Digitized by the Internet Archive

in 2024

[Link]
ORDINARY DIFFERENTIAL EQUATIONS
i
i]

‘
=
Weegee
 H. K. WILSON
Southern Illinois University, Edwardsville

ORDINARY DIFFERENTIAL EQUATIONS

Introductory and Intermediate Courses Using Matrix Methods

Wy ADDISON-WESLEY PUBLISHING COMPANY

Reading, Massachusetts - Menlo Park, California - London+ Don Mills, Ontario
This book is in the
ADDISON-WESLEY SERIES IN MATHEMATICS

Consulting Editor
Lynn H. Loomis

Copyright© 1971 by Addison-Wesley Publishing Company, Inc. Philippines copyright 1971 by

Addison-Wesley Publishing Company, 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 Congre
ss Catalog Card No. 78-133897.
For Jinny
a)
PREFACE

This textbook has grown out of lecture material which the author has assembled for
both elementary- and intermediate-level courses in ordinary differential equations.
The introductory material, consisting of topics from Chapters 1 through 6, was
originally designed for a group of mathematics majors who studied differential equa-
tions for a period of two quarters bridging the sophomore and junior years. The
presentation makes use of elementary matrix methods. This approach is taken in
order to reduce the student’s difficulties in relating his first course in the subject to
more advanced studies which he might later take. A formal course in linear algebra
is not assumed as background. The linear algebra which is needed is contained in
Sections 3.2, 3.3, 3.6, and 3.10 through 3.13. It is assumed, however, that the student
has had some exposure to determinants, the solution of specific systems of linear
algebraic equations by row elimination, complex numbers, and polynomial algebra.
A familiarity with the material in, say, the first edition of the algebra and trigonometry
textbook by Fisher and Ziebur is adequate background in this area.
The intermediate-level material, consisting of topics from Chapters 7 through 10,
is directed toward senior mathematics majors and engineering graduate students.
The understanding of it requires some familiarity with several notions from advanced
calculus and linear algebra. A few topics from elementary analysis have been in-
corporated into the exposition (Sections 7.4, 8.2, 8.3, and 8.4) to minimize prere-
quisites as much as possible.
One can base a variety of courses at different levels on the material in this book,
and several possibilities are suggested in the course outline table following the contents.
The asterisks in the table indicate sections which an instructor might wish to omit,
and the sections marked with a dagger may be omitted if the students have had
equivalent material in linear algebra or analysis. The courses labeled A through M
are first courses, and the others are intermediate or advanced courses. More specific
descriptions are given with the table. By his choice of material from Chapters 1 and

Vii
Viii Preface

2, which are intended primarily for motivation, the compiler of a syllabus can vary
the flavor of a course from very applied to almost pure. Sections 1.1, 1.2, and 1.3
must be covered, however; they teach solution methods for the linear equations
x = 0 and x’ = a(t)x + b(t). Sections 3.13 and 6.6, together with the proofs of
Theorems 4.1 and 4.2, and Lemma 6.1 may be omitted.
I should like to thank those of my colleagues who have read portions of the
manuscript and especially acknowledge the helpful suggestions of R. Kurth. The
ideas for several of the physical problems were suggested to me by M. B. Sledd.
Example 5 of Section 8.9 is based on J. Serrin’s treatment of the Blasius boundary
value problem, and the material in Section 9.8 is based on an illustrative problem of
L. Markus. Finally, I should like to acknowledge the contributions of my wife who,
in addition to providing much encouragement, has worked tirelessly in typing and
editing the manuscript and its revisions.

Edwardsville, Ill. H. K. W.
November 1970
CONTENTS

Chapter 1 Differential Equations and the Physical World

1.1 An object falling in vacuum

12 Linear friction
3 An object falling in an ence phere
1.4 The escape of a rocket .
1k An orbiting satellite :
1.6 A linear oscillator with one depres ne econ F
i, Linear oscillators with several degrees of freedom .
1.8 The series RLC circuit .
iES The parallel RLC circuit
1.10 An electronic oscillator

Chapter 2 Solution Methods for Special First and Second Order Nonlinear Equations

Dal Preliminaries
Ded Initial value problems
2s The separable equation Z
2.4 First integrals and implicit poltiens ;
Des The exact equation .
2.6 Integrating factors
Wey Reduction of order .
2.8 A soft spring oscillator

Chapter 3 Matrix Methods for Linear Equations with Constant Coefficients

3.1 Simply coupled systems of differential equations

Se Addition and scalar multiplication of vectors
3h8, Matrices . : :
3.4 Vector- and matrix- valued fonctions :

ix
x Contents

3) The exponential matrix 54

3.6 Determinants ; 60
Shi General solutions for ‘eae atrerential eins 63
3.8 The nth order homogeneous equation 70
3°9 The annihilation method Ie
3.10 Linear independence of vectors dy
Sell Linear algebraic equations 80
Sil Eigenvectors . 82
Sells} Jordan canonical form tor 2 x p) faaerees 84

Chapter 4 The Theory of Linear Differential Equations

4.1 Preliminaries 90
4.2 Initial value problems 91
4.3 The existence of solutions . 94
4.4 The first order scalar equation 99
4.5 Fundamental solution sets for the Homioseneste canon 103
4.6 Relations between solutions of the homogeneous equation 110
4.7 Solutions of the nonhomogeneous equation.. 115
4.8 Fundamental solution sets for homogeneous cautions win coment
coefficients 1A

Chapter 5 Solving Linear Equations with Laplace Transforms

Ss Definition of the transform 127

4 Solving homogeneous equations : 133
5)! Laplace transforms for some special fancdene : 157)
5.4 Solving nonhomogeneous equations . 143

Chapter 6 Power Series Solutions for Linear Equations

6.1 Analytic functions 147

6.2 Power series solutions otra Ord nae aesite 150
6.3 Regular singular points iW |
6.4 Euler’s equation : 159
6.5 Series solutions around recner arty: ponte : 161
6.6 Proofs for the series representation theorems 171
6.7 Bessel’s equation 176

Chapter 7 Qualitative Behavior of Solutions for Linear Equations

el Preliminaries : 182
Yer The homogeneous praion ah consent eoeticents ; 184
7.3 The homogeneous equation with periodic coefficients . 193
7.4 The homogeneous equation with continuous coefficients 204
“ES The nonhomogeneous equation 218
 Contents xi

Chapter 8 The Existence of Solutions

8.1 Preliminaries 224

8.2 Open sets and closed ae 225
8.3 Continuity of vector functions 229
8.4 Connected sets 25
8.5 A geometrical omrettion 234
8.6 The Lipschitz condition 238
8.7 The method of successive aDDIDURaTOnS 241
8.8 Maximal solutions . : 246
8.9 The estimation of escape times 249
8.10 Dependence on initial values 256

Chapter 9 Autonomous Systems

hil Preliminaries 263

oe The shift formulas 264
95 Phase portraits : 267
9.4 Divergence and Bendneont S neeatve eaten : 273
5 Limit sets of solution paths 278
9.6 Limit sets in the plane . 283
Dell Critical points in the plane 291
9.8 Phototropic platyhelminthes : 297
DE The equations of van der Pol and Tenntd 299

Chapter 10 Stability

10.1 Definitions 304

10.2 The stability of jinear ueie 310
10.3 The stability of equilibrium for puesta ee 315
10.4 The stability of periodic solutions 318
10.5 The direct method of Lyapunov 324

Further Reading . 341

Hints and Answers for Selected Exercises 343

Index 373
 Description

A Very few applications

B Linear applications only and no electrical circuits
(€ Linear applications only
D-H No electrical circuits
I-M General

N Systems of linear equations

O General, without Lyapunov stability for nonlinear equations
J? General
Q General, with more emphasis on ad hoc techniques
R Nonlinear oscillations for autonomous systems

es May be omitted
May be omitted if students have had necessary background
T

xX To be used

xii
 7

Table of Course Outlines

ABCDEFGHIJKLMNOPQR ABCDEFGHI JKLMNOPQR
IIT DE OOK KOSS IS OK KRIORI CHU XO XOXEXG OX OXON OO
NOO XOXOXO
O KONO A ROKK OS OO OOK SOK OK
LES NON SOOO XOONOXGXO
XENON OS SONOS XOXO XO XOXO KONO OX
1.4 XXXXXXXXK XX MOY OOOO ORORYO COCO
1S: POON OOK NOOK LEMOS, DOO O, OR OOOO, OOK
1.6 LOR OPOOOOOCOC COOOL: 6.6 * * * * # *# KK OK RK KR OK
ilo7/ ESD XONONONONE XOXOXO XO KOK te 6.7 Oe ere St as Sea
1.8 x XXX XK X
1.9 x XXX XX Tes! XXX X
1.10 ot AEE Sale | ee XXX X
7.3 XX XX
2.1 XXXXXKXXXXX XG EG Ee
De XXXXKXXX XK XX XK | 15) ey ees
7,3} XXXXXKXKXXKXX xXX
2.4 XXXXXKXXXKXX OXG Sel XXX X
Ds OO DOO. oO titt

OF xox Ox 8A ++ +H
2.8 x x XX! 8.5 XXX X
8.6 XX XX
3.1 XXXKXXKXKKXKXKXKKXKXKXX 8.7 XX XX
Se aimiei
aie) bot tet et 8.8 XXX X
oeetaia
te eter tet 8.9 XX XX
EN NO NO NOG NO NON ONO NON 8.10 xX XXX
BS) IPO OOK DS IOSD DOK
SO Meena
kl. Jt akeda tetas 9.1 XX XX
SON NON KON OXON ONG KOXKE NGG 9.2 xX XXX
BS DOCK ORO O.O,.OR IS 9.3 XXX X
39 * ee OK KR KK RK KK 9.4 YOSCOMOK

SLUeiie le tte eb TUT 9.5 XX XX

SDUU Sf SI Se ce Te a oH 9.6 XX XX
hg! A Fe ea a 9.7 XX XX
3.13 * * * FR ROK ROR OK OK KK 98 * OK OK
9.9 Shee

41 XXXXXKXKXKKXKXKXKKXKXX
AD X XOXOXO XOXOXO XOX XE 10.1 xX
43 XXXXXXXXKXXXXKXKXKXKXKX 10.2 xXX
44 XXXXXKXKXXKKKXX 1033 xX X
45 XXXXKXXXXKXKXKXKKXKXKXKX 10.4 xXX
ANS KDR YK OK ROK OOK OOK OS YK OKO 10.5 xXX
ANTE YK IER DDK EDDIE
DEDEDE IOS
48 *« * © OR RR KOK OK OK RK OK RK OK

5.1 * oe Ef) ES fy x
Ew) * CRS
Re be ee
ee * “pes
62 fo Se
5.4 * “ou ) Cae xe

xiii
CHAPTER 1

DIFFERENTIAL EQUATIONS AND THE

PHYSICAL WORLD

What functions equal their second derivatives? In asking such a question, one is
really ageing for the solutions of the aot equation en sink, coaht,
= ;
dp 7 * PONS oot
The equation is said to have been solved when ai// of the unknown functions x = ¢(t)
which satisfy it have been found. It is easy to guess a few solutions. For example,
x = e'and x = e ‘ equal their second derivatives. This is also true of x= 2e‘ + 7e~.
In fact, any function having the form x = c,e’ + coe*, where c, and cp are con-
stants, equals its second derivative. We shall see later that no other type of function
can have this property. twice aiff,
The remarks in the previous paragraph indicate that purely academic questions
can be a source of differential equations. Another, richly varied source of differential
equations has been the attempt to describe and study the physical world. One of the
oldest sources of differential equations has been the study of the motion of objects
and of fluids, but differential equations have also been used to describe electrical
effects, elastic deformations, and many other phenomena which are not directly
associated w:th motion.
This chapter consists of examples of physical situations which can be discussed
with the aid of differential equations. The examples are specifically intended to
illustrate origins for some of the differential equations occurring in later chapters.
Sections 1.1, 1.2, and 1.3 contain important solution techniques. Other sections
contain explanations of physical techniques which may be ofinterest to some students,
but not to others.
1.1 AN OBJECT FALLING IN VACUUM
Consider an object having mass m that is dropped from an altitude Ro, measured
from the center of the earth. Several questions might be asked about the motion of

1
 11
2 Differential equations and the physical world

the object. For example, what are its altitude and velocity at a given time? What
is its velocity at a given altitude?
Two of Newton’s laws are used to analyze this situation:

Newton’s Second Law of Motion. Relative to a nonaccelerated coordinate system,

the time rate-of-change of the momentum of a system in a given direction equals the net
external force acting on the system in that direction.

Newton’s Law of Gravitation. Two bodies, having masses m, and Mo, will exert
an attractive force on each other; the force acts along a line through their centers of
mass; and if r is the distance between their centers of mass, then the magnitude of the
force is proportional to m ym Bes

This force of attraction between two bodies is called gravitational force. The
magnitude of the gravitational force exerted on an object by the earth is called its
weight. It is given by Gmm,/r®, where m, is the mass of the earth, r is the distance
of the object from the center of the earth, and G is a proportionality constant. The
value of G, which we shall call the Newtonian gravitational constant, is 6.670 X
10~'! newton
- m?/kg?. ME/-a = GM /L) C= L3/ (m7)
The quantity g = Gm,/r® is called the acceleration due to gravity at altitude r.
The value of g at the earth’s surface varies since the earth is not perfectly spherical.
In Central America, g = 32.094 ft/sec”; in the arctic, g = 32.235 ft/sec”. Thus
the weight mg of an object having mass m depends on its altitude.
We shall assume for the study of our falling object that the earth is perfectly
spherical and that its radius R is such that g = 32 ft/sec” exactly. We neglect atmos-
pheric friction and the rotation of the earth. Further, we assume that the altitude Ro
from which our object is dropped is so near R that the weight mg of the object is
constant during its descent.
The time rate-of-change of the momentum mu of the object is m dv/dt, where v
is its velocity, and the net external force acting on it is the negative —mg of its weight
mg. By Newton’s second law of motion, then, we have

Oh as (1.1a)
Equation ([Link]) is solved for the velocity v by integrating each of its sides. This
yields
v(t) = Vo — gt, (1.1b)

where Uo is the velocity of the falling object at time ¢ = 0. The altitude

r(t) = Ro + vot — dgt?

is then found by integrating both sides of Eq. (1.1b) since u(t) = r’(t). Notice that
the initial velocity vo enters the solution (1.1b) as a constant of integration.
 Linear friction 3

The determination of v and r above is a special case of the following purely

mathematical problem: Given
1) a continuous function b on an intervala < 1 < w ’

il) m constants cy,..., Cn, and

ili) a point fo in (a, w),
find all functions x satisfying

——=b(t), a<t<w, (1.1c)

and

X(to) = C1 dFr (lo) = Ca, «2, dFo (to) = en.

n—1

(1.14)
This problem, which we shall call an initial value problem, is a basic topic in the
calculus. It is solved by integrating each side of Eq. (1.1c) n times and using the
data (1.1d) to evaluate the constants of integration. = i |
Sey le) pagel a 5% ea
* Teny “
EXERCISE 1.1
egy —|\ Ee . = VoOAe ss
1. Solve the following initial value problems. x = Kam t+c,t oe a cine C2=O
a) SY = GF te sO) = 2, sOy= il. NOR how tot= ton t a \e STE
lo) 2 = 6t + 2,5 ca) x(1) = Dexa2, se (hl) = 1 Il. -) Aa bids
Oe 0) = 0 = titan t-+£ (thu) +¢,
d) x = 1/0 — P), Ie <1, x) =2 35%, XE tA 1 A(t)+3
e) x = —27/ + 272 + 2%), x(0) = 3, x’) = 0, x"(0) = 1.
2. An object is dropped from the top of a building 1024 ft high. Exactly 1 second later,
another object is thrown downward with an initial velocity vo. If air friction is neglected,
how large must vo be for the objects to reach the ground simultaneously?
3. Suppose an object is falling in vacuum with a constant acceleration —g. Express its
velocity as a function of its altitude.
4. Suppose an object is traveling along a straight line in space with a speed of 60 mph
(88 ft/sec). What acceleration would be necessary to bring the object to a stop within
1 second? If the object has a mass of 5 slugs, how many pounds of force would have to be
exerted on the object to produce this acceleration? Interpret your results for an automobile
collision.
5. If air friction, bearing friction, and transmission inefficiencies are negligible, how large
an engine (in horsepower) would a 1600-lb dragster have to have in order to travel a quarter-
mile in 10 seconds? [Hint: 1 hp = 550 ft - lb/sec.]

1.2 LINEAR FRICTION

Suppose that we slide a rectangular block having mass m ona large table top (Fig. 1.1),
giving it an initial velocity vg > 0. Friction causes the block to come to a stop. Let
us analyze the situation and try to determine the length of time required for stopping.
 1.2
4 Differential equations and the physical world

Figure 1.1

We lump together the effects of atmospheric friction and the friction between the
table top and the block. In order to make an appropriate mathematical model, first
observe that the frictional force is in the direction opposite to that of the velocity and
tends to increase in magnitude as the velocity does. As a first approximation, then,
it is not unreasonable to assume that the frictional force is proportional to the negative
of some power of the speed |u|, that is,
Frictional force = —)v|v|', (1.2a)

where € > O and \ > Oare constants. The constants \ and e€ vary from situation to
situation. They depend on the material and shape of the block and on the range of
values that v may have.
A simplifying assumption which is frequently made for the frictional force (1.2a)
is thate = 1. We then say that the frictional force is negatively proportional to the
velocity v and call the friction /inear. Making this assumption and applying Newton’s
second law of motion to our block, we find that

vo @) = av), (1.2b)
where a = —)/m.
Equation (1.2b) states that the time rate-of-change of the velocity of the block is
proportional to the velocity at every instant 7. Integrating both of its sides, we do
not obtain a solution as we did with Eq. ([Link]). Such an integration merely produces
the relation

U(t) = Uo + af v(s) ds,

0
from which physical information is difficult to extract. Equation (1.2b) is solved in
another way. Certainly v(0) = vo > 0. Thus v(t) > 0 on some longest interval
0<+t< 7”. It is therefore permissible to write

DiGe foro) (0° <or (1.2c)

Fem
Integrating both sides of Eq. (1.2c), we find that
t t
In Of 20a — fads = ar OS wake its
Vo v(s) 0
Thus v(t) = voe™% and rt = +0.
Our expression of v(t) as a decreasing exponential seems to imply that the block
will never stop since e*’ # 0 for any t > 0. The apparent paradox is resolved by
noting that the mathematical model which we have constructed for the physical
situation is just not applicable to it for very small values of v(?).
sage
 Linear friction 5

The determination of v(t) above is a special case of the following purely mathe-
matical problem: Given
i) a continuous function a on an interval a < t < ,
il) a point fy in (a, w), and
ili) a constant c,
find all functions x satisfying

x’ = a(t)x, Cat a), (1.2d)

and
X(to) = ¢. (iezc)
If c = 0, then certainly x = 0 is a solution to the problem. If c ¥ 0 and if a solution
x exists, then x(t) ~ 0 on some longest interval, 19 < t < 7*, say. Thus

20) (a- ed Pe ,ORS +

t t

In | Oa ip oteTs

Consequently,
t

x(t) = c exp || a(s) as; to SS t< +o. (iets)

to
The zero solution is incorporated into the representation (1.2f) by allowing c to have
the value zero. We shall subsequently refer to this solution method as separation of
variables. It is applicable to physical problems in which the rate-of-change of an
unknown quantity is proportional to the quantity itself. In studying such problems,
it is helpful to remember that every solution of Eq. (1.2d) is of the form x = ce™
if a is a constant.

EXERCISE 1.2
1. Find all solutions to the following equations.
a) x’ = x. lo) oe = Sak
c) x’ = 3t2x. (eee —wi ax, t= 0:
er (SCC 1) et <a) 2c i) of = (COLDER Ke < ae
Re (1 — 7), lf] <1. A) = Or 7) et > 0,
2. One way of estimating the number of bacteria in a culture is to assume that the rate-of-
increase of the population at any time is proportional to the size of the population at that
time.
a) Suppose that a certain type of bacteria reproduces at a rate equal to one-third the
number of bacteria present. If a culture were started from a single individual, what
would be the size of the population 24 hours later?
b) The population of the United States was 5.31 million in 1800, 76.0 million in 1900,
and 179 million in 1960. If the population had increased at a rate proportional to the
size of the population, what should it have been in 1960? How do you account for
the discrepancy?
 Differential equations and the physical world 1.3
6

3. When a bank or savings and loan association advertises that it compounds interest
continuously, it means that interest is paid at the advertised rate in such a way that the time
rate-of-increase of a deposit is proportional to the amount on deposit. The proportionality
constant is the interest rate. Suppose that $1000 is kept on deposit for ten years at an interest
rate of 6%. How much greater will the yield on the investment be if interest is compounded
continuously rather than quarterly?
4. The real estate value of Manhattan was estimated at $13.4 billion in 1967. Suppose that
Peter Minuet had deposited his $39 in 1626 in a bank which paid 6% interest. Would he
have been able to buy Manhattan in 1967?
5. Anoctopus is kept in a 10,000-gal tank in an aquarium. The water in the tank is changed
continuously at a rate of 2 gal/min. If the octopus discharges ink into the water and the ink
mixes instantly, how long does it take for 90% of the ink to be removed from the tank?
6. Even a ventilation fan may not remove enough carbon monoxide from a closed garage
to make working on a running automobile safe. Suppose that an automobile engine produces
10 ft?/min of exhaust gases and that the exhaust gases contain 1 part carbon monoxide per
1000 parts of exhaust. How many cubic feet of air per minute must an exhaust fan move
through a closed garage containing the engine if the carbon monoxide concentration is to
be kept below 1 part per 1 million parts of air?
7. Instudying radioactive decay, it is frequently assumed that the rate at which a radioactive
substance decays is proportional to the amount of the substance present. Suppose the
amount of Carbon 14 in a prehistoric bone is now one-millionth of the amount that was
present when the creature from which it came was alive. Use the fact that Carbon 14 has a
half-life of 5440 years to date the bone.

1.3 AN OBJECT FALLING IN AN ATMOSPHERE

Notice that Eq. (1.1b) implies that a falling object can be made to strike the earth
with as high a speed as desired by simply dropping it from a sufficiently high altitude.
This prediction does not, of course, hold true for the actual situation. Objects falling
in an atmosphere under constant acceleration have terminal velocities due to atmos-
pheric friction.
Let us restudy the fall of the object discussed in Section 1.1 and attempt to take
frictional effects into account. Now we shall postulate (as we did in Section 1.2) that
the frictional force opposes the motion of the object and is proportional to its velocity.
Equation ([Link]) modified accordingly becomes

mvu'(t) = —mg + dv(2), (1.3a)

where } is a positive constant. We set a = \/m and write Eq. (1.3a) in the form
v'(t) — av(t) = —g. (1.3b)
Equations (1.1a) and (1.2b) are special cases of Eq. (1.3b), for which separation of
variables is an immediately applicable solution technique. A different technique,
which we shall call the method of integrating factors, can also be used to solve this
equation. The first step in the method is the integration of the coefficient —a of v(t).
 An object falling in an atmosphere 7

This gives —at. Then the function e~, which is called an integrating factor, is formed.
When Eq. (1.3b) is multiplied by the integrating factor, the left side becomes an exact
derivative; in fact,
d a7 oe —ax = See:
di COcmayi Oeste (1.3c)

The function v is found by integrating both sides of Eq. (1.3c) to obtain

v(t) = v(O)e + 2(ieee). (1.3d)

This equation predicts that no object, dropped with an initial velocity v(0) = 0, can
have a downward velocity exceeding g/a = —mg/.
The method of integrating factors will be used in Section 3.1 to solve the following
purely mathematical problem: Given
i) a continuous function 6 on the interval -~ < t < +m, and
ii) a constant a,
find all functions x such that
x = ax +7b@): (ise)
An integrating factor for this equation is e at
°’, and

£ (x(n) Svea K({)\v— dex x(t) =e. (1).

Thus

x(t) = e"'x(0) + ef e ““b(s) ds

0

represents every solution to the problem. Different solutions are obtained by assigning
different values to x(Q).

EXERCISE 1.3
1. Use the method of integrating factors to solve the following differential equations.
2D) oe) == Dee SE Bh b) xe = 2 e2"
©) x! = 2x + 3e~ 2? (al) Set OBS Hi,
2. Suppose an object falls from the top of a building 1225 ft tall. According to the physical
model of Section 1.1, what would be its speed when it reached the ground? How long would
the object take to reach the ground?
3. Assume that the object of Problem 2 has a terminal velocity of 280 ft/sec with air
friction taken into account. How long does it take for the object to attain 99% of its terminal
velocity? Compare your result with the time of fall and impact velocity in Problem 2.
4. A ball bearing is released from rest in a reservoir of oil. The deceleration due to fluid
friction is 1000 times the instantaneous speed (in meters per second). What is the terminal
velocity of the bearing? What time is required for the bearing to attain 99% of its terminal
velocity?
gO,
280% = SKN moh
oo a (N\
 1.4
8 Differential equations and the physical world

5. The motor ona boat can exert a maximum thrust of 44 Ib. Suppose the force of fluid
friction at any instant equals (in pounds) twice the speed of the boat at that instant. What is
the maximum cruising speed of the boat? If the boat weighs 160 lb, what time is required
to attain 99% of top speed? If the boat weighs 1600 Ib, what time is required to attain 99%
of top speed? Notice that the weight of the boat has no effect on its maximum speed;
however, the weight does affect the time required to attain that speed.
6. Suppose that two spheres having identical air friction characteristics are dropped from
a very great altitude. If one sphere is ten times as heavy as the other, which will strike the
ground first? Neglect wind, buoyancy, and the rotation of the earth, but take air friction
into account.
7. Suppose the rate of decrease in the temperature of water is proportional to the environ-
mental temperature less the temperature of the water. Two cups of water, one at 70°F and
the other at 210°F, are placed in a refrigerator (14°F). Which will freeze first?
8. Aman has $10,000. He wishes to spend it in equal monthly sums over his entire lifetime
so that none of it will remain for his heirs when he dies. He decides to deposit his money in
a bank which pays 6% annual interest (compounded continuously) and withdraw a fixed
amount each month. If his life expectancy is 400 months, what will his monthly income from
the investment be? How much more lucrative is this scheme than keeping the money in a
mattress and spending equal amounts monthly ?

1.4 THE ESCAPE OF A ROCKET

A rocket can be fired into space in such a way that it neither falls back to earth nor
goes into orbit, i.e., in such a way that it escapes. Just what is it that will cause a
rocket to escape? It seems plausible that the crucial factors are its speed and altitude
at the instant when the fuel is completely consumed (burn-out). The considerations
of Section 1.1 allow us to examine the escape problem to some extent.
We assume that the rocket is fired vertically from the earth and neglect the rotation
of the earth. For times ¢ > 0, we denote the rocket’s altitude (measured from the
center of the earth) by 7(f), its mass by m(t), and its exhaust velocity by —u(f).
If a mass —Am of burned fuel is discharged during the time interval ¢ to ¢ + At,
the net change Ap in the momentum p of the system consisting of the rocket frame,
the unburned fuel, and the discharged mass is

Ap = (v + Av)(m + Am) — (v — u) Am — mov.

Here, v is the speed of the rocket at time ¢, and v + Av is the speed at time ¢ + Art.
The average time rate-of-change of p is thus

Av A Am
= Thala St AD
At

and the instantaneous rate-of-change is

de agp Wetagee (1.4a)

 The escape of a rocket 9

We want to use Eq. (1.4a) to see what information Newton’s laws give about the
flight of the rocket. There will, of course, be a frictional force F(t) on the rocket as
long as it remains within the atmosphere. The only other force acting on the rocket
will be its weight, m(t) Gm,/r?(t). Notice that it is now incorrect to assume the weight
of the rocket to be 32m(t). Since rockets rise to great heights, the variation in weight
due to altitude must be considered. Newton’s second law of motion yields the
equation
Gm.m(t)
m(t)r’"(t) + u(t)m'(t) = + F(), (1.4b)
aC)
where 7’ = v and primes denote differentiation with respect to ¢.
The motion of the rocket prior to burn-out is difficult to analyze; if burn-out
occurs while the rocket is still within the atmosphere, then the situation is still com-
plicated. Let us suppose, however, that the fuel lasts until a time fy > 0 at which the
rocket has left the atmosphere behind. Then u(t) = 0 for t > to and F(t) = 0 for
t > to. The subsequent motion of the rocket is then described by the differential
equation
rif) = — 22. (1.4c)
We denote the velocity and altitude at burn-out by vy and ro respectively. How should
Vo and ro be related in order that the rocket escape? It is rather difficult to solve
Eq. (1.4c) for r(t) in terms of ¢t. Our question about vg and 79 can be answered, how-
ever, without an explicit solution of the differential equation.
We multiply Eq. (1.4c) by 7’(t) and obtain
d 2 d 2Gm,
— (r' = —|(——}]. 4
dt ('O) Al r(t) ) oe)
Then integration from fo to ¢ yields the relation

2Gm, 2 2Gm, |''”

(1.4e)
AS he ee Sar | :
Now suppose that the rocket escapes. Then r’(t) > 0 for all ¢ > to and r(t) > +
ast— +o. Letting t > +o in Eq. (1.4e), we find that

vo > V2Gm./ro. (1.4f)

Thus the satisfaction of inequality (1.4f) by vo and ro is a necessary condition for the
escape of the rocket.
Since a 1-kg mass has a weight of about 9.8 newtons near the earth’s surface and
since the earth is nearly spherical, with an approximate radius of R = 6.38 X 10° m,
we can use Newton’s law of gravitation
: Gm
weight = mass 5
€

to evaluate the constant Gm,, finding that Gm, = 3.86 X 10'* newton - m?/kg.
 1.5
10 Differential equations and the physical world

The worst possible choice for an escape altitude ro is the radius R of the earth
if Vo is to be as small as possible. Thus a lower bound for the speed required for a
“real” rocket to escape is ~/2Gm,/R = 11,200 m/sec. For escape altitudes ro > R,
the speed vo required for escape is, of course, less.

EXERCISE 1.4
1. If burn-out for a rocket occurs at an altitude ro = 100 km, what is its minimum escape
speed vo?
2. Is the satisfaction of the inequality (1.4f) sufficient for the escape of a rocket?
3. Suppose the burn-out altitude ro and velocity vo for a rocket are such that
2

Lore ey 0
2: ro
What is the maximum altitude to which the rocket can rise?
4. Explain why

ROOD ree area

du

| ene a

is a solution of Eq. (1.4c) for t as a function of r.

5. Consider a rocket in space for which gravitational effects are negligible, and suppose
that its exhaust velocity is constant, say u(t) = k. Express the velocity of the rocket in terms
of its mass. Qualitatively, how should one design a rocket so as to obtain the greatest speed
v(t) for a given initial mass mo = m(0)?

1.5 AN ORBITING SATELLITE

Let us now suppose that a satellite of mass m has been fired into space in such a way
that it goes into orbit. What can be said about the nature of the orbit? In the first
place, we may as well assume that the gravitational force between the earth and the
satellite does not cause the earth to move. Also, it is known from physics that the
satellite will move along a plane curve rather than a space curve. Consequently its
motion can be studied using two-dimensional coordinates with the earth located at
the origin (Fig. 1.2).

-atellite
(x, y)

Earth

Figure 1.2
 An orbiting satellite 11

The coordinate r = \/x? + y? is the altitude of the satellite (measured from

the center of the earth). Thus our analysis is similar to the analysis of the escaping
rocket in the previous section; rotation must now be taken into account, however.
The net force acting on the satellite has, as before, magnitude mGm,/r’, and the
force is directed toward the origin.
Let the cartesian coordinates of the satellite be (x, y). These coordinates have no
really interesting physical interpretation in this problem; measurements which can
actually be made on the satellite are made in terms of polar coordinates. The cartesian
coordinates are mathematically useful, however, because Newton’s second law of
motion is easier to apply in cartesian coordinates than it is in polar coordinates. We
assume that the orientation of the axes is such that

xX = To, y = 0, x’ = 0, y’ = v9

at the instant t = 0 when we begin to observe the orbiting satellite. Newton’s second
law of motion, applied in the directions of the x- and y-axes separately, yields the
equations

mx" = —m ue cos 6, my’ = —m s sin 6, (1.5a)

where
XT COSi0 and ve musi: (1.5b)

ME at
Equations (1.5a) and (1.5b) imply that

(59
and

ce ze (1.5d)
The term m,.G/r? in Eq. (1.5c) represents the downward acceleration of the
satellite due to gravity, and the term r6’? is called the centrifugal acceleration of the
satellite. The interplay of the forces corresponding to these two accelerations makes
the orbital motion possible: gravitational force tends to cause the satellite to fall and
centrifugal force tends to cause the satellite to rise. The expression —2r’6’/r in
Eq. (1.5d) is called the Coriolis acceleration of the satellite.
Equations (1.5c) and (1.5d) are the classical differential equations which are used
to describe orbital motion. The exercises below are concerned with extracting from
them physical information about the orbit of the satellite.

EXERCISE 1.5

1. Multiply Eq. (1.5d) by r? and integrate to find a relation between r and 6’. Use this
relation to eliminate 6’ from Eq. (1.5c). Then multiply Eq. (1.5c) by 2r’ and integrate to
12 Differential equations and the physical world 1.6

find a relation between r and r’. Finally use the fact that dr/dt = (dr/d@)(d6/dt) to show
that
dr re r ie 15
srl“) +a-a(Z)-1 , (1.5e)

where a = 1 — 2Gm,/rovd.
2. Write Eq. (1.5e) in the form
rT (7)

+ /r0 i
Uu
du
Uu
=|a (1.5f)
ulai{—) +0A-—-a-—-1
ro ro

Then evaluate the integral on the left by making the substitution u = 1/w and express r
in terms of 6. Note that the sign of a is important.
3. Show that the orbit of the satellite is an ellipse, parabola, or hyperbola when vo is
respectively less than, equal to, or greater than \/2Gm,/ro. The satellite will, of course,
escape from the earth if its orbit is a parabola or hyperbola. Explain why your results are
consistent with the discussion of the escaping rocket in Section 1.4. [Hint: Look up the polar
equations for the conic sections in a calculus book.]

1.6 A LINEAR OSCILLATOR WITH ONE DEGREE OF FREEDOM

Let us place a rectangular block having mass m on a tabletop and attach it to a wall
by a coil spring (Fig. 1.3a). We then stretch the spring xo units and release the block
at time ¢ = O with an initial velocity v). Where will the block be at each future time
tes

NVA

(a)
Figure 1.3

Let x denote the displacement of the block from its equilibrium position, that is,
from the position which it occupied when the spring was unstressed. The time rate-
of-change of momentum for the block is mx’’. We assume, as we did in Section 12:
that the frictional force on the block is of the form —)x’, where ) is a positive constant
and x’ is the velocity of the block.
We assume further that the force which the spring exerts on the block is of the
form —kx, where k is a positive constant. A spring that behaves in this way is called
a linear spring, and k is called its stiffness coefficient.
 A linear oscillator with one degree of freedom 13

Newton’s second law of motion requires that the total force —\x’ — kx on the
block equals mx’’.. Thus
x” + ax’ + wx = 0, (1.6a)
where a = \/mandw = k/m. Equation (1.6a) is called the damped linear oscillator
equation if a > 0.
To locate the block at each time ¢ during the motion, one must determine the
solution x that satisfies x = x9 and x’ = Vg when t = 0. This is called an initial
value problem, and methods for solving it are derived in Chapters 3 and 4. For the
time being, we shall merely state a solution procedure. Let B = 4w? — a?. Then

x = c,exp[—3(@ + V—£)] + coexp[—3(a-V—A)] if B<0,

x = (cy + Cot) exp [—4az] if brn; (1.6b)

x = exp[—al]- («,sin YP + Co cos YB‘ if Se 0:

where c, and cy» are constants which are uniquely determined by xo and vo.
If a = Oin Eq. (1.6a), then it has the form

x + w*x = 0, (1.6c)

and it is called the undamped linear oscillator equation. Each solution has the form

xX = Cc, Sin wt + C2 COs wt (1.6d)

for some choice of constants c; and cy. Conversely, the relation (1.6d) defines a
solution of Eq. (1.6c) for every choice of the constants c; and C2.

EXERCISE 1.6
1. Find the displacement x of the block in Fig. 1.3 if x(0) = 1, x’(0) = 0, and friction is
negligible.
2. Find the displacement of the block in Fig. 1.3 if x(0) = 0, x’(0) = 1, and friction is
negligible. How does one physically start the block into motion so that these initial conditions
are met?
3. Show that x = (e’#! + e—‘#)/2, where i? = —1, satisfies Eq. (1.6c) and the conditions
x(0) = 1, x’/(0) = 0. Deduce, on physical grounds, that cos wt = (e’*! + e*#‘)/2 and
sin wt = (e*”! + e+) /2i, assuming that Eq. (1.6c) and the initial conditions truly determine
the motion of the block. [Hint: Use the results of Problems 1 and 2.]

4. Assuming that w2 > a@2/4, find the solution of Eq. (1.6a) that satisfies x(0) = 1,
x'(0) = 0. How does one physically start the block into motion in such a way that these
initial conditions are met?
5. How much time is required for the block in Problem 4 to undergo one complete
oscillation?
 ils7/
14 Differential equations and the physical world

1.7 LINEAR OSCILLATORS WITH SEVERAL DEGREES OF FREEDOM

The principles explained in the last section can be used to analyze the motion of a
system consisting of several masses and springs. Such a system is said to have n
degrees of freedom if n coordinates are required for its description. For example, the
system in Figs. 1.4(a) and 1.4(b) has two degrees of freedom. In analyzing it, we
assume that the springs are linear with stiffness coefficients k,; and kz and that friction
is negligible. The coordinates w and y measure the displacements of the blocks from
the positions that they would occupy if the springs were unstressed.

Figure 1.4

Newton’s second law of motion is easiest to apply if we imagine that both blocks
are to the right of equilibrium and that y > w. The right-hand spring is then stretched
a distance y — w and it exerts forces +k2(y — w) on the blocks. The spring on the
left exerts a force —k,w on the first block. Thus
mw"
—kyw + ko(y — w) (1.7a)
and
may” = —k2(y — w). (1.7b)
Equations (1.7a) and (1.7b) are said to form a system of linear differential equations
for the unknown displacement functions y and w. In studying such a system theo-
retically, it is helpful to eliminate the second derivatives. This can be done by intro-
ducing the velocities x = w’ and z = y’ of the blocks as additional variables. The
equations can then be put in the form
a x ;
! =< ky aia ko ko
x = air a Ww =. is 3

aoe oe (1.7c)

baie La ed
Me Mo
This is the form in which linear differential systems will be studied in Chapters 3 and 4.

EXERCISE 1.7
1. If the spring with stiffness coefficient k; were removed from the system in Fig. 1.4(a),
how would the equations of motion (1.7a) and (1.7b) for the system be altered?
 The series RLC circuit 15

2. Let the block in Fig. 1.4(a) with mass mz be connected to a wall on the right by a third
spring with stiffness coefficient k3 in such a way that the third spring is unstressed when the
blocks are in equilibrium. How are the equations of motion for the system altered?
3. A large plate (mass M), resting on a frictionless tabletop, is connected to a wall by a
linear spring with stiffness coefficient k. A small rectangular block (mass m) rests on top of
the plate at its center. Assuming that friction between the block and plate will be linear if
the block moves on the plate, introduce coordinates and derive equations of motion for the
system.

1.8 THE SERIES RLC CIRCUIT

Figure 1.5 is a schematic diagram of an electrical circuit which is called a series RLC
circuit. The symbol labeled E denotes a battery which produces a voltage of E volts,
and the symbol labeled S denotes a switch, shown in the “‘off” position. The straight
lines in the diagram represent wires, and the dots (a, b, c, d, e) represent electrical
connections called nodes.

Figure 1.5

The symbols R and / in the diagram can be explained for our purposes by drawing
an analogy with hydraulics. The wires are like pipes that are filled with fluid. The
switch S corresponds to a valve, and the battery is analogous to a pump which main-
tains a pressure E. When the switch S is in the “on” position, the voltage E which
the battery produces causes the electrons in the circuit components to move in a
counterclockwise direction around the circuit, just as molecules of fluid under
pressure would move in a pipe.
Each electron carries a negative charge and, when 6 X 10!® electrons have
passed a point in the circuit, one says that 1 coulomb of negative electrical charge has
passed the point. If 1 coulomb of negative charge per second is passing a point in
the circuit, then it is conventional to say that a current of | positive ampere is passing
through the point in the opposite direction. Thus the electrical current J corresponds
hydraulically to the number of gallons of fluid per second passing a point on a pipe.
When analyzing a circuit, one does not, in practice, worry about the direction in
which the electrons are physically moving. One merely designates an arbitrary
direction for the current J. The algebraic sign of I(t) indicates the direction in which
the electrons are actually moving. When /(t) < 0, the electrons are moving in the
direction of J; when J(t) > 0, the electrons are moving in the direction opposite to J.
 1.8
16 Differential equations and the physical world

The symbol labeled R represents a resistor. A resistor in an electrical circuit acts

in the same way as does a crimp or partial blockage in a pipe. Just as a crimp impedes
the flow of fluid and causes a drop in pressure, a resistor impedes the flow of electrical
current and causes a drop in voltage. The unit of resistance is the ohm. If a current
of value /(t) amperes is passing through a resistor having a resistance of R ohms, the
voltage drop across the resistor equals R/(?).
The symbol labeled L represents an inductor, a device manufactured by winding
wire around a suitable form or core. (An electromagnet, for example, is an inductor.)
When an electrical current begins to flow through an inductor, a magnetic field is
formed. When the current J flowing through an inductor is increasing, some of its
energy is diverted into the production of the magnetic field. The current’s loss of
energy then causes a voltage drop across the inductor. If, conversely, the current J
flowing through an inductor is decreasing, the electromagnetic field weakens and the
energy released tends to produce a voltage rise across the inductor. The unit of
inductance is the henry. If the value of the current flowing through an inductor of
inductance L henrys is /(t) amperes, then the voltage drop (or rise) across the inductor
equals L dI(t)/dt.
The symbol labeled C in Fig. 1.5 represents a capacitor. To make a capacitor, one
fastens wires to two metal plates and brings the plates very close together but does not
allow them to touch. Electrical charge does not pass through a capacitor; its function
is to store electrical charge. Nevertheless, one may think of a time-varying current I
as passing through it. For illustration, consider the capacitor shown in Fig. 1.6. If
electrons are moving toward the left, atoms of metal in the left-hand plate of the
capacitor lose electrons and become positive ions. The right-hand plate acquires
excess free electrons and becomes negatively charged. Now imagine that the capacitor
is enclosed in a box, say, with one wire entering and another leaving it. An observer
merely detects a current J apparently flowing through the box.
Box
5 a eee ee =
: wales C Peres oes

a | ap NS | b

{sd
A aaJ
Figure 1.6

The voltage drop V(t) across a capacitor at any time f¢ is given in terms of the
amount of charge Q(t) stored at that time by the relation V(t) = Q(t)/C, where C
is a constant called the capacitance of the capacitor. (The unit of capacitance is
called the farad.) The voltage drop V(t) is expressed in terms of the current J flowing
through the capacitor by the relation
t

VQ) =o iw: i I(u) du,

0

where Vy is the voltage drop across the capacitor at time ¢ = 0.

