(x)=u(x7). Then
tu” +u=(§) (d?v/dx?) —(1/4x) (dv/dx)+v, and v satisfies the equation
1
v”——v'+4v=0 (6.89)
x
with x as independent variable. For (6.89), the relevant polynomial is g(A)=/?+4,
and its zeros = +2i are distinct. Thus the methods of Theorem | may be applied to
(6.89). According to Theorem 1, (6.89) has two formal solutions of the form
niexp (ix) cere” mec ad

k=0
where z and c, (k=1, 2,...) are determined recursively by substitution. Thus the original
equation (6.88) has two formal solutions of the form
eS
[t}"/? exp(t+2i,/t) Y-ot-*?, com
k=
and it can be shown that there exist solutions ¢ and w of (6.88) which have these formal
solutions as their respective asymptotic expansions as t>0o. Thus we see that in the
general case, the formal solutions are more complicated than in the case covered by
Theorem | in that the formal series involves fractional powers.
Exercises
8. Carry out a similar analysis for the equation ty” + y’+ y=0. Construct two linearly
independent formal solutions as t>o.
9. Find two linearly independent formal solutions as too for the equation y” + ty=0.
10. Find two linearly independent formal solutions as t>oo for the equation
y" +(—4)+ (c/t)+(@—m’)/t7] y=0 (confluent hypergeometric equation), and show

that if c=0 one solution is \/ ¢J, (it/2).
11. Give a complete discussion of the equation y’+(1+a/t+/t*?)
y=0 where «a,
B are real constants, both for ¢ near 0 and as too.
CHAPTER 7
Boundary- Value
Problems
7.1 INTRODUCTION
In the problems considered so far, we have singled out a particular solution

of a differential equation by specifying some suitable initial conditions.
Another important class of problems, called boundary-value problems,
involves the determination of solutions of a differential equation which
satisfy prescribed conditions at two given points. Such conditions are called
boundary conditions.
Example 1. The rotating string. To illustrate how boundary-value problems arise in

the mathematical formulation of physical problems, we consider the possible displace-
ments of a rotating string of length L. We denote by y(x) the displacement of the string
at x (0<x<L), and we suppose that the string is attached to fixed supports on the x
axis at x =0 and x=L. We wish to determine the possible displacements of the string as
a function of x (0< x <L) when it is allowed to rotate at a uniform (as yet undetermined)
angular velocity wm about its equilibrium rest position along the x axis. We assume a
tightly stretched flexible string with linear density p (that is, a mass of p units per unit
length) (see Fig. 7.1). One possibility is that the string remains in this equilibrium
Figure 7.1
278
Teh Introduction 279
position y=0. We wish to determine whether there are values of the angular velocity
(called critical speeds) for which the string can assume some other shape (called a
standing wave).
To arrive at a mathematical model for this problem, we shall use Newton’s second
law of motion together with the following physical assumptions:
i) The tension is large enough so that the additional stress introduced by

curvature of the string is negligible.
ii) The tensile force at any point acts in a direction tangent to the string and has a
constant magnitude 7.
ili) The angle between the direction of the string at any point and the equilibrium
position y=0 is small.
We examine the portion of the string between x and x ++Ax, where 0<x<L. The
mass of this portion of the string is p As, where s is the arc length measured along the
string; this mass is approximately p[(4x)*+(4y)*]’/”, if |4x| is small. The rotation of
the string about the x axis with angular velocity m produces an acceleration of
magnitude wy directed vertically toward the x axis.
Exercise
1. Use the argument by which equation (2.6) Section 2.2 was derived to show that
the rotating string produces an acceleration —@yu,, where u, is the unit radial
vector. (Here y replaces r and 0=at.)
Since the acceleration vector is vertical, only the vertical component of the tensile
forces on this portion of the string is relevant and this vertical component is
T [sin «(x + Ax) —sin «(x)]
where «(x) is the angle between the string at x and the horizontal (see Fig. 7.2). Thus
Newton’s second law of motion gives
T [sin «(x +Ax)—sin «(x)]= —pam*yAx [(2) | Gt)
aia Seeker oronieebeeeret estraCar

ths p: easees
wt
Figure 7.2
280 Boundary-Value Problems
If we divide the equation (7.1) by 4x and take the limit as 4x0, we obtain
d dy Zan
cEre [sin a(x)]+ p0%y|1+ (*) | =0. (2)
However tan «(x)=dy/dx and therefore

tan a(x) dy /| (2) | :
= = 1 + ss . We
na) [1+tan?a(x)]*/? dx dx oe
Substituting (7.3) into (7.2), we obtain the equation of motion of the string
ik \2|[(2) | beoe%y| 1+(2)J =0 (74)

which we may also write in the form
d7y dy 2512
ieae 2yy |
11+oF (=)
(— ==1()h Us
(7.5)
Exercise
2. Carry out the differentiation in (7.4) and derive the equation (7.5).
We observe that we cannot solve (7.5) by the methods we have studied, and we ask
the reader to accept the statement that the equation does not have solutions which can
be written in closed form. It is not even obvious that solutions of (7.5) satisfying given
initial conditions at x =0 exist on the whole interval 0<x<L, although this can in fact
be proved.
To simplify the complicated equation (7.5), we use the assumption (iii), which
implies that if y(x) is the displacement of the string, then |y’(x)|=|tan «(x)| is small for
O0<x<L. Then, replacing dy/dx in (7.5) by 0, we obtain the simplified model
eee
Fi PPO = 1.6
(7.6)
which we shall use for the differential equation governing the motion of the rotating
string.
The differential equation (7.6) alone does not specify the motion of the string com-
pletely. We have assumed that both ends of the string are attached to a fixed support.
This means that the displacement y(x) of the string must also satisfy the boundary
conditions
y(0)=0, —y(L)=0. (7.7)
Clearly y(x)=0 is a solution of equation (7.6) satisfying the boundary conditions (7.7);
this solution is called the trivial solution.
The mathematical problem is to determine the values of w for which (7.6) has a
solution not identically zero.on 0<x<L which also satisfies the boundary conditions
(7.7), and to determine the corresponding solution @. We shall see in the next section how
to solve this problem.
Wee Homogeneous Boundary-Value Problems 281
We remark that in assuming that dy/dx is small to derive the equation (7.6) we do not
rule out the possibility that (7.6) may have solutions for which dy/dx is not small. How-
ever, we may expect that only solutions with dy/dx small actually approximate solutions
of (7.5), and that the solutions of (7.6) for which dy/dx is not small need not have any
physical significance.
Exercises
3. Look up the derivation for the differential equation for the critical speeds of a
rotating shaft (see, for example, [11], p. 193).
4. Look up the derivation for the differential equation for the buckling of a column
under an axial load (see, for example, [11], p. 198).
We may impose other types of boundary conditions instead of requiring both ends
of the string to be attached to a fixed support. For example, we might assume that the
end at x =L is attached to a yielding support with a restoring force proportional to the
stretching. This gives a boundary condition of the form
Ty (L)=—ky(L)
instead of y(L)=0. The “limiting case” of such a condition is called a free end condition,
y (L)=0. This corresponds to a string unattached at the end x=L.
7.2 EXAMPLES OF HOMOGENEOUS BOUNDARY-VALUE PROBLEMS
Before developing a general theory for a class of boundary-value problems

we consider a number of typical examples in this and the next section.
Example 1. Let us examine the boundary-value problem consisting of the differential
equation
y"+Ay=0 (7.8)
together with the boundary conditions
y(0)=0, =y(n)=0 (7.9)

We wish to find all values of A for which there is a solution @ of (7.8) which is not
identically zero on 0<t<z and which satisfies the boundary conditions (7.9).
Exercise
1. Transform the problem (7.6), (7.7) for the rotating string into the problem (7.8),
(7.9) by making the change of variable t= x/L and letting A= pw*L?/Tn’.
The general solution of (7.8) can be written in the form
(t)=c, exp(i,/At)+c, exp(—i,/At) (7.10)
(see Section 3.4) for every A, real or complex, except 14=0. We may write
Ja=a+ ip with «, B real (see Appendix 3). We shall first show that the boundary con-
ditions (7.9) cannot be satisfied by a nontrivial solution of (7.8) unless eA is real, that is,
unless 6=0. With Ji= a+iB, (7.10) becomes
O=c.e. Fee
Imposing the boundary conditions (7.9), we obtain
o(0)=c, +c,=0,
(x) =e, 4 t" +-c,e8- *=0.
This is a pair of simultaneous algebraic equations for the constants c,, ¢,. By Theorems
1, 2 in Appendix 1, there is a nontrivial solution (that is, a solution with c,, c, not both
zero) of this algebraic system if and only if the determinant A of coefficients is zero. But
A=el@e7 iat __ eo Ba oian
The condition A=0 means that

eft olan _ 9 —buoian _ ()
or equivalently,
2Br 2ani
eck=—e"" = Cos 2am +1 Sin 20
Because f is real, e?9" is real. Therefore sin 2xn=0 and cos 2an=e74". The first of
these equations is satisfied if and only if <=n (=0, 1, 2, 3,...) and in this case the second
of these equations gives 1 =e?*; from this it follows that B=0.
Therefore the boundary-value problem (7.8), (7.9) has a nontrivial solution if and
only if Ja=n= 12352 (he scasenor—025 —Osimplics Ji=0, and is considered
separately below.) The corresponding nontrivial solutions are (from (7.10))
Of)=cje foe. (i= ieee):
However, from the first boundary condition, we know that c, +c,=0, so that
o(t)=c,(e"—e ™)=2ic; smut “(n=1, 2...)
or finally
(t)=A, sin nt (n=, 25::-) (7.11)
where A, is an arbitrary (real) constant. Note that A, remains completely undetermined
by the problem.
It remains to consider the possibility of anontrivial solution of (7.8), (7.9) for A=0.
In this case the general solution of (7.8) is
p(t) — Cy ate Cot
However, this solution satisfies (7.9) if and only if c;=c,=0 and thus there is no
nontrivial solution for A=0.
Exercise
2. Show that #(‘)=c, + cot satisfies (7.9) if and only if c; =c,=0.

Summarizing, we see that the boundary-value problem (7.8), (7.9) has nontrivial
solutions if and only if
72 Homogeneous Boundary-Value Problems 283
The corresponding nontrivial solutions are

@,(t)=A, sinnt (n=1,2,...)
where A, is any constant. The values of A (in this case A=n7?) are called eigenvalues
(or characteristic values) of the boundary-value problem, while corresponding
solutions (in this case A, sin nt) are called eigenfunctions (or characteristic functions)
of the boundary-value problem. The reader will note that the boundary-value
problem (7.8) and (7.9) has an infinite, but countable, sequence of real positive eigen-
values A=n7; these eigenvalues are not equally spaced and tend to oo asn—0o.
Exercises
3. Assuming that / is positive, we know that the general solution of (7.8) can be written
in the form
sin, /At
o(th=c, cos,/At+c,
where the factor . is inserted for convenience. Apply the boundary conditions
(7.9) to this form of the solution and determine the eigenvalues and eigenfunctions.
4. Compute the limit as 4-0 of the general solution $(t) in Exercise 3.
5. Show that |§ ¢,(¢) ¢,,(t) dt=0 if m#n and [5[¢,(0)]? dt=2A,7/2. [Hint: 2 sin nt
sin mt=cos (m—n) t—cos (m+n) t.]
6. Show that for the boundary-value problem (7.6), (7.7) for the rotating string there
is an infinite sequence of angular velocities w, =nn/L(T/p)'/? (n=1, 2,...), for which
there is a nontrival solution and a corresponding sequence of solutions ¢, (x)=,
sin (nzx/L) (n=1, 2,...), where the A, are constants which we cannot determine
from the problem.
We now observe that the rotating string problem posed in Section 7.1
and solved in Exercise 6 above has solutions which do not satisfy the as-
sumption made in obtaining the linearized equation (7.6), namely, that the
angle between the direction of the string and the horizontal is small. For
the solution @¢,(x)=A, sin(nzx/L), we have ¢/(x)=(nxA,/L) cos(nnx/L)
(n=1, 2,...), and this becomes large as n increases. This suggests that in
fact only the first few of the eigenfunctions are physically meaningful, and
the skillful reader may verify experimentally that this seems to be the case.
The mathematical question of whether the solutions of the linear boundary-
value problem (7.6), (7.7) approximate the solutions of the nonlinear bound-
ary-value problem (7.5), (7.7), or even whether there are values of w for which
the nonlinear problem has solutions, is an extremely difficult one to which
no satisfactory answer can be given here.
Example 2. Consider the boundary-value problem defined on the interval 0<t<z by
the differential equation
y"+dy=0 (7.8)
and the boundary conditions

y(0)=0 = y'(x)=0 (7.12)
By an argument similar to that used in Example 1 above, it may be shown that this
problem has no complex or negative eigenvalues. (You are invited to carry this out as
an exercise.) Assuming this fact, we need only consider 2>0. For A>0 the general
solution of (7.8) is
in,/At
o(t)h=c, cos. /At+c, = : (7.13)
Ji
For A=0 the general solution of (7.8) is
P(t)=cy +cat (7.14)
(see also Exercise 4 above). Using (7.13) for A>0, we have
¢'(th=-c, a: sin, /At+c, cos, /At.
Imposing the boundary conditions (7.12) we obtain
$'(0)=0 c, +c, =0,
'(n)=—-cy a. sin, /Am+c, cos Jin =0.
The determinant of coefficients of this algebraic system is

A=.// sin, /AE
and for the case A4>0, A=0 if and only if A=n? (n=1, 2,...). The case 42=0 is now con-
sidered separately.
Exercise
7. Show, using (7.14) and (7.12), that 2=0 is an eigenvalue of the boundary-value
problem (7.8), (7.12).
Thus the boundary-value problem (7.8), (7.12) has the sequence of

eigenvalues
A=n? (n=O50 ee)
and
O,\t)—A, cost aap =Os1 =)
as corresponding eigenfunctions. Note that ¢)(t) is the constant Ap.
Exercises
8. Show that [5 ¢,(t) ¢,(t) dt=0 if m#n, and that {5 [¢0(0)]? dt=24,2/2 if n>0,
while (§[¢(t)]? dt=2A)?.
9. Find the eigenvalues and corresponding eigenfunctions of the differential Eq. (7.8)
subject to the boundary conditions y(0)=0, y’(x)=0.
10. Interpret Exercise 9 for the rotating string (see Section 7.1).
1 Homogeneous Boundary-Value Problems 285
Example 3. Consider the boundary-value problem defined on the interval 0<t<z

by the differential equation
y’+Ay=0 (7.8)
y(0)=0, y(a)+ky'(x)=0 (7.15)
where k is a given, real, nonzero constant. By an argument similar to that used in
Example | above, it may be shown that this problem has no complex or negative
eigenvalues ifk 4 —1.
Exercises
11. Show that the eigenvalues of the boundary-value problem (7.8), (7.15) are the solu-
tions (if any) of the transcendental equation
tana A= Sys
12. Show that if /, is an eigenvalue of the boundary-value problem (7.8), (7.15), then the
corresponding eigenfunction is ¢,(¢)=A, sin \/4, t.
The above exercises show that the problem of finding the eigenvalues
of the boundary-value problem (7.8), (7.15) reduces to the problem of
solving the transcendental equation
tan. /An=—k./A. (7.16)

Although we cannot solve this equation explicitly, we can show that it
has an infinite sequence of solutions. As ./A x increases from (n—4) x to
(n +4) x, tan./Z 7 takes on all real values exactly once, with
lim tan/dn=—oo lim tan /An=+00.
JArn—1/2 Aen 1/2
Thus there is exactly one value \/A, between (n—4) and (n+4) such that
tan ./1,m=—k Ae This shows that there is an infinite sequence of
solutions 1, of (7.16) and hence an infinite sequence of eigenvalues of the
boundary-value problem (7.8), (7.15). In fact, if k>0, —k ./A decreases
as A increases and we may conclude that the solution of (7.16) between
(n—4) and (n+4) tends to (n—4), so that ./A, is given approximately by
(n—4), for large n. In a similar way, we may see that if k <0, RA, is given
approximately by (n +4) for large n (see Fig. 7.3).
In the above discussion, we have ignored the possibility that A=0 may
be an eigenvalue. Whether 1=0 is an eigenvalue depends on the value of
the constant & in the boundary condition at t=7.
Exercises
13. Show that A4=0 is not an eigenvalue of the boundary-value problem (7.8), (7.15)
unless k= —1.
Figure 7.3
14. Show that if k= —zx, 1=0 is an eigenvalue of the boundary-value problem (7.8),
(7.15) and $o(t)=Aot is a corresponding eigenfunction.
Example 4. Consider the boundary-value problem defined on the interval 0<t<x

y"+Ay=0 (7.8)
ay(0)+By(0)=0, — yy(x)+dy'(x)=0 (7.17)
where «, f, y, 6 are real constants. By an argument similar to that used in Example |
above, it may be shown that this problem has no complex or negative eigenvalues.
Exercises
15. Show that the eigenvalues of the boundary-value problem (7.8), (7.17) are the
solutions (if any) of the transcendental equation
in,/A
ce * = —(a6 — By) cos,/An.
Vi
*16. Show that the equation in Exercise 15 has an infinite sequence of real positive roots.
[Hint: If B50, the left side of this equation oscillates with amplitude pd ofA
which is large for large A, while the right side is bounded; thus there is a root
near each zero of the left side. If B5=0, the left side oscillates with amplitude
ay/./A, which is small for large A (or is identically zero if ~y=0); thus there is a root
near each zero of the right side.]
1k?5 Homogeneous Boundary-Value Problems 287
*17. Show that if ay=B5=0 but ad—fy#0, the eigenvalues of (7.8), (7.17) are
A=(n+4)? (n=0, 1, 2,...).
*18. Show that if «5 —fy=0 but either xy 40 or £540, the eigenvalues of (7.8), (7.17)
arenA=n? (n=071;.2,:2).
Example 5. Consider the boundary-value problem defined on the interval 0<1<2z

y’+Ay=0 (7.8)
and the periodic boundary conditions
y(0)—y(2z)=0, = (0)—y' (22) =0. (7.18)
By an argument similar to that used in Example 1 above, it may be shown that this
problem has no complex or negative eigenvalues. As in the previous examples, we use
the general solution
in, /At
(t)=c, cos,/At+c, ae (7.13)
of the differential equation (7.8) for 4>0. Examination of the boundary conditions
(7.18) shows that they are satisfied if fA is a positive integer. Thus there is an infinite
sequence of eigenvalues 1=n? (n=1, 2,...). Corresponding to the eigenvalue n? there
are two linearly independent eigenfunctions, namely, ¢,(t)=A, sin nt and y,(t)
= B, cosnt, for every choice of the constants A,, B, (n=1, 2,...). In the examples con-
sidered previously there was only one eigenfunction corresponding to each eigenvalue.
You should note that as there cannot be more than two linearly independent solutions
of the differential equation (7.8) for any value of A (Theorem 1, Section 3.3), there can-
not be more than two linearly independent eigenfunctions corresponding to any eigen-
value of any boundary-value problem for the differential equation (7.8).
The boundary-value problem (7.8), (7.18) has another eigenvalue, namely, A=0.
For A4=0 the general solution of the differential equation (7.8) is $(t)=c,+ ct. The
boundary conditions (7.18) give c,=0, and we see that Wo(t)= Bp is an eigenfunction
corresponding to the eigenvalue 1, =0 for every choice of the constant Bo.
Exercises
19. Show that there are no negative or complex eigenvalues for the boundary value
problem (7.8), (7.18).
20. Show that [3*,(t) p(t) dt=0, |3"Wa(t) W(t) dt=0 for n#m and 5" ¢,(t) Wn(t)
dt=0, where iA are the Seentuiclions of (7.8), (7.18).
21. Show that [3*[¢,()]? dt=74,7, [2"[W,(0)]? dt=nB,”, while
[$*[Wo(t)]? dt=20By’.
22. Show that A, sin ae cosnt is an eigenfunction of the boundary-value problem
(7.8), (7.18) corresponding to the eigenvalue A= n? for every choice of the constants
An Be
If you are acquainted with linear algebra, you should note that the
eigenfunctions of the boundary-value problem (7.8), (7.18) corresponding
to the eigenvalue A2=n* form a two-dimensional vector space which has

an orthogonal basis consisting of the functions sinnt and cosat (n=1, 2,...).
This is the content of Exercises 20-22 above. The eigenfunctions corre-
sponding to the eigenvalue 4=0 form a one-dimensional vector space
spanned by the constant function 1.
Exercises
23. Find the eigenvalues and eigenfunctions of the boundary-value problem defined
on the interval a<t<b by the differential equation y’+Ay=0 and the boundary
conditions y(a)=0, y(b)=0.
. Find the eigenvalues and eigenfunctions of the boundary-value problem defined
on the interval 0<r<1 by the differential equation y’+Ay=0 and the boundary
conditions y’(0)=0, p(1)=0.
. Determine all real eigenvalues and the corresponding eigenfunctions of each of the
following boundary-value problems. If the eigenvalues are roots of a transcendental
equation which cannot be solved explicitly, give the equation for the eigenvalues
and the form of the eigenfunctions.
a) y’+Ay=0, y(0)=0, y’(z)=0
b) y"+Ay= ere }=0; ines
c) yp’+Ay=0, y(0)=0, y(a)+ i
d) y"+2y'+(A4+1) y=0, y(0)=0, v(xsay
e) y’+(1+A4) y'+Avy=0, ae y(1)=0
f) t?y”—Aty'+Ay=0, y(1)=0, y(2)—y'(2)=0
g) YY —Ay=0, y(O=y(M=y"(0=y"()=0
26. Determine the eigenvalues and eigenfunctions of the boundary-value problem
defined by the differential equation
y—Ay=0
and each of the following sets of boundary conditions

a) y(0)=0, y”(0)=0, y(Z)=0, y’(L)=0
b) »(0)=0, »’(0)=0, »(L)=0, »’(L)=0
c) y(0)=0, y’(0)=0, y’(L)=0, y’” (L)=0
EXAMPLES OF NONHOMOGENEOUS BOUNDARY-VALUE

PROBLEMS—THE GREEN’S FUNCTION
Example 1. The physical problem of a rotating string to which a given external force
J (t) is applied leads to a nonhomogeneous boundary-value problem of the form
ytdy=f (t) (7.19)
with boundary conditions such as
y(0)=0 ~— y(n) =0 (7.20)
De) Nonhomogeneous Boundary-Value Problems 289
in the case of fixed end points. Solve this boundary problem explicitly. In (7.19), A is a
given constant which we assume to be positive.
We shall see that the problem (7.19), (7.20) is intimately related to the corresponding
homogeneous boundary-value problem
y"+Ay=0 (7.21)
with boundary conditions (7.20) (see Example 1, Section 7.2) and it will be essential to
distinguish the cases when / is an eigenvalue and when J is not an eigenvalue of (7.20),
Cre):
By the variation of constants formula (Section 3.7; see particularly Exercise 5) the
general solution of(7.19) is
1
(t)=c, cos,/At+c, et Vilae (t—s) f (7.22)
where we have used the general solution of the differential equation (7.21) given by
(7.13). We wish to have the solution ¢ defined by (7.22) satisfy the boundary conditions
(7.20). To satisfy the condition y(0)=0 we must clearly have c,=0. The boundary
condition y(z)=0 gives
sin ./A 1
AL aaa sin,/A(n—s) f(s) ds=0. (7.23)
a ah ialMe
If, for the given /, sin Sa m#0, Eq. (7.23) can be solved uniquely and we obtain
‘ey — fonsAa
‘ re
Therefore, from (7.22) we have

14
- aso [si —s) f(s) ds ss A(t—s) f(s) ds. (7.24)

aR OR Aas gts : _
We emphasize that the condition sin aA 1#£0 is precisely the condition that A is not an
eigenvalue of the homogeneous problem (7.21), (7.20) (see Example 1, Section 7.2).
If A=A* is an eigenvalue of the homogeneous problem (7.21), (7.20), that is, if
sin ./A* 1=0 and 4* =n’ for some positive integer n (see Example 1, Section 7.2), then
Eq. (7.23) is satisfied with an arbitrary choice of c, if and only if
[ro sin,/A*(x—s) ds=
or equivalently if and only if

us
|f(s) sin,/7* s ds=0. (7.25)

0
Exercise
1) Derive (7.25) [Aint sin 1*(n—s)=sin,/2* n cos A* s—cos./A* x sin hess
and sin ,/A* n=) |
If A* is an eigenvalue and the condition (7.25) is satisfied, the solution is
DYCds oe |rosin 7*(t—s)ds
where c, is arbitrary and cannot be determined from the boundary condi-

tions because sin,//* ¢/,/A* is an eigenfunction of the homogeneous
problem (7.21), (7.20). Thus if A=* is an eigenvalue, the nonhomogeneous
problem (7.19), (7.20) cannot have a unique solution. Either it has a family of
solutions if the condition (7.25) is satisfied, or it has no solution if the condi-
tion (7.25) is not satisfied.
If 2 is not an eigenvalue of the homogeneous problem (7.21), (7.20), then
the solution @, given by (7.24), of the nonhomogeneous problem (7.19), (7.20)
is unique. To show this, we suppose that w is another solution of (7.19)
satisfying (7.20). Then @— y is a solution of the homogeneous equation (7.21)
[Section 3.2] satisfying the boundary conditions (7.20). Therefore if ¢—w is
not identically zero on 0<t<z, A is an eigenvalue of the homogeneous
problem (7.21), (7.20), contradicting the assumption.
Suppose / is not an eigenvalue of the homogeneous problem (7.21), (7.20);
A#n? (n=1, 2,...) (see Example 1, Section 7.2). Then we may rewrite the
expression (7.24) for the (unique) solution ¢ of (7.19), (7.20) as
t t
=| |= f()as-| ae £(s) ds
fies me]
(0)
The first two terms in this expression may be combined.
Exercise
2. Show that
sin /A(t—s)_sin\/At sin,/A(n—s) _ sin,/As sin,/A(x—1)

ve: Jasin, Jin Jasinfin
oes Nonhomogeneous Boundary-Value Problems 291
Therefore
t
sin,/As sin, /A(n— t)

o(t)= | s) ds
2 a sin,/An Be
x (7.26)
|sin ,/At sin,/A(n—3) Os
: ryA sin . At
It is convenient to define the function

_sin ./ds sin Jan —t)
if OS s<t
Ss sin Jan
G(t, s, 4)= (7.27)
sin fat sin ./2(n—s)
= — ift<s<nx
) ASIN Are
and to write the solution (7.26) as
6()=| G(t,s,4) f(s)ds, (0<t<n). (7.28)

e)
The function G(t, s, 4), defined for 0<t, s<a, and A¢n* (n=1, 2,...) is
called the Green’s function for the nonhomogeneous boundary-value
problem (7.19), (7.20). When this Green’s function, which does not depend on
the forcing function f, is known, the solution of (7.19), (7.20) is given by
(7.28) for every forcing function /.
Exercises
3. Show that G(s, t, 2)=G(t, s, A) provided A#n? (n=1, 2,...).
4. Show that G(t, s, A) is continuous for 0<?t, s<x and (0G/dt) (t, s, 4) is continuous
for t#s but
0G 0G
lim |Cork S, ao (s—h, S i).Le
h>O+
5. Show that G(t, s, 2), considered as a function of ¢ for each fixed s, satisfies the
differential equation (7.21) except for ¢=s.
6. Show that G(t, s, A), considered as a function of ¢ for each fixed s, satisfies the
boundary conditions y(0)=0, y(z)=0.
It is possible to show that the Green’s function is unique, and that it is
completely determined by the properties given in Exercises 3-6, that is, there
is no other function having these properties (see [1]; [7]).
Exercises
le Discuss the case 4=0 of the nonhomogeneous boundary-value problem (7.19),
(7.20), that is, find the solution of the problem y’=f(t), y(0)=0, y(x)=0 and
determine the Green’s function H(t, s) for this problem. Show, that H(t, s)=
lim, 49 G(¢, s, 4), where G(t, s, 2) is given by (7.27). [Hint: Recall that lim,5
(sinax/x)=a.]
. Find the solution of the nonhomogeneous boundary-value problem y” + Ay =coskt,

y(0)=0, y(x)=0, where & is a given positive constant. [Hint: Treat the cases
A#k? and A=k? separately.|
. Show that the Green’s function G(t,s,A) for the boundary-value problem
y’ +Av=f(0), v’(0)=0, vy’(x)
=0 (see Example 2, Section 7.2) is given by
cos,/At cos,/A(m—5)
if0<t<s
JA sin,/An
G(t, s, A=
cos,/As cos,/A(m—t)
ifs<t<nz.
ye sin RLAgs
10. Show that the solution ¢ of the differential equation y’=/(t) satisfying the
boundary conditions y(0)+y(1)=0, y’(0)+y’(1)=0 may be written $(d)=§ G(t, s)
I (s) ds, where
that is, G(t, s)= —4—4|r—s].
So far we have only studied problems with homogeneous boundary

conditions. To illustrate the treatment of inhomogeneous boundary con-
ditions consider the following problem.
Example 2. Given the boundary-value problem defined by the differential equation
y’+Ay=0 (7.21)
with the boundary conditions
y(0)=A, y(x)=B (7.29)
where A and B are given constants. Show that this problem may be reduced to one in
which the differential equation becomes inhomogeneous and the boundary conditions
become homogeneous, thereby reducing the problem to one which we have already
handled, in Example | above.
Let g(t) be any function with two continuous derivatives on 0<t<z such that
g(0)=A, —_g(n)=B;
1kes) Nonhomogeneous Boundary-Value Problems 293
for example, g()=A+(B—A)(t/x) is such a function. Let y(t)= z(t)+g(t); then

y (d)=2' ()+9' (0), »" (Q=2" (t)+g" (0), and
y"()+Ay()=2"(N+9"(0+AZM+9(0].
Thus the original boundary-value problem is replaced by the problem
z"+Az=—[g"(t)t+4g()], 2(0)=0, z(x)=0
which is of the form (7.19), (7.20) with f(t)= —[g"(t)+Ag(d] and has a solution given
in Example 1. With the particular choice of g suggested above, the problem becomes
t
stim — a)A+B-A) iF z(0)=0, z(x)=0
Tt
so that f(t) = —A[A+(B—A) (¢/z)].
Example 3. Solve the boundary problem for
y’ +Ay=hit) (7.30)
y(0)=A, y(x)=B
where / is a given function and 4 and B are given constants.
We consider two simpler problems. First, we consider the problem with the homo-
geneous differential equation solved in Example 2. We let #, (t) be the solution of the
boundary-value problem (7.21), (7.29). Next, let p(t) be the solution of the boundary-
value problem (7.30) with homogeneous boundary conditions
y(0)=0, y(z)=0 (7.20)
which was obtained in Example 1. Then if 4 is not an eigenvalue of the homo-

geneous boundary-value problem (7.21), (7.20), the unique solution of the given problem
is p(t)=o1()+ ¢2(2).
Exercises
11. Show that $(t)=¢, (¢)+¢2(t) is a solution of (7.30), (7.29).

12. Show that the solution ¢ is unique if 2#n? (n=1, 2,...).
13. Solve the boundary-value problem
— y’+2y=1, y0)=4, y()=©6

14. Solve the boundary-value problem
y’+2y=1, y(0)=0, y(x)=1.

7.4 SELF-ADJOINT BOUNDARY-VALUE PROBLEMS FOR THE

OPERATOR L(y)=—y"
In Section 7.2 we have studied several different homogeneous boundary-
value problems associated with the linear differential equation
L(y)=—y"=Ay. (7.31)-
We observed that these problems had some features in common, namely,
(i) all the eigenvalues of each problem were real and (ii) each problem had
an infinite sequence of eigenvalues tending to +00. In this section we
shall examine the question of how general these properties are, that is, for
what boundary conditions they hold.
Let u(t) and v(t) be given functions (possibly complex-valued) which-
are continuous and have continuous first and second derivatives on an
interval a<t<b. We shall use u(t) to denote the complex conjugate of u(t).
We use integration by parts twice to attempt to evaluate |? Lu(t) 0(t) dt,
obtaining
b b b
fewo v(t) dt= -|v v(t)dt =—[u'(t) o(k+ | v(t) dt
where W(u, v)(t) denotes the Wronskian of the functions uw and d (see
Section 3.3).
Now, suppose that u and v both satisfy a pair of given boundary condi-
tions at a and b which make the expression W
(u, 0) (¢)]?=0. For example,
we may require both uw and v to satisfy separated boundary conditions of
the form
ay(a)+By'(a)=0, — yy(b)+dy'(b)=0 (7.33)
where «a, B, y, 6 are given real constants with at least one of «, B and at least
one of y, 6 different from zero. If 640, then the first condition in (7.33)
becomes y’ (a)= —(a/B) y(a) and, if u and v both satisfy this condition,
W
(u, 0) (a)=u(a) v'(a)—u'(a) 0(a) = —— [u(a) 0(a)—u(a) d(a)]=0. (7.34)
Exercise
1. If B=0 but «40, show that (7.34) remains valid for functions u, v satisfying
ay(a)=0.
7.4 Self-Adjoint Boundary-Value Problems 295
A similar argument using the second condition in (7.33) shows that

W(u, v)(b)=0. Thus we see that if w and v both satisfy the separated
boundary conditions (7.33), then W(u, 0) (t)]?=0.
Another type of boundary condition which may be imposed is periodic
boundary conditions,
y(a)—y(b)=0, — y'(a)—y'(b)=0. (7.35)

Exercise
2. Show that if uw and v both satisfy the periodic boundary conditions (7.35), then
W (u, d) ()]2=0.
The vanishing of W(u, 0) (t)]® in (7.32) has important implications for
the boundary-value problem on the interval a<t<b defined by the dif-
ferential equation (7.31) and the boundary conditions which imply
Wu, 6) (Qe =0:
Definition. A boundary-value problem consisting of the differential equation
(7.31) together with a pair of boundary conditions at a and b having the
property that W(u, 0) (t)]?=0 for any functions u and v both satisfying these
boundary conditions is said to be self-adjoint.
For a self-adjoint boundary-value problem the relation (7.32) becomes
|[Lu(t) 6(t)—u(t) Lv(t)] dt=0 (7.36)
provided u and v satisfy the boundary conditions. In particular, we have

already shown that (7.31), (7.33) and (7.31), (7.35) are self-adjoint boundary-
value problems. As the following exercises show, there are other self-
adjoint boundary-value problems associated with the differential equation
(7.31).
Exercises
3. Show that the boundary-value problem y’+Ay=0, y(0)+ y(z)=0, y’(0)+y'(z)=0
is self-adjoint.
4. Determine a condition on the real constants «, 6, y, 6 for which the boundary-value
problem y” + Ay=0, y(z)—ay(0)— By’ (0) =0, »’ (x)— yy (0) — dy’ (0) =0 is self-adjoint.
If uw is an eigenfunction of a self-adjoint boundary-value problem associ-

ated with the differential equation (7.31) corresponding to an eigenvalue
A, then by taking v=uw in (7.36) and using Lu=du, we obtain
b
(A—A) {uou(t) dt=0. (7.37)

a
By the definition of an eigenfunction, u(t) a(t)=|u(t)|? is not identically

zero. Therefore u(t) a(t)dt>0, and we may divide (7.37) by
{tu(t) a(¢) dt to obtain A=/. Therefore the eigenvalue / must be real. Of
course, this argument does not prove that the self-adjoint boundary-value
problem associated with the differential equation (7.31) has any eigenvalues.
We have, however, shown by a separate argument that there is an
infinite sequence of real eigenvalues for separated boundary conditions
(see Exercises 15, 16, Section 7.2) and for periodic boundary conditions
(see Example 5, Section 7.2).
For real A, the differential equation (7.31) has real coefficients and, in
view of the discussion in Section 3.4, we may restrict our attention in the
future to real solutions of (7.31). Thus we shall assume that all functions
are real and dispense with the complex conjugates in equations (7.32)
and (7.36).
In Exercise 5, Section 7.2 (and similarly in subsequent exercises 8, 20,
21 for other boundary-value problems) we showed that the eigenfunctions
b, (t)=A, sinnt (n=1, 2,...) of the boundary-value problem (7.8), (7.9)
satisfy the relation
T
Joo f(t) dt=0, neém.,

0
This was done by a direct evaluation of the integral. This property of eigen-
functions is true for general self-adjoint boundary-value problems, and in
order to establish this, we must define the concept of orthogonality of
functions.
Definition. Two continuous real functions f(t), g(t) defined on a<t<b are
said to be orthogonal on a<t<b if and only if
b
|f(t) g(t) dt=0. (7.38)

a
Thus Exercise 5, Section 7.2 shows that the eigenfunctions @¢,(t)=A, sinnt
(n=1, 2,...) are orthogonal on 0<t<n.
Let u be a (real) eigenfunction corresponding to an eigenvalue J and let
v be a (real) eigenfunction corresponding to an eigenvalue 14 of a self-
adjoint boundary-value problem for the differential equation (7.31). Then,
using Lu=Au, Lv =v in (7.36), we obtain
b
(A—p) [uo vt) dt—0.

a
7.4 Self-Adjoint Boundary-Value Problems 297
Since 4#p, we must have

b
(A—) [uo v(t) dt=0.

a
which says that the eigenfunctions uw and v corresponding to different

eigenvalues of the boundary problem are orthogonal on the interval
Gas8.
We summarize what we have shown up to this point in the following
theorem.
Theorem 1. For a self-adjoint boundary-value problem consisting of the

Ly=—y"=)y (7231)
and self-adjoint boundary conditions such as
ay(a)+By'(a)=0, — yy(b)
+dy’(b)=0 (7.33)
or
y(a)—y(b)=0, — y'(a)—y'(b)=0 (7.35)
(but not restricted to these), all eigenvalues are real and eigenfunctions
corresponding to different eigenvalues are orthogonal on a<t<b.
We have already shown in equation (7.34) that if w and v are eigen-
functions of the boundary-value problem (7.31), (7.33), with separated
boundary conditions corresponding to the same eigenvalue, then W(u, v)
(a)=0 (and also W(u, v) (b)=0). Since u and v also satisfy the differential
equation (7.31), it follows that uw and v are linearly dependent on
a<t<b (by Theorems 2 and 3, Section 3.3). Therefore, a self-adjoint
boundary-value problem for the differential equation (7.31) with separated
boundary conditions cannot have two linearly independent eigenfunctions
corresponding to any eigenvalue. This is not true in the case of periodic
boundary conditions (7.35). For, as we have seen in Example 5, Section 7.2,
there may be two linearly independent eigenfunctions corresponding to the
same eigenvalue. You should note that while the problem (7.31), (7.35) with
periodic boundary conditions is self-adjoint (Exercise 2), W(u, v) (a) 40,
W
(u, 0) (6)40 for functions u and v which satisfy the boundary conditions
Ta):
oe 1 says nothing about whether a self-adjoint boundary-value
problem for the differential equation (7.31) actually has any eigenvalues.
The examples considered in Section 7.2 suggest that there is an infinite
sequence of eigenvalues, but this has not been proved in the completely
general case.
Theorem 2. A self-adjoint boundary-value problem consisting of the dif-

ferential equation (7.31) and any self-adjoint boundary conditions such as
(7.33) or (7.35) has an infinite sequence of eigenvalues tending to + ©.
Theorem 2 has been established in Exercises 15, 16, Section 7.2 for
separated boundary conditions (7.33) and in Example 5, Section 7.2 for
periodic boundary conditions (7.35). If you are interested in the proof in
the general case you should consult [1], Chapter 10; [7], Chapter 7; or
[21], Chapter 2.
Exercise
*5. Consider the boundary-value problem
y"+Ay=0
m1, y(0)+my2y'
(0) +111 y(a) +11 2y (x)=,
m1 y(0)+mz2y (0)+n21y(2) +N22y (x)=0,
where the m;; and n;; are real constants. Show that this boundary-value problem
is self-adjoint if and only if
M4 4M 22 — M2442 =Ny 1 N22 —NQ4Nj2-
7.55 STURM-LIOUVILLE PROBLEMS
We now consider the more general second-order equation
do(t) y" +4, (t) y +a(t) y=Ay (7.39)

where do, a,, and a, are continuous and a)#0 on an interval a<t<b,
with homogeneous boundary conditions at a and b. However we make a
preliminary simplification. If we multiply the equation (7.39) by
-4 oe (|ae i)
wf) on |) fo |280]
and use the identity
oo 20) ]ontan(| 284) agty( |,

we obtain
a
75 Sturm—Liouville Problems 299
This equation may be written in the form

Ly=—(p(t) y') +4 (t) y=Ar(t) y (7.40)
where
and t
= ye (|aC i)
are continuous on a<t<b, p(t)>0 on a<t0 on a<t0
ona<t<b.
We shall now study boundary-value problems consisting of the differ-
ential equation (7.40) with p(t)>0, r(t)>0 on a<t<b, and either separated
boundary conditions
ay(a)+By(a)=0, — yy(b)+dy'(b)=0 (7.41)

where «, f, y, 6 are given real constants, or periodic boundary conditions
y(a)—y(b)=0, —y'(a)—y'(b)
=0. (7.42)
In the case of periodic boundary conditions we shall also require p(a)=p(d).
Such a boundary-value problem is called a Sturm—Liouville problem.*
Exercises
1. Show that, with L as defined in (7.40),
|[Lu(t)6(0) —u(t)Eo(t)] dt= pls) Ww, 8) (DE

for functions u and v which are continuous and have continuous second derivatives
on a<t<b. [Hint: Imitate the proof leading up to equation (7.32).]
2. Show that if u and v satisfy either the separated boundary conditions (7.41) or the
periodic boundary conditions (7.42) (with the additional assumption p(a)=p(b) for
periodic boundary conditions), then p(t) W(u, &) ()\,=9.
From Exercises 1 and 2 we deduce immediately that a Sturm—Liouville
problem is self-adjoint, in the sense that
b
|[Lu(t) (t)—u(t) Lv(t)] de=0 (7.43)
+ After the German mathematician Sturm (1803-1855) and the French mathematician
Liouville (1809-1887), who, independently, were the first to formulate these problems.
for every pair of functions u and v which satisfy the boundary conditions
(7.41) or (7.42). The boundary conditions are used only to establish the
self-adjointness condition (7.43). Obviously, as in the special case of the
operator L(y)=-—y" considered in the preceding section, any boundary
conditions which lead to (7.43), even if they are not of the form (7.41) or
(7.42), define boundary-value problems for which our results are valid.
From (7.43) we deduce, exactly as in Theorem 1, Section 7.4 that all
eigenvalues of a Sturm—Liouville problem are real. The orthogonality of
eigenfunctions of (7.40), (7.41) or (7.40), (7.42) corresponding to different
eigenvalues now takes a slightly different form from that of Theorem 1,
Section 7.4. Namely, let uv be a (real) eigenfunction corresponding to an eigen-
value 4 and let v be a (real) eigenfunction corresponding to an eigenvalue
u#A. Then, since Lu=Ar(t) u and Lv=yr(t) v, (7.43) gives
b
(A—p) |r(t) u(t) v(t) dt=0.
Since 44 yu, we have

b
|r(t) u(t) v(t) dt=0. (7.44)
a
This motivates the following definition:

Definition. Functions u and v satisfying (7.44) are said to be orthogonal
on a<t<b with respect to the weight function r(t).
If r(t)=1, as it is in the differential equation y”+Ay=0 considered in
Section 7.2, orthogonality with respect to the weight function r(t) reduces
to the ordinary orthogonality defined earlier.
We have now developed the following analog of Theorem 1, Section 7.4
for the general Sturm—Liouville problem.
Theorem 1. For a Sturm—Liouville problem, consisting of the differential
equation
L(y)= —(P(4) vy) +4 (0) y=Ar(d) y (7.40)
where p, q, ¥ are continuous on a<t<b, and p(t)>0, r(t)>0 on a<t<b,
ay(a)+By(a)=0, — yy(b)+dy'(b)=0 (7.41)

where a, B, y, 6 are real constants or
y(a)—y(b)=0, — y'(a)—y'(b)=0 (7.42)

(with p(a)=p(b) for the boundary conditions (7.42)), all eigenvalues are
Ués Sturm—Liouville Problems 301
real and eigenfunctions corresponding to different eigenvalues are orthogonal

on a<t<b with respect to the weight function r(t).
Actually, we have proved Theorem | for any self-adjoint boundary-
value problem, not merely for the particular boundary conditions (7.41)
or(7.42).
Exercise
3. Show that there cannot be two linearly independent eigenfunctions corresponding

to the same eigenvalue for the differential equation (7.40) with separated boundary
conditions (7.41).
4. Determine a condition on the real constants a, 8, y, 6 for which the boundary-
value problem defined by the differential equation (7.40) and the boundary con-
ditions y(z)—ay(0)— fy’ (0)=0, y’(z)—yy(0)—dy’(0)=0 is self-adjoint. (Compare
Exercise 4, Section 7.4).
The question of existence of eigenvalues for the general Sturm—Liouville

problem is more difficult than for the particular equation y”+Ay=0. We
shall state an analog of Theorem 2, Section 7.4, but we refer you to a more
advanced source, such as [1], Chapter 10; [7], Chapter 7; or [21],
Chapter 2 for the proof.
Theorem 2. The Sturm—Liouville problem consisting of the differential
equation (7.40) and the boundary conditions (7.41) or (7.42) (with p(a)=p(b)
in the case of (7.42)) has an infinite sequence of eigenvalues tending to + ©.
Exercises
5. Test whether each of the following houndary-value problems is self-adjoint
a) ys +Ay=0, y(0)+2y’(0)+2y(n)=0, y(0)+y(x)—y'(x)=0

b) y’+y’+y+Ay=0, y(0)=0, y(1)=0
c) (ty’)' +2y+Ar?y=0, y(I)—y'(1)—y'(2)=9, v(I) + (2) +y'(2)=0
d) y’+Ay=0, »(0)—y(x)=9, y'(0) + y'(x)=9
e) y'+Ay=0, y(0)—y(1)=0
f) iy’—Ay=0, y(0)—y(1)=0
L(y)= —(p(t) vy) +4(t) y=Ar(t) y

where p, q, r are real continuous functions on a<¢0, r>0 and where
[p(t) r(t)]'? and p’(t) are continuous on a<t<b. Show that the changes of
are
variables (known as the Liouville transformation)
a
change this differential équation to
(*) — 244 (2) u= pu, O<z<n

where
ate) 21 —KKC)
Kate
f(2)=[p() ro)".
The form of the equation (*) is useful in studying the asymptotic expression for the
eigenvalues as no, for the Sturm—Liouville boundary-value problem; see [21],
jos 13s,
*7, Transform each of the following differential equations to the Liouville normal form
a) ?y"+t)'+(At? —p’) y=0
b) (1-2?) y’—2ty’+Ay=0
c) y’—2ty'+(A—1) y=0
*8. Find the eigenvalues and eigenfunctions of the boundary-value problem
y’+4r(t)y=0, y(0)=0, ~— p(x) =0

where
Are the eigenfunctions and their derivatives continuous at t=7/2? [Hint: Find the
solutions of the differential equation satisfying respectively y(0)=0 and y(z)=0,
and choose 4 to make the solutions match at t=7/2.]
*9. Find the eigenvalues and eigenfunctions of the boundary-value problem
y"+Ay=0, y(0)=0, »(5)=0, y(x)=0.
10. Show that the boundary-value problem

y’+dAy=0, = y'(0)+2y'(z)=0, — v(x) =0
is not self-adjoint and has no real eigenvalues. Show that there is an infinite sequence
of complex eigenvalues.
7.6 A Singular Boundary-Value Problem 303
11. Show that every complex number / is an eigenvalue of the nonself-adjoint boundary-
value problem
y’+Ay=0, — y(0)—y(n)=0, ——y'(0)

+y(n) =.
12. Show that the nonself-adjoint boundary-value problem
y'+dy=0, 2yQ@)—y(m)=0, —-2y'

+y(n)=0
(0)
has no eigenvalues, real or complex.
7.6 REMARKS ON A SINGULAR BOUNDARY-VALUE PROBLEM
In many applications we are faced with boundary-value problems of

Sturm—Liouville type for which the hypotheses that p, g, r are continuous
and that p(t)>0 and r(t)>0 for a<t<b made earlier are not satisfied. For
example, some problems involving cylindrical symmetry lead to a boundary-
value problem consisting of the differential equation
L(y)=—(ty) += y=dty (7.45)

z
where m is a known integer, together with boundary conditions at r=0

and t= 1. Comparing this with equation (7.40), we see that p(t)=1t, g(t)=m?7/t,
r(t)=t and the hypotheses of Theorems 1 and 2, Section 7.5 are certainly
not satisfied at t=0. Thus these results cannot be applied directly. Such
problems are called singular boundary-value problems.
Even thought the hypotheses of Theorems 1, 2, Section 7.5 are not
satisfied, we may consider the integral
|[Lu(t) 5(t)—u(t) Lv(t)] dt
which played an essential role in the theory, where the integral may be
improper. For the particular case of
L(y)=-(n
+= y
we obtain, after integrating by parts,
=u(1) o'(1)—u'(1) o(1)— lim e[u(e) 0’ (e)—u'(e) o(e)]

EOE
=W (u, v)(1)— it eW (u, 0) (é).
We therefore see that we may impose a homogeneous boundary condition

on u and v at t=1, such as u(1)=0, v(1)=0 or yu(1)+du'(1)=0,
yv(1)+6v’(1)=0 to ensure that W(u, v)(1)=0. If, in addition, we require
u(t), u'(t), v(t), v(t) to be bounded’ on any interval containing t=0, we
will have
and therefore
|[Lu(t) 5(t)—u(t) Lv(t)] dt=0 (7.46)
Since this relationship was all that was used in proving Theorem 1, Section
7.5, this result remains valid for the equation (7.45) and boundary condi-
tions of the form
y(1)=0, y(t), y(t) bounded at t=0 (7.47)
or
yy(1)+ dy’ (1)=0; y(t) and y(t) bounded at t=0. (7.48)
Observe that the condition that y(¢) and y’(t) be bounded at r=0 is of a
different type from the boundary conditions imposed previously. More-
over, this condition is inherent in the differential equation and arbitrary
values cannot be assigned at t=0. It is possible to prove an analog of Theorem
2, Section 7.5 for such problems, but we consider only the special case
(7.45), (7.47).
+ This requirement explains why in Chapter 6 we were interested in solutions analytic

at a singular point; for example, Jo(0)=1, J,(0)=0 (n=1, 2,...); see Section 6.9.
7.6 A Singular Boundary-Value Problem 305
Carrying out the differentiation in (7.45) and multiplying by ¢, we obtain

the equation
t?y"+ty’ +(At? —m?) y=0 (7.49)
which is a form of the Bessel equation that has been studied in Section 6.9.
We recall from Exercise 23 that the general solution of (7.49), with m an
integer or zero, is
(t)=CyIm(/At)
+02Km(/ At).
We recall also, see Theorem 3, Section 6.9, that the solution Kea t) is
unbounded on any interval containing the origin. Thus if f(t) is to be a
bounded solution of (7.45) on 0<t<1, c, is chosen zero and
$(t)=CrIm(a/At). (7.50)
Note that ¢’(t)=c, a, Tea t), and, from the power series representa-
tion 6.62, Section 6.9, ¢’(t) is also bounded on 0<t<1. Therefore (7.50)
satisfies the differential equation (7.45) and the second condition in (7.47).
It remains to determine / so that the first condition in (7.47) is satisfied, that
is, we wish to determine / so that
Im(x/A)=0. (7.51)
By the analog of Theorem 1, Section 7.5 for this problem, we know that A
must be real.
Instead of invoking an analog of Theorem 2, Section 7.5, we can give
an independent proof of the fact that Eq. (7.51) has an infinite sequence
of real positive solutions. This proof has already been outlined in Exercises
17, 18, Section 6.9. Let this sequence of solutions of(7.51) be uw? (n=1, 2,...).
(These solutions are tabulated in [11] or [12].) The corresponding eigen-
functions of the boundary-value problem (7.45), (7.47) are
Pn(t)= And m (Unt), (=e)
By the analog of Theorem 1, Section 7.5, these eigenfunctions are orthogonal

on 0<r<1 with respect to the weight function r(t)=+. This fact was also
established independently in Exercises 19, 20, Section 6.9.
The above considerations suggest how one may attack a singular
boundary-value problem. Whether this approach is successful depends
on how the coefficients in the differential equation behave near the singular
point. The interested reader is referred to [1], Chapter 10; [5], Chapters
9,10; or [19], Chapter 5 for other problems of this type.
Exercises
1. Show that the Legendre equation

(1—t7) y’—2ty' +Ay=0
can be written in the form Ly= —((1—77) y’)'=Ay.
2. Recall that r= +1 are regular singular points while t=0 is an ordinary point for the -
Legendre equation. Consider the singular boundary-value problem consisting of
the Legendre equation together with the condition that the solution be bounded on
—l<zt<1. Show that this problem has an infinite sequence of eigenvalues
A=n(n+1)(n=1, 2,...) and compute the corresponding eigenfunctions. [Hint:
See Example 2 and Exercises 9-17, Section 6.3.]
*3. Prove that the analog of Theorem 1, Section 5.5 for the boundary-value problem
in Exercise 2 is valid. [Hint: Consider
|[Lu(t) o(t)—u(t) Lo(t)] dt
and imitate the analysis given above for the Bessel equation.]
7.7 NONHOMOGENEOUS BOUNDARY-VALUE PROBLEMS AND

GREEN’S FUNCTION
In Section 7.3 we discussed nonhomogeneous boundary-value problems

for the operator L(y)=—y"”. We showed that there was a unique solution
provided the corresponding homogeneous problem had no solution and that
this solution could be written as an integral involving the Green’s function.
In this section, we shall consider the analogous problem of solving a non-
homogeneous boundary-value problem corresponding to a more general
Sturm—Liouville operator.
Let us begin by studying the boundary-value problem defined by the non-
homogeneous differential equation
L(y)= — (P(t) v) +4(t) y=Ar(t) y+ f(0) (7.52)

y(a)=0, y(b)=0. (7.53)
We shall assume throughout that p, g, r, and f are continuous on a<t0, r(t)>0 for a<t<b. We shall establish the following
result.
Theorem 1. The nonhomogeneous boundary problem (7.52), (7.53) has a

unique solution @ for a given value of 4, provided the corresponding homoge-
neous boundary-value problem
L(y)= —(p(t) ¥)+4(t) y=Ar(t) y (7.54)

HS Green’s Function 307
with the same boundary conditions (7.53) has no nontrivial solution (that is,
provided that A is not an eigenvalue of (7.54), (7.53)). This solution @ may be
written in the form
b
(=| G(t, s,A) f(s)ds, — (a<t<b) (7.55)

a
where the function G(t, s, 4), called the Green’s function for the problem (7.52),
(7.53), has the following properties.
i) G(t, s, A) is a continuous function of (t, s, 4) for a<t, s<b and for 4 not
an eigenvalue of (7.54), (7.53).
ii) (6G/ét) (t, s, A) is a continuous function of(t, s, A) for t#s and for i not
an eigenvalue of (7.54), (7.53); moreover
«eG 0G 1
lim E (s+h, s, A)-= (s—h, s, ’)|--
h>O+ p(s)
ili) G(t, s, A), considered as a function of t, satisfies the homogeneous dif-
ferential equation (7.54) for each t except t=s.
iv) G(t, s, A), considered as a function of t, satisfies the boundary conditions
(7.53) for each s, a<s<b.
v) G(t, s, A)=G(s, t, A) ifa<t, s<b and if d is real but not an eigenvalue of
(7.54), (7.53).
Proof. Let $,(t, 4), b2(t, A) be the solutions of the differential equation
(7.54) such that
b:(a, =1, $4(a,4)=0, $2(a,A)=0, $5(a,a)=1. (7.56)
Then ¢, and @, exist on the whole interval a<t<b by Theorem 1, Section
3.1 and are linearly independent on a<t<b by Theorem 1, Section 3.3.
Using Theorem 3, Section 3.3, we calculate their Wronskian
Wd, 2) (9 =exn/‘ |me a) rtd

t
The general solution of the nonhomogeneous differential equation (7.52)

may be obtained from the variation of constants formula (3.34), Section 3.7.
It has the form, obtained after several steps which are omitted here,
aN aerate nee |Peeve

p(a) J
—1(t, A) b2(s, A)] f(s) ds. (7.57)
In view of (7.56), the boundary condition ¢(a)=0 gives c, =0. The boundary
condition ¢(b)=0 gives
caalbs2) |[al 2)46,2) (6.2) als 2]£19)dd=0. (7.58)

The equation (7.58) may be solved for c, if $2(b, 2)40. This condition
$2(6, 4)4O0 merely says that 2 is not an eigenvalue of the homogeneous
boundary-value problem (7.54), (7.53), for if ¢2(b, 4)=0, then $(t, 4) is
an eigenfunction, while if @,(b, 4)40 there is no solution w of (7.54) with
w(a)=(b)=0. Thus the hypotheses of the theorem assure ,(b, 1)40,
and we may solve (7.58) for c,. Substituting the value for c, obtained from
(7.58) into (7.57), we obtain the solution of the nonhomogeneous boundary-
value problem (7.52), es)
1 [loa Be G.b ereeal yayas

wi |Aa 6-62
=a |Cea NECA ee ONG:
+g |etre 2)— $1(b,A)ba(s,A]f(s) ds

A)] f(s) ds
on [2 (t, A) #1 (s, A)— 4 (t, A) p2(s,
Combining the first and third terms, we obtain the form (7.55) as desired, with
1
Loi (t, 4) b2(s, 2) b2(b, A)— b(t, A) b2(s, 4) 1 (0, A)|
P(a) $2(b, A)
G(t, s, A)= ifs<t
1
P(a) 2(b, A)
Lb2(t, 4) b1(s, A) Pa(b, A)— a(t, 4) b2(s, A) $1 (b, A]
iis 0%
(7.59)
From this explicit representation (7.59) it is easy to verify that G(t, s, 4)
has the properties (i}{v) given in the statement of the theorem. The only
part of the theorem not yet proved is the uniqueness of the solution (7.55)
7.7 Green’s Function 309
of the nonhomogeneous problem (7.52), (7.53). Since the difference between

two solutions of this nonhomogeneous problem is a solution of the
homogeneous problem (7.54), (7.53), the assumption that / is not an eigen-
value implies that the difference between two solutions of the nonhomo-
geneous problem (7.52), (7.53) is identically zero, that is, the solution (7.55)
of the problem (7.52), (7.53) is unique. I
Exercises
1. Verify that the function G(¢, s, A) given by (7.59) has the properties (i}{v) listed
in the statement of Theorem 1.
2. Show that if 4 is an eigenvalue, the equation (7.58) becomes
b 10 a f2(s, 4) f(s)
p(a)
oads=0.
3. Show that when 4 is an eigenvalue, the boundary-value problem (7.52), (7.53) has
a solution if and only if fis orthogonal to the eigenfunction #2(t, 4) on a<t<b.
4. Solve the boundary-value problem (7.52), (7.53) when / is an eigenvalue and f is
orthogonal to the eigenfunction ¢,(t, 4) on a<t<b. Show also that the solution
is not unique in this case.
When one attempts to construct a Green’s function for a boundary-value

problem with more general boundary conditions than (7.53), the above
method is not the easiest approach. We therefore look at the question of
solving a nonhomogeneous boundary-value problem from a slightly different
point of view. We begin by constructing a function with the properties (i}{v)
of Theorem | and then we shall use this function to form the solution of
the nonhomogeneous boundary-value problem.
We consider the boundary-value problem defined by the nonhomogene-
ous differential equation
— (p(t) y’ +4(t) y=Ar(t) y+ f(0) (7.52)

and the general separated homogeneous boundary conditions
ay(0)+
By (0)=0, — yy(b)
+dy'(b)=0. (7.60)
We assume, as before, that p,qg,r, and f are continuous and p(r)>0,
r(t)>0 for a<t<b. Let w,(t, 4) be a solution, not identically zero, of
(7.52) which satisfies the boundary condition ay(a)+fy'(a)=0 and let
W(t, A) be a solution, not identically zero, of (7.52) which satisfies the other
boundary condition yy(b)+dy'(b)=0. If A is not an eigenvalue of the
corresponding homogeneous boundary-value problem (7.54), (7.60), then
the solutions W, (t, A), z(t, 4) are linearly independent on a<t<b.
Exercise
5. Show that if w,(¢, 4) and W(t, 4) are linearly dependent on a<1<8, then / is an
eigenvalue and w, (t, 4) (or W(t, A)) is a corresponding eigenfunction of the homo-
geneous boundary-value problem.
The most general function of (t, s, 4) which satisfies the differential equa-
tion (7.54) for a<t<s and the boundary condition at a has the form
G(t, s, A\=c,(s, 4) y(t, A), WEST eek
Similarly the most general function of (¢, s, A) which satisfies the differential
equation (7.54) for s<¢<b and the boundary condition at 6 has the form
G(t, s, A)=c2(s, A) Wo(t, A), s<t<b.
Thus the function G(t, s, 4), defined by
Git, Sah _fels

ae A)Wile,
Wa (tr),
asess
s<t<b
is the most general function having the properties (iii), (iv) of Theorem 1.
The symmetry property (v) implies
c1(s, A) Wi (t, A)=G(t, S, A)=G(s, l, A)= C(t, A) W2(s, A)
if t<s. This gives

Cy (s, A) C2 (t, A)
W2(s,4) Wilt, A)
and since the left side of this equation is independent of ¢ while the right
side is independent of s, both sides must be independent of both s and ¢.
Now, we may write c, (s, A)=kw(s, A), co(t, A)=ky, (t, A) and
_ Jkia(s,
A) Wilt, A), astss
AU tee s<t<b
It is easy to verify that this function G has the property (i) of continuity
in (t, s, A) for a<t,s<b if A is not an eigenvalue. The partial derivative
(6G/ét) (t, s, A) is continuous except for +45 and
Atay E (st+h, s, . (s—h, s, |
= lim [ks (5, A) Wals-+h, 2)—kha(s, DV (oh, a]

=k. (s, 4) Wo(s, A)—Wa(s, 4) Wi (s, a)]
=kW(W1, Wo) (s)

7.7 Green’s Function 311
where W(W,, W2) (s) is the Wronskian of w,, W, as defined in Section 3.3.
By Abel’s formula (see Theorem 3, Section 36),
W.Va)
(6)=Wi. Wa)(a)p(s)
2) a
Thus we may satisfy condition (ii) of Theorem 1, by choosing

k=1/p(a) W(W,, 2) (a). We now have the function
W2(s, A) Wi (t, A)
Ger a)S p(a) W(1, Wo) (a) (7.61)

Ww,(s Ay. ) Wo(t ; A) ; s<t<b
p(a) WW, W2) (a)

defined for a<t, s<b, A not an eigenvalue, which has the properties (i}{v)
of Theorem |. If we form the function
b
()= |Gls.) Fd
a
i W(t,A)
pla) Ws. Va) (af 54).
aw a RTATE
a, A). (7.62)
we may easily verify by direct substitution that ¢ is the unique solution

of the nonhomogeneous boundary-value problem (7.52), (7.60) if 2 is not an
eigenvalue of the homogeneous problem (7.54), (7.60).
Exercises
6. Show that $(t) given by (7.62) satisfies the differential equation (7.52) and the
boundary conditions (7.60) if A is not an eigenvalue.
7. Show that the problem (7.52), (7.60) has a unique solution if J is not an eigenvalue.
The Green’s function which we have constructed here reduces to the one
constructed by a different approach in Theorem | when we take S=6=0
in the boundary conditions (7.60). In fact, the following is true.
Theorem 2. There is only one Green’s function for the boundary-value

problem (7.52), (7.60) which is continuous for a<t,s<b, A not an eigenvalue,
and such that the unique solution of the problem (7.52), (7.60) is given by (7.62)
for every continuous function f(t).
Proof. We assume that G,(t,s, 4) and G,(t, s, 4) are two such Green’s
functions. Since 4 is not an eigenvalue of the homogeneous problem, the

problem (7.52), (7.60) has a unique solution ¢(t) and by uniqueness we must
have
|G.(E 5:4) f(s) is=| G,(t, s, A) f(s) ds
or equivalently
b
|[G,(t, s, A)—G,(t, s, A)] f(s) ds=0 (7.63)

a
for a<t<b, and for every continuous function f, We choose the particular
function
JS=G. (t, S; A)—Gyr(t, S, A)
for any fixed ¢ (a<t<b) and for / not an eigenvalue. Then (7.63) becomes
|[G, (t, S, A)—Gy(t, S, A)]? as=0
and since G,(t, s, 4)—G,(t, s, A) is continuous, this implies

Gi(GS, A)= Goss A) ax<s<b
for every t (a<t<b) and / not an eigenvalue. This proves the uniqueness
of the Green’s function. I
Exercises
8. Verify, by using (7.61), the previously constructed Green’s functions in Section 7.3,
namely, equation (7.27) for Example 1, and Exercises 9 and 10.
*9. Show that the Green’s function for the problem
—(ey + y=dty+£0)
Z
y(1)=0, y(t), and y’(¢) bounded at t=0 (see Section 7.6) is

1
Im(a/ As)eee le Ret
July Em 29) Im(./ As) Kin(/A)— Km(x/ At)
G(t, s, A= 1
Jal fipLlmls/20) Int/45) KB) —Inla/21) Kl23)Inl/B ts

10. Construct the Green’s function for each of the following boundary-value problems
and use it to solve the boundary-value problem.
a) y"-y=0, — y(0)=0, y(1)=0
ed Green’s Function 313
b) y’+4y=?, y(0)=0, v(1)=0

c) y"+av=f(), —_y(0)=0, »(1)=0
d) y’+Ay=sin 2x1, y(0)=0, y(1)=0
11. For each of the following boundary-value problems, determine at least approxi-
mately the eigenvalues (if any exist) and the corresponding eigenfunctions. Also
determine the Green’s function.
a) —y"+2y'’—y=Ay, y(0)=0, y(1)=0
b) —y"—4y’—4y=)y, y(0)=0, y’(L)=0
“ips Solve, if possible, each of the following boundary-value problems.
a) y’+Ay=0, = y(—1)=0, y(1)=1
b) y"+4y=0, —_y(1)—y'(0)
c) —y"—4)'-—4y=1y, (0)
d) y"+Ay=t, y(—1)=0, y()=1
e) y’+Ay=sint, y(1)—y’
(0)
f) y’+4y=cost, —_y(0)=0, y(x)=
g) y’'+4y=cost, —-y’(0)=0, y’(z)=0
h)y’+y=1, y(—L)=0,y'(D)
i) y’+y=1, y(0)=0, y(1)=0
CHAPTER 8
Existence Theory
In Chapter | and 2 we have constructed mathematical models for various

physical problems. These mathematical models were all ordinary differential
equations with initial conditions. We have seen that for such a model to serve
as a plausible representation of the physical problem, it should have exactly
one solution. In Theorem 1, Section 1.6 for first-order equations, Theorem 1,
Section 2.5 for second-order equations, and Theorem 2, Section 2.5 for equa-
tions of higher order, we have stated results of this type. In the first two
sections of this chapter we shall give proofs of these existence and uniqueness
theorems. In Section 8.5 we shall generalize these results to systems of
differential equations. In addition, we shall investigate in Section 8.3 the
question suggested in Section 1.6 of the size of the interval of existence of a
solution. In particular we shall prove an existence theorem for linear systems
(Theorem 1, Section 4.2) and for second-order linear differential equations
(Theorems | and 2, Section 3.1) in Section 8.6. As an application of several
ideas in existence theory, we consider the qualitative behavior of the motion
of a damped nonlinear pendulum in Section 8.4.
8.1 EXISTENCE OF SOLUTIONS
We begin with the problem of proving the existence of a solution @ of the

y=f(t,y) (8.1)
satisfying the initial condition
(to)=Yo (8.2)
on some interval containing f 9. This is called a Jocal problem, since it is
concerned only with existence of solutions near the initial point (tg, yo). To
treat this problem we make certain hypotheses on f in some rectangle
314
8.1 Existence of Solutions 315
centered at (to, Yo). This will mean that we can apply the local result of
this section at every point in a region D in which / satisfies these hypotheses.
Suppose fis continuous in D and that (fo, yo) is an arbitrary point of D.
The first step in our development is the observation that the initial-value
problem (8.1), (8.2) is equivalent to the problem of finding a continuous
function y(t), defined in some interval J containing fo, such that y(t) satisfies
the integral equation*
t
r=r0+ |f (s, y(s)) ds, tel. (8.3)

to
This equivalence is made precise as follows.

Lemma 1. If ¢ is a solution of the initial-value problem (8.1), (8.2) on an
interval I, then satisfies (8.3) on I. Conversely, if y(t) is a solution of (8.3)
on some interval J containing to, then y(t) satisfies (8.1) on J and also the initial
condition (8.2).
Proof. If ¢ is a solution of (8.1) on J satisfying (8.2), we have
? (=f(t, (0), tel

and integrating from f) to any ¢ on J we obtain
#0) #(00)= |£6.66) as

Imposing the initial condition (8.2) we see that @ satisfies (8.3).
Conversely, if y(t) is a continuous solution of (8.3), then by the continuity
of f, y(t) is differentiable. Thus by the fundamental theorem of calculus
applied to (8.3) we have that y(t) satisfies
y(j=f(t,yd), teJ
and putting t=fy in (8.3) we have y(to)=Yo. This completes the proof. I
Lemma | permits us to establish the existence of a solution of (8.1), (8.2)
by proving the existence of a solution of (8.3). This is important because
integrals are in general easier to estimate than derivatives.
Exercises
1. Determine the integral equation equivalent to the initial-value problem
y=tt+y*, = y(0)=1.
* Equation (8.3) is called’an integral equation (of Volterra type) because the unknown
function appears both under and outside the integral sign.
316 Existence Theory
Ze Prove that the initial-value problem
y’t+g(t,y)=0, yO)=yo, y(0)=Zo (8.4)

where g is continuous in some region D containing (0, yo), is equivalent to the
integral equation
t
woe +at-| Movouors (8.5)

0
[Hint: To show that if ¢ is a solution of(8.4) on / then ¢ satisfies (8.5) on J, integrate

(8.4) twice and use the fact that
ds hate.oe) dt= fe-nots ao a

[{farsoen icha-({f
To prove that a solution of (8.5) is a solution of (8.4), proceed as in the proof of

Lemma 8.1. But now you will need to use the formula
t t
d 0H
+ |a0 s\ds=H(é, i+ | Fo s) ds
0 0
which is easily proved by the chain rule, assuming only that H, 0H/0t are continuous
on some rectangle containing s=1=0. ]
. Construct an equivalent integral equation to the initial-value problem
y't+wy=glt,y), — y(0)=y'(0)=z0
assuming that g is continuous in a region D containing (0, yo) and where np>0
is a constant. [Hint: Assuming a solution @ of the differential equation on an
interval J which satisfies the initial conditions, apply the variation of constants
formula (Section 3.7). To prove the converse, proceed as in Exercise 2. Answer:
t
ts, sin u(t—s :
y(t)=yo cosut+— sinut+| wn g(s, y(s)) ds.]
(0)
. Prove that if @ is a solution of the integral equation

co
y(t)=e"+a |sin(t—s) a ds
t
(assuming the existence of the integral), then ¢ satisfies the differential equation
y" +(1+a/t?) y=0 (see Eq. (6.82), Section 6.11).
Returning to the main question of proving the existence of solutions of

(8.3) (and thereby of (8.1), (8.2)), we outline a plausible method of attacking
this problem. We start by using the constant function ¢9(t)=yo as an ap-
proximation to a solution. We substitute this approximation into the right

side of (8.3) and use the result
P1(t)=Yo + f(s, $o(s)) ds
as a next approximation to a solution. Then we substitute this approxima-

tion ¢,(t) into the right side of (8.3) to obtain what we hope is a still better
approximation @,(t), given by
#:()=v0+ |f(s, $1 (s)) ds
and we continue the process. Our goal is to find a function ¢ with the property
that when it is substituted in the right side of (8.3) the result is the same
function @. If we continue our approximation procedure, we may hope that
the sequence of functions {¢,(¢)}, called successive approximations, converges
to a limit function which has this property. Under suitable hypotheses this
is the case, and precisely this approach is used to prove the existence of a
solution of the integral equation (8.3).
Exercise
5. Construct the successive approximations to the solution @ of the differential

equation y’==—y which satisfies ¢(0)=2. Do these successive approximations
converge to a familiar function, and if so is this function a solution of the
problem?
We will consider the problem (8.1), (8.2) first with fand df/dy continuous
on a rectangle R= ({(t, y)| |t—tol<a, |y—yo|<b} centered at (tf, yo). We
assume that f and Of/dy are bounded on R (if, as often happens in practice,
the functions f and 0f/dy are continuous on R, the closure of R, defined
by R={(t, y)| |t—tol<a, |v—yo| <5}, then they are necessarily bounded
on R), that is, that there exist constants M>0, K>0 such that
Ue <M, ie(, nsx (8.6)

for all points (t, y) in R. If (t, y,) and (¢, yz) are two points in R, then by
the mean-value theorem, there exists a number 7 between y, and y, such that
f(t y2)-ft n= (t, n) (v2—1).

Since the point (t, 7) is also in R,

of < K, and we obtain
éy (t, n)
f(t, ya)—S(t vil SK ly2—-W1| (8.7)

valid whenever (t, y,) and (¢, yz) are in R.
Definition. A function f which satisfies an inequality of the form (8.7) for all
(t, y1), (t, ¥2) in a region D is said to satisfy a Lipschitz condition in D.
The above argument shows that if fand éf/éy are continuous on R, then
f satisfies a Lipschitz condition in R. It is possible for f to satisfy a Lipschitz
condition in a region without having a continuous partial derivative with
respect to y there, for example, f(t, y)=t|y| defined in any bounded region
containing (0, 0). In this chapter, we assume the continuity of 0f/¢y for sim-
plicity, but we could instead assume that f satisfies a Lipschitz condition
without substantial changes in the proofs.
Example 1. If
f(¢, y)=y'/? in the rectangle R= {(¢, y) ||¢|<1, |y| <2}, then
f does not
satisfy a Lipschitz condition in R.
To establish this, we only need to produce a suitable pair of points for which (8.7)
fails to hold with any constant K. Consider the points
(t, y,), (t, 0) with —1<¢<1, y,>0.

Then
fiy)-f 0) is ania
y,—90 Pere ar es
Now choosing y, >0 sufficiently small, it is clear that K=y,~7/? can be made larger
than any preassigned constant. Therefore, (8.7) fails to hold for any K.
Exercises
6. Compute a Lipschitz constant K as in (8.7) and then show that each ofthe following
functions f satisfy the Lipschitz condition in the regions indicated.
a) f(t =P +y%, (ty) |lls], yl <3}.
b) f(t, y)=p(t) cosy+q(t) siny, {(t, y)||t|< 100, |y|<0oo}, where p, qg are con-
tinuous functions on —100<r<100.
c) f(t, y)=t exp (—y?), {(t, y) ||t| <1, ly|< oo}.
7. Show that f(t, y)=7¢\y| satisfies a Lipschitz condition in the region {(¢, y) ||t|<1,
ly] <0}.
We have already indicated that we will use an approximation procedure
to establish the existence of solutions. Now, let us define the successive
approximations in the general case by the equations
Polt)=Yo
bj10=Vot |f(s, (3) ds, 7=0,1, 2,0. (8.8)

Before we can do anything with these successive approximations, we
must show that they are defined properly. This means that in order to define
j+1, On some interval /, we must first know that the point (s, ¢;(s)) remains
in the rectangle R for every s in J.
Lemma 2. Define « to be the smaller of the positive numbers a and b/M.
Then the successive approximations , given by (8.8) are defined on the interval
I given by |\t—to)|<«a, and on this interval,
16;(t)—yol<M|t—tol<b, j=0,1,2,.... (8.9)

Proof. The proof is by induction. It is obvious that ¢o(¢) is defined on J
and satisfies (8.9) with ;=0 on J. Now assume that for any j=n>1, @, is
defined and satisfies (8.9) on / (then, of course, the point (t, ,(¢)) remains in
R for te/). Then by (8.8), ¢,4, is defined on J. To complete the proof we
need to show that for re/, $,+,(¢) remains in R, or analytically that ¢,,4,
satisfies (8.9) with
7=n + 1. But from (8.8), the induction hypothesis, and (8.6)
we have
t
ae (9-yol-||Leona < |
|Ir ls, by(s)l ds
<M |t—to|<Mas<b.
This establishes the lemma. I
In order to explain the choice of « in Lemma 2, we observe that the

condition |f(t, y)| <M implies that a solution ¢ of (8.1), (8.2) cannot cross
the lines of slope M and — M through the initial point (to, yo). The relation
(8.9) established in the above lemma says that the successive approximations
# do not cross these lines either. The length of the interval / depends on where
these lines meet the rectangle R. If they meet the vertical sides of the
rectangle (Fig. 8.1), then we define « =a, while if they meet the top and bottom
(to, Yo + b)
(lo af a, Yo)
Figure 8.1
of the rectangle (Fig. 8.2), then we define «=5/M. In either case, all the
successive approximations remain in the triangles indicated in the figures.
We can now state and prove the fundamental local existence theorem.
(lo. Yo + b)
(fo + a, Yo)
mae
Figure 8.2
Theorem 1. Suppose f and @f/éy are continuous and bounded on the

rectangle R and satisfy the bounds (8.6). Then the successive approximations
$j, given by (8.8), converge (uniformly) on the interval I: \t—to|<«, to a solu-
tion @ of the differential equation (8.1) which satisfies the initial conditions
(8.2).
Proof. Lemma 2 shows that the successive approximations @; are defined
on the interval /. The next step is to estimate the difference between ¢;
and @;,,. We work on the interval tg <?<?f +a to the right of fo, but the
argument can easily be modified to give the result on the interval tg —a<t<fp.
We define
r(Q=ldj41(0-—¢,(O1. f=0, 1, 2,....
Then, using the definition (8.8) and the Lipschitz condition (8.7), we have
r(Q=lO;41(0-O(Ol= [76 6,0) 6,-(9)]

To
<|IF. d)-F 6. d,s) d

to
Tt
<K 16,09) 8)-6) ds
=K [ri-u6)San he ees (8.10)

The case j=0 is slightly different. We have, from (8.6)
rol)=10s()— dolAl=|]Fs bo(9)

; (8.11)
< |1/6.os) ds-<M(19.

From (8.10) and (8.11) we will Peres induction that
Sipe (=0,1, 2.2 sto st<te +o: (8.12)

(j+1)!
The case j=0 of (8.12) is already established. Assume that (8.12) is true for
j=p-—1 for some integer p>1; then (8.10) gives, on using the induction
hypothesis,
t t t
MR? Ft= o) oe 7.2

rlOSK [r-16)d<K [rp-a6)ds<k |
to to
= ME (Gt)
3 SEMare!
(p+1)!
which is (8.12) for j=p. This proves (8.12).
Exercise
8. Prove the analog of the inequality (8.12) for the interval tp -«<t< tp.
Combining (8.12) with the result of Exercise 8, we have
MKi|t—t./*! M[K|t—tol}/** M(Ka)*?

A= (jab) geek a KG a1)! K(j+1)!’
(=O 2stol on 9813)
It follows from (8.13) that the series }°%9 r,(¢) is dominated on the interval
|t—to|<a by the series of positive constants (M/K) )'%9(Ka)/**/(j+1)!,
which converges to (M/K) (e8*—1). By the comparison test (Section 6.2),
the series 7 9 r;(t) converges (in fact, uniformly) on |t—t|<a. In view of
the definition of the r,, this implies the absolute (and uniform) convergence
on sab pine
ae ane dim see t)—,(t)].
Since @,(¢ )+ oo [bm+1(—bm(2)], this also proves the con-
vergence of i Ue {,(t)} for oe t in the interval J to some func-
tion of t, which we call ¢(t). We will show that this function ¢(t) is continuous
and satisfies the integral equation (8.3) on J.
From the definition of $(¢),
therefore,
11) 4,(0= ¥ (nei) 00(0]

Now, from (8.13),
() <) Iba 1()— dy (0) a) int)

aye
(Ko)"** vn) oe (8.14)
n+1)!~ K (j+1)! = n!
_M (Ka)*! Ka
It is an elementary exercise to see that

Keyes
—0 as joo.
> (Sin
To prove the continuity of f(t) on J, let e>0 be given. We have
p(t+h)—o()=o(t+h)—$,(t+h)+¢,(t+h)—
¢6;(0+ 6;(0—¢ (0, and thus
IP(t+h)— P(t) <|h (E+ h)— Gj(t+ A) +1b)(¢+h)— 9, (0)

+|6;(t)—$(t)| S$2e,+|6;(t+h)— 6, (0)
by the above estimate. Choosing / sufficiently large and |h/| sufficiently small,
and using lim,..,, é; =0 and the continuity of the ¢,(¢), we can make
IP(t+h)—(i)<e.
We now wish to show that the limit function ¢(f) satisfies the integral
equation (8.3). We will do this by letting jo in the definition (8.8) of the
successive approximations and by showing that
im |£( 46) ds= [P0566) (8.15)

Once this is done the proof of the theorem is completed by applying
Lemma 1. To prove (8.15), we have, using the Lipschitz condition (8.7) and
the estimate (8.14),
fro =f(s,(9) 4ex|/0-» (s)] ds

gallate
ae
and this approaches zero as jo for every t on J. This establishes (8.15)

and completes the proof. I
Incidentally, we have also established the following useful consequence
in the course of the proof of (8.14).
Corollary. The error committed by stopping with the jth approximation @,(t)
satisfies the estimate
6-105
Gary
M (Kay Gs |
for every t on I.
Exercises
9. Construct the successive approximations to the solution ¢ of the differential equa-

tion y’=y which satisfies ¢(0)=
10. Construct the successive approximations to the solution ¢ of the problem in
Exercise 9, but using $ (t)=cost instead of ,(t)=1. Do these successive
approximations converge, and, if so, what is their limit?
11. Construct the successive approximations ¢p, ¢,, ¢2, ¢3 to the solution ¢ of the
differential equation y’=cosy which satisfies @(0)=0.
12. Consider the integral equation
er eee |neo (8.4

of Exercise 2 where g(t, y), (g/0y) (¢, y) are continuous on the rectangle R:|t|<a,
ly—yo|<b. (Thus, they are automatically bounded on R.) Let |g(t, y)| <M,
\(6g/Oy) (t, y)| <K for all (t, y)eR. Define
Pol(t)=Yo ;
$, (t)=Yo+Zot— |(t—s) g(s, 6, -1(s)) ds (eee als

0
Show that
a) the ¢, are well defined for |t|<«, where
b a
on
; eS min
j a)by
rel M= IZol
Z +M
b) Show that {¢,,} converges to a solution of the integral equation (8.4) on |t|<«.
This together with Exercise 2 establishes the existence of solutions of the initial-
value problem y”+g(¢, y)=0, y(0)=Yo, y’ (0) =Zo.
13. Consider the integral equation
ioe)
y(t)=e" +a |sin(t—s) osds

t
of Exercise 4. Define the successive approximations
fo(t)=0
$, (t)h=e" +a |sin(t—s)
a) Show by induction that

eae 1
ld, (t)oT Oye 1 (t)| S (nD I sepreecor Wilk A ccce

ae Wy I
Since ¢,(t)=o0(t)+(¢1 — 0)+-::+(¢,(t)— $,- 1(0) this shows that the @¢, are
well defined for 1 <t<oo, and {¢,} converge uniformly for | <t<oo to a con-
tinuous limit function @.
Sy Show that the limit function satisfies the integral equation.
ie)— Using
IPn(t)<
161 ()— Po(QI+- - +1Pn(t)—
bn-1 (Ol
and the above estimate for |¢,,(t)— ¢, — , (¢)|, show that the limit function satisfies
the estimate
lp(t)|<el", 1<t<oo.
This together with Exercise 4 supplies the missing steps which were assumed
in the justification of the asymptotic series solution of the equation
, a
in Section 6.11.
We have suggested in Section 1.6 that Theorem | is not the best possible
result of its type. Under the hypotheses of Theorem 1, we also have uniqueness
of solutions of (8.1), (8.2), as we shall prove in the following section. However,
we may have existence of solutions without uniqueness. In fact, the following
result is true.
Theorem 2. Suppose f is continuous on the rectangle R, and suppose
If(t, ¥)| <M for all points (t, y) in R. Let « be the smaller of the positive numbers
a and b/M. Then there is a solution $ of the differential equation (8.1) which
satisfies the initial condition (8.2) existing on the interval |\t —to|<«a.
We shall make no attempt to prove Theorem 2, as its proof is con-

siderably more difficult than the proof of Theorem 1. It cannot be proved
by the method of successive approximations, as the successive approxima-
tions may not converge under the hypotheses of Theorem 2. A proof may
be found in [7], Chapter 1, Theorem 1.2. The hypotheses of Theorem 2 do
not guarantee uniqueness, as in shown by the following example.
Example 2. We have previously (Section 1.6) discussed the equation y’=3y?/?, with
f(t, »)=3y7"?, (Af/ey) (¢, y)=2y7 1/3. Since Af/dy is not continuous for y=0, we cannot
apply Theorem | to deduce the existence of a solution of y’=/(t, y) through the point
(0, 0). By the method of Example | we see that f does not satisfy a Lipschitz condition
either. Since fis continuous in the whole (¢, y) plane, we can apply Theorem 2 to this
problem. In fact, there is an infinite number of solutions through (0, 0). For each constant
c>=0, the function ¢, defined by
0, —o<t<c
6.0=4 (op c<t<a
is a solution of y’=3y*/* through (0, 0). In addition, the identically zero function is a
solution of this initial-value problem.
Exercises
14. Do the successive approximations for solutions ¢ of y’ =3y?/? with ¢(0)=0 converge
to a solution?
15. Do the successive approximations for solutions of the problem considered in
Exercise 14, but using
0, O<t<l
solt)=| tr 1<t<oa
instead of ¢(t)=0, converge to a solution?
If fand of/dy are continuous on a region D, not necessarily a rectangle,

then given any point (¢o, vo) in D we can construct a rectangle R lying
entirely in D with center at (to, Yo). The hypotheses of Theorem | are then
satisfied in R and Theorem | gives us the existence of a solution o(t) of
y'=f(t, y) through the point (¢o, yo) on some interval about fo. In fact, this
solution may exist on a larger interval than the one constructed in the proof of
Theorem 1. It can be shown that the solution exists for all values of ¢ for which
the points (t, f(¢)) lie in D. This result will be established in Section 8.3.
In Section 1.4 when we solved the /inear equation of first order
y +p(t) y=4(t),
where the coefficients p, g are continuous on an interval J, we demonstrated
that the solution passing through the point (to, yo) exists not just near fo
but on the whole interval J. We did this by giving an explicit representation
for the solution and showing that this representation is valid for all ¢ in J.
The reader may now observe that this initial-value problem is equivalent to
the integral equation
t t
#0)= v0 |06 (6)ts—[a6) Ly

to to
The successive approximations
Pol(t)=Yo
t t
$110 =v0~| vl) 446) =| 46) ds J=0,1, 2.

are defined for all ¢ in J. The reader may also observe that the proof of
Theorem | applied to this problem gives the existence of a solution to the
first-order linear differential equation on the whole interval J. While no new
information is gained by this observation, it will be of considerable
importance when we discuss the existence of solutions of linear systems in
Section 8.6. For linear systems there is no explicit representation of solu-
tions, nevertheless, the method of successive approximations will yield a
global existence result, that is, the existence of a solution on the whole interval
on which the coefficients are continuous.
Exercises
16. Consider the differential equation y’=y, with the solution ¢()=e' satisfying the
initial condition $(0)=1. Letaet) be any polynomial in t and define
po(t)=
netic
1( Keil, DOsoc
a) Is limy.,, $,(t)=(t)? Prove your answer.

b) What general statement can you make concerning the initial “guess” do (t)
and the convergence of the successive approximations? What happens if the “‘1”
in the definition of ¢,(t) for k>1 is replaced by some other function, say p(t)?
17. Consider the differential equation
= f(t, y)
where f and of/dy are continuous in a region D in the (t, y) plane, and let
(fo, ¥o) be a point in D. Let G be a bounded subregion of D containing (to, yo) and
let G be the closure of G. Define
M=max |f(t, y)}.
G
Through (¢o, 9) construct the lines 4B and CD of slope M and — M respectively,

as shown in Fig. 8.3. Now, construct vertical lines H/, JK intersecting the ¢ axis
at a and f respectively, so that the isosceles triangles HOJ, JOK are contained in
G. Let o(t) be any continuous function defined on «<t<f such that the set of
points {(2, po(¢)) |«<1<f} is contained in G. Define
b,(=r0t | o(s)) ds, a<t<B
and generally
bl0=r0+ |f(s, dy—1(s)) ds, esi 2, noo8 0 SSS)0h
Figure 8.3
Prove that the successive approximations {¢,(¢)} converge (uniformly) to the

(unique) solution $(t) of y’=f(t, y) with (to)=yo on «<t<f, and estimate the
error |(t)—¢,,(¢)| on «<¢<f in terms of
K,= max {|¢o(t)+¢,
(t)|} for m=0
a<t1.
é |éy
[Hint: Follow the proof of Theorem 1.]
(Remark: The same result holds if f satisfies a Lipschitz condition.)
18. In the notation of Exercise 17, suppose that a function o(¢) has been found
satisfying the hypotheses of Exercise 17, and also constants k>0, 6=0 are known
such that
I: (t)—o(tl<Kit—tol’, a<t<f.
Show that
+1 K3 K3 + | a<t<f
PO Seach ees
where K, is as in Exercise 17. Can you generalize this result? [Hint: Assume an
estimate for |,,(t)— @m—1(¢)| on «<t<f and compute an estimate for |@(7)—@,,(¢)|
on a<t<f.]
19. Let #(t) be the solution of y’=?+y? on 0<r<1, with ¢(0)=0. Show that
1 1
60-(; ae ’)=O000IS5¢., 0st=1,
[Hint: In the notation of Exercise 18, let ¢9(¢)=2°/3, and compute @,(¢) and
|, (t)—o(t)|. Then apply the result of Exercise 18 to the differential equation
y'=0?+y? on the closed rectangle {(t,y)|0<r<1, |y|<A} for some suitably
chosen A>0. Such a choice is A =0.345, and this gives K, =0.690. ]
20. Let g(t) be a differentiable function on 0<t<z. Show that the differential
equation
my tal) y=Ay
has a solution @,(t, 4) on 0<t<z such that
sin,/At
+M,
$1 (t, A=
Ji
$4 (t, A)=cos \/At+M3
where |M,|<kK/A, \M|<K//2 on 0<t<nz for some constant K. [Hint: Consider
the integral equation
P(t, er sin ,/A(t—s) q(s) @(s, A) ds
and use successive approximations as in Exercise 13 to prove the existence of a

bounded solution ¢, (¢, A). Then estimate
M,=(1/,/A) |sin,/A(t—s) q(s) 6; (s, 4) ds.

0
Differentiate the integral equation to obtain the desired estimate for $j (t, A).]
21. Show that the differential equation considered in Exercise 20 has a solution
$2(t, A) on O<t<nz such that
par(t, A)=cos./At+M,
1
$3(t, A= a) sin Ji +5 cos Jit
|a) ds+M,
0
8.2 Uniqueness of Solutions 329
eral air :
where |M,|<K/,/A, |M4|< Rie on 0</<z for some constant K. [Hint: Proceed
as in Exercise 20, using an appropriate integral equation.|
(Remark :Exercises 20 and 21 above coupled with the Liouville transforma-

tion, Exercise 6, Section 7.5, are the starting points for the study of the
asymptotic behavior of eigenvalues and eigenfunctions of general Sturm—
Liouville boundary-value problems; see for example [21], pp. 110-114.)
8.2 UNIQUENESS OF SOLUTIONS
The second property which we wish a mathematical model for a physical

problem to exhibit is uniqueness. Thus our next goal is to prove that, given
suitable hypotheses, there is only one solution of the differential equation
(8.1) which satisfies the initial condition (8.2) (Section 8.1). We have seen
by examples that the assumption of continuity of f is not enough to
guarantee uniqueness. On the other hand, we have stated that the
hypotheses of Theorem 1, Section 8.1, are enough to guarantee uniqueness.
The main purpose of this section is to prove that assertion. The principal
tool in the proof is the Gronwall inequality.
Lemma 1. (Gronwall inequality): Let K be a nonnegative constant and
let f and g be continuous nonnegative functions on some interval a<t<f
which satisfy the inequality
fW<K+ |foes
fora<t<f. Then
1 ()<K exp(|als) 4s)

fora<t<f. ;
Proof. Let U(t)=K+/J;, f(s) g(s) ds, and observe that U(«)= K. By hypothesis
f(t)<U(t), and by the fundamental theorem of integral calculus,
U'(t)=f(t) g(t) for «<t<f. Because g(t)>0, we obtain U'(t)<U(t) g(t) for
a<t<f. We multiply this inequality by exp(—J‘ g(s) ds) and apply the
identity
t
U’(t) exp (—fas is)U(t) g(t) exp (-|a(s) i)

to obtain
Integration from « to ¢f gives

t
U(t) exp (-[aits)U(a)<0,
or, since f(t)< U(t) and U(a)=K,

t
F()< USK exp(|o() 4), a<t<B,
which is the desired inequality. I
Exercises
1. Let K, and K, be positive constants and let f be a continuous nonnegative

function on an interval «<t<f which satisfies the inequality
t
HSKi+K: |ods
Show that f(t)<K, exp(K, (t—a)).

2. Find all continuous nonnegative functions f on 0<t<1 such that f(t)<Jo f(s) ds
for0<r<l.
3. Let f(t) be a continuous nonnegative function satisfying the inequality
t
f(t)}<K,+e(t-—a4)+K, {ro ds,
on an interval «<1<f, where ¢, K,, K, are given positive constants. Show that
é
f(t)<K, exp(Ky(t— a) += [exp K,(t—a«)—1].
2
[Hint: Consider U(t)=K, +e(t—a)+K3 [i f(s) ds.]

4. Find all continuous, not necessarily differentiable, functions f(t) such that
t
Lf (t)]? =|f(s)ds — (t=0).
5. Let r(t) bea nonnegative continuous function which satisfies the inequality
t
r(t)<a+K [ro [r(s)]? ds

0
8.2 Uniqueness of Solutions 331
for r>0, where a>0 and k>0 are constants, and f is a nonnegative continuous
function for t>0. Show that ;
r()<al| ~Ka| f(s) is|
for all t>0 for which fj f(s) a ae

Sa Define R(t)=a+K fo f(s) [r(s)]* ds, so that R(0)=a, R(t)>r(d, and
R'(t)=Kf(d) [r(O]? < Kf (t) [R(tae Proceed to separate variables and integrate,
just as in the proof of the Gronwall inequality. |
Now we are ready to prove the desired uniqueness result.

Theorem 1. Suppose f and @f/¢y are continuous and bounded on the rectangle
R defined by the inequalities |t—to|<a, |y—Yo|<b. Then there exists at
most one solution of (8.1) satisfying the initial condition (8.2).
We recall that under the hypotheses of Theorem 1 we have already
established the existence of at least one solution @ of (8.1), (8.2) existing
on the interval |t—t)|<«a, where « is defined in Theorem 1, Section 8.1. We
also recall that the hypotheses of Theorem | imply the inequality (8.7) (the
Lipschitz condition).
Proof of Theorem 1. Suppose that @, and @, are two solutions of (8.1),
(8.2) which both exist on some common interval J containing f9. Since, by
Lemma 1, Section 8.1, every solution of (8.1), (8.2) also satisfies the integral
equation (8.3), we have t
bl=vo+| £6 2(s)) ds
to
for every t in J. Subtracting these two equations, we obtain
= [EF ba(9)P6506)
Taking absolute values and using (8.7), we have
t
lA]. #00)-F160
<K fos (s)| ds (for teJ)

where |(6f/@y) (t, y)|<K for all (t, y)eR. Taking first the case r>/) and
then t<f, the Gronwall inequality now implies for both cases that
|f2(t)—,(t)| <0. Since |,(t)—¢,(d| is nonnegative, we have |,(t)
— $,(t)|=0 for all tin J, or $3 (t)=¢, (0) for tin J. Thus there cannot be two
distinct solutions of (8.1), (8.2) on J, and this proves uniqueness. |
It is not necessary to assume as much as continuity of éf/éy to ensure
uniqueness. It is clear from the proof of Theorem | that the Lipschitz
condition (8.7), which follows automatically from the continuity of éf/¢y,
could be used in the hypothesis of Theorem | instead of the continuity of
cf/éy without changing the proof. It is possible to prove uniqueness of
solutions under considerably weaker hypotheses, but in most problems
Theorem | is applicable and such more refined results are not needed.
Exercises
6. State and prove a uniqueness theorem for solutions ofthe initial-value problem
y’+g(t, y)=0 y(0)=Yo y (0)=Zo

where g is a given function defined on a rectangle R: |t|<a, |y—yo|<0b. Refer to
Exercise 12, Section 7.1. [Hint: Under appropriate hypothesis if @, and ¢, are both
solutions of the initial-value problem existing on some interval |t| <a, we have
o;(t)=Vo + Zot -|¢—s) g(s, ;(s)) ds (k=1, 2)
for |t|<qa; subtract the two equations corresponding to k=1 and k=2, then use
the Lipschitz condition (8.7), obtaining
t t
(s(0)~ #0) sk |(—s)$1(s)—2(s)| ds<ko [i600 ds

0 0
and finally appeal to the Gronwall inequality.]
7. Prove the Osgood uniqueness theorem: Suppose that the function /(¢, y) satisfies the
condition
Lf (t, v2) - F(t yi) SA(ly2 — yl)
for every pair of points (¢, y,), (¢, y2) in a region D. Suppose that the function
h(u) is continuous for 0<u<a for some «>0, that h(u)>0, and that
i du
im |——-=0o
e-0+ | A(u)
Then through each point (fo, Yo) in D there is at most one solution of the equation
y'=f(t, »). [Hint: Suppose $, and ¢, are two solutions with ¢,(to)=2(to)=Yo.
8.3 Continuation of Solutions 333
Define w(t)=$2(t)—¢,(¢) and suppose y/(‘}#0. Then show that |w'(t)|<h(\W(0))

<2h(|W(z)|). Suppose y(t,)40, for some ¢;>%o and let u(t) be the solution of
u'=2h(u) satisfying the initial condition u(t,)=|y(t,)|. Show that u(t) is strictly
positive for fo<¢<t,. Show that w’(¢,)<u'(t,) and therefore y/(t)>u(t) on some
interval to the left of 7;. Then show that w(d)>u(t) for t)<¢<t, and obtain a
contradiction. ]
8. Show that the functions h(u)=Ku* («>1), h(u)=Ku log |u|. h(u)= Ku log |u| log
|log|u| |,and so on, satisfy the Osgood condition of Exercise 8 above. (The case
h(u)= Ku is, of course, the Lipschitz condition.)
9. Let f(t, y) and g(t, y) be continuous and satisfy a Lipschitz condition with respect
to y ina region D. Suppose |f(t, y)—g(t, y)|<e in D for some e>0. Let ,(t) be a
solution of y’=f(t, y) and let ¢3(t) be a solution of y’=g(t, y) such that
|2(to)— 1 (to)|<0 for some to and some 6>0. Show that for all ¢ for which
$,(t) and ,(t) both exist,
(b2(0)—4,(01 deKhl 4S (ekFo1)

where K is the Lipschitz constant. [Hint: Show that |,(d)—¢,(A|<6
+ |,[K|¢2(s)—¢,(s)| +e] ds and complete the argument by a slight generalization
of the Gronwall inequality (Exercise 3).]
8.3 CONTINUATION OF SOLUTIONS
The existence theorems of Section 8.1 state that under suitable hypotheses
there is a solution of a differential equation that exists on some, possibly
small, interval. The question to be studied in this section is whether this
solution in fact exists on a larger interval.
Example 1. Consider the first-order initial-value problem
y=y, o()=1
whose solution ¢ can be written explicitly as ¢(t)=1/(1 —2) (see Example 1, Section 1.3).
Clearly the solution exists on — 00 <t<1. Suppose we try to determine the interval of
validity of the solution as given by Theorem 1, Section 8.1. Here R is the rectangle
R={(t, y)||t|<a, |y—1|<5}, and « and M of Theorem 1, Section 8.1, are given by
M=max,(y”)=(1+5)*, «=min(a, b/M). The largest value of the positive number
b/M=b/(1 +5) is 1/4.
Exercise
1. Using calculus or otherwise, show that
b 1
a ee
Thus, no matter how a and b are chosen, we have «<1/4. If we take a> 1/4 and
b=1, so that b/(1+b)?=1/4, we obtain «=1/4, and Theorem 1, Section 8.1, gives
the existence of a solution for |t|< 1/4.
Suppose that f and Of/dy are continuous in a given region D in the

(t, y) plane. Let (to, yo) be a given point in D. Consider the initial-value prob-
lem
y=f(ty) (8.16)
V(to)=Yo- (8.17)
By the definition of a region, there exist numbers a>0, b>0 such that the
rectangle R= {(t, y) ||t—to|<a, |y—Yo| <5} centered at (to, yo) is contained
in D. By Theorem 1, Section 8.1, there exists a number «>0 and a unique
solution @ of (8.16), (8.17) existing on the interval |t—1o|<«a (see Fig. 8.4).
A
Figure 8.4
Now consider the point (t9+, (to+a)) as a new initial point. Since
this point is in D (it cannot be on the boundary of D because the closed
rectangle R is contained in D), there exist numbers a, >0, b, >0 such that
the rectangle R, ={(t, y)| |t—(to+a)|<ay, ly—G(to+a)|<b,} centered at
(to +, P(t o+«)) is contained in D (see Fig. 8.5). Consider the differential
equation (8.16) subject to the initial condition
V(to +a) =P(to +a). (8.18)

By Theorem 1, Section 8.1, there exists a number «,>0 and a unique
solution of the initial-value problem (8.16), (8.18) existing on the interval
|t—(to +.x)| <a, (see Fig. 8.5). Because of the uniqueness
O(t)\=W(t), (tota—a,<t<to+a). (8.19)

We may now define the function ¢ on the interval tg —a<1t< to tata, by
: eae to —A<t<tyta
W(t), totacst<totat+a,
Figure 8.5
original initial-value problem (8.16), (8.17) on the interval tp-—a<t<ty

+a+«a,, a larger interval than the interval |t—t)|<a on which we origi-
nally constructed the solution ¢. This solution ¢ is called the continuation
to the right of the solution ¢ to the interval tg-—a<t<typ+a+q,.
Exercises
2. Show that the function ¢ satisfies the integral equation (8.3) on the interval
fo —AStSlo t+.
3. Consider the solution ¢ of Example | above, which has been shown to exist on the
interval —1<+<+. Consider now the continuation of ¢ to the right obtained by
finding the solution wy through the point (%, $). Show that on any rectangle
R,={(t, y) |(t-Zlsa, |y—$|<b}, M=maxg, y?=($+5)?. Deduce, similarly to
Example 1, that «, =7;. This now gives existence on —3<1<j¥5.
4. Continue, similarly to what was done in Exercise 3, the solution ¢ of Example |
to the left of the point (—4, 4).
In general, the solution ¢ may be continued to the left from the point
(to —«, (tp —«)) in a similar manner, as suggested by Exercise 4.
Exercise
5. Formulate the continuation to the left.
In Example | and Exercises 3 and 4 we have indicated how to continue

the solution ¢ of y'=y’, where $(0)=1, to the right and to the left of the
interval |t|<a, with «=min(a, b/M)=1/4. How far can the continuation be
extended? To answer this question, and the corresponding general question
for the initial-value problem (8.16), (8.17), we need the following auxiliary
result.
Lemma 1. Suppose f and of/0y are continuous in a domain D and suppose
|f| is bounded in D. Let $ be a solution of
Y(to)=Yo (8.17)
existing on some interval y<t<06. Then lim,.;— (t) and lim,.,, 6(¢) exist.
Proof. Let t, and t, be any two points on the interval y<¢<6 with t, <fp.
Then, since ¢ satisfies the integral equation (8.3),
Subtraction gives
and the assumption |/(t, y)| <M for (t, y)eD now gives
lb(t2)— (ti) SM (to—t4). (8.19)
Since the right side of (8.19) tends to zero as f, and f, both tend to 6 from
below, the Cauchy convergence criterion shows that ¢(t) tends to a limit
as ¢ tends to 6 from below. We obtain the proof that #(¢) tends to a limit as
t tends to y from above in an analogous way by letting ¢, and r, tend to y
from above in (8.19). 1
In view of Lemma 1, we can define $(6)=lim,..;- $(t), 6(y)=lim,.,+ $(d
and we have the solution ¢ defined on the closed interval y<t<6. The con-
tinuation process described above can be repeated provided that the graph
of the solution remains in a region in whichfand df/éy are continuous, and

this continuation process is carried out by means of rectangles remaining
in D, precisely as R and R, were used above.
Theorem 1. Suppose that f and Of/0y are continuous in a given region D
and suppose f is bounded in D. Let (to, Vo) be a given point of D. Then the
unique solution $ of y'=f(t, y) passing through (to, Yo) can be continued until
its graph meets the boundary of D.
Remark. The most important hypothesis in Theorem | is that fis bounded
in D.
Proof. Suppose the solution ¢ cannot be extended up to the boundary of
D, but can be extended to the right only to an interval tp <t<?. By Lemma 1,
o(t)=lim,.,- (2) exists. If (7, f(2)) isa boundary point of D, we are finished.
If not, there is a rectangle centered at (7, p(#)) lying in D. But now we can
continue the solution ¢ to the right of 7?by the method given above, resulting
in a contradiction. This shows that ¢ can be extended up to the boundary of
D. An analogous argument shows that ¢ can also be extended to the left up
to the boundary of D, and this completes the proof. I
The reader should note that in the proof of Theorem 1, it is possible
that f= +00. If f= +00, then it is possible that lim,.;_ |b(t)|=0. If t<c,
since f is assumed to be bounded on D, we must have lim,_,;_ |(t)|<0.
In many problems the hypothesis made in Theorem | that fis bounded
in D is not satisfied if D is taken as the largest possible domain of definition
of f. For the equation y’=y? considered in Example 1, if D is taken to be the
whole (t, y) plane, the function f(t, y)=y? is not bounded. We may handle
this problem as follows.
Example 2. Consider the solution ¢ of the equation y’=y? such that $(0)=1. Here D
is the whole plane. If we consider the equation on the region D4= {(t, y) |—00<t<o,
ly|<A} instead of D, then f(t, y)=y? is bounded on D, by A’, and Theorem | is
applicable. It shows that the solution ¢(t)=1/(1—2) can be continued for those values
of tsuch that |p(¢)|=1/(1—1)<A, or — 00 <t<1—1/A. However, since this is valid for
arbitrary A >0, p(t) can be continued to the interval — 00 <t<1, but not to the interval
—o<f<l.
In Exercises 3 and 4 above, the solution @ was “built up” on a union

(or sum) of contiguous (adjacent) closed intervals. If such a continuation
can be accomplished in a finite number of steps, the resulting solution will
be valid on a closed interval. If, however, an infinite number of steps is
required, as in fact the case for the equation in Example 2 above, then the
resulting solution may be valid only on an open interval. This can happen
because a union of an infinite number of closed intervals need not be closed.
For example, the union of the intervals 1—(1/n)<t<1—1/(n+1), (n=1,
2,2.) isthe wterval O<7<1-
In applications, a very common situation is the case where the domain

D is the whole (t, y) plane. We start with an arbitrary infinite strip
D,z={(t, y) |—0<t<oo, |y|<A} containing the initial point (to, yo) and
employ the argument of Example 2. By Theorem 1, the solution @ can be
continued to the boundary of D,. Clearly, either this gives the existence of
the solution on — 00 <t<oo or there exists a ? such that lim,.; 6(¢)=
A; in
the second alternative we enlarge the strip D, by increasing A. It may happen
that by choosing A finite but large enough we obtain existence of the
solution ¢ on — co <t<oo. In this case, ¢ is also bounded on — co <t<oo,
since |¢(t)|< A on —0«<t<oo. If, however, the solution ¢ reaches the por-
tion |y|=A of the boundary of D, for every A >0, there are two possibilities.
Either ¢ exists for — 00 <t<oo but is not bounded, or there exists a finite
number T such that lim,_.,, |¢(¢)| = 00. We can now summarize.
Corollary to Theorem 1. If D is the entire (t, y) plane and iff and Of/éy are
continuous in D, then the solution $ of y'=f(t, y) can be continued uniquely
in both directions for as long as |\p(t)| remains finite.
A similar result holds for the case where D is the infinite strip a<t<b,
ly|<oo, and f, 0f/dy are continuous in D. Note that this corollary does not
say that if D is the entire (t, y) space, then every solution can be extended to
the interval — 00 <t<oo.
Exercises
6. Show that the solution ¢ of y’= —y? with ¢(1)=1 exists for 0<t<co but cannot
be continued to the left beyond r=0.
7. Show that no solution other than ¢(t)=0 of the equation y’ = — y* can be extended
to the interval —0<t<oo.
8. Formulate, as another corollary to Theorem 1, the result on continuation if f is
defined on the infinite strip D={(t, y) |a<t<b, |y|< oo}.
There are special cases where we can prove a global result directly from
successive approximations. We have already pointed this out for the first-
order linear differential equation in Section 8.1. We shall see in Section 8.6
that, for linear systems in general, global existence can be proved directly
from successive approximations. Other cases are illustrated by the following
exercises.
Exercises
9. Let f be a continuous scalar function defined on the domain D= {(t, y ||t—to|<a,

\y|< oo}, where « is an arbitrary positive number. Suppose there exists a constant
K=K(a)>0, depending on « but not on y, such that
Ks
? (t, y)| <K(q).
Show that the successive approximations

Pol(t)=Yo
bje=v0+| £6 j(s)) ds (i=Oh lee)
converge to a (unique) solution of the initial-value problem y’=f(t, vy), v(to)=Yo.-

Also show that
M aK
IPOs lvol += (e =: |t—tol<a,
where K= K(x) and M=max,,_,,) <. |f(¢, Yo)|. [Hint: The trick now is to get around
the fact that f(/, y) itself is not necessarily bounded on D, even though f(t, yo) is
bounded.
i) Induction easily shows that each ¢;(j=0, 1, 2,...) is well defined for |t—to|<a.
Then
IP: ()— Pol(tl< ||r6 Yo)| ds|<M| t—tol.
ii) Show by induction that

MK? |t—t,}/*?
lpj+i1(t)-O,(NlS G+)!
iii) Deduce the uniform convergence of {@;} to a limit function @ on |t—t|<a.

Then for |t—tfo|<«a,
j
| ;(t)—Yol= Da py (t)— Px-1( a |p, (t)— bx-1 (0
© M Ki- ei M Ka
y pe (Cx 1
al)
1M
from which we obtain the desired bound for |(¢)| on taking the limit.
iv) Show that the limit function ¢(t) is a solution of the initial-value problem.|
10. Let f(¢, y) be a continuous function on the whole (f, y) plane. Suppose that df/dy
is also continuous and suppose that for every «>0,
of
_ (t, y) <K=K(q), lt] <a, ly|<oo.
dy
Show that for every given (to, Vo) the equation y’=/(t, y) has a unique solution
#(t)on — 0 <t<oo such that (to) =o. [Hint: By Exercise 9, |p (¢)| is bounded on
every interval |f—f)|<a; now apply the corollary to Theorem 1. ]
11. Show that the equation
y' =p(t) cosy+q(t) siny,

where p and q are continuous functions on — oo <t<oo, has a (unique) solution
g on —c© <t<oo satisfying the arbitrary initial condition $(to)=Vo. [Hint: Use
Exercise 10.]
8.4 THE NONLINEAR SIMPLE PENDULUM
In Section 2.2 we constructed a mathematical model for the motion of the

simple pendulum released from rest. We assumed that the only external
forces were a constant gravitational force of gravity and a force of friction
proportional to velocity, and we obtained an initial-value problem of the
form 5
y” +2ay' +k? siny=0, y(0)=yo, y' (0)=0, (8.20)
where a and k are given positive constants, and yo is a given constant which
may be assumed nonnegative without loss of generality. Since this initial-
value problem cannot be solved explicitly, we considered only small oscilla-
tions. That is, we assumed that y remains small and that sin y may be replaced
by y. Thus, instead of (8.20) we considered the initial-value problem
x" +2ax'+k*x=0, x(O)=yo, x (0)=0 (8.21)
which, being linear with constant coefficients, is easily solved by the methods
of Section 3.4. Its solution is
M(=ean E cos(k? —a?)1/? bez ae

aye sin(k? —a’)'/? |
if the system is lightly damped, that is, if a<k. If we let

k? —a*=w7>0, A=ky,/w, 6=arctana/o,
then we can rewrite this solution as
x(t)= Ae“ cos(wt—64). (8.22)
From this expression, we can derive physical predictions about the motion
of the pendulum, for example that it oscillates with period 2z/m and
amplitude which decreases exponentially from A to zero.
However, there is no reason beyond intuition to believe that the solution
of the simplified problem (8.21) is a good approximation to the solution
of the original problem (8.20). In addition, even (8.20) is only an approxima-
tion to the real problem. The derivation of (8.20) neglects all forces other
than gravity and friction, and involves various other simplifying assump-
tions. A precise mathematical model for the motion of the pendulum would
be an initial-value problem of the form
y+ 2ay' +k? siny=f(t, y, y), viOl= vo. ¥ (O)=0), -9(8.23)
where the term f(t, y, y’) includes all effects neglected in the derivation of
(8.20).* It is reasonable to assume that f(t, y, y’) is small, for otherwise (8.20)
could not be a useful model for the motion of the pendulum. However, it
* When there is no friction (a=0), a discussion of the nonlinear pendulum is more

difficult. The interested reader is referred to [2], Chapter 6, for one aspect of this topic.
8.4 The Nonlinear Simple Pendulum 341
cannot be assumed that an explicit expression for f(t, y, y’) is known. Thus
it does not make sense to ask for an explicit representation for the solution
of (8.23). The proper question to ask is whether for a given class of
functionsfwhich are suitably small, the problem (8.23) has a solution which
is described approximately by the solution (8.22) of the simplified problem
(8.21). In other words, the mathematical problem is one of existence of
solutions for a class of differential equations and of the qualitative behavior
of these solutions, rather than one of explicit solution.
Let y(t) be the solution of the initial-value problem
y"+2ay'+k’y=p(ty,y), yO)=yo, y(0)=0. (8.24)
We will assume that there exists a constant M>0 such that
Ip(t, y, y’IS<M (yl? +1y'?) (8.25)

for t>0 and for sufficiently small |y| and |y’|. We remark that (8.23), which
would appear to be the proper problem to study, is a special case of (8.24)
with p(t, y, vy)=f(t, y, y’)—k? (sin y—y). Since |sin y—y|<|y|?* for small |y|,
the function f obeys the same condition (8.25) as does p.
We wish to show that as ft 00, y(t) behaves in the same way as does x(‘),
which is given by (8.22). From (8.22) we see that |x(t)| is no greater than a
constant multiple of exp(—azt). We begin by making the changes of variable
x =exp(—art) win (8.21) and y=exp(—az) v in (8.24). Then (8.21) becomes
u"+wu=0, u(0)=yo, wu (0)=ayo (8.26)
with solution u(t)=A cos (wt—6). Also (8.24) becomes
v” +@*v=exp(at) p[t, exp(—at) v, (exp(—at) v)']
=a(t,,0), )=Yo, ¥ O)=ayo. ey
Using (8.25) and (exp(—azt) v)’=exp(—at) (v'—av), we see that
la(t, v, v’)|=lexp(at) p(t, exp(—at) v, (exp(—art) vJ’)|
<exp(at) M (\exp(—at) v|? + |exp(—at) (v' —av)|’) (8.28)
<exp(—at) M (|u|? + |v’ — av’)
< Lexp(—at) (\v|* +|v'|’)
for some constant L>0.
Theorem 1. If (8.28) is satisfied, then there exists a constant B>0 such that
every solution v(t) of the initial-value problem (8.27) with |yo| sufficiently small
satisfies
lv(t)|<B, lv’ (t)|<B (8.29)
for all t=0.
Proof. We consider (8.27) as if it were a linear nonhomogeneous problem
to which we apply the variation of constants formula (see Section 3.7,
especially Exercise 5). Since the unknown function v appears in the non-
homogeneous term in (8.27), we obtain an integral equation for v(t) rather

than an explicit formula; this integral equation is
t
v(t)=u ()+- [sn[w(t—s)] q[s, v(s), v'(s)] ds. (8.30)
The equation (8.30) may be differentiated to yield

t
v' (t)=u'(t)+ |vs [w(t—s)] g[s, v(s), v'(s)] ds. (8.31)

0
When we solve (8.26), we see from the explicit solution that |u(t)|<c,
\u'(t)| <<c for t>0, where c is a constant which can be made arbitrarily small
by making |yo| sufficiently small. In fact, |u(2)| <A =kyo/a, |u'(t)| <w@A=ky®,
and we can take c=kyy if m>1, and c=kyo/a if o<1. Now we use (8.28)
to estimate in (8.30) and (8.31), and we obtain
\v(t)| Set {u[s, v(s), v'(s)]| ds
<e+— |Lexp(—as (\v(s)|?

+ |v’(s)|?) ds,
® (8.32)
|v’ (t)| se+ {la
[s, v(s), v’(s))| ds
<e+| Lexp(—as) (|v(s) (s)I*) ds.

+e’|?
0
We let K=L if o> 1, K=L/o if w<1; and we add the two inequalities in
(8.32) to obtain
lv(t)| + |v’ (| <2c+2K [exo—as) (|v(s)|? +|v’(s)|?) ds

yt (8.33)
<2c+2K [exo— as) (|v(s)|+|v’(s)|)? ds.

0
We now let r(t)=|v(¢)|+ |v’ (|, so ona
10.s2e+2K [esas
—as) [r(s)]? ds,
10)
and apply the generalization of the Gronwall inequality derived in Exercise 5,

Section 8.2 to obtain
re) 2e/|1—4Kc |exo(-a9 is
Since Jo exp(—as) ds<|¥ exp(—as) ds=1/a, we have r(t)<2ac/(a—4Ko),

provided that c<a/4K. We now let B=2ac/(a—4Kc), and then r(t)=|v(t)|
+|v'(t)| 0 provided that |y9| is chosen small enough that c<a/4K.
This completes the proof of Theorem 1. !
By returning to the variation of constants formulas (8.30) and (8.31), we
may now obtain more precise information about the behavior of v(t) and
v(t) for large ¢.
Theorem 2. If (8.28) is satisfid, then for every solution v(t) of (8.27) with |yo|
sufficiently small, there exist constants A, 6, C>0 such that
v(t)=A cos(wt—5)+h(t),
where
|h(t)|
<C exp(—at), |h’(t)| <C exp(—at)
for t=0.
Proof. We rewrite (8.30) and (8.31) as
v(t)=u(t) te {si[w(t—s)] q[s, v(s), v'(s)] ds

0
-= |sin[@(t—s)] q[s, v(s), v'(s)] ds (8.34)
v’(t)=u' (t)+ |cos[w(t—s)] q[s, v(s), v'(s)] ds

0
(oo)
- |cos[@(t—s)] q[s, v(s), v'(s)] ds (8.35)
respectively. These formulas are valid if the infinite integrals converge.

Since
lsin [w(t —s)] q[s, v(s), v'(s)]| <la(s, v(s), v'(s))|
< Lexp(—as) (\v(9? + '(9))< 2LB? exp(—a9,
using (8.28) and (8.29), and since |g exp(—as) ds converges, the integrals
in (8.34) converge. A similar argument proves the convergence of the
integrals in (8.35). Now we define

(o.@)
a=ug+4| sin[@(t—s)] a[s, v(s), v'(s)] ds

0
=u(t)+sinwt o [a[s, v(s), v’(s)] cosas a|
—cosat 3 {a[s, v(s), v’(s)] sinws ash (8.36)
Then a(t) is a linear combination of solutions of the linear homogeneous

differential equation u”+m?u=0 and is therefore itself a solution. Now
(8.34) becomes ©
o(t)=a(%)— |sin reneer ha fitets\ a (s\| a8
and it is easy to verify from the definition of w(t) that (8.35) becomes
00
v’ (t)=a' (t)— |cos[w(t—s)] q[s, v(s), v’(s)] ds.
We may write these relations as v(t)=a(t)+A(t), v'(Q)=a'

(1) +h'(d), where
|h(t)| << C exp(—at), |h'(t)|< C exp(—ar) for all t>0 with some constant C>0.
Since every solution a(t) of u”+@?u=0 has the form a(t)=A cos(ct—4),
the proof of Theorem 2 is complete. I
We now return to the original variables x =exp(— at) wand y=exp(—at) v

and apply Theorem 2. This yields the following result.
Theorem 3. Let y(t) be the solution of the initial-value problem
y"+2ay'+k*y=p(ty,y), yO)=y0, y'(0)=0 (8.24)

with a<k and yo>0. Let w* =k*—a’. Suppose that p(t, y, y’) satisfies
P(t, ys VIS M (ly? + yl?) (8.25)
for t=0 and for |y\, |y'| sufficiently small. Then there exist constants
A, 0, C>0 such that
y(t)=A exp(—at) [cos(wt—4)+h(t)] (8.37)

where
|h(t)| <C exp(—at), |h’(t)| <C exp(—at)
for all t>0.
Exercise
1. Show that the constant C in Theorem 3 may be taken as the larger of 2LB?/aw and
2LB?/a.
The reader will note that the formula (8.37) says that every solution of
the “‘correct” problem (8.24) with |yo| sufficiently small behaves like some
solution of the idealized linear equation (8.21). It does not, however, say
that the solution of (8.24) behaves like the solution of the initial-value
problem (8.21) which is composed of the idealized differential equation
together with the same initial conditions as those in (8.24). The solution of
(8.21) is x(t)=A exp(—ar) cos (wt—9), while the solution of(8.24) is approxi-
mated by 4 exp(—az) cos(wt—5). To justify the use of the simpler problem
(8.21) in place of the true problem (8.24), we would have to show that A is
close to A and 6 is close to 6.
From the relation
A cos(wt—6)= yo cosat 490 sin cot
we see that i
A?=yé+(ayo/w)? and 6=arc tana/o.
If we define
oO
1 1 :
d, =5 |q[s, v(s), v'(s)] cos@s ds, d= in |qLs, v(s), v'(s)] sinas ds,
(0) 0)
then from (8.37) we see that

A cos(wt —4)=A cos(wt—6)+d, sinwt +d, cosat
=(2 4 )snort od) cosat.

7)
From this we obtain
ze ayo é 2ayYo
Ha (P44 + (Yo + dz)? = A? +—~— dy +i2 + 2daV+d
o2
co) (8.38)
. dyo/m+d,
tan —————_
o0=a rc..
Yo+d,
Using (8.28) and (8.29) we have
|ts. v(s
16,v (ids <2 8? |exp(—as)ds = 2B /a.
0 0
Thus
ldj|=2LB-/a, [Id |= 2LB*a.
Since B can be made arbitrarily small by making |yo| sufficiently small, |d,|
and |d,| can be made arbitrarily small. Now we see from (8.38) that
A*— A? and tand—tanéd (and hence 6—6) can be made arbitrarily small.
Thus we have obtained the following result to complete our study of the
damped pendulum.
Corollary to Theorem 3. The constants A and 6 in the expression (8.37) for
the solution of the nonlinear problem (8.24) are approximated by the amplitude
A and phase angle 6 obtained by solving the linear problem (8.21). These
approximations may be made as close as desired by taking |yo| sufficiently small.
8.5 EXISTENCE THEORY FOR SYSTEMS OF FIRST-ORDER

EQUATIONS AND HIGHER-ORDER EQUATIONS
We now wish to consider the extension of the results of Sections 8.1—8.3 to

systems of first-order equations of the form
y ={(¢ y) (8.39)
where y and f are vectors with n components and where ¢ is a scalar. Such
systems were introduced in Chapter 4, and the reader is referred to
Section 4.5 for the necessary background, basic definitions, and notation.
Weare first of all concerned with the problem of existence and uniqueness of
solutions of (8.39). Such a result was stated as Theorem 1, Section 4.5, but
without proof. Before proceeding, we remind the reader that because of the
equivalence of single scalar differential equations of nth order and systems
of first-order equations (established in Section 4.5), every result which is
established for (8.39) has an immediate interpretation for an nth-order
scalar equation, or for that matter a system of such equations of any order.
In what follows D will represent a region in (n+ 1) dimensions (recall
that a region is a set in (t, y) space with the property that given any point
(to, n) in D, the “box” B= {(¢, y) ||t—to|<a, ly—n|<b} lies in D, provided
a and b are sufficiently small). Let f be continuously differentiable with
respect to ¢ and the components of y at all points of D, which we denote by
fe C,(D), and suppose that there exists a constant K>0 such that the norms
of of/dy; satisfy
Lal(t,y)|<k (j=1,..., n) (8.40)
for all (t, y) in D. As usually happens in practice, such an inequality is
automatically satisfied if feC,, for example, on the closed “box”
B={(t, y) ||t—to|<a, ly—n|<b} for some fixed positive numbers a and 5,
or on any closed, bounded set in (n+ 1)-dimensional space. It then follows
that for any points (ft, y), (¢, z) in D we have the inequality
f(t, y)—f(t, z)|< Kly—z|. (8.41)
This may be seen by applying the mean-value theorem to each variable
8.5 First-Order and Higher-Order Equations 347
separately and then using (8.40), or by the following, more elegant argument.
Define the function G by
G(c)=f(t, z+0(y—z)) (0<o<1)

and consider f(t, y)—f(¢, z). We have
f(t, y)—f(t, z)=G(1)—G(0)= |eo do.
By the chain rule, letting f,,=0f/0y; (j=1,..., n), we have
G'(o)=f,,(, z+ o(y —2)) (vi—z,) +f, ( 2+.(y —2)) (v2 —22)

ee thy (Zt O(¥—Z)) (Va)
Using the bound (8.40), we find
1
fle yf 2).<|1G (0) dos K yl hy.aal 4

0
+1¥n—2nl}
=K ly—z
which is (8.41). A function f satisfying an inequality of the form (8.41) for
any points (f, y), (t, z) in D is said to satisfy a Lipschitz condition in D with
Lipschitz constant K. A function f satisfying (8.41) need, of course, not be
of the class C, and all the remarks made in the simple case of scalar functions
apply here.
We begin with the problem of existence.
Theorem 1. Let f and of/dy;(j=1,..., n) be continuous and bounded on the

box B={(t, y)| |t—to|<a, |y—n|<b}, where a and b are positive numbers,
and satisfying the bounds
If(t, y)/< ct
of(t,y) 2k Gin) (8.42)
pov
j
for (t, y) in B. Let « be the smaller of the numbers a and b/M and define the
successive approximations
do(t)=n
(8.43)
o)=n+| peer a
Then the sequence {;} of successive approximations converges (uniformly) on
the interval |t—to|<a to a solution (t) of (8.39) which satisfies the initial
condition ®(to)=N.
We remark that if the hypothesis holds on the closed box B=

{(t, y) ||t—tol <a, ly—n| <b}, then (8.47) is automatically satisfied and
the conclusion holds on |t—t9|<«. The choice of « is suggested by the same
reasoning as in Theorem 1, Section 8.1.
The proof is step by step and line by line the same as the proof of Theorem
1, Section 8.1, with the scalars f, @, v9 replaced by the vectors f, @, n, and
in obvious places the absolute value is replaced by the norm. We remind the
reader that the first step is to establish the equivalence of the initial-value
problem with the integral equation
t
vio=ne |f(s, b(s)) ds (8.44)

to
and then work with (8.44). This is the analog of Lemma 1, Section 8.1.
Exercises
1. Give a detailed proof of Theorem 1. (The reader is urged to carry out this proof
with care, in order to appreciate the usefulness of introducing vectors.)
2. By writing the scalar equation y” =g(t, y, y’,..., y"~) as a system of n first-
order equations (see Section 6.2), apply Theorem | to deduce an existence
theorem for this scalar equation.
3. Given the system y,=y2+y241
Ya=Yi-y2—1.
Let y=(3!) and let B be the “box” {(¢, y)||t|<1, ly|<2}. Determine the bounds
M, K in (8.42) for f and ¢f/dy, for this case. Determine « of Theorem 1. Compute the
first three successive approximations of the solution (7) satisfying the initial con-
dition )(0)=0, @=($').
We remark that Theorem 2, Section 8.1 also has an analog which is
easy to state. So far as uniqueness of solution of the system (8.39) is
concerned we have the following analog of Theorem 1, Section 8.2.
Theorem 2. Let the hypothesis of Theorem 1 be satisfied in the box B with
center at (to, n). Then there exists at most one solution ® of (8.39) satisfying
the initial condition (to) =.
Exercise
4. Prove Theorem 2. [Hint: Let be another solution of the same initial-value

problem and suppose the solution @ and exist on some common interval J.
Then 6 and th both satisfy the integral equation (8.44). Using this and (8.41) show
that t
6) —W(OL<IK [eo-vo Aivalifonvandde inh:

to
8.6 Linear Systems 349
The conclusion now follows as in the scalar case (Theorem 1, Section 8.2), by using
the Gronwall inequality (Lemma 1, Section 8.2).]
The situation with regard to continuation of solutions for systems is

also analogous to the situation in the scalar case. By initiating the proofs of
Theorem 1, Section 8.3, and its corollary, replacing absolute values by
norms, we obtain the following results.
Theorem 3. Suppose that f and of/éy;(j=1,..., ) are continuous in a given
region D and suppose f is bounded in D. Let (to, n) be a given point of D.
Then the unique solution @ of y'=f(t, y) passing through (to, n) can be
continued until its graph meets the boundary of D.
Corollary. If D is the entire (t, y) space and if f and of/éy,; (j=1,..., n) are
continuous in D, then the solution of y' =f(t, y) can be continued uniquely in
both directions for as long as \(t)| remains finite.
Just as in the scalar case, if f is not bounded in D we may consider smaller
regions in which f is bounded, apply Theorem 3, and then let these smaller
regions tend to D to obtain a continuation theorem (see Example 2,
Section 8.3).
Exercises
5. Obtain the analog for systems of Exercise 17, Section 8.1.