 The series RLC circuit 17

Having explained the symbols in the diagram, we now derive a differential

equation which the current J in the series RLC circuit must satisfy. To do this, we
employ Kirchhoff’s voltage law.

Kirchhoff’s Voltage Law. The sum of all the voltage drops across the components in
a Series circuit is zero.

Now assume that the switch S in Fig. 1.4 has been in the “on” position for a very
long time. Then the capacitor is fully charged and there is no current flowing in the
circuit. Consequently, there is no voltage drop across the inductor or resistor, and the
initial voltage Vo across the capacitor equals —E. At time t = 0, the switch S is
thrown into the “off” position, and current begins to flow around the circuit as the
capacitor discharges. Thus /(0) = 0 and, by Kirchhoff’s voltage law,
t

Li'(@t) + RIG) +c ibI(u) du + Vo = 0. (1.8a)

Letting t — 0+, we find that /’(0) = E/L, since Vp) = —E. Finally, differentiation
of (1.8a) yields the differential equation

Wrae” 1ae
R!
lt=
ah
(1.8b)

for the series RLC circuit. The current in the circuit is described by that solution J
which satisfies the initial conditions

JOy=0, FO)=E/L.
Notice that Eq. (1.8b) is mathematically identical with Eq. (1.6a) for the damped
linear oscillator. Inductance is analogous to mass; resistance is analogous to a co-
efficient of friction; the reciprocal of capacitance is analogous to a stiffness coefficient;
and current is analogous to displacement.

EXERCISE 1.8

1. Ifa battery which produces a voltage E is switched into a series circuit containing a
resistance R and an inductance L at time ¢ = 0, the current J satisfies the initial condition
1(0) = 0. Compute J(2).
2. If the inductance L in Problem 2 is replaced by a capacitance C, the initial condition for
the current is (0) = E/R. Find J(0).
3. Suppose R, L, C, and E in Fig. 1.5 have the values R = 104 ohms, Z = 1 henry, C =
2 « 10—8 farad, and E = 100 volts. Use Eq. (1.6b) to find /().
4. Consider the series RLC circuit of Fig. 1.5 with R = 0. Show that the current oscillates
indefinitely when the battery is switched out of the circuit. What is the frequency of the
oscillations?
 1.9
18 Differential equations and the physical world

1.9 THE PARALLEL RLC CIRCUIT

Figure 1.7 is a schematic diagram of a circuit which we shall call a parallel RLC
circuit. AS was the case with the circuit in Section 1.8, we assume that the switch has
been left in the ‘‘on” position. Then the capacitor is fully charged, and there are no
currents flowing in the circuit. At time ¢ = 0, the switch is turned off, and the currents
x, y, and z begin to flow.

Figure 1.7

To analyze the circuit, we shall need to use not only Kirchhoff’s voltage law but
also Kirchhoff’s current law.
Kirchhoff’s Current Law. The sum of the currents entering any node equals the
sum of the currents leaving that node.
Kirchhoff’s current law (applied at node d, for example) tells us that x + y = z.
We regard the circuit around nodes a, d, e, fas a series RL circuit. By Kirchhoff’s
voltage law,
Lx’ — Ry = 0: (1.9a)
The initial condition for x is x(O) = 0.
We similarly regard the circuit around nodes a, b, c, d as a series RC circuit and
find that t
aaah z(u) du — E+ Ry(t) = 0. (1.9b)

The initial condition y(0) = E/R for y is found by taking the right-hand limit
t— 0+. Differentiation of the equation gives
C'z+ Ry’ = 0. (1.9c)
Using the fact that z = x + y, we then write Eq. (1.9a), Eq. (1.9c), and the initial
conditions in the forms fe
iS LT» (1.9d)

D Sing apa: (1.9e)

x0) =0, y(0)= = (1.9f)

 An electronic oscillator 19

To determine the values x(t), y(t), and z(t) of the currents at times ¢ > 0, we must
therefore solve a system of differential equations subject to initial conditions, that is,
we must solve an initial value problem.

EXERCISE 1.9
(The first three problems refer to the circuit in Fig. 1.7.)
12 eshow that x ©) = E/E:
2. Differentiate Eq. (1.9d), eliminate y from Eq. (1.9e), and obtain a differential equation
for x.
3. Use Egs. (1.6b) and the results of Problems 1 and 2 to find x(a), y(0, and z(#).
4. The capacitor in Fig. 1.8 is charged so that the voltage across it is E. The resistors
have a common resistance R, and the switch S is closed at time ¢ = 0. Formulate an initial
value problem for the currents x and y.

Figure 1.8

1.10 AN ELECTRONIC OSCILLATOR

Engineers have long known how to construct vacuum tube circuits which will produce
periodically varying voltages even though the voltages supplied to the circuits are
constant-valued. Such circuits are called electronic oscillators, and one of them is
diagrammed schematically in Fig. 1.9. It was van der Pol who first gave a satisfactory
explanation of the way an electronic oscillator works. He did this by deriving a
differential equation for the voltage V, between nodes a and d in the circuit shown.

Figure 1.9
20 Differential equations and the physical world 1.10

This differential equation, which we will rederive below, has a periodic solution
toward which every other solution tends with increasing time. Thus the voltage V,
is periodic for all practical purposes.
The indicated vacuum tube, which is called a triode, has three internal elements
called the plate, the grid, and the cathode. These are connected into the circuit at
nodes p, g, and c respectively. The voltage V, is called the grid voltage, and the
voltage from p to c, which we shall denote by Vj, is called the plate voltage.
In the triode, electrons move from the cathode to the plate, creating a positive
plate current I, which flows from plate to cathode. The grid is a wire screen through
which electrons leaving the cathode must pass if they are to reach the plate. A few
electrons will actually strike the grid rather than pass through its openings. These
collisions are detected as a current flow within the tube from the grid to the cathode,
but the magnitude of this current is so small in the context of the present discussion
that we shall ignore it. If the voltage V, is negative, the electrons, which are negatively
charged, will be repelled as they approach the grid. The flow of current J, will thus
be impeded. If, on the other hand, V, is positive, the grid will attract approaching
electrons, accelerating them so much in the process that they pass through the open-
ings in the grid and continue onward toward the plate. In this case, the flow of plate
current J, is aided.
The two inductors in the diagram are wound on the same core. Consequently
they have a common magnetic field, and a change in current through one coil produces
a voltage across the other. The extent to which this interaction occurs is described
quantitatively by a constant M, called the mutual inductance of the coils. The voltage
from p to e equals MI’ — L,I;, where primes denote differentiation with respect
to time f. The voltage from d to a equals LI’ — MI}. We shall assume, as is the
case in an actual circuit, that the inductances M and L, are so small that MI’ — L,J;,
has a negligible value. Since Kirchhoff’s voltage law, applied around the nodes p, c
d, e, p, yields the relation
Vez b= 10 Myre
we may therefore take V, = E.
For each type of triode, there is a function y = ¢(x) and a constant u > O such
that J, = ¢(V, + pV,). The function ¢ is called the characteristic of the tube, and
wis called its amplification factor. In our case, V has the constant value E. Thus
I, = W(V,), where ¥(V,) = (E+ wV,). The function y is called the transfer
characteristic (of the tube) at voltage E. If one knows the type number of the tube,
he can look up its amplification factor and graphs of its transfer characteristics in
tube manufacturers’ handbooks.
For the oscillator being studied, a transfer characteristic of the type graphed in
Fig. 1.10 is required. The horizontal asymptote indicates the maximum plate current
I, possible for the tube, and the inflection point occurs at V, = 0. We then approxi-
mate ¥(V,) by a cubic polynomial ¥(0) + aV, — 8V}/3, where a and 8 are positive
constants.
 An electronic oscillator 21

Sel

Figure 1.10

The sum of the voltage drops around the nodes a, b, d, a satisfies

Li'(t) + Co /I(t) dt + RI(t) — MI{(t) = 0. (1.10a)

Since V,(t) = C~' f I(t) dt and I(t) = (a — BV7)V;, Eq. (1.10a) can be put into
the form

ey g eM a RG) ny?
Mae RE g VO
g Lt lee at (1.10b)
.

This is the differential equation which describes the operation of the oscillator circuit.
Its appearance can be simplified dramatically by the substitutions w = LC, € =
(Ma — RC)w, A? = €/wMB, and V,(t) = Ax(wt). The simplified form of Eq. (1.10b)
is then
x” + ex? — 1x’ + x = 0. (1.10c)
This equation is called van der Pol’s equation, and we shail see in Chapter 8 that, if
€ > 0, it has a periodic solution toward which every other solution tends as f
increases.

EXERCISE 1.10
(All the problems below refer to the circuit in Fig. 1.9.)
1. In order that the circuit oscillate, it is necessary that the graph of the transfer character-
istic have a sufficiently steep slope at V, = 0. How steep should the slope be?
2. Given that the voltage V, is periodic, explain why the current J is periodic.
3. To use the oscillator as a source of periodic voltage for some external device, it is neces-
sary to couple it to the device. How could one do this without changing the differential
equation (1.10b) for the circuit? [Hint: Make use of the mutual inductance phenomenon. ]
4. Make the formal change of variables x = s, x’ = 1/y in van der Pol’s equation and
reduce it to a differential equation involving only the first derivative dy/ds.
CHAPTER: 2

SOLUTION METHODS FOR SPECIAL FIRST AND

SECOND ORDER NONLINEAR EQUATIONS

2.1 PRELIMINARIES

A differential equation of the form

Mee f (PO ee eae) (2.1a)
is called an nth order, normal, ordinary differential equation (ODE). It is called nth
order because its derivative of highest order is an nth derivative. It is called normal
because x” is given explicitly in terms of the other symbols in the equation. It is
called ordinary because no partial derivatives occur in the equation. Most of the
differential equations derived in Chapter 1 are of this type.

Example 1. An equation for the distance x traveled by a body falling from rest under
a constant acceleration g is (x’)” = 2gx. This is a first order, ordinary differential
equation. It is not normal. The same physical phenomenon is also described by the
equation
Xl =N/ 26x. (2.1b)

This equation is normal. If we interpret it as a special case of Eq. ([Link]), then

f(t, x) = V2gx. ||
The differential equation (2.1a) is called linear if f has the special form

IG x, Bee O od) ae) cs —ay(t)x"—? ee An—1(1)x’ a An(t)x =r b(t),

where a1, ..., Gn, 6 are functions defined on an interval a < t < w. In this case, it
is customary to write the equation in the form

x™ 4 ay(t)x—P + +++ + an_ (tx! + an(t)x = b0). (2.1c)

An equation that is not linear is called nonlinear.

22
 Preliminaries 23

Example 2. The equations in Chapter 1 which have the form (2.1c) are Eqs. (1.1a),
(1.1c), (1.2b), (1.2d), (1.3b), (1.3e), (1.6a), (1.6c), and (1.8b). Equations (1.4c) and
(1.10c) are nonlinear. ||
Example 3. Let us consider a viscous fluid which flows parallel to the x-axis in the
xy-plane with a constant speed U,,. A flat plate is immersed in the fluid in such a way
that its edge coincides with the positive x-axis (Fig. 2.1). The plate creates a wake,
which alters the velocity of the fluid at each point p in the plane. If p has coordinates
(x, y), where x and y are positive, let u(x, y) and v(x, y) denote, respectively, the
horizontal and vertical components of the fluid velocity at p.

Fluid flow

eee ss
|
Figure 2.1
Flat plate

It has been found by experiment that the fluid has velocity zero at the surface of
the plate, that is,
tix, 0) — tx 0) for 59 Pl Oe (2.1d)

The wake is, of course, not very pronounced at large distances y from the plate. This
physical observation is expressed mathematically by the relations
ime x,y) =U, lim. v(x, y= 0 (2a1€)
y>+o Uae
for each fixed x > 0.
In one mathematical model for viscous fluid flow, it is found that wu and v satisfy
the partial differential equations

ge OE ae = 0. (2.1f)

The constant v is called the (kinematic) viscosity of the fluid. To find the velocity of
the fluid at points with positive coordinates (x, y), one tries to solve the differential
equations (2.1f) subject to the boundary conditions (2.1d) and ([Link]). Such a problem
is called a boundary value problem.
Blasius studied the problem (2.1d, e, f) by seeking a solution of the form

u(x, y) = Usr'(t), t= y(2vx/Un)~""”. (2.1g)

Substitution of these relations into Eqs. (2.1d, e, f) produces the boundary value
problem
fea res =. (2.1h)

r(O) = 0, r'(O) = 0, r(t)— 1 as tro+on. (2.11)

 pape
24 Solution methods for special first and second order nonlinear equations

We shall show that this boundary value problem has a solution in Section 8.9.
The values of this solution can be used in conjunction with the formulas (2.1g) to find
the velocity of the flowing fluid. Here, however, we merely comment that this is
another example of a nonlinear equation which occurs in an attempt to study the
world around us. ||

EXERCISE 2.1
1. Which of the following equations are linear?
2) oe! = Phe ls) Se = fae.
(ep ante amon dix. = fa ox: |
e) x” + w*x = 0. f) x” + w2x + x? = 0.
g) x” + di — cos fx = 0. h) x’ + — cos/)sinx = 0.

2. Which of the following equations are normal?

a) rn’ +. rr’ =0- b) sin'@’) -- x = ¢.
c) x’ + sin-! @ — 4) = 0. d) V1 — (x)? = tx.
ex” =x+tvi1 — (x’)2. f) x sinx’ + xc? = 0.
3. An object falls 64 ft to the earth’s surface. Find a reasonable value which its impact
velocity cannot exceed.
4. Show that Blasius’ equation (2.1h) has solutions r(t) = mt + 6 which are linear func-
tions even though the equation itself is nonlinear. Find a solution r of the equation which
satisfies the initial conditions r(2) = 1, r’(2) = 4, and r’’(2) = 0.
5. Substitute the relations (2.1g) into Eqs. (2.1d, e, f) and derive the boundary value problem
(2.1h, i). [Hint: Use the chain rule.]

2.2 INITIAL VALUE PROBLEMS

Up until now, we have used the expressions “solution” and “‘initial value problem”
rather informally. In this section, we shall give them precise meanings.
By a solution of Eq. (2.1a), we shall mean a function x = ¢(t) such that

Oo (DG oO. Os. le 8)

at every point of an open interval T,~ < ¢ < tg. One poses an initial value problem
when he gives a time ¢y together with n constants c,,..., cy, and asks for a solution
x = ¢(t) of the equation which satisfies the initial conditions

oli) = C1, '(to) = Co, «.- 3 (Ege ten

Chapter | contains several examples of initial value problems. In many cases, it
is necessary to know the initial state of a physical system as well as its differential
equations if one is to predict its future behavior. This is not always the case. For
example, the initial values of the currents in a well designed electronic oscillator are
completely irrelevant to its physical operation. The general phenomenon alluded to
in this statement is called asymptotic stability, a topic which is presented in Chapters
7, 9, and 10.
 Initial value problems 25

In this chapter, we consider Eq. ([Link]) only for the cases n = 1 and n = 2.
Naively, one would hope to perform upon the equations x’ = f(t, x) and x’ =
S(t, x, x’) a finite number of operations from the calculus and obtain finally the
symbol x on the left and a simple combination of elementary functions of ¢ on the
right. We saw in Chapter 1 that this can sometimes be done. It cannot always be
done, however. Consider, for example, the equation x’’ + sinx = 0. One can
easily check that x = 0 is a solution, but no other e/ementary solutions are known.
Nevertheless, the equation has infinitely many solutions: precisely one for each of its
initial value problems. The truth of this assertion is a consequence of the theory
developed in Chapter 8. For completeness, however, we briefly discuss here the
existence of solutions for the equations x’ = f(t, x) and x” = f(t, x, x’).
Notice first that the graphs of solutions for x’ = f(t, x) are curves in the real
tx-plane, and these curves must lie in regions for which the function /f is defined. For
example, the graph of a real solution for x’ = \/x — ¢ must lie in the region D
defined by the inequality x — ¢ > 0. The graph of a solution for x’ = 1/tx must
lie in one of the four regions defined by the inequalities x > 0 and tx < 0.
The following theorem, which we shall call an existence and uniqueness theorem,
is actually a corollary of Theorem 8.3 (Chapter 8). We shall therefore merely state
and explain it here. The theorem, which is illustrated in Fig. 2.2, guarantees that
many initial value problems
x =f (bX), X= Xo when bets (2.2a)
have uniquely determined solutions.
Theorem 2.1. Let f and df/dx be continuous on a region © in the tx-plane and let
(to, Xo) be a given point interior to D. There is one and only one solution x = ¢(t),
Te <t< 74", ofx’ = f(t, x) with the following properties:
1) $(to) = Xo
ii) (t, (2)) is a point of D forts” <t < T4°;
iii) either |t| + |(t)| ~ +o or (t, (1)) approaches a boundary point of D as
t— Ts and as t—T¢".
Example. For the equation x’ = 1/tx, f(t, x) = 1/tx. Let xo and fo be positive
constants so that (fo, xo) is in the first quadrant of the tx-plane. Both f and
of (t, x)/ax = —1/tx® are continuous on the interior of the first quadrant; we may
therefore choose it as the region D of Theorem 2.1. The theorem guarantees that the
initial value problem
1
X=
a =
2 x = Xo when bes (2.2b)
tx

has a solution x = ¢(f). It is found by integrating ¢(1)¢’(t) = t~} to obtain

o (ft) — x = 2 In |t/tol;
which implies that
o(t) = +(xo — 2 In |t/tol)"””.
 9494
26 Solution methods for special first and second order nonlinear equations

Figure 2.2

Since (t, (2) is contained in the first quadrant, (7) is given by the positive radical.
As t—> 0+, (1) > +o. As t— to exp (x2/2), the point (¢, ¢()) approaches the
point (to exp (x2/2), 0) on the ¢-axis. For this initial value problem, the interval
Tan < tere" is therefore 0-7 Tyexp:(G/2)) ||
A result similar to Theorem 2.1 holds for equations of the form
x Xe): (2.2e)
Theorem 2.2. If f(t, x, y), Of (t, x, y)/dx, and of (t, x, y)/dy are continuous on a
region D in real three-dimensional space and if (to, Xo, Yo) is a point interior to D, then
there is a unique solution x = $(t), Ts < t < T¢*, of Eq. (2.2c) with the following
properties:
i) $(t0) = Xo, $'(to) = Yo:
ii) (¢, 60), 6’) is in D for tT4— < t < tet;
iii) either \t| + |6()| + |e’O| > +a or (t,¥(1), ¥/(D) approachesa boundary
point of Das t— Tg" and ast—T,~.
In the remaining sections of this chapter, we shall concern ourselves with methods
for solving x’ = f(t, x) and x’’ = f(t, x, x’) in terms of elementary functions. Only
special cases are considered, and the reader is cautioned that the collection of tech-
niques is not exhaustive. The most complete catalog of solution methods has been
assembled by Kamke [11].*

EXERCISE 2.2
1. Carefully graph the solution x = ¢() found in the example above.

* Bracketed numbers refer to the books listed in the Further Reading section at the end of
the text.
 The separable equation 27

2. Find the function f and a region D for each of the initial value problems listed below.
a) x’ = (1 — x? — 72?)!/2, x» = 0 when ¢ = 0.
b) x’ = t/(x? — #2 — 1), x = 0 when ¢ = O.
c) x’ = 1/(1 — x? — (%’)?), x = 0 and x’ = 0 when ¢ = 0.
d) x” = dn d/(x? — (’)?), x = 1 and x’ = 0 when ¢ = 1.
3. The solution to the initial value problem

oe! ss Deel. xe = 0 and el when ia—a0)

is x = tan?f. Illustrate Theorem 2.2 by sketching the locus of the parametric equations
We tanto ve= SOC1 i. 2 = 11, — 1/2 em oem2.

2.3 THE SEPARABLE EQUATION

In Section 1.2, we solved the first order linear equation x’ = a(t)x, x # 0 by writing
it in the form x’/x = a(t) and integrating. In the last example, we solved an initial
value problem for the equation x’ = 1/tx by writing xx’ = 41, x = ¢(t) and
integrating.
These differential equations are examples from a class of differential equations
called the separable equations. The equation x’ = f(t, x) is called separable if the
values of f have the form f(t, x) = a(t)b(x), where a and 6b are continuous.
To solve an initial value problem
xX sau), X= XG when b= to; (2.3a)

posed for a separable equation, one first determines whether or not b(xo) = 0. If
b(xo) = 0, then x = x9 is a constant solution of the initial value problem. If
b(xo) ¥ 0, then b(x) # 0 for |x — xo| sufficiently small. One then separates the
variables and integrates to obtain the relation

/ Mam i a(s) ds. (2.3b)

Example 1. To solve the initial value problem
Xen x: XS XH when ft = 16, (2.3c)

write (In x)x’ = ¢f and integrate to obtain

x ae

/ Inudu = / Saks
xo to

Evaluation of the integrals yields the relation

im
xInx
— x= XpInXo — Xo 5 Se (23q))

and the solution x = ¢(f) to the initial value problem (2.3c) must therefore satisfy
the algebraic equation (2.3d). It is, unfortunately, not possible to solve this equation
for x as an elementary function of f. One nevertheless analyzes it to study the
solution ¢. ||
 P48)
28 Solution methods for special first and second order nonlinear equations

Example 2. Let us solve all the initial value problems

x! = 2tx?, ~ x=Xxo j\when ft = fo.
If xo = 0, then x = Xo is the unique solution. If xo # 0, then
9 t

itu—* du = 2 | sds
ro to

and

—— = 15),
ml
Example 3. Let an object of mass m be dropped in vacuum from an altitude x9 and
assume that the acceleration g due to gravity is constant. The altitude x of the object
then satisfies the initial value problem

Xie = N/ 6 cy ee), x = Xo when t = 0. (2356)

One solution of this initial value problem is the constant solution x = xo. This is
not the solution that describes the descent of the object however. We find the physi-
cally meaningful solution by evaluating an improper integral; specifically, we write
© t
? du
tim Deo — mp2 ~ =f,oe
and find that [2(xo — x)/g]!/? = ¢. Consequently, x = xo — gt?/2 is the re-
quired solution. It is important to note here that the initial value problem (2.3e) has
two solutions. Theorem 2.1 is not contradicted, however, since

<g(xo — x)? = —g12e(to — x2?

is discontinuous at x = xo. ||
Example 4. Chemical engineers frequently need to compute the time required for a
storage tank to empty. To illustrate the principles involved, let us consider the
conical tank in Fig. 2.3. In a time interval ¢ to t + At, the surface of the liquid
(density p) falls from a height h to a height h + Ah. The corresponding volume of
liquid enters the drainpipe and occupies a cylindrical region of height As and radius
a < R. From the geometry of the cone, we see that

V(t) — Viet Af) = 3°) — FO + AD).

On the other hand,
V(t) — Vit + At) = ra? As.
Thus

1 (t+ Ad) — rt) _ stirs As

3 At Sowa Ag
 The separable equation 29

Figure 2.3

and, in the limit as At — 0,

£ (0) = —3a’v,

where v = ds/dt is the velocity of efflux. Provided the orifice is in a flat bottom, the
velocity of efflux of a nonviscous fluid from a tank of otherwise arbitrary shape is
known to be v = +/2gh, where h is the depth of the liquid. Thus

r’r! = —a’x/2gh.
Since h = r — a, it follows that

r’r! = —a’? 2g(r — a), r(0) = R.

The student will be asked to find the time required to empty the tank in an exercise
below. A word of caution about the use of such a mathematical model is in order:
do you think the model is a good one if the tank happens to be filled with a very
viscous fluid? ||

EXERCISE 2.3
1. Solve the following equations:
a) x = V1 — x?, b) x’ = @ +. ~*)/( + #),
cx = 12x2/3, d= x/(t? +2¢+1)¢>-1.

2. Find the time required to empty the conical tank of Example 4.

3. What time is required to empty a cylindrical tank into a vertical pipe affixed to its
bottom if fluid friction is neglected?
4. Consider a chain that is suspended by its ends. Construct a rectangular tx-coordinate
system with origin at the lowest point and ¢-axis tangent to the chain. The shape of the
chain will be given by the solution of the initial value problem x” = CV/1 + x’2, x = 0
 2.4
30 Solution methods for special first and second order nonlinear equations

and x’ = 0 whent = 0. Here C > Oisa constant characteristic of the particular chain used.
Identify the curve of suspension.
5. The chemical law of mass action states that, at constant temperature, the rate of forma-
tion of a compound is proportional to the product of the concentrations of the reactants.
Suppose a manufacturer of magnesium hydroxide wishes to produce it by the reaction
MgO + H20 — Mg(OH)e.

A mole of MgO is 2.4 times as massive as a mole of water. Thus the mass ratio of the reac-
tants for complete reaction is 2.4/1. Suppose the process is started with a mixture of 2400 kg
of MgO and 1000 kg of water. If x denotes the number of kilograms of Mg(OH)2 at time f,
then according to the law of mass action,
2.4 3%
aele 2400 — —
k ( (1000——]-
34 x) x)

Express x in terms of t and k. Suppose 1 kg of Mg(OH)z is produced during the first hour
of the reaction process. How many hours are required for the manufacture of the batch?
6. Describe all curves x = ¥(f) which have constant curvature w’’(A)/(1 + ¥/2(0)?/?
and are tangent to the f-axis at the origin.
7. Amotor boat of mass M is cruising in a straight line on a calm lake. The motor is shut
off and the boat decelerates in a straight line due to the friction of the water. If x denotes the
distance of the boat from the spot at which the motor was stopped, then x’ is its speed.
Discuss the motion of the boat if the frictional force is proportional to
Ae KeemeO <ce cals b) x. cy Gtts 6.0)

2.4 FIRST INTEGRALS AND IMPLICIT SOLUTIONS

We have defined a solution of the differential equation

See ENN ee ee ee) (2.4a)

to be a function x = ¢(f) that satisfies the differential equation at every point of an

open interval. A solution of this type is called an explicit solution, and the process of
finding explicit solutions is called solving the equation explicitly.
For many specific cases of the differential equation (2.4a), finding all the solutions
explicitly is a very difficult task. Nevertheless, useful information can sometimes be
obtained from first integrals and implicit solutions.
A first integral for Eq. (2.4a) is a function y = U(t, x1, X2,..., Xn), defined
on the same domain as f, such that
d is
i U(EGO), ano (=O (2.4b)

for each explicit solution ¢.

Example 1. In Section 1.6, we considered a block having mass m which was moving
on a horizontal surface under the influence of a linear spring with stiffness coefficient k.
 First integrals and implicit solutions 31

We showed that, in the absence of friction, the displacement x of the block satisfies
the undamped linear oscillator equation

x’ +0*?x=0, ww? =k/m. (2.4c)

The block’s kinetic energy is my?/2, where y = x’, and its potential energy is
kx?/2. The total energy
U(x, y) = Bkx? + Ymny?
of the block is a first integral for Eq. (2.4c). To see this, let x = $(f) denote an
arbitrary solution of the differential equation so that

¢’'(t) + wo(t) = 0.
Then

U($(), #'() NZ £ (60 + w°9*(Y)

© (26'o") + 02460")
= m¢'(t)(¢’"(t) + w(t) = 0. ||
Example 2. Every differential equation of the form x’’ = f(x) has

U(x, y) = $y" — | 1) dx (2.4d)

for a first integral if f is continuous. One can discover this for himself by observing
that
$e) — f(4)e'O = 0 (2.4e)
whenever x = ¢(f) is a solution of the differential equation. The left side of Eq. (2.4e)
is an exact derivative, and it therefore has the equivalent form

4 ony? — 4 |ow at = 0.
Thus U is indeed a first integral for x” = f(x). ||
In studying a second order equation
He as Me
4g SRS ce)Bb (2.4f)

it is frequently worthwhile to look for first integrals. Suppose, for example, that we
wish to find a solution x = ¢(t) to some initial value problem for Eq. (2.4f). Ifa
first integral U is known, then the defining relation

7d Ul oO, ye
#') = 0
may be integrated to yield U(t, ¢(2), ¢’(t)) = ¢ for some constant c. Thus
UGA xe) =-¢ (2.4g)
32 Solution methods for special first and second order nonlinear equations 2.4

is a non-normal, first order differential equation having x = ¢(t) for a solution. If

Eq. (2.4g) can be solved algebraically for x’ in terms of ¢ and x, say
x’ = f(t, x), (2.4h)

then x = ¢(f) will be a solution of the resulting normal equation.

Example 3. We saw in Example | that

2 (OY + 074%(0] = 0
for each solution x = $(t) of the undamped linear oscillator equation (2.4c). The
specialization of Eq. (2.4g) to this case is therefore
(x')* + w*x? =e:
Solving for x’, we find the analog
dx = — w2x?)1/?

of Eq. (2.4h). This equation is separable. We solve it by writing

+ |Sane |e dx

and evaluating the integrals to find that

cos! (wx/+/c) = +(wt + 7),

where Y is a constant of integration. The solution is more commonly displayed in
the form
C F
sie ae (wt + Y) = c, Sinwt + c2 cos wt,

where
c¢, = —Vc(siny)/w, c2. = Vc (cos Y)/w.
We therefore conclude that every solution x = ¢(t) of the undamped linear oscillator
equation has the form ¢(t) = c; sin wt + cg cos wt for some choice of constants c,
and cy. Conversely, one verifies by substitution that x = c, sin wt + Cc, cos wt is
a solution for every choice of the constants c; and cy. This is precisely the assertion
which was made in Section 1.6 about the solutions of x” + w?x = 0. ||
It is natural to ask at this point whether or not there is a notion of second or third
or fourth integral for nth order differential equations. In fact, there is such a notion.
A kth integral, 1 < k <n, forx™ = f(t, x, x’,...,x@7) is a function U, defined
where f is defined, such that
1

ee U(t, 60), o'@), ash 0

 The exact equation 33

for every solution x = ¢(t). One computes a kth integral (when he can) by performin g
k integrations. This is the origin of the name.
Consider now a first integral of a first order differential equation, that is, a
function U such that
d
7 Ul ) = 0 (2.4i)
for each explicit solution x = ¢(/). Integrating Eq. (2.4i), we have U(t, (t)) = c
for some constant c. Thus each explicit solution x = (ft) is also a solution of the
equation
UGIX) =" (2.4j)

For this reason, Eq. (2.4j) is called an implicit solution of the original differential
equation. The value of the constant c is determined by an initial condition x = xo
when ¢ = fo.

Example 4. The relation

2
SETN J 8 ce
t
bo Xo
In x9 — Xo — 6
2
was shown to be an implicit solution of x’ = t/In x in Example | of Section 2.3. ||

EXERCISE 2.4

1. Find a first integral and implicit solution of the differential equation (1.4c) r’’ = —Gm,/r?
derived in Section 1.4.
2. Find first integrals and implicit solutions for each of the following differential equations.
a) exe — 1/0 b) x’ = 2t/(1 + cos x).
c) x= (1 + x8)1/2, d) x” = —k?/x8.
e) x’ + sinx = 0. f) x” = —kxlt?,

3. Verify that U(t, x,y) = x? + y + cit + ce is a second integral of x’”’ + 3x?x’’ +

6xx'? = 0 for every choice of the constants c; and cg. Find the solution of the differential
equation that satisfies x = 1, x’ = 0, and x’ = 0 when ¢ = 0.
4. Solve the differential equation x’ — w?x = 0, where w > 0. [Hint: Pattern your com-
putations after those in Examples 1 and 3.]

2.5 THE EXACT EQUATION

Consider again the general first order equation

x! = £(i, x), (2.5a)

where f and df/dx are continuous on a region © in the /x-plane. Theorem 2.1
guarantees that every initial value problem consisting of Eq. (2.5a) and the initial
condition x = X9 when ¢ = fg has a solution if the point (fo, xo) is interior to .
 7)
34 Solution methods for special first and second order nonlinear equations

This theorem does not, however, guarantee that every solution of (2.5a) can be
expressed in the form
U(t, x) = c¢ (2.5b)

for some one function U and various constants c. The particular equations (2.5a) for
which this can be done are called exact equations. More formally, one says that
(2.5a) is exact if and only if there is a function U defined on D such that

é U(t, x) = 0 (2.5c)

for each of its solutions x = ¢(f).

What form must f have if Eq. (2.5a) is to be exact? Well, if the equation is exact,
then a relation of the form (2.5b) has to hold for each of its solutions. Performing the
indicated differentiation, we find that
aU A
Bp 3) Tae a 0,

which implies that

“i oU(t, x)/dt
oU(t, x)/dx
aU(t, x)/dx # 0. Thus an exact equation must have the more specific form

»_ _ MG,x), (2.5d)
where M(t, x) = OU(t, x)/dt and N(t, x) = dU(t, x)/ax for some function U.
To decide whether or not an equation given in the form (2.5d) is in fact exact,
one must determine whether or not there is a function U such that dU(t, x)/dt =
M(t, x) and dU(t, x)/dx = N(t, x). The next theorem gives the most common way
of making this determination.

Theorem 2.3. Let M/N, 9M/dx, and dN/dt be continuous for a < t < w and
a<x <b. Then anecessary and sufficient condition that Eq. (2.5d) be exact is that
OM(t, x)/dx = ON(t, x)/dt fora <t<wanda<x<b.
Proof. Assume first that the equation is exact. Then there is a function U such that
dU(t, x)/dt = M(t, x) and dU(t, x)/dx = N(t, x). Since M and N have continuous
first partial derivatives,
aM a a°U - a°U _ aN
ax) = axa; age age
Conversely, suppose that 0M/dx = dN/dt. To prove that the differential equa-
tion is exact, we shall verify that

(OES) = (iN(t, u) du + | M(s, Xo) ds, (BAS

 The exact equation 35

a<t, to < w,a< x, Xo < 5b, is a first integral. First note that

aU * 9N
OL (t,x) = [ Or (t,u) du + M(t, xo)

“aM
= ifiy (t,u) du + M(t, xo)

= M(t, x) = M(t, Xo) ale M(t, Xo)

I M(t, x)
and

a (i, x) = N(t, x).

Then
d au
i UG x)= “ai (t,x) + x ee CST)
,oU mI MG) ;
ENG)
7= 20;

and the theorem is proved. ||

To solve an exact equation, it is frequently more efficient to compute U directly
with indefinite integrals than to compute it by mechanically applying formula (2.5e).
What is actually wanted is a function U such that dU/dt = M and dU/dx = N.
Integration of the latter equation with respect to x gives

U(t, x) = /N(t, x) dx + g(t), (25h)

where g is an unknown function. Differentiation with respect to ¢ produces, in turn,

the relation

M(t, x) = aU(t,x)/at = /oe Ce eb:

But dN/dt = 0M/dx. Thus

e'() = MUt,x) - /om bx d= h(O),

where / is a Known function. It follows that

UG, x) = iexa x) dx + io dt

is the desired first integral.

Example 1. The equation x’ = —(x? + 2°)/2xt has M(t,x)=x?+ 2° and
N(t, x) = 2tx. Consequently, 0M/dax = 2x = dN/dt, and

UEx) = iixe x) dx + g(t) = ix? + g(t).

36 Solution methods for special first and second order nonlinear equations 7a)

On the other hand,

Mx xt + 2’.
x? + g(t) = a Cy)

that
Thus g’(t) = 1”, g(t) = t°/3 + ¢, and it follows
2
U(t,x) = tx? + ae
is a first integral of x’ = —(t? + x”)/2tx. ||
Example 2. Let a denote a constant and let b denote a continuous function. The
equation
ae“'x — b(t)e™
x= = eat
(2.5g)
is exact with M(t, x) = ae“x — b(tje% and N(t, x) = e”, since dM(t, x)/dx =
ae“ = AN(t, x)/dt. In this case, Eq. (2.5f) has the form

UG x)= [ etas = ex + g(t),

which implies that

ae“'x — b(t)e™“ = ae“x + g'(t).
Consequently, g’(t) = —b(t)e™, and it follows that

U(t, x) = ex — ion div="c

for some constant c. Cancelling the factor e*’ in Eq. (2.5g), one finds that every
solution of x’ + ax = b(f) is given by

Xe Cie. /b(t)e™dt

for some constant c. Compare this result with the material in Section 1.3. ||

EXERCISE 2.5

1. Solve those of the following equations that are exact.

2
je a Octane , lap 2g
ee 2xt Dee as
x + 21 oaag
Cex = — . dy. es :
2x+t ye 2xt

Vie Se ha 2
12 x+ £2
 Integrating factors 37

g) x= — XOO Ga pe ea4
3
COS xt
+ tcos xt 12 + 3xe"')
2 3
: 2i eal a | , x
i) xi = — —_____.. ze :
) x2+t+1 ae! tln xt

2. A system of differential equations having the form

7
dx
——— F
Fes =dyS = Gx,y)
—_—_ =

is called a planar Hamiltonian system if there is a function H such that F(x, y) = 0H(x, y)/dy
and G(x, y) = —dH(x, y)/dx. Verify that each solution x = $(4), y = W(t) of a planar
Hamiltonian system satisfies H(x, y) = c for some constant c.

2.6 INTEGRATING FACTORS

The differential equation

x x” cos xt
x’ =
2 sin xt + xtcos xt

is not exact. If we multiply the numerator and denominator of its right side by x,
however, we obtain an equivalent equation,

x? cos xt
x’ = — = >] (2.6a)
2x sin xt + x2tcos xt

which is exact. Applying the solution procedure for exact equations, we find that
each solution of Eq. (2.6a) satisfies x” sin xt = c, for some constant c.
Similarly, if we write the linear equation x’ + ax = b(t) in the form
pl) Bee) :
iv ]
(2.6b)
then we find that it is exact if and only if a = 0. We saw in Example 2 of the last
section, however, that
ae™ — b(t)e”
x = — eat

is exact for all values of a. This equation differs from Eq. (2.6b) in only one way: its
numerator and denominator contain the common factor e“’.
As these examples suggest, it is frequently possible to make a first order equation

exact by multiplying the numerator and denominator on the right by the value of a
function p called an integrating factor. To find u(t, x), one requires that

2 (M(t, ult, 9) = 2(NG, Yale >))

 Solution methods for special first and second order nonlinear equations 2.6
38

and tries to solve this partial differential equation for y(t, x). When the differentiations
are performed, it is found that » must satisfy the equation
on , OM _ 4, 9p ON ; (2.6c)

Solutions of Eq. (2.6c) can sometimes be calculated by use of the following trick:
Set u(t, x) = a(x)b(t). Then the equation reduces to
a(x) , oM DO ON.
uesa(x) “i oxen a b(t) a Ot Ged)

One solves Eq. (2.6d) for a(x) and b(t) (when he can) essentially by inspection.
There is a general guideline for its solution, however: Set one of the ratios a’(x)/a(x)
or b’(t)/b(t) equal to the value of a known function of x or ¢ in such a way that there
results an equation involving only one independent variable. This equation is then
solved for the remaining unknown a or b.
Example 1. Let x’ = — M(t, x)/N(t, x) with M(t, x) = —x? and M(t, x) = t ln xt.
For this case, Eq. (2.6d) becomes

eae)
as (x) Seay =14+ 1nxt+
b'(t)
Gy In xt.

Setting b’(t)/b(t) = —1/t, we obtain a linear equation

a’'(x) + 3x7! + x7)a(x) = 0

for a and find that a(x) = x7? exp (1/2x”). Thus p(t, x) = tf
1x—3 exp (1/2x?”) is
an integrating factor for the original differential equation, and we are assured that the
equivalent equation
; polel!22?
xs 5
a er maalinoct
is exact. ||
Example 2. The equation x’ = —(2x? + 4xt?)/(4x7t + 22°) is not exact. To find
an integrating factor wu of the form u(t, x) = a(x)b(t), we must choose a and b so that
7x3 a'(x) = b(t)
(x2 Ants) AO) ONS (47 PyA(t) + 2r.2 (2.6e)

Setting a’(x)/a(x) = 1/x reduces Eq. (2.6e) to

4x?
x* +4 27? == (4x9)
z b'(t)
BD)”| (2.6f)

from which it follows that b’ QO) = I/t. Thus a(x)= x, b(t) = t, and p(x, t) =
xt. The equation x’ = —(2x*t + 4x?t3)/(4x31? + 2xt*) is exact and has the same
solutions as the original equation for xt 4 0. ||
 Reduction of order 39

EXERCISE 2.6
1. Use the integrating factor e** to solve the equation x’ = —(1 + tx)/t?:
2. Find an integrating factor for each of the following equations.
3
Qin Pees atx ; ae x8 ap Bi
Pe 2t + 212x2 2 ange eye
[pce 6x" + 6tx 3 [ag a
XG
.
ae Oxt + 472 os 1+ xe
avers cot x — fo 6 + 6sin x f
x+ft 2t2 + 3r(sin x + cos x)
3. Verify that the equation x’ = — M(t, x)/N(t, x) has an integrating factor p(t, x) = b(t)
if (OM(t, x)/ax — ON(t, x)/dt)/N(t, x) depends only on ¢.
4. Verify that x’ = —M(t,x)/N(t,x) has an integrating factor u(t, x) = a(x) if
(ON(t, x)/dt — OM(t, x)/dx)/M(t, x) depends only on x.
5. Construct a nonseparable, nonexact equation x’ = —M(t, x)/N(t, x) which has
g(t, x) = e+) for an integrating factor.