8.6 LINEAR SYSTEMS
The analog of Theorems 1, 2, 3, Section 8.5, for existence and uniqueness

of solutions of /inear systems is simpler than the general case already treated.
Since the successive approximations can be defined on the same interval as
that on which the coefficients are continuous, and since they converge on this
interval, we obtain a global existence theorem directly, without having to
study continuation of solutions. In addition, uniqueness of solutions follows
from continuity of the coefficients. We may sum up the basic existence and
uniqueness theory for linear systems in a single theorem.
Theorem 1. Let A(t) be an nxn matrix and g(t) a (column) vector with
n components which are both continuous on some interval I. Then for every
to in I and for every constant vector n, the initial-value problem
y=A(thy+g(t), y(to)=n (8.45)

has a unique solution existing on the interval I.
Proof. By the analog for systems of Lemma 1, Section 8.1, the initial-value
problem (8.45) is equivalent to the integral equation
=n+ jr
[A(s) g(s)] ds. (8.46)
Let J be any closed finite subinterval of J containing fo. (If J is closed, take
J=I.) We proceed as in Theorem 1, Section 8.1, or Theorem 1, Section 8.5,
by defining the successive approximations
Ho(t)=
(m=1, 2,...). (8.47)

6,..0=n+| [A(s) ,(s)+g(s)] ds,
First we prove by induction that each vector function @,(f) is well defined
and continuous on the interval J. Clearly $9 (¢) is well defined and continuous
on J. Suppose ,(¢) is well defined and continuous on J. Then A(s) o,(s)
+ g(s) is a continuous vector on J, and its integral form f, to ¢ is a continuous
vector on J; hence from (8.47), @;(¢) is a continuous vector on J. Thus, by
induction, each @,(¢) is well defined and continuous on J.
Since J is a closed finite interval on which 4 and g are continuous,
|A(2)| and |g(z)| are bounded on J. We let K and L be constants such that
|A(sI<K, = [g(s)ISL, — seJ
and we define M=K|n|+ZL. Now, proceeding exactly as in Theorem 1,
Section 8.1, we can establish the analog of (8.12), namely
—MKi(t—to"
ldj+1(t)-O,(Hi < (8.48)
Gea)!
for j=0, 1, 2,..., teJ. The only difference is that here (8.48) is valid for all
tin J. The remainder of the proof of existence of a solution (t)=lim,.. ., 0;(2)
on J proceeds exactly as in Theorem 1, Section 8.1. Since J is an arbitrary
closed finite subinterval of J, this solution exists on the whole interval J.
To show that the solution (ft) is unique, we suppose that W(t) is another
solution of the initial-value (8.45) on J. Then
w=n4 |(4096 )+-8(s)] ds
=1+ |[40 ¥O)+860] as

8.6 Linear Systems 351
Subtraction gives
t
$Hiee |AS eeveas

to
We take norms and use |A(s)|<K to obtain
0) WO) SK |16()—WOi ds
The Gronwall inequality (Lemma 1, Section 8.2) now gives |@(t)—w(a)| <0.
Since |O(‘)—(z)| is nonnegative, we have |(t)—wW()|=0, or b(t)=W(v).
Thus there cannot be two distinct solutions of the initial-value problem
(8.45), and the proof of the theorem is now complete. |
Exercises
1. By converting to the equivalent system, prove that if p, g, fare continuous on an

interval /, then the initial-value problem consisting of the second-order equation
y"+p(t) y'+4(t) y=f(0) and the initial conditions y(to)=yo, =y'(to)=Zo, where
t él, has a unique solution existing on the interval /.
2. State and prove an existence and uniqueness theorem for linear differential
equations of arbitrary order n.
When we study Laplace transforms in Chapter 10, we will need to

know that if the nonhomogeneous term g(t) is of exponential growth at
infinity, that is, if there exist constants M@>0 and c such that
lg(t)| <M exp(ct)

for all sufficiently large t, then all solutions of the system y’=Ay+g(?)
are of exponential growth at infinity. We may now prove the necessary result.
Theorem 2. Consider the initial-value problem
y=Ayt+g(t), y(O=n (8.49)

where A is a constant nxn matrix, and g(t) is continuous for 0<t<0o and
of exponential growth at infinity. Then the solution o(t) of (8.49) (which exists
for 0<t<oo and is unique by Theorem 1) and its derivative '(t) are of
exponential growth at infinity.
Proof. The initial-value problem (8.49) is equivalent to the integral
equation
t t
vig=nt |Ay(s) s+] 26 ds. (8.50)

0 0)
By hypothesis, there exist constants M>0 and c, and a time T such that
\g(t)| <M exp(ct), Ve (8.51)
We may assume c>0, since increasing c increases the right-hand side of
(8.51) and does not affect the truth of the inequality. We may rewrite (8.50) as
FE ii t t
y(t)=n+ |Ay(s) ds+ |g(s)ds+ |Ay(s) ds+ {209ds. (8.52)
We have shown in Theorem | that y(t) is bounded on the closed interval

0<t<T. Using this fact and the continuity of g(), we see that there exists
a constant K such that . ie
n+ |ay0 is+| 209 ds|<K. (8.53)
Now, taking norms in (8.52) and using (8.51) and (8.53), we obtain
ly(t\|<K + [lay ds | a0)ds

T T
sK+| lA ly (s)| dv | exp(cs) ds

fe a
M
—K +||A| |y(s)| ds anges [exp(ct)—exp(cT)]
T
M
< K+— exptcr+| |A| ly(s)| ds.
vi
Multiplying by exp(—cr),
we have ;
M
ly(t)|exp(—ct) <(Kexp(—et)+*)4-| Lal ly(s)] exp(—ct) ds
<(iot-0)e) aoten(-cda
T
a F (8.54)
since exp(—ct)<exp(—cs) for t>¥. Since c>0, there exists a constant L such
that Ke(—ct)+ M/c<L for t>T. Thus (8.54) becomes
t
ly («)| exp(—ct)< +a ly(s)| exp(—cs) ds.

T
8.7 Dependence on Initial Conditions 353
We let r(t)= y(t) |exp(—ct)>0, so that
r(t)<L +f r(s) ds,

T
and apply the Gronwall inequality (Lemma 1, Section 8.2) to obtain

r(t)<Lexp[|A|(t—T)],
or
ly(t)| < Lexp(—|A| T) exp[(|A|+c) ¢],

which shows that |y(/)| has exponential growth at infinity. Since y’(t)=
Ay(‘)+8(0),
ly’ ()|SIAI ly(2)| +g ()| S|A] L exp(—|A| T) exp[(|A|+c) t]+M exp(ct)
<(|A| Lexp(—|A| T)+M) exp[(|A|+c) ¢],
and thus |y’(t)| also has exponential growth at infinity. |
Exercises
3. State and prove a theorem of the nature of Theorem 2 concerning solutions of a

linear equation of order n, Y+p,y""+---+p,y=f(t), where p,,..., P, are con-
stants and /(f) is of exponential growth at infinity.
8.7 DEPENDENCE ON INITIAL CONDITIONS
A solution (ft) of the system of differential equations
y=f(¢, y) (8.55)
passing through the point (¢), n) depends not only on ¢, but also on the
initial point (to, n). When we wish to emphasize this dependence, we write
the solution as (t, fo, n). We will show that under suitable hypotheses
depends continuously on the initial values and in fact that is a continuous
function of the “‘triple” (t, to, n) (actually (n+2) variables). As in the
previous sections of this chapter, we make no attempt to prove the most
refined result of this type.
Theorem 1. Suppose £ and 0f/0y; (j=1,...,) are continuous and bounded
in a given region D, with
If(t, yI<M,
of(t, y)
a 2k ee (8.56)
J
Let (t) be the solution of the system (8.55) passing through the point (to, )
and let (t) be the solution of the system (8.55) passing through the point
(79, H). Suppose o(t) and y(t) both exist on some interval «<t<p. Then to
each ¢>0 there corresponds

6 >0 such that if \t—?|
<6, |to —?9|<6, |W—h| <0,
then
Ib(t)—W(ii<e, «<t<fB,a<i<f. (8.57)
Proof. Since ¢(t) is the solution of (8.55) through the point (fo, n), we have,
for every t, a<t<f,
t
(t)=n+ |f(s, (s)) ds. (8.58)
Since y(t) is the solution of (8.55) through the point (29, 9), we have for
every t,a<t<f,
t
W(t)=H | f(s, W(s)) ds. (8.59)
Since
subtraction of (8.59) from (8.58) gives
(0) vijmn—ai+ |Hts. (9) #6, WO)] as [15.6 ds,

to to
and therefore
()—Y(l<in—Ai +||f(s, 6()) —£(s, W(s)) ds] + fi (spas. (8.60)

Using (8.56) in the eran side of (8.60), we wri
(WO <in—al + K |(5) —W(s) ds] + M leg tol.

H iacicicd al outhic sie
Ib()—W(9)
<0 Mo+K|1b(9) WO) ds. (8.61)
to
The Gronwall inequality (Lemma 1, Section 8.2) gives |(t)—w(d)|
8.7 Dependence on Initial Conditions 355
<6(1+M) exp(K |t—?9|)<6(1+ M) exp(K(B —«)), since |t —79|< B—«a. Since

u
w-¥OI <||ms W(s))h ds <M|t—t|<Mo

t
if |t—7?|<6, we have
Ib (t)—W (7) <1 (t)—W (| + (2) -—W(2) <5 (1 + M) exp[K (B—a)]+5M.

Now, given ¢>0, we need only choose 6<é/{M+(1+M) exp[K(B—«a)]}
to obtain (8.57) and thus complete the proof. |
Theorem | shows that the solution (t, to, n) of (8.55) passing through
the point (fo, 9) is a continuous function of the “‘triple”’ (¢, to, n). It is pos-
sible to show that if the initial point (7), H) of W(t) is sufficiently close to the
graph of the solution $(t), then there is a common interval on which both
solutions (¢) and W(t) exist (see [7], Chapter 1, Theorem 7.1).
In practice, the solutions o(¢) and w(t) are often known to remain
in a closed bounded subset of D. Thus the hypothesis that f and of/dy,
(j=1,...,) are bounded is automatically satisfied on this subset, even
though it may not hold on D, and this is all that is needed in the above
proof.
Continuous dependence on initial conditions is true under considerably
weaker hypotheses than those of Theorem 1. In fact, it can be shown that
uniqueness of solutions by itself implies the continuous dependence of
solutions on initial conditions (see [7], Chapter 2, Theorem 4.1). Under
stronger hypotheses, more can be proved; under the hypotheses of Theorem
1, it can actually be shown that (tf, to, n) is continuously differentiable with
respect to ¢) and the components of y (see [7], Chapter 1, Theorem 7.2).
The technique used to prove Theorem | can be applied to establish the
following result.
Theorem 2. Let f(t, y) and g(t, y) be defined in a region D and satisfy the
hypotheses of Theorem 1. Let (t) be the solution of y'=f(t, y), y(to)=n and
W(t) the solution of y'=g(t, y), y(to)= nN existing on a common interval
a<t<pP. Suppose |f(t, y)—g(t, y)|<e for (t, y) in D. Then the solutions
(1), W(t) satisfy the estimate
Id()—W
(| <m—Al exp(K |t— tol) +e(B—a) exp(K|t—tol), a<t<f.
Exercise
1. Prove Theorem 2. [Hint: Write the integral equations satisfied by (¢) and
W(t), and subtract to obtain
+ |[£(s, W(s))—g(s, W(s))] ds.

to
Then take norms, use the hypotheses, and apply the Gronwall inequality to
obtain the result.]
Theorem 2 says roughly that if two differential equations have their

right-hand sides “‘close together,” their solutions cannot differ by very
much. Theorem 2 also gives a different type of continuity property, indicated
by Exercise 2 below.
Exercises
2. Let f,(t, y) be a sequence of vector functions converging to f(t, y) in the sense that
f(t, y)—f,.(t, y)| <«, for (t, y) in D, with ¢,0 as ko, and let f and f, (k=1, 2,...)
satisfy the hypotheses of Theorem 2. Let n, be a sequence of constant vectors
converging to n. Let ,(¢) be the solution of y’=f;,(¢, y), y(¢o)=1 (K=1, 2,...),
existing on a<t<f, and let (t) be the solution of y’=f(t, y), y(to.)=n existing
on «<t<f. Show that lim,_,, W,())=(t) for «<1t<f.
3. Let f, of/éy;(j=1,...,n) be continuous in a region D. Let (ft, to, n) be the
solution of y’=f(t, y) for which @(¢o, fo, n)=n, and suppose that @ is differentiable
with respect to each of its (n +2) variables.
a) Show that 6o/@n; (t, fo, n) is the solution of the linear system w’=f,(¢, @(¢)) w
for which
op
an;(to, to. N)=e;,
the unit vector with jth component | and other components 0 (j=1,..., 7).
(Here f, denotes the matrix (6/;/0y,).)
b) Show that 0/01 (¢, fo, n) is the solution of w’=f,(t, @(¢)) w for which
op
at, (to, to, N)= —F(to, 0).
c) Show that
ae (t, to, n) f(to, n)=0.

CHAPTER 9
Numerical Methods
of Solution
In many problems the only effective method for obtaining information

about the solutions of a differential equation is to use a numerical approxi-
mation procedure. In this chapter we discuss some numerical methods,
beginning with the elementary Euler method and continuing with some other
methods which are better suited to high-speed computation. As we can
give only a brief introduction to this extensive subject, we refer the reader
to other books, for example [10], [12], [15], [17] for a more complete
discussion of many ideas.
9.1 THE EULER METHOD
We have seen in Example 2, Section 1.4, that the differential equation
y +2ty=sint
subject to the initial condition $(0)=yp9 can be “solved” for y=o(t).

The result is $(t)=e""yyt+e ‘ye sins ds. However, if we want the
numerical value of ¢(1), this “solution” is not very useful since it gives
1
p(1)=e 'yote' fe sins ds

0)
and the integral {j e* sins ds cannot be evaluated exactly. Of course, we

can approximate the integral by the trapezoidal rule or Simpson’s rule
and obtain a numerical value which approximates ¢(1).
Another difficulty,-more serious than this one, arises in the simple
pendulum model, for which we found in Exercise 10, Section 2.4 that the
357
358 Numerical Methods of Solution
displacement (ft) is defined by solving the following equation for @:

p(t)
L\1? dd dh
() |(cos#(t)—cosOo)'/2
(9)
It is not possible to evaluate the integral on the left side of this equation
in terms of elementary functions. A numerical approximation for thé
integral is impractical since the upper limit of integration is variable.
Even if a numerical approximation could be obtained, we would still have
the problem of solving for the implicitly defined function ¢. Also, the
integral is improper which adds to the difficulties.
In fact, in practical problems, the most common situation is that no
usable expression for the solution can be found at all, even though it can
be shown that there is a unique solution. When we meet such difficulties
we must often resort to the use of numerical approximations from the
start. In this section, we shall develop one approximation method and
we shall indicate some of its refinements which are frequently used on
electronic computers later in this chapter.
Let us consider the solution of the first-order differential equation
y'=f (ty) (9.1)
through the point (fo, yo). We assume that f satisfies the assumptions of
the existence and uniqueness theorem (Theorem 1, Section 1.6) in some rec-
tangle R= {(t, y) ||t—to|<a, |y—yo| <b}. Then the theorem assures us that
there is a unique solution ¢ of this problem existing on some interval
|t—to|<a, where «<a. We wish to find a numerical approximation for
the number $(t)+7), where T is specified and |7|<« (see Fig. 9.1). We
shall do this by a construction which makes use of the geometric inter-
pretation of the solution suggested in Section 1.5. To be specific, we suppose
that t)+ 7 is to the right of to, 0< T<«, and divide the interval [t), to + T]
into n subintervals by specifying intermediate points t9<t,<t,<---<t,=
to+T. In practice, these points are usually equally spaced, but this is not
necessary. Now start at (to, Yo). We know that the curve y=@(t) passes
through the point (to, yo) and that its derivative at (to, Vo) is f(to, Vo). Since
we do not know ¢(¢), we cannot follow it to ¢,. Instead, we pretend that the
solution is a straight line Ly with slope f(to, yo) and we follow this line to f,.
The premise, of course, is that, if t, is close enough to fo, the error made by
following the straight line segment instead of the solution (called the trunca-
tion error) is not too large. We know from analytic geometry that the
equation of the straight line Lo through (¢9, yo) with the slope my =f (to, Yo)
iS Y=Vo+(t—Lo)f(to, Yo). We compute y,, an approximation for ¢(t,),
by substituting t=¢, in the equation of Lo. This gives
V1 =Yo+(ts —to) f (tos Yo)=Yo

+Mo (ty —to).
9.1 The Euler Method 359
ly t, t, ls - t=f4+T t+a
a —— Qa fe!
Figure 9.1
We now pretend that the solution y=¢(¢) passes through (¢,, y,). If it
did, its tangent at (t;,y,;) would have slope m,=f(t,, y,). Let L, be the
straight line through (¢,, y,) with slope m,. We know that the equation
of L, is y=y, +(t—1,) f(t,, y;). We now proceed along the straight line L,
and compute our approximation y, to (t,) by substituting r=r, in the
equation of L,. This gives
Yo =Vp+(to—-t1)
f(ts, yi).
In general, of course, y, is not likely to be $(¢,), and o’(t,) is not necessarily
m,. Thus we are retaining the error made in the first stage and possibly
compounding it by using the wrong slope. For the moment, let us not
worry about this error. We can continue the process until, at the nth stage,
we obtain an approximation y, for (7).
This construction, which is called an iterative procedure, can also be
expressed analytically as follows. Having computed the approximations
y,(i=1, 2,..., k), we compute y,,, from the formula
Vut1 =Vert (teas —ty) f (tes Vad (9.2)

This relation is an example of a difference equation. If we denote by w(t)
the ‘‘approximate’”’ solution constructed by this method, then
W(t)=y,+(t—ty) f (tes Yad)

on the interval 4,<t<t,,,, for k=0, 1,...,.n—1. The graph of w is the
“polygonal curve” in Fig. 9.1. It can be regarded as the solution of the
difference equation (9.2), and the method amounts to approximating the
differential equation (9.1) by the difference equation (9.2).
If to+T is to the left of to, the same procedure can be used, with the
obvious modification that we go to the left along each line segment. In
practice, the interval [to, to+7] is usually divided into subintervals of

equal length h, so that h=T7/n. The formula (9.2) then becomes
Ver =Ye thf (ty, Ye). (9.3)

This technique of approximating solutions is called the Euler method or
the Cauchy polygon method. It is suitable for high-speed computation; in
practice, the methods most often used are refinements of the Euler method.
Example 1. Let ¢ be the solution of y’=y passing through (0, 1). Compute ¢(1)
using equally spaced mesh points with h=0.1.
Here f(t, y)=y and the iterative formula for y,,, becomes
Ve+1 =y,t+hy,=(1 +h) y, = (1.1) Vx (k=0, Ines 9)
and we can tabulate our results as follows:
k tk Vk Vist
= lly,
0 0.0 1.000 1.100

0.1 1.100 1.210
2D 0.2 1.210 11338)
3 0.3 1331 1.464
4 0.4 1.464 1.610
5 0.5 1.610 LATA
6 0.6 eerie 1.948
V 0.7 1.948 2.143
8 0.8 2.143 Zon
9 0.9 DAS 5y) 2s )3)
10 1.0 De 93
The exact solution is given by ¢(t)=e' and thus ¢(1)=e. We have obtained the
approximation 2.593 for e. Note that we have actually calculated (1 + 1/n)" with n= 10.
Exercise
1. Use the same procedure with h=0.05 to see how much better the approximation
becomes.
In Exercise 1, we are calculating (1+1/n)" with n=20. Since

lim, ...(1 + 1/n)"=e, it is clear that we can approximate e as accurately as
desired by using the Euler method with sufficiently small A and obtaining
the approximation (1 +h)" or (1 +1/n)" with n=1/h.
A numerical approximation is useless unless there is some way to esti-
mate the error of the method. What is actually needed is an error bound
—a maximum possible value for the error.
In the example, we rounded off all entries to three decimal places. This
introduces an error at each stage, called the round-off error, which may
91 The Euler Method 361
accumulate from step to step. It can be shown that we can decrease the
round-off error at each stage by keeping more decimal places.
There is another error in the Euler method, called the truncation error,
caused by the use of straight lines to approximate the solution curves.
We must be careful to distinguish between the /ocal truncation error, which
is the error that would be introduced in going from the value y, at ¢, to
the value y,,, at t,,, if y, were exact, and the cumulative truncation error,
which is the actual error in the value of @(t) + 7) caused by the approxima-
tion. You should note that the cumulative truncation error is not simply
the sum of all the local truncation errors. Since there may be an error already
present in y,, there may be an additional error in y,,, caused by using the
wrong slope at ¢,, as well as the wrong value of the solution.
As an example, let us calculate the local truncation error of the Euler
method. The differential equation
v=f(t y)
has (ft) as its exact solution. This means that
¢' (th=f (t, (0) (to StSty +a).

Now, (t,) is the exact value t,, }(t,+,) is the exact value at t,,,, and
blud-ola)= |eam |F0.00)a

Thus the exact value at t, 4, 1s given by
tk+1
blual=ole)+ |£6 6() ae. (9.4)

tk
The formula for the approximate value ),+, at t,+1 1S
Vert =Vet (tes 1 — te) (ths Ved: (9.5)
This formula is usually applied with all the subdivisions of equal length h,
so that t,4,—4,=/A. It then becomes
Mf (tes Ve):
Vier 1=Yut (9.6)
In this formula, y, is the approximate value at ¢,. The local truncation
error T;, is defined as |$(¢,+1)—¥x+1| under the assumption that (ty) =Vis
that is, that the approximate value at ¢, is exact. By subtracting (9.5) from
(9.4) we see that
tk+1
n=] |£660) dls) Mt)

For convenience of notation, set f(t, $(¢))=F(t). Then

te +1
=| |Fl)dt—(es—n) Fla),
tk
We can use the mean-value theorem to get an upper bound for the error.
We have
F (t)—F
(t,) =(t—t,) F’(sx)
where s, is some point between ¢, and f. From this, we obtain
tk+1 tk+i
(arm |(Fa)H(e—
19)Pod)a
=(ty41—&) F (t,)+ |(t—t,) F’(s,) dt
and we find that the local truncation error is given by

tk+1
n| |(t—t,) F’(s,) dt
tk
Suppose that M= max _|F’(t)|. Since

to<t<to+T
of
F'(t) rl
of Hy of
AU) Barwal $(1)) O()=2-(, AU) BieeA P(t) F(t, OO)
if f, Of/6t, and Of/éy are bounded on the rectangle R, a bound for the num-
ber M can be calculated explicitly from the expression
Of Of
M<max/——(t, y) +max ay y) max|/(t, y)|
where the maxima are all taken over the rectangle R. Note that this
expression does not depend on the particular (unknown) solution ¢. Then
we have tkh+1
M
T,.<M |(w dt=> (teri hy)’ (9.7)
tk
In the standard case of equal subdivisions, t,,,—t,=h, and we have
shown that the local truncation error of Euler’s method is at most 4Mh?.
If the cumulative truncation error were the sum of all the local trunca-
tion errors, then, since there are N=T/h steps, this error would be at most
1Mh?-N=4TMh.
We will show (later in this section) that, although this reasoning is invalid,
the cumulative truncation error is in fact no greater than a constant
multiplied by A. This shows that we can reduce the cumulative truncation
error by making the value of 4 smaller (that is, by increasing the number
of subdivisions).
The bound obtained here for the truncation error is usually far larger
than the actual error which occurs when the method is applied to a
specific problem. This is a common phenomenon in numerical analysis.
The method we have used to estimate the truncation error suggests an
obvious refinement—use a more sophisticated approximation for the
integral, such as the trapezoidal rule or Simpson’s rule. In fact, the use
of Simpson’s rule leads to an approximation method known as Milne’s
method (see for example Section 9.3) which is widely used for computation.
It has a cumulative truncation error no greater than a constant multiplied
by At.
The fact that the truncation error involves a positive power of h implies
that for small values of h, we obtain a high degree of accuracy. Roughly
speaking, the higher the power of h, the more accurate the method. It
would be possible to devise very complicated approximate procedures
with truncation errors no greater than a constant times h!’, but it is more
effective and faster on a computer to use a relatively simple formula, and
a very small value of h.
The method we have discussed can be adapted to second-order differ-
ential equations, but the adaptation involves some complications that we
shall not discuss here.
Exercises
2. Use the Euler method to estimate (1), where ¢ is the solution of y’=1+y passing
through (0, 0). Use equal spacing with h=4,,,—t,=0.1. Also calculate (1)
exactly by solving the equation, and compare the results.
3. Repeat Exercise 2 for the differential equation y’+2ty=sin¢ with initial condition
~(0)=0. Note that you will need to use tables to find sin 0.1, sin 0.2, etc., and that
you will not be able to compare your answer with the exact solution of the
equation.
4. Use the Euler method to estimate the value of $(1) where ¢ is the solution of
dy
—=2ty+1—2t?
dt
passing through (0, 0). Use equal spacing with h=0.1.
An approximation method, such as the Euler method, is subject to errors

of two quite different kinds. One is the round-off error, caused by the fact
that the values actually calculated are rounded off to a certain number of
decimal places or to a certain number of significant figures, and are thus really
approximations to the values given by the approximation method. The
errors so introduced may accumulate as we proceed from one stage to the
next. The analysis of round-off errors, being largely statistical in nature, is
quite difficult. We shall not discuss it further at the moment, except to suggest
that an approximation method, in order to be useful, ought to have the
property that a smaller round-off error at each stage produces a smaller
cumulative round-off error. Since we can reduce the round-off error at each
stage by keeping more decimal places in the calculations, this property
would give some control over the cumulative round-off error. It will turn
out that we can check rather easily whether this property, called stability of
the numerical method, holds for any given numerical method or not.
The other type of error is the truncation error, caused by the approxima-
tion in the method itself. For the Euler method, this arises in the use of
(t.+1—%) f(t, &(t,)) aS an approximation to the integral |**' f(s, o(s)) ds.
We have defined the local truncation error to be the error introduced in
going from the value y, at ¢, to the value y,,, at t,,,, assuming the value
y, to be exact rather than an approximation. We have shown that the local
truncation error of the Euler method (9.1) is no greater than $Mh?. The
constant M can be obtained from bounds for the function f and its
first-order partial derivatives in the region under consideration.
We have suggested that the cumulative truncation error of the Euler
method is not greater than a constant multiplied by h. The cumulative
truncation error is defined to be the actual deviation of the approximation yy
from the true value $(to+ T)= (ty). (This tacitly assumes that all numbers
can be computed without round-off error.) Let us now estimate this
quantity. We define
E,=|0 (te) — ysl k salsa
to be the cumulative truncation error at the kth stage.
Theorem 1. If f is continuous and has continuous first-order partial

derivatives with respect to t and y for ty<t<ty+Nh and all y, then the
cumulative truncation error of the Euler method with step length h is no
greater than a constant multiple of h.
Proof. Since, by (9.4)

Tk+1
eR Sue |Ronen
and by (9.6)
Vat =e thf (ths Va)

we obtain
tk+1
P(te+1)—Ve+1 =P (ty) —Iet |FS(s, b(s)) ds—hf (ths Vx)
Srsient |FG sane ne

thf (th, (te)
—hf (ths Ve)-
Thus, taking absolute values, we obtain
IP (t+ 1)—Ve+ stb) n+ |f(s, P(s)) ds—hf (t,, &(t,))
thf (ty, b(t) Ff (ths Vidl-

Using the definition of E,, we have
tk+1
oe <E,+| | Feeney Caeen (9.8)

+h |f (te $(t,))—SF(tes Vodl-
The local truncation error estimated previously is
Ik+1
| ONTO) NOMEN)
tk
and we have shown that this is no greater than 4Mh*. Since we have
assumed that f has a continuous partial derivative with respect to y, the
mean-value theorem for derivatives shows that there exists a constant L
such that
IF (tes P(ti))—F (tes Ye S L1G (th)
—Vel = LE,
(see Section 8.1, for a similar calculation). When we use these estimates
in (9.8), we see that
Ex+1<E,+4Mh? +hLE,=E,(1+hL)+3Mh? k=0,...,N. (9.9)
It is now easy to use induction to estimate E,.
Exercise
5. Show that the inequality (9.9) implies
disuse)
Ey 2
hLy*—11
Mi?, k=0,1,...,N (9.10)
hL
(Note that Ey =0.)
To complete the argument, we require an elementary inequality.

Lemma 1. Ifp>0, q>0, then (1+p)4<e4.
Proof. For p>0,e?>1+p. Taking (natural) logarithms, we obtain
p>log(1+p). It follows that pg>gq log(1+p)=log(1+p)%, which implies
e’4>(1+>)*, as desired. | E
Now, we apply this lemma to give (1 +AL)*<e**", and then (9.10) yields
5 Mh, k=0,1,...,N. (9.11)
If we define a new constant M’= M(e'? —1)/2L, this becomes
Ey<M‘h
which completes the proof of Theorem 1. 1
Since the cumulative truncation error of the Euler method is no greater

than a constant multiplied by 4, we can make this error as small as we wish
by making the length A of the subintervals sufficiently small, that is, by
making the number AN of subintervals sufficiently large. However, it is more
efficient to use a more sophisticated approximation method whose
cumulative truncation error is no greater than a constant multiplied by
some higher power of h. There is a balance to be sought here between the
more involved computation of a more sophisticated method and the more
rapid shrinking of the cumulative truncation error as / is decreased. In
the next two sections, we shall examine some refinements of the Euler
method which are sufficiently simple for effective calculation, yet sufficiently
accurate that a satisfactorily small error can usually be obtained with a
reasonably small number of subintervals. You should remember that a
large number of subintervals means that a large number of separate
calculations must be made, and a large number of simple calculations may
take longer, either by hand or by computing machine, than a smaller
number of more complicated calculations. However, each problem and
method must be examined with the particular computer available in mind.
The Euler method may be illustrated from the point of view of the
computer by a flow diagram (Fig. 9.2). Such a diagram indicates the
processes to be carried out by the computer. An instruction such as
t, +h->t, means that the next step in the iterative procedure is started
by replacement of 4+h by 44, (that is, K+, =t,+h). The question
(t,>to+T7?) is included in the flow diagram to show the process by which
the computer decides when to stop iterating the procedure.
Read in h, yo, to, fo + T
bt
Olea) &
Se = f(tes Ye)
i& thot
Ye ar hf, Yr
Record ti, yx
&>t+T? No
Figure 9.2
Exercises
6. Use the Euler method, with h=0.1, to approximate the value for t=1 of the
solution of the differential equation y’=t+y passing through the origin.
7. Use the Euler method, first with h=0.2, and then again with h=0.1, to find an
approximation to #(1), where @ is the solution of y’=7o(t? +”) such that $(0)=1.
Can you make any estimate of the accuracy of the answers?
8. The differential equation y’=3y/? has an infinite number of solutions through the
origin, as we have seen in Sections 1.6. Suppose we try to use the Euler method to
approximate one of these solutions, and repeat the process with a sequence of step
sizes {h,} which decreases to zero. Does the sequence of approximations converge
as n—0o? If so, to which solution does the sequence converge?
9. Use the Euler method, first with h=0.2, and then again with h=0.1; to approxi-
mate (1), where @ is the solution of y’=ty?—y such that (0)=1. How small
must / be chosen for the Euler method to give an approximation which is correct
to two significant figures?
9.2 THE MODIFIED EULER METHOD
The Euler method, although easy to apply, has too large a truncation
error to be of much use in the actual calculation of numerical approxima-
tions. This relatively large truncation error is due to the use of the crude
approximation h/(t,, y,) to the integral {i<*' f(s, b(s)) ds. By using a better
approximation to this integral, we may expect to obtain a numerical method
with a smaller truncation error. One obvious improvement is to approxi-
mate the integral by the length of the interval multiplied by the value of
the integrand at the midpoint of the interval (see Fig. 9.3), rather than by
Af
ty = fre)
tea ti+2
Figure 9.3
the length of the interval multiplied by the value of the integrand at one end
of the interval. To do this, we integrate over two subintervals rather than
divide each subinterval. Using (9.4), we write
k+2
P(tk+2)=P(ty) + | f(s, P(s)) ds (9.12)

tk
and we approximate the integral in (9.12) by 2Af(t, +1, 6(4,+1)). This leads
to an approximation method given by the iterative formula
Vu+2=Ve+2Hf (the 1. Ver): (9.13)
This method is called the modified Euler method, and the method of
approximating the integral used in the modified Euler method is called
midpoint quadrature. It is sometimes used in the numerical approximation
of definite integrals. We shall see (Theorem 1 below) that the modified
Euler method has a significantly smaller local truncation error than the
Euler method.
The modified Euler method expresses y,,, in terms of y, and 44.
Since it involves two subintervals, it is called a two-step method. This
introduces a difficulty which does not arise in the Euler method. In order
to begin the approximation procedure by taking k=0 in (9.13) to calculate
2, we need not only the given initial value yo but also y,. However,
9.2 The Modified Euler Method 369
y, is not given by the initial conditions and we must use some other ap-
proximation procedures, called a starting method, to calculate y,.
One method of obtaining a value for y,, which is frequently used if the
function f is analytic, is to use a power series expansion about fg (see
Section 6.1). Power series expansions are not very useful for numerical ap-
proximations because their convergence near the ends of the interval of
convergence may be very slow. However, if ft, is close to fo, a power series
expansion may provide an accurate yet easily obtained approximation
for yj.
A second method of obtaining a value for y, is to use a one-step method,
such as the Euler method. However, to improve the accuracy we subdivide
the interval [f, ¢,] into smaller subintervals. Another approach, probably
the most commonly used one, is to use a Runge-Kutta method. The basic
idea of the Runge-Kutta methods is to obtain as small a truncation error
as possible in an explicit one-step method. This requires subdivision of
the interval [7), ¢,] and, thus, the Runge-Kutta methods may be regarded
as refinements of the method, suggested above, of subdividing the interval
[to, t; ]and using the Euler method. We shall discuss Runge-Kutta methods
in Section 9.5; for the moment we wish only to point out the need for starting
methods and to suggest some possibilities.
Example 1. Let us estimate e by using the modified Euler method with h=0.1 to
approximate (1)=e, where ¢@ is the solution of y’=y such that #(0)=1. We begin by
using a power series expansion to estimate $(0.1)=1+0.1+3(0.1)*+---. This gives the
value y, = 1.105, correct to three decimal places. Now, we can use the iterative formula
(9.13). Using h=0.1, f(t, y)=y, in (9.13), we obtain
Vat2=Vet+O2V41-
It is convenient to tabulate the calculations as follows.
li Ve Ver1 Ve t0.241=Ve+2
0.0 1.000 1.105 i 22

0.1 1.105 12 1.349
0.2 12211 1.349 1.49]
0.3 1.349 1.491 1.647
0.4 1.49] 1.647 1.820
0.5 1.647 1.820 2.011
0.6 1.820 2.011 Dn)
0.7 2.011 DD 2.455
0.8 2022. 2.455 DNS
0.9 2.455 Dail
1.0 DANS
We obtain the approximation 2.713 for e, which is considerably better than the ap-
proximation 2.593 obtained in Section 9.1 by the Euler method with the same number
of mesh points.
We have suggested that the modified Euler method is more accurate

than the Euler method for one particular problem. Let us prove that this
is true in general by calculating the local truncation error of the modified
Euler method.
Theorem 1. If f is continuous and has continuous partial derivatives with
respect to t and y of the first and second orders for to<t<tg+Nh and
for all y, then the local truncation error of the modified Euler method with step
length h is no greater than a constant multiple of h?.
Proof. Let ¢ be the solution of the equation y’=f(t, y) which satisfies the
initial condition y(to)=yo. Then, by (9.4),
tk+2
HAE: | POLIOES (9.14)

tk
The approximations to this solution are defined by (9.13). The local

truncation error 7, is defined to be |f(t,42)—),42|, under the assumption
that the approximate values at 4, and ¢,,, are exact, that is, that $(t,)=y,
and that $(t,41)=),+1. Subtracting (9.13) from (9.14) and assuming that
P(t) =Ves P(t.+1)=Ye+1, we see that
tk+2
| | f(s, $(s)) ds —2hf (ty41, b(t +1))}-

tk
For convenience of notation, we set f(t, p(t))=F(t). If F is twice differ-

entiable, which is the case if fhas continuous second-order partial derivatives
with respect to ¢ and y, we can use Taylor’s theorem to write
F (s)=F(tya1)+(8—
thai) F(teai)+ Se
(s—tea1)” _,
(th SSS tha25 e<o<tk+2)-

Integration with respect to s from t, to t,4 gives
k+2 tk+2 tk+2

Sy ea:
Fs) ds—20F (25) [|(tos) Pltces) ds [Bee (jas.
tk tk tk
However,
tk +2
(Ste a) F’ (t, 44) ds she )) [(te+2—te+1)? —(—thas)?]
=4F" (t,, 1) [h?—h?]=0.

9.2 The Modified Euler Method 371
Thus the local truncation error 7; is

tk+2
2
5 F’(é) ds
a
tk
Suppose that M=max ,, <;<,,+7 |F” (|. There is such a constant M since the
assumptions of the theorem imply that F has a continuous second derivative.
Then
tk+2
M
tk
and Theorem | is proved. |
Theorem | suggests that, if A is small, the modified Euler method

represents an improvement in accuracy over the Euler method. This
improvement, however, may be more apparent than real. The constant
M in (9.15), being a bound for |F”(t)|, involves bounds for the second-order
partial derivatives of f, while the corresponding constant in the truncation
error for the Euler method involves only bounds for the first-order partial
derivatives of f. This means that the constant in the bound for the
truncation error for the modified Euler method may be considerably larger
than the constant in the bound for the truncation error for the Euler
method. However, if / is sufficiently small, the additional factor h makes
the modified Euler method more accurate than the Euler method. The
essential point is that the power of 4 is more important than the constant,
but only if h is small enough.
You should note that we have not yet estimated the cumulative
truncation error for the modified Euler method. Rather than doing this
now, we state the fact that this cumulative truncation error is no greater
than a constant multiple of h?, and defer the proof to the next section,
where we shall indicate a general technique for passing from the local
truncation error to the cumulative truncation error.
Exercises
1. Use the modified Euler method, with h=0.2, to approximate the value for t=1 of
the solution of the differential equation y’=t+y passing through the origin.
Repeat the problem with h=0.1, and compare the two results with the result
obtained in Exercise 2, Section 9.1.
2. Use the modified Euler method, with h=0.1, to find an approximation to ¢(1),
where ¢ is the solution of y’=75(t? +7) such that ¢(0)=1. Compare the result
with the result obtained in Exercise 7, Section 9.1.
3. Use the modified Euler method with h=0.1, to approximate the value ¢(0.5) of the
solution @ of y’=y—t such that ¢(0)=0. Compare your answer with the one ob-
tained by the Euler method with h=0.1 and the one obtained by explicit solution
of the differential equation.
4. Draw a flow diagram (analogous to Fig. 9.2, for the Euler method) for the modified
Euler method.
Another possible approximation method is to use trapezoidal integra-

tion, that is, to approximate °
f(s, o(s)) ds by :LS (te P(t)

+S (teri, P(te+1))J
which is the area of the trapezoid bounded by the line segment joining
[tes A(tis O(ty))) to thei, f(te41. O(te+1))] and the three lines y=0, t=t,,
oy = fil, o(l))
tear
Figure 9.4
t=t,,,; see Fig. 9.4. This approximation leads to the iterative formula
hive 4
Ven =p aa) (Ate, Ve)
+ fete View alle (9.16)
This method, called the improved Euler method, gives y,., implicitly
rather than explicitly. There are methods, as we shall soon indicate, for
dealing with implicit formulas. For some problems the improved Euler
method is useful.
9.3 THE MILNE METHOD
A more accurate approximation than the Euler and modified Euler

methods is obtained by the use of Simpson’s rule to approximate the
integral (}**? f(s, $(s)) ds. Simpson’s rule gives the approximation
:[SF(tes P(t)
+46 (thor, P(tes
+S (thea, (ty +2))] (9.17)
9.3 The Milne Method 373
for this integral and leads to the iterative formula

i § :
Ver 2=Vie v3 Lf (tis Ve) +4S (tevas Vex) +S (Geos Vu+2)] (9.18)
known as the Milne method of approximation.

The use of the approximation (9.17) is suggested by an attempt to ap-
proximate the curve y=f(t, (¢)) by a parabola on the interval [t,<t<t,4>].
As before, we write F(t)=/(t, ¢(t)) for convenience of notation. Now, we
determine the constants a, b, c so that the parabola
y=atb(s—tys1)+¢(8—tha1)?
with vertex at (¢,.,, a) and axis along the y axis passes through the three
y
legt
Figure 9.5
points (t,, F(t), (te+1> F(te41))s (e+2, F(e+2)); see Fig. 9.5. Since +.
— ty, =t+1—t,=h, the conditions that these points lie on the parabola are
F(t,)=a—bh+ch’,
Pa )\=a-; (9.19)
F(t, 42)=a+bh+ch’.
We can solve the equations (9.19) for the three constants a, b, c and
then use te42
[a+b(s—tys1)+¢(s—tk+1)7] ds
tk
as an approximation to J/**? F(s) ds.

Since
tk +2
[a+ b(s—tes1)+e(s—th+1)?] ds
b tk+2
2
H[est3o—nad #5 bt =2ah+— ch3 (9.20)
tk
we do not even need to solve (9.19) for 5. It is easy to obtain
F (ty+2) —2F (tess) + F (ty)

G=F (t44) C=
2h?
and when we substitute these values in (9.20), we obtain (9.17) as an
approximation for the integral {/**? F(s) ds=|**? f(s, o(s)) ds.
Example 1. Let us use the Milne method (9.18) with A=0.1 to solve the same problem
as the one discussed in Example 1, Section 9.2. As before, we begin by using a power
series expansion to obtain the approximation y, =1.105. Now, the formula (9.18), with
S(t, y)=y, h=0.1, becomes
Viet 2 = Va +30(Ve
+4 +1+ e+ 2):
We can solve this for y,
4 , obtaining
29, 42=31
yy +4 e412
Now, we can tabulate the calculations as follows.
ty Vk Ve+1 SVE
t+TOV +1=Ve+2
0.0 1.000 1.105 1.069 + 0.152=1.221

0.1 1.105 1.221 1.181 +0.169=1.350
0.2 22 1.350 1.305+0.186=1.491
0.3 1.350 1.491 1.443 + 0.206 = 1.649
0.4 1.491 1.649 1.594+ 0.227 = 1.821
0.5 1.649 1.821 1.763 + 0.251 =2.014
0.6 1.821 2.014 1.947 +0.278 =2.225
0.7 2.014 2.225 2.153 +0.306 =2.459
0.8 2225 2.459 2.378 + 0.339 =2.717
0.9 2.459 DPATANG
1.0 DANG
The accuracy of this approximation, as compared to the accuracies of the approxima-

tions obtained earlier by the Euler method and the modified Euler method, suggests the
usefulness of the Milne method in practice.
Although the Milne method is implicit (that is, the iterative formula (9.18)
which defines the method expresses the approximation y,,, which is being
calculated implicitly in terms of y, and y,4,) this causes no difficulty because
the differential equation y’ =y is linear, and this makes it possible to solve for
Ye+2 explicitly. However, the use of the Milne method for nonlinear dif-
ferential equations requires some means of dealing with an implicit iterative
formula.
The usual technique is to solve (9.18) by another iterative procedure at
each step. One uses some other method to obtain a first approximation
Ve+2 tO P(% +42) and then calculate a second approximation y,,, from the
formula
ess . .
Ve +2 =Ver 3 [Lf (tes ve) +4F (thoi Ver tS (tha, Jn+2)]- (9.21)
For more accuracy we may iterate this procedure, substituting y,,, in the
right-hand side of (9.21) and using (9.21) to calculate a third approximation,
and continuing in the same way. It is possible to prove that if fis continuous
and has continuous first-order partial derivatives with respect to ¢ and y,
then the approximations obtained in this way converge to a solution of the
difference equation (9.18) for h sufficiently small. However, it is usually
impractical to carry out this iteration procedure at each stage of the
approximation method. Normally, the most efficient procedure is to calculate
De+2, Called the predictor, by some explicit method, such as the Euler method
or the modified Euler method, and then use (9.21), called a predictor-corrector
formula, to calculate the corrector y,4 5.This value y,, 5 is the one used in
continuing to the next stage in the Milne method.
Exercises
1. Use the Milne method with h=0.2 to approximate the value for t=1 ofthe solution
of the differential equation y’=/+ y passing through the origin. Repeat the problem
with h=0.1, and compare the two results with those obtained in Exercise 1, Section
a2
2. Use the Milne method, with =0.1, to find an approximation to (1), where @ is
the solution of y’=75(t?+,y7) such that #(0)=1. Compare the result with that
obtained in Exercise 7, Section 9.1.
3. Draw a flow diagram (analogous to Fig. 9.2, Section 9.1, for the Euler method)
for the Milne method.
4. Use the Milne method with h=0.2 to approximate (1), where ¢ is the solution
of y’=y-—tr such that ¢(0)=1. Compare your answer with that obtained by
a) using the Euler method with h=0.05,
b) using the modified Euler method with h=0.1,
c) finding the exact solution by the methods of Chapter 1.
5. Find an approximation to #(0.5), where ¢ is the solution of y’=y+y? such that
$(0)=1, by using the Milne method with h=0.1, and with the modified Euler
method as predictor.
6. What value of / is needed to obtain an approximation correct to two significant
figures, using each of the Euler, modified Euler, and Milne methods, for $(1),
where $(t)=e~" is the solution of y’+2ty=0 such that #(0)=1?
The estimate of the local truncation error for the Milne method is con-
siderably more difficult than the error estimates we have obtained previously.
Obviously, the local truncation error depends on the accuracy of the predictor
used. The simplest way to estimate this error is to split it into two parts. First,
we estimate the local truncation error in the formula (9.21) under the as-
sumption that the predictor }, > is exact. In addition, we assume as before
that y, and y,,, are exact in our estimate. Then we calculate the additional
error introduced in (9.21) by the error in the predictor. This, of course, will
depend on the method used to obtain the predictor.
Let ¢ be the solution of the equation y’=f(t, y) which satisfies the fal
condition y’(to)=yo. Then, by (9.4),
tk +2
Hiy=ol™ |Hens (9.22)

tk
The local truncation error of the Milne method is defined to be 7,

=|6(t,.42)—¥,+2|, calculated under the assumption that the values y, and
Yee in (9.21) are exact. Subtracting (9.21) from (9.22) and assuming
P (ti) =Ves P (te +1) =Ve+1, we obtain
Ik +2
h
=| | f (s, &(s)) as—; LS (the Va) H4S (teas, Ver tS (tetas Dx+2)]}. (9.23)
tk
For convenience, we set f(t, ¢(t))= F(t), as before, and we split 7, into two
parts U, and V,. The term U, is calculated under the assumption that
Pe+2 is exact, that is, ),4.=—(t,42). Thus,
tk+2
u,-||Fo
(s) ds 5 [F(t +4F (te) + Flee al] (9.24)
The term V, is the additional error introduced by the assumption

Vu+2 = b(t +2). Thus,
h
V= 3 |S (let 2» Plta+2)— F (tes 2, Jx+2)I- (9.25)
Exercise
7. Show that 7,<U,+4+ V,.
Let us estimate U,; this is simply the truncation error in Simpson’s rule.
In doing this, we must assume that f(t, y) a continuous partial derivatives
of all orders up to the fourth. We define G(t)={* F(s) ds on [t, t+2], so
that G’(t)= F(t), and then we define the new Aa 3ug[0, A] by
P(t)=G(ty 4 PGi). LF (tes: +t) +4F (ty+1)

+ F (t,41—7)] O<t<h. (9.26)
Comparing (9.24) and (9.26), we see that U,=|P(h)|. Differentiating (9.26)

repeatedly, we obtain
P(t) =F (t.41+1)—F (G41 —-1)— 3 LF (1 $2) + 4P he)
aC aie LF art F (eno),
P= 2 Feit) —2F (4 1) + F419)
- (Ptr tt) F(a Oe ia
PYO=3LF (hai t2)—F (ei —9)-5 LF" (t.+1 +7)
pre
tee y=-e)|
mt T m Wa
Ip Oe (ta. +t)-F (t,41—T)].
By the mean-value theorem, there exists €, t,<t4,—-T<€<h4,ttSh42,

such that
F" (tes +1)—F" (tee 1 — = 20F™ (2).
Thus
P()\=—47-
FF (e). (9.28)
We observe from (9.26), (9.27) that P(0)=P’(0)=P’(0)=0. Therefore,
Pre)= |Pr) do; P()=| Po) do; p()=| Po) do. (9.29)
0) 0 0)
Suppose that M= max _ |F(t)|. Then

to<t<to+T
—-M<F®(E)<M. (9.30)
From (9.28) and (9.30), we obtain
—$07M <P" (1)<30°M.
Integrating and using (9.29), we obtain
—§t°M <P" (t)<8t°M.
Repeating the integration twice more, we obtain
—gt*M <P’ (t)<75t*M
— 51°
M <P(t)<59t°M.
The final step in the calculation of U, is the use of U,=|P(h)| to obtain
. U,<o5Mh. (9.31)
This quantity is called the local truncation error of the Milne method, rather
than 7,, even though 7, is the quantity which actually interests us.
To estimate V,, we will only need to use the fact that f(¢, y) is continuous
and has continuous first-order partial derivatives with respect to ¢ and y for
to <t<tg+T and all y. This implies that there exists a constant L such that
f(t y- F(t NIsLly—dI (9.32)

for to <t<to+T and all y. (See Eq. (8.7), Section 8.1). Let us assume that
the truncation error of the method used to calculate the predictor j,4 is no
greater than Ah? for some positive integer p and some positive constant A.
For example, if the Euler method is used, p=2 and if the modified Euler
method is used, p=3. Then, using (9.25) and (9.32), we see that
Lh A
Vi a IP (th+2)—Veral-
Since |P(t,+2)—J_42| 1s the truncation error of the method used to calculate

Jy+2, we have
Lh
Vi <>Z (Ah). (9.33)
Since 7, <U,+V,, we obtain, using (9.31) and (9.33),
Then,
A
te 5 ace peck. (9.34)
If p is less than 4, the second term on the right side of (9.34) dominates for
small h, while if p is greater than or equal to 4, the first term on the right side
dominates. If we define g to be the smaller of the integers 5 and p+1, then
we can find a constant B such that
M
50 5
h?+ AL
3 enp+1Mececge offq
for small A and, thus,

T, < Bh‘.
Notice that this error estimate is not improved by improving the method
used to calculate the predictor beyond a certain point. If we can use a
method with local truncation error no greater than a constant multiple of
h*, then 7, is no greater than a constant multiple of A>.
In order to estimate the cumulative truncation error for the Milne
method, we give a general method for passing from the local truncation
error to the cumulative truncation error. We shall formulate this result
only for one-step methods, but there is no serious difficulty in extending
it to two-step methods such as the Milne method, provided/ has sufficiently

many continuous partial derivatives.
Theorem 1. Suppose that if f is continuous and has continuous first-
order partial derivatives with respect to t and y for tpy<t<tg+T and all y.
Suppose y, (k=1,..., N) are the approximations calculated by some one-step
method with step-length h to the solution $ of y'=f(t, y), Y(to)=Yo-. Suppose
that the local truncation error of the method is no greater than s. Then the
cumulative truncation error is no greater than a constant multiple of é/h.
Proof. For this proof it is convenient to define $, to be the exact solution
of the differential equation such that
Py (ty) = Vr (k=0, Weed
Thus, ¢o is the solution ¢ which satisfies $(t9) =o. In this notation, the local
truncation error 7; is given by
T= |Va+i — $y (tzssl=lPn is (tes 1)— x (tee a) (k=0, Ie N (9.35)

(See Fig.9.6.)
y | ¥ = beri)
(tiga, Ver) de
Figure 9.6
The cumulative truncation error is
Ey =|yy —$(ty)|=ln (tw) —— Po( ty)|< ene —x-1(ty)|. (9.36)
We must estimate |, (ty)—@,—1(¢y)|. This quantity is the difference at

ty between two solutions of (9.1) whose difference at ¢, is the local truncation
error 7,_,<é, because of (9.35). This difference can easily be estimated
by means of the Gronwall inequality (Lemma 1, Section 8.2). Since
b(t)=YotJi, f(s, G(s) ds, we can write
tn
Pr —1 (tw) =x —1(t,)+|f(s, Px—1(s)) ds (k=1,..., N).
Subtracting and using (9.32) and (9.35), we obtain

tn
lpi (tw) — Px —1 (tw)| <i da-sle+ |If (s, bx(s))— f(s, bx -1(s))| ds
tk
<Tth| 1$u(3)
—Oy-1(9)| ds
set |ids()~ 1) dst esGokeotein):

Now, the Gronwall inequality gives
|x (tw) — bx- 1(y)| Se exp L(ty —t,)=e exp L(N—k)h (k=1,..., N).
We substitute this into (9.36) and find
N N
Ey<ée ), expL(N—k)h=eexpLNh ¥\ exp(—kLh).
k=1 k=1
The geometric series })f_ , exp(—LA) with first term e~ and ratio e" ™ has
sum
ane 1—e7NLh 1—e NL
3 je hh]
Thus
eLNh_ 4
NEE 1
and since e“"—1>Lh, Nh=T, we obtain
eo 1
[Bry
oye
or
Me
<<
erty,
where M =((e"7 —1)/L). |
Theorem | shows that if the /ocal truncation error of a one-step method
is no greater than a constant multiple of h?, then the cumulative truncation
error is no greater than a constant multiple of 4?~'. As an analogous
result can be proved for two-step methods (and for methods involving any
9.4 Stability, Consistency, and Convergence 381
finite number of steps), we now see that if fhas sufficiently many continuous
partial derivatives, the cumulative truncation error of the modified Euler
method is no greater than a constant multiple of h?, and the cumulative
truncation error of the Milne method (with a predictor whose local trunca-
tion error is no greater than a constant multiple of h*) is no greater than a
constant multiple of h’*.
Obviously, the accuracy of approximation is improved if a method whose
truncation error involves a higher power of / is used. On the other hand,
to obtain this improvement in accuracy, a more complicated approximation
procedure is required. The benefits of improved accuracy may be offset by
the computational difficulties. For many problems, the Milne method
represents a good compromise between accuracy of the formula and con-
venience of computation. The reader should remember that the truncation
error depends on f/, and can always be reduced by making / smaller.
The fact that the Milne method is implicit causes (as we have seen)
some technical problems in its use and in the estimate of its error. There
is another disadvantage, which we shall discuss in the next section. For
some differential equations it is unsuitable because of difficulties arising
from the round-off error, and other methods are more useful. One such
method is the Adams method, which is obtained by the use of interpolating
polynomials in approximating {{*" f(s, o(s)) ds (see [10]). In implicit form
this is given by
h
Vera =Ver2 7 a4 LF (tes Vi — SS (tears Ver s)+ LOS (tes 25 Ye+2)]
+9 f (ty+3> Vers) (937)
which has a local truncation error no greater than a constant multiple of

h°. An explicit Adams formula is
h :
Ves Veron s [Sf (te Vi) 16S (thet, Vers)
+23S (toa, Ve+2)] (9.38)
whose local truncation error is no greater than a constant multiple of A‘.

While the Adams methods are no more accurate than the Milne method,
and are three-step methods (which require more computation), they are free
from the disadvantages of the Milne method.
9.4 STABILITY, CONSISTENCY, AND CONVERGENCE
When we use a numerical method to obtain an approximation to the

solution of a differential equation, we are trying to find a set of numbers
y, (k=1,..., N) defined by the method. However, we actually round off
all numbers calculated to a specified number of decimal places, or to a spe-
cified number of significant figures, and thus we actually obtain a slightly
different set of numbers z, (k=1,..., N). The difference r,= |z,—y,| is called
the round-off error. If we retain more decimal places or significant figures,
we obtain another set of numbers 2, (k=1,...,) with round-off error
7, =|2,—Yx|. If |zy —Zy| is small whenever all |r, —7,| (k=1,..., N) are small,
that is, if a small change in the round-off error at each stage produces only
a small change in the final result, then the approximation method is said
to be computationally stable. The property of stability depends on the ap-
proximation method and has nothing to do with the suitability of the method
for a particular problem. For example, the formula yo=1, y,+;=3), is a
ridiculous attempt to approximate the solution ¢ of y’=0 such that ¢(0)=1,
but it is stable in the sense just defined.
A numerical method is said to be consistent with a differential equation
if the solution of the differential equation satisfies the approximation
scheme except for terms which tend to zero as h-0. A numerical method
is said to be convergent if the approximations tend to the actual solution
as h>0. You will note that all methods presented above are convergent.
Example 1. Consider the Euler method as an approximation for the solution ¢ of
the initial-value problem, y’=y, #(0)=1.
We wish to show that it is both consistent and convergent. Consider the exact
solution $(t)=e'. At a mesh point t,=”h, we have, from the Euler method,
Vn+1 =(y, hy.) =Vas 1 —(1 +h) Yn-
Substituting the exact solution, (t,)=e", we have

P(tra1)-(1 +h) b(t,)=e"* P"—(1 +h) e™
=e"™(e"—1—h).
Since nh=t we have |e""| < K, where K is a constant, for any finite ¢. Also, |e*—1—h|<
< Mh? for some constant M. Therefore
ID(tn+1)—(L+h) O(t,)|
SK ih?
for some constant K,>0. Since K,h? tends to zero as h-0, the Euler method (for this
problem) is consistent.
To show that the method is convergent, we solve the difference equation
Yn+1=(1 +h) Via Yo=1.
We obtain, by an easy induction, y, =(1 +4)". Since t=nh, we consider lim,_,9(1+)"=

lim, .o(1 +/)"=e', which, of course, is the exact solution.
Exercise
1. Show that the modified Euler method for the problem considered in Example 1
above is consistent and convergent.
Obviously, a method which is not convergent is useless for obtaining

numerical approximations. It is possible to give conditions for stability
9.4 Stability, Consistency, and Convergence 383
and consistency of numerical methods which are easy to verify. Then the
following result, for whose proof we refer the reader to more specialized
works such as [8], is of great importance.
Theorem 1. A stable, consistent finite difference approximation method is
convergent.
Of course, as we have seen in the preceding sections, the rate of con-
vergence of the approximations to the actual solution as h->0 depends
on the truncation error, and this must be estimated for each method.
To unify our previous considerations, we consider the difference equation
p Pp
2S Ver ih y Bif (their Yuet i) (9.39)
with «, #0, and with « and f, not both zero. For example, in the Euler meth-
od (9.6), p=1; a, =1, a =—1; B,=0, By=1. The modified Euler method
(9.13) 4s_ziven by (9.39) with p=2; o,=1, «,=0, o5=—1; B,=0, 6. =—2;
B)=0. The Milne method (9.18) is given by (9.39) with p=2; a,=1, a, =0,
&) = —1; B, =4, B; =4, By =3. The implicit Adams method (9.37) is given by
(9,39) with, p=3; ¢,—1, o,——I, o,—0,.0,=0; 8.=9/24 (p,—19) 24.
B, = —5/24, By=1/24, while the explicit Adams method (9.38) is given by
(9.39) with p=3; «,=1, o,=—1, «,=0, a =0; £;=0, B,=23/12,
B,=—16/12, By=5/12. The difference equation (9.39) is said to define a
multi-step method. The integer p is called the rank of the method. An explicit
method is characterized by the condition £,,=0.
It is convenient to define the two characteristic polynomials
p=5 af! and a(0)=>Bi

Thus, for the Euler method (9.6),
pli —oe a rand) Sol)=r

For the modified Euler method (9.13),
p()=C?-1 and a ()=2L.
For the Milne method (9.18),
p(j=C?—-1 and =o (0) = 3(0? +46 +1).
For the implicit Adams method (9.37),
pd=C-2 and 6 (=
(90° +190? 50 +1).
For the explicit Adams method (9.38),
p(j=c?-(? and =o (C) = 7(2367-160
+5).
Two basic results, for whose proofs we again refer the reader to [8], are
the following criteria for determining whether a method is stable and con-
sistent.
Theorem 2. The multi-step method (9.39) is stable if and only if all roots of
the polynomial equation p(¢)=0 satisfy |\¢|<1 and those roots with absolute
value | are simple roots.
Theorem 3. The multi-step method (9.39) is consistent ifand only if p(1)=0
(that is, 1 is a root of p(¢)=0) and p'(1)=a(I).
Exercises
2. Verify that the methods (9.6), (9.13), (9.18), (9.37), and (9.38) are all both stable
and consistent.
3. Is the method (9.16) obtained by means of the trapezoidal rule stable and con-
sistent?
If a method is both stable and consistent, then =1 is a root of p(¢)=0.

If there are other roots of p(f)=0 with absolute value 1, as happens with the
modified Euler and Milne methods but not with the Euler and Adams
methods, then another error may appear. There may be an error in the
initial value or starting value which tends to grow larger and, if the actual
solution tends to zero as t becomes large, this error may eventually
become large compared to the solution. This phenomenon is known as
numerical instability or conditional instability. The possibility of numerical
instability is the reason why the Milne method is not always suitable, despite
its small truncation error. What may happen is that the Milne method, when
applied to a differential equation such as y’=y one of whose solutions is a
negative exponential, may give rise to a difference equation approximation
whose solution includes a term which grows like a positive exponential.
Then, as the solution becomes small, the numerical approximation may
grow.
We have devoted our attention mostly to the truncation error and have
given very little discussion of the round-off error. The study of the round-off
error is largely statistical and quite difficult. We make no attempt to carry
out such a study, but content ourselves with a few general remarks on the
relationship between truncation error and round-off error.
We may reduce the truncation error in a given problem by using a more
refined approximation method or by using a smaller step-length. In either
case, the number of computations becomes larger, and we must expect the
round-off error to increase. If the round-off error increases faster than the
truncation error decreases, then the accuracy of the final result is not im-
proved, and the additional effort is wasted. In practice, the most efficient
procedure is usually to choose the step-length so that the round-off error and
truncation error are roughly the same size.
oS Runge-Kutta Methods 385
9.5 RUNGE-KUTTA METHODS
In Section 6.1 we considered the possibility of finding a power series ex-

pansion for a solution of a nonlinear differential equation. At that time, we
discarded this method because of computational difficulties and because of
the difficulty of proving convergence of the series obtained. Now, however,
let us return to this idea with a slightly different point of view. Instead of
seeking a power series expansion, we use the first few terms of the series
expansion to obtain a numerical approximation. Also, we divide the interval
under consideration into subintervals of length h, just as we have done in
developing finite difference methods in the preceding four sections.
Let ¢@ be the solution of the differential equation
y=f(t, y) (9.1)
which satisfies the initial condition
$(to)=Yo-
If we assume that ¢ has a continuous second derivative on the interval
[to, to +J], then we can use Taylor’s theorem to write
Geta
P(t)=P(to)
+(t— Lo) b' (to)+ 2)
¢" (6) (9.40)
where tp <&<t. If t; =t) +h, this becomes
h? ”
(1) = (lo) +h (to) +55 p"(¢).

Using (9.1) and the initial condition $(to)=yo, we have
an
b(t1)=Yo thf (to, yo)t5? (¢). (9.41)
The idea behind our present approach is to obtain the approximation
y, to d(t,) by neglecting the term (h7/2!) $”(€) in (9.41). Thus,
Yi=YotMf (to, Yo):
If we divide the interval [to, f9+7] into N subintervals of length h by
defining the partition points t,=tf)+kh(k=0, 1,..., MN), we can use this
procedure to obtain an iterative formula
Vnr1 =Ve thf (tee Ye) (9.42)
with local truncation error (h?/2!) ” (€). The formula (9.42) is, of course, the
Euler method which we have already studied in Section 9.1.
Now let us use, instead of (9.40), Taylor’s theorem with three terms plus
an error term. Then
$()=(to) +(-t0) 4 (to) + —- O"(to)+ GH". 9.43)

We can evaluate $"(t,) by differentiating

b (=f(t,(9)
obtaining
b' (D=h(t A) tHe OO) & (D=h( OO) +H(4 6) £(6 6).
Proceeding in the same way as above, we obtain the approximation formula
h7
Yat =Vathf (te Vid) NH [fi(tes Yi) +L (tee Yu) F (ts Ved] (9.44)
with local truncation error (h°/3!) @”’(é). The method (9.44) is a plausible
means of obtaining numerical approximations. However, it suffers from the
disadvantage that its use requires the calculation of derivatives of f,We could
develop analogous procedures using higher-order Taylor approximations,
but these would require the calculation of higher-order derivatives of f-
The Runge-Kutta procedure is an attempt to obtain formulas equivalent
to Taylor approximations which do not involve derivatives of f. The most
frequently used Runge-Kutta formula is
h
Yar = Vere (Pi + 2P2 +23 + Pa) (9.45)
where
h hp
Pi=S (te Yi) p= S(tr5 nett)
h h
r= S(t nate), Da ies y+ hp3).
The term ¢(p; +2p,+2p3+ 4) represents an ‘“‘average”’ slope of ¢ over

the interval [t,, t,+,]. The term p, is the slope at ¢,. The term p, is an
approximation to the slope at the midpoint of the interval obtained by
means of the Euler method. The term p,; is a second approximation to
the slope at the midpoint. The term p, is an approximation to the slope
at t,,, obtained by means of the Euler method with slope p3. The formula
(9.45) is an explicit one-step procedure, whose application does not require
the calculation of any derivatives of f Since it requires an evaluation of f
at four different points, it is still rather cumbersome. However, it is
equivalent to a five-term Taylor formula and can be shown to have local
truncation error no greater than a constant multiple of h°. We omit the proof
of this fact, because it is quite involved technically. However, we shall
suggest the idea behind the proof in Example | below, where we obtain a
Runge-Kutta formula equivalent to a three-term Taylor formula.
Because the Runge-Kutta formula (9.45) is rather cumbersome, it is
seldom used by itself. Its chief value is as a starting method to obtain
starting values before changing to a multi-step method such as the Milne
oS Runge-Kutta Methods 387
method. We recall that before we can use the Milne method (9.18), we
need not only the given initial value yg but also the value y,. The deter-
mination of y, is usually carried out by a Runge-Kutta method.
The fact that the Runge-Kutta formula (9.45) is rather cumbersome may
be illustrated by the complexity of its flow diagram (Fig. 9.7).
This diagram has even been simplified slightly by the suppression of the
notation a, =+, a,=4, a, =4, a,=$, «, =0, a, =4, a,=4, «7, =1. In an actual
calculation, this information would also have to be fed into the computer.
Read in h, yo, to, f + T
bt
Yo Yr
&£+hot,
yr t h®->y,
f@, 1) Pp
@+ap—®
4,+ah—or
Ye + ah, > 7
Record &, yx
Figure 9.7
Example 1. Let us demonstrate how to obtain a Runge-Kutta formula equivalent to

a three-term Taylor formula which agrees with the formula (9.44) except for a term
no greater than a constant multiple of h*. We are seeking a method of the form
Ver1=Yethl[af (tes Ve) + bf (e+ ah, y+ Bhf (te. Vx))] (9.46)
where a, b, «, B are constants to be determined. By Taylor’s theorem for functions of

two variables we may write
f (tet ah, Ve+ BMS(tesVi)=S (the Va) + HAL (tis Va)

+BME (ths Ye) Sy(ths Ye)+RH? (9.47)
where R is a remainder term involving the second-order partial derivatives of f, When
we substitute (9.47) into (9.46) we obtain
Veer =Veth(atd) Sf(ths Ya) +N? [abhi
(tesVe) + BOS (tes Yu) Sy(ter Ya)]+ ROM? (9.48)
Comparing (9.48) with (9.44), we observe that if the constants a, b, a, B are chosen so that
a+b=1, ab=>, pb=3, (9.49)
then the approximations obtained by (9.44) and (9.48) differ only by the term RbhA?.
Thus (9.46), with the constants satisfying (9.49), is a Runge-Kutta formula of the
desired type. For example, we may choose a=3, b=3, «=1, B=1, to obtain
h
Vert Vers LS (tes Ve) AS (tet1sVet
hf(tes Yed))J (9.50)
as a Runge-Kutta method equivalent in accuracy to a three-term Taylor formula. The

reader should observe that (9.50) is equivalent to the improved Euler method (9.16),
with the Euler method used as a predictor.
Exercises
1. Use the Runge-Kutta method (9.50) with h=0.2 to approximate #(1)=e, where
¢ is the solution of y’=y such that #(0)=1.
2. Use the Runge-Kutta method (9.45) with h=0.2 to obtain an approximation to
the same value $(1)=e treated in Exercise 1. Compare the accuracy and labor
involved with those of previous attempts to solve the same problem: Example 1,
Section 9.1 (Euler method), Example 1, Section 9.2 (modified Euler method), and
Example 1, Section 9.3 (Milne method).
3. Use the Runge-Kutta method (9.45) to obtain a starting value #(0.1) for the solu-
tion ¢ of the equation y’=r+ y’/? with each of the following initial conditions
a) $(0)=0
b) $(0)=1
c)$(0)=10
d) $(0)=100
4. Obtain an approximation to ¢(0.1), where @ is the solution e' of y’=y such that
~(0)=1 by using
a) the Runge-Kutta method (9.45)
b) the Runge-Kutta method (9.50)
How small must / be taken to obtain approximations of comparable accuracy by
using the Euler method with a subdivision of the interval 0<1<0.1?
5. For each of the following initial-value problems obtain an approximate value of

(1), where $(¢) is the (exact) solution, using step size h=0, | by (i) the Euler method,
9.6 Systems and Equations of Higher Order 389
(ii) the modified Euler method, (iii) the Milne method, using the Runge-Kutta
method for starting values and predictor-corrector formulas, (iv) the Runge-
Kutta method.
a) y'—4y=1-1, $(0)=1
b) Y= +y’, (0)=0 (compare Exercise 19, Section 8.1)
c) y =e, (0)=0 [Hint: Use a table of exponentials. ]
6. Repeat the calculations of Exercise 5, but with step size h=0.05, and compare the
results with those of Exercise |. [Suggestion: Do not attempt this exercise unless you
have a high-speed computer available or are unusually eager to do hard computing. |
9.6 NUMERICAL METHODS FOR SYSTEMS AND

EQUATIONS OF HIGHER ORDER
Everything that we have done in this chapter has been for first-order
equations, but the methods are equally suitable for systems of differential
equations, although not entirely without difficulty. We can write a system
yi =f; (t, Vises Yn)
y. =f. 2(t,{,y V1 seeeo Vn ) (9.51)
Vn = DAG Vioeees Yn)
in the form
y'=f(t, y) (9.52)
with y the column vector with components (j,,..., y,) and f(t, y) the column
vector with components (f;(1, ¥1,.-+. Yn), folly Vis =+->Vn)s
1-9 Falls Vio+-+> Va)
(see Section 4.5). We can apply the approximation methods developed in
this chapter to the system (9.51) by applying them to each component of the
vector equation (9.52).
u'=0v, v’=g(t, u, v). (9.53)
This can be written in the form (9.52) with
(1). (oes)
The Euler method applied to (9.53) leads to a pair of iterative formulas
Ug+1 =U, + hv, (9.54)
Ve 1 =U +NG (ths Uns Vx)
For a system (9.53) we are given initial values v9 and vo, and once we have found both
u, and v,, we can use (9.54) to compute y+, and % +1.
Example | illustrates how finite difference methods can be adapted to

systems. Since an equation of order higher than the first can always be
written as a system (Section 4.1), such equations can also be treated

numerically. Note, for example, that the system (9.53) is equivalent to the
second-order equation u” =g(t, u, uv’). Every method and theorem presented
for a single equation has an analog for systems, though the analog is not
always obvious. For a complete discussion, we must again refer you to one of
the more thorough treatments of the subject. ‘
Exercises
1. Find iterative formulas analogous to the modified Euler and Milne methods for
the system (9.53).
2. Use the Euler method with h=0.1 to approximate the value $(z/2) of the solution
@ of the second-order equation y” + y=0 such that $(0)=0. (Note that $(¢)=sin¢,
so that the exact value of (7/2) is 1.)
3. Use the Euler method with h=0.1 to estimate the smallest positive value t at which
the solution wy of y’ + y=0 such that w(0)=1 vanishes. Use your answer to estimate
a value of z, and suggest some ways of obtaining a more precise estimate. [Hint:
Since y/(t)=cost, the exact value of t is 2/2.]
4. Consider the solution ¢ of the differential equation y’+(g/L) y=0 such that
$(0)=5, $’(0)=0, where g, L, 0) are given positive constants. (This is one model
for the simple pendulum obtained in Section 2.2.) Use the Euler method with h=0.1
to estimate the value of #’(t), where t is the first value of ¢for which ¢(t)=0. Also
estimate 7, the first value of t>0 for which $(t)=@). Estimate the same
quantities using the Milne method with h=0.1.
5. Repeat the calculations of Exercise 4 using the more precise model
y"+(g/L) sin y=0.
6. Find iterative formulas analogous to the Euler, modified Euler, Milne, and Runge-
Kutta methods for the system
yi =i (t, ¥1, Y2)
Y2=Sr(t V1» Y2).
7. Specialize the results of Exercise 5 to the linear system
Vi =411(t) V1 +4 2(t) v2 +b, (t)
V2 = 421 (t) V1 +492(t) v2 +bo(t).
8. Generalize the results of Exercises 5 and 6 to n-dimensional systems of the form
y'=f(¢, y)
and
y =A(t) y+b(t).
9. For each of the following initial-value problems obtain an approximate value of
(1)=(¢1(1), 62 (1)), where (t) is the (exact) solution, using step size h=0.1 by (i)
the Euler method, (ii) the modified Euler method, (iii) the Milne method, using the
Runge-Kutta method for starting values and predictor-corrector formulas, (iv) the
Runge-Kutta method.
9.6 Systems and Equations of Higher Order 391
a) Vj =2y, +ty2 ¥1(0)=1

Yo=ty; ¥2(0)=0
b) yi = Be | +2y> = l Vi (0)=0
Yo= —2),+3y2 y2(0)=1
10. For each of the following initial-value problems, obtain an approximate value of
(0.5), where $(r) is the (exact) solution, using step size h=0.1 by (i) the Euler
method, (11) the Runge-Kutta method.
a) y”—2y?=0 $(0)=1, ¢'(0)=—-1
b) y’+2y?=0 o(0)=1, ¢’(0)=-1
[Hint: Reduce to a system of two first-order equations and the results of
Exercise 3.]
CHAPTER 10
The Laplace Transform
In this chapter we shall study a method of solving initial-value problems

for linear differential equations and linear systems of differential equations.
This method does not enable us to solve any problems in differential equa-
tions which we could not solve by the methods already studied in Chapters
3, 4, and 5. It does, however, provide a simple and often employed technique
for solving many problems which arise in applications, particularly those
involving linear equations with constant coefficients. One reason for its use-
fulness is that it enables us to find the particular solution of the differential
equation which satisfies the initial conditions directly, rather than first
finding the general solution and then using the initial conditions to deter-
mine constants.
10.1 INTRODUCTION
For every function f(t) of a suitable class, we define the Laplace transform
F(s), also denoted by #(/f), by
oo) A
Fi)= |e*s0 dt= lim ferro dt, (10.1)

A>
0
where it is understood that we restrict our considerations to those values of

the complex parameter s and to those functions f for which the above limit
exists.
Example 1. To calculate the Laplace transform of the constant function 1, we evaluate

(oo) A
, = 1 e SA
é-* d= lim |e" = dt= in| == = | (As>0). (10.2)
A> A> oo RY Ss
0 t@)
392
10.2 Basic Properties of the Laplace Transform 393
Thus #(1)=1/s. Clearly, the integral does not converge for @s <0.
Example 2. We can calculate the Laplace transform of e*' by almost exactly the same
process as that used in Example 1.
: ae 1
#(et)= [enter dt= [eons di= (As>z). (10.3)
S—Zz
0 0
Notice that the only difference between the integrals evaluated here and in Example 1
is that where there was an s in Example 1, there is an (s—z) in Example 2. This is an ex-
ample of a general principle to which we shall return later.
The operator defined by (10.1) is linear in the sense that if f,(¢) and f(t)
have Laplace transforms F,(s) and F,(s), respectively, and if a and b are
constants, then af; (t)+ bf, (¢) has Laplace transform aF, (s)+bF;(s).
The idea behind the Laplace transform method is very simple. It will be
shown that every solution of any linear homogeneous differential equation
with constant coefficients has a Laplace transform. Also, the Laplace
transform of the derivatives of fcan be expressed in terms of the Laplace
transform of f and the value of f and its derivatives at t=0. This then
means that if ¢ is the solution of a linear differential equation with constant
coefficeints which satisfies some given initial conditions at t=0, the Laplace
transform of ¢ satisfies a linear algebraic equation rather than a differential
equation. When we have solved this algebraic equation, we need only find
the function whose Laplace transform is the solution of this algebraic
equation. This may often be facilitated by tables of Laplace transforms.
Of course, in order to be sure that the function found by this procedure is the
same as the function ¢, we need a uniqueness theorem for Laplace trans-
forms, to the effect that two different functions cannot have the same Laplace
transform.
10.2 BASIC PROPERTIES OF THE LAPLACE TRANSFORM
In defining the Laplace transform by Eq. (10.1), we must impose some con-
ditions on the function f(t) to assure the convergence of the integral. We
consider functions f(t) defined for 0<t<oo which grow sufficiently slowly
near ¢=0 to assure the convergence of the integral at zero (the integral may
be improper); we also require that the function f(t) grows sufficiently slowly
for large ¢ to assure the convergence of the infinite integral for some values
of the complex parameter s; finally, we require f(t) to be integrable over every
closed subinterval of 0<t< oo. This leads us to the following definition.
Definition 1. A function f on 0<t<oo is said to be of exponential growth
at infinity if it satisfies an inequality of the form
If (Ql< Me", (10.4)

394 The Laplace Transform
jor some real constants M>0 and c, for all sufficiently large t.
Now, we define the class A of functions on 0<t<oo which are
i) absolutely integrable at zero (that is, lims.o+ | |f(0)| dt exists for
sufficiently small a>0);
lil) piecewise continuous on 0<t< oo;
lii) of exponential growth at infinity.
This is the class of functions for which we wish to define the Laplace trans-
form. Clearly, the functions 1, ft, 7" (n a positive integer), sint, cost, e*' for
any complex z are in the class A, but exp/?? is not.
Theorem 1. If f is a function in the class A, the integral \f e~™ f(t) dt
converges absolutely for all complex numbers s with sufficiently large real part.
Proof. For small t, |e~*'| is bounded, and therefore the assumption thatf
is absolutely integrable at zero implies the convergence of the integral
{2 |e“(0 dt for every 6>0. If 2s=c, |e “|=e%, and then (10.4) yields
le“ f (t)| <Mer Coe
for ¢>7, where T is some number greater than zero. Therefore, if ¢>c, the
infinite integral [Ff |e “f(d)| dt converges because its integrand decreases
exponentially to zero, and we have
(c)
M
fens (t) dt < [ier dt<M [ee Rate ene
o—c
T it
Finally, because
f ispiecewise continuous on 6<1<T, {jf |e “f(t)| dt exists,
and this completes the proof of the theorem. |
The Laplace transform F(s) of any function f(t) in the class A is defined
by Eq. (10.1) for all complex s with sufficiently large real part; we will usually
be concerned only with real values of s. Sometimes we denote the Laplace
transform by the symbol #, to emphasize that the Laplace transform is an
operator, which associates the function F(s) with the function f(r). Thus, we
write
F(s)=
2if (}.
As already remarked, this operator is linear.
Example 1. If f(¢) is a complex-valued function of the class A, f(¢)=u(t)+iv(t), where
u, v are real, we have
LYS Q=L {u(t + wn} =f {u()} + iF {v(}.
From the definition, it is clear that the Laplace transform of a real-valued function is
real for real s. Thus, # {u(¢)} is the real part of #{f(t)} and # {v(t)} is the imaginary
part of f {f(t}. If z=a+if, and if f(t)=e, we have
E ' 1
L (e*)= FL fee" = Y {e* cos Bt+ ie" sin Bt} = ee
! 1 s—a+ip s—a+tif
(10.5)
~s—a—if s—a—if s—atiPp (s—a)?+p?
When we take real and imaginary parts of (10.5) with s real, and then let s be complex
again, we obtain
sS—o
Se COS ph =e,
as Te tees), (10.6)
Ae =P
In particular, taking «=0, we obtain
s : B
= Pes aap . SB ae ea (As>0). (10.7)
We could use the same direct approach to compute the Laplace trans-
forms of other functions such as ¢* and re’, but we can calculate these trans-
forms less laboriously by using some additional general properties of
Laplace transforms.
Suppose fis in the class A. If we differentiate (10.1) with respect to s under
the integral sign, we obtain formally
[e@)
F(s)= ~ [erty dt. (10.8)

0
Since fis of exponential growth at infinity, satisfying the bound (10.4),

tes"
f(Dix Mte> enor
where o=&s, and the integral (10.8) converges absolutely for ¢>c. From
this we can prove that the relation (10.8) is valid if the real part of s is
sufficiently large, that is, if 2s>c. In fact, we can use the same argument to
justify repeated differentiation under the integral sign, which gives the
formula oO
FY (s\=(—1)" [ter dt © (ee): (10.9)

0
Comparing (10.9) with the definition of the Laplace transform, we see that we
have sketched the proof of the following result.
Theorem 2. The Laplace transform of a function f in the class A has
derivatives of all orders and these derivatives are given by (10.9). If f(t) belongs
to the class A, then t*f(t) also belongs to the class A for every positive integer
k, and its Laplace transform is given by
Sef=(-I GLO} (K=L2.). (10.10

, k
If we apply (10.10) with f(t)=1 and use (10.2), we obtain
=(-0S ()=z5 (As>0;k=1,2,...). (10.11)

If we apply (10.10) with f(t)=e*' and use (10.3), we obtain
‘ eg k!
Lite =(y a (+)-—n (As> Aik = 12). (L002)
By letting z=a+if and taking real and imaginary parts, we can obtain
the formulas
kK!@[(s—a)+ip]**?
ees cos Bt} = (ane ae ?
; (10.13)
yyy EF ls aye ip
2 ire sin
pi = [e—a?4 pee
LAE giaan
a CAG ect ye au
ee (10.14)
By evaluating the indicated real and imaginary parts in (10.13) and (10.14),
we can obtain explicit expressions for these Laplace transforms.
Exercises
1. Find the Laplace transforms of ¢ cosft and ¢ sin Bt, using (10.14).
2. Find the Laplace transforms of t cos ft and t¢sin ft directly from the definition.
3. Calculate the Laplace transforms of
a) cos* Br. [Hint: Use the half-angle formula. }
b) sin Bt cos Br.
4. Calculate the Laplace transform of the function f given by
0) Oat<1)
r= (1<t<2)
On eS2):
5. Calculate the Laplace transform of the function f given by
f= sin 2t
0
reeew.
O<t<nx
(t>z).
The relation between (10.2) and (10.3) suggests that the multiplication of a
function by an exponential does not affect its Laplace transform except for
causing a translation of the independent variable. This is in fact a general
property.
Theorem 3. If the function f in the class A has Laplace transform F, then
the Laplace transform of e'f(t), for any constant a (real or complex) is
F(s—a).
Proof. It is easy to verify that e“f(t) also belongs to the class A. The
only part of the verification which is not completely obvious is that e“f(t)
is of exponential growth at infinity. If / satisfies (10.4), and if «=a, then
lef (0) <e*Met= Me®@tt
and thus e“f(t) is of exponential growth at infinity and has a Laplace trans-
form, which we may calculate directly. We obtain
ie.@)
Lie i (i= |e Fei (t)dt= eS

9 f(t) dt=F (s—a),
0 ee
and the result is proved. I
Exercises
6. Use Theorem 3 to derive (10.6) from (10.7).