2.7 REDUCTION OF ORDER

In Example 2 of Section 2.4, we showed that any solution x = ¢(t) of

x = f(x) = 0 (2.7a)
satisfies

‘|wor — /CO)LO a = 0.
Integrating this equation, we find that

et) = + E+2 /fo) ¢'O ar)

for some constant c. Thus x = ¢(f) also satisfies the separable first order equation

eerie
A generalization of the technique above produces the same result for the more
general equation
Xf (Oo) (2.7b)
To illustrate the generalization, let x = ¢(t) be a solution of Eq. (2.7b) and let
y= ¢'(t). If ¢’(@ # 0, then
dx _ @ xd) de dy 1a)
Poe ee Wigs dre didtedx
Substitution of these identities into the equation reduces it to a first order equation
dy _ fy),
dx y
8
 9X54|
40 Solution methods for special first and second order nonlinear equations

Now suppose that Eq. (2.7c) can be solved fory in terms of x, say y = g(x). Then,
since x = $(t) and y = ¢/(f), it follows that @ satisfies the separable first order
equation x’ = g(x). This equation, when solved, yields an implicit solution of
Eq. (2.7b).
Example 1. Setting y = x’ in the equation x” = —x'/x, we have y dy/dx = —y/x,
or dy/dx = —x~+. Thus y = —In |x| + c, where c is a constant. It follows that
U(x, y) = y + In |x| is a first integral of x’’ = —x’/x, and

ae
¢ — In|x|
is an implicit solution. ||
Example 2. Let us solve the boundary value problem

= 0, x = 0
x! +. 2x'x when 7=0 “and 4— lt as°%— 4.

We formally make the substitution x’ = y, x’ = y dy/dx and reduce the differential

equation to dy/dx + 2x = 0. Its solutions are of the form y = —x? +c. Since
x—last—> +o, x’ = y—-—1+c. This can only happen for c = +1. Thus

| dx ke
wae
and
1 [eer

Since x = 0 when t = 0, k = O also. Consequently, x = (e7’ — 1)/(e?’ + 1). ||

EXERCISE 2.7
In Problems 1 through 3, solve the indicated initial value problems for x in terms of ¢. In
Problem 4, solve for ¢ in terms of x. In Problems 5 through 18, find (first) integrals of the
equations.
Le ex 8 = 0x0) =x Oar
2x DN x = 10, xr /4) = Le 4) = 2)
3. x" + x’? cot x = 0, x(0) = 7/3, x'(0) = —2/V3.
AS x! + 2x! Sx ys == 0 x(0) = (0x O)i=rt.
5. x" + e* = 0, 6x 07 He er 0!
7. x" — xx' +x =0. 8. x” + f(x) = 0, f continuous.
9. x’ + xsecx’ = 0. 10>x" + x/3x’ = 0.
11. x! xx 2S Ov x= 0) 12. x!” + xx’? — S5xx’ + 4x = 0.
13. x" + xx’? +x =0. 14, lex Fe eet 0:
ley, oe! NE Se SS so) = (0). 16. x!’ + xx’? = 0.
17 oe ee) te 18. x” + x3e-*’ = 0,
 A soft spring oscillator 41

2.8 A SOFT SPRING OSCILLATOR

Figure 2.4 shows a spring which has been stretched x units. The force — f (x) which
the spring exerts on the point p is called its restoring force. A restoring force —f(x)
must satisfy the condition xf(x) > 0 for all sufficiently small nonzero values of x.
This inequality is just a mathematical way of saying that the restoring force opposes
compression (x < 0) and extension (x > 0) of the spring. Assuming that f is con-
tinuous, one can see from the intermediate value theorem that f(0) = 0, that is, the
spring exerts no restoring force when it is unstressed. If there is a positive constant k
such that f(x)/x < k for all sufficiently small nonzero values of x, the spring is called
a soft spring. A soft spring resists extension and compression less forcibly than a
linear spring does.

hu NT LWW ——> f(z)

Figure 2.4

Now consider a 1-kg block which moves horizontally on a frictionless surface

under the influence of a soft spring with restoring force —f(x) = —x + x, |x| < 1.
Newton’s second law of motion gives rise to the differential equation
ee = x = 0 (2.8a)
for the displacement x of the block from equilibrium x = 0.
Multiplying Eq. (2.8a) by y = x’ and integrating, we obtain the first integral

U(x, y) = By" + 3x? — Bx’. (2.8b)

The quantity U(x, y) is the total energy of the moving block: y*/2 is its kinetic energy
and x?/2 — x*/4 is its potential energy. The implicit differential equation
4
MEA gt = =2F 2E=x’()+ x70) — x*00)/2, (2.8c)

for x is called an energy conservation law by physicists.

To find an implicit solution of Eq. (2.8a), one solves for x’, separates variables,
and integrates, obtaining the relation
dx
= 2.8d
PRE xa2pe a)
Equation (2.8d) is a perfectly good implicit solution, yet physical information about
the motion of the block is difficult to extract from it.
Another way to study Eq. (2.8a) is to plot Eq. (2.8c) with x’ = y in the xy-plane.
A picture showing all the graphs obtained from
x?

yo-x+5+2E bls,
42 Solution methods for special first and second order nonlinear equations 2.8

for all constants E > 0 is called a phase portrait. In this context, the xy-plane is
called the phase space for the differential equation.
To make a phase portrait, one first finds constant solutions of the original dif-
ferential equation. This is done for Eq. (2.8a) by setting x = c, x’ =0, x” =0. The
resulting algebraic equation c — c* = 0 shows that x =0, x = 1, and x = —]1 are
constant solutions. In the phase plane, y = x’. Graphically, then, we depict the
constant solutions by plotting the points (0, 0), (1, 0), and (—1, 0) (Fig. 2.5). These
points are called critical points.

Figure 2.5

The next step in the construction of the phase portrait is to plot each one of the
curves y? = 2E — x? + x*/2, |x| < 1 that has a critical point on it. If x = 0 and
y = 0, then E = 0 and we must plot y? = —x?(1 — x?/2), |x| < 1. The only
point satisfying these conditions is the critical point (0,0). If x = +1 and y = 0,
then 2E = 4 and we must ploty? = (1 — 2x? + x*)/2 = (1 — x?)?/2 for |x| < 1.
Doing this, we obtain the curves C; and C, shown in Fig. 2.6.

Figure 2.6

Now suppose the block is put into motion with initial conditions x = xo and
Y = Yo which satisfy
4
Pats eeears!
yotx Ape a
 A soft spring oscillator 43

and let x = ¢(t), y = $’() denote the corresponding solution of Eq: @.8a). If we
plot this solution parametrically, its trace will be the closed curve in Fig. 2.6. There
are four physical motions possible for the block.
The first (second) motion is initiated by stretching (compressing) the spring until
x = 1 (x = —1)and then releasing the block with velocity y= x’ = 0. The velocity
of the block remains zero and the displacement remains x = +1. The spring is
thus too soft to return the block to the position x = 0.
The third (fourth) type of motion occurs when the point (xo, yo) is in the upper
(lower) half-plane. The trace of the equations x = ¢(f), y = $/(f) is then the curve
labeled C, (Cz). Thus ¢(f) > +1 and ¢’(t) > 0 as t > +0, and the block must
move toward one of its extreme stationary positions x = +1 as time passes. The
orientation arrowheads point to the right (left) in the upper (lower) half-plane,
since x, = y-
If the block is at rest with x = +1, say, and if it is very slightly displaced to the
left, it will move to the stationary position x = —1, its displacement and velocity
functions x = ¢(f) and y = ¢’(f) tracing the curve Co.
If the block is set into motion with initial conditions x9 and yo such that
0 < yo + x% — x§/2 < 1/2, then the curve y? = 2E — x? + x*/2, |x| < 1, is
an oval C3 lying inside the curve y? = (1/2) — x? + x*/2. Since there are no
critical points on this oval, it is the complete trace of the corresponding solution
x = ¢(t), y = ¢’(t) of Eq. (2.8a); and it corresponds to a periodic motion of the
block.
The period of motion can be found with the aid of Eq. (2.8d). To do this, we
take xo = 0 and 0 < po = V2E < 1/\/2. By symmetry, then, the period T for
one complete oscillation satisfies
ete dx a Te

[QE — x2 + x4/2pl2 4
By analyzing this integral carefully, one can show that T— +o as E— 1/4 and
T—0Oas E—O0. Thus the period of oscillation depends on the total energy of the
block, and hence on the amplitude 21/?E/* of the oscillation. Such a dependence is
characteristic of nonlinear oscillators. In linear oscillators, amplitude and period are
completely independent.

EXERCISE 2.8

1. A spring with a restoring force f(x) = —x — x? is an example of a hard spring. Sketch

a phase portrait for the equation x’”’ + x + x? = 0,
2. Sketch a phase portrait for the equation x” + x!/3 = 0.
3. Sketch a phase portrait for the equation x’’ + xe?” = 0,
CHAPTER 3

MATRIX METHODS FOR LINEAR EQUATIONS

WITH CONSTANT COEFFICIENTS

3.1 SIMPLY COUPLED SYSTEMS OF DIFFERENTIAL EQUATIONS

In applications, one frequently needs to study simultaneous differential equations.

(These are usually called systems of differential equations.) For example, we showed
in Section 1.7 that the motion of two oscillating blocks is described by a system

Wa 3% p
Xe = ea kee 4. Ka F

my my,
Jos ue
|ip ae
ko
eee
ko
eS,

ata Mog ms»

of differential equations.
Systems of differential equations involving many unknown functions can occur.
It is therefore desirable to have a systematic scheme for keeping track of the data and
unknowns in such a system.
The purpose of this chapter is to set up such a scheme and apply it to the study of
differential systems having the form

Ma gt a Oe
(3.1a)
Xn = GniX1 Sa AnnXns

where the a;,;’s denote constants.

If $1,..-, ¢n are differentiable functions, defined for —a < t < +00, we shall
say that they form a solution for the system (3.1a) if the substitution x; = $,(0),...,
Xn = n(t) reduces each of the equations to an identity in ¢. We shall, conversely, use
the expression so/ution to mean such a collection.

44
 Simply coupled systems of differential equations 45

The system ([Link]) will be studied by our making a change of variables so that it
assumes the form

Yi = Ary + M172 5
yo = Moye) TT2y3- .
(3.1b)
Ya-1 = An—1Yn—1 + Vn—1Yns
yn = AnYns
where the \,’s are real or complex numbers and each 7; is either zero or one.
We illustrate a solution method, called the method of integrating factors, for the
system (3.1b) by explicitly calculating the solutions for the case n = 3, y; = x,
Yo = y, ¥3 = z. The system is then
ca Nyx s Yiy 5

y’ = Aeoy + Y 22,

i \3Z.

The last equation is solved by the method used in Section 1.2 to give z = Ae‘,
where A is a constant of integration. The second equation is then written
y= hoy = AY xe.
One multiplies by the integrating factor e~*2", as was done in Section 1.3, and observes
that
d (ve?) = Axe

dt
There are two cases. If \3 # Xo, then one finds upon integrating that

y = Bed2! + Avge*s'/(X3 — do),

where B is a second constant of integration. If 43 = Xz, then
y = Bez! + Avote®s,
The final equation will have either the form
x! — yx = 7, Ber2! + AV 2e%3'/(Ag — Az), =F Dv,
or the form
x! — yx = 7 Bed2? + Av 1Yote*’, when No = Na.

One multiplies the appropriate equation by the integrating factor ei! and proceeds
as before. In performing the integrations, we find that five cases arise: \y = \2 = Xs,
1 = \o # X3, Ay = A3 # No, AQ = AZF Nt, A1 FAQ# \3 ~ Ay. The student

will be asked to solve a specific equation for x in each of the five cases as an exercise
 LP.
46 Matrix methods for linear equations with constant coefficients

below. Notice that the calculations would be considerably simplified if one or both
of the constants Y;, Y3 were zero.
A system of differential equations having the special form (3.1b) will be called a
simply coupled system. This chapter begins with the development of vector/matrix
notation so that systems of the form ([Link]) can be written compactly. We shall then
turn our attention to changes of variables that reduce systems of the form ([Link]) to
simply coupled systems of a very special type called Jordan systems.

EXERCISE 3.1

Solve each of the following differential systems.

i =x ‘ Ds BRS te ‘
y= Sy 2, y= 2 ocks 22
z= oz. a 2z.
3. x =5x+ y , 4X eee ;
y’ = Sy + Ze y’ = 3y _ ae

z' = 2z z' = 2z

ee + oy ’

y= 2yi 2:
nS Pe

3.2 ADDITION AND SCALAR MULTIPLICATION OF VECTORS

If m real or complex numbers are written in a vertical (horizontal) array, the resulting
mathematical object is called a column (row) n-vector, and the numbers so arrayed
are called its components. For example,

1
3 X2
a S z and FO tox) = X3
©
eon —X X1X34X3 — ACL
/ — x2)
X9

are column 4- and 3-vectors, respectively. Notice that f(x,, x2, x3) has for its com-
ponents the values of three functions of the variables x1, x2, x3. For every f,

[Ei Vleet aT
is a row 5-vector. We shall use the unqualified word “vector” to mean a column
n-vector. Column vectors will be denoted by lower case boldface letters a, b, ¢,
¢(t),.... When a vector is to be regarded as a row vector, we attach to it the super-
script T. Thus a’, b’, c”, 7(2), .. . are the vectors a, b, c, #(f), ... written as rows
rather than columns. If x is a (column) vector with components xj,...,Xn, We
shall write x = [x1,...,Xn»]’ for it as a matter of typographical convenience. The
set of all n-vectors with complex (real) components will be denoted by V;,(@)(Vn(@)).
 Addition and salar multiplicati
of vectors
on 47

LAx = [x;,...,
and
%)y = [y1,..., yn] denote two n-vectors, One says
that x = y if and only if x, = y,fork =1,...,n. The swn of x and y is by defini-
Hon x+y = [4,4 1,---,%,
+ Yn}. Vector addiis
tion
associative and com-
that is,
mutative,
K+ Vt+D=K+y¥)4+z and x+y=y+x
for al) n-vectors x, ¥, and z. The vector 0 = [0,..., Of with all components equal
to zex0 is called the zero vector. By —x, we shall mean the vector[~x,,..., —x,J’.
LA a be any number. The product vector ax isthe vector [ax,,..., ax,J", and
the multiplication so defined is called scalar multiplication. We shall refer to real or
complex numbers 25 scalars. Mt is easy to verify the following properties of scalar
multiplication for arbitrary vectors x and y and arbitrary scalars a and b:
i) l-x = x,
ii) (abjx = athx),
ii) (a + bx + y) = ax + ay + bx + by.

EXERCISE 3.2
1, Construct 2 row vector and a column vector using the numbers 2, 7, 3i, »/2, 0 for

K+Y ;
& . =
aot miy+zi,thnx=-?,y=27,z=7
won!
—n
|
Zz

Ne) 4 NY \\ i \l ~»
~
\l
Ww
UA & NS, N
&lA
J a

We
ee
eas|
‘ene

| 5 i 0 deg de
2 2=7('= 3):
=
Sal —-~ oN = x10 4 y13 +2]1], ten x= y=

0 0
3 1 0 outs
M1, if = *10|+ y131+z
hen x= 2,y=27,
]1| , z=? wer”) bs
AY
WS
ae
SS
0 0 Z
48 Matrix methods for linear equations with constant coefficients ApS!

In each of the Problems 12 through 15 below, there are given three vectors p;=
Lx; Z;]’, 7 = 1, 2, 3, and a simply coupled system of differential equations. (1) Check
that the components x;, y;, and z; of each p; satisfy the given differential equations. (2) Let
C1, C2, c3 denote arbitrary constants. Form the vector q = cipi + c2Pp2 + cap3 and
check that the components of q also satisfy the differential equations. Are they the solutions
produced by the method of integrating factors?
12. pi = [e**, 0, 0]7, po = [0, e—*, OJ”, ps = [0, 0, e2t]T,
Ke ee ce
13. pins [ent 0, 0)ee p2 = [0, ees Oe ps3 [0, 16°%, eles
XC — ey) Vl ee

X=
OX ny ey Dylans
15. pi: = [cost + isin#, 0, 0]", po = [0, cost — isins,0]", p3 = [0, 0, e?']”.
x’ = ix, y’ = —iy, z’ = 2z. [Note: e* = cost + isint, i? = —1.]

3.3 MATRICES
If mn real or complex numbers are arranged in a rectangular array with m rows and
ncolumns, the resulting mathematical object is called anm X n matrix. For example,
: t 0
A= eel and BQ) = 2) sire
i Orel oil 7

are 2 X 3 and 3 X 2 matrices respectively. An n-vector, as defined in Section 3.2, is

just ann X 1 matrix; a row n-vector is a 1 X n matrix.
Anm X n matrix is called square if m = n. Thus

c-[ a
isa 2 X 2 square matrix. The only nonsquare matrices that we shall have occasion to
use in this book are row and column vectors. We therefore use the unqualified word
matrix to mean ann X n square matrix in all subsequent discussions. The integer n
will be called the size of the matrix.
In certain discussions, it is necessary to give designations to the elements of a
matrix. The symbolism A = [a;;] means that a;; is the number in the ith row and
jth column of A. We shall also write A = [a,,...,a,] for the matrix with columns
a; = [aij,..., jn)’, 1 <j <n. In using this notation, one thinks of A as n column
vectors written side by side.
Let A = [a;;] and B = [b,;] denote two n X n matrices and let a and b denote
scalars. One says that A = Bif and only if a;; = b,; fori, j= 1,...,n. We define
the sum of A and B to be the matrix (4 + B) = [a;; + b;;]. Matrix addition is
associative and commutative, that is,

A+@B+C)=(A4+8)+C and’ (44°=82)

@ 4)
 Matrices 49

for all X n matrices A, B, and C. The matrix 0 with all elements zero is called the
zero matrix. If A = [a;;], we denote [—a;;] by —A.
If A = [a;;], the product aA is by definition [aa,;;].
Then 1- A = A, (ab)A =
a(bA), and (a + b)(A + B) = aA + bA+ aBe bB.
The matrix A” obtained from an n X n matrix A by interchanging its rows and
columns is called the transpose of A. If A = [a;;], then A? = [a;,].
An n X n matrix may be multiplied from the right by an n-vector x to give a
product n-vector Ax. This multiplication is defined by the equation
a male fay qe oe
AXA) | Te: P= (3.3)
Ani +++ Ann}|Xn An1X} Se OS eee es

As a numerical example,
|! ;H ns E + i).
AS TN Sie 43
A fundamental property of this multiplication is the distributive property: If x
and y are any n-vectors and if A and B are any n X n matrices, then
(A + B)(ax + by) = aAx + bAy + aBx + bBy

for all scalars a and b.

The notion of matrix-vector multiplication leads to a convenient notation for
systems of linear equations. For example, the algebraic equations
4x + l6oy — 23z = 0,
Seo 4y+ 7z
= 0,
x + z=0

may be compactly written in the form

LSE
The differential equations (3.1a) can, in view of Eq. (3.3), be written as
,
x ral
: = 3 5
AG Ani +++ Ann|| Xn

If A = [aj,..., a] and B are any n X n matrices, the product BA is defined

by the equation BA = [Ba,,..., Ba,]; that is, the jth column of BA is the vector
Ba;,1 <j <n. For example,
Pees Gm lil ei eiocns3 V/2 + 2i 2x+2 6+7
Orme 2 Ore Bayes] = 2e 2y 14
Lea RSD) zee itet+3V2 x+y+3z 104+ 37
50 Matrix methods for linear equations with constant coefficients 3.3

Matrix multiplication is not commutative. An example is provided by the products

Nees WA ii 8 Be Al ee
Cel | eter | anaes al ae
and
be SO) RS Be|a dy
1ST 10 els |ee?
If A = [a;;] is ann X n matrix, we call the set of elements {a11, d22,-.-5 Ann}
the principal diagonal of A. The matrix J with ones on its principal diagonal and
zeros elsewhere is called the identity matrix. It has the property that AJ = IA = A
for every square matrix A. The matrix B is called the inverse of the matrix A if
AB = BA= J, If, for'example,
Lie cost sin | ae a bee —sin it
—sin t cost Sine 7 COST

then AB = BA = I. Thus B is the inverse of A.

Some, but not all, matrices have inverses. To illustrate this, let

Anaifetond|
1 —2
rane Zieh
OY
Then AC = 0, the zero matrix. If A had an inverse, say B, the computation

C=1C=](6AC = BUC) = B20 = 0

would be valid. Since C ¥ 0, A does not, in fact, have an inverse. When the inverse
of a matrix A exists, it is customary to denote it by A~’. A matrix which has an
inverse is called nonsingular, and one which has no inverse is called singular.
Letting A, B, C denote arbitrary n X n matrices, one may summarize the com-
putational rules for matrix multiplication as follows:
i) (AB)C = A(BC);
ty ZIRE Pe Ze
iii) AA~! = A7!A = J, whenever A~! exists;
iv) A 6) SB Ac.

EXERCISE 3.3
Perform the indicated operations.
27070)16
en Oe 1
0 0 1
 Vector- and matrix-valued functions 51

2, 0) Oak ik -@ 1-1 1]//1 0 1

5 f2 Te1 1 O1© it OHO T ©
Y © So © i 0-1 2{/0 1 O
6 16 — 4 —3 1 0)) ie 45 —3
7 1 —2 “le
—1 4 8. F i 10-1 2
QO @ Y @® aio @ fi
Oster 0:14 Owr1ae0 [2
9. f 0 1 [squared] 10. F 0 1 [cubed]
0 0 0 Oy © @

Express the following systems of equations in matrix notation.

ih, 8? eS Dye te gy, 3? ss Sy Se aap pd? ST
eX a2 sel Ze) oe ly, — Zane — ytDZ,
13. byx} |= aiix1 + aiox2 + a13x3,
Xg = 421X1 + a2Q2x2 + a23Xx3,
% w |= A31X1 + A32X3 + A33X3.

14, Solve the following differential equations.

lg e Ce a
mek Sy i Olli Xe el ON es

Za © OQ Alike 4h 0 0 —i||z

Perform the indicated matrix multiplications in Problems 15, 16, and 17.
Tews) eos eeeoe 91 a = 0
feed Ctwee S| 66 15 36 Bate
canine Ty) aaa en
ee? e341 fae260r 54 ae78:1) 62> al
fone te 4eOe 131-282-6191 ||.7 41) 1)
omcures||— 12 22709'41)|| oe =01
ee 4 420 % les
17.4=]4 3 1|| 10 9 1 Sr
2a 24 —3 A ; 0 8
18. Solve the differential equation

Siena” = Aa,
where A is the product matrix of Problem 15.
19. Work Problem 18 taking A to be the product matrix of Problem 16.
20. Work Problem 18 taking A to be the product matrix of Problem 17.

3.4 VECTOR- AND MATRIX-VALUED FUNCTIONS

The vector concept is useful in ordinary differential equations because it makes

possible the simultaneous discussion or manipulation of several functions. When it
is necessary to work with n functions ¢1,...,¢n of the real variable f, it will be
convenient to array the values of the functions as a column vector $(t) = [¢1(), ---
52 Matrix methods for linear equations with constant coefficients 3.4

on(t)]’. The correspondence ¢ — ¢(t) is called a vector-valued function of the real

variable t or, more simply, a vector function of t. The function itself is denoted by ¢,
and its value at the point ¢ is denoted by ¢(t). The functions ¢,..., ¢» are called
the components of @, and we write @ = [$1,..., ¢n]’-
Vector-valued functions of several independent variables are defined similarly.
et
SiGe Os es Iio ae es

denote the values of n functions fi, ...,f, at the point (%1,..., Xn, 1). We identify
the n-tuple (x1, ..., Xn) with the vector x = [x1,..., Xn] and define
fi, s+ ey Xns t)

{(x,1) = :
Iris +3 Xn;5 t)

The correspondence (x, t) — f(x, t) is called a vector-valued function of the vector

variable x and the real variable t and is denoted by f.
Matrix-valued functions are defined analogously. If, for example, the symbols
a; (37 = ly... @) denote n” functions defined on an interval a < t < w, one
forms the n X n matrix A(t) = [a;;(t)] and the correspondence t — A(f) is called
the matrix function A.
In order to avoid undue wordiness, we occasionally write “the function ¢(t)”
or ‘‘the function f(x, t)” for the respective phrases “the function @ with value ¢(t)
at ¢” or “the function f with value f(x, 7) at (x, ¢).”’, When this convention is likely
to cause confusion, however, we shall not employ it.
We call a vector or matrix function continuous if each of its components or
elements is a continuous function, and we differentiate or integrate a vector or matrix
function by performing the desired operation on each of its components or elements.
If, for example, a(f) = [27, sin 7, 1’ and f(x, y, z) = [xy, x* y+ zz. then
t ue 2
/ d 0 i“
; a(s) ds
d. — — ot Tees Fi
—— a(t) = cos t 3 ov {x.y %)
— = s ’

GPS)
oy ’ ’ ait se ice
0 Oz ~) 3 ;
2

The notion of vector function allows us, therefore, to write the differential
systems ([Link]) and (3.3) in the condensed form

x’ = AX, (LHC)
where x = [x1,...,%n]” and A = [a;;]. A solution of this equation is then a
vector-valued function @ such that $/(t) = A@(?) for all t. One can regard (LHC) as
one vector equation or as the system ([Link]) of n scalar equations. It will be con-
venient for our purposes to refer to it as a first order, n Xn, linear, homogeneous
differential equation with constant coefficients A = [a;;].
 Vector- and matrix-valued functions 53

In the subsequent study of (LHC), use is made of the product rules for dif-
ferentiating matrix and vector functions. These are much the same as the product
rule for scalar functions, but the factors must not be commuted.
Theorem 3.1. Let A and B denote differentiable matrix functions defined on an
interval a < t < w and let x denote a differentiable vector function defined on the
same interval. Then
i) (AMx()) = A'Ox() + A(x’, and
ii)(ABW)!
= A()BO) + ABC.
Proof. \f A(t) = [ai;(t)] and x(t) = [x1(t),..., Xn(O]", then the ith component of
A(t)x(t) is Dik=1 Gin(t)xz(t), and
LY aula) = > alex) + anlOxLO
k=1 k=1 k=1
by the usual product rule. But >°j_, a/,(t)x;,(t) is the ith component of A’(t)x(t)
and >o7_1 axz(t)xi(t) is the ith component of A(t)x’(t). This establishes part (i)
of the theorem. To prove part (ii), write B(t) = [bi(t),..., bn»(]. Then
(AMBOY = (AMD), ---, (AMb»D)]
= [A’(t)bi(t) + Abi (D), ... , A’Obn»(D) + AMD, (CD)
= [A’()bi(2), ..., A’ br()] +--+ * + (AMD, ..., Ab]
= A’(t)B(t) + AC)B'(A). ||

EXERCISE 3.4

1. Compute (d/dt)[t?e?'/2, te?', e?']".

2. Compute (0/0x)[x? + xy + y?, x? — y3,x — y]?.
3. Compute
Aen 3 peed i
OM, Ss = || CK
hese V1 + 52

4. Let (4 = [x(), y(d, z(]" be a 3-vector with ¢’(t) = 0 for all t. If o(0) = 0, find
x(t), y(t), and z(z).
5. Let o(1) = [x(), yO, z@]"_ be a 3-vector with ¢’(4) = ¢(t) for all ¢. If 6) =
[1, 3, 7]7, find x(4), yO), and z(z).
6. Find a 2-vector $(1) = [x(4), y()]* such that $’()) = —¢(d and $(0) = [1, 0].
7. Find a 2-vector x(‘) = [x1(4), x2()J" that satisfies the equation x(‘) = x(0) +
{5 x(s) ds, where x(0) = [3, 5]”. [Hint: Differentiate the equation by components. ]
8. Find a 2-vector x(t) = [x1(), x2()]" that satisfies the equation x(‘) = x(0) +
J (t — s)x(s) ds, where x(0) = [1, 0]”. [Hint: Differentiate the equation by components.]
9, Let x() = [x1(), x2()]" satisfy the equation
t

x(t) = x(0) + i.
3x1(s)1x9(s)
+ S260] ds,
where x(0) = [1, 1]7. Find the components of x(7¢).
 BES)
54 Matrix methods for linear equations with constant coefficients

10. Let (1) = [x(d, y(O, z(O]’ satisty the equation

t

o(t) = $(0) + | [2x(s) + y(s), 2y(s) + 2(s), 2(3)I" ds,

where (0) = [1, 1, 1]’. Find the components of $(7).

3.55 THE EXPONENTIAL MATRIX

Let us call a matrix of the form

Nae VG 0

B= ‘
Yn-1
0 Wee =

bidiagonal and designate the set of elements {¥1,...,Yn—1} a8 its superdiagonal.

The general simply coupled system ([Link]) of Section 3.1 has, in matrix notation,
precisely the form y’ = By, where B is an arbitrary bidiagonal matrix.
Two special types of bidiagonal matrices are of interest here. The first type is
called a Jordan block matrix. One says that B is a Jordan block (matrix) if itis 1 X 1
or if it has the form
DA line Ace 9
“lle (3.5a)
0 ee

all diagonal elements being equal and all superdiagonal elements having the common
value 1. The second type is merely called a Jordan matrix. One says that a bidiagonal
matrix Bis a Jordan matrix if its diagonal and superdiagonal elements may be grouped
into Jordan blocks without rearranging them, Jordan blocks with equal diagonal
elements being contiguous. The matrix B in Fig. 3.1 is a Jordan matrix. The dashed

Sl) 0 (On Ga OmOmCIEO

0, 3¥4°L080!0 -0 OM
0031000000
0.070
2 10) OnOmom0
B=|0-00 01m 110 0 O
Og 020g 0 Our GOOmO
OOO ORDO tania
0,.0%)050
20m 0r Omen
OOM 050) 0 OnOmOms
 The exponential matrix 55

lines indicate the required grouping into Jordan blocks. We shall denote a Jordan
matrix B with Jordan blocks B;, 1 < j < k, on its diagonal by either

By, 0
By

or B = diag [B,, Bo,..., By].

In Section 3.1, we showed how one can solve equations with bidiagonal, hence
Jordan, coefficient matrices by the method of integrating factors. A great many cases
can arise for the integrations when one uses the general notation

yil’ M4 Ae alone 0 al

eal adh | (cal (3.5b)

ass Yn—1

Va 0) AE, aS Va

since the integration procedure at each stage depends on which 7,’s equal zero and
which equal one.
There is a standard procedure which enables one to write from memory the
solutions of Eq. (3.5b) without actually performing the integrations. We now develop
this procedure.
Recall from calculus that if a is any number, the exponential function e®’ has a
Maclaurin series expansion
1+ at+ at?/2! +--++ a*t®/k! +--
which converges to it for all ¢. In the currently popular calculus texts, the exponential
function e” is defined to be the inverse of the logarithmic function t = a7} f u7! du.
It is nevertheless possible to define
fea}

Xi
PLL as a fe
—|
&

and deduce its properties with power series techniques.

In Section 1.2, we showed that x = e“c, is a solution of
x’ = at (355C)

for every constant c, and that every solution x = ¢(¢) of Eq. (3.5c) has the form
¢(t) = e“cy for some constant cy. Since matrices can be squared, cubed, and in
general raised to any power, it is not implausible that a similar result should hold for
the general equation
x’ = Ax, (LHC)

where A isan X n matrix and x is an v-vector.

 3.5
56 Matrix methods for linear equations with constant coefficients

We shall show in the next aie that the matrix power series
k
T+ a+ 5syoaoe ftp (3.5d)
converges to ann X n matrix for 4 t and that the series may be differentiated term
by term to obtain
o o k
ay ae Sy (3.5e)
The sum of the series (3.5d) is, not implausibly, denoted by e“'. Equation (3.5e) can
then be succinctly written as

A — | . = —

Thus
Foyle hipaa MO} ik
ett. itcom as oe hae onc Ee Ss agape

Onn ay Pe Opies
ee ea 0
[L 0 ee hOg Tee ee
E ert 0 T

0 ox
Example 2. To find e4’ when A = E \ compute

oe he 3 eee BS Niet
A =|; ol A kK eal ’ A =| * |:

Then
Nupe yl yk x ie f 3 ye-1k-1

Bete 5s KY (1)! 9) eon a ea

k=1 Nate oy kyk
are ! aE0 ay
Vee
ete aay
a Sf= olf ‘| !
Theorem 3.2. The series (3.5d) converges for all t and it may be differentiated
term by term.
Proof. The matrix series (3.5d) actually consists of n? power series displayed in a
Square array (see Examples 1 and 2). We shall use the comparison test to show that
 The exponential matrix 57

each of these n? component power series converges for all t. To do this, write A = [a;;]
and denote the sum )°7_; 50”, |a;;| of the absolute values of the elements of A
by |A|. Now the ijth element of A? is
(451415 + °° + + Gin@nj),
and summation of the inequality
|ai141; = Rae =F QinQn;| < la; | : lay, = Ig \din| 5 lay,;|

SSE oe ere OR ee are (ee)

for i, j = 1,..., m produces the inequality |A?| < |A|?. It follows by induction
that |A*| < |A|* for every k > 2. Thus no element of the matrix A*t*/k! has ab-
solute value larger than the number |A|*|t\*/k!. But >°fo |A|*t*/k! converges to
the number e'4'"'" for each t. Thus each series of elements in Eq. (3.5d) converges
for every ¢.
It is shown in the calculus that if }°f-o cz, and >°f_9 d; are convergent series,
then for any real numbers a and 8, the series )-f_9 (ac; + Bd,) converges, and its
sum is @ >of-0 cy + B Dik-0 dy. From this observation, one may conclude that
og k+l yk 2 gkyk—l
Veo) ey a)
k=0 KL k=o Kk! k= (k — 1)!
Because e“* is just the sum of n” convergent power series arranged in a square
array, each element of e“‘ has a derivative which may be calculated by differentiating
its series. Thus, for any constant vector c,

—eA'e = c= >) AS c= ||
dt dtjao k! ka1 (k — 1)!
Corollary 3.2. For any constant vector ¢, x = e*‘e is a solution of x' = Ax.

Do not misinterpret the corollary: it does not say that every solution x = @(t)
of x’ = Ax has the form ¢(t) = e“’c for some constant vector c. We shall show that
this is true for Jordan matrices in the next theorem, but a proof that it is true in
general will not be given until Section 3.7.
Anticipating this result, however, we comment here that the problem of solving
x’ = Ax is really the problem of finding e*’. If A is not a Jordan matrix, it is usually
not very easy to compute e“! using only the defining power series (3.5d). Series
substitution, however, works very nicely for the computation of e”' if B is a Jordan
matrix. We verified by substitution in Example 1 above that e?' = diag [e*"’,...,
e")] for each diagonal matrix B = diag [\1,..., An]. Similarly,
je}
ib fie hen eae
(k — 1)!
et Ol ae ew (3.5f)
<aieiy ah as t
Oe.0F SR: I
58 Matrix methods for linear equations with constant coefficients AYS)

for each k X k Jordan block matrix (3.5a). This was verified for k = 2 in Example 2.
The general verification goes through in the same way, but the computations are more
elaborate.
The next theorem allows us to immediately write the solutions of Eq. (3.5b)
without performing any integrations.
Theorem 3.3. Let B denote ann Xn Jordan matrix and let y = Y(t) denote an
arbitrary solution of
y’ = By. (3.5g)
Then (t) = e®'c for some constant vector c.
Proof. First assume that B = diag [\1,..., An] is a diagonal matrix. Then each
component y; of y satisfies

Vit) = Wit), Lsign. (3.5h)

Separating variables and integrating, we find that there is a constant c; such that
y(t) = edito,. Thus
V(t) eritc, edit 0 (651

yO={[ : |=] : |= ay > |= ele,

Yn(t) e : nog 0 ent Cn

where ¢ = [cy,.... Cal--

Next assume that Bisa k X k Jordan block matrix (3.5a). Then the components
Wy; of p satisfy
Vi(t) = AWi() + ve(4) ;
Yo(t) = AWo(t) + ¥3(0) ;

Oe BPE Oy
vi(t) = Av; (Z).
Applying the method of integrating factors, we find that there are constants
C1, 2 5 Ce such; that
Wilt) = ec, ;
Yr-1(t) = ev epes + tec, 3

- Mt Mt Mee ye
¥idt) = ec, + te C2 tasins eo (ea Ck.
In matrix form,
let) eS ie Sie
ie . C2
W(t) =e:
t
0 i sete Ck (3.5i)
But (3.5i) is precisely the equation y(t) = e?’c, where ec = [cy,..., Cl
 The exponential matrix 59

Finally assume that B = diag [B,,..., B;,..., By] is a general Jordan matrix.
Then the system y’ = By actually consists of k independent subsystems u; = B,u;,
where the components of y are the components of all the u;’s. Since each B; is a
Jordan block, uj = e?%'c; for some constant vector ¢;. Thus

ePit 0 Cy

W(t) = 1. a Ls
0 eBxt Ch

But e?' = diag [e?'’,..., e?*]. Thus y(t) = e?’c, where the components of ¢ are
the components of all the c;’s. ||

Example 3. If B = diag [B;, Bo, B3, B4] is the matrix in Fig. 3.1, then y’ = By
has the form
ae ah)
uj =]0 3 Llu, uy = 2uo,
Om OFS
1 x 1 O
i= [; {|U3, u=|0- 9 Illus,
0 O

where uy = Lyi, Yo. Val’, Ug = [ya]’, us = ys, yel’; us = [V7 ys; yo]’. Since

Ft a
eBit mr e2t 1 t : eB2t = en.

0 O 1
l ae)
eBst sale I eBat af nt F
Oe Oy 0 |

each solution of y’ = By has the form given in Fig. 3.2. ||

v1 Crumtic. enioe |Oe mnOF 0=20) o | a

Yo Owe Yer |cho SS ee 0 Co
Ys EO Oe ae
V4 Omen ON 202s co ORO ge nOberO™ —LO0P sii'cs
ys) =|0 0 On Mier te a0) 0 0-ailice
Ve 0 0 Ome Og LOme es 102 50 2-08 alle
y7 iy 1) 0 Quen Oie HOM eo eere tars 1er litc;,
Vs ONO 0 O MnO suheOe Orne™ ie" | 303
Vole .0 0 Ome 0 emo sO = Fe" ||co!
 Matrix methods for linear equations with constant coefficients 3.6
60

EXERCISE 3.5

1. Compute the first three terms of the series for estat

1 3 DP Om 4
=|; a and if hla |\Q #2 Zhe
if 2a

2. In parts (a) through (k), there is given a matrix B. Find the matrix e®' either by drawing
an appropriate analogy with Examples 1 and 2 or by substituting the matrix B into the
exponential power series.

270 170
9 |Tv 1 d)|0et7 0
»|0 3 »|Oni | 0m atone

oe it r 0 O xr 1 O
a) iO ze © f) tO am i g)|0
7 1
00fn OY OD as 0 On

aie OO 7 1 ONO
07 0 O x IO) ae OY ©)
Nowy et De Ont
000 - 000f
1.0) 0 ahti
hy 2 OF aie.
Dlo on 0 Sa aria
0 00cm Oy (0) ) a

fl) Be! = Sire. WH = hy, i! == We,

lo) oe? = See db gy, DY = Shy ae
@)) §2 a5 Bb WW ee Dp sb we = De
dG) 2a yy yee

4. Substitute the matrix (3.5a) into the exponential power series and derive the formula
GSt):
5. Show by power series substitution that if B= diag[B,,...,Bn,], then e?! =
diag [eP ite cea":
6. Let P denote a nonsingular matrix. Verify by power series substitution that P—1e4*P =
e(P-1AP)t.

3.6 DETERMINANTS

There is associated with each n X n matrix A a (real or complex) number det A

which is called the determinant of A. There are several equivalent definitions for
determinant. We shall adopt an inductive definition. If A = [a,;;] is ann Xn
matrix with n > 2, we shall write A;; for the n — 1 X n — 1 matrix obtained from
A by deleting its ith row and jth column.
 Determinants 61

Definition
1) If A = [ay,]isal X 1 matrix, we define det A = aj}.
li) Assume that determinant has been defined for n—1Xn-—1 matrices,
n > 2, and let A = [a;;] denote ann X n matrix. We define

det A= a33(—1)!*1 det Ay, + +++ +-ayn(—1)!4* det Ain. (3.6a)

Example. If A = [a,;] is a 2 X 2 matrix, then Ay; = [ao] and Aj. = [ao].
Thus det A = a11@22 — G21Qo9. If

411 @12 443

A=|d21 Qg2 23
431 432 433

isa 3 X 3 matrix, then

Ay = ie a, Aj2 = is al: and Ay3 = Be a |°

432 433 431 433 431 432

Thus

A229 a3 ao, a a a
Get A= ars det| <= 12 det : ae —- 413 det a 22 . ||
432 433 431 433 431 432

We shall assume the following theorems which are proved in linear algebra
courses.