7. Use Theorem 3 to derive (10.13) from (10.14).
8. Using the results of Exercise 1, find the Laplace transforms of te“ cosft and
te™ sin Bt.
9. Let f(t)=(t) for 0<t<a, and let f be periodic with period a, so that f(t+a)=/(t)
for 0<t<oo. Show that
‘ 6e “d(t) dt
gs Be ee (2s>0).
[Hint: Write
LAF (Qp=foe “f(t dt+fare“ f()dt+--,
and transform each integral so that the range of integration is [0, a].]
10. Let f be the ‘square-wave function” given by f(d))=1 (0<t<a/2), f(j=—-1
(a/2<t<a) with f periodic of period a. Sketch the graph and find the Laplace trans-
form of f.
11. Let f be the “square-wave function” of Exercise 10 and let g(¢)=fpf(s) ds. Sketch
the graph and find the Laplace transform of g.
12. Find the Laplace transform of each of the following functions:
sin wt, O0<t<n/o
On t/w<t<2n/w
and f(t+2n/w)= f(t) for0<t<o.
*b) [> e-“ sin(t—u) du
*c) PP? [Hint: 2 ('?)=F
GB/2) s 7/*euse Theorem 2.]
d) ¢? sin ot.
The usefulness of the Laplace transform in finding solutions of a dif-
ferential equation depends on the fact that the Laplace transform of the
derivative of a function can be expressed easily in ferms of the Laplace
transform of the function,
Theorem 4. Let f be a function in the class A whose derivative also

belongs to the class A, and let the Laplace transform of fbe F. Then
Lf ()} =5F ()—F (0). (10.15)
Proof. We simply apply the definition of the Laplace transform and
integrate by parts. This gives
oe) A
visoj=|[e*r dt= lim foro dt
= lim Jesrco}s+ [se*r |

0)
co
=—f(0)+s |ef (t)dt=sF(s)—f

(0),
where we have used the fact that lim,.,, e “4f(A)=0 for &s sufficiently
large. |
The result of Theorem 4 can easily be extended to derivatives of higher
order. It is convenient to introduce some additional notation. For each
positive integer k, we define A* to be the class of functions in A which have
continuous derivatives up to order k on (0, 00) and whose derivatives up to order
k also belong to A. Thus, the hypothesis of Theorem 4 is that f belongs to
the class A’.
Theorem 5. If f belongs to the class A‘ for some positive integer k, and if F
is the Laplace transform of f, then
L{f()} =F (=f
0) 91-2 Q)=- =f (j=1,2,...,k).
(0) =F) (10.16)
Proof. We can prove (10.16) most easily by induction on j for any fixed k,
remembering that the induction procedure cannot be carried out for j>k
since the hypotheses do not guarantee the existence of the Laplace trans-
forms of derivatives of order higher than k. The case j=1 of (10.16) was
proved in Theorem 4. If (10.16) has been established for f(t), we can write
fi*) (0) as the first derivative of f(t) and apply Theorem 4. We obtain
FE Vie) a0)
=s[s/F(s)—s/~1f (0)— += f9"Y(0)] — f(0)
=r 'F(s)—s/f (0)— sf U- Y(0)—f (0),
and this is Eq. (10.16) with jreplaced by (j+ 1). Thus, Theorem 5 is proved by
induction. I
The formula (10.16) for calculating the Laplace transforms of the

derivatives of a function is the key to the solution of linear differential
equations by means of Laplace transforms. We shall study this subject in
more detail in Sections 10.4 and 10.5. However, we give some simple ex-
amples here which we have already solved in Chapter | to illustrate the
main idea.
Example 2. Let us find the solution ¢ of the familiar first-order differential equation
y +ay=0, (10.17)
$o(0)=yo, (10.18)
where a and yo are given constants. From our previous solution in Section 1.4 we know
that the solution ¢p is in the class A'. Thus, we should be able to find ¢, by Laplace
transforms. Let Yo(s)=#(o). Using (10.15), we may take the Laplace transform of
every term in the equation
fo(t)+ago(t)=0,
satisfied by @o, and we obtain
SYo(s)—$o(0)+aY¥(s)=0 or (sta) Yo(s)=yo,
using (10.18). This yields
¥o(s) ane
sta
and the only remaining problem is to find a function which has this expression as its
Laplace transform. As we have seen in Example 2, Section 10.1, ype ™ is such a func-
tion. By direct verification and application of Theorem 1, Section 1.6, we know that
it is the only solution of Eq. (10.17) satisfying (10.18). However, as motivation for more
complicated problems, it is useful to look at this a little differently. At this stage, we
do not know that it is the only such function, but if we accept the truth of the statement
that two different continuous functions cannot have the same Laplace transform, then
we conclude that ¢9(t)=yoe “. Since ype“ does belong to the class A’, our reasoning
is valid, except for the uniqueness statement which will be stated precisely in Section 10.3.
Example 3. Now, let us apply the same method to find the solution ¢ of the non-
homogeneous equation
y tay=f (i), (10.19)

$(0)=Yo. (10.20)
Here
/ is a given function belonging to the class A. The problem in (10.19) and (10.20)
can be solved by the method of Section 1.4, and from the solution it is obvious that if
fis in the class A, the solution ¢ is in the class A’. We proceed as in Example 2; we may
assume that ¢ belongs to the class A’ and we let Y(s) be the Laplace transform of ¢
and F(s) the Laplace transform of f. Using (10.15) we take the Laplace transform of
every term in the equation
satisfied by @, and we obtain

sY(s)—(0)+aY¥(s)=F(s) or (sta) Y(s)=yo+F(s),
using (10.20). This yields
F
Von)
Sa) Sd
We must now find a function which has this expression as its Laplace transform. We
have already seen in Example 2 that y)e “ has Laplace transform yo/(s+a), but we
have no method as yet of finding a function whose Laplace transform is F(s)/(s+a).
This suggests that we will need a method of finding a function whose Laplace transform
is the product of two given functions. In this case the given functions are F(s) and
1/(s+a), which are the Laplace transforms of f(t) and e -“, respectively. This problem
will be studied in Section 10.3. For the present, we remark that the method of Section
1.4 gives
t
Bl)=yoere+
|mo” Fu)du,
10)
and this suggests that |) e ““~ f(u) du should have the Laplace transform F(s)/(s+a).
Example 4. In Examples 2 and 3 we have used the Laplace transform to solve linear
differential equations with constant coefficients. To show that the method is less
useful for linear equations with variable coefficients, let us attempt to find the solution
W of the equation
y' +2ty=sint, (10.21)
V.O)=yo- (10.22)
In Example 2, Section 1.4, we found that
t
W(t)=Yo exp(—t?)+exp(—t’) Jew sins ds.

0
When we try to obtain this solution by the use of Laplace transforms, we meet with
serious difficulties. If we let Z(s) be the Laplace transform of w and take the Laplace
transform of every term in the equation
y (t)+ 2ty(t)=sint,
satisfied by w, we apply (10.10) and (10.15) and we obtain
sZ(s)— (0) —2Z'(s)= (10.23)

s*+1
While in Examples 2 and 3 we obtained an algebraic equation for the Laplace transform
of the desired solution (and thus simplified the problem), here we have a differential
equation for the transform of the solution which is no simpler than the original
problem. This example suggests that ihe usefulness of the Laplace transform method is
limited mainly to equations with constant coefficients.
10.3 The Inverse Transform 401
10.3 THE INVERSE TRANSFORM
The examples of the previous section have indicated the need for finding a
function with a given Laplace transform. The function f(t) whose Laplace
transform is F(s) is called the inverse Laplace transform of F(s). The inverse
Laplace transform is, as we shall prove later in this section, linear. We must
consider the following questions:
i) If we know that F(s) is the Laplace transform of a function f(¢), how
can we compute the inverse transform
/ from a knowledge of F’?
il) Is the inverse transform of a given function F unique?
An answer to the first question can be given in a theoretical way by means
of the so-called complex inversion formula (see, for example, [20], p. 66).
However, the derivation and application of this formula require a knowledge
of real and complex analysis. Therefore, we confine ourselves to a more
elementary approach that will enable us to find the inverse transforms of
some functions commonly arising in applications.
We have seen in Example 3, Section 10.2, that it would be useful to have a
general method of calculating the inverse Laplace transform of a product
of two functions each of whose inverse Laplace transforms are known. Let
us now determine whether there is a means for doing this.
We assume that we are given F(s) and G(s), and that we can find functions
f(t) and g(t) in the class A whose Laplace transforms are F(s) and G(s),
respectively; thus F(s)=Y#(f(t)) and G(s) =f (g(t). Our problem is to deter-
mine the function h(t) whose Laplace transform is the product F(s) G(s), if
such a function exists. If there is such a function, then
ih) = |e “h(t) dt
0 (10.24)
=F {s) a()=| “ro du [ate dv,
0)
where &s>o for some real number o. If F(s)=L {f(t} for As>a, and
G(s)= Lf {g(t)} for s>B (a, B real), then o=max(c, f). Since each integral
converges absolutely for #s>oa, we may write the product of the two
integrals on the right side of (10.24) as a double integral, obtaining
|e “h(t) =| |e+ F(y) g(v) du dv, (10.25)

0 0 0
for As>o.
We write this as an iterated integral,

00 00 00
[emmy a= |ot) |ferent audv ;
this can be justified under our hypotheses because of the absolute conver-
gence noted above. Making the change of variable u+v=t in the inner
integral, we obtain
fo.@) oO
fewny di— ato) {ero) at|

8
8ot
=|er Ef
re—nae tv|dt (&s>c), (10.26)
where the interchange in the order of integration can again be justified.
Exercise
1. Obtain the limits of integration in (10.26). [Hint: Draw a sketch of the region of
integration.|
Equation (10.26) states ¥ {h(t)} =L []o f (t—v) g(v) dv], which suggests
that the solution to our problem is
m= |ree) a dv. (10.27)
By reversing the argument we have just completed, we may prove that the
Laplace transform of the function A(t) defined by (10.27) is F(s) G(s), as
desired.
Exercise
2. Show that iffand g belong to the class A and / is defined by (10.27), then
oO
e “h(t) ir=| f(uje~™ du |a0 e * dv.

8
ot 0 0
We may summarize what we have sketched in the following theorem.

Theorem 1. Let f(t), g(t) belong to the class A and let F(s)=L {f(t},
G(s)=2{g(t)}. Then [fs flt—v) g(v) do] =F (s) G(s)
The function / defined by (10.27) is called the convolution of fand g, and is
sometimes denoted by h=fxg to indicate that the convolution of two
functions resembles a product in many ways.
Exercises
3. Show that f*g=g*f for any two functions f, g in the class A.
. Show that (/*g)*h =/*(g*h) for any three functions /, g, h in the class A.
. Show that /*0=0 for any function f in the class A.
&
Nm. Show that if fand g are defined for — 00 <t<oo but are both identically zero for
t<0, then their convolution /*g can be written as [_% f(t—v) g(v) dv.
=
7. Show in two ways that ¥ {{§ f(t) dt} a ATE)
ire
The convolution integral in Exercise 6 is sometimes used to define an

operation analogous to multiplication for certain classes of functions. The
reader should be careful not to carry the analogy too far; for example, it is
not true that f*l=/, since for an arbitrary function f we have
|tes is=| Fo do#f (t).
The properties of the convolution make it possible for us to calculate some

inverse Laplace transforms which would otherwise be difficult or impossible
to find.
Example 1. Find the inverse Laplace transform of
1s? 1)=[1(s—1)] [1s+ 1).

Since #~' {1/(s—1)}=e', and Y~*{1/(s+1)}=e', we see that
t t
e(3 i)=[erst dune! | et

se
0 0
There is another way to calculate this particular inverse transform. If we decompose

1/(s* —1) into partial fractions, we find
1 1 (= 4)
24 IGS) (+4) 2\s—1 os 1’
2,
and thus, by the linearity of Y~' (this will be proved following the corollary to Theorem
|
Gale |
1
Sy Cece:
Example 2. Find the inverse Laplace transform of 1/[s?(s*+1)]. We have seen, in

Eq. (10.11), that Z~1(1/(s?))=2, and in Eq. (10.7), Y~*(I/(s?+1))=sint. Now, by
Theorem 1, #~ *(1/s?(s?+1)) is the convolution of tand sin t, which is
io t t
[c-w sin u dua |sinu du |u sinu du

1) 10) 0
=t—tcost+tcost—sint
=t—sint.
We may also obtain this result by partial fractions. For
1 1 1
(al) Set
and hence, by either method
ges : =f} gt =t—sint

s#(s7+1)) 2 Ser Oe oa
The above examples suggest that if we wish to find the inverse transform of
a rational function N(s)/D(s), where N and D are polynomials with the
degree of N less than the degree of D, then we can take advantage of partial
functions to decompose N(s)/D(s) into a sum of terms whose inverse
transform is easily found.
Exercises
8. Find the inverse Laplace transform of each of the following functions. (We indicate
the partial fraction decomposition for the convenience of the reader.)
s?—6 al) oO
a) +4s7+35 5 ue 1 eee
b) 1 = A FeB i Cs+D
Sis Less es
16 A Bs+C Ds+E
c) s(s?+4)? ager To 44e
BE
d) a where f(t) is in the class A and ¥ {f(1)} =F(s).
sta
€
) F(s) ; where f(t) is in the class A and # {f(t)}=F(s).
s?+1
9. Find the solution ¢ of the integral equation
t
(+ [tu o(u) du=1.

0
[Hint: Take the Laplace transform of every term and use Theorem 1; then solve for
L£(¢) and finally find ¢.]
This is as far as we can go with systematic techniques for constructing

inverse transforms, without using the inversion theorem. Beyond this the
most useful aids in finding inverse transforms are a good table of Laplace
transforms (a brief one may be found at the end of the chapter), experience,
and luck.
Let us now turn to question (ii), namely, the uniqueness of the inverse
transform. We state the following basic result without proof—a proof may
be found in [20].
Theorem 2. Let f,, f, be continuous on 0<t<co and belong to the class A,

and let their Laplace transforms F,(s) and F,(s), respectively, exist for As>o.
If F,(s)=F,(s)
for real s>o, then f,(t)=f,(t) on 0<t<oo.
Theorem 2 justifies the method used in Examples 2 and 3, Section 10.2,
where we found the Laplace transform of an unknown function and then
found this unknown function by taking the inverse transform. If the inverse
transform were not unique, we would not know that the process of taking the
inverse transform returns us to the original unknown function.
Another consequence of Theorem 2 is the linearity of the inverse Laplace
transform: If the continuous functions /; (¢) and f,(¢) in the class A are the
inverse transforms of F,(s) and F,(s), respectively, and if a and b are con-
stants, then the inverse transform of aF; (s)+bF>(s) is af, (t)+ bf; (¢). In order
to prove this, we need only observe that af, (t)+ bf, (t) has Laplace transform
aF ,(s)+5F,(s) and use Theorem 2 to show that af, (t)+f,(t) is the only
continuous function in the class A with Laplace transform aF, (s)+5F,(s).
The problem of inversion of Laplace transforms presents many practical
as well as mathematical problems. As a simple example, we give the following
theorem which shows that not all functions F(s) can be the Laplace transform
of a function f(t) in the class A.
Theorem 3. If f belongs to the class A and if F(s) is its Laplace transform,

then lim,...,, F(s)=0.
Proof. Since f belongs to the class A, it satisfies an inequality of the form
(10.4) (see Section 10.2). Then, letting c= 2s, we have
Ciniee: alt
and it is clear that limy,..., F(s)=0. |

We have actually proved a little more than we claimed. Not only have we
shown that F(s) tends to zero as soo, but in fact that |sF(s)| remains
bounded as soo.
The question of which functions are Laplace transforms is a difficult one,
and we cannot in this brief treatment give a more precise answer. In
practice, we only apply Laplace transforms when we can prove the existence
of an inverse transform by finding it explicitly, but the question remains an
important one because it is often impossible to find the inverse transform
explicitly.
Exercise
10. a) Show that if f(t) belongs to the class A’ and has Laplace transform F(s), then
lim, SF (s)=/(0). [Hint: Use Theorem 4, Section 10.2, and Theorem 3. ]
b) Generalize part (a) to a result for functions in the class A‘ by using Theorem 5,
Section 10.2, and Theorem 3.
10.4 APPLICATIONS TO LINEAR EQUATIONS

WITH CONSTANT COEFFICIENTS
Our main purpose in developing the Laplace transform has been to apply
it to the solution of linear differential equations. We have suggested how
this may be done by examples in Section 10.2, where we considered first-order
differential equations. For equations of higher order, the general idea is the
same, but there are some technical problems which arise when we try to find
the inverse transform. In this section we shall discuss these technical
problems and the means of dealing with them by a collection of examples.
We shall concentrate on equations of the second order, but we shall also
indicate the minor additional problems which arise for equations of higher
order.
In each of the examples involving a second-order differential equation we
shall be seeking the solution @ of an equation of the form
Agy’ +a,y"+a,y=f (t), (10.28)

where do, a;, @ are constants, which satisfies the initial conditions
dO)=yo, o (O)=N1. (10.29)

Under suitable hypotheses on the function f, we may apply an existence
theorem, such as Theorem 1, Section 3.1, to conclude the existence of a
unique solution ¢ of Eq. (10.28) which satisfies the initial conditions (10.29).
When we attempt to find this solution by means of the Laplace transform,
we can expect to obtain it only if it and its derivatives up to the second order
10.4 Linear Equations with Constant Coefficients 407
are of exponential growth, that is, if it belongs to the class A?. Thus, to
conclude the existence of a solution which can be found by means of the
Laplace transform, we need a theorem on the growth of solutions of linear
differential equations with constant coefficients. Such a result is proved in
Theorem 2, Section 8.6. This theorem applied in the present context says
that every solution of a linear nonhomogeneous differential equation of order n
with constant coefficients, whose nonhomogeneous term is in the class A,
belongs to the class A".
Example 1. Let us use the Laplace transform to find the solution ¢ of the equation
y”+3y'+2y=0 which satisfies the initial conditions $(0)=1, $’(0)=1. We let Y be
the Laplace transform of ¢ and take the Laplace transform of the equation
" (t)+30'(t)+2(t)=0,
satisfied by @. Using (10.16), we obtain
s*Y(s)—s$(0)—'(0)+3[sY(s)—$(0)]+2Y(s)=0;
now, using ¢(0)=@'(0)=1, we obtain
(s*+3s+2) Y(s)=(s+3) 6(0)+ ¢'(0)=s+4.
Thus,
Sia 2 st+4
¥(s) (10.30)
s?4+3s+2 (s+1)(s+2)
In order to take the inverse transform in (10.30), we must simplify the expression (s+ 4)/
(s+ 1) (s+2). We use the method of partial fractions to accomplish this. Letting
s+4 A B
(s+ 1) (s+2) Sel ee
for all s, we observe that
AZ s+4 . st+4
= —— =
PEE Es FZ) so—-1 S$2
Boe
+ _
(en so
a
st4
so —2 (s+ 1) (s+2) s+
Thus,
2 Oe?
Y(s)= (s+1)(s+2)
st1 s+2
Now, we may use (10.3) and the linearity of the inverse transform to find a function
whose Laplace transform is Y(s), and because of the uniqueness of the inverse transform
(Theorem 2, Section 10.3), this function must be the desired solution ¢. We see that
$(t)=3e '—2e~*'. We could, of course, have obtained this solution by using the
methods of Section 3.4, but this would have involved first finding the general solution
of (10.28) and then substituting the initial conditions (10.29) to determine the constants.
The Laplace transform does not solve problems which would otherwise be unsolvable,
but it does provide an easy, practical method of solution for many problems.
Exercise
1. Verify the solution $(t)=3e ‘—2e~~! to Example 1 by direct substitution.
Example 2. Let us now use the Laplace transform to find the solution @ of the equa-
tion y’+4y'+4y=f(t) which satisfies the initial conditions #(0)=1, ¢'(0)=2, where
S(t) belongs to the class A. We let Y be the Laplace transform of ¢, F the Laplace trans-
form of f, and we take the Laplace transform of the equation
b"(t)+4¢'(1)+4e()=f(t),
satisfied by ¢. Much as in Example 1, we obtain
s*Y (s)—s(0)—$'(0)
+4[s¥ (s)— (0)]+4¥ (s)=F(s),
or
(s?+4s+4) Y(s)=(s+4) 6(0)+¢'(0)+
F(s)
(s+4)+2+F(s)=s+6+F(s).
Thus,
s+6+F(s) s+6 F(s)

Y(s)= = i
(5) s*+4s+4 (s+2)? (s+2)?
Again, we use partial fractions to simplify this expression. We must find constants A
and B such that
s+6 A dy B
(Go) se (s+2)?
We observe that
s+6
B= Lo
lim ——
aa (s+ 2)*=
) lim
ete (s+6) ) =4,
A s+6 B s+6 4 s+2 1

s+2 (s+2? (842)? (5+2)? (s+2)? (s#2)? s+2
which implies A =1. Now,
Vas See 6 1 4 F (s)
= + 5
(s+2)? s+2 G42" G+)

and we may use (10.3), (10.12), and Theorem 1, Section 10.3, to find the inverse trans-
form. This yields the solution.
t
bl)me rede 2+ [tu e “8 OF (udu.

0
Exercise
2. Verify the answer to Example 2 by direct substitution.
Examples | and 2 give us enough insight into the method to enable us to

solve the general linear homogeneous equation of order n with constant
coefficients. Let ¢ be the solution of the equation of order ,
ie (y) = agy” aie a,y" ee Sa a,y = 0 > (10.3 1)
(ao, 4,,..., d, constants) which satisfies the initial conditions
$1) =Vo. 6=I 6 OV=Ypr. (10.32)

We let Y be the Laplace transform of ¢ and take the Laplace transform of the
equation
agp (t)+a,p"" Y(t)+---+a,6(0=0.
Weobtain
ao[s"¥(s)—s"~ y=
8 y=]
aq [Sa VS) aS"5 7Vo real
+a,—1[sY(s)—yo]
+a, Y (s)=9,
or
p(s) ¥(s)=ao(s"" "Yo s" *y1 t+ + Yn=1)
+4, (s"~7 yo + e+ + Vy 2) Fe +n 10> (10.33)
where p is the characteristic polynomial of the linear operator L,, (see Section
3.4).
p(s)=dos"+a,s" 1 +++-+a,.
We may write (10.33) in the form
p(s) Y(s)=a(s),
where g is a polynomial of degree at most (n—1) on the right-hand side of
(10.33). (The degree could be less than (n—1) if the given initial value yo is
zero.) This polynomial g is linear in yo, 1,.--, Ya—1-
We now have
y (5) 40)
p(s)
and the next step is to separate the rational function q(s)/p(s) into partial
fractions. If the roots of the polynomial p are z,,..., 2, of multiplicities
M,,..., M,, respectively, then we can write
P(s)=4(s—21)"*(s—22)"?---(s—2,)"*.
The process of separating into partial fractions gives
q(s) _ ay1 Aim a1 tae a2 m2
p(s) s—2Z (s—z,)" s—Z, (s—z>)”?
kl = Ak,m,
us a ty (s—z,)"*
where the constants d,,,..., A, m, May be calculated. We may now take the
inverse transform of Y(s) using (10.3) and (10.12). We obtain the solution
(
t)=a; (ee
) sa
ay
(m,
er
m
=
-1 t
1) !
Oe Zit©
k1
teapot Ak, my peat etd,
(m,— 1) 5
which is, of course, the same as that obtained in Section 6.4. If some of the
roots z,,..., Z, are complex but the coefficients do, ..., a, of L are real, then
we may express the solutions in terms of real functions just as in Section 3.4.
We remark that if in place of Eq. (10.31) we consider the equation
LiAy)=f (0),
where f belongs to the class A, we handle the additional term by using the
convolution exactly as in Example 2 above.
The Laplace transform is used a great deal in the study of electrical
circuits. We have already mentioned in Chapter | the linear electrical circuit
consisting of a capacitance C, a resistance R, and an inductance L connected
in series. The voltage v(t) across the capacitance may be described by the
equation
1
Lv"v + Rv'
+ v=0.
Uae
Consider such a circuit with an external applied voltage A cos(kt+a), such

as might arise from connecting the circuit to a source of alternating current.
Then the voltage v(t) would be described by the equation
eal
Lv" + Ro’ + v=A cos(kt+a). (10.34)
In studying Eq. (10.34), we find it convenient to write the nonhomogeneous

term as the real part of a complex exponential, solve the corresponding
complex equation, and then take the real part of the solution, just as we
suggested in Section 3.4. To do this, we define the complex number
b= Aci: (10.35)
then
A cos(kt+a)=A[Ae'“*?] = B(Ae'e)
=Abe™.
Thus, instead of (10.34), we consider the complex equation
1
Ly" + Liars Verve. (10.36)
Let Y be the Laplace transform ofthe solution ¢ of(10.36) which satisfies the
initial conditions $(0)=yo, 6'(0)=y,. Then, using (10.3) and (10.16), we
see that
LL? ¥(9)—s¥o—
11+ R[SY() yo]+zY(9)= b
s—ik
or
1 b
(i+ Rs+4) Y (s)=———
+ L(syp + y1) + Ryo. (10.37)
s—ik
Let z,; and z, be the roots of the polynomial p(s)=
Ls? + Rs+1/C, so that
Ls? + Rs+1/C=L(s—z,) (s—z2). Then we can write (10.37) as
Y(s)= b Elona) ees

L(s—z,)(s—22)(s—ik) =L(s—z,)(s—22)
When we separate into partial fractions, we obtain

M N P
Y(s) Sarr:
s—ik s—z,
ts
s—Z,
,
with
b b
ee a Diez) (= 2) PR
When we take the inverse transform, we obtain the solution of (10.36) in the
form $(t)= Me™ + Ne?" + Pe". If L, C, R are all positive constants, a
reasonable hypothesis in applications, then the roots z, and z, of the poly-
nomial p(s) have negative real part. (Why?) We let ¢,(t)=Me™, ,(t)=
Ne*'+ Pe". Then $(t)=¢,(t)+¢,(t) and ¢,(¢) tends to zero exponen-
tially as t+ +00. The term @,(t) is called a transient, because its effect dies
out, and the term @,(f) is called the steady state. As we wish to concentrate
on this steady state, we do not examine the transient term further. (For this
reason we did not actually compute the constants N and P.) The steady state
is [b/p(ik)] e'. To calculate the corresponding voltage v, (t), we must take
the real part of this expression. We define the transfer function
CW= say.
Then
v,(t)=A[bC(k) eM] =R [ Ae’*| C(k)| gree) a
=A A] C(k)| eee)
= A|C(k)| cos[kt+a+argC(k)].
When we compare this steady-state output voltage v,(¢) with the input
voltage A cos(kt+«), we see that the effect of the circuit has been to multiply
the amplitude by the gain function |C(k)| and to introduce a phase lag arg C(k).
Note that v, is independent of the initial conditions. The transfer function
C(k) is determined by the electrical circuit, but depends on the frequency of
the input voltage. The essential principle in tuning a radio is to adjust the
circuit, usually by varying the capacitance C, to maximize the gain function
for a given k. In other electrical applications, it is necessary to vary k to
maximize the gain function for given values of L, R, and C.
The reader is warned that we have assumed that L, R, and C are constant
and that the circuit is linear in the above discussion. For time-dependent or
nonlinear (vacuum tube) circuits, the resulting differential equations cannot,
as a rule, by solved by means of the Laplace transform, and other methods
must be developed. It may still be reasonable to define a gain function and a
phase lag, but these will no longer be given by a transfer function.
The Laplace transform is sometimes used to solve linear nonhomogeneous
differential equations whose nonhomogeneous terms do not belong to the
class A. While the solutions obtained in those cases are then purely formal,
it is possible to show that they are actual solutions in a more general sense.
The justification requires a more sophisticated approach, such as the theory
of distributions. Here, we shall only give an example to indicate the nature
of the problem.
Let us again consider an electrical circuit consisting of a capacitance C,
a resistance R, and an inductance L connected in series. However, now let
us attempt to determine the behavior of the circuit if a large external voltage
is appled over a very short time interval. Let this external voltage be
defined by
O,(j=We (0<t<o)
0 (t>0o).
Then the circuit is governed by the differential equation
1
Ly" + Ro'+— U—0_(1).
Since 6,(¢) is certainly in the class A, we can treat this equation by taking
Laplace transforms, using
Be =| Genes (10.39)
Exercise
3. Find the steady-state solution $, of Eq. (10.38) satisfying the initial conditions
$(0)=Yo, $’(0)=3,. (By steady-state solution we mean, as above, the difference
between the solution and those terms ¢, in the solution which (because L, R, C
are positive constants) tend to zero as t> + 00.)
tml
Figure 10.1
It is of interest to consider the “limiting behavior” ofthis electrical circuit

as ¢—0. The applied voltage 6,(t) does not tend to a limiting function in
the usual sense (see Fig. 10.1) when o-0, but we can think of the limit as
an impulse at t=0 whose integral is 1, and call this a ‘‘generalized function”’
6(t) (also the Dirac delta function). This would lead to the differential
equation
1
Lv" +Rv'+e v=6d(t). (10.40)
Wetreat this equation by Laplace transforms, proceeding as if we could write

: ae
# {5(t)}=lim FEp {5,(t)}= lim
rile
—-— = 1. (10.41)
a70 a0 Os
The reader should compare (10.41) with Theorem 3, Section 10.3, which states
that for a function fin the class A, lim,.,, “f(t)=0, and should note that
there is no contradiction here since 6 does not belong to the class A, although
0, does.
If we let Y(s) be the Laplace transform of the solution @ of (10.40)
satisfying the initial conditions #(0)=yo, $’(0)= yo, we obtain, using (10.41)
and writing p(s)=Ls* + Rs + 1/C=L(s—z,) (s—z) as before,
L[s*¥ (s)—syo— 1] + RESV(9)—Yo]+aVO)=1,

1
(12 +845) Y(s)=14+L(syo+yi)+RYo;
(10.42)
1 L(syo+y1)+RYo
Y(s)= L(s—z,)(s—z2) L(s—21)(s—22)
This can be separated into partial fractions in the form
A B
Vis)= =e ;
Dierai S—Z>
which leads to the solution (t)=Ae*"+ Be". Observe that if L, R, C

are all positive, so that z, and z, have negative real part, lim,.., (t)=0.
Observe also that the output voltage ¢ is a continuous function, even though
the input voltage 6 is not.
Since the method used in this example has not been justified, the answer
obtained must be verified by direct substitution. Such a verification requires
an understanding of the meaning of 6(t). Let us also point out that we could
obtain the same solution without using the Laplace transform by means of
variation of constants (see Section 3.7). This would also involve the same
nonrigorous calculations, and would also require verification by direct
substitution.
Exercises
Find, using Laplace transforms, the solution ¢ of each of the following differential
equations which satisfies the given initial conditions.
4. y’"—y=0, AeA fo
5. y’—Sy'+6y=0, Pa? Pa!
6. y"—6
+11y'y"
—6y=0, $0)=40'(0)=0, 670) =1
7. y" by =e, $(0)=¢'(0)=0, ¢’(0)=1
8. y’-9y=e’, $(0)=1, $'(0)=0
9. y’—9y=sint, $(0)=1, ¢’(0)=0
10. y’+4y=sin
21, o(0)=1, g'(0)=1
11. y’-y=0, $(1)=0, $'(I)=1
[Hint: Begin by making the change of independent variable t = t—1 to move the initial
time to zero.|
12. y’—9y=f(t), ¢(0)=1, $'(0)=0, where f is in the class A. [Hint: Use the
convolution.|
13. y’+4y=f(t), 6(0)=1, o'(0)=1, where
f is in the class A.
14. Plot the gain and transfer functions of the linear differential operator defined by
each of the following:
a) L(y)=y" +4y'+3y
b) L(y)=y
+4y'+4y
c) L(iy)=y"+y
15. Find the steady-state solution of the differential equation y’ +9y=e*".
“16. Find the steady-state voltage in an electrical circuit with L=1, R=10, C=%
with an applied voltage 96(t).
17. Find the solution of each of the following initial-value problems by means of
Laplace transforms:
a) y"+4y'+4y=4cos2r, y(0)=2, y'(0)=5

10.5 Applications to Linear Systems 415
b) y’+a@*v=f(t), (0)=0, (0) =0, ~~where / is the function in Exercise

12(a), Section 10.2.
—c) y"+y=H(t), yO)=1, y'(0)=0, where H(t)=3 for 0<1<4, H(t)=2t-5
for 1>4.
rc
— d) y"+2y'+y=24+(t—-3) U(t—3), y(O)=2, y'(0)=1, where U is the unit
step function
SN. J6S0s
*18. a) Show that if (rf) is a solution of the Bessel equation of index zero,
tye + y + y= 0,
then the Laplace transform Y(s) of #(t) satisfies the first-order differential equation
dY
(s*+1) —+sY=0
ds
regardless of the initial conditions prescribed.
b) Solve the equation obtained in part (a) and use Exercise 10, Section 10.3 to show
that Y(s)=¢(0) (s?+1)~‘/. Explain why ¢’(0) cannot be prescribed.
c) By expanding Y(s) in powers of 1/s, show that
¥(s)=4(0) ¥ (-1)
Be
j=
1x3x5---x(2k—1)
Vk ! ved 1
g2k atek,
d) Show that
60=00) (lV age

(oe) t2*
[Note that the step from (c) to (d) is purely formal. However, (d) is a “‘rigorous”’ solution
of the Bessel equation, as can be justified by direct substitution and theorems on power
series (see Section 6.9).|
*19. Each of the following equations defines a function #(¢). Find ¢(t) by using
Laplace transforms. [Note that before we do this we should actually first prove that
¢ belongs to A. This can in fact be done by estimating |¢(t)| and using the Gronwall
inequality (Lemma 1, Section 8.2). Alternatively, we could just obtain the answer
and then verify that it belongs to A.]
a) $(t)=40?
—[o b(t—t) edt
b) (t)=C+JoO(c) sin(t—t) de
c) P(t)=1+2
fp b(t—1) cost dt
- 4) d'()=sintt
fyo(t—1) costdr, —(0)=0
e) $'(t)=t+Jo
o(t—t) cost dt, (0)=4
f) fo @(t—t)e ‘dt =t
10.5 APPLICATIONS TO LINEAR SYSTEMS
The methods used in the previous section to obtain solutions of a linear

differential equation with constant coefficients can be used with no essential
change to obtain solutions of a linear system with constant coefficients. We

shall not go into great detail in the solution of linear systems, but shall only
give a few examples to illustrate the method. The applicability of the
Laplace transform method to each of the systems considered follows from
Theorem 2, Section 8.6.
Example 1. Find the solution @=(@,, #2) of the system
Yi=2yi+Y2, Y2=—Yit4yr,
which satisfies the initial conditions ¢,(0)=0, $2(0)=1. Also find a fundamental
matrix. We let Y;(s)\=L{¢,(0}, ¥(s)=L {¢2(2)}. When we take Laplace transforms
in the equations satisfied by $,, 2, we obtain
SY, (s)—
$, (0)=2Y, (s)+ Y,(s)
SY)(s)=$2(0)= — Y%(s)+4¥2(s)
or
(s—2) ¥1(s)— ¥,(s)=¢$1(0)=0
¥, (s) +(s—4) Yo()=$2(0)=1.
Solving for Y, (s)and Y,(s), and computing the inverse transform, we obtain successively
hence
@;()=te”, $2(t)=e™ +te*.
To find a fundamental matrix we find the solution y=(W,, w2) for which w,(0)=1,
W2(0)=0. Proceeding as above we obtain
(s—2) Vi (s)— ¥%(0)

=, (0)=1,
Y, (s)+(s—4) ¥Yo(s) =W2(0)=0,

the solution of which is
s—4 1 1 —1
WO) |ayes
Gay?
Therefore, ,(t)=(1—2) e*', W2(t)=—te*', and by Theorem 2, Section 4.3, a fun-
damental matrix is
The reader should compare this solution with that of Example 2, Section 5.5, for the
initial condition imposed here.
Exercice
1. Verify the solution obtained in Example | above by direct substitution.
By taking the Laplace transform in Example 1, we reduce the problem of

solving a system of differential equations to the problem of solving a linear
system of algebraic equations. This is a substantial simplification, as
algebraic systems can be solved explicitly by an easily carried-out procedure
if the order of the system is not too large.
Another type of problem for which the Laplace transform is very useful
is a system of linear differential equations with constant coefficients of order
higher than the first. While such a system can be reduced to a system of
first-order equations by the methods of Chapter 4, it can also be solved
directly by means of the Laplace transform without the need for this reduc-
tion.
Example 2. Find the solution (¢,, $2) of the system
yi-—2y,—2y,=0, yy—2y,+y2=—2e™,
t
which satisfies the initial conditions ¢,(0)=3, $3(0)=2, ¢2(0)=0. We let Y, and Y,
be the Laplace transforms of ¢, and @5, respectively. When we take Laplace transforms
in the equations satisfied by ¢, and @5, we obtain
[s*Y,(s)—3s—2]—2[sY, (s)—3]—s¥,(s)+2Y,(s)=0
[5% (9)-3]-2%
(9)sH3(9)=— >
or
(s?—2s) Y,(s)—(s—2) Y,(s)=3s+2—6=3s—4

2 3s+1 (10.43)
—2) Y,¥,(s)+s¥5(s)
(s—2) Yo (5) = area Se ey
Solving the system (10.43), we obtain
Y,( 3s7—4s—1 pet n 1 F 1

Liat corr Grrl ten ee st1 s—2’ (10.44)
» 1 1
Y,(s)= =
(st1)(s—1) s—-1 s+1
We take inverse transforms of (10.44) to obtain the solution
di(t)h=e+e*, d2(t)=e'—e'.
Exercise
2. Verify the solution obtained in Example 2 by direct substitution.
To solve the problem of Example 2, we could proceed as follows. Trans-
form the system to an equivalent system of three first-order equations, find a
fundamental matrix for this system, and finally impose the initial conditions.
It is clear that in most simple problems such as the one in Example 2, the use
of Laplace transforms does give an answer more quickly.
Example 3. Find the solution (¢,, #2) of the system
yit2yitvityety2=0, Wityty2=0,
which satisfies the initial conditions ¢,(0)=1, $1 (0)=0, ¢2(0)=1, @2(0)=0. We let
Y, and Y, be the Laplace transforms of $, and $3, respectively. When we take
Laplace transforms in the equations satisfied by ¢, and ¢,, we obtain -
[s? Y, (s)—s]+2[sY, (s)—1]+ % (s)+[s? ¥,(s)—s]+[s

¥(s)— 1] =0
[s¥,(s)—1]
+ %(s)+[s¥(s)]=0
or
(s?+2s+1) Y,(s)+(s?+5s) Y,(s)=2s+3
(10.45)
(s+1) Y,(s)+sY,(s)=2.
When we attempt to solve the algebraic system (10.45), we find that it is inconsistent!
Hence, there is no solution to the given system of equations which satisfies the initial
conditions.
The Laplace provides an alternative means, independent of the develop-
ment of Chapter 5, for constructing a fundamental matrix for the system
y =4y. (10.46)
where A is an arbitrary n x n constant matrix. This is done as follows.
We will say that if f(t) is a vector function with n components defined on
0<t<oo, then feA if and only if each component of f is in the class 1; we
write Y {f}=|P exp(—st) f(t) dt. The analog of formula (10.16) holds for
vector functions (the proof is exactly the same as in the scalar case). It follows
from Theorem 2, Section 8.6, that if @ is a solution of (10.46) with @(0)=n,
then ge A! for any initial vector n. Let Y(s)=# {o}. Taking Laplace trans-
forms of both sides of (10.46) and using the initial condition, we obtain
sY(s)—n=AY(s).
Thus,
(s{—A) Y(s)=n. (10.47)
The system (10.47) is a linear nonhomogeneous system of n algebraic equa-
tions in m unknowns, namely the components (Y;,(s), Y2(s),..., Y,,(s)) of the
vector Y(s). Clearly, if s is not equal to an eigenvalue of A, det(s/— A)#40,
and (10.47) can be solved uniquely for Y(s) in terms of n and s by Cramer’s
rule. From it, since det(s]— A) is a polynomial of degree n, it is clear that
Y(s) is a vector whose components are rational functions of s and linear in
(11, 125+++5 Mn), the components of y. Hence, each component of Y(s) can be
expanded in partial fractions (the denominators will be integral powers of
(s—A,), where A; is an eigenvalue of A). Doing this we can then invert Y(s) to
find the solution @(f) corresponding to any initial vector y. Letting y succes-
sively take on the values
1 y 0
1 ;
up = 1 »N2= : ae65 1, 0
0 0 I
(or any other n linearly independent constant vectors in @, which form a
basis), the solutions ,, 5,..., @, used as columns of the matrix ® generate a
fundamental matrix of y’= Ay, such that &(0)=/.
Example 4. Construct a fundamental matrix ®(¢) (for example, the one with ©(0)=/)
for the system y’ = Ay, where
Be 1
A=W 0 1
Sa
Compare your result with Example 3, Section 5.5.
We use the method outlined in the preceding paragraph. In that notation with
n=3, we have
or
(s1—
A)¥(s)=n,
s—3 1 —1] | y,(s) ny
as = y2(s)| =| 2
—1 1 s—2}[ys(s)} [ms
Expanding det (s/— A) by the first row, we have
det (sE — A)=(s—3) [s(s—2)+1]+[12(s—2)+1]
—[—2+s]=s*—5s*
+ 8s—4=(s—1) (s—2)’.
By Cramer’s rule, we have
iy Ll —1
GO| Gy 8 - 1
Yolst=
(5) (s—1)(s—2)?
_m[s(s—2)+1]—n2(s—2+ 1)+n3(—1+5)
= (6-1)6-37
_m(s—1)—m +N3
(s—2)
=1 s—2
¥,(s)= 3
n (2s—3)+n2(s* —5s+5)+n3(s—1)
(s—1) (s—2)? ‘
s-3 1
det] —2 s 1
se ol —1 1 m3} _m(s—2)—no(s—2)+n3(s?
—38+2)
a(s)= (= 1$s=2) (s—1) (s—2)?
1-2 3
TP ike
It is convenient to substitute specific values for y,, 2,73 at this point, rather than
waiting until after taking the inverse transform as was suggested in the general
procedure. When we take 4, =1, 7, =0, n;=0, we obtain
s—1 A B
TA) S paa\ ae2) =D)
Then (s—1)=A(s—2)+B, which gives A=1, B=1. Thus
1 I :
MG) = 3? Gy and y, (t) =e" + te? =(1+42) CE.
Next
2s—3 A B C
¥i(s)\= =
2(3) (s—1)(s—2/ Re ES
and 2s—3=A(s—2)?+B(s—2)(s—1)+C(s—1), which gives A=—1,
B=1, C=1.
Thus,
Also
¥,(3) 1 1 1
Si =
: (s—1)(s—2) s—2 s-—1

and
y3(t)=e"—e'
Thus
(1+1) e™
, (t)= |(t+ 1) ee
e2t_et
is that solution of y’= Ay for which ¢, (0)=e, =[1, 0, 0].

Now, consider the case y, =0, n,=1, n;=0, then
Y, (s)= — and = yy, (t)= —te”*,

with
s?—Ss+5=A(s—2)?
+B(s—1) (s—2)+C(s—1)
which implies |=A, 4+ B=1, so that B=0, —1=C.
1
nS) ——
2($ s—1 (s—2/
and
y2(t)=e'—
te”!
—1 1 1
Y;(s)= =
(s—1)(s—2) s—1 s—2
and
y3(t)=e'—e7".
Thus
te?!
o,(t)= tet
et
W()=——5; and yi(=te™,

¥5(s)=
2 (s = : 2) and iy,2 (t)=te”" >
l 2t
Y; Ss er and y3(t)=e
Thus,
te"
3(t)= | te”
e2t
is that solution of y’=

Ay for which ,(0)=e3. Therefore
(1 +t) e?! —te te

(t)=[41
(0),b2(t), O3()]= |(1+) e*—e e'—te* te"
e2t—et a—ert et
is the fundamental matrix which is the identity at t=0.
Exercises
Find the solution [,, 2] of each of the following systems of equations which satisfies
the given initial conditions:
3. yity2=0, $1 (0)=1, 62(0)=0
Vive =
4. y,+3y3=0, 1 (0)=1, 6,(0)= -1

ioe ages AU
5S. Vit 3yi t2yityoty2=0, yet 2yity2—y2=0,
b,(0)=1, 4(0)=—1, $2(0)=0
6. 1 — 391, Foye V2— 292 =9, Vi Sit y2 =e,
7. A mechanical system consisting of two coupled springs, assuming no friction and

a periodic external force, is governed by a system of differential equations.
myi+kyy,—k2(y2—y1)=0
my; +kz(y2—y1)=A cosat,

where m,,™,k,,k,, and A are positive constants. Find the motion of the
system corresponding to the solution (¢,, 2), where $,(0)=1, ;(0)=0,
2(0)=0, 63 (0)=0.
* 8. For a system as in Exercise 7 with k, =k,=k, A=0O, define the kinetic energy
T=4m,y\? +4mzy;” and the potential energy V=k/2(yj —2y, 2 +y3). Show that
the total energy T+ V is constant. [Hint: Show that (d/dt) {T(t)+ V(d}=0, using
the differential equations. ]
9. Construct a fundamental matrix for the systems y’= Ay, where 4 is each of the
following matrices:
if 2 2 -3
a) a-|) | 8) a-|; rei
3 -1 -4 2
2 3 -2 -4
2 -1 -3 2
1 5: eel Mas)
10. Solve Exercise 10, Section 5.6, by Laplace transforms.

10.6 A TABLE OF LAPLACE TRANSFORMS
We present here a brief table of Laplace transforms, with references to this

chapter where the transforms listed are derived. It is understood that F(s)
denotes the Laplace transform of f(t), that G(s) denotes the Laplace trans-
form of g(¢), and that in transforms involving an arbitrary function f or g,
the functions involved belong to a suitable class to guarantee the existence
of the transform.
10.6 A Table of Laplace Transforms 423
Function Transform Reference
1
1 : (10.2)
e a (10.3)
Se
S
cos ft +B? (10.7)
sin Bt
B
+p? (10.7)
e“ cosBt sine (10.6)

(s—a)?
+ p?
(&
aesin pt (s—a)?+
B 2 (10.6)
A a (10.11)*
S
ee ee (10.12)
(s—z)
t*f (t) (—1)* F®(s) (10.10)
; k!(s+ip)k*?
t cos Pt [(s—o)? + Bee? (10.14)
gi: k!(s+ip)k*?
t* sinBt [6-04 (10.14)
fae k![(s—a)+ip]*** t
t*e* cosBt (s—0) +B (10.13)
He sinBt ills
w eee
[(s—a)? + p?]** 1
(10.13)
e"f (t) F (s—a) Theorem 3, Section 10.2
f'(t) sF(s)—f (0) (10.15)
f? (t) S/F (s)—s!~* f (0)—---— f 9 (0) (10.16)
t
[re ate) F(s) G(s) Theorem 1, Section 10.3

0
|rea F(s)/s G(s) Exercise 7, Section 10.3

(0)
t For the purpose of carrying out this calculation, assume that s is real, and then
let s be complex again after obtaining the answer.
* Formulas involving k! are valid when k is not an integer if k! is replaced by I’(k+ 1).
APPENDIX 1
Determinants and Linear

Systems
We consider the linear system in n equations and ” unknowns Ww, W,..., Wy:
Q1414W, + 4j2W2 + °** + 44,W,=Cy
Az,W, + 9W2 +++ +42,W,= C2
Qn1W1 + G,2W2 72> OW, Cy
where a,; and c,(i,j=1,...,n) are given real or complex numbers. If

C1 =C,=-+:=C, =0 we say that the system (1) is linear homogeneous, other-
wise it is called nonhomogeneous. By a solution of the system (1) we mean
a set of n numbers (real or complex) w,,..., w, with the property that if we
replace w; by w; on the left-hand side of (1) for i=1,..., n, the left-hand side
of the first equation will reduce to c,, the second to c3,..., the last one to c,,.
To discuss the solvability of the system (1) it is convenient to introduce
the determinant A of coefficients denoted by
Mii Gat PO ee
Qg1---A22 ee Arn
A=
ant An2 ee Ann
Here A is the number defined by
A=) (+) Qo +++9Ani,

A2i99
where the sum is taken over all indices i,,..., i, such that i,,i,...,i, is a
permutation of the numbers 1, 2,..., and where the + sign is used if the
424
Appendix 1 425
permutation is even and the — sign is used when the permutation is odd.
Thus Nt (=. A=, 1495 —@,4>); and if n=3
A = Gy 1422433 — Ay 1423439 +44 273431 — Ay 771433 +Ay 3421039 — 443097031.

The first principal result concerning the system (1) is for the homo-
geneous system (c; =c,=---=c,=0).
Theorem 1. If cy=c,=-:-=c,=0 and if A=0, there exists a solution of(1),

W1, W2,.-.-, w,, Such that not all w,, are zero.
Proof. We prove the result only for n=2; the general case is usually
treated in courses in linear algebra. Thus, consider
a14W, +a,.W,=0
Az1W, +a22W2=0
with A=a,,4,,—a,2a,,=0. If a,,40 take w,=1; then w,=—a,,/a,,
from the first equation. Moreover, substituting these values in the second
equation, we obtain
a12 1
Ay1{ ——— +422 )}=— (441422
— 421442) =0.
a ai
Thus, if a,, #0, Ww, = —a,,/a,, and w,=1 isa solution with w,, w, not both
zero.
Similarly if a,, =0 and a,, 40, W, = —(a)/a2,), W,=1 is a solution and
if a,, and a,, are both zero w, =1, w,=0 1s a solution which takes care of
all the possibilities and completes the proof. I
We now turn to the full system (1). The basic result is the following.
Theorem 2. (Cramer’s rule.) If the determinant A of coefficients of (1) is
not zero then the system (1) has a unique solution W,,..., W,. This solution is
given by
A,
wW,=— k= ttt
k A ( )
where A, is the determinant obtained by replacing the kth column of A (that

1S, Qizs Dans «++> Unk) OV C15 C25+++5 Cn, And leaving the remaining columns un-
changed. In particular, if cy=C,=-::=¢,=90, then the only solution is
W,=W,=---=w,=0.
Proof. We again give the proof for n=2 only and we suppose first that there
exist W,, W such that
Gy 1Wy + 412W2=Cy (2)
A71W1+422W2=Cp.
Then we multiply the first equation by a, and the second by —a,, and add,
426 Appendix 1
obtaining
ee A,.
W,A=c,ay2 — 44202 =
Dn boy)
Similarly, we multiply the first equation by —a,,, the second by a,, and
add, obtaining
~ 44, C4
WA
=— 47 1C, +4440.=
Az, C2
Since A 40 it must be the case that w,=A,/A and w,=A,/A. If we now define
the numbers w, and w, by these relations, we see by direct substitution that
they satisfy (2). This completes the proof. |
Combining Theorems | and 2, we easily obtain the following result.
Theorem 3. The linear system (1) of n equations in n unknowns has a unique
solution if and only if the determinant of coefficients A is not zero.
APPENDIX 2
Polynomial Equations
et
PAZ)=@o2 1 Ojo gaye, (1)
be a polynomial of degree n>1 with real or complex coefficients. Let
z=b be a root of the equation p,(z)=0.
Theorem 1. If z=b is a root of multiplicity k of the equation p,(z)=0, then
Pa(b)=p,,(b)=
pn(b)=---=pp (6) =0
but p (b) £0.
Proof. Since z=b is a root of multiplicity k
Pn (2) =o (z— b)*q(z)

where q(z) is a polynomial of degree n—k with g(b)#0; also, since b is
a root, p,(b)=0. Differentiating p,(z) we have
Dn (Z) = aok(z —b)*~* q(z)

+ ao(z—b)* (2)
and thus if k—1>0, p;(b)=0. If, however, k=1
Pn (2) = 404 (2) + 4o(2—) a'(2)

from which p,(b)=a q(b) #0. This proves the result for k=1.
To establish the general case we recall the formula
m(m—1
( ) pom oer
Ng feet 43
fg)
CG) =f gmp ge 5
for finding the mth derivative of the product fg of two m times differentiable
functions f and g. Applying this formula, using f(z)=a9(z—5)*, g(z)=4q(2),
427
428 Appendix 2
we obtain
Pn(z)=ao[k(k— 1)---(k—j+1) (2—b)*

4q(z)
+ terms which have higher powers of z—b as factors].
Therefore, for any k>1, lettingj assume values 1, 2,...,k—1, k, we find-
P,(b)=p, (b)=--- =p*~ ?(b)=0, but p (b)=k !agq(b)
#0 and this completes
the proof. 1
APPENDIX 3
Complex Numbers
and Complex- Valued
Functions
We assume that the reader is familiar with the definition of a complex number
z as an ordered pair (x, y) of real numbers with the definitions of addition
(1, V1) +(%2, V2) =(%1 + X2, V1 + 2)

and multiplication
(1, V1)" (%2, Va) =(%1%2

—V1V2, X1V2+X2y1)
for any complex numbers z,=(x,,
y,) and z,=(x2, y2), as well as the
familiar associative, distributive, and commutative laws for addition and
multiplication. It is customary to denote the complex number (0, 0) by 0,
and the complex number (1, 0) by 1. Note that z+0=z for every complex
number z and z:1=z. For z=(x, y) we define —z&(—x, —y) and z-'=
[x/(x?+y?), —y/(x?+y?)] if z#0; then z+(—z)=0 and z-z~1=1. As is
customary, we define z,; —z, =z, +(—2z,) and z,/z,=z,z3 ‘ ifz, #40.
It is also customary to identify the complex number (a, 0) with the real
number a and to denote (0, 1) by 7. Then, using the rules of addition and
multiplication, we have
z=(x, y)=x(1, 0)+y(0, 1)=x+iy.
The number x is called the real part of z, written x = Zz, and the number y
is called the imaginary part of z, written y=.%z. We also have
i7 =(0, 1) (0, 1)=(—1, 0)= —-1.
We define z, the complex conjugate of z=x+iy, by
Z=X—1y.
The product zZ7=x?+ y’, and we define the norm, or absolute value, of z as
tae) Sey)
429
430 Appendix 3
The absolute value has the properties

i) |z|>0O and |z|=0 if and only if z=0
ii) |—z|=|z|=[2|
iii) |2,22|=|2,|-|Z2|
iv) |Z, +2Z2|<|24|+
|z0I
v) |Z: -22/= |Iz1|—Izol |
It is often convenient to introduce polar coordinates for complex numbers
as follows. Let z=(x, y) be any complex number and let (r, 6) be the polar
coordinates of the point (x, y) in the plane. Then x=rcos@ and y=r sin6,
and we may write
z=x+iy=r(cos6+i sin)
with
r=(x-4-y-). o—l2), ape arctan =,
The reader can easily verify that if z,} =r, (cos0, +i cos6,), z3=rz (cos. +
isin@,), then, by the definition of z,z, and the trigonometric addition
formulas,
242, =1yrz[cos(6, +02)+i sin(O, +62)].
Taking z;=z,=z we obtain z*=r?(cos20+isin20), and by induction we
see that z”=r"(cosn@+i sinn@) for every positive integer n.
A complex function f (of the real variable 7) defined on some real
interval J is a rule (or correspondence) which assigns to each real number
t and J exactly one complex number /(f).
Example 1. Let z be a given complex number, z=«+if (a, B real) and consider the
function f given by f(t) =e, — 00 <t< oo. Thus to each real number ¢ the correspondence
f assigns the complex number /(t)=e%*'". (This function is discussed in the next
paragraph.)
Each complex-valued function f defined on some interval J gives rise to

two real functions, u, v, each defined on J by the relations
uQ)=BUf()], eQ= FL], tind,

where &[f(t)] is the real part of f(¢) and 4 [f(t)] is the imaginary part of f(¢).
Then we can write f (t)=u(t)+ iv(t). We sometimes write u=@ f, v= f. For
example, if f(t)=(¢?—1)+icost, t real, then u(t)=Z[(f(t))]=t?—1 and
v(t)=4[f (t)]=cost. The function e”,*=a+if8 plays an important role.
For this function we have the important identity, due to Euler, which states
that
fQ=e" =e" P= e4 (cosperismpes a CO 1 OOF
Appendix 3 431
Thus its real and imaginary parts are

u(t\=2[f(t)|=e cosBt, v(t)= SL f(t) |=" sin pt.
Any derivation of this formula requires some knowledge of complex analysis.
It is, however, not incorrect from the logical point of view to take Euler’s
identity as a definition of e*' when z is a complex number, and this is what is
done here.
Exercises
1. Using Euler’s identity as a definition, prove that e@'*72" =e" ¢72', where z,, z, are
given complex numbers.
2. For each of the following functions find the function u=&f, v= Sf:
= 2+ 3it b t)=
a) f(t)=e , t real ) f(t) a t real
°) f()=(2+
Seen8? ae
| AeSit odie
yen
If f, g are complex-valued functions defined on some interval /, their
sum (f+ g), product (fg), quotient ((//g) (g(t) #0)) are defined respectively by
the relations (f+ 9) (t)=f(t)+9 (4), 9) Q=/(0)-9 (9, (1/9) =f (d/9 (2) exactly
as in the real case. It should not come asa surprise that the definitions of limit,
continuity, derivative, and integral for a complex-valued function f, and
hence the calculus of such functions, is completely analogous to the real
case. We shall summarize the appropriate definitions and properties.
Definition. A function f is said to approach a (possibly complex) number L
as t approaches some number to (written lim,.,,f(t)=L) if and only if
lim,.,,{f()-L|=0 (where |f(t)—L| denotes absolute value).
Exercise
3. Give an e—o definition of the statement lim,.,, f(t)=L.
It is clear that all properties of limits which hold in the real case hold in
the present case as well. An important consequence of this definition is the
following property which does not arise in the strictly real case, but which is
very easy to prove:
Lemma. Suppose f is a complex-valued function of the real variable t, with

u=Rf, v= Ff. Then lim,.,, f(t)=L=a+ iB if and only if
lim u(t)=lim Z[f(t)]=a« and lim v(t)=lim ¥[f(t)]=f.
t>to t>to t>to t>to
We say that a complex-valued function f defined on an interval J is

differentiable at t=f, in J if and only if
lim Slot :=F {to) (h real)
h->0
432 Appendix 3
exists. We define the value of this limit to be the derivative of fat r=f) and
we denote the derivative of f at t=tg by f(to). More useful from the
computational point of view is the following equivalent definition (the
equivalence is established by the above lemma): f is differentiable at t=to
if and only if the real functions u=&f,v=¥f are differentiable at t=to;
the derivative offat t=tp is given by :
f (th=w (to) +iv' (to).
Example 2. Prove that for every complex number z=«+if, and for all r, (e**)’=ze?zt
(same formula as in the real case).
By Euler’s formula e‘ =e” cos fr+ie™ sinBt. Since the real functions e™ cos fr,
e™ sinfr are differentiable for all t, we have by the above equivalent definition of
derivative
(e*)’ =(e* cos Bt) +i(e” sin fry’
=ae™ cos Bt— Be™ sin Bt +i(xe™ sin Bt+ Be” cos fr)
=(«x+if) e™ cos Bt+(—B+ia) e* sinBt
=(a+ iB) [e* cos Bt+ie™ sin Br] =ze™.
It is clear that all properties and rules of differentiation familiar from the
study of real functions hold in the case of complex-valued functions. This
includes, of course, derivatives of higher order.
Finally, we say that a complex function £, whose domain is the real interval
a<t<hb, is integrable on [a, 6] if and only if both real functions on u=Af,,
v=¥ f are integrable on [a, b], and in this case we define
b b
|ro ar=|u(t) aril v(t) dt.

a
This definition gives all the usual rules of integration.
Example 3
r/2 r/2 r/2
e
et it= |cost dt+i |sintdt

.
ce) 0
t=nx/2
—icost =1+i.
t=0
APPENDIX 4
The Exponential
Matrix
In order to justify the definition of exp M, where M is any nxn matrix,

we need the concept of convergence of a sequence of matrices. To measure
distances between numbers, we use absolute values. To measure distances
between matrices, we introduce the concept of the norm of a matrix.
We define the norm (length) of amatrix A, denoted by ||A||, by
n
|Al= dX laijl. (1)

es
that is, as the sum ofthe absolute values of all the elements. We readily verify
that the matrix norm satisfies the following properties:
i) |A+B) <A +B
ii) | ABI <All (Bi
for matrices A, B of complex numbers. The above norm is convenient for
our purposes; other matrix norms satisfying the properties (i) and (ii) are
possible.
Exercise
1. Prove properties (i) and (ii) of || A].
We now use the matrix norm (1) to define the concept of convergence of a
sequence of matrices.
Definition. The sequence of {A} converges to the matrix A, where A® and
A belong to F,,, ifand only if the sequence of real numbers {\|A—A™'||} has
limit zero, and in this case we write
{A544 or lim AM=A.

k> oo
433
434 Appendix 4
Clearly, because of the definition of the norm, this means that {A} A
if and only if the sequence {a‘?} of complex numbers, representing the
element in the ith row and jth column in the matrices {A}, converges to the
element a;; of the matrix A as ko, for each of the n* elements (i,7=1)..., 7).
To prove this, note that |a —a,;|<||A®
— Al] for i, j=1,..., 7
A matrix function A(t) is a correspondence that assigns to each point t of an
interval I one and only one nxn matrix A(t). Using the remark following the
definition of convergence of a sequence of matrices, we see that it is consistent
to say that a matrix function A(t) is continuous, differentiable, or integrable
on an interval / if and only if each of its n? elements q;;(¢) is continuous,
differentiable, or integrable, respectively, on J. We say that a_ series
Vo U, of matrices converges if and only if the sequence {)%-o U,} of
partial sums converges. The limit of this sequence of partial sums 1s called the
sum of the series.
Combining the definition of convergence of a sequence of matrices with
the Cauchy criterion for sequences of real or complex numbers, we can es-
tablish the following result:
Lemma 1. A sequence {A,} of matrices converges if and only if given a
number ¢>0, there exists an integer N=N(e)>0 such that ||A,,—A,||<é
whenever m, p> WN.
We are now ready to prove that the definition
M? M!
exp M=I+ LSE se oS pa
a aR es
makes sense for every nxn matrix as follows. We define the partial sums
M M? M*
S.=1+7, Nag (2)
Using the matrix norm, we have for m>p

m M* m \|.M||*
ISn-S1=|, 5. ple [aren (3)

! A !
Exercise
2. Use the properties of the matrix norm to justify the calculation (3).
[Note that the calculation is possible since the sums in (2) are finite.]
Since, for any matrix M, is a real number, we can form
eM (IMIF/K!).
Hence, the sum
m
eh, (|| MII‘/k 3)

Appendix 4 435
is a partial sum ofa series of positive real numbers that is known to converge.
By the Cauchy criterion for series of real numbers, we see that given ¢>0 we
can choose an integer N>0 such that the right-hand side of (3) is less than ¢
whenever m, p> N. By (3),
{S.Opl) 2 for
m, p>N.
By Lemma 1, the series converges, and thus exp M is well defined for every
nxnmatrix M.
APPENDIX 5
Generalized Eigenvectors,
Invariant Subspaces, and
Canonical Forms of Matrices
In our study of a linear system of differential equations with constant

coefficients, y’ = Ay, when we treated the general case where A has fewer than
n linearly independent eigenvectors in Section 5.6, we stated a theorem of
linear algebra without proof. This appendix is devoted to the proof of that
theorem, and some interesting algebraic consequences.
In order to prove the desired theorem, we must introduce the concept of a
direct sum of subspaces of a vector space. Let V be a vector space and let XY
and Y be subspaces of V. We make the following definitions.
Definition 1. The vector sum of the subspaces X and Y, denoted by X+Y,
is the set of all vectors ve V of the formv=x+y for some xEX, ye Y.
Definition 2. The intersection of the subspaces X and Y, denoted by XX Y,
is the set of all vectors ve V which are in both subspaces X and Y.
It is easy to verify that ¥+ Y and Xn Y¥ are subspaces of V.
Exercise
1. Let X and Y be subspaces of a vector space V. Show that Y+ Y and XN ¥Y are also
subspaces of V.
For any subspaces X, Y of V, the vector 0 belongs to XN Y.

Definition 3. The sum X+Y is said to be a direct sum, written X@Y, if
XAY={0}.
Exercise
2. a) If X and Y are finite-dimensional subspaces of a vector space V, show that
dim(X+ Y)+dim(X q Y)=dim

X +dim Y.
436
Appendix 5 437
[Hint: Begin with a basis for ¥7 Y and extend it to a basis for ¥ and a basis for
all
b) Deduce that the sum X-+ Y is direct if and only if
dim(X + Y)=dim X + dim Y.
We now extend the definitions of vector sums, intersections, and direct
sums to any finite number of subspaces. Let X,, X>,..., X;, be subspaces of a
vector space V.
Definition 4. The vector sum

k
», Xi
i=1
of the subspaces X,, X>,..., X;,is the set of all vectors ve V of the form
k
v=) x
i=1
for some x;€X,, where i=1, 2,..., k.

Definition 5. The intersection
k
aes
i=1
of the subspaces X,, X>,..., X;, is the set of all vectors ve V which is in each
of the subspaces X,, X>,..., X,-
Definition 6. The sum
is a direct sum, written
if and only if
for everyj, where j=1, 2,..., k.
Exercises
3. Let X,, X>,..., X;, be subspaces of a vector space V. Show that if

k
v=> AG:
i=2
is a direct sum and if X,+ W is a direct sum, then

k
2» x;
i=1
is a direct sum.
438 Appendix 5
4. Show that a vector space V is a direct sum of subspaces X,, X,..., X4;
k
V= ® Xj;
t=1
if and only if every vector ve V has a unique representation
where x,eX, for i=1,,2,...,.X.

5. Let X;,1,---, Xisn, be a basis for the subspace X;, where i=1,..., k. If
k
a= ® Xi>
t=1
show that:|5j1.25 Xie) Xoies 1s Xone Ma Bayle a Dasis tory
We recall that if 7 is a linear transformation of complex n-dimensional

space into itself, then corresponding to any basis {v,, V2,...,¥,} of the
space there is an nxn matrix that represents the linear transformation T
with respect to this basis. The elements a;,(i, j=1,..., m) of this matrix A are
defined by
Ty,;= » AjiV jp (le n). (1)

j=
Corresponding to a different basis {w,, W2,..., W,} of complex n-dimensional

space there is another matrix B that represents 7. Such matrices A and B
are similar; that is, by definition there exists a nonsingular n xn matrix P
such that B= P~'AP. The proof that A and B are similar may be found, for
example, in [3, Section 5.7]. We write 4~B to denote that A and B are
similar.
Exercise
6. Show that (i) A ~ A, (ii) A~ B implies B~ A, (iii) A~ B, B~ C implies A~C.
The study of canonical forms of matrices involves the choice of a basis

of complex n-dimensional space relative to which the matrix of a given
linear transformation takes a particularly simple form. Equivalently, this
study involves the determination of a matrix of a particularly simple form
that is similar to a given matrix.
We first discuss the case of an nxn matrix A with precisely n linearly
independent eigenvectors.
Theorem 1. Let A be annxnreal or complex matrix. Let ¥,, V>,..., V, ben
Appendix 5 439
linearly independent eigenvectors corresponding to the eigenvalues 41, Ay, ..., An

(not necessarily distinct). Then there exists a nonsingular matrix P such that
Di WOK as aub.0
: ape
PAP Oy
Sg2 S:| a
lee:
Of eee
that is, A is similar to the diagonal matrix B.
Proof. Define the matrix
Pav v,].
By hypothesis, the eigenvectors V,, V>,..., ¥, are linearly independent; thus,

P is nonsingular. Now,
PT AP=P ALN eens v,]
=P oA Avanaes Ay,,|
= /P>! LAyvine AWnls
where in the last step we have used that v, is an eigenvector corresponding to

A,, SO that Av,=A,V,, where k=1,...,n. But, by matrix multiplication,
since P-'P=J and since v, is the kth column of P, P™ 'v, is the kth column
Oni k= 02.2. n. Vherefore;
i> AIAG
Poe alan
Ge one
The converse of Theorem | is also true:
Theorem 2. If the complex nxn matrix A is similar to a diagonal matrix D,

then A possesses n linearly independent eigenvectors.
Proof. Since A is similar to a diagonal matrix D, there exists a nonsingular
matrix Pe@,,, Such that 1
il
ee Vy Be ds
dn
Since P is nonsingular, the columns of P are linearly independent. Suppose
that
Pal\V5 V5 --+5 v, |.
We will show that v, is an eigenvector of A corresponding to the eigenvalue

d,. From P~'AP=D we have AP=PD. Writing this out, we obtain, by
matrix multiplication,
440 Appendix 5
AP=A[Vj, V2)«++5 Yn]=[A¥1,---) AV],

dy
d
PD=|Vievee el ‘
d,
=iVis OV oY:
Therefore
[Avis AV, .0+, AV, |= (di
¥i, do¥ay---, 4% |
and so Av, =d,v,, where k=1,..., n. Thus, v, is an eigenvector corresponding
to the eigenvalue d,. As already remarked, the eigenvectors v,, V2,..., V, are
linearly independent. I
Since eigenvectors corresponding to distinct eigenvalues are linearly inde-
pendent, we obtain the following useful result.
Corollary. If a matrix A has a distinct eigenvalues 4,,...,2,, then the
corresponding eigenvectors form a basis of €,, and A is similar to the diagonal
matrix
Ay
B= Aa
An
The reader will observe that both A and P~ ‘AP in Theorem | have the
same eigenvalues /,,..., 4,. This suggests the following result.
Theorem 3. If A, B are complex nxn matrices and if A and B are similar,
then A and B have the same characteristic polynomials and hence also the same
eigenvalues.
Proof. Since A and B are similar, there exists a nonsingular matrix P such
that B= P~'AP. We compute
det (AI — B)=det(AI — P~' AP)=det(AP~ 'P—P~' AP)
=det P~'(AI— A) P (2)
=det P~' det(AI
— A) det P=det(AI
— A),
where we have used several properties of determinants. Since the eigen-
values ofA are the roots of the equation det (A /— A) =0 and the eigenvalues
ofB are the roots of det(A /— B)=0, the result follows from (2). |
Corollary. Let v be an eigenvector of the complex n x n matrix A correspond-
ing to the eigenvalue 4. Let P be a nonsingular nxn matrix. Then P~'Vv is an
eigenvector of P~' AP corresponding to the eigenvalue 2.
Proof. Since Av=Av, we have
(P-*AP) (Po W=P> pave v=) Petal
Appendix 5 441
Exercise
7. Justify each step in (2).
The converse of Theorem 3 is false since, for example, the matrices
ie) it il
4-(q | and B=|, |
both have a double eigenvalue 1, but are not similar.

Example 1. For the matrix
a
ea
find a matrix P such that the matrix P~' AP is diagonal.
“(}
We compute that the eigenvalues of A are 3+5i, and that
are eigenvectors corresponding to the eigenvalues 3+5i and 3—Si respectively. As in

Theorem 1, we define
LS:
P=[u, |; if
Since det P40, P is nonsingular; further,
Hence,
rer | oreretd eae |fae

-] ae Mt fie, de Sillfyion
‘a 345i S=Sr (ii

at =6=37 Sori t
, fori: 2 0 23451 20
ete) e610 |e 0. esr
The reader will note that the diagonal elements of P-' AP are precisely the eigenvalues
of A, in accordance with Theorem 1.
Example 2. Compute the eigenvalues and eigenvectors of the matrix
Slat
4 3
and find a matrix P such that P~' AP is diagonal.

We have
det (AI — A) =det te Z|

=4?—41+3=(4-3)(A—1).
442 Appendix 5
“() =
Thus 4=3 and J=1 are eigenvalues. Corresponding eigenvectors are
respectively. Observe that both the eigenvalues and the corresponding eigenvectors are
real. Moreover, the vectors u and v form a basis for two-dimensional Euclidean space.
early
Clearly, :
P=[u, v]= ae
is a matrix such that P- ‘AP is diagonal.

As Example 2 suggests, if both A and all eigenvalues /,,..., A, are real,
then in Theorem | and its corollary and Theorem 3, the matrices P and B
are real. We have the following real analog to Theorem 2. Jf a real nxn
matrix A is similar to a real diagonal matrix, then A has n real linearly
independent eigenvectors.
Exercise
8. Prove that the matrices
i @ i i
4-(q if B-| |
are not similar.
To discuss the situation when there are fewer than n linearly independent
eigenvectors, we introduce the concept of generalized eigenvector. If for some
value A and some p>1, there is a vector y such that
(A—AIPv=0 but ~-(A—AIP"' v0, (3)

then v is said to be a generalized eigenvector of index p corresponding to the
generalized eigenvalue 4. When p=1, J is an eigenvalue and v a corre-
sponding eigenvector. We note that since u=(4—J/)?" 'v40, and since
because of (3), (A—Al) u=(A—Al)’"v=0, the “generalized eigenvalue” 2
must be an eigenvalue of A with a corresponding eigenvector u.
Lemma 1. If v is a generalized eigenvector of index p, then the vectors v,
(A—Al) v,...,(A—AD)?~'v are linearly independent.
Proof. If the given vectors are linearly dependent, then one of them, say
(A —AI)*v for some k(0<k<p—1), can be written as a linear combination
of the later ones:
Dil
(A—ADFv= Yo ¢(A=alyy.
j=kt+1
Since (A —A/)4v=0 for qg>p, application of (A—AI)?~1!~* to both sides of

this equation gives (A —A/)’~'y=0, which is a contradiction, proving the
linear independence of the given vectors. I
Appendix 5 443
Given an eigenvalue 4 of A, we consider the subset X of complex n-

dimensional space consisting of all generalized eigenvectors corresponding
to A together with the zero vector. It is easily shown that X is a sub-
space of @,. Let r be the largest index of any such generalized eigenvector,
so that (A—A/)’x=0 for every xeX, but (A—Al)""'y40 for some yeX.
Since, by Lemma 1, X contains at least r linearly independent vectors, r
is finite (in fact, r<n). The integer r is called the index of X, and dim Y>r.
The subspace X may also be described as the null space of the linear trans-
formation (J— AJ)" (or the null space of the matrix (A —A/)"). Let Y be the
range of the linear transformation (T—A/)". We recall that the range of
the linear transformation (T—AJ)" is the set of vectors y such that
(T—AI)'z=y for some complex vector z. By the theory of linear algebraic
systems, dim ¥+dim Y=n. Next, we observe that X and Y are disjoint sub-
spaces; for if vyis in both X and Y, then (T—AJ/)'v=0 and v=(T—AJ/)'u for
some uce@,. Thus, (T—A/)?"u=(T—A)'v=0, and thus u is a generalized
eigenvector; that is, ue X. But since r is the largest index of any generalized
eigenvector corresponding to A, there exists an integer /, 0</<r such that
(T—AI)'u=0; hence also v=(T—AI)"u=0. This shows that X and Y are
disjoint subspaces of complex n-dimensional space whose dimensions add
up to n. It follows that @, is the direct sum of X and Y (written @,=X@® Y),
that is, that every vector v has a unique decomposition v=x+y, with xe X,
ye Y.
The subspaces X and Y are invariant under T, that is TxeX for every
xeX and Tye Y for every ye Y. To see this we observe that X and Y are
invariant under (7 —J/), and it follows easily that they are invariant under T.
Next, suppose that the subspace X has dimension k, and let v,,..., v, bea
basis of ¥. Then Y has dimension (n—k), and if v,+1,..., V, is a basis of Y,
the fact that complex n-dimensional space is X¥ © Y implies that v,,..., v;,
Vi+1>---> V, IS a basis of @,. Since X is invariant under 7, 7v,;e¢X, where
(i=1,..., Kk), and since ¥,,-.., Vz 1s.a basis of X,
j=
Note that the sum is from j=1 to j=k, not to j=n. Similarly,
Tv)
f ay,
ji" (=k ln)
j=k+1
Comparing these formulas to (1), we see that the matrix A of T with respect
to the basis v,,..., v, of G, is a “block diagonal” matrix,
tld iO
ae Nae a
444 Appendix 5
Here, the k x k matrix A, represents the restriction 7, of the transformation

T to the subspace X and the (n—k) x (n—k) matrix A, represents the restric-
tion 7, of T to the subspace Y.
Since for the block diagonal matrix A,
det (A —AI)=det(A, —AI,) det(A, —AI,),
where J, is the k xk identity matrix and /, is the (n—k) x (n—k) identity
matrix, the characteristic polynomial of A (or of T) is the product of the
characteristic polynomials of A, and A, (or of T, and of 7,).
On X, the transformation T has only the eigenvalue J. For, if w is an
eigenvalue and v a corresponding eigenvector in X, and if r is the index of X,
then
0=(A—Al)! v=(A—-Al!! (A—AD) v=(A—Al)* (Av—Av)
=(A-AD!* (u—A) v=(u—A) (A-AI 8v=-- =(u-A v,
and A=. Thus, X contains all the eigenvectors of A corresponding to the
eigenvalue A, but no other eigenvector. In particular, if the linear transforma-
tion T has only one distinct eigenvalue A, then the subspace Y (the range of
(7—AI)") contains no eigenvectors of T. Thus, the restriction 7, of T to Y
has no eigenvalues or eigenvectors. Therefore, Y must consist of only the
zero vector, and X=@,. Observe that in this case, the linear transformation
T—AI (or the matrix A—AJ) is nilpotent: For every vector xe@,,
(A—Al)'’x=9.
We have now developed the algebraic machinery needed to prove the
following basic theorem, which is used in Section 5.6.
Theorem 4. Let 1,,43,..., 2, be the distinct eigenvalues ofa matrix A, with
multiplicities ny, Nz,..., Ny, respectively. Then complex n-dimensional space
is the direct sum of the subspaces X,, X>,..., X, of generalized eigenvectors
corresponding to the eigenvalues A, A ,..., A,, respectively. The subspace
X; is invariant under A, has dimension nj, and (A —/,1)""x =0for every xe X;,
where
j= li 2.4.
Proof. We prove the result by induction on k, the number of distinct

eigenvalues. The case k=1 has been covered in the remarks preceding the
statement of the theorem. Now, suppose that the result has been established
for every matrix with fewer than & distinct eigenvalues, and let A have k dis-
tinct eigenvalues /,,..., A,. On X,, the transformation T corresponding to
the matrix A has only the eigenvalue A,. Let Y, be the range of (T—1,/)",
where r, is the largest index of the generalized eigenvectors corresponding to
A,. The characteristic polynomial of A is
Appendix 5 445
As we have seen in the remarks preceding the statement of Theorem 1, the

characteristic polynomial of A is the product of the characteristic polynomial
of A on X, and the characteristic polynomial of A on Y,. The characteristic
polynomial of A on X, contains all the factors (A—/,) in p(A) because X;
contains all the eigenvectors of A corresponding to 4,, and it contains no
factor (A—A,) with ; #1 because X, contains no eigenvector of A correspond-
ing to an eigenvalue 4,#A,. Therefore, the characteristic polynomial of A
on X, isexactly +(A—A,)". Since the degree of the characteristic polynomial
of a linear transformation is equal to the dimension of the space, this shows
that X, has dimension n,.
We have seen that complex n-dimensional space is the direct sum of X,
and Y,. By the induction assumption applied to the restriction T, of T to the
subspace Y,, which has only (k — 1) distinct eigenvalues 15, /3,..., 2,, Y, is
the direct sum of X3, X3,..., X,. It now follows from Exercise 3 that complex
n-dimensional space is the direct sum of X,, X>,..., X,,, and the theorem is
proved by induction. I
By the argument preceding the statement of Theorem |, we now obtain
the following interpretation of Theorem 1.
Corollary 1 to Theorem 4. The matrix A of the linear transformation T
relative to a basis of complex n-dimensional space made up of bases of the
subspaces X,,..., X; is a block diagonal matrix
A,
A= A,
A,
where A, is ann; xn, matrix that represents T on X; for j=1,...,k.
Corollary 2 to Theorem 4. If A is a complex nxn matrix, then there exists
a nonsingular matrix P such that P~' AP has the block diagonal form given in
Corollary |.
The Jordan canonical form of a matrix is obtained from the above repre-
sentation by choosing bases of the subspaces X,,..., X;, in a suitable manner.
This requires a careful study of nilpotent transformations. A linear transfor-
mation L such that L"=0 but L’~!#0 is said to be nilpotent of index r.
We recall that the subspace X; is the null space of the transformation repre-
sented by the matrix (A—A,I)/, where r, is the largest index of the general-
ized eigenvectors corresponding to /;. Since X; is invariant under (A —A,J),
we may regard (A —/,/) as a linear transformation on X; that is nilpotent of
index r,.
Let L be a nilpotent linear transformation of index r on a vector space X
of dimension n. Then there is a vector u such that L'u=0 but L’ ‘'u40.
446 Appendix 5
By Lemma 1, the chain of vectors u, Lu,..., L’~‘u is linearly independent.