Theorem 3.4 (Laplace). Let A denote ann X n matrix. Then

det A = a;,(—1)'*! det Ajy + ++ + ain(—1)'* det Ain (3.6b)

for eachi = 1,...,n. IfiA# k, then

ai3(—1)**! det Axi ae seas te din(—1)**™ det An = 0. (3.6c)

Theorem 3.5. If A and B aren X n matrices, then

det A? = det A (3.6d)
and
det AB = (det A)(det B). (3.6e)

The number det A;; is called the minor of the element a;;, and the signed minor
(—1)'*? det A;; is called the cofactor of the element a,;;._ Equation (3.6b) is called
the Laplace expansion for det A along the ith row.
We defined det A above as the value of the Laplace expansion along the first row.
Laplace’s theorem implies that any row could have been used. In fact, since the
columns of A are the rows of A’, it follows from Eq. (3.6d) that

det A= a,,(—1)14 det Ay; +++ + anj(—1)"*? det A,; (3.6f)

62 Matrix methods for linear equations with constant coefficients 3.6

for each j = 1,...,”. Equation (3.6f) is called the Laplace expansion for detA
along the jth column.
It follows easily from Laplace’s theorem that the determinant of a triangular
matrix (one with only zero elements above or below the principal diagonal) is the
product of the elements on the principal diagonal. If A is a large nontriangular
matrix, the evaluation of det A by Laplace’s theorem can be a very laborious task.
The difficulties can be mitigated somewhat, however, by taking advantage of zero
elements. For example,
6 16 —23 6 16
det} —1 —2 7T/= (-it?-3-det| _f | - 42:
Ora 3

If a matrix A has no zero element, one can construct a matrix B which has at least
one zero element and satisfies det B = det A. This is a consequence of the following
theorem.
Theorem 3.6. If the n X n matrix B is obtained from the matrix A by adding a
scalar multiple of a row (column) ofA to another row (column) of A, then det B =
det A.
Proof. We prove the assertion for rows. If A = [a,;] and
aii 400 Ain

‘Rie a1 + Ami gs) An + Amn

: : 9

An1 6 As Ann

then B = CA, where C is the matrix obtained from the identity matrix J by adding
the product of \ and its mth row to its kth row. It follows easily from Laplace’s
theorem that det C = 1. By Eq. G.6e),
det B = (det C)(det A) = det A. ||
At least one element in the kth row of B can be made equal to zero by choosing
\ properly. If, for example, a, # 0, the choice \ = —a;,1/am , reduces the first
element in the kth row to zero.
Example. Let

Multiply the first row of A by —3 and add the result to the second row of A to obtain

| ers a
B, =|0 —-1 —-3
PP yi TA
 General solutions for linear differential systems 63

The matrix B, has the same determinant as A. Now multiply the first row of
B,
by —i and add the result to the last row of B, to obtain

re ee.
By =|0 -1 —3
0-i 0
We have det By = detB, = det A. Thus

det A = 1 (= det]7) sa = —3j. ||

EXERCISE 3.6

Find det P for each of the following matrices P.

Le ly 3
PP = |)—2— 3 —1)- Prey ecea)
fa. 5 eRe)
6-2 1 [-2 1 3
SO P= \—7 1 1) 4, P=) 3-1] =8)2
3 0-1 Ex 0 3
In Problems 5 through 10 below, ) is a constant. The determinant detP will be a poly-
nomial in }. For what values of ) will det P = 0?
r= ht 2G 9 Ae ee rG 9
Wee Ppet|) 66 62515)\ = 3641 - Cre 4 29: 18 |-
2 0 Gk 42 eee l ey
26a ea 54 78 ey eS A pee 0
fee |S 6) \ = Of) |: 8, P= 10 eae
12 27 Ae =A a Sleek
pe 2 3 Dany 3 0
Peer | 24-4 =15 |: {One e eeset—e tOn |
ay TOs ray I aes 8 Ln ieee

3.7 GENERAL SOLUTIONS FOR LINEAR DIFFERENTIAL SYSTEMS

Having digressed a bit to review determinants in the last section, we now return to
the study of the differential equation
x’ = Ax (LHC)
that we began in Section 3.5.
The student has made changes of variables in many mathematical problems,
e.g., in evaluating integrals. A common method for changing variables in (LHC) is
to make the substitution x = Py, where P is a nonsingular matrix and y is the new
vector variable. This yields Py’ = APy, or
Ve (Ps APy. (3.7a)
64 Matrix methods for linear equations with constant coefficients Sul

Thus x = ¢(f) is a solution of (LHC) if and only if y = P~'$(t) is a solution of

Eq. (3.7a). This result can be applied to find all the solutions of (LHC). Its applica-
tion hinges on the following theorem which is due to Jordan.
Theorem 3.7. Let A denote any n X nmatrix. There exists a nonsingular matrix P
such that P~' AP is a Jordan matrix.

Now let us assume that we wish to find some solution x = ¢(t) of (LHC). We
make the change of variables x = Py, where P is the matrix in Jordan’s theorem.
The equivalent equation (3.7a) then has the form
View By, (3.7b)

where B = P~!AP is a Jordan matrix, and y = P~'¢(t) is one of its solutions. By

Theorem 3.3 in Section 3.5, however, this solution of Eq. (3.7b) has the form y =
e®'c, for some constant vector c,. Thus
x = o(t) = Py = Pe®c, (3.7c)
is the solution of (LHC) that we were seeking. We simplify this solution by writing
out its exponential series to find that

x = PU+ Br+ 4B7t? + ---Je,

= PIP-1P + (P7!AP)t + 4(P~1 AP)??? + + Jey
= P[P-1P + (P"1AP)t + 3(P~1A?P)t? + ++ Jey (3.7d)
= PP~ [I+ At + 4A72? + ---]Pe, = Ie4'(Pe)
=e;
where c = Pey. We have thus formulated a generalization of Theorem 3.3 for
arbitrary equations x’ = Ax.
Theorem 3.8. If x ¢(t) is a solution of x' = Ax, then $(t) = e4‘c for some
constant vector c¢.
The expression x = e“‘c, where c is an arbitrary constant vector, is called a
general solution for x’ = Ax.
Several comments are in order here. First, we have not mentioned how one might
go about computing the inverse P~' of a given nonsingular matrix P. A discussion
of the matter has been deferred until Section 3.11 because the only property of P~!
that we need here is P-1P = PP—! = J,
Second, we shall not give a proof of Theorem 3.7 here. Its proof is rather com-
plicated, and we shall leave it to the linear algebra courses in which it is usually
presented. In order that the matrix P not seem so much a deus ex machina, however,
we shall consider the problem of computing it for 2 X 2 matrices in Section 3.13.
Finally, we comment that Theorem 3.8 is the converse of Corollary 3.2 in Section
3.5. Thus the problem of solving x’ = Ax is really the problem of computing

A RTL At SAP EEA F (3.7e)

 General solutions for linear differential systems 65

Problem 1 of Exercise 3.5 was included to show that computation of e4¢ by power
series substitution can be very complicated. Also, it is difficult to obtain enough
cae directly from the power series (3.7e) to graph, say, the solutions x =
4te. One can nevertheless make very profitable use of Theorem 3.8. To do this, we
Hoe the notion of eigenvalue for the matrix A.
Let us subtract a variable \ from each diagonal element of a matrix A. The
matrix notation for this operation is A — \J. The determinant det (A — XJ) will
be an nth degree polynomial in the symbol \ (see Problems 5 through 10 of Ex-
ercise 3.6). The solutions of the polynomial equation

det (4 — 07) = 0 (3.7f)

are called eigenvalues or characteristic values for the matrix A; det (A — J) is called
the characteristic polynomial for A; and the equation itself is called the characteristic
equation for A. Eigenvalues are called proper values or latent roots in some books.
If \ = Xo occurs exactly m times as a solution of Eq. (3.7f), we shall call it an
m-fold eigenvalue or an eigenvalue of multiplicity m. An eigenvalue of multiplicity one
is called a simple eigenvalue.
Example 1. Let B denote a Jordan block matrix, say

Ge ee)

B=}; ~ a
0 neal

Then det (B — dJ) = (a — )*, and B has but one eigenvalue \ = a of multi-
plicity k. ||
Example 2. Let B = diag [B,,..., B,] be a Jordan matrix; then

B— >I= diag [B, — M,...,

By — A]
and
det (B — AJ) = det (B; — AI) ~~~ det (B; — vd).

Thus the eigenvalues of a Jordan matrix are precisely its diagonal elements. ||
The final theorem of this section provides a method for solving x’ = Ax, pro-
vided the eigenvalues of A are known. To prove it, we need a lemma.

Lemma 3.9. Let A, B, and P denote n X n matrices. IfP~'AP = B, then A and

B have precisely the same eigenvalues and these occur with precisely the same multi-
plicities.
Proof. Since P~*AP = B,
P(A — ADP = PO1AP — )PP >= B— 0.
By Theorem 3.5, Eq. (3.6e),
det P~1- det (A — XJ): det P = det (B — DJ).
 aa
66 Matrix methods for linear equations with constant coefficients

By the same theorem, it follows from the identity P~'P = J that (det P~*)(det P) =
Leas
det (ALT) det BN
and the lemma is proved. ||
Theorem 3.9. Suppose that the matrix A has eigenvalues \1,..., x with multi-
plicities my,...,m, respectively. If x = $(t) is any solution of (LHC) xt = Ax;
then
k

OL) = ery eter ey anes (3.7)

r=1

for some choice of constant vectors ¢;;.

Proof. Letx = 9(t) denote a solution of (LHC). By Theorem 3.8, there is a constant
vector ¢ such that @(t) = e4’c. Now let P be such that P~'AP = B is a Jordan
matrix. Then, reversing the computation (3.7d), we find that $(t) = Pe®‘ce,, where
c, = P~'c. Now each nonzero element of e?’ has the form r¥e**', where u > 0
is an integer and ), is an eigenvalue of B. Thus every element of the product matrix
Pe®' has the form p(t)e™, where p(t) is a polynomial in t. Performing the multiplica-
tion Pe®'c, we finally conclude that each component ¢,(f) of ¢(f) has the form
k

Pre tast to Sit Je, (3.7h)

r—1

where the Y,,;’s are scalar constants. This proves the theorem, for we have only to
define c,; = [Vrj1, Vrja) +++» Yrjn]’ to obtain the solution (3.7g). ||
Example 3. Let us find the solutions of the system
Bali DO meealiiexs
yi =|0 2 O}lyi-
Zz 0) Jt S§i\lLz

The eigenvalues of the coefficient matrix are \ = 2, \ = 2, \ = 3. In this case, the

solution (3.7g) assumes the specific form
x Ay B, Gy
Z\ = |Asle™ Bel ze? Cae
y Az Bs C3
Substituting this relation into the differential equation, we obtain
Ay By, B, Ch
2|A2|e* + 3| Co} e%
+ 2! Bo| te*+ |By| e”?
the Be B3 Ce
2. Ore 2° Oi PBs 20 SEG;
=|0 2 O}| Aste +/0 2 01) Bolte? +10 27 01|'C,)e%.
Oe seas On 1 BB O13 | ies
 General solutions for linear differential systems 67

Thus

2A, FP By 2B, NEG

2A, + By.|e*’ + |2B,| te?! + |3C,| e?

2A3 + B3 2B; 3C3

ZA yt 2B, + Bs 2G +5 Cz
= DAS eo 2 Be fee a 263 (eae
Aoi Ae B. + 3B3 Co+ 3C3
Equating coefficients of the corresponding exponentials yields the equations

PAY B= 2A7 As, 2B, = 2B, + B3, 3Cy = 20;

=p Cs,
2A> —+ Bs = 2A 5, 2Bo = 2Bo, 3C, = 2E5,

242 + B3 = Ao + 343, 2B3 = Bo + 3Bs, 3C3 = Co + 3Cz,

which we solve simultaneously to find that
A, 1s arbitrary, B, is arbitrary, C, is arbitrary,
A 2 B,, Bo = 0, Co 7a 0,

A ee Sa Bi, Bz = 0, C3 Cr .

Thus

x Ay B, Ci
y| =| —B,|e7?+] 0 |re?+1] 0 Je**. ||
Z B, 0 Ci

Example 4. By introducing the auxiliary variable y = x’, one can convert the damped
linear oscillator equation
x” + ax’ + w?x = 0 (3.71)

into the system

Hl 2 Ee =| Et (3.7))
The eigenvalues of the coefficient matrix are given by the quadratic formula
2s is VAD
Wi ee ae (3.7k)
They will be complex, real and equal, or real and distinct as 8 = w? — a?/4 is posi-
tive, zero, or negative. We consider here the cases 8 ~ O and leave the case B = 0
to the exercises.

cl-filesign anon
By Theorem 3.9, each solution of the system (3.7)) has the form

le MAT ee Bales:
68 Matrix methods for linear equations with constant coefficients 3E7/

We differentiate this equation and substitute the result into the system to obtain
\,A eu ao NeBie. => Age! a Byer?!

and
NeA se 2 + AoBoer?! = —(wA, + aA>)e™! = (w?B, + aBy)e",

The coefficients of corresponding exponentials must be equal. Thus

141 = Aa, \oB, = Be (3.7m)

yA, = —(w?A; + Aa), and \oBy = —(w?B, + aBz). (3.7n)
Substituting the values of \; and dg into Eqs. (3.7m) and (3.7n), we find that

Ap = —5 + Vo?/4—
ow?Ay and By = — 5 — Vo2/4 — o By.

Thus A, and B, are arbitrary and it follows from Eq. (3.71) that
x = Aye! + Byed2!, (3.70)

If 8 < 0, this solution may be written in the form

x =e 42 [Ayevislé + Boe Vial],

If 8B> 0, then dy, \2 = —(a/2) + iv/8 and
x ep Agene Bea va,
Euler’s identity e = cos @ + isin @ can now be applied to express x in terms of
sines and cosines:
x = e “#2, siny/B t + co cos v/B Zt], (3.7p)
where cy = A; + B, and cg + (A; — B,)i. The solutions (3.70) and (3.7p) are
precisely the solutions claimed for Eq. (3.7i)—(1.6a) in Section 1.6. ||

EXERCISE 3.7

1. Solve the differential equation x’ = Ax if

se eid, 8 9
wae E | ya=|_' At

eth alee 2

Se70u0
lei | PP
eee

Al
e) 4=|=1 20 f)A=]1 01
DO 2 Soe
TOY «4 =249 ==Ann
g) A= | — 24) 45. h)4 =|110) 941
cee) tt Ae oeel
 General solutions for linear differential systems 69

2. a) Show that each component of every solution of

‘|te 2 |X
y Ey
approaches zero ast > +0,
b) Let A be ann X n matrix. Under what circumstances will each component of every
solution of x’ = Ax approach zero as t > + ?
3. Find a solution of the equations

w | —-1 0 O-—I1I|iw
ait om OO My BON ix
yl2= OP OR= 2a eco liy
Z Jl (il WL || 4

that satisfies the boundary conditions

z0)=1, yO) =1, Od. and w(t) 0 ~~ as t— +o,

4. Solve the damped harmonic oscillator equation x’? + ax’ + w?x = 0 assuming that
a = "40)",
5. A flat plate, 1 m long and having a mass of 2 kg, rests on a large table. Resting atop
the plate, at its center, is a small cube with a mass of 1 kg. At time ¢ = 0, the plate is struck
on its left edge and the cube is struck on its right face, giving the plate an initial speed of
4 m/sec and the cube an initial speed of 1 m/sec. Thus the plate slides to the right, and the
cube moves on top of the plate to the left. Suppose that the frictional forces between the
cube and the plate and between the plate and the table top are proportional to velocity, the
proportionality constants being —4 newton: sec/m and —8 newton: sec/m respectively.
How far will the cube have moved relative to the table top when it has come to rest? Where
on top of the plate will the cube come to rest as its motion ceases?
6. In Section 1.7, we derived the system (1.7c) of differential equations which describes the
motion of two oscillating blocks. Find a general solution of the system (1.7c) if ki = 9,
kg = 12, m, = 3, and mz = 4. Suppose the block on the right is pushed 1 unit to the left
and is released with initial velocity zero. Find y() and w(t). How many seconds will elapse
before the system returns to its initial position for the first time?
7. Solve the initial value problem
R 1 1
ae ne = ee ee ee
which was derived for the parallel RLC circuit in Section 1.9, when
ayeL = 4, C= 10-° R = 1000, £ =*600.
bye "4,6C = 10-8)-R ='6005 Ei= 600.
el = 2.88, C = 10°, R= 1200, £ = 600.

8. Two blocks having unit mass are connected by a linear spring with stiffness coefficient
k = 2 and rest on a frictionless surface. The block on the right is held fixed, the block on
the left is moved 1 unit to the right. The blocks are released with zero initial velocity. If
70 Matrix methods for linear equations with constant coefficients 3.8

the spring is L > 1 units long when unstressed, how much does it stretch and compress as
the blocks oscillate? [Hint: Use the differential equations that you derived in Problem 1,
Exercise 1.7.]

9. Consider the system of blocks and springs described in Problem 2 of Exercise 1.7,
assuming that the masses of the block and the stiffness coefficients of the springs are equal
to one. The block on the left is moved 1 unit to the right and set into motion with unit
initial velocity while the block on the right is held fixed. Explain why neither block can
execute a periodic motion. What sort of initial conditions give rise to periodic motions?

3.8 THE NTH ORDER HOMOGENEOUS EQUATION

An equation of the form

x dx?) es apa ia. — O (LHC-n)

is called an nth order, linear, homogeneous differential equation with constant co-
efficients. The damped linear oscillator equation

x” + ax’ + w*x = 0, (3.8a)

which we solved in Example 4 of the last section, has this form. We solved the
oscillator equation there by converting it into a system and using Theorem 3.9. Since
it is computationally inefficient to go through this procedure for each specific equation
of the form (LHC-n), let us, once and for all, convert the general equation into a
system, describe its solutions, and thus provide ourselves with a direct solution
method.
Let x = y(t) denote some solution of (LHC-n) that we wish to compute. We
introduce the auxiliary variables x; = x, Xo = x’, ..., Xn = x‘*— and obtain
the equivalent system

0 1 0 0
0 0 1 0
ces : aA Ee hax (3.8b)
0 0 0 1
mh mi) Ie Bo 1G

where x = [x1,...,Xn]’. Notice that x = [y(t),...,~°-@J]? is a solution of

the system (3.8b). It is not hard to verify that

pQ) = AX + ayn” b+: >> +a ha, = 0 (3.8c)

is the characteristic equation for the coefficient matrix. Let \1,..., , denote its
solutions and let 7, .. . , m, denote their respective multiplicities. By Theorem 3.9,
the solution x = [¥(t),..., ~~ (]" has the form
k
\— s, (rn foeeet hi eg YS
7 |
 The nth order homogeneous equation 71

for some constant vectors ¢),;,Co,,..., Cred DLCUMY a eee ee Vd Fea oe wi)
denote the first components of ¢;,, Co;,..., Cm Then
k
(nan
p=
oe ty, et (3.8)
We state this result as a theorem.
Theorem 3.10. Each solution of
NP A Re eel. Quy 50 (LHC-n)
has the form (3.8d) for some choice of the constants ¥ 1, ... 5 Vm,r-
Let us henceforth call Eq. (3.8c) the characteristic equation for (LHC-n). The
polynomial p(\) may be found directly without converting the differential equation
into a system. One makes the substitution x = e* and obtains e*p(\) = 0. For
example, the characteristic polynomial for x’ + 2x” + 7x’ + x = 0 is p(d) =
3 + 2\2 + 7A + 1. It follows from this general observation that x = e is a
solution of (LHC-n) if p(\;) = 0, that is, if \; is any solution of its characteristic
equation.
Let us write the characteristic equation in factored form
pO) = = AM = AI = AN" = 0
and notice that
d’p men
2 Ar) = 9, sie a 9 Oe) = 0! (3.8e)
dN
Verbally this observation is expressed by saying that the multiple zeros of a poly-
nomial are also zeros of its derivatives. Keeping this in mind, we substitute
x = tte, 4 > 0, into the left side of (LHC-n) and obtain
BQ), __ ot a "PQ) | Zeer dp) 2
fe Ge = 1) a 2(u — 2)! dav? ees

SiS GET So panes

If we now set \ = ), in (3.8f) withO < uw < m, — 1, then it follows from Eqs. (3.8e)
that the entire expression equals zero. Thus all of the expressions
x= ert x = tet, x = t7e!, eo ee tte, (3.8g)
yr = 1,...,k define solutions of (LHC-n); equivalently, each term in the sum (3.8d)
defines a solution of (LHC-n). There are n of these solutions and we denote them for
the moment by $1,...,¢n- If 8i1,..., Bn are constants, then x = 6i9i(¢) + °° +
Bnon(t) also defines a solution since

Xa OY A be ay yx" +b ax
= Ss Bi ($9(t) + arg P(t) + +++ + Gn—164() + ano(d)) = 0.
j=
 Matrix methods for linear equations with constant coefficients 3.8
1p

Identifying the B’s with the 7’s in the sum (3.8d), we then see that it defines a solution
(LHC-n) no matter how the 7’s are chosen. We state this observation, which is the
converse of Theorem 3.10, as Theorem 3.11.

Theorem 3.11. The expression

k
a) Oe, ee (3.8h)
a

defines a solution of (LHC-n) for every choice of the constants V tr, Vary «+ + » Vm,r-
In view of Theorems 3.10 and 3.11, we shall call (3.8h) a general solution for
(LHC-n).
Example. Let us find a solution x = y(t) of x’” — 4x” + 5x’ — 2x = 0 that
satisfies ¥(0) = 1, y’(0) = 0, ¥’’(O) = 0.
The characteristic equation is \? — 4\? + 5A — 2 = 0, with solutions \ = 1,
) = 1, and \ = 2. Theorem 3.10 guarantees that the solution will be of the form

W(t) = (c1 + Cote’ + cze”,

where C1, C2 and c3 are constants to be determined. We calculate y(t) and y’(t) and
obtain the equations
Cy + cg3=1=y7(0),
Cy + co + 2c3 = 0 = y'(0), and
C1 + 2c2g + 4c3 = 0 = y’(0).
Thus ey =0Nc> ="—2..C-.— Ieeand

Y(t) = —2te’ + e?*. ||

EXERCISE 3.8

In each problem below, find a general solution of the indicated differential equation. If
initial conditions are given, find the solution that satisfies the conditions.
il, ge ae Abe? ae abe = (0), Daxtie-
2x ae =e (0)!
3, ae” Se Be 0), 4. x’ + 2ax' + a*x = 0.
Sax”? = 3x Bx = 0: 6. x” + 3x" + 3x’ +x = 0.
Tope ee ae Tx x= OF (0) = 1 xO) ale Ole
Soc — i te O erae(()) ae, 1 (Os On tO) poten
9, xO) - 2x" 4+ x% = 0; 10!" 4.2” = x! = 0,77 = 4.
Li XO?) 8x 12) 20,
12 x2) De 6x = 150 ON Aine lr noe
13. x% + 6x!" + 10x” + 6x’ + 9x = 0.
14. 25x’ — 15x’ — 4x’ = 0.
 The annihilation method 73

15, x9) + (4 + S5ix’” + (—2 + 20x" + (—24 + 20i)x’ — 24x = 0. [Hint: Try \ =
—2 twice.]
16. Solve the initial value problem LCI” + RCI’ + I = 0, 1(0) = 0, 1’(0) = E/R, which
was formulated in Section 1.8.
17. A very flexible cable of length L is held so that half its length hangs over the edge of a
high ledge. If the cable is released from rest, how much time will pass until it falls free from
the ledge?
18. A very flexible cable is draped over a very small pulley with two-thirds of its length on
one side. If the cable is released from rest, how much time will pass until it falls free?

3.9 THE ANNIHILATION METHOD

In the last section, we developed a method for solving the homogeneous linear equation
SO Nae sla So al (apenie meme he ool 0 (3.9a)

One factors its characteristic equation

Me Ge 2 ES An—1h + an = 0

to obtain
(A = hy)™ aeiceee (0 = yA ee = (0)

and writes a general solution

k

Se I a ee ae yee (3.9b)
rl!

In applications, there also occur differential equations of the form

OO LX ae ee lg ee | ax” (1), (3.9c)
where 6 is a continuous function that is not identically zero-valued. Such an equation
is called a nonhomogeneous linear equation with constant coefficients, and the func-
tion b is called a forcing function. The annihilation method is a technique for solving
Eq. (3.9c) when the value of 6 has the form

b(t) = A,b,(t) a a Ambit), (3.9d)

where each A; is a constant and each 6,(f) is given by one of the expressions
te, eeeet tk3 sin wt, t**coswt, t*e% sinwt, t**e%' cos wt. (3.9e)
It will be convenient in the discussion to refer to an expression ¢,f\(t) + °** +
Cmfm(t), Where the c;’s are constants, as a linear combination of fi(2),.-- NI EAUS
We may then say that the annihilation method is a technique for solving Eq. (3.9c)
when A(t) is a linear combination of the values (3.9e).
The technique is best explained by way of examples. Suppose, then, that we wish
to solve the equation x” + x = 7”. We differentiate three times and obtain
x™ + x’ = 0. Every solution of the first equation satisfies the second, but the
 Matrix methods for linear equations with constant coefficients 3.9
74

second equation has solutions that are not solutions of the first. By Theorem 3:10;
each solution of x® + x’’’ = 0 has the form
X= Cy Col 1 eal cue” iciee (3.9f)

We substitute this expression into x” + x = ¢? and obtain

de, — Cyc —cge ey Cot Cal cue ee ar
Thus, if Eq. (3.9f) is to satisfy the original equation, it is necessary and sufficient that
c3 = 1, co = 0, and cy = —2. It follows then that each solution of x’’ + x = 12
must be given by .
X= —2 +77? + ce’ exe” (3.92)
for some choice of the constants c4 and c;. Conversely, Eq. (3.9g) is a solution of
the differential equation for every choice of c4 and cs.
In general, the annihilation method consists of three steps: first, the nonhomo-
geneous is reduced to a homogeneous equation by differentiation, scalar multiplication,
and subtractions; next, a general solution of the resulting homogeneous equation is
found; finally, this general solution is substituted into the nonhomogeneous equation
and numerical values are determined for all but n of its constants. Theorems 3.12
and 3.13 below imply that this procedure will work in all specific cases, for they
imply that each nonhomogeneous equation whose forcing function (right side) is a
linear combination of the functions with values (3.9e) can always be reduced to a
homogeneous equation.
We motivate the method with several more examples which illustrate procedures
for reducing certain nonhomogeneous equations to homogeneous ones.
Example 1. If b(t) = c, a constant, differentiate Eq. (3.9c) once to obtain
x2Ftd fe Qe py s* 4 ga ane =O: ||

Example 2. If b(t) = e, differentiate Eq. (3.9c) once to obtain

xen, PF aie Sa beet Cee + ax’ = ver, (3.9h)

Now multiply Eq. (3.9c) by v and subtract the products from Eq. (3.9h) to obtain
Cae el Crlen On C Ren yee hee | |
Example 3. If b(t) = sin wf, differentiate Eq. (3.9c) twice to obtain
Hee ay) eat gag! oo ten) Se (3.91)
Multiply Eq. (3.9c) by w” and add the products to Eq. (3.9i) to obtain
x rt 2 ) ai axe ae (ao a Ox” a nee

PF (Qn 0" 2 dn 2)x "Oda 1x 1 @ Ga = Omen

Example 4. If b(t) = te” and n = 3, differentiate Eq. (3.9c) once to obtain
x™ 4 ayx! + agx"” + agx! = ev + ter, (3.9))
 The annihilation method 75

Now multiply Eq. (3.9c) by Y and subtract the products from Eq. (3.9j) to obtain

et (dye) Xe te (Ge — 8Va)’ (ag — Yag)x’ — agYx = eY, (3.9k)

Differentiation of Eq. (3.9k) gives
x (a; — Y)x™ + (ay — Yay)x’” + G3 — Yaz)x” = agyx’ = Yee GN)
Now multiply Eq. (3.9k) by Y and subtract the products from Eq. (3.91) to obtain
Be (Gye )X ag — ay YX! (as — 2a) Yds)”
+ (—2a3Y + Y7a2)x’' + a3y*x = 0. ||
Now let us verify that Eq. (3.9c) can always be reduced to a homogeneous equa-
tion if the values of 6 are given by (3.9e). We first observe that each of the expressions
displayed there can be obtained from
Bare ae a Bape om. (3.9m)

by specializing the values of By, Bo, k, a, and w. This is a consequence of Euler’s

identity
ett — e% cosy + ie* sin y.
It is therefore sufficient to validate the annihilation method for the case

b(t) = Byte! + +++ + B,thre’, (3.9n)

The next theorem further simplifies the problem.

Theorem 3.12 (The Superposition Principle). Let ay,...,@, and cy,..., Cr be

constants and let b,,..., 6b, be continuous functions. If the functions ~; are solutions
of the equations
Or ge ele pen oct Aa, (t)
forj=1,...,7, then ¥@) = cil) + °°+ + crf,(2) is a solution of
x7 ax) st a ax + a, (x) = bd),
where b(t) = c1b,(t) +--+ - +. ¢,b,(0).
Proof. We verify the assertion by computation.

POO) + a) Ho + ra + ad
= j=xyCVS) F aWF PD Fo + an) + ans)
r

= 2) eb) = 5). I
2

It follows from the superposition principle that if the annihilation method is

valid for Eq. (3.9c) with b(t) = Br"e7’, it is also valid when O(7) has the more general
form (3.9n).
 Matrix methods for linear equations with constant coefficients 3.9
76

Theorem 3.13. Let» > 0 bean integer and let8and denote any complex numbers.
The equation
x gx eg, a td, = oree (3.90)

can be reduced to a homogeneous equation of order (n + k + 1) by (k + 1) dif-

ferentiations, each followed by a suitable scalar multiplication and subtraction.

Proof. Differentiate Eq. (3.90) to obtain

OED eG + ots Ean aix 4 a,x = Bue e+ bie | Gp)
Then multiply Eq. (3.90) by 7 and subtract the products from Eq. (3.9p) to obtain

x@TD 4 @, — Yx™ 4 +++ + Gy, — Van_1)x’ — VOnx = Bpt ev,

The power of ¢ on the right is thereby reduced by one. Repeating the process (k — 1)
times, we obtain an equation of the form

Seo ICTS Spee NE Faroe | Og ene teen nC

We proceed next as we did in Example 2 and finally obtain a homogeneous
equation. ||

Example 5. Let us find a solution x = y(t) of the equation

x” —x=t+e! (3.9q)
that satisfies the conditions Y(0) = 0 and y’/(0) = 0. We differentiate the equation
and eliminate e’ from the result ofthe differentiation to obtain x/” — x” — x/+x=
1 — t. After two more differentiations, we find that x — x% — x/” + x’ = 0.
All solutions of this equation have the form
xX = Cy + Cot + cge’ + cate’ + cs5e—, (3.9r)
by Theorem 3.10.
The solutions of Eq. (3.9q) are also of the form (3.9r), but not all the constants
are arbitrary. To discover which constants are arbitrary and which are not, we make
the substitution (3.9r) in Eq. (3.9q) and obtain the relation

x” — x = —cy — te + 2cge’ = t+ '-.

Equating coefficients, we find that c; = 0,cy = —1,andc4 = 4. Thus the solutions

of Eq. (3.9q) are given by

x = —tt+ Fre’ + czet + cs5e', (3.9s)

where cz and cs are arbitrary. If there exists a solution x = y(t) with ¥(0) = 0 and
¥'(0) = 0, it will be given by (3.9s). We solve the algebraic equations 0 = (0) =
cz + cs, 0 = ¥'(0) = —3 + cs — eg to find that cz = 4 and c,; = —4. The
required solution is therefore y(t) = —1 + $te'+ 4sinhz. ||
 Linear independence of vectors 77

EXERCISE 3.9

Solve the following equations.

De Axess pl: 2, X04
xX eed,
8) oe de ea See An xx
t= De.
By Oc) SE se aaa 6. x” +x = et,
Tex aX = tet, 8.x” +x = 2? + 3et,
Dx -- xX = sin ZF. IO, a segs = Sse
ieee — 7 SinoTF, 1D, se So se er Se
1330 4=x = e' sin 7: 14. x’ + x = e* sin 3¢.
Se Xx = fe" Sin f; 16. x’ +x =f+ e'sint.
De at-Ox. Ox! = Axe eo, 18.0 x" 6x" Ox ax = fe=*.
19. x. 6x!” + Ox’ -- 4x = 3te—* + Je*.
20. Consider a cylindrical object of radius r, height A, and density p which is partially sub-
merged in a vat of liquid having density w > p. If the object is depressed and released at
time ¢ = 0, the depth x of the bottom of the cylinder satisfies the differential equation
ptr2hx" = prr2hg — wrr2gx, where g is the acceleration due to gravity. Suppose x(0) =
a > Oand x’(0) = 0. Find x. If the object were a buoy on a lake, would it be possible to
tell its weight with the aid of binoculars and a stop watch?
21. If a coil of inductance L, a capacitor of capacitance C, and a light bulb of resistance R
are connected in series across a voltage source of magnitude Eo sinwf, the differential
equation for the current J in the circuit is

LI’ + RI’ + 1/C = wEo cos wt.

What values ofZ and C will cause the bulb to burn most brightly ? Could a dimmer be built
using these principles ?
22. Consider a metal block of mass m which is constrained to move on a horizontal friction-
less surface under the influence of a linear spring of stiffness coefficient k and a time de-
pendent force of magnitude Fo sin wt. The differential equation for the displacement x from
equilibrium is mx’ + kx = Fo sinwf. Assume that w = /k/m and solve for x. What
happens as tf > + © regardless of the values of x(0) and x’(0)?
23. If friction is considered in Problem 22, the appropriate differential equation for the
displacement x is mx’ + ax’ + kx = Fosinwt, a > 0. What value of w will give the
largest amplitude to the displacement?

3.10 LINEAR INDEPENDENCE OF VECTORS

A set {a;,..., 8m} of n-vectors is called /inearly dependent if there are constants
C1,..+, Cm, not all of which are zero, such that

cay = koi + Cman = 0.

A set of vectors that is not linearly dependent is called linearly independent.

78 Matrix methods for linear equations with constant coefficients 3.10

In Exercise 3.2, Problem 6, the reader showed that

I 2 3
eedel eee
7 8 9
The indicated vectors are therefore linearly dependent.
It is proved below that every set of n + 1 or more n-vectors is linearly dependent.
It is useful to know (and the reader is asked to verify the assertion in the exercise
below) that every set of vectors containing the zero vector is linearly dependent.
The definition of linear dependence above can be phrased more compactly by
introducing a little more terminology: If a;,..., a, are n-vectors, an expression of
the form c,a; + °°* + Cmam, Where c1,...,Cm are constants, will be called a
linear combination of the vectors. Thus {a;,...,4,} is linearly dependent if and
only if there is a nontrivial linear combination of its elements that equals the zero
vector.
The following short theorems are useful in the discussion of linear independence.

Theorem 3.14. If {a,,...,@m} is a linearly independent set of n-vectors, then no

nonempty subset of {a1,..., Am} is linearly dependent.
Proof. Assume for contradiction that the set {a;,..., az}, 1 << k < m, is linearly
dependent. Then there are constants c,,..., c,—not all zero—such that cya; + -:-
+ c;,a; = 0. Thus

Cia a cpap Oa eit 02, = 0,

and it follows that {a;,...,a,,} is linearly dependent, a contradiction. ||

Theorem 3.15. The set {a,,..., Am} of n-vectors is linearly dependent if and only
if at least one element equals a linear combination of the others.
Proof. Assume {a1,...,4m} is linearly dependent. Then there are constants
C1,---,Cn—not all zero—such that cya; +--+ + cman = 0. We may assume
without loss of generality that c; # 0. Then
a; = Se CS 9.9.2) é
Cy C1 m

Conversely, if aj, say, can be expressed as

CW tied St Witenes,
for any constants ky,..., km, then

ay hoes ek ae
and {a1,...,@,} is linearly dependent. ||
Let us say that a set {y1,..., y,} of n-vectors spans V,,(€)—or is a spanning set
for V,,(C)—if every n-vector can be expressed as a linear combination of its elements.
 Linear independence of vectors 79

A spanning set that is linearly independent is called a basis for V,(@). If u; is the
n-vector with ith component equal to one and all other components equal to zero,
then it is easy to verify that {u,,...,u,} is a basis for V,(€). We shall call it the
standard basis.
A fundamental result from linear algebra is the following theorem.
Theorem 3.16. Any set {¥1,.--,Yn41} of n+ 1 n-vectors is linearly dependent.

Proof. Suppose for contradiction that {y1,..., Yn, Yn41} is a linearly independent
set of n-vectors and let X = {x),...,Xn} denote the standard basis for V,,(C).
(Note that it is not assumed that x; = u;.)
Since X spans V,,(@), the set {y1, x1, ..., Xn} is linearly dependent. Thus there
exist constants bj, c1,..., Cn—not all zero—such that

biyy, tex, +-°+ + c2xn = 0. (3.10)

It is not true that cy = --- = c, = 0, for this would imply that y; = 0 and would
contradict the linear independence of {y1,...,¥m41}. We may therefore assume
that the standard basis vectors have been arranged in such a way that c; ~ 0. Then
x, can be expressed as a linear combination of the vectors y;, X2,..., Xn by merely
solving Eq. (3.10). Since {x;,...,Xn} spans V,(C), it follows that Y, =
Ry XG ex coesealso, —Ihis-being the case, the set 4yi,[Link],- 9) Xp) 1S
linearly dependent. Repeating the argument n-times, we find that Y, = {y1,..., Yn}
spans V,,(C). This contradicts the independence of {y1,..., ¥n41} since there must
now exist constants c;,..., Cn such that yniy = C1y1 +°°* + CnYn- ||

Theorem 3.16 implies, in less abstract terms, that every system of algebraic
equations
a11X1 9° OimXen = 0,

An1X1 + °:° + QnmXm = 0

has a nonzero solution provided m > n + 1.

We close the discussion by remarking that the notion of linear independence
applies to a set {a;,...,@m} of vectors. Reference to the vectors aj,..., Am as
being linearly independent rather than to the set of vectors is a fairly widespread
convention, and we shall occasionally use it.

EXERCISE 3.10

Which of the following sets of vectors are linearly dependent? (Note that this is really a
question of solving linear algebraic equations.)

18 6 9 18 6 9
1. |—66], |—18/, |—36 2. |—48], |—12], |—18
2 0 9 42 12 18
80 Matrix methods for linear equations with constant coefficients 3.11

ese camer bo
‘Labbe BERE
3.11 LINEAR ALGEBRAIC EQUATIONS

If A is ann Xn matrix and b is a n-vector, the equation Ax = b is called homo-

geneous if b = 0 and nonhomogeneous otherwise.
The homogeneous equation Ax = 0 always has at least one solution: the vector
x = 0. Later, we shall need to know conditions under which Ax = 0 has a nonzero
solution. The basic facts are contained in Theorem 3.17.
Theorem 3.17. If any one of the following statements is true, then so are all
the others.
i) The columns of A are linearly independent.
ii) Ax = 0 has only the solution x = 0.
iii) A~' exists.
iv) det A ¥ 0.

Proof. It is convenient to show first that statements (i) and (ii) are equivalent. Let
the columns a,,...,a, of A be linearly independent and suppose that x =
[c1,..-,¢n/’ is a solution of Ax = 0. Then cia; +---+ c,a, = 0.” Since
@1,.--, a, are linearly independent, c¢, = -*-= c¢, = 0. Thus x = OF "if con-
versely, Ax = 0 has only the solution x = 0, it follows from the definition of linear
independence that the columns aj,...,a, of A are linearly independent.
We complete the proof by showing that statement (ii) implies (iii), that (iii)
implies (iv), and that (iv) implies (ii).
Suppose that statement (ii) is true. Let u; (¢ = 1,...,m) denote the standard
basis vectors defined in Section 3.10. Then a,;, the ith column of A, satisfies a; = Au;.
By Theorem 3.16, the vectors u;, a;,...,@, are linearly dependent (i = 1,..., 7).
Thus, there exist constants d;;, ... , din Such that

uy = dja, oF ar ae Gan

Let x; = d;,u; + ---+d;,u; for i = 1,...,”. Then u; = Ax;. Were there to

exist a vector z; such that u; = Az,, then 0 = A(x; — z,). Since statement (ii) is
true, it would follow that x; = z,. Thus there exists
a unique matrix X = [xy,..., Xn]
such that AX = J. To show that X is an inverse for A, it is also necessary to show
that XA = J. This is done by observing that A(XYA — I+ X) = J. Since there
is only one matrix X such that AX = IJ, X = (XA —I+4 X). The last equation
reduces to XA = J. Therefore statement (iii) is true.
 Linear algebraic equations 81

Next, assume that (iii) is true. Then 4A~! = I. By Theorem 3.5, 1 = det J =
det (AA~') = det A- det A~!, and det A ~ 0. Thus statement (iv) is true.
Finally, suppose that (iv) is true. For each = 1,..., 1, let
b; = [(—1)!*/ det Ay;,..., (—1)"*# det Ans]
and let B have b; for its jth row. By Eqs. (3.6b), (3.6c), and (3.6f), BA = (det A)I.
Now suppose Ax = 0. Then 0 = BO = BAx = (det A)x. Since det A ¥ 0,
x = 0. Thus statement (ii) is true. ||
The matrix B in the last theorem is the transpose of the matrix of cofactors of A,
that is, if B = (6,;), then
b;; = (—1)'*) det Aj:.
If det A # 0, then A~' = B/(det A). This equation is a useful representation
of A~' and provides an algorithm for its computation.
It follows from Theorem 3.17 that the truth of any of the following contrapositive
statements implies the truth of all the others.
i’) The columns of A are linearly dependent.
il’) Ax = 0 has a nonzero solution.
iii’) A is singular (A~! does not exist).
iv’) det A = 0.
The nonhomogeneous equation Ax = b may have no solution or infinitely many
solutions. Notice that x = 0 is not a solution since b ¥ 0 for a nonhomogeneous
equation.
Theorem 3.18. The nonhomogeneous equation Ax = b has a unique solution if
and only ifA~* exists.
Proof. If A~' exists, then x = A~'b is the unique solution.
Suppose, conversely, that A~1 does not exist. We show that if Ax = b has a
solution x = d, then it is not unique. Let x = ce be any nonzero solution of Ax = 0.
Then A(c + d) = Ac + Ad = 0+ Ad =D. Thus x = c+d is a solution of
Ax = b different from x = d. ||
To actually solve a system Ax = b, it is most efficient, for the problems in this
book, to use some method of elimination of variables such as reduction to triangular
form.

EXERCISE 3.11
Find all vectors x satisfying Ax = b given the following values of A and b.

18 6 9 21 6 9 =
1. A = | —66 —21 —36|, b = O. 2. A=|— —18|], b=
—12 42
2 0 3 42 12 18 UW
Col
eSel
82 Matrix methods for linear equations with constant coefficients 3.12

Dye yal es 6 = 6n— 40

3. A=] —28 —63 —91, b =} —7]. A SAC 5 LOM Tee den}
WG Pi BS9) 3 —4 -3 -1

—6 -—4 0 —2 G7 3
yy Zhe |) iO 7 illo i= Ble 6. b0.
A =|—24 —7 —15|,=
—4 —3 —] —] —4 -—2 0)

7. Calculate P—! for the matrix P in Problem 1 of Exercise 3.6.

8. Calculate P~1 for the matrix P in Problem 2 of Exercise 3.6.

9. Calculate P—! for the matrix P in Problem 3 of Exercise 3.6.

10. Calculate P—! for the matrix P in Problem 4 of Exercise 3.6.

3.12 EIGENVECTORS
Let us recall that the eigenvalues of an n X n matrix A are those values \y,..., xz
of \ that satisfy the characteristic equation
det (A — AJ) = (A — Ay)™ + A — A)” = 0. (3.12a)

Thus a given matrix A will always have at least one eigenvalue, and it can have n
distinct eigenvalues. Now let \; denote an eigenvalue of an n X n matrix A. Since
det (A — i,J) = 0, it follows from parts (ii) and (iv) of Theorem 3.17 that there is
at least one nonzero vector x such that
(A= x= 0. (3.12b)
Such a vector is called an eigenvector of A corresponding to the eigenvalue \;. Eigen-
vectors are sometimes called characteristic vectors or proper vectors.

Example. The eigenvalues of the matrix ; | are the values of \ for which

3— F
det | 3 1 2 vanishes. These are the zeros of the polynomial (3 — A)(1 — A) —

15, namely \; = 6 and \y = —2. The corresponding eigenvectors are obtained by

solving the equations

ae 5 2} = (2) at ee mae ea Om
3 1 — 6|| xe 0 3 1+2]||x2} [0

The solutions are not unique. The vectors

Pi = al?
§ and p2 = 6 i
=

are eigenvectors for any nonzero choice of the scalars a and b. Notice that Pp; and p>
are then linearly independent. ||
 Eigenvectors 83

For a general n X n matrix A, any scalar multiple of an eigenvector is again an

eigenvector. If the matrix A has the property that every set of two or more elgen-
vectors is linearly dependent, we shall abuse the language a bit and say that A has
but one linearly independent eigenvector. With this convention, we may say that a
matrix A has from one to n linearly independent eigenvectors.
If an eigenvalue \; of A has multiplicity m,, then there will be at least one, and
no more than m; corresponding linearly independent eigenvectors.
Theorem 3.19. If \1,..., x are distinct eigenvalues of the matrix A and if
Pi,---, Px are (respectively) corresponding eigenvectors, then {p,,..., Px} is linearly
independent.
Proof. Observe that the set containing only the vector p, is certainly linearly in-
dependent. Suppose for contradiction that the set {p;,...,pj;—1i} is linearly
independent, but that the set {p;,..., p,} is linearly dependent for some j(2 < j < k).
By Theorem 3.15, there are constants cy,...,cj;—1 such that pj = cypy + -°* +
c;—1Ppj—1. If this equation is multiplied from the left by (A — \,/), the equation
0 =, Qi Pi Se 2 * Cai Ap — Adj
results. Since {p,, ..., P;—1} is linearly independent and the eigenvalues are distinct,
it follows thatc,; = --: = c;_; = 0. But this implies that p; = 0,acontradiction. ||
Two matrices A and B are said to be similar if there is a nonsingular matrix P
such that P~'AP = B. In this case, the matrix P is said to represent a similarity
transformation and it is called a similarity matrix. A matrix D is called diagonal if
each element not on its principal diagonal is zero. We shall occasionally denote
diagonal matrices by displaying only the diagonal elements, for example, D =
diag [\1, \2,.--,5 An]. The matrix A is called diagonalizable if it is similar to a diagonal
matrix.
Theorem 3.20. Ann X n matrix A is diagonalizable if and only if it has n linearly
independent eigenvectors.
Proof. Let P = [pi,.--,Pn] denote a matrix with n linearly independent eigen-
vectors of A for columns and let \1, . . . , An denote the corresponding (not necessarily
distinct) eigenvalues. Then
AP = A[pi, $44 -» Pn] = [Ap1, as -» Apr] 3 [\iP1, Te a5 XnDal = jc DE

where D = diag [\j,..., An]. It follows that P~'AP = D.

If, conversely, A is diagonalizable, there exists a nonsingular matrix P =
[pi,--->Pn] and a diagonal matrix D = diag [\1,...,\n] such that P~’AP = D.
Then
[Api,..-, APn] a A[Di3 sau Pal AP PD [\1P1, +++» AnPnl.

Thus p; is an eigenvector ofA, and); is the corresponding eigenvalue (j = 1,..., 7).

Since P is nonsingular, it follows from Theorem 3.17 that pi,..., Pn are linearly
independent. ||
84 Matrix methods for linear equations with constant coefficients Ay I)

We saw, in Lemma 3.9, that similar matrices have precisely the same character-
istic polynomial, hence the same eigenvalues. The eigenvectors of similar matrices
are related by the following theorem.

Theorem 3.21. Let P~'1AP = B. Then x is an eigenvector of A corresponding to

an eigenvalue ), if and only if y = P—'x is an eigenvector of B corresponding to dj.
Proof
(B — ,Dy = (P~1AP — djI)P7x = P71(AP — 2,P)P"x
= P-1(A —A,I)(PPax) =P (A el)

EXERCISE 3.12
In each of the problems below, there is given a matrix. In Problems 1 and 2, there are also
given two vectors. Check that the vectors are eigenvectors and find the corresponding
eigenvalues. In the other problems, find the eigenvectors and eigenvalues of the given
matrix.

1 ol 1 1 25 <3 1 3
7 OR a fea OM pe | yFame 5)ale I fd

aeQO #4 < |
1
5. EE2 |
24 6 9 26. 334 “S73
Oni 665-150 36 ir ne -18 8. |—28 —61 —91
oS TES) 42 12 te ee eee
Sey at 2 m1
Oni Ome Onn -nA = Lid hay (1
Annet 2
3 1 oar6)
12. Let A; denote an eigenvalue of the matrix
0 1 0 0
Pa hao 1 0
we SS UR Rag

Compute the eigenvectors of A corresponding to );.

13. Which of the matrices in Problems 3 through 11 can be diagonalized?

3.13 JORDAN CANONICAL FORM FOR 2 X 2 MATRICES

In Section 3.7, we solved the equation x’ = Ax by applying a theorem of Jordan
(Theorem 3.7) which we restate here for convenience: If A is any n X n matrix, there
is a nonsingular matrix P such that P~'AP = B is a Jordan matrix. It is customary
to call B a Jordan canonical form for A, and one says that the matrix P reduces A to
Jordan canonical form. Here we wish to examine the structure of P.
 Jordan canonical form for 2 < 2 matrices 85

Let us assume that A is given and that we have in hand a matrix P that reduces A
to a Jordan canonical form B. Since A and B are then similar, Lemma 3.9 implies
that they have precisely the same eigenvalues. But the eigenvalues of B equal its
diagonal elements since Jordan matrices are triangular. Thus the eigenvalues of A
form the principal diagonal of B.
Now let us write P = [pi,...,p,] and B = diag [B;,..., B,,], where Bis a
v; X v; Jordan block matrix
ee! 0
B; = i1
0 dj
for j= 1,...,m. It may happen that some of the Jordan blocks have equal diagonal
elements; thus the listing \1,...,A;,..., Am Of eigenvalues does not imply that
they are distinct.
We write the equation P~'AP = B in the form AP = PB and consider, for
notational simplicity, only the first y; columns of P:
AP Alpi,
a. Piss0 ol
Axl 0! 0
1|
|
0 o) ale eT
= EB = Api, os sys 07] Me EF Paeree tag = ue

0) orale By

=i Di, eAdD2 Piss 2 AtPoy te Dp, ais as 1:

Thus the first vy; columns of P satisfy the equations

Ail py 0, (A = Al)ps = Pi. 5-5 (A Ai), = py,i-» 2134)

The other columns of P must, of course, satisfy a similar sequence of equations.
Notice that the columns pj, p,,41,---» Py,+---4+»,-;+1 are linearly independent
eigenvectors of P.
Now let us suppose that A is a given X n matrix and that we do not know, but
wish to find, a matrix P that reduces A to Jordan form. A plausible method for
determining P is the following: (i) the eigenvalues and eigenvectors of A are com-
puted; (ii) sequences of equations of the form
Me 0A i ex, (A Xan Xn hv ons G13)
are solved for eacli eigenvalue \ of A; and (iii) the columns of P are then chosen
from the vectors computed in step (ii).
We shall now prove that this procedure can always be implemented for 2 X 2
matrices A. Let A be given and let \;, \2 denote the eigenvalues of A. There are two
cases:
86 Matrix methods for linear equations with constant coefficients Ay |)

CASE I. The matrix A does not have two linear linearly independent eigenvectors,
that is, every eigenvector of A is a scalar multiple of some one fixed eigenvector p;
and the eigenvalues \; and \g are equal.
CASE II. The matrix A has two linearly independent eigenvectors pj, po, correspond-
ing respectively to the eigenvalues )j, Xo.

Consider Case I. We begin to solve the sequence of equations (3.13b) with

\ = \, and find that x; must be an eigenvector of A. We therefore take x; = pj.
Then we attempt to solve the second equation, which now has the form (A — \jJ)x2 =
p,. An apparent difficulty is that (A — \,/) is singular; thus it is conceivable that
the equation might not have a solution for x2. Even if it did have a solution, Theorem
3.18 guarantees that it would not be unique. The corollary to the next theorem shows
that there is in fact a solution for xo.

Theorem 3.22. Assume that the Case I hypotheses hold and let q be any nonzero
vector which is not an eigenvector of A, that is, q ~ 0 andq # cp, for any constant c.
Define y by y = (A — XA )q. Then y is an eigenvector of A; hence y = cp, for some
constant c # 0.

Proof. Let pz ~ 0 be any vector which is linearly independent of p,. Then, by

Theorem 3.16, we can express q as a linear combination q = c,p; + Cope with
C2 # 0. To prove the theorem, we must show that (4 — \yJ)y = 0. We do this
by writing y = d,p; + d2pe and showing that d2 = 0. Notice that we may also
write y in the form

y = (A — Aql)q = (A — Aq] )(CoPi + CoP2) = (A — AyD )coPo.

Thus

(A — dqT)cops = dip; + dopo

or, equivalently,
d

Denote the matrix on the left by B and note that it is nonsingular. Equation (3.13c)
then tells us the pp = d,B7'p,/co. Thus
(be
A1P2 = MB p= “1B'Ap,

a:
Ch
2
[a+ (+2)C2 r/o,
i dz\ 3-1
. {t+ (u+2)p Jp.
acd ds
ah Pi + (x Se 22)Po,
 Jordan canonical form for 2 < 2 matrices 87

and it follows that

dy
e Dis; Cees
es Po = 0. (3.13d)

Since p2 was chosen linearly independent of p,, Eq. (3.13d) implies that d> = 0. |
Example 1. Every eigenvector of

is a scalar multiple of p, = [l, 1]”, and the eigenvalues are \; = \y = 3. Let

q = [91,92] be any nonzero vector which is not an eigenvector of A, that is, which
does not satisfy gq, = qo.

(A 37 ||
7 2
omenEERE$(G5— le
qi)
M —% 214 3(92 — 41)
The last vector is an eigenvector of A, just as Theorem 3.22 predicts, since it can be
expressed in the form $(¢2 — q1)p1._ ||
Corollary 3.22. Under the hypotheses of Case I, the equation (A — 4I)x2 = py
has a solution.
Proof. Let q ¥ 0’be any vector that is linearly independent of p;. By Theorem 3.22,
y = (A — X,/)q is an eigenvector of A, that is, there exists a constant c ¥ 0 such
that (A — AJ)q = cp;. Define xg = q/c. Then (A — A,J)xe = pi, and the
equation is solved. ||
Continuing the study of Case I, we rename the vector x of the corollary py and
summarize the results of the discussion with the equations

(A = Ait)pi = 0, (A — dil )Po = Pi. (3.13e)

Since p; and py» are linearly independent by construction, it follows from Theorem
3.17 that P = [p1, pe] is nonsingular. Thus the equation
pees |
AP = A[pi, po] = [Ap1, Apo] = [AiP1, A1P2 + Pi] = ‘abs |

implies that
a Poee ee BE
P AP =|": A, (3.13f)

Now consider Case II]. We set \ = \,; and x; = p, in Eq. (3.13b) and attempt
to solve the equation (A — \4)x2 = pi. We shall see from Corollary 3.23 below
that this cannot be done.

Theorem 3.23. Assume that the Case II hypotheses hold and let q be any nonzero
vector that is not an eigenvector of A corresponding to \;. Define y = (A — d1)q.
Then y is an eigenvector of A corresponding to dj, 1 # J.
88 Matrix methods for linear equations with constant coefficients 3.13

Proof. Since p; and pg are linearly independent, we may express q in the form
q = CiPi1 + CoP2, where co # 0. Then

(A — AWDy = (A — Ag)(A — Aa = (A — AIA — Af )(C1P1 + CoP2)

= ¢€1(A — dol)(A — dql)pi + €2(A — AVI)(A — Aol )Pe2
= €1(A — dol)0 + c2(A — AxJ)0
=o
Example 2. The eigenvalues of

4-[
are \; = 6 and \» = —2. Choose any nonzero vector x = [x1, Xe)’ and form
the product
y = (A — 61)x = [—3x1 + 5x2, 3x1 — 5x2)’.
If 3x, = 5x2, then y = O and x is an eigenvector of A corresponding to \; = 6.
If 3x, ¥ 5x9, then y ¥ 0 and
Ay = [6x1 — 10xo, —6x, + 10xg]’ = —2y.
Thus y is an eigenvector of A corresponding to \y = —2 just as Theorem 3.23
predicts. ||
Corollary 3.23. Under the hypotheses of Case II, the equations (A — ;I)x2 = pi
(i = 1, 2) do not have solutions.
Proof. Suppose (A — d,J)xe = p; has a solution x» = q. Since (A — X;x)q is an
eigenvector of A corresponding to ij, j ¥ i, (A — \;x)q = cp;. Thus p; = cp;,
with c ¥ 0, a contradiction. ||
Continuing the study of Case II, we note that the eigenvectors p; and pz satisfy
the equations
(A — dyJ)p; = 0, (A — Aol )p2 = 0.
If P = [pi, po], then
AP = [Api, Apo] = [\1P1, AoPo]
= [P1, Po]: diag [\1, \2] = P diag [\q, do).
Since p; and pg are linearly independent, P is nonsingular and P~'AP = diag [\4, do].
Thus A can be reduced to Jordan form under the hypotheses of Case IJ. There is
one further question to be answered for Case II: Would it be possible to find a matrix
QO = [q1, qe], different from P, such that
ae
ovao=[%) 4.)
21 At ae

The answer is no, for the columns of Q would have to satisfy the equations

(A — dil)qi = 0 (3.13g)
 Jordan canonical form for 2 < 2 matrices 89

and
(A — Aol )qo = qu. (3.13h)
To satisfy Eq. (3.13g), q, would have to be an eigenvector of 4 corresponding to ).
According to Corollary 3.23, Eq. (3.13h) would then have no solution.

EXERCISE 3.13
1. For each of the following matrices A, find a matrix P that reduces A to Jordan form.

= (4 3 es 8 9
a) 4= (5 Al ya=|_! a

y= 92 ‘98 VAL

2. Let A bea 2 X 2 matrix and let c denote any constant vector. Show that there exist
constants k; and ke such that

a) eA'e = kypie*!! + kopoer2!

if A has two linearly independent eigenvectors p; and pe corresponding, respectively, to
eigenvalues \1 and \2; and

b) ete = kypye™! + ko(tp1 + peje"

if every eigenvector of A is a scalar multiple of py and )j is the associated eigenvalue.
[Hint: Use Eq. (3.7d) in the proof of Theorem 3.8.]
3. Explain why the matrix
Zain O
A= c3 1
O @ 4

is not in Jordan form. Solve the system x’ = Ax anyway. [Hint: Use the method of in-
tegrating factors.]
4. Construct a 3 X 3 matrix A that is not in Jordan canonical form, and a matrix P
such that

5. Solve sequences of equations such as (3.13b) to find, for each matrix A below, a matrix P
that reduces A to Jordan canonical form.
CHAPTER 4

THE THEORY OF LINEAR

DIFFERENTIAL EQUATIONS

4.1 PRELIMINARIES

The student has learned in Chapter 3 how to reduce the problem of solving either
the equation (LHC) x’ = Ax or the equation
yi? gx VELOSO x a0 (LHC-n)
to an algebraic problem: that of finding the zeros of the associated characteristic
polynomial and, in the case of (LHC), solving certain systems of linear algebraic
equations. Let us now consider the equations
x’ = A(t)x + b(t) (L)
and
x™ 4 a(x P + +++ + an_ix’ + anx = BC), (L-n)
where A, b, a},..., 4m and b are continuous functions on an interval a < t < w.
These are called linear differential equations with continuous coefficients. If b(t) = 0,
we say that the equation (L) is homogeneous and denote it by (LH) x’ = A(f)x.
Otherwise (L) is called nonhomogeneous. Analogous terminology holds for (L-n),
where
xo AO)xe7P Se = 5 ap ax! a, =O. (LH-n)

Given a nonhomogeneous equation (L) or (L-n), the associated homogeneous equa-

tion (LH) or (LH-n) is called its complementary equation.
If an element of A(t) in (LH) or one of the coefficients a;(t) in (LH-n) is a non-
constant-valued function of ¢, one cannot usually hope to express solutions as simple
combinations of elementary functions, as is the case with (LHC) and (LHC-n). It
is nevertheless of interest to study the relations between the solutions of a given
linear differential equation, for these relations allow one, in some cases, to deduce
the qualitative properties of individual solutions.

90
 Initial value problems 91

It is sufficient for most of our purposes to prove theorems for (L) and specialize
them to (L-n). A connection between the two equations is established by setting
Mie x, Xo = ss ty = Xe Then (1-n) has the form
x, + GE) pact ke An—1(t)xX2 + a,(t)x, = b(t),
and one may write it in the form x’ = A(t)x + b(t), where
Xen e ele BG) =s10F OSB)
and
0 lL earaclc 0

LOS ye
—a,(t) ... —a,(t)
If x = y(¢) is a solution of (L-n), the corresponding solution of the associated system
is therefore x = [¥-(1), V'(), ...,¥° PO).

4.2 INITIAL VALUE PROBLEMS

The existence question for solutions of x’ = Ax was handled in Chapter 3 by

our indicating a method for computing the solutions. In the next section, we shall
show that (L) x’ = A(t)x + b(t) has solutions even though one might not be able
to express them as elementary functions. It is assumed throughout the discussion
that A and b are continuous on an intervala < t < w.
A solution of (L) is to be a function @ such that @’(t) = A(t)d(t) + b(t) for
a<t< w. Now let xg be a given n-vector and let ty be a point of (a, w). One is
said to have posed an initial value problem for (L) when he asks for a solution x =
¢(t) of (L) that satisfies the initial condition (to) = xo. An initial value problem
for (L-n) may be posed directly in the form
Bae =F aire Creech acre An(t)x ar b(t),
C= Chk = IC, when t = to,
where c,,..., Cy are given numbers and fo is a given point of (a, w).
The student has solved initial value problems in previous exercises. For example,
Problem 11 of Exercise 3.12 is to find a solution to the initial value problem
sep 24 6 9}| x x(0) —5
y| =|—42 —9 —18]| y], y(O)| =| 5].
Z ADR Digee2 bi|\| z z(0) 3
In Example 5 of Section 3.9, we saw that there is one and only one function that
satisfies the initial value problem
x’ —-x=t+ e, x(O)s=20 ox(0) — 0;
namely, x = —t + 4te' + $sinht.
 4.2
92 The theory of linear differential equations

Since each solution of (L) is the solution to some initial value problem for (L)
and vice versa, one can discuss the existence of solutions to (L) by showing that the
initial value problem
x’ = A(t)x + D(t), xX = Xo when t = 0, (IVP)

where fo in (a, w) and x9 are given, has solutions. Interestingly enough, (IVP) cannot
have more than one solution. To show this, it is convenient to introduce the notions
of rectangular norm and equivalent integral equation.
If x = [xq,..-,Xn]’ is an n-vector, we define the rectangular norm |-| by |x| =
|x,| +--+: + |x,|. This norm has a number of properties that follow directly from
the definition of absolute value:
i) |x| > 0, and |x| = 0 if and only if x = 0.
ii) [x + y| < |x| + lyl.
iii) [cx] = |c|- |x| for all numbers c.
iv) If @ is a continuous vector function, then

ifo(s) ds) < / |o(s)| ds} .

v) |Ax| < |A|- |x|, where |A| denotes the sum of the absolute values of the
elements of A.
If x = ¢(?) is a solution of (IVP), then, by integrating, one verifies that x = (tf)
is a solution of
t

X= Xo + / {A(s)x + b(s)} ds. (1)

0

Conversely, one differentiates (I) to show that each of its solutions is a solution of
(IVP). In this sense, (I) is an equivalent integral equation for (IVP).

Theorem 4.1. Suppose A and b are continuous on a < t < w and let to in (a, w)
and Xo be given. Then (IVP) has no more than one solution.

Proof. Suppose that x = ¢(f) and x = (‘) are solutions to (IVP). Then, from (1),
t

g(t) — Y(t) = / A(s)i¢(s) — ¥(s)} ds. (4.2a)

0

Let to < t< c < w, where c is fixed. Since |A(s)| is a continuous real-valued func-
tion, it attains a maximum ||A|| on [fo, c]. When properties (iv) and (v) of the
rectangular norm are applied to Eq. (4.2a), the inequality

lst) — WO) < Hail | lp(s) — Y(s)| de (4.2b)

results. Denote the integral in inequality (4.2b) by u(t). Then w(t) < ||A| u(t).
 Initial value problems 93

This inequality can be solved by the methods used in Sections 1.2 and 1.3. Multi-
plication by the integrating factor e—''4'!" gives rise to the inequality
d(u(t)e|'4""")
/dt < 0,
which implies that u(t)e—!'4''* is nonincreasing. Thus
wpe AaNe < u(t oe |!Allto = 0,

and u(t) = ie lo(s) — ¥(s)| dt = 0 for tp) < t <c. Inequality (4.2b) implies, in
turn, that |(¢) — ¥(t)| = Oforto < t < c. Similarly, one shows that |6(1) — ¥(1)| =
Ofora<a<tKX fo. Since a and ¢ are arbitrary, (1) = Y(t) fora <t<w. ||

EXERCISE 4.2
In Problems 1 and 2, identify the point x = (x1, x2) in the x1x2-plane with the 2-vector
x = [x1,x2].
1. How would the points a = (ai, a2) and b = (61, b2) have to be situated in the x1xe-
plane in order that |a + b| = [a] + |b]?
2. Given that x = (x1, x2), a = (1,0), and b = (—1, 0), describe the set of points x in
the x1x2-plane that satisfy |x — a] + |x — b] = 4.
3. Overestimate |Ax| by |A] - |x| if

a) 4=( bee teats =| _1 eee (oe

4 0 1
C)A S12 1-E7 _ 0 x= [1,7
7]?.
0 310 a

4. Let x and y be continuous vector functions on the interval a < ¢ < b and let ||x||,
llyll, [x + y|| denote the maximum values of |x(2)|, |y@|, [x@ + y(d| on [a, 6]. Show
that |[x + y|| < |[x|| + llyl.
5. Let A and x be continuous ona < ¢t < band let ||A||, ||x||, ||Ax|| denote the maximum
values of |A(d)|, |x()|, and |A()x(d| on [a, 5]. Show that
a) ||AMx(|| < |All - x@I, b) [A(x < JA] - [IxlL,
c) |A(@x(| < |All > [xl d) ||Ax|| < |All - [Ixll-
6. Let A and B denote n X n matrices. Show that
a) |A + Bl < |A| + |B, b) |AB| < |A|- |B].
7. Let A and B denote continuous n X n matrix functions on the interval a < t < band
let ||A|| and ||B|| denote the maximum values of |A(f)| and |B()| on [a, 5]. Show that
a) ||A + Bll < ||Al| + ||Bll, b) ||ABl| < |All - ||Bll.
8. Solve the following inequalities for x in terms of f.
AD) oe! SPA. I) oe S Ane
(O) od << (Oye. d) x’ < a(t)x, a continuous.
94 The theory of linear differential equations 4.3

9. Let u denote a positive-valued, continuous function of t. Prove Gronwall’s inequality:

u(t) < A + B fj u(s) ds implies u(t) < Ae®', where A and B > 0 are constants.
10. Find the solution x = $(f) of the initial value problem x” + 2x’ + 2x =0, x =
—(1/\/2) and x’ = 0 when ¢ = 0. What is the first time T > 0 at which ¢(7) = 0? Is
O(L) = OF
11. Let ¢ and T have the same values that they have in Problem 10. Solve the initial value
problem
TW se oy =O, vel and y =o) when je—ele

12. The blocks in Fig. 4.1 are not connected. Each block has a mass of kg. The spring
is linear, 10 m long when unstressed, and has stiffness coefficient 2 newtons/m. The fric-
tional force between each block and the table is proportional to velocity. The constant of
proportionality for the left-hand block is 1/2 and that for the right-hand block is 3/2.
Suppose the right-hand block is pushed 1/./2 meters to the left and is released with zero
initial velocity. How far will the right-hand block travel during the first second after the
blocks separate?

Win r
Figure 4.1

13. Prove each of the identities below by considering suitable initial value problems for the
equation x’ + x = 0.
a) sin (—/) = —sint. b) cos (—2) = cost.
c) sin (tf + a) = sinfcosa + cosf sina, a any constant.
d) cos (t + a) = costcosa # sint¢ sina, @ any constant.
e) sin (t + 27) = sint. f) cos (t + 27) = cost.
g) d(cos t)/dt = —sint. h) d(sin t)/dt = cost.
i) sin? ¢+ cos? t = 1.

4.3 THE EXISTENCE OF SOLUTIONS

One can show that (IVP) has a solution x = ¢(t) by constructing a solution to (I).
The construction is formally simple: one defines ¢0(t) = xo and
t

One tye (|
{Aol + bY} ds, KS 0. (4.3a)
Then, if
Jim es = $0 exists (4.3b)
and if t

lim
k—-+Loo
/: {A()ox(s) + b(s)} ds = / {A(s)9(s) + b(s)} ds, (4.3c)
it follows that x = 9(f) is the desired solution.
 The existence of solutions 95

The sequence {¢;} defined by Eq. (4.3a) is sometimes called a sequence of

successive approximations to ¢, and this method of constructing a solution ¢ is called
the method of Picard.
Example 1. We calculate successive approximations for the solution x = g(t) of
x’ = atx that satisfies ¢(0) = 1.
The integral equation (1) takes the form
t

oO) = 1+ / 2s $(s) ds
0

in this case. The successive approximations are

dM=1, mM=1+, $2.) =14+ 2? 4 24/2,
and in general
ont) =Lttt
+ 0"/n.
The reader should observe that limy_,..¢n(t) = e” is the solution to the initial
problem. ||
Example 2. To calculate successive approximations for the solution x = y(t) of
x’’ + x = 0 which satisfies the conditions ¥(0) = 0, ¥’/(0) = 1, we first convert the
differential equation into a system by defining y = x’. Then

bI-L4 ob:
and the corresponding integral equation is

PRs WT Jol-1ee ollyel™

le
t

be
We indicate the calculation of the successive approximations without further
comment: ;
0 0 O10 t
bolt) = Hi oi) = 9]a iLie 4Re i Ar
t
0 Oe nits 5 t ;
b2(t) = Hq: i Ae 4He Fi, . 2,
0 01 s _[t- ee
t

Gal) = H = I,EE 4[ - TA ag Fe— 2/2!

In general,
t— 19/31 + +++ + (-D Pk — uF kezal,
goi(t) = ie 77/2! a ee Cpir/Ck)!

and
t — 19/3! +--+ + (-1)
tt 1/Qk — )
g21r—1(t) = ooo Koel,
1 — 27/2! + +++ + (-1)%0?"-7/Qk — 2)!
96 The theory of linear differential equations 4.3

Notice that ;
iene [ee‘|
kw cos t
The student knows from his work in Chapters 1, 2, and 3 that the first component
Y(t) = sin ¢ is the solution to the initial value problem as posed. ||
To verify that the convergence requirements (4.3b) and (4.3c) are satisfied when
A and b are continuous, we apply several standard results from advanced calculus
which we combine into one lemma without proof. The statement in the lemma that
> ¢=0 f, converges means that each component series converges.

Lemma 4.2. Let {f),| denote a sequence of vector-valued functions that are defined
and continuous on an interval a < t < c. Suppose that there exists a sequence of
numbers M;, such that |f,(t)| < M, for all t in [a,c] and all integers k > 0. If the
series > ~~09 M;, converges, then the series >\y—o f(t) converges for every t in [a, c].
If f(t) = doe=0 f(t), then f is continuous on [a, c] and
t t

/ f(s)ds = >> | f(s) ds

to k=0 ¥ to

for every choice of to and t in [a, c].

Example 3. Let f,(t) = (cos 2rkt)/(4k? — 1), O< t< 1, k>1. It is easy to
check that 4k? — 1 > k? for k > 1. Thus |f,(2)| < 1/k? = M, for 0.47.1
and all k > 1. By Lemma 4.2, )\~o (cos 2rkt)/(4k” — 1) converges to a con-
tinuous limit function f on [0, 1]. Further,

cos — = sin 2rkt

as
EAS) dies »af4k2 — S= 2 ake =
for any ¢ in (0, 1].
Example 4. Find power series expansions, valid for |r| < 1, for In(1 + #) and
Lane eh:
Notice that

d
an el +
ae!
t) =
=e
fey, = Far (—1)'t
kik .
if |t| <jile

and
d <1 < We 2k :
pe ta FaSle= 2A it ba] coe.

since the indicated series are geometric series with common ratios —t and —t2
respectively. Let us assume that |t| < r < 1, where r > 0 may be as near one as
desired, and let us define
(Aye
f.(4) = (e€ 1\rre*
;
 The existence of solutions 97

Then If,(0)| = \z|* a lz|2" x rk 25 pek eh re 2 Pr) = 27", and ein = i f a

By Lemma 4.2, the two series above may be integrated term by term over subintervals
of-lt| <7. Thus
o : peti

In (i =) = - (eat) eae oar

and
n~ A

Since r may be taken as near one as desired, the series are in fact convergent for
lel <1. |
To apply Lemma 4.2 to the sequence {¢,} of successive approximations defined
by Eq. (4.3a), we introduce the difference function A, = $%41 — 9» for all integers
k>0. Then y41 = Xo + Sy 9Ay and limy ,4« v(t) exists if and only if the
series > ¥~9 A(t) converges.
Theorem 4.2. Assume that A and b are continuous on a < t < w and let to in
(a, w) and Xv be given. Then (IVP) has a solution.
Proof. Leta <a< to, t< c < wand denote the maxima of |A(t)| and |b()| on
[a, c] by ||A]|| and ||b|| respectively. It follows from Eq. (4.3a) that

Ai(t) = o1(t) — Xo = / {A(s)xo + b(s)} ds (4.3d)

and
t t

Ariilt) = / A(s){ox(s) — o¢—1(5)} ds = / A(s)A;(s) ds, Keaet, (4.3e)

0 0

If M =||A|| - ||xol| + ||b||, then Eq. (4.3d) and the properties of |:| imply that
|Ai(@|| < M|t — tol. (4.3f)
Similarly, Eq. (4.3e) implies that

acco <| ftvatl-lavcolas| < ae-ttall-| fJs — solas| = Hel

In general,
C1 C k+1
=< mu lll eee, a0): (4.3g)
[Axii()| < M- (Al Leste
CADE (kK + 1)!
We may as well assume that ||A|| # 0, for otherwise the successive approximations
converge trivially. Then, if M; denotes the expression on the right of (4.3g),

: a
R+V
||A]|
iA aye
aly M
pllAli(c—ay
a aan) All 1).
98 The theory of linear differential equations 4.3

By Lemma 4.2, (fo Ax(t) converges and limy + $n(t)= o(t) exists, that is,
(4.3b) is established.
Now let & = Ag, — b for each integer kK> 0. Then -

/ lim i {eos ae y
{A(s)on(s) + b(s)} ds = if = (E-41(s) — Ex(s)){ ds. (4.3h)
tg N+

ACs) A;,(s), and it follows from

A(s)ibk4i(s) — bx(s)y =
eh ss
inequality (4.3g) SO
that

4@ ay] < Mall=eat

A A ee ee
Since )\f-9 M Mire —at= Me'lAll—®) Lemma 4.2 may be applied to Eq.
(4.3h), with f,; = £41 — &, to obtain
t t is t
| {A(s)9(s) + b(s)} ds = | Eo(s) ds + & / {Er41(5) — Ex(s)} ds
t t

= J £o(s) ds + lim he{A(s)$n41(s)

to N+
+ b(s)} ds
J to

—< / £o(s) ds.

This establishes (4.3c). Since [a, c] was an arbitrary closed and bounded subinterval
of (a, w), (4.3b) and (4.3c) hold for every ¢ in (a, w) and the proof is complete. ||

EXERCISE 4.3
1. Compute the first three successive approximations for solutions to the following initial
value problems. Try to guess the limit function that the approximations will ultimately
approach.
ae — aux (O)y—s le where astasconstant:
b) x’ = f(x, x(0) = 1, where f is continuous. [Hint: Let F(t) = f§ f(s) ds.]
QO) eS e— il, e@) = © Gar = oe Il ax) = th
ee ar Ories XO) he ls Fe te ‘ =
2 3 7 k 4HE Ex + H eels ie

ofty =o0 3}Lyl’

alo} Gol-[)
Lx® 1 »bT-[5
y sib)
=9) 15 \[y1™ bel
yO) 4(il
2. Find constants M;, such that >°?_, M; converges and lf.()| < M, for a<t< bb,
given that
sin kt
k2
a) f,(t) =
cos kt
k2
 The first order scalar equation 99

e
k2
b) iO) = Pale ra),
b= il
k2
c:
pil
c) HOPS ft} a=0, b=}:
k+1

sin?" t
d) f.(0) = Ped ’ a=0, b= a

eee

3. Let ao be any v-vector and define a; = (1/k)az_1 for k > 1. How can one tell that the
series >? 9a, converges?
4. Let Yo()= 0 and define ¥x(4) = fo x16) + 1] ds fork > 1. If dD) = v() —
Yx—1(4), show that >°%_1 (4) converges to a limit ¢(/ for all ¢ and compute fj 6(s) ds.
What initial value problem does ¢ satisfy at t = 0?
5. Consider the initial value problem

ET-[o lbh bol- Lh

y 0 2}Ly!’ Ly») 1
Use a method analogous to that used in Problem 4 to show that successive approximations
converge to its solution on an interval of the form |t| < constant.
6. Solve the initial value problem

fx) = + x =.0, x =0- and. x =1 when ¢=0

by making the change of variables x(t) = y(s), s = In¢f. Does your result contradict
Theorem 4.2?

4.4 THE FIRST ORDER SCALAR EQUATION

The reader has learned to solve a special case of the scalar equation
x’ = a(t)x + b(t) (L-1)

in Section 1.3. In general, one computes the solution that satisfies the initial condition
x(to) = Xo by first forming the integrating factor exp (—fi a(s) ds]. When (L-1)
is multiplied by this factor, the result is
nit t t

x'(t) exp Ey a(s) a — exp Pedi a(s) as- a(t)x(t) = exp lel a(s) as- b(t).
' , ; (4.4)
100 The theory of linear differential equations 4.4

The left side of Eq. (4.4) is the derivative of the product x(t) exp (—f; a(s) ds].
Thus (4.4) can be written as
t

é («0exp ek a(s) as|)= exp Ey a(s) as|- b(t),

and an integration gives

NE LeXO | a(s) asXo + exp || a(s) as|. / exp aL a(s) as|- b(u) du.

The reader may check that the integrating factor is a solution of the complementary
equation
x’ = alh)x (LH-1)
and that

x=" exp if.a(s) is|- [ox ek a(s) as|-Bw) ds

is a particular solution of the nonhomogeneous equation; namely, the unique solution

that equals zero when ¢ = fo.
In order to motivate the subsequent sections of this chapter, we shall give here a
discussion of (L-1) that can be generalized to linear systems (L) x’ = A(t)x + b(2).
For the scalar equation (L-1), some of the discussion will appear almost trivial. The
reasons for its content and arrangement will become apparent later. It is assumed
that a and b are continuous on an intervala < t < w.
1. Theorem 4.2 guarantees that (LH-1) has solutions. Let x = y(t) bea solution
and suppose that the only constant c for which ¥(t)c = 0 on a < t < w is the con-
stant c = 0. Then x = y(f) is called a fundamental solution.
Exercise: Find a fundamental solution of x’ = x/(1 — 17, —o <t< 1.

2. A solution x = y(t) of (LH-1) is a fundamental solution if and only if Y(t) ¥ 0

for every f in (a, w).
Exercise: Prove this.

3. No fundamental solution of (LH-1) can have both zero and nonzero values; thus
a solution whose value equals zero at one point is the identically zero solution.
Proof. Let x = Y(t) be a solution of (LH-1) and let fo in (a, w) be arbitrary. Then
¥'(t) = a(@)Y(.) and, solving with an integrating factor, one finds that y(t) =
W(t) exp Le a(s) ds]. Thus ¥(¢o) = 0 implies y(t) = 0 for every tin (a, w). ||
4. Let x = y(t) be a fundamental solution of (LH-1) and let x = ¢(t) be any other
solution. Then there is a constant c such that $(t) = ¥(f)c. Conversely, (t)c is a
solution of (LH-1) for every value of the constant c.
 The first order scalar equation 101

Proof. Since ¥(t) ¥ 0 for a < t < w, the quotient (t)/y(2) is differentiable. In
fact,
(O4O) = WOO — ¢OWO)V720
= Wao — oa(NV))/¥7(D = 0.
Thus $(t)/y(t) = constant. Verification of the converse assertion is left to the
reader. ||
If x = ¥(f) is a fundamental solution of (LH-1), the expression x = ¥(t)c, where
c denotes an arbitrary constant, is called a general solution of the equation since it
is a solution for every value of c and every solution of (LH-1) can be obtained from
it by giving c an appropriate numerical value.
Exercise: Let x = $(t) and x = y(t) satisfy the initial value problems x’ = (cos f£)x,
x = 1 when ¢ = O and x’ = (cos ft)x, x = 1 when t = 7/4, respectively. Find a
constant c such that ¢(t) = Y(t)c.
5. The difference of two solutions for (L-1) is itself a solution for (LH-1).

Proof. If ¢’ = a(t)p + b(t) and y’ = a(t)y + b(2), then ( — yy = alg — ¥). ||

This observation allows one to find all the solutions for (L-1) if he can find just
one (a particular) solution. In fact, let x = y¥(f)c denote a general solution of (LH-1)
x’ = a(t)x and let ¢, be any function that satisfies (L-1). Then any solution of (L-1)
can be obtained from x = ¥(t)c + ¢,(t), which is called a general solution, by
giving c the correct numerical value.
Exercise: Find a particular solution ¢, of the equation x’ = x + | by inspection.
Write a general solution and find the solution x = ¢(f) that satisfies the initial
condition ¢(0) = 3.
6. Suppose x = y(t) is a fundamental solution of (LH-1). Then

p(t) = vo [ [(s)/W(s)] ds, to in (a@,o),

is a particular solution of (L-1).

Proof ;

op(t) = (0) / [b(s)/Y(s)] ds + b()

t

= au | [b(s)/W(s)] ds + b(t) = A(Db,(4) + BC). ||

This technique for finding a particular solution is called variation of parameters.

Exercise: Find a particular solution of x’ = x + e’sin ¢.