We will form a basis for X consisting of several chains of this type. If r=n,
then we have a basis of X consisting of the vectors u, Lu,..., L’~‘u. If r<n,
let U, be the subspace of X spanned by these vectors. For every v¢U,,
consider the chain v, Lv,..., LS’ ‘v, where each vector L?v, 0<p<s-—l1, lies
outside U, but L’ve U,. We choose a v that maximizes the length s of this
chain, and we let r,<r be the length of this maximal chain. Then vy, Ly,...,
L”?~1y are outside U, but L’veU,. Since u, Lu,..., L’~ ‘wis a basis of U,,
we may write
rae
Lv) pan, (4)

7=0
We apply L’"? to both sides of this equation and use L'u= L’v=0 to see that
pas r2—1
O= v= tL ic ar ae
j=0 j=0
Since L’-"u, L’~"?**u,..., L’~ ‘uw are linearly independent, c;=0 for j=0, 1,
...5 ("2 — 1). Thus, (4) becomes
Ret
ive Dmete (5)

f=
We define
r=
Hy) cea:
Aiea
Then it is clear from (5) that L’u,=0. On the other hand,
Du, =Lv—= Yo cdi*Sa

[|
(k=0, 1,...5 (72-1).

jer
Since L‘*/~"u is in U, but L*v is outside U,, L*u, is outside U, for k=0, 1,
..., (f2 — 1). Thus, every nonzero linear combination of u,, Lu,,..., L7~'u,
is outside U,. Let U, be the subspace ofX spanned by up, Lup,..., L’?~1u,;
then U, and U, are disjoint. The direct sum U,® U, is invariant under L.
Exercise
9. Prove the last statement.
If this direct sum is not all of X, we construct a maximal chain outside

U,®U, by the same method. Continuing in this manner, we can write X
as a direct sum of a finite number of subspaces U,, U>,..., U,, each of which
is spanned by a chain of the type given above. Thus, we have proved the
following result.
Appendix 5 447
Theorem 5. Let L be a nilpotent linear transformation of index r, on a

vector space X of dimension n. Then X has a basis of the form
BEE cig EAL Me LO Mig Te eee
_ —3? = — —
with
Petgeto sare) “and “EL a-0) (k=dy2,
1),
To construct the matrix B that represents L with respect to the basis given
by Theorem 5 we denote the basis elements respectively by v,, V>,..., V,-
Then we have
Ly, =0, Lv, =%,,..., LV,,=V,,-1;
Lv,,+:=9, EY. eo Ve aii seres LN 5 Vee
From the definition (1) of the matrix ofa linear transformation with respect to
a given basis, we see that B has the form
By
Beale? ) (6
B,
where B, is the r, xr, matrix given by
Oe Oy eet
B= 0
1
0 0 0
The Jordan canonical form of a linear transformation T is now obtained

by combining Theorems 4 and 5. According to Theorem 4, we can decompose
complex n-dimensional space into a direct sum of subspaces X,,..., X,. On
the subspace X,, the transformation T—/,J is nilpotent of index r;. We now
use Theorem 5 with L= T—A,/ to construct a basis for the subspace X;. The
matrix of the transformation T—/,/ restricted to XY; has the form (6). Thus,
with respect to this basis, the matrix of the restriction of T to X; has the form
Cy
C,
C;
448 Appendix 5
where C;, is the r; x r; matrix

i EO
0
Ce 0
1
0 nits
This gives the following important result.
Theorem 6. (Jordan Canonical Form) Let T be a linear transformation
of complex n-dimensional space with eigenvalues A,,..., A, of multiplicities
Ny,-.-, Ny, respectively. Then there exists a basis of complex n-dimensional
space relative to which T is represented by a Jordan canonical matrix
A,
A= a
A,
Here A, is ann, xn, matrix that has all diagonal elements equal to 4;, and that
has chains of \’s separated by single 0’s immediately above the main diagonal,
and all other elements zero.
Corollary. Every matrix is similar to a Jordan canonical matrix.
We remark that the length of the chains of 1’s in A; depends on the
integers r,,..., 7, 1n Theorem 5. It can also be shown that, except for the
order of the blocks 4;, the Jordan canonical form is unique.
Bibliography
1. G. Birkhoff and G. C. Rota, Ordinary Differential Equations, 2nd ed. (Ginn, Boston,
1961).
2 . F. Brauer and J. A. Nohel, Qualitative Theory of Ordinary Differential Equations
(Benjamin, New York, 1969).
3: F. Brauer, J. A. Nohel, and H. Schneider, Linear Mathematics (Benjamin, New
York, 1970).
. H. Bremermann, Distributions, Complex Variables, and Fourier Transforms
(Addison-Wesley, Reading, Mass., 1965).
. N. de Bruijn, Asymptotic Methods in Analysis (North-Holland, Amsterdam, 1958).
. R.C. Buck and E. F. Buck, Advanced Calculus, 2nd ed. (McGraw-Hill, New York,
1965).
. E. A. Coddington and N. Levinson, Theory of Ordinary Differential Equations
(McGraw-Hill, New York, 1955).
. C.A. Desoer and L. A. Zadeh, Linear Systems Theory (McGraw-Hill, New York,
1963).
. A. Erdélyi, Asymptotic Expansions (Dover, New York, 1956).
. P. Henrici, Discrete Variable Methods in Ordinary Differential Equations (Wiley,
New York, 1962).
. F. B. Hildebrand, Advanced Calculus for Applications (Prentice-Hall, Englewood
Cliffs, N. J., 1962).
. F. B. Hildebrand, Jntroduction to Numerical Analysis (McGraw-Hill, New York,
1956).
_ F. Jahnke and F. Emde, Tables of Functions with Formulae and Curves (transla-
tion) (Dover, New York, 1945).
. W. Magnus and F. Oberhettinger, Formulas and Theorems for the Special
Functions of Mathematical Physics, 3rd ed. (translation) (Springer, Berlin, 1966).
_ B. Noble, Numerical Methods, Vol. 2: Differences, Integration, and Differential
Equations (Oliver & Boyd, Edinburgh, 1964).
_ W. Rudin, Principles of Mathematical Analysis, 2nd ed. (McGraw-Hill, New York,
1964).
. J. Todd, A Survey of Numerical Analysis (McGraw-Hill, New York, 1962).
. W. Wasow, Asymptotic Expansions for Ordinary Differential Equations (Wiley,
New York, 1966).
_E. T. Whittaker and G. N. Watson, A Course of Modern Analysis, 4th ed.
(Cambridge Univ. Press, Cambridge, 1927).
20. D. V. Widder, The Laplace Transform (Princeton Univ. Press, Princeton, N. J.,
1946).
Psa K. Yosida, Lectures on Differential and Integral Equations (Wiley [Interscience],
New York, 1960).
449
Answers to Selected Exercises
1.1 Sloss
30 log(100
log 2
10(log 80—log 5). log 2
min k=
log 2 10
DISS
0.2 min
95.3 ounces
1.2
FSF
CO
FK
re
NOOO
1l+e"?
= f (tos ¥o)=0 is necessary. Consider ¢” and note that $” (to)= f;(to, Yo) if
Ff (tos Yo) =0.
Where $(t)= —t, max if t<0, min if t>0.
1.3 $(0)=1
(0, —1)
The (t, y) plane
Must be real-valued
Integral doesn’t converge, consider ¢(t)=0.
(5/4—2)" 1, |e] <./5/2

—(c? — 12) 12
e = 125/3/3 4t.13//3
0
JP=8+y2
Clay wee.
At —tot+Yo
450
Yot(to+ 1)
VYo(t—to) +to(t+ 1)
1
log [log (1 +e’) —log 2+]
=>a
9 VomA a= 6381+2>
xe
Wis 2 ett Sa
to3)/3
1
Yotl
tan (t+ 7/4)
1+?
1
log
(1 —t)
sin” '(t? + 7/2)
a) y=kx
b) y7=x-+k
c) 2x*+y?=k?
d) y?=4cx
10e~ 3 kg
w(x)=KL—L,./K? —2Kx
[yi +(t—t,)]°
[yd/?+(t—to)]?
f,(t), «0 where fp (t)=0,
ae
L=|4 t—a),t>o (a>0)
1.4 tel 4277!
Fy -pit-to) (0-4)
p P
Ap TE tpg
t/2—5/2t
(Aj 5— 8) St
(1/t?) (t—to+
toYo)
1
Wd
wW
2YADA1
3 sin t—
.
2 sint
1 t .
t+—| log——to+toyo
t to
1
10. Aeris. 1)
t ly 1
iL.
Rie ore
452 Answers to Selected Exercises
i il
(Ol 3 9 |ila
e)
ieee
14. (t+l)e*
16. e(2t—2)+e'(2+a)
a
17. tan t+———
cos t
18. e(1+a)-1
19. All solutions tend to zero if 1<0; all solutions are constants if A=0; all
solutions grow exponentially if 2>0.
20. The limits are as follows.
a) 0
b) co
ei l/2
d) 0
e) 0
22. (1—t—e™'*')"}
24, 27 9
27. limit is b/a
elec 3
31. 3—-21.9 kg
(2
1.6 Lc) No
2. b) Is not a region; all others are regions
4. a) (b) continuous
c) f continuous, g not continuous
0, yes
D is whole plane
ies
GO
A
SENa) i) Yes ii) Yes
b) Theorem does not apply
ty Al}
2.1 2. s(t)=
cos0
3. x(t)=[v9 cosO] t+xo
y(t)= 2 t?+[v 9sinO] t+ yo
do |?
2.2 T =mg cos0 +m] |
2.4 Apply existence and uniqueness theorem twice
y(t)}=G/2gR*t+ RP)?
y()=4(e2"+e72)
y(t)-=4(e*—e-)
Applicable
y(t)=e"
y()=(1-1)-
A is [Note that hypotheses of the uniqueness Theorem
Ji+e are not satisfied.]
2.5 a) ¢,=—1,¢,=1
a) All(¢, y, y’)
d) {(t, y, y'): v4, y’ #0}
n) All (t, y, y)
Sol There is a unique solution of the initial-value problem on any interval not
containing —1 or 0.
3.2 a) t<0,0<t
Cc ) t>0
oO ) All t
3:3 a) Dependent
c)) Independent
e) ) Independent
f) Dependent
h) Independent
0
) Pi (to)=62 (to) =9
A re linearly independent
3.4
sint+cost
cos 2t+4 sin 2t
ill. ) Cc, sin3t+c, cos3t
d) cy exp[(—1+,/2) it]+c, exp[(—1—./2) it]
13)
1e(as0 Bron)
h) (cp cst)ie "*
(C,il (0k) ——=
Vayian cides, gon

D a P Ey
2 Si 2
b) a,d
16. a) A sin(./2t+a)
13 a:
b) sin | 3t + arcsin ——
3 13
By) S) Dine
3.5 1.2) c.e'+(c cos vee sin ew
c) (cy +c 2t) cost+(c3+cz,t) sint
3 ms. t
e) cy sint+c, cost+exp( )(«sin ~+C4 COS ‘)
2 2
oe V3, : a t
exp ° C5 Sin > +Cg Cos 5
3. a) cy +cot+ a +c¢,¢3
oF :We\+C,4 COS.‘|a )
33 BHO \
c) e, exp 7 —At+c, xp-w At
+c; cos 7 —At+c, sin 7 —At
3.6 oy fave
t+ i
c) t/2 log at —1
4. cyt+c,t?+c3t7' (t>0)
n?>—1/4
7. u’+{1— 5 uO
t
SBE 4. a) (cost) (log |cost|)+t sint+c, cost+c, sint

t
c) c, cos2t+c, sin 2t—4 cos 2t | f(s) sin 2s ds

to
t
+4 sin 2t J f(s) cos 2s ds

to
9. a) e ‘(c, cos,/3t+c, sin ./3t) 4 e,e" +45 te?"

LW (Stet tae zat Vo Cree eee
13. a) c, cos2t+c, sin 2t—it cos2t
c) cye ‘'+c,te7 oie tg? —6t+12—4e!
e) cye*+c,e° 7°2t
+4176 —zkte!
a) c, cost+cy sint—t sint—cos t log|sin ¢|
b) c, cos(2 log|t|)+c, sin (2 log|t|)
+4 sin log|e|
d) cye**+c,te*
+4e!
f) cyt? +c¢5t? log|t|+4
h) cyt[cos log|t|+i sin log|t|]+c,t~ ' [cos log|t|—i sin log|t|]
Jj)) —log|cost|—sin log|sect+tant|+c, cost+c, sint+c,
5
(#) . /10
30 cos ,/2g/15t+ 180(,/15/2g) sin ./2g/15t

75 2g I5\ 2g 1
30— cos /—t+180| /—J]sin /—t+ cos 2t.
2g —60 15 2g 15S 2g—60
a) 0=S cos4t (degrees)
b) 2/z cycles per sec
c) 27/9 ft
d) a ft/sec, accel
=0 ft/sec?
2D) b) =
—(R/2L)t
it R2 = Ps ; 1 R2 1/2
c) n(t8) ema sin(a= t

“Eo slings 'l
26. to) (cos at —coswt), where w Sarg provided w#a
A(k2—k2) akA
3.8
(BRP +a ia wern ag
Amplitude is A [(k6 —k*)? +.a7k?] 1?
4.1
sol} ok -[] =—sint cost
—cost —sint
1a. g(u) iu),where g(u) =e?" e ?f (s) ds
2,
456 Answers to Selected. Exercises
0 1 OE 0
i ||/-7 0 a6" F=85 é
14. a) w= ome fo etl oO
1S 0 2 —13 cost
1
»(0)=| 6
1
4.2 1. a) Unique solution on 0<t<o
b) Unique solution on 0<t<3
T T
4. a) ) Unique solution a Se
6. a) Largest interval not containing 0

b) Largest interval not containing 0 or 3
20. cye7*+c,(t+34)
4.4 2. (—#% cost—4% sint + 3e7'(t—1), —2 cost
++ sint —3e*)
6. a) (cost) log|cost|+t sint+c, cost+c, sint
t t
c) cy cos2t+c, sin2t—
cos 2t |f(s) sin2s ds+4sin2t |f(s) cos2s ds
to to
11. a) e ‘(cy cos, /3t+c, sin, /3t)+c3e7'

+pste*"
12. a) c, cost+c, sint—t sint—(cost) log|sin¢|
b) cye* +c,te*' + fe"
d) elV2/2)t J2 Re
G, COS =f --C>b COS——__f
: D ‘ 2
+c3
;
sin
—— t+c,t
2
sin
2
——t
: 2 ss 2
+ ew V2i2)t oh v2
Gri 5 COS tnt
=2 Col
6 COS 4 ails
=e
Sey
Sil Sea Shit
2
E 2 2
15 wy Eeert
Je.
tana L ee
pee R\~1?2 =(R/2L)t oi 1 EatR12
) (2 *) ‘ sin( x) ?
aE» 1
14. (cosat—coswt) where @ 2 = Le provided w4a
L(w—a)
4.5 3. yr =3yt—2yiyat 33
Y= —2y{ + 2yiy2—2y2
9. c(—e ‘,e~‘), c any nonzero constant
Ite 2
5.1 2) a(c | i )
Cy @ ol
Sy? 1.
1
= =t3a| |
Ss3 ie c) A fundamental matrix:

ene exp(2+,/7) t exp(2—,/7) t
te 3
— See, (2+./7)t sa sp Tat
$e Sst! tT expat fi) NL ee

t
t4c+t4
| s~4b(s) ds, where t—ty>0
to
5.4 15. Node, not an attractor

Node, attractor
Saddle point, not an attractor
6.1 3h o(t)=(t—1)+(t—1) +5 |
aes
co 1-3-5.--(2k—3) ;
6.3 5h
a ee ee eee
2(—1)"(2m +1)=a (1—4m) (6—4m)--(4(k=m+4) |ar
ee,
o()=1-a(2+1) =ta(a+1) (0+2)(a+3)-Ae
Po(t)=1, Pi(t)=t, P2(t)=—2(1—32’),

()=-2(0-3°), Pa(Q=8(1-
1027+94)
1h
118) P(i)=> «t, where co=1, cy=2,
k=0
1+k(k— 1
ae if k>0
ee en
6.5 Me a) ¢,|t|?+c2|t\-?
c) €;lt|+c5|¢7
4
e) (cy +2 log|e)) |¢|+e3 || *
t=O irreg. sing. pt.
t=O reg. sing. pt.
8. t=O reg. sing. pt.

10. t=0, 1 reg. sing. pt.
12. t=Oirreg. sing. pt, t=+1 reg. sing. pt.
6.6 9. cy, (t)+cr¢2(t), where

© iy.
o:)=1+ 2 es i.
sspeerfieS EM 9]
7 wai gle(k—-2)k oY
valid
on —o <t<0 or 0<t<a@
eo)
13. o=ll 2 ale) e, j=1,2

where
1 ee 4 3
ee ae A ga a ae ie Yas.
2(k—1+z,)
¢,.(z;)= rie 25: k 1(z;)
with f(z)=z?+z+1 valid on 0<|t|<0o

15. oilt)=|t =
Ya
Cee where c,= aan
—ke,-1
A
f (z)=2?—-1, valid on 0<|t|< 00
17. None, as t=0 is an irreg. sing. pt.
6.7 3. b) Two solutions valid for 0<|t|<0o; the one corresponding to the root
z=0 of the indicial equation is analytic at t=0.
c) One solution, analytic for all t.
e) If 2y is not an integer or zero, two solutions valid for 0<|t|<o; if
2y is an integer or zero, one solution valid for 0<|t|< oo, and if y=0
or y is an integer, this solution is analytic at t=0.
g) Two solutions valid for 0<|t|< 00, neither of which is analytic at t=0.
4. a) d@=ltl"* Yo os*, where ¢o=1, c,=0,

k=0
oy, Ck-2 ae ees

C= f (k+4) A ) Zz 4
00 k=
c) OE 3. ne where co=1, cer = FG)cm
f=:
f) Ifc is not a negative integer or zero,
ab ala+l)b+1),
ele 3 1-2:¢ (e+1) 5
=F(a,b,c:t) for |i}<1
If c is not a positive integer, there is a second solution
|t\'~°F(a—c+1, b—c+1, 2—c; 2), also valid for 0<|t|<1. Thus, if c
is not zero or an integer, the general solution is
AF (a. b, c; t)+ B\t|)~°F(a—c+1, b—c+1, 2—c; t).
of 1 in each case.
6.8 ile a) o,(t)=lt > ¢,(z;)* with j=1, 2, z,=4,

kK=0
c,(2;) Z
750
2 Cx+1(2;) A a
f (z;+1+k) e f (2) 4
zs 5)
b) o,(t)= Vet. with c,= oer

=0 C
pa(th=lt] )) dyt* +, (t) log|t|, with by =2co,

k=0
by 1 —2(K+1)
Ces
b= for k>1
: (k+ 1) ee
ce) (=I Yc, (z,) t with j=1, 2,
k=0
Se
ates
ee
eae au?
f(k+z))
f (2)=2?—z-2
d ) b=K*
j=l + Yec,t*
=
1-0.
with c,=0,c,=—*“~., =F
Cee
(k>2), f(z)=27+42+3;
balt)=lel-? Y. byt+ad,(t) logit, with

k=0
1 bys 1 —4(2K
+4) Cy44
res
bo =a 0, 3 —
7W
b,=bs=---=0.
a) Indicial equation: z(z—1)+3z=z?—-3z=0;
1 1
a fi
ANemma ee >
Lo@)
P(th=alt Y ct*+B lel? cn:

k=0 k =0
e) 27=0; d()=a Y cyt

oO
+Ble k=0Y bytk+B logic k=0

k=0
Y cyt!
i) 0 is an irregular singular point.
(ce) Qk
k
a) by ()=(e'? ¥ (Qk+)!'
|142 5|—1)F 1-4-7--.3k—5)

o) dull dati ok!
cl
‘
,
d) dy ()=14+t+e, b2(t)= lai It|< 1

= 2 (kt Dik? $2(t)= —4,(t) log |¢|
+7 i>oYore ire|
1
where Hy Ua 2 see
a al
Y
k) $1 ()=Idl e, 62 (t)=¢1(8)log|t|+1- r=2 ——— t*
(k—1)!
1
where Healt
m) ,(t)=J2(t), f2()= Y2(0)
6.9 73). ee p(kt) if p is not an integer or zero;
ool
ivp(kt)+¢2Y, (kt) if p is an integer or zero.
24. ar
Dis a Je Aes (1)
for)
€,t7J2(t)+c9t? ¥,(t)
c) c ny )+¢p1_4)9(t)=e,t *? cosht+c,t~*? sinht
d) etJ oC) Fests 2 (A)=e costa, sini,
where
EQ 0ty hy (t)= —n/2[J, (it) +i, (it)]
6.10 6. p(2)=0, p'(0)=2, q(00)=4' (20) = 4" (10) = 4" (wo) =0
a) Z;=a,Z,=5
ee 42k) Eos
d) Cy 2X Pk+ DI! +cot, |t|>1
ef (ee eee
a) $,(t)~ + + |
Miele “Ast 218i)
2()~—— Oo
|1-—+——_-..
12 1232
118¢ 2!(8t)?
1.2 4. Cy +Cyt
Jan =(n+4)
$,(t)=A, sin(n+3)t, n=O, 1,2,--
t—
Wee ¢,(t)=A, sin ro a n=1, 2,---, where b>a

: b— b-—a
2D a) A,=(n+3)*, ¢,(t)=A, sin(nt+4)t, n=O, 1,...
c) $,(t)=sin \//,t, where A, is a root of
JSan= —tan Saat
e) No real eigenvalues
f) A=0, do(t)=t-1
4
t
a) i=(¥) » bn (t)= Aysin, ally Pee
b) sin a fae cosh “fi, L— cos Gps sinh ./1,L=0,
$, (th= aon [sin 4/2,¢ sinh 4/2,L —sin AL sinh 4/2,t]

sinh 4/2,L
The Agk?:
L#k?: $()=
o(t)= — ——
cos ,/At
i k
jo ewe
it (cos ./An—cos kn) +
(Ak?) sin /2n Ws
Panes $e sinkt tsinkt
k LO c arbitrary
1+cos 2) sin ./2t

| n /2) —4+ cos Jat+4
2 sinn ./2
COS) 20
Caag te
DICOSty/ 216
7.4 “d—By=1
Tks: p(x) (45 — By) = p(0)

Ja,=4n,
|. sin 8nt, 0<t<5,
o,(t)= .
[4,sin 4n(nx—2), astm (=, Dooce
Ted 10. c) cos J1#0
=o08v/4s sin/A(1=1) Renee

J cos,/2
G{t.5,4)=
_ COS At sin,/2(1—s)
JA cos,/A
cos A sin,/At+sin /A cos,/At

123 a)
sin2 ,/2
if sin2./A40
4 eof esa ELE in fro
,/A cos./AL—2 sin,/AL
fee cos. /AL—2 sin, /AL £0

f) No solution
g) ccos2t+4 cost, c arbitrary
» cos
-
os!1)—1) sint—cost+1
sin |
8.1 fi y=1+| {s?+[y(s)]*} ds

sinp(t—s
é ias ‘ pa u.
3. Annee el g(s, p(s) ds
Ins )
9. do(t)=1. (= a j!
14. Yes, to $(t)=0

15. Yes, to d,(t)
tye dal 0 SK LKB 1ot Fae
K,|t—tol }
9.1 7. h=0.2: 6(1)=1.13+.03

1)=1.140+.015
9.2 2. 1.1459
9.3 2. 1.1463
9.4 Beeyies
9.5 3. a)
) 1.1076
) 10.3238
) 101.0075
) i) 34.411
ii) 59.938
iii) 64.858
b) i) 0.293
ii) 0.3451
iii) 0.350245
1V ) 0.350234
9.6 if Euler: yy n41=Yi,nthhi lt» Vivn: Yo,n)
Yo,n+1=Ya,nthfo (tn Vin: Yo, n)

10.2 3:
s?+4p
0:ideee
1
SS
1
R)
aie)
“eran
‘ yk
:
er)
s>0
1 —2 exp(—as/2)+exp(—as)
Rs>0
s(1—exp(—as))
12
2) (s? +o?) (L—e 9”)
1
% (s*+1)(s+1)
ey ga eis
c) $52 (*) ; Ks>0
4)2m (3s* — w?)

Seay
10.3 He a) —2+3e'+}e°*!
c) 1—cos 2t—t sin2t
t
e) Jsin(t—u)
f (u) du
cos 2t
10.4 3: @. Ne ge ON en tt — ent)
Se! —e+4he 3t
v3 t
fre — see (os 3t+ V3sin 3 nce)
ds (25e7' + 23e" **)—§ sint

1(et-1_g- (0)
2
cos 2t+4 sin2t+

3J sin2(t—x)
f(x) dx
(0)
jse7'
+a cos3t+ sin 3t
a) 2e7"(1+t)—} sin 2t
c) 3—2 cost+2[(t—4) sin(t—4)] U(t—4)
_ a) —14+2t+2t?7+e°*
b) t? +35t°
c) 1+2te’
d) at
e) 44+317+7,t*
10.5 bi(t)=2t+1, $2(t)=—2t
bade "+90 'eahe™, dall)=Hert—e%)
#(t) is any function for which #,(0)=1 and $(0)=0;
bi(t=e—J 10) ep,(u)du

Index
Abel’s formula, 79, 145, 311 Binomial theorem, 166

Absolute convergence, 166, 215, 234, 262 Birth rate, 3
394, 402 Block diagonal matrix, 443, 445
Absolute value, 429 Boundary, 31
Acceleration of gravity, 37 Boundary conditions, 278, 281, 283, 284
Adams method homogeneous, 292, 293
explicit, 381, 383 nonhomogeneous, 292, 293
implicit, 381, 383 periodic, 287, 295, 297, 298, 299
Air resistance, 44, 110 self-adjoint, 297
Amplitude, 92, 93, 112, 286, 340 separated, 294, 295, 297, 298, 299, 309
Analytic, 215, 216, 231, 250 Boundary-value problems, 278-313
at infinity, 264, 267 homogeneous, 281-288, 289, 304, 309
function, 215, 216, 231, 250 nonhomogeneous, 288-293, 306-313
Annihilator method, 106 nonself-adjoint, 303
Applied voltage, 99, 111, 155 self-adjoint, 293-298, 299, 301
Approximate solution, 359 singular, 303-306
Approximation methods, 357-391 Buckling of a column, 281
Asymptotic behavior, 85, 192, 194-197,
340-346 Canonical form, 438
Asymptotic expansion, 271, 272, 274 Jordan, 445—448
Asymptotic series, 271, 275, 324 of real 2 x 2 matrices, 183-184
Attractor, 192 of 2 x 2 matrices, 180-182
Auxiliary equation, 84 Capacitance, 26, 83, 99, 110, 129, 155, 412
Cauchy criterion, 336
Basis for sequences of matrices, 435
of solution space, 137 Cauchy polygon, 360
of vector space, 137, 140, 438 Cauchy’s inequality, 225, 242
Bernouilli equation, 24 Center, 191, 192
Bessel equation, 130, 232, 244, 251, 253-263, Chain rule, 45, 347
264, 270, 274, 305 Characteristic equation, 88, 89, 94, 229
linearly independent solutions of, 256-260 equal roots of, 84, 89, 90, 91
Bessel functions, 253-263, 275 Characteristic functions (see Eigenfunctions)
behavior as t0, 263 Characteristic polynomial, 84, 93, 170, 174,
of first kind, 255 201, 409, 444, 445
of index zero, 258, 260, 415 Characteristic values (see Eigenvalues)
of second kind, 258, 260 Chemical processes, 18, 27
zeros of, 261 Circuit (see Electrical circuit)
Binary fission, | Comparison test, 216, 227, 243, 321
466 Index
Comparison theorem, 261 asymptotic behavior of, 329

Competing species problem, 195-197 linearly independent, 287, 297, 301
Complex conjugate, 429 orthogonality of, 297, 300, 305
Complex numbers, 429-432 Eigenvalues, 165, 169, 170, 182, 192, 198,
Complex-valued functions, 85, 394, 429-432 207, 440, 441, 442, 443, 444
Complex-valued solutions, 86 asymptotic behavior of, 329
Compound interest, 5 of a boundary-value problem, 283, 285,
Computational stability, 383 286, 287, 288, 290, 300, 309 -
Conditional instability, 384 complex conjugate, 173, 183
Confluent hypergeometric equation, 252, distinct, 180
264, 266, 277 infinite sequence of, 286, 297, 298, 302
Consistency of numerical methods, 381-385 multiple, 171, 181, 197
Continuation of solutions, 333-339, 349 real, 297, 300
Continuity simple, 170
of matrix functions, 116 with negative real part, 192
of vector functions, 116 Eigenvectors, 165, 169, 170, 175, 182, 207,
with respect to initial conditions 353-356 438, 439, 440, 441, 442, 443, 444
Continuous dependence of solutions, Einstein, 60
353-356 Electrical circuits, 25, 83, 99, 110, 111, 206,
Convergence 410, 412, 414
of integrals, 394 Electromotive force, 25
rate of, 383 Equilibrium, 194, 196, 279
of sequences of matrices, 165, 433-434 Equivalent system, 123-126, 131, 143-144,
of sequences of vectors, 160 151-153, 158, 179, 187, 192, 204, 205
of successive approximations, 321 Error bound, 360
uniform, 320, 339 Error of a method, 360
Convergent method, 381-385 Errors, 323, 360-365, 370, 375, 384
Convolution, 401, 402, 414 Escape velocity, 55
Corrector, 375 Euclidean distance, 160
Cramer’s rule, 102, 105, 140, 425 Euclidean length, 159, 160
Critical point, 185, 193, 195 Euclidean space, 158
Critical speeds, 279, 281 Euler, 430
Current, 25, 110, 111, 129, 155 Euler equation, 228, 230, 264, 265
Cylindrical symmetry, 303 Euler method, 357, 360-367, 382, 383, 386
388, 390, 391
Damping consistency of, 383
critical, 91 convergence of, 383
light, 91 Euler’s constant, 261
(see Overdamping) Euler’s formula, 430
Death rate, 3, 4, 193 Euler’s identity, 430
Dependence on initial conditions and Existence and uniqueness, 32, 33, 58, 59, 130,
parameters, 353-356 314-356, 358, 406
Determinant, 424-426 for scalar equations, 314-339
of coefficients, 284 for second-order equations, 60, 66
Difference equation, 359 for systems, 346-356
Differentiability, 6 Exponential growth at infinity, 351, 353, 394
of matrices, 116, 118 Exponential integral function, 271
of vector functions, 116, 118, 134, 161 Exponential matrix, 165, 433-435
Differentiated series, 215
Dimension, 81, 135, 198 Flow diagram, 367, 372, 387
Dirac delta function, 413 for Euler method, 367
Direct sum of subspaces, 436, 437, 446 for Milne method, 387
Direction field, 28, 29 for modified Euler method, 372
Distance function, 160 Force, 37
Distributions, 412, 413 external, 109, 111
gravitational, 38, 45
Eigenfunctions, 283, 288, 295, 302, 309 periodic, 111
Index 467
Forced oscillation, 112 Index (of subspace), 443

Formal series, 225 Indicial equation, 234, 237, 238, 240
Formal solutions, 225, 241, 267, 269, 270, double roots, 237, 246
2S roots of, 240, 242, 245, 249, 250, 251, 265
Free-end condition, 281 Indicial polynomial, 234, 237, 252
Frequency Inductance, 26, 83, 99, 110, 129, 155, 412
applied, 26, 38, 109, 111 Induction, 69, 173, 219, 226, 235, 243, 246,
natural, 38, 109, 111 254, 319, 321, 339
of vibration, 110 Inductor, 25
Friction, 44, 109 Initial conditions, 2, 8, 12, 40, 41, 46, 156
Functions Initial-value problem, 8, 40, 41, 119, 120,
analytic, 215, 216, 231, 250 128, 156, 314-356
bounded, 317 Integrability
complex-valued, 85, 394, 429-432 of matrices, 117, 161
unbounded, 112 of vector functions, 117, 161
vector-valued, 116, 118, 134, 161 Integral equations, 316, 322, 324, 328, 335,
Fundamental matrix, 139, 140, 142, 147, 348, 350, 355, 403
148, 149, 167, 174, 197, 416, 419 of Volterra type, 315
Fundamental set of solutions, 80, 137, 144, Integration by parts, 294, 303, 396
146, 154 Interpolating polynomial, 381
Fundamental theorem of calculus, 21, 148, Intersection (of subspaces), 436, 437
315, 329 Interval of convergence, 215, 239
Invariant subspace, 443
Gain function, 412, 414
Inverse function theorem, 19
Gamma function, 254-256, 423
Inverse Laplace transform (see Laplace
General solution, 80,85, 91, 94, 98, 100, 107, transform)
184, 307 Irregular singular point, 231
amplitude, phase shift, form, 91 at infinity, 264, 267, 269, 272, 275
Generalized eigenvectors, 442, 443, 444 Iterative formula, 359, 373, 390
Generalized function, 413 implicit, 372
Geometric interpretation of solutions, 7
Geometry of first-order equations, 34 Judicious guessing, 106, 111, 179, 180
Global existence, 326, 338, 349
Gravity, 44, 49, 110 Kinetic energy, 422
Green’s function, 288, 291, 292, 307-313
Kirchhoff’s law, 26, 129
properties of, 291, 307
uniqueness of, 291, 307, 308, 312 Laguerre polynomial, 244
Gronwall inequality, 329, 331, 332, 333, 343, Laplace transform, 351, 392-423
349, 351, 353, 354, 356, 379, 380, 415 basic properties, 392-397
Growth rate, 3, 193 of derivatives, 396
Half-life, 4 existence of, 394
Hermite polynomials, 220 inverse, 400, 416, 417
High-speed computation, 357 inversion of, 400-405
Hooke’s law, 39, 109, 128 of a product, 401, 402
VH6pital’s rule, 24 of a rational function, 402, 403
Hypergeometric equation, 232, 233, 244, tables of, 393, 404, 422, 423
251, 252, 264, 265 uniqueness of, 393, 406
Legendre equation, 221, 231, 233, 244, 252,
Identity theorem 265, 306
for analytic functions, 224 polynomials, 222
for power series, 215 Length of a vector, 159, 348
Imaginary part of a complex number, 429 Linear combination, 137
Implicit function theorem, 13 Linear dependence, 71
Implicit methods, 372, 381 Linear differential equations, 65
Improper integrals, 57 of first order, 20, 50, 79, 96, 325, 398-400
Improper nodes, 188, 190 . homogeneous, 69, 79, 393
Improver Euler method, 372, 388 nonhomogeneous, 69, 70, 99
468 Index
of order n, 99, 124, 406, 407 Method of majorants (see Majorants)

second order, 40, 123, 143, 204, 405 Method of undetermined coefficients (see
with analytic coefficients, 217 Undetermined coefficients)
with variable coefficients, 95 Midpoint quadrature, 368
Linear differential operators, 68-70, 81 Milne method, 363, 372-381, 383, 386, 387,
Linear independence, 71, 72, 74, 75, 77, 136 388, 390
Linear systems, 114-156 Modified Euler method, 368-372, 388, 390
of algebraic equations, 170, 393, 418, Momentum, 37, 39
424-426 Multi-step methods, 383, 384
applications of Laplace transforms, consistency of, 383, 384
415-422 stability of 383, 384
with constant coefficients, 165-211
existence and uniqueness, 349-353 Natural frequencies, 111, 127
homogeneous, 134 Negative semiorbits, 92, 93
initial-value problem, 128 Neumann form of second solution of Bessel
nonhomogeneous, 147, 177 equations, 261
Linear transformations, 438 Newton, |
Linearity Newton’s law
of inverse Laplace transform, 405 of cooling,5
Linearly independent sets first law, 37
of eigenfunctions, 297, 301 of gravitation, 44, 48, 49
of eigenvectors, 136, 173, 175, 197, 198, second law, 37, 44, 45, 128, 210
438, 439 Nilpotent, 444, 445, 447
of formal solutions, 269, 276 Node, 188, 192, 195, 197
of functions, 216 Nonhomogeneous boundary-value
of solutions, 74-82, 84, 89, 90, 97, 98, 100, problem (see Boundary-value
104, 105, 136, 138, 139, 154, 202, 225, problem)
233, 235, 236, 239, 243, 247, 249, 257 Nonhomogeneous linear equation (see
of vectors, 135, 136, 173, 175 Linear differential equation)
Liouville, 299 Nonsingular matrix (see Matrix)
Liouville normal form, 302 Norm
Liouville transformation, 301, 329 of exponential matrix, 434
Lipschitz condition, 318, 320, 322, 325, 327, of a matrix, 433-435
331, 333, 347 of a vector, 159, 348
Lipschitz constant, 318 Null space (of a linear transformation), 443
Local problems, 314 Numerical approximation methods, 357-391
Adams, 381, 383
Majorant series, 225—227 Cauchy polygon, 360
Mass, 37 Euler, 360-367, 382, 383, 386, 388, 390,
Mass-spring system, 37, 65, 83, 110, 111, 127 391
Mathematical models, 1—5, 8, 25-28, 37, 40, improved Euler, 372, 388
47, 314, 340 Milne, 363, 372-381, 383, 386, 387, 388,
Matrix, 115 390
continuous, 116 modified Euler, 368-372, 388, 390
diagonal, 439 Runge-Kutta 369, 385-389, 390, 391
differentiable, 116 Numerical instability, 384
fundamental, 139, 140, 142, 147, 148, 149,
167, 197 Ohm’s law, 25, 26, 129
integral of, 117 One-sided derivative, 6
nonsingular, 171 One-step method, 369, 378
similar, 187, 209, 438, 442 Operation, 68, 394
Matrix differential equation, 206 linear, 68, 69, 392
Matrix function, 434 Orbit, 185, 188, 189, 190, 191
Matrix norm, 434 Ordinary points 231
Mean-value theorem, 317, 346, 362, 366 at infinity, 264
Method ofjudicious guessing (see Judicious Orthogonal trajectories, 17
guessing) Orthogonality, 296, 309
Index 469
of eigenfunctions, 297, 300, 305 Riccati equation, 99

with respect to a weight function, 300 R-L-C circuit, 25
Oscillation, 65, 85, 91, 156 Rocket flight, 49, 54, 55
damped, 85, 91, 156 Rotating shaft, 281
Osgood uniqueness theorem, 332, 333 Rotating string, 278-281, 283
Overdamping, 91 (see Damping) Round-off error, 364, 381, 382, 384
cumulative, 381, 382
Partial fraction decomposition, 15, 402, 403, Runge-Kutt a method, 369, 385-389, 390, 391
406, 407, 409, 413
Particular solution, 100, 102, 107 Saddle point, 189, 192, 197
Pendulum, 37, 43, 55, 65, 83, 93, 110, 127 Salt solution, 27, 28
damped, 340-346 Schwarz inequality, 159
linear. 95.127 Second linearly independent solution, 245,
nonlinear, 66, 340-346 246, 248, 250, 251, 259
Period, 93, 340 Self-adjoint, 293
Periodic function, 24, 396 Semiorbit (see Postive semiorbit,
Permutation, 424 negative semiorbit)
Phase log, 412 Sequences
Phase plane, 91, 92, 93, 185 of matrices, 433-435
Phase portrait, 185, 188, 189, 190, 191, 192 of vectors, 160
Phase shift, 92-93 Series, 166, 167, 212, 214, 217, 220, 221, 239
Poincaré, H., 271 of matrices, 165
Polar coordinates, 430 Simple harmonic motion, 38
Polygonal curve, 359 Simpson’s rule, 363, 372, 373, 374
Polynomial equations, 427-428 Singular boundary-value problems, 303-306
Population, 1, 2, 3, 15, 17 (see also Boundary-value problems)
Positive semiorbit, 92, 93 Singular points, 228-266
Potential energy, 422 at infinity, 263-266
Power series, 167, 212, 214, 217 Solution, 6, 7, 118
coefficients of, 212 in closed form, 212, 220
convergence of, 220, 221 graphical, 28, 29
generalized, 239 by Laplace transforms, 406-422
Power series expansion, 369, 385 linearly independent, 136
Power series solution, 217-277 real form, 91
Predator-prey problem, 192-195 Solution curve, 7, 156
Predicator, 375, 388 Solution matrix, 139, 142
Predicator-corrector, 375 Solutions
Probability, 2 asymptotic behavior, 85, 192, 194-197,
Projectile, 42 340, 346
Proper node, 189 bounded, 91
Punctured interval, 239 complex-valued, 85, 86
in power series, 217-277
Radioactive decay, 4 real, 137
Radius of convergence, 214, 215 about a regular singular point, 232-253
Ratio test, 216, 219-227, 236, 243 unbounded, 109, 112
Real part, of a complex number, 429 Spiral point, 191, 192, 195
Recursion formulae, 218, 224, 240, 247, 248 Spring
274 linear, 18, 19, 38-43, 109
Reduction of order, 95 nonlinear, 41
Region, 5, 30 Spring constant, 39
Regular singular point, 230, 231, 232-253 Square-wave function, 397
at infinity, 264 Stability (of a numerical method), 364,
Relativity, 60 381-385
Resistance, 25, 83, 99, 110, 129, 155, 412 Standing wave, 279
Resistor, 25 Starting method, 369, 386
Resonance, 38, 111, 112, 156 Starting values, 369, 384, 386, 388, 389, 390
Resonant frequency, 112, 156 Steady state, 27, 411, 414
470 Index
Stretching transformation, 276 local, 361, 362, 363, 364, 370, 379-381
Sturm, 299 of Milne method, 363, 375-379, 380, 381
Sturm-Liouville boundary-value problem, of modified Euler method, 370-371
298-303, 329 of predictor, 375
Sturm-Liouville operator, 306 of Runge-Kutta method, 369
Subspace, 135, 174, 197, 199, 201, 436 in Simpson’s rule, 376-378
invariant, 443, 444, 445, 446 for successive approximations, 323
Successive approximations, 317-323, 325, Tuning a radio, 412 2
326, 327, 328, 339, 347, 350 Two-dimensional systems 180-192
Superposition, 69, 70, 106, 210 Two-step methods, 368, 381
Systems
of first-order equations, 156 Undetermined coefficients, 106, 111,179, 180
of linear algebraic equations, 170, 393 Uniqueness
418, 424-426 of inverse Laplace transform, 405
of linear differential equations, 114-156 of solutions, 329-333, 339, 348, 349
numerical methods for, 389-391 of solutions of nonhomogeneous
triangular, 122 boundary-value problems, 290, 311
Variables separable, 9, 10, 12, 18, 20, 22, 53,
Tangent, to a curve, 7 54556, 1218331" ;
Tangent field, 28 Variation of constants formula, 101, 102
Taylor, formula, 385, 386, 387, 388 148, 149, 151, 152, 154, 177, 228,
Taylor serires, 213 289, 307, 343
Taylor’s theorem, 385, 386, 388 Vector algebra, 414
Tension, 45, 279 Vector functions, 116, 118, 134, 161
Transcendental equations, 285, 286, 288 Vector space, 134, 288
Transfer function, 411, 414 axioms for, 134
Transient, 27, 411 basis for, 135, 136
Trapezoidal rule, 363, 372 dimension of, 135, 137
Triangular matrix, 122, 173 finite-dimensional, 135
Triangular system, 122 Vector sum (of subspaces), 436, 437
Trivial solution, of boundary-value Vectors, 115, 136
problems, 280 derivative of, 116
Truncation error, 384 integral of, 117
cummulative, 361, 362, 363, 364, 365, Voltage, 25, 26, 129, 155, 411, 412
379-381
of Euler method, 364, 365 Weber form of second solution of Bessel
of explicit Adams method, 381 equation, 261
of implicit Adams, method, 381 Wronskian, 76-69, 84, 86, 97, 143, 294, 311
b
so nudalee
fgcone ae *

Brauer, Fred - Nohel, John A. - Ordinary Differential Equations - A First Course

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Brauer, Fred - Nohel, John A. - Ordinary Differential Equations - A First Course

Uploaded by

Copyright:

Available Formats

a i

Copyright © 1973 by W. A. Benjamin, Inc. Philippines copyright 1973 by W. A. Benjamin, Inc.

Differential equations are essential for an understanding of many important

study Chapters 4 and 5 of this book in a first course in differential equations.

Madison, Wisconsin F.B.

Chapter 1 Introduction to First-Order Differential Equations

Chapter 2 Introduction to Second-Order Differential Equations

Chapter 3 Linear Differential Equations

Chapter 4 Linear System of Differential Equations

4.2 The existence and uniqueness theorem 130

Chapter 5 Eigenvalues, Eigenvectors, and Linear Systems with Constant Coefficients

Chapter 6 Series Solutions of Linear Differential Equations

Chapter 7 Boundary-Value Problems

Chapter 8 Existence Theory

Chapter 9 Numerical Methods of Solution

Chapter 10 The Laplace Transform

Appendix 1 Determinants and Linear Systems 424

or ern ucstips wit

‘luis je eane bad fran, “a

ina # sue i ORC aoe ps ox

1.1 SOME POPULATION AND

factory. For example, if the food supply to a population was suddenly

We recall that d(log |y(t)|)/dt=y'(t)/y(d). Thus the above equation can

and therefore integration from zero to ¢ yields

log|y(¢)| — log |y(0)| =at.

or, since A>0,

y()=ay(t)—by(), yO)=A. (1.5)

Example 1. Assume that a radioactive substance decays at a rate proportional to

Suppose that y(0)=A. Then, exactly as in (1.2), the solution is

J According to Newton's law of cooling, the rate of change of temperature of a

1.2 FIRST-ORDER EQUATIONS

* We postpone a precise definition of region to Section 1.6.

+ By differentiability of a function at an end point of an interval, we mean the existence

Such a function ¢ is called a solution of (1.8) on: the interval /. For

1. Sketch these functions for c=0, +3, +1.

is a solution of this equation on each of the intervals —coo<t<—1, —l<tr<l,

from which we get 5S—Sce=1+ce, and therefore c=4e1. Thus

is the desired solution.

* 8. Apply Exercise 7 to determine the maxima and minima of solutions of

We shall now study some methods of finding solutions of certain types

1.3 EQUATIONS WITH VARIABLES SEPARABLE

We shall say that a differential equation of first order

f(t, y)=9(t) hy). (1.10)

Example 1. Solve the initial-value problem

y=y?, — v(to)=Vo (1.11)

This integral is easily evaluated; its value is

u u ¢ (to) 3 bl” blto)

or, solving for (rt),

rectangle D=((t, y) |a<t<b, c<y<d} (see Fig. 1.4). Let @ be a solution

|fam |204 ((515)

of (1.15) must be investigated again as above. As before, the constant func-

Since ((0)=y,=1, letting v= (s) as in Example 1, we obtain

|ay |uen ea:

Example 4. Solve the initial-value problem

y=ay—by*, y(0)=A, (1.16)

d (t)=ap(t)—bo* (1). (1.17)

Integration from 0 to r yields

To evaluate the integral we use partial fractions

$ (t) $ (t) $(t)

1, JO) 1), |abo

Substituting this result in (1.18) yields

Nas eran rey eee

In the simple model discussed in Section 1.1, the solution continues