102 The theory of linear differential equations 4.4

Example 1. Find a general solution for the equation x’ = 2tx + e”. A fundamental
solution to the complementary equation is exp dh 2s ds = eae A particular solu-
tion is ¢,(t) = e” fi (e)(e”) ds = te”. Thus a general solution is x = ce” +
fe" |||
Example 2. Let 5 denote the sawtooth function of Fig. 4.2. The equation x’ =
—x + b(t) can have at most one solution of period 7. Suppose, on the contrary,
that there are two solutions x = ¢(t) and x = y(t) having period T. Then x =
¢(t) — Y(t) is a solution of x’ = —x. Thus there is a constant c such that ¢(t) —
y(t) = ec. If c #0, then ¢(t) — ¥(t) ~ +o as t— —o. This means that at
least one of the functions is not periodic (a continuous periodic function is bounded).
Thus c = O and ¢(t) = Y(t). This result can also be established by solving the equa-
tion if one cares to perform the variation-of-parameters algorithm on the sawtooth
function. A better way of solving x’ = —x + b(t) is given in the next chapter. ||

b(t)

Figure 4.2

EXERCISE 4.4

1. Solve the following injtial value problems. If you can find particular solutions by
inspection, do so.
a) ee es OM Wheni aa): b) x’ = 2tx -- t, x = 0 when ¢ = 0:
c) x =x+e,x=1whent=0. d) x = (@/d + 38, x = 0 when ¢ = 1.
eC) x =" ek cosit, x — 1) when — 0!
f) x’ = —(tanax + sint, x = 1 when ¢ = O.
g) x’ = (sec¢)x + sect + tant, x = 0 when ¢ = 7/4.
h) x’ = 21x + 4,x =1 whent=0. i) x’ = x + cost, x = 1 when ¢ = O.
j) x’ = x + b(X), x = 0 when ¢ = 0, where b is the sawtooth function of Example 2
and —(7/4) <t < (7/4).
2. Show that the equation x’ = (sin #)x has a nontrivial periodic solution.
3. Show that the equation x’ = (sin? yx has no nontrivial periodic solution.
4. Show that the equation x’ = x + sin thas a periodic solution. Is there more than one?
5. Does the equation x’ = —x + sin? t have a periodic solution?
 -

Fundamental solution sets for the homogeneous equation 103

6. Suppose 6 is a real-valued, continuous function such that b(f) > c ¥ 0 as t > Soe
Prove that every solution of x’ = —x + b(A) approaches c as t— +o. [Hint: Use
L’Hospital’s rule on a general solution.]
7. Let i denote the complex unit (i? = —1). What conditions must a continuous function
b satisfy if x’ = ix + (A is to have a solution of period 27?
8. Suppose 6 is a continuous function such that b(f) > 0 as t— +0. Prove that every
solution of x’ = —x + b(1) approaches zero as t—> +. [Hint: There are two cases.]
9. Try to write down a general solution for the equation in Example 2 by using variation
of parameters.
10. Discuss the behavior of the various solutions for x’ = —x + e-tsintast— +o.
11. Discuss the behavior of the various solutions for x’ = —x + sintast— +o.
12. We showed for illustration that the differential equation in Example 2 could have at
most one periodic solution. Could it have a periodic solution at all?
13. Suppose a current-limiting resistor is connected in series with an inductance ofL henries
across a source of voltage Eo sinwt. The resistance of the resistor decreases as it heats,
and in this particular application, the time-dependence of the heating is assumed to be such
that the resistance is given by R(2 + e~‘). What is the largest absolute value that the
current can reach at any time during operation of the circuit? It is not necessary to solve
the differential equation of the circuit in order to answer the question. Just examine it.

4.55 FUNDAMENTAL SOLUTION SETS FOR THE HOMOGENEOUS EQUATION

Here we concern ourselves with the generalization of the remarks in Section 4.4,
items (1) through (3), to equations (LH) x’ = A(t)x, where A is a continuous” X n
matrix function on the intervala < t < w.
The student should, throughout the discussion, keep clearly in mind the reasons
for the discussion: namely, to represent the solutions of systems such as

phere (22) OCU lee =A 0 allie

ales| ee Oe 1 -y or = bev ets —6
Z 1 —2 OJLz y (eee See
in familiar mathematical forms or to obtain qualitative information when no appro-
priate solution method is apparent, it is necessary to understand the structural
relations that hold between solutions of the system. In this connection, it should be
pointed out that constant-coefficient equations need not be as easy to study as text-
book problems might lead one to believe: ifn > 2, finding the zeros of a characteristic
polynomial
I te Ge aah aig meeee + an

exactly can be quite difficult. .

Let $1(1), ..., ¢n(t) denote any n solution vectors for (LH). It is convenient to
write them side-by-side as the columns of a matrix ®(t) = [¢1(4),..-, $,(t)]. The
104 The theory of linear differential equations 4.5

matrix function ® is called a matrix solution for (LH) because ®’(t) = A(f)®(1).
Suppose that the only constant vector ¢ for which ®(7)e =0 ona <t<w is the
constant vector c = 0. Then @ is called a fundamental matrix solution, and one says
that its columns 1, ..., » form a fundamental solution set for (LH). The quantity
det B(1), a < t < w, is called the Wronskian of the solutions $1,...,¢n. Now let
$1,--+5%m denote any vector- or scalar-valued functions which are defined on
a <t<w. They are called linearly dependent over (a,w) if there are constants
C1, -+ +5 Cm—not all zero—such that c1@1(t) + +++ + CmOm(t) = 0 for every ¢ in
(a, w). Otherwise, they are called /inearly independent over (a, w). In these terms,
one can say that solutions $;,...,@, of (LH) form a fundamental solution set if
and only if they are linearly independent over (a, w).

Example 1. The vectors

$1(t) a [2e”*, ae OPe

$2(t) = [cos t, cost — sint, —2cost — sin?#]’, and

$3(t) = [sint, sint + cost, —2sin¢ + cos f]”
are solution vectors for the system

58 |// 2 W iiiiss
ylo= Tt 0 Nyt
Z 1 —2 O]]z
and

Je" cos t sin¢

&(1) = | e? cost — sint sint + cost
0 —2cost —sint —2sint-+ cost

is the corresponding solution matrix. It happens that @ is a fundamental matrix

solution, but we shall verify this later. ||
If det (t) does not equal zero at any point of (a, w), then must be a funda-
mental matrix solution for (LH). To see this, suppose there are constants c},..., Cn
satisfying
€191(t) 42> + CnGn(t) =0

on (a,w). In matrix notation, we have ®(1)c = 0, where ¢ = [ci,...,¢n]”. Since

det (7) ~ 0 on (a, w), it follows from Theorem 3.17 of Section 3.11 that @—'(t)
exists for every ¢ in (a,w). Then 0 = ~1(Ad(e = Ie = c.
On the other hand, suppose that det ®(t) equals zero at some point t = fo of
(a, w). Then there is a nonzero solution ¢ of the algebraic equation ®(to)e = 0. It
is easy to check that x = ¢(t) = ®(Z)c is a solution of (LH) with (to) = 0. Note
that x = 0 also satisfies the equation. By the uniqueness of solutions to initial value
problems, x = (ft) is the identically zero-valued solution. Hence @(t)e = 0 on
 *

Fundamental solution sets for the homogeneous equation 105

(a, w), and it follows that & is not a fundamental matrix solution. We have therefore
proved the following theorem.
Theorem 4.3. The matrix solution ® is a fundamental matrix solution for
x’ = A(t)x if and only if the Wronskian det ®(t) does not equal zero at any point
of (a, w).
Example 2. The relations x = sin e’ and x = cos e! denote solutions of x! — x’ +
e”'x = 0. If the differential equation is written as a system

Ml "i|oe 1KE (4.5a)

then

Oe sin e! | eit ewer | cos e!

e' cos et —e' sin e?
are solution vectors for the system. A matrix solution ® is obtained by writing ¢,
and 2 side-by-side:

sino e t cos e t
P(t) = : 2
© ibcose’ —e’sin «|
Since the Wronskian det 6(t) = —e’ # 0, © is a fundamental matrix solution. ||
When it is necessary to check that a Wronskian det (t) does not equal zero on
an interval a < t < w, the process may be laborious since the elements of ® are
functions. The matrix @(t) in Example 1 is an illustration. It is enough, however, to
check the Wronskian at one convenient point of (a, w) since det ®(f) is either iden-
tically zero on (a, w) or is never zero on (a, w). This is a corollary of the following
theorem.
Theorem 4.4 (Abel-Liouville-Jacobi Formula). If ® is a matrix solution of
x’ = A(t)x, then ,
det &(t) = det (ty) exp |/ TrA(s) as, (4.5b)

where to and t are arbitrary points of (a,w) and TrA(s) is the sum of the diagonal
elements of A(s).
Proof. The proof is similar for all > 2. To keep the notation simple, however, we
consider the case n = 2 only. Let

so that
oo =f] mi oo -[
x(t) u(t)
Loe bes ae
106 The theory of linear differential equations 4.5

and write
a(t) b(t)
A= Be a ;
Since @; and @» are solutions of (LH),
x = ax + by, u’ |= au+
bo,
y =cx+ dy, v’ = cu + dv.
Then

d x ul se)
=oF det ® == de det k Al+ det
de * A ,

7 ax + by au+ bv 58 u
= det y D |+ et]. 3 cu + dv

atl Cltoes, a
ax au x Uu

= (a + d) det is “|= JTrA - dete.

The equation £ et &(t) = TrA(t): &(t) can be solved with the integrating factor

exp [—f,, TrA(s) ds] to obtain

t

det (1) = det $(to) exp |

i TrA(s) as|- |

Corollary 4.4. Either the Wronskian det ®(t) is identically zero or it is not zero
for any t in the intervala < t < w.

Example 3. The Wronskian of the matrix solution ® in Example | is

Dee cos t sint
det | e7? cos ¢ — sint sint+ cost |-
0 —2cost—sint —2sint-+ cost

To check that det &(1) ~ 0 for —w < t < +o by evaluating the determinant as
a function of ¢ is unnecessary labor. By Corollary 4.4 it is enough to check the
determinant when ¢ = 0 to find that det (0) = 5. ||
If the system (LH) x’ =A (4)x arises from an nth order equation
x™ = —a,()x—Y —--- — a,(t)x, (LH-n)
by means of the substitution xj = x) x5 =x, se xn ethene Abels
formula (4.5b) is t

det b(t) = det (to) exp | a,(s) as| (4.5c)

 Pa

Fundamental solution sets for the homogeneous equation 107

since 0 1 ie 0
A ss % - .
© 0 0 itaek|
—Aa,(t) =O, 1t) 6.00 —ay,(t)

The scalar equation (LH-n) can of course be studied by writing it as a system

(LH). We did this in Sections 3.7 and 3.8. Note, however, that if ¥i,..., W,», are
solutions for (LH-n), then the corresponding matrix solution for the associated
system will be
Pirvmwants We
aN Kiscig ae me
/ /

ye ee. ae)

The functions yj, . . . , W will be linearly independent and will be called a fundamental
solution set for (LH-n) over a < t < w, if and only if @ is a fundamental matrix
solution of the system. As a matter of practice, one need not write a scalar equa-
tion as a system to check that a set of solution functions is fundamental: the
Wronskian may be written down directly and evaluated at a convenient point.
Example 4. Two solutions of the equation
]
xt ” +t GA
Ape x ae 0, if > 0).

are x = \/t and x = V/t logt, t > 0. Their Wronskian is

Vt (Vt log 1)
det b(t) = det} _| 1 1 :
= Sai - log ‘
QN/t G ti aON Gt
Since det (1) = 1, the functions are linearly independent over the interval
0<t< +o. ||
As a matter of completeness, there is a question which we should answer. Are
there n linearly independent solutions of (LH) defined over (a, w)? Yes. Let fo in
(a, w) be arbitrary. By Theorems 4.1 and 4.2, there is a unique solution ¢; of (LH)
which satisfies the initial condition $;(to) = ux, where u; is the vector which has
kth component one and every other component zero (k = l,...,). Clearly
b(t) = [61(1), . . ., 4 (2)] is a solution matrix for (LH). Since det (0) = det J = 1,
Abel’s formula implies that det (4) ~ 0 on (a,w). By Theorem 4.3, ® is a
fundamental matrix solution and ¢$;,...,@n are linearly independent over (a, w).
We state this result as a theorem.
Theorem 4.5. If A is continuous on the interval a < t < », then
x" = A(t)x (LH)
has a fundamental matrix solution ® on the interval a < t < w.
108 The theory of linear differential equations 4.5

The results above for x’ = A(#)x are analogous to results for the scalar equa-
tion x’ = a(t)x, as indicated in Table 4.1.

Table 4.1

py’ = a(t) £’ = A(NP

Definition of ce= 0
10 lies c = 0
= 0 BP impAee implies
(Ac(Ac = 0 impli
fundamental solution

Characterization of t inin (a,(a, w)

Of for any ¢
i (a, w) detet B(¢ ~ 0
&(r)
fundamental solution V(t) ¥ 0 for any ¢ in

Permanence of (ft) = 0 on (a, w) or det (ft) = 0 on (a, w) or

fundamental solution | Y(t) ¥ 0 on (a, w) det &(t) ¥ 0 on (a, w)

Abel’s formula can be used to construct a fundamental solution set for the
second order equation
x” + a,(t)x’ + a2(t)x = 0 (LH-2)

provided one nonzero solution x = y(t) is known. Suppose that y,(t) ¥ 0 for
ty < t < ty. By reasoning analogous to that in the proof of Theorem 4.5, one can
be assured that there is a second solution x = W.(t) defined for ty < t < tg which
is linearly independent of y;. By Abel’s formula,

70) ¥2(0) /
t

det |; ; = ~ ds|> 4.5d

Vit) ¥5(2) nas to ae ( )
where Co is the value of the Wronskian at tf = fo. A little rearrangement puts Eq.
(4.5d) in the form

ya) — BO
Vi)
c
Y2(t) = 7D exp BP a,(s) as|. (4.5e)

Equation (4.5e) can be solved for y(t) by the method of integrating factors, although
it might be necessary to express the solution in terms of an indicated definite integral.
Example 5. The function x = (1 + on is the solution of the equation
pels 4. x! 2 i

l = oe ATS
1s ore °
which satisfies the initial conditions x = 1 and x’ = 0 when ¢ = O. The solution
x = y(t) which satisfies the initial conditions y(0) = 0, y’/(0) = 1 will satisfy the
equation
t

xy'(t) — x'y@) = exp| (eee

ae
las
Ts3 x=(Ud+ Ae
 -

Fundamental solution sets for the homogeneous equation 109

which can be written in the form

2t l
Wee
BY ty te
Upon solving the last equation with the aid of the integrating factor (1 + 17), one
obtains y(“) = ¢/(1 + t?). ||

EXERCISE 4.5
In Problems 1 through 20, find a fundamental set of solutions for the given equation or
system of equations. In Problems 11 through 20, one solution is given.
1. x’ + 4x’ + 3x = 0. Dill oe Phree tee EO)

By eM Se Bee SE ihe oS Zhe 10) 4. x!” + 3x" + 3x’ +x = 0.

5. x!” + 2x’ + 2x = 0. 6. EY 2x! Bx 2x! 4 2x = 0.
Re epee eap g Sfics[Goss|| 8]:
y 3 IJly

Li»
y 0 2ILy

EY-L3 5}
; ae ul’ 0 1 Ol} u

y a y w —1 3 -1]|w
6 4 1
pt x = 0, a2 2-0

3 1 1
12.x"
tt
+ pallet
ox + se,
gx —
Ota =->?
2 b> 8

13. $i + ax=0, —sSinl (Ine) nO:

14. x’ + (tan px’ + (tant — 1)x = 0, x=e, Se

Sh aeWPS. d + 2tan’ 2 t)x es =0, = sect,

x = a8 5a

16. x” — 4tx’ + (4 —2)x =0, x=e'.

yx” tv
1 1
e (Et 4-2)x-0,
ih ate Gum
2 Vies eT >
Vt At an, =?
4 Wale 1-t
ie ek — x= . = ——

d+ 7) — 2) 1+t
2, 1 Sahl
19. x ” Leis
ieee) [Link]
Gian = Ox = sin (4)
—j>o f¢ > 0.

20. x” — ox’ + 47x = 0, =COS i), 7 S> O:

21. Let g be a continuous function defined for -~» <t< +% and suppose that the
equation x” + q(x = 0 has a solution x = ¥(¢) such that limy,4. |W(D| + WO] = 0.
110 The theory of linear differential equations 4.6

How can one be sure that no solution other than a scalar multiple of y behaves in the same
way as t— +0?

4.6 RELATIONS BETWEEN SOLUTIONS OF THE HOMOGENEOUS EQUATION

We observed in item (4) of Section 4.4 that if x = y(t) is a solution of x’ = a(t)x,

then so is x = ¥(t)c for every constant c. If x = ¥(f) is a fundamental solution and
if x = ¢(ft) is any other solution, we showed that there must exist a constant c such
that $(t) = ¥(t)c. These results generalize to systems and higher order scalar equa-
tions in a natural way. The generalizations are listed below as several short theorems
which are fundamental facts about linear combinations of solutions for linear
homogeneous equations
x’ = A(t)x. (LH)
The student should keep in mind that the linear combination c1@,(t) + +++ + Cnon(t)
assumes the form ®(f)e in matrix notation.

Theorem 4.6. If ® is a matrix solution of (LH) and if C is a constant matrix, then

X(t) = B(£)C is a solution matrix for (LH).

Proof. X'(t) = ®(t)C = A(DME(C = A(HX(t). ||

Example 1. The matrix

&(1) = ee 4
is a solution matrix for

; e2! 5te2!
eel
a) hh 9 2t 15 2t 4 2t 2t
The matrix X(f) = E ont | = |3 a i : yee |is also a

solution matrix. ||
Corollary 4.6a. If ¢1,...,4n are solutions of (LH) and if c1,...,¢n are any
constants, then
o = cid + * 4-2 Gig,

is also a solution.

Proof. Take
C1 0 0

ee ee :
Cy 0 0
in Theorem 4.6. ||
 -

Relations between solutions of the homogeneous equation 111

Example 2. The vectors $,(¢) = [e”,0]” and (1) = [5te?*, e?4]? are solution
vectors for
7] = ON =) x].
Veen OFe2\inRy
The vector
6G) = Al2t |4.3 ee2t 2t
of ies Bae 2t

is also a solution vector. ||

Corollary 4.6b. If ¥1, ..., Wn are solutions of

x™ = —a,(t)x™—P — +++ — an_i(t)x’ — an(t)x (LH-n)

and if cy, ... Cn are any constants, then

yo="eyi + ** > EF Can

is a solution also.

Proof. Convert (LH-n) into a system. Then

pull ih =Sen ys
V1 eases Vn
/ y

yo) we yord

is a matrix solution. Now take

in Theorem 4.6 and consider the first row of C. ||

Example 3. The functions y(t) = 1/(1 + 2”) and y2(t) = (1 + 17)? are solutions
of the equation
eens ly
= 0.
a+ ep
/

The function y3(t) = ¢°(t* + 32? + 3)/(1 4 2°) is also a solution of the equation
since ¥3(t) = Yo(t) — yi). ||
If a fundamental set of solutions for (LH) or (LH-n) is known, then any other
solution can be constructed from this fundamental set by forming linear combinations.
The next theorem directly generalizes item (4) of Section 4.4 to systems.

Theorem 4.7. If ® is a fundamental matrix solution for (LH) and if X is any

matrix solution, then there is a constant matrix C such that X(t) = ®(1)C.
112 The theory of linear differential equations 4.6

Proof. We need only show that &—1(1)X(t) is a constant matrix.

"/ar ead n= Pe PRE
Zh OX) = B® OXO + & OGAO
-o "4 a) usX(t)
(1) X() + &7? dt
I —@1()A() + &1DADX( = 0.
Therefore &~1(t)X(t) = constant matrix. ||
Example 4. Consider the system x’ = 2x + y, y’ = 2y. Two solution matrices are
e2t te2*
B+ He -—T+ a
&(t) = E 1 and X(t) = et —2e2t
e
Since det (0) = 1, @ is a fundamental matrix solution. Thus ~'(f) exists and
| me 3 -—7

a constant matrix. ||
In the applications, one is frequently interested in finding a vector solution which
satisfies some initial value or boundary value problem. Any solution can be found
if one knows a fundamental solution set.
Corollary 4.7a. If ¢ is any solution of (LH) and © is a fundamental matrix solution,
there is a constant vector ¢ such that o(t) = ®(ft)e.
Proof. The proof is virtually identical to the proof above.

es
5 (® oO) I= he—P 0)Aue ~1¢4) 40FO
F OF "80 + ATO
—& "(DAG(D) + (NAG(Y = 0.
Thus ®~ '¢(t) = constant vector. ||
Example 5. Find that solution @ of the system
; t l /
1 t
ES
a ee) ea een or
which satisfies the initial condition (0) = [7, 3]”, given that [1, ¢]” and [z, 1] are
linearly independent solutions.

+0-o[]] +f
By Corollary 4.7a, the required solution is of the form

The constants c; and c2 are evaluated by solving the equations

sl} [ols f
 -

Relations between solutions of the homogeneous equation 113

for c, and Co, and

o(t) = [7+ 34,3477". ||

An analogous result holds for the mth order scalar equation

ean (ee tee Git) ka, (tx. = 0. (LH-n)

Corollary 4.7b. If p is any solution of (LH-n) and if ¥1,...,Wn is a fundamental
set of solutions, then there are constants C1, ..., Cn such thatw = cy, t-++ + Cavan:

Proof. \f (LH-v) is converted into a system, then

eV wy ame el
is a vector solution and
Vey,
, /
OD = vi ire Wh

yp?
=
) ia yp sce
1)

is a fundamental matrix solution. Now apply Corollary 4.7a. ||

Example 6. The functions y(t) = sin (e’) and W2(t) = cos (e’) form a fundamental
solution set for the equation x’’ — x’ + e*x = 0. The solution x = y(t) that
satisfies the initial conditions y(In 7/2) = 1, ¥’/(Inw/2) = 1 will be of the form

v(t) = cz sin (e’) + ce cos (e’).

The constants c, and cz are found by setting t = In 7/2 in the equations

V(t) = c, sin (e’) + c2 cos (e’), ri
y(t) = ce! cos (e’) — cee’ sin (e’). 6)
When this is done, one finds that

W(t) = sin (e’) — £cos (e’).

Notice that Eqs. (4.6) comprise a system of algebraic equations for c; and Co,
and that the determinant of the matrix of coefficients is just the Wronskian of ~;
and Wo. ||

EXERCISE 4.6
In Problems 1 through 20 there is given an initial condition. Find that solution of the same
numbered equation in Exercise 4.5 which satisfies the initial condition. Use your previous
solutions as starting points.
ih, 3) = ty AO: = 1h 2. x(0)'= 2, x'(0) =:1.
Se e(O) See) 15x O)-=1. Ay x(0). =1.0,. x’ (0) =s1,-—"'(0) = 0,
 4.6
114 The theory of linear differential equations

Sy, 20) = i, 37) = 1. 6. x(0)-= 0, XO) = 1,2’O =H 0, x70) ]4:

i, AO) = 2, WO) = ©, , SACO)) = Bs XK) = 1
9. x(0) = 2, yO) = 3. 10. u(0) = 1, v0) = 0, w0) = 0.
1x0) = 2,0 (seat: 121) eae
1, sD) = 2, el) = 3 at sO) = 2, 364) = 3;
ilSy, 2t@) = 5 kh@O) = & 16. xO) = 0>x’@) = 4.
7h, afl) = OF sO) = Se; 18290) — 1. ©) —
19527 — ee (2/7) — se 20. x(0) = 0, x/(0) =s1,. xO) = Lex’) =20:
21. Find a constant matrix C such that ¥()) = ®()C if

a) &(/) = [. a
1
el, Wi) = i
0
ive1 | elt
et! “Sie 0 e#t 5(1 + de** 3e*
()) eX) =|) © ert MO |, %4e) =| © en @) |e
0 0 e8t 0 0) 4e3¢

22. Let denote ann X n matrix function, differentiable on the intervala < t < w. Dif-
ferentiate the equation ®—!(A)@(1) = J to prove that fa = —6—!’6-1,
23. The switch in the circuit of Fig. 4.3 is thrown at time ¢ = 0. The currents x and y
satisfy the initial value problem
, LR, MR> MR, LRe
agesee pr 2 * IF t
[eal ee eee ean

x(0) = E/Ri, yO) = 0.

If L = 2, M = \/3, Ri = Re = 1, express x and y as functions of time. At what time is
the current y largest?

L Re

Figure 4.3

24. What value must the positive constant w have in order that the boundary value problem
x’ + w?x = 0, x(0) = 0, x(1) = 0 have a nontrivial solution?
25. What conditions must the real constants a and b satisfy in order that the boundary value
problem x’’ + ax’ + bx = 0, x(0) = 0, x(+) = 0 have a nontrivial solution?
 Solutions of the nonhomogeneous equation 115

4.7 SOLUTIONS OF THE NONHOMOGENEOUS EQUATION

It is assumed throughout this section that a fundamental solution set reo
for
x’ = A(tx (LH)
is known, and we write ® for the fundamental matrix solution which has the oS
for columns. Then
$1 = A(t)dr, .--5 Grn = Aton,
or more compactly,

®'(t) = A(t)®(2).
If just one solution @, of
x’ = A(t)x + b(t) (L)
is known, then every solution can be expressed in terms of ¢,(f) and ®(). An analo-
gous assertion holds for the nth order scalar equation. As an illustration of the
technique consider an example.
Example 1. Let us find the unique solution x = y(t) of the equation

Si A ae = , (4.7a)

which satisfies the initial conditions y(1) = 1, ¥/(1) = 0.

We shall see in a later section that the functions

¥i(t) = Vtsin ie In ) and = Yo(t) = Vt cos (a In )

form a fundamental solution set for the complementary equation x’’ + 1~?x = 0.
A particular solution y, of Eq. (4.7a) can be found by inspection: y,(t) = ¢. Notice
that y, is not the solution to the initial value problem which we want to solve since
Yp(1) = Land yi(1) = 1.
If we define u(t) = Y(t) — ¥,(t), however, it is easy to check that u’’(t) +
t- u(t) = 0. That is, uw is a solution of the complementary equation. Thus, by
Corollary 4.7b, there are constants c; and cg such that u(t) = cyi(t) + coo(t).
This implies that Y(t) = ciWi(t) + coW2(t) + t. To solve the problem as posed,
then, we need only determine the numerical values of c; and cy by solving the
equations
1 = WU) = civi(l) + Cove) + I,
0 = Wd) = ci) + cova(l) + 1,
and conclude that

y(t) = — Se Visi (Xm ‘ +t. |I

116 The theory of linear differential equations 4.7

Before discussing methods for finding particular (specific) solutions, let us

formalize the technique of Example 1.

Theorem 4.8. If x = $,(t) is a particular solution of x! = A(t)x + b(t), then

every other solution x = $(t) can be expressed in the form

& = Cio, + °° + Cabn + bp = PC + Op, (4.7b)

for some suitably chosen constant vector ¢ = [C1,..., Gils

Proof. Let x = ¢(¢) be an arbitrary solution of (L) and define

W(t) = P(t)e + p(t), ¢= B'(o)ld(to) — d(to)],

where fo in (a, w) is arbitrary. Since
V(t) = ® (Hc + o,(t) = AP()c + ,(¢)
= A®(t)c + Ad, (t) + b(t) = AY) + b@),
x = y(t) is a solution. But (to) = (to). By the uniqueness of solutions to initial
value problems y(t) = $(1), that is, (4) = b(t)c + ¢,(t). ||

Corollary 4.8. If 1,...,%n are linearly independent solutions of (LH-n) on

(a, w) and if p is a particular solution of (L-n), then any other solution y of (L-n) can
be expressed in the form
Y= Cito Can ep, (4.7c)
for suitably chosen constants C1, C2,... 5 Cn-

Proof, Writeld = WaW')3.25 07

a 1s Go Wee ee ae
Vio ++) Un
af Mio
oa pire 1)

Then, by Theorem 4.8, there is a vector ¢ such that ¢(/) = S(A)e + ¢,(t). Equation
(4.7c) is the first row in this matrix equation. ||
Theorem 4.8 and its corollary generalize item (5) of Section 4.4 to systems of
equations and higher order scalar equations.
We shall refer to the formulas (4.7b) and (4.7c) as general solutions for (L) and
(L-n) respectively. A solution method for these equations may then be summarized
as follows: To find the general solution of a nonhomogeneous equation, one first
solves the corresponding homogeneous equation and then finds any particular
solution of the nonhomogeneous equation.
The method of annihilation (Section 3.9) combines these two steps for equations
of the form
x™ 4 ayx™Y 4 +++ + ay yx! + ax = B(2),
 Solutions of the nonhomogeneous equation 117

where the a,’s are constants and the function b has one of the forms (3.9e). For
other types of equations, the method of annihilation is not so useful. There is, how-
ever, a method for finding a particular solution of a nonhomogeneous equation which
can be applied whenever the coefficients of the differential equation are merely con-
tinuous. It is a generalization of the variation-of-parameters technique given in item
(6) of Section 4.4.

Theorem 4.9 (Variation of Parameters). If ® is a fundamental matrix solution for

x’ = A(t)x, then

$,(t) = &(t) / @\(s)b(s)ds, a<t, to <a, (4.7)

is a particular solution of x’ = A(t)x + b(t). It satisfies the initial condition $,(to) = 0.
Proof

$1,(t) = &'(t) / @—1(s)b(s) ds + &(1)}@—*(N)b(t)

A(t)®(t) / ® *(s)b(s) ds + b(t) = A(t)o,(1) + b(t). ||

Corollary 4.9. Any solution x = $(t) of x’ = A(t)x + b(t) can be expressed
in the form
t

p(t) = P(t)e + P(Z) / b~*(s)b(s) ds,

fo
where c is a constant vector and ® is a fundamental matrix solution of x' = A(t)x.
Example 2. Consider the system

br -La oli] *El

sal
=
OF xe .0
,
(9
4.7

It is easy to check that

$i(t) = hee | and

cos ¢t
a(t) = ee |
are solution vectors for Eq. (4.7e) and that
sint cost
Ok eae —sin |
is a fundamental solution matrix. Now

sil ten COST

or Ee —sin t]’
so
2 sins coss||O| | cos
2 1(s)hG) = lane —sin |H i. pies 4
118 The theory of linear differential equations 4.7

and

[; & ZA “(s)b(s)ds = [| css] i, ie ne |

» L—sin s ds = ir gi ds COS Dees

Finally,
= sin ¢ . cos? sin t
p(t) I x0 | 4” *(9)b(s) ds = bee —sin iFe t) — ‘|

_ bee cos” t— cos |a ! — cos |

sin ¢ sin ¢

The reader might have noticed that ¢,(t) = [1, 0]’ is also a particular solution which
could have been found by inspection. ||

Theorem 4.10. If ¥1,..+ 5 %n are linearly independent solutions of (LH-n) on

a<t<_, there are functions 0,1, ..., Un such that

Vp = 011 +°°* + Una

is a particular solution of (L-n). In fact, one need only solve the equations

|vit)... nl) | |0 |
: : es alan (4.7f)
ee ir SENG) b(t)
for v(t), ...,0,(2) and integrate to find vx(t), .. . , Up(t).

Proof. The values v,(t),...,U,(t) are the first through nth components of
ie @—'(s)b(s)ds in formula (4.7d). To see that we can compute them from
Eq. (4.7f), set
t
COR eHOl = / &'(s)b(s) ds.
Then i
[vPi@),..., v4)" = & (b(t) and &(A)[v,(t),...,v,(O)° = b(2).
The last equation is (4.7f). ||

Example 3. We find a particular solution of the equation

xX Uses taney Sa) 2 eh <my
Two linearly independent solutions of the complementary equation are x = sinf
and x = cos?t. Thus, we solve the equations

sint cost|[v;(t) 0
be fe Ssin |Ew zs Be | (4.7g)
 Solutions of the nonhomogeneous equation 119

for v;(t) and v(t) to find that

Ot) = sine and v(t) = —(sin £) tan ¢. (4.7h)
It follows that v(t) = —cost and vo(t) = sint — In|sec¢ + tan t|, where the
constants of integration (which are arbitrary) have been taken equal to zero. A
particular solution is therefore

Y,(t) = sin t(—cos t) + cos ¢(sin ¢ — In |sec¢ + tan ¢])

= —(cos t)In |sect + tanz?|. ||
Example 4. The variation-of-parameters algorithm can be used to study problems
which do not fall exactly within the scope of the existence theory as formulated in
Sections 4.2 and 4.3. Let us find a function which is continuous on the interval
—«o <t< +o and satisfies the differential equation
t
0 i a ae id (4.71)

for all t ¥ 0. Equation (4.7f) has the form

sinft cost||v;| _ 0
cost —sint||v,| |—lift<0, +lift>0
Thus v{(t) = —cos?¢ and v,(t)= sint if ¢ < 0, but vi) = cost and v(t) =
—sint if t> 0.
General solutions of Eq. (4.71) are
x ll Asint+ Bcost — 1 if t<=0

and

x = Csint+ Dcost+ 1 if eS @.

We “paste” the two functions together to form one continuous function by requiring
that B — 1 = D +1, that is, by requiring that B = D+ 2. Thus
Be ston (Di. 2)(cos2) —— 1, L-—20s
* = lCsint
+ D(cos 1) + 1, 7 0,
is a solution to the problem as posed for any choice of the constants A, C, and D.
Notice that the unique solution of the initial value problem
x! x = t/|t\, ve O ema! when t= —1/2 (4.7j)
is
x = —sint-+ cost — l, —o <it< 0. (4.7k)

Each of the functions

—sint + cost — l, pee A (4.71)
= ; ;
Csint — cost + l, t> 0, C arbitrary
 The theory of linear differential equations 4.7
120

satisfies the initial conditions, is continuous for —# < t < +o, and satisfies the
differential equation except at t = 0. Yet none of these functions qualifies asa
different solution because none satisfies the differential equation at t= 0. Ina
more general existence theory, (4.71) is admissible as the solution to the initial value
problem (4.7j) if C is chosen so that x’ is continuous at ¢ = 0 (see Problem 11 in
Exercise 4.7). ||
Although the general procedure is rather complicated, the method of annihilation
can be adapted to yield particular solutions of systems x’ = Ax + b(t), where A is
a constant and b(¢) has special forms. We shall not discuss the general situation here,
but it is convenient to know that the system x’ = Ax + ce” has a particular solu-
tion of the form x = pe if \ is not an eigenvalue of A. The vector p may be de-
termined by substitution (which amounts to solving the algebraic equation
(A — d\J)p = —c for p).
If Wy; ..., veare solutions of x’ = Ax biG), .2.,x = Ax) ete
spectively, then Wy= Y; + °°: + Wm is a solution of x’ = Ax + b(t), where b =
b, +-::+b,. This fact is called the Superposition Principle. Using it, one can,
for example, find a particular solution of the equation x’ = Ax + csin wt if iw is
not an eigenvalue of A. This is done by finding particular solutions of the equations

—itwt
x’ = Ax + i cen, x= AX= Dce

and adding them.

EXERCISE 4.7

In Problems 1 through 10 there is given an initial value problem and (in most cases) a funda-
mental set of solutions to the corresponding homogeneous equation. Solve the initial value
problems by variation of parameters.
il, se¥ tee = SiN, KO) = il, XO) = ©.

i, oe! = pe ea RECT, SKM) = O, sO) i,

4 6
3.x" x taxes, x) =3, ¥0)=0, WO=7, WO =?.
; 5 ; 1 2 2
An yt! oe aod ci at as 74008 (*)yy oe (2)SiG) 5d (2)= 0,
(al
Wid) = LLy (*)’

Clete} Bolla)
CU) Dol-Lih eo -[3} #0 -[8]
 Fundamental solution sets for homogeneous equations with constant coefficients 121

oe fee arte el e x@)}_ | 1 e' sint

is H 7 Ee a le ba = |_i]: ot) = & cos |

e sint

[3 eel —t I1||x 1 x(0)| _ |0 1 t

: 3 aes
| t+ L) EaBKE oto = [1]. oo =|]
Bald it @) -+ x : x(1) 1

OM a 1 Oe 1 hy Peet 16 |p)! 101s

Zz 0 0 1jLz 1 z(1) 1

t t 1
gilt) = |0]> got) =|F|> o3(0=] 4
0 0 t

ae E00 Ie t x(1) —2
10. = alot P-Olly| +1 ly@l=| 4b
Zz Peeper |e t z(1) 3

t 0 0
gift) =] 1]> got) =|t]}> g3(t)=] 0]-
0 1 t

11. Show that there is a unique, continuously differentiable function x = (A) defined for
—«°o <t< +o, which satisfies the differential equation x’’ + x = t/|t| for t ¥ 0 and
the initial conditions Y(—7/2) = 0, ¥/(—7/2) = 1.
12. Suppose the metal block of Fig. 1.3 is initially at rest with the spring (unit stiffness
coefficient) uncompressed. It is pushed to the left by a force of magnitude b(f) = ¢,
0<t < 7/2, b(t) = 0, t > 7/2. Describe the motion of the block if there is no friction.

4.8 FUNDAMENTAL SOLUTION SETS FOR HOMOGENEOUS EQUATIONS

WITH CONSTANT COEFFICIENTS

The theory developed in Sections 4.5 and 4.6 guarantees that homogeneous equations

x’ = A(t)x (LH)
or
aL)x ea ee ge ax 4 a,(ty = 0 (LH-n)
have fundamental solution sets and that any solution can be expressed uniquely as a
linear combination of fundamental solutions. Here we wish to connect these results
with the solution methods derived in Sections 3.7 and 3.8 for constant-coefficient
equations
x’ = Ax (LHC)
122 The theory of linear differential equations 4.8

and
XO gy DE oe te gy a, nO, (LHC-n)

Consider the nth order equation first and suppose that its characteristic equation
has the factored form
Oka NY) ec Oe Se Cee)
mter-tmt+-:--+m =n. We saw in Section 3.8 [Eqs. (3.8e) and (3.8f)]
that
x Set Keates ee ate ee (4.8a)

are solutions of (LHC) for r = 1,...,k. Thus Eqs. (4.8a) yield m, solutions cor-
responding to the eigenvalue \;, m2 solutions corresponding to the eigenvalue
Ao, ..-, and m; solutions corresponding to the eigenvalue )x,.
Let us show that the 7 solutions defined by Eqs. (4.8a) are linearly independent
by proving that
k

ss (1, Se LY or = Rig = Et eal EY =0 (4.8b)

Us|

implies Vy,.= Yor ="? = Ym, = 0 for.r = 1,...,k.) To do this) werassume

that the characteristic values \; = a; + i8; are arranged in such a way that a; <
ag <-+** < ax. We multiply Eq. (4.8b) by e~” and obtain

aie ton- 2 ae ye,)

k—1
ODS Cie + tar Et Yee exp [Oe = A = Os
1

Now

lexp [(A, — Ax)é]] = lexp [Car — ax)t]- exp [i(B, — Bx)¢]]|

exp [—(a, — a,)t]- [cos” (8, — Bx)t + sin? (8, — B,)t]*/?
= exp [—(ax — a,)t].

If wu> O and n > 0, products of the form vt“e~"' > 0 as t— +o by L’Hospital’s

rule. Thus, letting t > +o in Eq. (4.8c), we find that

lim (az top o> it i) = 0.

t>+o

This is only possible if ¥1, = Yor = ++: = Ym,z = 0. Knowing this, we replace
Eq. (4.8b) by
ee 1

é (M1, et Yore Sis ie ee eae = 0,

—

multiply by exp [—)x_1/], and repeat the argument. After doing this k times, we
conclude that all the Y’s are zero. Thus we have proved the following theorem.
 Fundamental solution sets for homogeneous equations with constant coefficients 123

Theorem 4.11. The n functions defined by Eqs. (4.8a) JOT \aetien eeform.a
fundamental solution set for (LHC-n).
Note that Theorem 4.11 and Corollary 4.7b together constitute a different proof
of Theorem 3.10.
Next, consider the equation x’ = Ax, where A is ann X n matrix. If \ and p
form an eigenvalue-eigenvector pair for the matrix A, then x = pe is a solution of
the equation since
Ax — x' = Ape — dpe = (A — N)pe™ = 0.
We shall call such a solution an eigensolution. Suppose now that A has n linearly
independent eigenvectors pi, ..., Pp» with corresponding eigenvalues \,..., Xn.
Then
S(O) 1 [pies a. Hapa 2]
is a solution matrix and, since det (0) = det [p,,..., pn] ~ 0, it is in fact a
fundamental solution matrix.

Theorem 4.12. If A is ann Xn matrix with n linearly independent eigenvectors

Pi,---,5Pn and corresponding eigenvalues \y,..., Xn, then
x= pie’,t Say X = Pre Ant

form a fundamental solution set for x' = Ax.

Now let us turn our attention to equations x’ = Ax which do not have n
linearly independent eigenvectors.
Theorem 4.13. Let A be a 2 X 2 matrix which does not have two linearly inde-
pendent eigenvectors and let \, and p, be an eigenvalue-eigenvector pair for A. Then
x’ = Ax has a fundamental solution set of the form

x = pie’, x = (tpi + pode, (4.8d)

where
(A — Aql)po = Pi. (4.8e)
Proof. Theorem 4.12 guarantees that x = p,e"’ is a solution of x’ = Ax. Corollary
3.22 of Section 3.13 guarantees that Eq. (4.8e) has a solution for py and that pp is
linearly independent of py. Thus x = (fp; + p»)e’ is defined and

x’ — Ax = 4 (tpi + poe! + pied! — Atpye*! — Ap2e*'

= —te'(A — dyDypy — e?"[(A — dol)p2 — pi] = 9.
The solutions (4.8e) are linearly independent since their Wronskian equals det [p1, po],
which is nonzero, when ¢ = 0. ||
We shall not prove a more general version of Theorem 4.13 for n X n matrices
A, but a few words of clarification for this case are in order. It was established in
Section 3.7 (Theorem 3.8) that each solution of x’ = Ax has the form x = eA
124 The theory of linear differential equations 4.8

for some constant vector c. Consequently, &(t) = e4’ is a solution matrix for the
system. Since ®(0) = J, the identity matrix, and det J = 1, it follows that ® is a
fundamental matrix solution. Its columns, therefore, form a fundamental solution
set. We wish to examine these columns and thereby discover the forms of these
fundamental solutions.
Recall, from Eq. (3.7d), that &(t) = e4’ can be expressed in the form (1) =
Pe®' if P-1AP = B. Let us assume that P has been chosen, as in Theorem 3.7, so
that B is a Jordan matrix
Bea diag (Ba, aac Das ea

where B; (j = 1,...,m) is v; X v;, a Jordan block with the eigenvalue \; of A on

its diagonal. [Note: It may happen that some of the blocks have equal diagonal
elements.] The columns of B which contain the elements of B; are the columns
numbered
Cae TsFP Ps

where o = vy +-++:+-+ y;_,, and the correspondingly numbered columns of e?¢

have the form

The o + Ist through o + v,th columns of Pe?! are thus

Deqie

(Po4it + Dae
0?
(peasTI + Poot + pes) eM Ge)
: fi-} Sc
Po+1 @;— 1! erent hae Po+r;-10 2 Po+v; Cale

and the n solution vectors defined by (4.8f) for 7 = 1,...,m thus comprise the
fundamental solution set which we wished to find.

Example. Consider again the system

Xx

Vij s| OR 20) yale (4.82)

7és
 -

Fundamental solution sets for homogeneous equations with constant coefficients 125

which was solved in Example 3 of Section 3.7. We found there that the system has a
general solution of the form

x Ay By Cy
y| =|}—By, le +] 0 |te +] 0 |e (4.8h)
ta B, 0 Cy

where A;, By, and C, are arbitrary constants. Let us rearrange Eq. (4.8h) and write
it in the form

[x,y,z]? = Aypie*’ + Bi (pit + po)e* + Cipze**, (4.81)

where

|Oi [1, 0, OF |
Uy [0, =k i and |
UE ihe 0, Wie

The matrix P = [pj, po, p3] has the property that P~1AP is a Jordan matrix

Zee 0
OFF240
O03

Taking o = 0, j = 1, A; = 2, and v; = 2, we may therefore identify the solution

vectors p,e”’ and (pit + pz)e” in Eq. (4.8i) as the solution vectors (4.8f). Similarly,
taking o = 2, j = 2, \; = 3, and v; = 1, we identify the solution vector p3e* in
Eq. (4.81) as the vector (4.8f). ||
If the equation (LHC) is obtained from (LHC-n) by means of the substitution
X1 = X,Xq = X’,..., Xn = x [see Eq. (3.8b)], then it turns out that the only
constant vector in (4.8f) that has a nonzero first component is p,,1. In this case, the
first components of the solution vectors in (4.8f) coincide with the functions (4.8a)
with m; = v; and k = m. A proof of this assertion would constitute another proof
of Theorem 3.10 so we will not pursue it here. It is important to note, however, that
x’ = Ax does not generally have solutions of the form x = r*e*'e with k a positive
integer.

EXERCISE 4.8

In Problems 1 through 5, use the answers for Exercise 3.12 to solve the differential equation

a
for the indicated values of a, b, c and d.
PT-[2 Se
1@a@=3,b5=1,¢=0, d= 2. 2G = Ae ba 3 0 = 2,40 =),
3. a= 6,b=2, c= 5, d= 3. 4-€a@= 0b —1,¢c=—1,7d=0.
Sea =2,6=3 ,¢= 2.
= —3,d
 The theory of linear differential equations 4.8
126

6. Find a solution of x” + 5x’ + 4x = 0 which satisfies x(0) = 1, x’(0) = 0.

7. Find a solution of x’” + 2x” + x’ + 2x = 0 which satisfies x(0) = 0, x’(0) = 1,
x"(0) = 0.
8. Find a solution of x” + x’ + x = 0 which satisfies x(0) = 0, x’(0) = 2
9, Find a solution of x” + x” + 2x’ + 2x = 0 which satisfies x(0) = 0, x’(0) = 0,
(Ogle
10. Find a solution of the system
se | 24 6 9\|x
y| =| —66 —15 —36]] y
Z D 0 Dez

which has x(0) = 0, y(0) = 0, z(0) = 4.

11. Find a solution of the system

i 24 6 9\|x
= |—42 —9 —18]| y
A) Vas AL |||

which has x(0) = —1, y(0) = 2, z(0) = 4.

12. Compute the eigensolutions for the differential system of Problem 5, Exercise 3.7.
Discuss the physical characteristics of the motion if it is described by an eigensolution.
13. Repeat Problem 12 for the system in Problem 6 of Exercise 3.7.
14. Repeat Problem 12 for the system in Problem 8 of Exercise 3.7.
15. Repeat Problem 12 for the system in Problem 9 of Exercise 3.7.
16. Find a set of real-valued fundamental solutions for each of the following differential

vol-tL2 ae
ETL ab
S-Ee)
rae falc

4 -
oo T-fL-IB xa

Z
ee

)
Ol

@
es

Alliiz

w | it 4 1 O|| w
x -1 1 01 3%
Bae On One lay
Zz 0 0-1 1 Z

17. The equation x’”” — 3x’’ + 3x’ — x = 0, has as a fundamental solution set the func-
tions e’, re’, t2e’. Construct the corresponding matrix solution of the associated system.
CHAPTER 5

SOLVING LINEAR EQUATIONS

WITH LAPLACE TRANSFORMS

5.1 DEFINITION OF THE TRANSFORM

If \ = a + i@ is any complex number, we shall call a the real part of \ and B the
imaginary part of \. We denote them by a = Red and 8 = Im. Now let f be
a scalar- or vector-valued function such that fe f(t) dt exists for allt > 0. We
shall call a function which satisfies this condition positively integrable. If, for a
positively integrable function f, the improper integral
+0 fo foo
fe dt = : Wiea COS BEAL =o) 4 f(tje—*' sin Btdt — (5.1a)
converges, we shall call the function £f of \ which it defines the Laplace transform
of the function f.
Example 1. Let c be any complex number and write E,(t) = e°’. Then
+o ,
£EA(A) = /f e'e—™dt = wees
= =)

provided Rev’ > Rec. ||

An important general property of the Laplace transform is its linearity: if £f(A)
and £g()) exist and if a and b are constants, then
+o +00
aLf(r) + b£g(r) I a fide dt + b i g(t)e—™dt
0
+0

= ii [af (1) + bee dt = Llaf + bg](r).

This fact, in conjunction with the result of Example 1, allows us, for example, to
compute the Laplace transforms of the sine and cosine functions.

127
 5.1
128 Solving linear equations with Laplace transforms

Example 2. Let S,(f) = sin wt and C,(t) = coswt. Then

B —twt ae dwt
Si)
i
= ]ae"
dwt
— 58 and C(t) = 3e° +

Thus, for Re \ > 0, we have

1 1 l 1 w
LSa(\) = 5s is Oi eee
and
pot wet Awe
BGAN) 15 ae gga eee goes ain
Example 3. Let f(t) = e“sinwt and g(t) = e* coswt, where a and w are real
numbers. Then
1 a tw —tw
f(t) = 7° (ea = tea),
and

i 1 1 w
Oe ‘lS (a+ iw) X= tasl- (X — a)? + w?
Similarly

Neve
ee) = 3[>a - ee | Se !
Frequently, for a given positively integrable function f, one wishes to deduce
that the integral (5.1a) converges, but he wants to do this without actually evaluating
the integral. This end can be achieved by demonstrating the existence of constants
M, n, and fo such that |f()| < Me™ for tp) < t < +o. In this case,
[fe™| = e—“[cos? Bt + sin? Bt]/?| f()| < Me~*™*", X= at if.
Since
+0

Me dt. a >a,
0
converges, it follows from the comparison test that
+00

f(be— dt
0

converges for Re \ > 7». A function f which can be dominated by an exponential

function in this way is said to be of exponential order at infinity, and one customarily
abbreviates this statement by writing f(t) = O(e") as t— +o. Thus, positively
integrable functions which are of exponential order at infinity will have Laplace
transforms which are defined for Re } sufficiently large.
 Definition of the transform 129

In this chapter, we shall use Laplace transforms to solve the differential equations

x’ = Ax + b(t) (5.1b)
and
De igh Vays Bg a! Avge b(t). (5.1c)
We assume, throughout the chapter, that A and a,,...,a, are constants and that
b and 5b are piecewise continuous, that is, have at most a finite number of discon-
tinuities in any bounded interval and have finite left- and right-hand limits at any
point of discontinuity. The existence theory of Chapter 4 applies to Eq. (5.1b) and
Eq. (5.1c) with a slight modification when b and b are merely piecewise continuous.
We admit a continuous function as a solution if it satisfies the differential equations
everywhere except at the discontinuities of b and b. In the case of the nth order
equation (5.1c), we also require that each solution have n — 1 continuous derivatives.
In order to use Laplace transforms to solve linear differential equations, we must
know that their solutions, in fact, have Laplace transforms.
Theorem 5.1. Assume that b is positively integrable and of exponential order at
infinity. Then every solution of
x’ = Ax + b(t)

is of exponential order at infinity, hence has a Laplace transform.

Proof. Let |A| denote the sum of the absolute values of the elements of A and let
M > 0 and 7 > |A| be such that |b(1)| < Me” for 0 <t< +o. If x = o(f) is
any solution of the differential equation, then
t

o(t) = e4e + i;e4'—b(s) ds,

0
where c = (0), by Corollary 4.9 and Theorem 3.8. Recall from Section 3.5 that
le4*| < e'4' for t > 0. It follows from this fact that
t
|p(2)| < eA! ae val el Altt—s) ons ds

== sale lpr
Picletaije 6ers0F al
Since every nth order equation (5.1c) can be written as a system (see Section 4.1),
we have an immediate corollary to this theorem.
Corollary 5.1. Assume that b is positively integrable and of exponential order at
infinity. Then every solution of

x” + dix + ey =e imei —- anx = b(t)

is of exponential order at infinity, hence has a Laplace transform.

130 Solving linear equations with Laplace transforms Shi

In view of Theorem 5.1 and its corollary, we may now turn to the problem of
computing the Laplace transforms for solutions of linear equations. The fundamental
tool for doing this is the following theorem.
Theorem 5.2. Let ¢,¢',...,¢”» be continuous and of exponential order at
infinity and let ¢” be piecewise continuous. Then, for 1 <r <n, £o” exists and
is given by
SHO) SN LOO) SNe ak aid Ne,
where

Ce ¢(0), (Co) ¢'(0), coon (Go = 6” (0).

Proof. The proof is by mathematical induction. Fix an integer value k for r, say
r = k, where 1 < k <n — 1, and define y(t) = ¢ (4). By hypothesis, there are
constants M and 7 such that |y(t)| < Me™ for0 < t < +o. Assume that Red > 7.
Then

i;We dt = We +» | we at (5.1d)
0 0

We note that |y(r)e—*”| < Me~“®* *—* and that £Y(A) exists for Re \ > 7. Thus,
letting 7 — +o in Eq. (5.1d), we have

LY’) = —¥(O) + ALP). (5. le)

If k = 1, then Eq. ([Link]) is the assertion of the theorem. Thus the basis for the
induction is established. Next, suppose that the theorem is true forr = k,1<k<
n— 1. Then £¢“ exists and is given by
LN) SIN LOOM AN Cie ee (5.1f)
By Eq. ([Link]), £6%T (A) exists for Re > 7, and
LH*TPVCQY) = ALGOA) — cet. (5.12)
Combining Eq. (5.1g) and Eq. (5.1f), we conclude that

BOCK) SE NEPG(N) NC ee ce) eon

that is, that the theorem is true for r = k + 1. By the induction principle, it is
therefore true for 7 —" Ieee nell
Now let us see how this theorem is applied.
Example 4. Take the Laplace transform of both sides of the equation

ete (5.1h)
to obtain
72x — Ax(0)' = x10) 28x 10:
 Definition of the transform 131

It follows that

(? + w”)£x = dx(0) + x’(0)

and

a (0)
Wena (0)
(x2 + w?)

I x(0)L[cos wt] + x'@)—— Lf[sin wf]

= L[x(0) cos wt + x’(0) sin wf].
(6)

It turns out that the equality of these Laplace transforms implies the equality of
the untransformed functions. Thus

x = x(0) cos wt + x’(0) sin wt

(69)

is a general solution of Eq. (5.1h). This general solution is more convenient for solving
initial value problems at t = 0 than the general solution x’ = c, cos wt + cg sin wt
(with c; and cg arbitrary) since x(0) and x’(0) appear explicitly in it. ||
Applying Theorem 5.2 to Eq. (5.1b), we obtain
ALx — x(0) = ALx + Lb(A) or (A — AI)£x = —x(0) — Lb).

If \ is not an eigenvalue of A, then det (A — AJ) ¥ 0 and one may write

£x = —(A — dT)“ (x0) + £bQ)). (5.1i)
Recall from Section 3.11 that
Fe BOD)
(A-My'=—>>
y= 50)
where p(\) = det (A — XJ) is the characteristic polynomial of A and the element
b;;(\) in the ith row and jth column of B(A) is the cofactor of the element in the
jth row and ith column of (A — J). Equation (5.1i) can therefore be written in the
more meaningful form
&x= _— BQ)
75) (x(0) + £b()). (5.1)

Example 5. Theorem 5.2 applied to the system

(fier een
gives

(2 )-FO)-(t e+e)
 5.1
132 Solving linear equations with-Laplace transforms

Po) te] -b8]-ei 0

from which it follows that

Since

Pu Ge?
ctl NaaieadSE:

the Laplace transforms £x and Ly are given by

2? — 2r+ 2

1-vr -l
Ve co thie le lane (lea “pes In): |
Ly (2 Sn De) re ee
Applying Theorem 5.2 to the nth order equation (5.1c), we find that

iN yhbleak aac Marae eal he hep ot ia eee re

LX =
NM” + ay) + +++ + an_ib + an
(5.1k)
where
(0) = 6p), =a ae ee ee Olea
Notice that the denominator in Eq. (5.1k) is precisely the characteristic polynomial
p(d) for the differential equation (5.1c). If we define aj = 1 and co = O, then we
can write Eq. (5.1k) in a more easily remembered form

pAjsx = >> e auen-t) "+ Lb). (5.11)

=0 \k=0
Example 6. The Laplace transform for an arbitrary solution of

x!” + 3x" + 3x'+x=0 (5.1m)

satisfies

ox = x(O)\?
XOM ++ C'O SO) Sr
(x'(0)
+ 3x(0))r (0) x'(0)+ 3x(0
+ 3x'(0) 3x(0))| 1 Gan)
A? + 3A*7 + 3A4+ 1

EXERCISE 5.1

1. Compute the Laplace transform of a general solution for each of the following differential
equations.
a) x” + 4x’ + 3x = 0. b) x + 2x 4+x = 0.
c) x!” + 5x! + Tx! + 3x = 0. GQ) 207 i Be aa X eet Os
@) x tox! ee on) fy x09) + 2x!" + 3x" + 2x’ + 2x = 0.
 Solving homogeneous equations 133

ale
2. Compute the Laplace transform for each of the following functions.
a) t b) sinh¢ c) cosh ¢
d) e‘ sin 3¢ e) e?’ sint f) e?' cos 3t
g) e°' cos 2t h) e“ sinh ¢ i) e sinh (if)
j) e?¢ cosh t k) e cosh 2r 1) sin? ¢
m) e*' sin? ¢ n) te?
a ea
DfO={y ig aap ae ae
‘yes ere)
if = ae&.
nso b | a>b>0.
0 if |f—al>ob
qg fO=n+1, feet Sneed. 0.
3. Show that the function e” is not of exponential order at infinity.
4. Suppose that f is piecewise continuous and of exponential order at infinity. Show that
Lf() — 0 as |A| — +0. [Hint: Use the triangle inequality for integrals.]
5. Suppose that £f exists and define g(t) = f(at), a > 0. Show that

Le) =a LG").

6. Suppose that f is of exponential order at infinity and that {> t~!f(d dt exists. Define
g(t) = t~! f(t). Show that Lg(A) exists for real \ and
+00

Lg() a Lf (u) du.

[Hinte™ /t-= Sr? ent du fOr Nes 07 0,1

7. The function sinc (t) = (sin f)/rt is called the interpolation function. Use the result
of Problem 6 to find £ sinc (A).

5.2 SOLVING HOMOGENEOUS EQUATIONS

Let us find a general solution for the equation

x” +-3x" + 3x/+x=0.

By Example 5 of the last section,

. x(O)\? + (x'(0) + 3x(0))d\ + (3x0) + 3x’(0) + x0) |

Qt (5.2a)
134 Solving linear equations with Laplace transforms See

The partial fraction decomposition theorem guarantees that there are constants
A, B, C such that
x(0)r2 + (x0) + 3x(0))A + (3x(0) + 3x’) + x’(0))
a+ DF
A B om
= Oca > Grp) a OMeaans
Taking a common denominator on the right and equating the numerators, we find
that
x(O)\? + (x'(0) + 3x(0))\ + (3x) + 3x/(0) + x’(0))
= AX? + QA44+ B+ (A+ B4+C).
In order that this equation be an identity in \, it is necessary that
x(0) = A, —x')4 3x0) = 24 + BB, 3x0) 3x0) x70) = Ae ec
that is, it is necessary that
A=x(0), B=x0)+x(0), C= (x(0)+4+ 2x0) + x’()).
With this choice of A, B, C, we have
x(0) x(0) +: x10) xO) 2x OPE Oy
SX + (5.2b)
Atel Chee Oss 12
To find x, we shall first identify the terms on the right of Eq. (5.2b) as the Laplace
transforms of known functions. This is accomplished with the aid of the following
theorem.
Theorem 5.3. Let k = | be an integer and let f be positively integrable and of
exponential order at infinity. Then g(t) = t*f yhas a Laplace transform &g given by

£g(d) = (—1)" a~ of).

Proof. Since d*e—**/dy* = (—1)*t*e—™, we have
+2 r = d®e =
£8(0) = [ eMf() dt = i (Si ep ae
It is permissible to interchange the order of differentiation and integration since f is
of exponential order at infinity. Thus
oo

£e() = (-1)' a ya = -'L0). |

To apply this theorem to Eq. (5. eb) recall that Aes = (+ 1)7'. By the
theorem, £[te~']= (A + 1)~? and g[t?e—]= 2( + 1)~3. Thus
£x = x(O)g[e~] + (x) + x'Oelte] + 4(x(0) + 2x0) + x") Lfe2e~4]
L[x(O)e~ + (x(0) + x’(O))te™ + 4(x(0) + 2x'(0) + x'’(0))t?e~4].
 Solving homogeneous equations 135

We should like to conclude from this equation that

x = x(O)e* + (x0) + x'(0))te~ + $(x(0) + 2x’(0) + x(0))12e~*. (5.2c)
That this conclusion is correct is a consequence of Lerch’s theorem.
Lerch’s Theorem. If &f = Lg, then f(t) = g(t) for every point t at which both f
and g are continuous.
Given a function ¢ of \, a function f of ¢ such that £f(\) = (A) is called an
inverse Laplace transform for ¢, and one writes f = £&'¢.
Why would one want to solve the equation x’ + 3x” + 3x’ + x = 0 with
Laplace transforms anyway? After all, it is very easy to invoke Theorem 3.10 and
immediately write down a general solution
x = cye $+ cote* + cat?e—. (5.2d)

The answer is that Theorem 3.10 does not explicitly specify that cy = x(0), co =
x(0) + x’(0), and cz = (x(0) + 2x’) + x’(0))/2 as Eq. (5.2c) does. To deduce
this from Eq. (5.2d), one must differentiate the equation twice to find x’ and x”,
set ¢ = 0, and solve the resulting algebraic equations for c1, C2, Cg in terms of x(0),
xO) ex(0),
The labor-saving advantage of the Laplace transform method of solving initial
value problems for homogeneous equations is the elimination of the need for com-
puting derivatives of the solution at zero. To see this, consider the equation
Fee ei aoe eg ee le 0: (LHC-n)
We saw in Section 5.1, that the Laplace transform £x of its general solution satisfies

DQ)e =) CN (Ca + ac) Ne (Cy ain FE a_i),

where
PO a= tC a) een)
is the characteristic polynomial of the corresponding homogeneous equation. The
partial fraction decomposition for £x will be of the form

Ya Vmy1 Vik Vong

Ce ROUEN Ha 1 OS oD Q— wo
= < G0. ig Se see. a see eae -- Be
a a

One evaluates the 7;,’s in terms of cy,...,¢n- It then follows immediately from
Theorem 5.3 and Lerch’s theorem that
k — Ay
ae (Vir = Yort aiaeanale ae Vege @ Ye ‘
r=1

This equation has precisely the same form as the general solution (3.8d). The dif-
ference is that here the Y;;’s are evaluated in the process of making the partial fraction
decomposition rather than by substitution, as was the case in Section 3.8.
136 Solving linear equations with Laplace transforms SZ

We saw in Section 5.1 that any solution x = ¢(f) for x’ = Ax satisfies

where c = (0) and

PO) = Oe) ee Oe
is the characteristic polynomial for A. Making the partial fraction decomposition

50) _ 55] Bin 0 wa ak is |:

PO cn ee.

we conclude from Theorem 5.3 that

k
+ Borel ee,
eo(t) = yy (Bye tees
r=1
This equation is the same as Eq. (3.7g) with c;, = B;,c.

EXERCISE 5.2

1. Compute an inverse Laplace transform for each of the following functions.

» 1 b) A+ 1 ) 3A + 2
EE OED. oe ye)
1 SP Il 1
Oe = Anas 2 Ce, are ) ey4
2A — 1 24 — 1 1
2) 244 ear ) ep 2 + APD
i)3A. + 6A
=
A3 + 3A2 + 4A
+4
+4 2
jee1 ———
At == 27-1

» Eo pa a! 2 al n2

2. Compute Laplace transforms for the following functions.

E) (@) = &. b) f@ = re".
Oo) (@ = f sitep. ab) KO) = FP CON ay.
ie 1 0 eee
CI (aay (2007) eI a aaa 2s
0 ve i SS

3. Combine the initial conditions listed as (a) through (j) below with the corresponding
differential equations in Problem 1 of Exercise 5.1 and solve the resulting initial value
problems.
a) x(0) = 1, x’) =. b) x)= 25557 ©), =" 1;
c) x(0) = 1, x’) = 1, x”) = 1. d) xO) = 0,7 x’) = 1, x") = 0.
 Laplace transforms for some special functions 137

e) x(0) = 1, x’/(0) = —1. f) x(0) = 0, x’(0) = 1, x’) = 0, x'"(0) = 4.

g) x(0) = 2, y(0) = 0. h) x(0) = 2, y(0) = 1.
1) x(0) = 2, y(O) = 3. J) uO) = 1, vO) = 0, w0) = 0.
4. Compute an inverse for each of the following transforms by applying Theorem 5.3.

1 x paps
Cero) >) O2 + 9p o 02+ 97
4) =I yh e210 fi 3
Oe Nene Way aay TE 293

5.3 LAPLACE TRANSFORMS FOR SOME SPECIAL FUNCTIONS

In this section, we shall compute Laplace transforms for several special functions.
The Heaviside Unit Step Function H, defined by

bo fors 7e0:
Ol 0 ior PEC

is illustrated in Fig. 5.1(a). Its Laplace transform is LH(\) = 1/\. The modified
step function A,(t) = H(t — c), c > 0, has its discontinuity shifted c units to the
right of the origin. Its Laplace transform is £H,(\) = e—/d.

H(t) Hi ((ti—=s0)

b——__—_—

t
|
oe
|
|
(a)
Figure 5.1

Now let f be any function which is defined for —0 < t < +o asin Fig. 5.2(a).
The function f,(t) = f(t — c) is defined for all t, and its graph to the right of c is
congruent with the graph of f to the right of the origin (Fig. 5.2b). In working with
Laplace transforms, however, one is usually not interested in the values f(t) of the
function f for negative ¢. Thus it is convenient to introduce the function H.f.
with values

fat) = H(t — Of — 0) = ffeet oe

It is defined for all t ¥ c and it preserves the properties of f(t) for t > 0.
 Solving linear equations with Laplace transforms BES)
138

AAC) f=)

Figure 5.2

Theorem 5.4. If &f exists, then Hf. exists and is given by

LHF (A) = ec“ SfQ).

Proof
$00 +o
if er HG of — ce) at i eo VG — ea
0 c

+00

ond / e—F(u)du I= e~’Sf(r). ||

0

Example 1. A block having a mass of | kg rests on a frictionless surface and is con-

nected to a vertical wall by a linear spring with stiffness coefficient 1 newton/m
(Fig. 5.3.). At time ¢ = 0, the block is at rest in the equilibrium position, and it is
put into motion by a force of 1 newton acting to the left for 1 second only. Let us
determine the subsequent motion of the block.
The force function F can be succinctly described in terms of the unit step function
by
F@= —H@)+ H¢ — 1).

The displacement x of the block from equilibrium thus satisfies

x’ +x = —H(t)+ AC — I), x(0)- == (0) 30; (5.3a)

The Laplace transformation applied to this equation yields

72x — Ax(0) — x0). + Sx = —Xa 8 te

Thus
asiat |
es ig
os:
fn a
MN <—+—_+- F(t)

Figure 5.3
 Laplace transforms for some special functions 139

To determine x, we note that

1 Se lgal =e
NOIO) Te See
Thus
aN =
eee CaN ee d
mn Lr? 2» 142
GHa(N\)s en &C = SRO) COD
where C(t) = cos¢ and H,(t) = H(t — 1). By Theorem 5.4, however,
e*SC(A) = £H,C,()).
Thus
£&x = L(A; — H+ C — A,C,]Q)
and
x = A(t — 1)— A(t)
+ cost — A(t — 1) cos(¢t — 1)
Pee ee COs.7. Lote Os< t= 1h
~ (cost
— cos(t — 1) for oe 81)
Nonhomogeneous linear equations in which the forcing term is a periodic function
occur frequently in applications. The next theorem indicates a rapid method for
computing the Laplace transform of a periodic function.
Theorem 5.5. Let b denote a piecewise continuous scalar- or vector-valued function
with minimal period T > 0. Then &b exists and is given by

foe ™b) dt.

Lb(\) = are Rens .0;
1 —+€,
Proof
oo stg (n+1)T
£b(d) = / e a(n) dt = >> . e b(t) dt
10) n=0/n

T
T iM e 42Mp(r 4+. nT) dr
T
=: Ss ol e“'b(r) dr
0

== e—b(r) dr 2 (Cea
A n=0
140 Solving linear equations with Laplace transforms 5.3

Example 2. Let f denote the periodic step function illustrated in Fig. 5.4. Then
T/2 —yt ft! Nt

ofr) = E= iy (2 dt —
eae ie Z dt

ee ee ee N= Go
—AT/2)2 —AT/2
F tanh(7). i
NC cane) Mise ene)
One can also use Theorem 5.5 to compute the Laplace transform of the sawtooth
function 6 illustrated in Fig. 5.5.
The next theorem describes a more efficient method for doing this, however.

i)

E >< »— -— »—— -—

$$
____—_—_—__—_» |
TE
2
—H -—c —— -— -— »—

Figure 5.4

Theorem 5.6. Let f be a positively integrable function which is of exponential order

at infinity and define g(t) = ik f(r) dr. Then &g exists and is given by
Soh paar Cay
Proof
+ t

g(\) = / one i f(t)

dr dt
0 (0)

+0 +o
sl f(r) / edt dr
+0

= )} . f(rje—™ dr = rf (a). ||

b(t)

Figure 5.5
 Laplace transforms for some special functions 141

Example 3. Consider the sawtooth function b illustrated in Fig. 3.07 FOL ts

T/4-—-n1/2, n = 051,25... .5 6”) exists and

1 for Osh 1/4:

b(t) =4—1 for TAT <a 81/4,
1 for STA AT:
Using Theorem 5.5, one computes
_ {AT /2 —\P}4
SQ) a4 — de tae le
By Theorem 5.6,
l Rees
Lb(A) = Cae te ey || (5.3b)

If f and g are integrable, one defines a new function f * g, called the convolution
of f and g by

1 AON | roc — 1) dr.

A basic result for the Laplace transform of a convolution is given by the following
theorem.

Theorem 5.7 (The Convolution Theorem). Let f and g be positively integrable and
of exponential order at infinity. Then &f*g exists and

Lf*Q(A) = LFA): Lg).

Proof
oo t

Lf*Z(r) = i oar if
f@)g(t — 1) drat

+ +x
= i! ro | e—'g(t — 7) dt dr
0 T

fe fo
f(r)e—* ‘| eg(u) du dr
0 0

Ef (r)- Ler). ||
Example 4. Let £A(\) = 1/X2(\2 + 1). Since £[4]Q) = \~” and [sin J) =
1/0? + 1), LAQ) = S[sin JO): L[JQ) = eff Gin) — 7) dr]Q). By Lerch’s
theorem,
at

AQ) = ik(t —7)sinrdr =t—sint. || (5.3c)

 5.3
142 Solving linear equations with Laplace transforms

It is not very hard to verify that f * g = g* f. The integral (5.3c) in the last
example is harder to evaluate when written in the equivalent form
t

/ 7 sin (t — 7) dr.
0
It is thus worthwhile, when setting up a convolution integral, to give some thought
to the order of convolution.
We conclude the section with a dual for Theorem 5.4.

Theorem 5.8. Assume that &f exists and define g(t) = e“f(t). Then Lg exists
and £g(\) = £fQ — a).
Proof
+e 2
£g(r) = i e g(t) dt = [ eo
OF dt = Sf — a). ||
Example 5. Since £[sin ¢](\) = 1/(\? + 1), £fe“ sin JQ) = 1/[A — a)? + 1]. ||

Example 6

g[t*e™ sin f(\) = (—? © sper! sin f(A) = (—1)* d : - ||

a dyk aXe (Xia)

EXERCISE 5.3

1. Compute inverse Laplace transforms for each of the following functions.

=), —2x

4) ee b) ev Ory ) RSR Saas

XZ 2h 4-2 AZ + 24 + 2 A2e%
+ 4er
oe Hes 710%
d) 4-1 . AZ(A— 1)2 ») 246 4) 29
2. Let A and c denote positive constants. The function f(f) = AL[H() — H(t — o)] is
called a gating function. Compute its Laplace transform.
3. Let A and m denote positive constants. The function f(t) = mtH(h/m — H(t) is
called a ramp function. Compute its Laplace transform.
4. Given below are several functions f and corresponding intervals 0 << ¢t<T. Let b
denote the function of period T which satisfies b() = f(),0 < t < T. Find £6 in each
case.
NyG@=sni O<'<e
Def @) =" {sin th Olen
ce) f@® = A[A() — H(t — od], 0 < t < c, where and care positive, T = 2c.
d) f(@) = mtH(h/m — t), 0 < t < h/m, where A and m are positive.

5. Compute an inverse Laplace transform for each of the following functions.

S leven tr b) p ™
iN Sas NOS we
 Solving nonhomogeneous equations 143

oN
Cee Ee A |
ACL — e72A) ) wOEEe)
e
e) ————_
tanh d f) Cuctanhi(he
ms =>)
mn CD — 2/2
6. Use Theorems 5.6 and 5.7 to find inverses for the following transforms.
1 1
xQ + 3) 02 +1)
a Ny —————$_______

1 1
©) 3202 $1 D QP 4 12
=3) =4),
e —4
ys
ACA2 + 1)
JN ACL — e744) © 2
ee
1 1 —
g) ACAZ
+ 2A + 1)
h) AZ (A2Cee)
+ 2A+ 1)

5.4 SOLVING NONHOMOGENEOUS EQUATIONS

One can use Laplace transforms to efficiently solve nonhomogeneous linear equations
when the forcing function is such that variation of parameters is difficult or un-
pleasant.

Example 1. Let us try to compute a general solution for the rather innocuous looking
equation
x’ = —x + did), (5.4a)
where 6 is the triangular wave function shown in Fig. 5.5 (with T = 4). This equation
was briefly examined in Example 2 of Section 4.4, where it was shown that there
could exist at most one periodic solution (also see Problems 1(j), 9, and 12 of
Exercise 4.4).
A general solution
t

x=e'et+ il e°b(s) ds (5.4b)

0
of Eq. (5.4a) is easily found. To use this general solution, however, one needs to
evaluate the integral, and this is a rather complicated task using classical methods.
Here, we shall solve the differential equation by finding a particular solution x =
$(t) with the aid of Laplace transforms and adding this particular solution to a
general solution x = e ‘c of the complementary equation x’ = —x.
Taking Laplace transforms in ¢’(t) + ¢(t) = b(2) gives
PAF rch) «

and from Example 3 of the last section,

Dea a DYany
WHOS a As eo")
144 Solving linear equations with Laplace transforms 5.4

ae Gis—aeme) rs lee ee eee

A2(A ]
Te
To compute ¢ from £4, notice first that the denominator in Eq. (5.4c) is 1 — ES
Theorem 5.5 says that the Laplace transform of a periodic function g has the form
eocye Lose
T At
Oe
lee
where T is the period of g. This suggests that ¢ might be periodic with period 4 for
some particular initial value a = ¢(0). Suppose it is in fact periodic. Then the
numerator in Eq. (5.4c) satisfies
4
aye — Cie een lem eee
I CONE al ee aie) A+ D
and, defining f(1) = ¢(t) for 0 < t < 4, f() = 0 for t > 4, it follows that
4
Jenssen toy i De sini em cae
Sf) = i e™'$(1) dt = at ery \2(A + 1) (5.4d)
The theorems in Section 5.3 can be used to deduce f from Eq. (5.4d). Notice
first that, if h(t) = —1 + 4+ e, then
1 1 1 1
LAA) = =a et x
+I MAF
By Theorem 5.4,
ELLf(D] = akf[e] — a&[A(t — Ye'*4] + SAD] — 2£[A(t — DAC — 1]
+ 2£[H(t — 3)hCt — 3)] — £[ACt — 4A(t — 4)]
and, by Lerch’s theorem,

f(t) = ae — aH(t — Ye*** + AC — 2A(t — DAC — 1)

ASOT (6 3G 3) eH ee ee C7)
for ti< 4514 1F 3. Sinces/.@¢)==0.10n f a4. it olows that

0 = ae — ae *** + h(t) — 2h(t — 1) + 2A(t — 3) — A(t — 4)

= e"[a(1 — e*) + 1 — 2e + 2e? — e*]
or, equivalently, that
=
Qe lim Pete ae.
eae . (5.4f)
Thus
—-l+t+e%+ ae for Os ole
Of) =\3 —t ee + Ger — aes for I< 3. GAe)
—S5+t+e%s ae — Qett! 4 Jet 3 for St<aieed
 Solving nonhomogeneous equations 145

Since @ was assumed to be periodic, its values for t > 4 can be determined from the
relation ¢(t) = ¢(t + 4) beginning with 0 < ¢ < 4.
We have actually shown that if Eq. (5.4a) has a solution x = ¢(¢) of period 4,
then it is given by Eq. (5.4g), where a has the value (5.4f). One checks by substitution
that the function ¢ so defined is actually a particular solution. A general solution of
(5.4a) therefore has the form x = e~‘c + (ft), where c is an arbitrary constant. ||
Laplace transforms also provide an efficient solution method for nonhomogeneous
equations when the method of annihilation is applicable but very complicated.

Example 2. Let us compute the solution x = ¢(t) of

x = 2 + 2x = fe’ sin?

which satisfies ¢(0) = ¢’(0) = 0. Note first that

1
Sint) = 544° I
£fe’ sin (A) = eerceeae
and
ics
£[te’ sin 7](\) =
d
5 |p Cn
1 eee a
201)
[A2 — 2d + 2)? :

Consequently,
ge 10.
8 RE erroneous ~2

Now let f satisfy

LF) = [\? — 2A + 27°”.
Then
Lard
£60) = — 5 HES)
and, by Theorem 5.3,
g(t) = at f(t).
Since
£fe’ sin t](\) = DS EE) et

however, it follows from the convolution theorem that

7OQ= / [e” sin r][e’” sin (t — 7)] dr

0
ile Sint
se A — lot
ste cost.

Hence

o(t) = ite’ sint — 3t7e’ cost. ||

146 Solving linear equations with Laplace transforms 5.4

EXERCISE 5.4
1. Use Laplace transforms to solve the following differential equations.
BY ae? ia bail. beaN 0s
ec) x” +x = 1 +7. d) x” +x =342¢42’.
©) xe! = 9G =e" ly 6eY Se ge = 172%
8) x! ey = 17 4 Be", h) x” -2x = fsin 27.
)) 567 abe SO ial ye j) x’ +x =rt+e'sint.
k) x!” + 6x" + 9x’ + 4x = te~. 1) x6" Oy de le Seale ae
2. Let f denote the periodic step function shown in Fig. 5.4 of the last section. Solve
oe Se SE KOE
3. A1kg rectangular block is resting on the top of a cart and is connected to one end of the
cart by an unstressed spring with stiffness coefficient 1 newton/m (Fig. 5.6). Friction between
the block and the cart may be neglected. Suppose that the cart is moving at a speed of
1 m/sec to the right and that it is uniformly stopped in 1 second. Describe the motion of
the block.

es

| LAW
©) ©
Figure 5.6

4. The circuit depicted in Fig. 5.7 represents a full-wave rectifier and filter circuit. Its
purpose is to convert 117 V alternating current into an approximately constant voltage V
across the resistor R. Compute the Laplace transform for V if the voltage from node a to
node b is E|sin wt|,w = 1207. [Hint: The currents x and y satisfy the differential equations

eee R Eo
a Txt zy t Z/sin od,

; 1 1
Ee ites Page
and V = —Ry.]

\e—=Elsin 1207¢| =~ =

Figure 5.7
GHAPTER.6

POWER SERIES SOLUTIONS FOR

LINEAR EQUATIONS

6.1 ANALYTIC FUNCTIONS

If the coefficients of a normal linear differential equation

x Aa) x (6.1a)
or
eat) 2 oe an()x = 0 (6.1b)
can be expressed as power series on an open interval, then the equation’s solutions
can also be expressed as power series on the same interval. Some of the techniques
involved are presented in this chapter.
Example 1. Consider the initial value problem
x + 2tx’ + 2x = 0, x0) = 1, x0) =0 (6.1c)
and suppose that its solution is expressible as a convergent power series x =
De ea esince x = > | kat’ © and x = 52, k(k — law, it follows
that

x! + 2tx’ + 2x lI aek(k — 1)a,t"~=a a kat? + Si Dp

> l| bo k=1. k=0

II
Ms
Il °
(kK + 2) + Dazaot* + Dy2kayt DyDayar

ll
=iMs
Il
(K+ MK + Danse + 2k + Daa =0.
as

The last power series is nets zero by assumption. All its coefficients therefore
equal zero, and the a;’s must satisfy the recursion formula
ee ae at Wit? ete wih 20) (6.1d)
147
148 Power series solutions for linear equations 6.1

Since x’/(0) = a; = 0, it follows from Eq. (6.1d) that a, = 0 if k is an odd integer.

With this case excluded and k = 2m > 0, Eq. (6.1d) takes the form

Ag¢m+1) = —dom/(m a8 We m ze 0. ([Link])

This equation applied recursively to itself yields

Agm+1) = aly -E 1)!, 1 > 0.

Consequently,
; (6.1f)

and we have shown that if the solution of the initial value problem (6.1c) is repre-
sentable as a power series, then that series is precisely (6.1f). Since the steps in the
derivation of the series are reversible, one can show that it represents the desired
solution by merely demonstrating convergence with the ratio test. Note, incidentally,
that x = e~ is the solution of (6.1c) and that (6.1f) is its Maclaurin series. ||
Recall now that function f is said to be analytic at a point to if its values can be
expressed by a convergent power series >-%_ a4(t — fo)* on an open interval around
to. It is shown in calculus that if f is analytic at fo, then

(Gy = DS feo)
k!
(t — to)",
=0

that is, a, = f"(to)/k!, k > 0. This means that the power series expansions for
analytic functions around various points fo are precisely their Taylor series around
the same points. The function f is called analytic on an interval if it is analytic at
every point of the interval.

Example 2
i) If f@ = sint, then f is analytic on the interval —wo < t < +o.
ii) If f(@) = 1°/%, then f is not analytic at the point ¢ = 0. Since it does not
have a second or higher derivative at ¢ = 0, it has no Taylor series around
t= 0.
ii) If f@ = eV", ¢t ~ 0, and f(0) = O, then Jf has a Taylor series around
t = 0: namely,

O Of 4077/2) ices 42079/k) 2 oe =

Since the functional values f(t) do not equal the sum of the Taylor series on
an open interval around t = 0, f is not analytic att = 0. ||

With the definitions above in mind, then, we shall say that a vector function f
or a matrix function A is analytic at f9 (or on an interval) if each of its components or
 Analytic functions 149

elements is analytic at ¢ (or on that interval). If such is the case, we write

io) (k) ee) k

OO
ey ak
for tin a neighborhood of fo.
Now let us regard n scalar power series
(eo) wo
k k = :
Dea Cn tg) es ts to) aay).
k=0

as a vector power series

ioe)

Dd ax(t — to)" (6.12)

k=0
where a; = [ayx,---,@nt]’, k > 0. The radius of convergence of the series is taken
to be the smallest radius of convergence of any of its component series, and it can
be found by examining the component series separately or by applying Lemma 6.1
below. If the series (6.1g) has a nonzero radius of convergence R, then each of the
component series converges to a value f;,(t) for |f — fo| < R and one calls f(t) =
[f1) +---+ fr(D) the sum of the series. Analogous remarks can, of course,
be made for matrix power series )-%_9 Az(t — fo)".
With regard to computations involving vector and matrix power series, let us
note that f above may be integrated and differentiated within its interval of con-
vergence |f — fo| < R by performing the desired operation on the terms of the
series. If g(t) = 02-9 bs(t— to)” for |t — to| < R, then f(t) = g(t) for these
values of ¢ if and only if a, = b; for all k > 0. It is shown in the calculus, that if f
and g are scalar-valued, then
ee) k
fHzQ— >» (bsarbs-o)( — to)’, |t — tol < R. (6.1h)
k=0 t—=(

This series is called a Cauchy product and, if

AQ) =3>) Ant =) > [ft ol = R,

k=0

then one can show that

leo} ik

Aa) = [=O
(Y Adi )te=to)!
N70

for |t — to| < R. We conclude the preliminary discussion with a lemma which is
useful in determining the interval of convergence for vector power series.
Lemma 6.1. Suppose the coefficients aj, of the series
iva)

Daealo— ty) (6.1i)

k=0
 Power series solutions for linear equations 6.2
150

satisfy an inequality of the form

k
(kT Napa |< ML a,| eee, (6.1j)
r=0

where M and L are positive constants. Then the series converges for |t — to| < L.
Proof. Apply inequality (6.1j) to itself recursively to obtain

lai] < M|aol,

M
|as| < ap (ll+ ML)|ao|,
ML(\ + ML)-:-(kK + ML) |aol .
|ax+4| <
(k + 1)! Lk+1
If M;, denotes the right side of the last inequality for k > 0, then the series
> 1-0 M(t — to)” converges for |f — to| < L by the ratio test. Specifically,

Me at) Sete aM vena

lim ; = jem aol
pce Ap (=e G rerio a Oe| L
Since |a,(t — to)*| < M,|t — fol”, the series (6.1i) converges by the comparison
test. ||

6.2 POWER SERIES SOLUTIONS AROUND ORDINARY POINTS

In this section, we consider the problem of expressing solutions of linear differential

equations as power series. Consider throughout the discussion a differential equation
Ke A(X, (6.2a)
where each element of A is analytic on an interval a < t < w. The interval (a, w)
will be called an interval of analyticity for Eq. (6.2a) and each of its points will be
called an ordinary point for the equation.

Theorem 6.1. If to in (a, w) is arbitrary, then any solution x = $(t) of Eq. (6.2a)
may be expressed in the form
ie)

ot) = > Ct — fo). (6.2b)

and the radius of convergence of this series is not less than the smaller of (to — a)
and (w — to). Further, $(to) = ¢o and the other c;,’s may be found by substituting
the series (6.2b) into the differential equation and equating coefficients of like powers
of (t — to).

Proof. Since A is analytic at to, one may write A(t) = ~_9 Ax(t — to)x, and the
matrix series converges for |t — to| < R = min {to — a, w — to}. Now let co =
$(to) and determine ¢;,, for k > 0 from the recursion formula
1 k
Chi = pare 7,Ue (6.2c)
 Power series solutions around ordinary points 151

If the series (6.2b) converges, then it must satisfy Eq. (6.2a) since

$(t) — A(t)$(t) I Ms kex(t — to)? — (= Ax(t — oS c(t — to)

k=1 c
k=0 k=0
C

7) o k
= ket — 1) — 5 ic Aci) — to)”
k=1 k=0 \r=0
k
( + 1 )ex41 — pe eC = Py = (I)
k= 0 i—

Convergence: Choose an arbitrary positive number L < R. Since

Dd Arlt — to)"
k=0
converges for |t — fo| < R, there is a constant M such that |A,(t — to)*| < M for
|t — to| < L and all kK> 0. In particular |A;| < ML~*. Now take norms on
both sides of the recursion formula (6.2c) to obtain

lex4i] < i ({Ax| leol + °-- + Aol lex) < ag eo “7b [eg|).

Multiplication of this inequality by (k + 1)L**! yields

k

(K+ 1" en43] < MLD Ler,

r=0

and it follows from Lemma 6.1 that the series (6.2b) converges for |f — fo| < L,
hence for |f — tol < R. ||
Example 1. The coefficient matrix of the system

lee ieee oh
can be written in the form A(t) = Ay + Ait, where
0 1 ee)
Ao= |i | and lk il

Let us compute the solution [x, y]” = S~%o ext” which satisfies the initial condition
x = 1, y = Owhen? = 0. The recursion formula (6.2c) adapted to this problem is

C1 = Alo, (k + Wenas1 = Arce-1 + Aotk = ee

o-E o=QHBI-Gh =-s8)s0l-

Thus

C4
152 Power series solutions for linear equations 6.2

and the power series expansion for the solution has the form

cll hles ais aes

eas hsot (6.2e)

t+ 5 =t
3 «©

Example 2. Find the unique solution x = y(t) of x” — 1/(1 — Dx = 0 that

satisfies ¥(0) = 0,y/(0) = 1. Herea = —1 <t <1 = wand df) = YAO)
for —1 < ¢ < 1. Set x’ = y and write the equation in the form

Dl-[ets abl
The matrix A(t) may be expanded in the form

0 1 m
A(t)
See
= .a :
~ Ag+ os:Axt,
k

where

Ao = F ql and A, = f 0 for kis ts

0 Ler

The desired solution x = y(z) is the first component of the vector (ft) = fo ext”,
where

Ch41 = B Aces, @ ike =20imacs = OMe

ae

Performing the algebraic computations for k = 0, 1, 2, 3, we find that

co= [i], = {9}: =| i): AL «=|

cto (ebeli Be
Thus

-| i+ 13/6 + 74/124 -
1+ 42/2 + £3/3 + 18/24.
 Power series solutions around ordinary points 153

The solution y then has a convergent power series expansion for —1 < 14 < 1
which begins
Wt) = t+ 29/64 4/12 + 71°/1204+---. ||
The example above illustrates more than Theorem 6.1. First, one frequently
cannot deduce an explicit formula for the coefficients of the solution series and one
must then content himself either with knowing only finitely many terms of the solu-
tion series or with working with a recursion formula. Second, it is unnecessary labor
to convert a scalar equation

rate ee e a(t) xi—=0; Oe hic, (6.2f)

with analytic coefficients a,,..., a, into a system in order to express its solutions
as power series. The scalar version of Theorem 6.1 is presented as a corollary.

Corollary 6.1. If ty in (a, w) is arbitrary, then any solution x = y(t) of Eq. (6.2f)
can be expressed in the form
io)

V(t) = DO
=0
ex(t — to)", (6.2g)
and the radius of convergence of the series is not less than the smaller of (to — «) and
(w — to). Fork = 0,...,n — 1, cy = W(to)/k!, and the other c;s may be found
by substituting the series (6.29) into the differential equation and equating coefficients
of like powers of (t — to).

Example 3. We use Corollary 6.1 to find the unique solution x = y(f) to the initial
value problem x” + (1/t)x = 0,0 < t < +00, x = 0 and x’ = 1 when t¢ = 1.
First, we express the function 1/t as a power series in (1 — ¢). Elementary
facts about the geometric series guarantee that

]
C2) iGeet 195 10S
AME
>

A solution of the form

X= eh), co
= 0, cy)
=1
k=0
is assumed to exist and is substituted into the equation

lee (= (SiG = D) 3 = 0.
k=0
This procedure yields

- k(k — I)ex(t — 1)°-? + bs (—1)*@ — D'\(d c.(t — »')= 0.

k=2 k=0 k=0
 Power series solutions for linear equations 6.2
154

Thus
oo 7) k

Dk + Yk + Vero — DY + »
=)
(os
“=
("ei — 1) =
k=?

In order that the coefficients of like powers of (¢ — 1)* be equal, it is necessary that

Ch2 = ee a ef sb (-1)~ k = 0. (6.2h)

The c;’s for k > 2 are computed recursively from the formula (6.2h), starting with
the initial conditions cy = ¥(1) = 0, c, = ¥’/C1) = 1. They are
af — 1 —
Cp = ON 96a Sag Cas, CCS = 5oe

Thus the power series expansion for Y(t) around the point fo = | is

Gh ee. 62%
6

_
ies

yvo=@-yn-E5V ye
ex 3 a 4 os 5

The student should compare this solution method with that suggested in Problem 1
below. ||
Several comments can be made about Example 3. First, no explicit formula for
the coefficients c; in terms of k was obtained. Had we not carried out the computation
to cg, it would have been tempting to assume incorrectly that c, = (—1)*/3- 2*~?
for k > 3. Second, the series (6.21) cannot be used to compute y(t) for t > 2 because
it does not converge outside the interval 0 < t < 2 (see Problem 3 below). These
difficulties do not occur for every differential equation. A classical equation for which
series solution works very neatly is discussed in the next example.

Example 4. For every constant v, the equation x’’ — 2tx’ + 2vx = 0 is called
Hermite’s equation. All its solutions can be expressed in the form x = > f~xo cxt*,
and Corollary 6.1 guarantees that each solution series converges for —a < t < +o.
The c;’s are determined by substituting the series into the differential equation

DS k(ke—=leet "a ket= =e Opa

k=2 k=1 k=0

= Sh 2 kee Dent= 2 Qkegt Qvegt”

> Il to =1 k=0

= (2-1-2 + 2vey)t°tio
+ (kK + 2)(k + enzo + 20 — k)ex}t*

Thus

Co lo77 £0 and Cky2 od 2v —

= eae k) i ¢t a
 Power series solutions around ordinary points 155

If x(0) = co and x’(0) = cy are arbitrary, then

J) TICS ye
2 ae ae Day 0? C= a — C2 ae,
a 2) Cos

eee 297C = 20 4)
Gass CGF
5 = 6! Co; ea 5

A n2'vy — 2)---(v + 2= 2k)

Cok ( 1) (2k!) Cos k we L;

and

one 20 — ate
1 ees C=
= eeé 2° a ue a 3). -

re _ ue20-5
; Le LeDe we = 3) = Se mee

ink VI 1 3 le
oe SI SS Geta et eS eet
A general solution is then

ts
= ¢oj1 =ipy |
ae | ze Ny Qk)!
Feet
See
OC 1)
eee OaQk +1)!
ee ek) tmap

where co and c, are arbitrary. ||

Once the unique solution x = @(f) to the initial value problem

= A(t)x, a Ee ON Ke Cg when (ea (6.2})

is known to be analytic, one can formulate a more natural way of computing it. The
idea is this: #(f), since @ is analytic at fo, is given by the Taylor series
oe) ifs
o®) ig) SS
2
2? (a k!
for |f — fo| < R = min {w — fo, t9 — a}. We substitute ¢(1) into the differential
equation and compute the successive derivatives by differentiating both sides,
obtaining
I! cy = (fo) = A(fo)€o,
2! Co = A’(fo)eo + A(to)e1,
3!¢3 = A (to)eo + 2A’ (toler + A(to)er,
a
Ato )exee
(Kk +1) cry = ait)Bie--
os
156 Power series solutions for linear equations 6.2

This computational arrangement has two advantages: it is easy to remember, and it

is not necessary to expand A(f) into a Taylor series at the beginning of the computation.
An analogous procedure applies to the equation

x™ + a(x Y +++: + ao(t)x = 0.

EXERCISE 6.2
1. If an equation with analytic coefficients has the form

Pyiag al OB EO) 0,

POY Sg a
it is frequently easier to compute solution series by first writing

bo(t)x” + bi@x’ + bo(t) = 0.

Solve the initial value problems in Examples 2 and 3 by writing the respective differential
equations in the form
a) =x— = 05 b) @ — 1)x” + x” +x =0.

2. There are given below two recursion formulas. Assuming that the power series
> v1 cx” which they generate are nontrivial, what is the radius of convergence of each
series if cg > O and lim, cx/cr41 exists?
1 1
a) Cr41 = rage b) cepa = cx + rageat

3. Prove that the series in Example 3 does not converge if f > 2.

4. Solve the following equations in which a and 5 ¥ 0 are constants and m > 0 is an
integer.

2) a2” S= pype = @). b) x” =— tx’ + bx = 0;

c) x” + at™t1y’ + bi™x = 0.
5. An example of an equation with a finite interval of analyticity about ¢ = 0 is
x’ + x/(1 — 1?) = 0. Solve the equation using power series in ¢. What are the intervals
of convergence? (The technique of Problem 1 is helpful here.)
6. Legendre’s equation
a7 2t are vy + 1).
epee epee Goa
illustrates the fact that a power series solution may converge for all ¢ even though the co-
efficients of the differential equation have discontinuities. Solve Legendre’s equation on
the interval |¢| < 1 and show that there is a polynomial solution P,(¢) if v is a non-negative
integer. What is the interval of existence for P,(r) as a solution to an initial value problem
pense! Alt rq = 0), Ali na =] 2%, airel aie rq = —V?
7. Solve the initial value problem

xt (eh sl) ee. x(0) = 0, x'(0) = 1.

 Regular singular points 157

8. Solve the systems of equations

Pll allyl
y t OjLy

her lb
oa ion 0 if |||8

9. Solve the system of equations

10. Solve the initial value problem

coe 0, y+ ty” — 2y +x = 0,
x0=0, xO@=1, yoO=0, yO=0, y"O=0.

6.3 REGULAR SINGULAR POINTS

Consider the equation
x” + a,(t)x’ + ao(t)x = 0, he se es Too) (6.3a)
where a ~ —o and either a; or dg is not analytic at a. The point a is called a
singular point for the differential equation.
Example 1. The differential equation
x at x/t = 0, 0<-t< +0, (6.3b)
which we considered in Example 3 of Section 6.2, has a singular point at t = O since
a(t) = 1/t is not analytic att = 0. ||
If it is possible to express a,(t) and ao(ft) in the forms

a= GopbO, a = Gopi bl.

where b, and by are analytic at t = a, then a is called a regular singular point. In
this case, it is customary to write the differential equation in the non-normal form
(t — a)?x" + (t — a)by(t)x’ + b2(t)x = 0, Waal 0): (6.3c)
Example 2. Multiplying Eq. (6.3b) by t?, we see that it becomes
(t — 0)?x" + (t — 0):0x’ + tx = 0.
The modified equation has the form (6.3c) with
o=— 0,0 = + 0,b1() = 0, and b(t) =f.

Thus a = 0 is a regular singular point. ||

A singular point which is nor regular is called an irregular singular point.

158 Power series solutions for linear equations 6.3

Example 3. The differential equation 1°x’’ + x = 0,0 < t < +o, has an irregular
singular point at f = 0. It is not of the form (6.3c). ||

In the rest of this chapter, we shall develop a method for expressing solutions of
the equation (6.3c) in terms of roots, logarithms, and power series 2/9 Cx(t — a)
around the singular point a. Note that the power series used in Section 6.2 were of
a different character in that they were series >)fo cx (t — to)* around ordinary
points f9 rather than about singular points a.
Why should we want to develop such a method? We illustrate some motives
with two more examples.

Example 4. In Example 3 of Section 6.2, we derived a power series expansion (6.21)

for the solution x = y(t) to the initial value problem x” + x/t = 0, x = 0 and
x’ = 1 when ¢ = 1. We showed there that the series converges for 0 < t < 2,
and the student has shown (Exercise 6.2, Problem 3) that the series does not converge
for 2 < t< +o. On the other hand, Theorem 4.2 guarantees that x = Y(t) is
defined on the whole interval 0 < t < +o. This is no paradox; it just happens that
the power series (6.21) equals the values of y only part of the time.
Suppose now that, for some reason or another, we wish to find Y(t) for t > 2.
If we could find a fundamental solution set ¥;, Y2 in a form valid for0 < t << +m,
then we could express y(t) as a linear combination, say Y(t) = cyWi(t) + Cove(d),
and this representation for y would be valid for the entire interval O < t << +m.
It would be of little use to expand y(t) and wW2(Z) in series about an ordinary point
ig > O because the radius of convergence R of the resulting series will be T = fo,
the distance from fo to the nearest singular point t = 0. Power series expansions
around the singular point avoid this difficulty. ||

Example 5. Ordinary differential equations with regular singular points frequently

occur in applied mathematics when one is trying to solve a boundary value problem
for a partial differential equation. Let us, for example, think of r and @ as plane
polar coordinates and try to find a function z = u(r, 6), 0 <r < 1,0 < @ < 2r,
which satisfies the Helmholtz equation

0°z 1 0z 1 07z
or? "a r or =" r2 062
+ 7z=0 (6.3d)
and the boundary condition u(1, 6) = 1, 0 < @ < 2x7. The desired solution z is
interpreted geometrically as a smooth surface in three-dimensional space lying above
the unit disc in the plane. Because of the cylindrical symmetry and the independence
of the boundary condition u(1, 6) = 1 from 6, it is not implausible to suspect that
Eq. (6.3d) might have solutions of the form z = ¢(r). For such a solution to exist,
it is necessary that

2¢"(r) 7) +7r°n? = 0. (6.3e)

4 o(7) +? Sr)
 Euler’s equation 159

Thus x = (7) is a solution to the boundary value problem if it satisfies the differential
equation
r?x!’ + rx’ + r?r2x = 0 (6.3f)
and the condition
lim x exists.
r>0

Equation (6.3f) has a regular singular point at r = 0. In Section 6.6, we shall find
a pair of linearly independent solutions x = Jo(Ar) and x = Y (Ar) for it in terms
of series around the singular point r = 0. The series expansions will allow us to
deduce that lim,.o,Jo(Ar) exists and lim,9, Yr) = —o. Thus z=
Jo(Ar)/Jo(A) will satisfy the boundary value problem. If a more general boundary
condition u(1, 6) = f(6), O < @ < 2m had been given, the analysis would have
been more complicated. Such topics are treated in Fourier series and partial differen-
tial equations courses. See Weinberger [10]. ||

6.4 EULER’S EQUATION

In the next section, we shall develop a method for expressing solutions of

(t — a)?x"” + (t — a)by(t)x' + bo(t)x = 0 (6.4a)

in terms of roots, logarithms, and power series of the form >°f~_» c.(t — a)* under
the assumption that 6; and bz are analytic on an interval a < t < w. The formulas
involved in the study of Eq. (6.4a) are simpler if the equation is converted into a
2 X 2 system. The variable change is different from that used in Section 4.1. We set
SH Xo = — a)x’. (6.4b)

Then xj = X2/(t — a) and

=[l — dM)’ -— b2(O)x

x6 = (t—a)x”’ +x
(f= a)
—_ (b= bi@ be Oi@)x_
-_ (t — a) (t — a)
The system x’ = B(1)x/(t — a), with x = [x1, x2]" and
0
BOE Ree 7s 1ae (6.4c)
is therefore equivalent to Eq. (6.4a).
In what follows, we shall consider an equation

x foyer
= 7 BOX (6.4d)
where B is a 2 X 2 matrix function which is analytic fora < t < w. The theorems
which are proved for this equation will then be specialized so as to obtain information
160 Power series solutions for linear equations 6.4

about Eq. (6.4a). Let us agree to write

B(t) = >> Bt — a)”, lt-—al<@m-—a.

k=0
We shall see below that the nature of the solutions of Eq. (6.4d) depends heavily on
the eigenvectors and eigenvalues of B(a) = Bo. An indication of the nature of this
dependence can be had by considering the equation for the case B(t) = B, a constant.
The equation
yf = Bx (6.4e)

is solved explicitly by setting (t — a) = e* and defining y(s) = x(t). Since

dx Gy GS GN eee
dt dsdt ds (t— a)
the transformed version of Eq. (6.4e) is
y’ = By. (6.4f)
If B has two linearly independent eigenvectors p; and py with respectively corre-
sponding eigenvalues \, and dg, then a general solution of Eq. (6.4f) is (by Theorem
4.12)
Y = cypie®® + copse’,
where c,; and cg are arbitrary constants. The corresponding general solution of
Eq. (6.4e) is
xX = cypi(t — a)! + copo(t — a),
where (t — a)*s = exp [dj In (¢t — a)] for j = 1,2. If, on the other hand, the
eigenvectors of B are all scalar multiples of p;, then B has a double eigenvalue )}.
We define py by
(A — NI)p2 = Pi. (6.4g)
Then, by Theorem 4.12,
y = cypie™® + co(po + spije’,
where c; and cg are arbitrary constants, is a general solution for Eq. (6.4f), and

X = ¢ypi(t — a) + co(po + piln(t — a))(¢ — a) (6.4h)

is the corresponding solution of Eq. (6.4e).
In solving Eq. (6.4e), we made use of the identity

exp [\; In (t — a)] = (¢ — a).

This identity might appear a bit strange if \; = a; + i8; is a complex number. No
confusion will result, however, if we keep in mind that, for t > a,

exp [\j In (¢ — @)] = exp [a; In (t — a)]- exp [i8; In (t — a)]

= (t — a)*i(cos In (t — a)*i + isin In (t — a)*s).
 Series solutions around regular singular points 161

When 5,(t) = 6; and bo(t) = bz are constants, Eq. (6.4a) has the form

(t — a)?x" + (t — a)byx’ + box = 0, (6.41)

and it is usually called Euler’s differential equation. The matrix B of the associated
system (6.4d) has, by Problem 12 of Exercise 3.11, two linearly independent eigen-
vectors if and only if the characteristic equation

AA — 1) + 514+ bg = 0 (6.4j)
has distinct roots. Thus, a general solution of Eq. (6.4i) has the form
x = ¢y(t — a) + c(t — a)
if \, and Xz are distinct roots of the characteristic equation and has the form

X= (cy + coln(t — a) )(t Says

if \; is a double root.
Notice that the characteristic equation (6.4j) can be found without converting
Eq. (6.41) into a system: one sets x = (t — a)* in the differential equation and
divides by the factor (t — a)* after performing the indicated differentiations.

EXERCISE 6.4
1. Describe the behavior as t > 0+ and as t— +o of the solutions to the following
Euler equations.
a) t2x" + 3tx'’ +x = 0. b) ¢2x” — tr’ +x = 0.
Cte) =x = "0. d) 72x” +x =0.
ear + 2ty! x = 0,

2. Describe the behavior as t— 0+ and as t— +o of the solutions for the systems

x = Bx/t if
yg _ (2-1 = har bo

el pi
a B=|7y ie oa=| 4 |

6.5 SERIES SOLUTIONS AROUND REGULAR SINGULAR POINTS

Let us turn our attention again to the equation x” + x/t = 0,0 <1t< +o and
try to find any solution of it in a form which is valid over the entire interval
0<1t1< +o. Corollary 6.1 is applicable only about interior points fo of the
interval 0 < t < +o; but might there not be a power series about ¢ = 0, with
infinite radius of convergence, which satisfies the equation for t > 0? Suppose there
were. It would have to be of the form x = w(t), where

ere Dee k=0

(6.5a)
162 Power series solutions for linear equations 6.5

We substitute the series (6.5a) into the differential equation and obtain

kth = Deg”? + Dent l= 0;

i = 2

which, upon baitiat cn becomes

+ Dy[(k + 2k + Dense + crpilt®=

For ~, to be a solution, it is necessary that cy = 0 and that
Ck+1
Ch4-2 (k + Vk hy k= 0, (6.5b)

The constant c, is arbitrary.

Assuming that the c;’s are chosen according to the rule (6.5b), what will be the
radius of convergence R for the solution series? By the ratio test,
Ch+1 = lim (kK + 2)(K + 1) =
= lim
k>o Ck+2 kw

Thus convergence is assured for all t. Computing the c;’s, we obtain the relations

on = 0; c,seis arbitrary,
re Co 2 Be x] Bee C3 Se eee
3-9 ef
372-20] :

= a eS O
paint S16 GRMN ic Pack acre,
and in general

= Wick:
+ =
eu Be
Consequently,
oo k
OLj= Ds Sel Gilera (6.5¢)
5
where we have taken c; = 1.
Since cy, was arbitrary, any power series in ¢ which satisfies the equation
x" 4+ x/t = 0 must be a scalar multiple of the series (6.5c). Thus we cannot
expect to express a second solution x = Wo(t) of x’ + x/t = 0 as a power series if
W1 and W» are to be linearly independent. At the end of this section, we shall compute
a solution

Walt) = 1+ ¢ — $f? + she? — r#o er" -+ (In?) y (=1)""? ee 1)!

of x" ++ x/t = 0 which is linearly independent of ¥,. We shall state in this section,
and prove in the next section, two theorems with corollaries which describe the
solutions of
1
 Series solutions around regular singular points 163

and
(t — a)?x” + (t — a@)b,(t)x’ + b2(t)x = 0, (6.5e)
when B, b,, and bs are analytic for a < t < w. The student should keep in mind
that if the system (6.5e) comes from the equation (6.5d) by way of the substitution
x = [x, (t — a)x’]’, then
0 ]
oe re ie et
and the solutions of Eq. (6.5e) coincide with the first components of solutions of the
system (6.5d).

Theorem 6.2. Suppose that B is analytic for a < t < w and that the eigenvalues
A, and Xz of B(a) neither are equal nor differ by a nonzero integer. Then
1
x’ = —— B(t)x, SE <igss
(t — a)
has two solutions with the forms

o:(t) = (t — a)" YS x(t — a),

— (6.5f)
o2(t) = (t— a)? DO dx(t — a)"
k=0

and they are linearly independent over a < t < w. The coefficients ¢; and dj. can be
evaluated by substitution. -
Example 1. Write the coefficient matrix for the system

~
hmlm
Re

in the form B(t) = By + By,t, where

1 0 fo 1
cae ] and Bi =|; A

The eigenvalues of By are \; = 1 and A, = 4. They do not differ by an integer. The

system therefore has solutions of the form (6.5f). To find a solution corresponding
to 4; = 1, we substitute

x= 9) =2) > gt = © qt?!

k=0 k=0

into the differential equation rx’ = Box + B,tx and find that
x

» en(A + bet = > Boe, t**? 5F > Byc,t**?.

k=0 k=0 k=0
164 Power series solutions for linear equations 6.5

To simplify this expression we change the index in each sum, setting r = k + 1

in the sum on the left and r = k + 2 in the sums on the rou This gives

ye Cort -> Bot;— 1a + SE B 1Cr— site

r=1 r=2

Consequently,

(By — Weo + ) [(Bo — rDer-1 + Bier—2)t” =

T=2

and it follows that

(Bo — lJ )cot = 0 and (Bo = TL) Gre = —Byic,_9, Ig Zz Poe (6.5g)

We solve the first of equations (6.5g) by taking ¢g = all, 0]”, where a is an arbitrary
nonzero constant. Solving the second of these equations, we find that

C= —
0
1
=
0 C;_9, r > Ds (6.5h)
2r — |

We compute the coefficients c, recursively from the recursion formula (6.5h) starting
with cy = afl, 0]” and find that

Gemlll ime pede sells ae FUR TS

= : Zid
oi(t) = 1 é ely 1 :

0 Otae a 3 poe
— YK See 2 4 ee

To find a solution corresponding to \2, = 4, we substitute

X = 65(1) se eed = ede

k=0 k=0

into the differential equation tx’ = Box + B,tx and group like powers of ¢ to obtain

(Bo — 31)do + Dy
=1
[Bo — (K+ Bde + Bide_i}t =
Thus

(Bo— 31)do=0 and (Bo— (K+ 9) = —Bide_1, k>1. (6.5)

The first of equations (6.5i) is solved by taking dy = B[0, 1]", where @ is an arbitrary
nonzero constant. Solving the second of these equations for d,, we obtain

py
aa
d.=|, dion aed (6.5)
aE
 Series solutions around regular singular points 165

The d;’s are then computed recursively beginning with dy = 6[0, 1]’, giving

feces Ol ot Spr.
$2(t) = meeGite Ol eteesGl- taOfte
2143/2 ate 347)? fo -2%
as (a ae 4,712 an oh |

Since a second order equation (6.5e) can always be written as a system (6.5d),
Theorem 6.2 can always be used for its study. It is convenient, however, to state the
results for the second order equation directly.

Corollary 6.2. Suppose that b, and bz are analytic for a < t < w and that the
roots \; and Xz of the characteristic equation

AMA — 1) + Aby(a) + bola) = O (6.5k)

neither are equal nor differ by a nonzero integer. Then
(t — a)?x!" + (t — a)b,(t)x’ + bo(t)x = 0
has two solutions with forms

Wil) = @ — a)" k=0

DY) alt — a)",
Oy =a Daa
and they are linearly independent over a < t < w. The coefficients c;, and d;, can be
evaluated by substitution.
The corollary needs no formal proof. One merely converts the second order
equation to a system x’ = B(t)x/(t — a), notes that (6.5k) is the characteristic
equation for B(a), invokes Theorem 6.2, and displays the first components of the
resulting solution vectors as ¥; and y2. A scheme for remembering the characteristic
equation (6.5k) without forming the matrix B(a) is the following. Set x = (¢ — a)’
in the left side of the differential equation, divide out (¢ — a) after simplifying,
and then set ¢ = a.

Example 2. The equation

Pela (ty = 05 oy > 0, OO P< 4, (6.51)

is called Bessel’s equation of index v. Let us solve it for the case vy# n/2, where n
is a non-negative integer. The characteristic equation is

KOe DEP eNe Lee (Ola) =? =? =. 0.

Since v ¥ n/2 for any integer n > 0, the roots \, = v and \2 = —v do not differ
by an integer.
 6.5
166 Power series solutions for linear equations

We substitute the series

x= ye Gr
k=0
into Bessel’s equation and find that

SV thot kt Det’ + OS + kent’ + p2 I i 22 Wo

k =0 k=0 b==(0) =

DO (e+ kb)? = lex + cea}

= (w+ Det k=2 = 0.
Since vy —4, cy = 0. The other coefficients are determined from the recursion
formula
an Ch 2
ORG= pa >
eleeZe Pl,

Thus c, = 0 if k is odd. If k is even, we have k = 2m, m = 0,1,2,..., and the

recursion formula gives
(—1)"co A S
Cam * Jemmi(v + 1)---( + m) Ndh a i

It is customary to choose cy = [2” fo ew’ du]. With this choice for co, the
resulting solution
sgh (DVS
x= co(5) Lary (=1)" we) \e
is denoted by J,(t) and is called the Bessel function of the first kind of index v. Corre-
sponding to \x = —v, we obtain

ase (E\ itis (=1)" Ne

Salo ee rsnccram)!
which, with co chosen as before, is denoted by J_,(t). ||
When the eigenvalues \; and \» of B(a) are equal or differ by an integer, the
solutions of the system (6.5d) are as described in the following theorem.
Theorem 6.3. Suppose that B is analytic for a < t < w and that the eigenvalues
dy and \» of B(a) either are equal or differ by an integer, say }; — \» = M> 0.
Then
i. 1
x= C= Bes or Wien)

has one solution of the form

oi@) = @ —a)* k=0

2) at — a)" (6.5m)
and a second solution of the form

go(t) = (t — a)? D di(t — a) + €9,(1) Int — @). (6.5n)

 Series solutions around regular singular points 167

Here € is a constant which, along with the coefficients ¢;, and d;, can be evaluated by
substitution.
Example 3. Consider the system tx’ = (By + 1B,)x, where

seer ol pa OREO!
Bo = E | and B, = E 3

The eigenvalues of By are \; = \y = m7. As in Example 1, we substitute x =

ko ext" into the differential equation and find now that

l
kik!
k ’ ke Olees
kk!
where a is an arbitrary nonzero constant. Thus
es | peer

gift) = filer
k=0

To find #2, we substitute

eo . n ] pete
x = d,.t°+™ + e(Int
» : ( Me, k |kk!= |

into tx’ = Box + Bix and obtain

jm iI ioe
(Bo a mI )do +. ae [Bo os (k + w)1]dy, — By,dy_} aa clia pet = (I).
el : 'k!
We must therefore have (By — J)d,y = 0 and
a eg
[Bo a (k =F aw)Id), = —By,dy_y aP alles for k = Il.

The solution of the latter equation for d, is

da
: : d;_; + €
= eee
I kale
l k!k!
k 0 =— J

If we were to choose € = 0, then we would have to take dy = ALI, O]" for some
nonzero constant 8 and the d;’s would, except for a scalar factor, equal the c,’s.
This would give #;(t) over again. The choice of € is otherwise arbitrary so we shall
set €= —1 and take dy, = [—1,0]’. Then
11
Sia fs 22 00 22H
a: = loli = ees
 Power series solutions for linear equations 6.5
168

and
1 I 2
dg =
32.92. + 32-21* 3-313! . 3

| I
epee es
The other d,’s are found similarly. ||
We state the scalar version of Theorem 6.3 as a corollary.

Corollary 6.3. Suppose that b, and bz are analytic for « < t < w and that the
roots \; and Xz of
AMA — 1) + 4i(@) + bola) = 0
either are equal or differ by an integer, say \y — \2 = M > 0. Then

(t — a)?x" + (t — a)by(t)x’ + bo(t)x 0, 61S

ff K @,

has one solution of the form

ive}

Wilt) = (t— a)” DO a(t — a)" (6.50)

k=0
and a second solution of the form

i= 2 oe %,Tee a) eye) (6.5p)

Here € is a constant which, with the coefficients cy, and d;, can be evaluated by
substitution.

Example 4. We shall compute a solution of x’’ + x/t = 0 which is linearly in-

dependent of
E k
i= =
G1)k+l aaa
t ;
wie ee (6.5q)
The indicial polynomial is \(A — 1). The solution y,, given by the series (6.5q),
corresponds to \ = 1, the zero with the larger real part. From Corollary 6.3, the
desired second solution 2 has the form

Y(t) = 1° e aut + e(in gt) me(aie

a 1!
Thus
Mh = = Lede k—2 = _4k+1 ie
es es eC arcs) EGE = Dik — 2)!
4 2¢ 2 ae ae ra (—18ty I
t 2 DE (ie =e =,RED
 Series solutions around regular singular points 169

and
F 7 é ie ee % (Sie
AOE Yee ae Te rine aly
(6.5r)

ADoce ee asa
heE-p eS”
—eE a aban.

Notice that the In ¢ terms cancel out as Corollary 6.3 predicts. Now combine the
series on the right in Eq. (6.5r) to obtain

Wet) + po(t)
S= 2e— et dy tO- [K+ Dhdgs +
RG)
ik!
dE
manera
k
ae

When the terms are equated to zero, the result is

€ = —dp

eee =e Gees eel

eer) Pa Cate kik!
1 k42rf 2k + 1
pa Gre Geet) thee kel.
= ————_ };-—g. —_— a D =

With the choice dy) = —e = 0, the recursion scheme gives y; again. Choose do =
—e = land d, = 1, say. Then

1 3 5
d= 54 | 1 = Gia} = 7
Alo 5 S)
ee 18
ee! kg? ALA
48k oil 1728
Thus W,2 has the form

Slag Sis 47 4 = k41 i

=
Mates glo5 agli re —
17737 o+---—(nt
EPEC—1 —_______.
re sm
EXERCISE 6.5

1. Each of the systems below has two linearly independent solutions each of which can be
represented by convergent power series in ¢. Compute these solutions.

Se
7 =k
om
© ,ee’ | ape +147) Atet+ aide
—
&we
wS
Se
a
a Geer
ee
Ss eas") @+) JLy
170 Power series solutions for linear equations 6.5

2. Write down the characteristic polynomial for each of the following equations and find
its zeros.
a) 6f2x" + tx’ + (1 + 60x = 0. b) 22x” + tx’ + (f — 3)x = 0.
c) 22x" + tx’ + (1 + Dx = 0. d) t2x"" — tx’ + (2+
Dx = 0.
e) tx’ + x/+ x =0. f) t2x" + 3x’ + 1 + dx = 0.
g) 4r°x"” +x = 0. h) tx’ + 2x/ +x= 0.
i) 422x" + 4tx’ = (1 + 40x = 0.

3. Find two linearly independent solutions for each of the equations in Problem 2.
4. Show that t = 0 is a regular singular point for the equation (sin fx” — x’ + x = 0.
5. Find a nonpolynomial solution of Legendre’s equation

(1 — #?)x"" — 2tx’ + pv + 1)x = 0

on the interval 1 < ¢ < 3.

6. Use Theorem 6.2 to find a fundamental solution set over 0 < ¢ < + for each of
the following systems of equations.

oft =42 Ie} Of emartel|iae

o[f-34" SE
7. Use Theorem 6.3 to find a fundamental solution set over 0 < t < +o for each of
the following systems of equations.

>) bifer are |e

oft =3[5 allel
8. Use Theorem 6.3 to find a fundamental solution set over 0 < t < +0 for each of the
following systems of equations.

= ale all}
i) —
—Us
Sats
|
Jee eeoas al Fi)
oT -7
9. The solution x = ¥(/) to the initial value problem

MO 19! a a 0, weil, os! SO) Wien ¢=—O

is not analytic at the point ¢ = 0. The solution is, however, representable as the sum of a
series of elementary functions. Compute several successive approximations to the solution
and conjecture the form of the series. Make a change of variables x(t). = y(s),-s =f)
which will reduce the given initial value problem to a regular singular point problem. Then
 ~

VIA UA Che Ut ice,Non ensenhiihlite Carpio Th

Wh) = Z ASS.
99, arn 49 the, whebitrns A Os,vpbemn

bn
Nae i I,
19 429 iin: Al = ¢, AL) = bb), Pb) = AD) 05:8 29y Then
634

C4 CUS RE VOR. PHMD, UII IID IAAI IDS, EVE SLIMES

V0 prove Venn 6/1, we ba wihks Ce witihtion & =~ of 9 tt — a =
Wi. 0 Pip
os

~ a = Aas 7 Aa ~ af, fa)

Wists f
Aye Byo~ hh, Ay
= BB; lo ko, WD) = F tht — of
We tes tan Git £4, G4) Kis 4 GHD SLI SIME

11= Zé ~ wf, 6%)

Tika Wi grows Coit toe bust, whettions GL) hits,
$iistintion A tus vet0s 690) tn Ex, 4a) Bows
a Ea

£ es - f= t+ E (Em) -a-

Vix His be 2 tod hotly 8 1, hts mexomaty on8 oie Gs

Bs — Sts = At =
mah, tak > 1, 6a
gz
bis Eat Ate Eebate
Vo whe Gre het A tase cys, we I t,t on egos SB, vorsespontang
Wie T9 whe Gee sek cgi, nie Gis Ag ~ 11 = By — Cn + Sb Soe
e cpt iz aut A teeter bs + bios wy mega
k > 1,64 (4, -)#O
wh (Ay — SS sis, Ths Cee 6/8 con be cries seossdy Irom

om — hie Em GSe)
172 Power series solutions for linear equations 6.6

It follows that
wn

oi(t) = (t — a)y,() = ((— a) DS at — @)*

k=0

satisfies x’ = B(t)x/(t — a) formally.

Since \; — A» is not equal to an integer, the same procedure produces a second
formal solution

o2(t) = (t — a)*y2(t) = (t — a)” pe d,(t — a)",

where do is an eigenvector of By corresponding to Xo.
We show that the series for @,(¢) converges by applying Lemma 6.1 to the re-
cursion formula (6.6c). Let L be a number satisfying 0< L<w—a. Since
Yeo An(t — a)* converges for |t — a| < w — a, there is a constant Q such that
|Ax(t — «)*| < Q for all |f — a| < L and all k > 0. In particular, |A,| < QL.
From the recursion formula, we see that
fen
feel = [Go kl | Oe n—
lea ie (6.6d)
Since Ag — kl is a matrix of the form

a—k b
@ ck

its inverse will be of the form

1 is—k -—b |.
(a—k)d—k)—be| —c a-—k
As a consequence, each element of k(Ay — kI)~' is bounded for all k > 0. Let
us say |k(Ay — kI)~'| < P for k > 0. Then inequality (6.6d) implies that
(Roms

L*klex| < OP = leqEan.

10)

But this inequality has the same form as the inequality (6.1j) in Lemma 6.1. The
lemma implies that y;(t) = \if~o ex(¢ — a)* converges for |t — a| < w — a since
w — a — L may be arbitrarily small. The proof that yo(f) = ?_> dk(t — a)”
converges is analogous.
Next let us show that $; and @» are linearly independent over a < t < w.
Write 4; = a; + 78; and \»y = ag + iBy, where a; > ao, and let cy and cy be
constants such that

ci(t — a)y,(t) + cot — a)2yo(t) = 0 (6.6e)

fora<t<w. Then

C1(t — ayy, + Coy2(t) = 0 (6.6f)

HOS fy Oe BY
 Proofs for the series representation theorems 173

If ay > a9,

(¢ — a)1™2) = (¢ — a)®1™2(cos (8; — Bo)(t — a) + isin (8, — B2)(t — a))

(6.62)
approaches zero as ¢ approaches a. Then coyo(a) = cody = 0. Since dy is an eigen-
vector, C2 = 0 and it follows from Eq. (6.6e) that c; = 0 also.
If a; = a, then 6; # By since \; ¥ Ao. We see from Eq. (6.6g) that
(t — a)*:—*2 does not approach a limit as t— a+. Since yo(t) > yo(a) as teat,
Eq. (6.6f) implies that c; = 0. But c; = 0 implies that cy = 0 also, and the proof
of Theorem 6.2 is thus complete.

Next we prove Theorem 6.3. Let \; — \> = M > 0. The solution

Dn

oi(t) = (t— a) >) ex (t — a)

k=0

is constructed in the same way as the solutions in Theorem 6.2. (Keep in mind that
Co iS an eigenvector of B(a) corresponding to 4.)
To construct ¢2(t), we make the substitution

x = (t — a)*y + e¢i(4) In (¢ — a)
in (t — a)x’ = B(t)x and find that

(t — a)y’ = [B(t) — dolly — €(t — a) *¢1 (0).

This equation simplifies when we set \2 = \; — M and ¢,(t) = (t — a)y,(0),
yilf) = Deo cx(t — a)”. In fact,
(t — ay! = [B(t) — dolly — € k=0
DY ex(t — a)". (6.6h)
The coefficient matrix B(t) can be expanded in the form

B(t)— Aol= (Bo — Ao) + Bit — a) + BA(t—a)? +°°-, =|t-—al <<w-a.

Define Ay = Bo — dol and A; = B; for k > 1. Then zero is an eigenvalue of Ao

and any corresponding eigenvector is an eigenvector of By corresponding to Xo.
We shall show that Eq. (6.6h) has a power series solution
ioe}

y() = 2 de(t — a)", (6.6i)

which converges for a < t < w. This will prove the theorem.
When the series (6.6i) is substituted into the differential equation (6.6h), the
result is

aioe A, dy ee ry ae Dylt arawea

3 kd,(t — a)*= me (6.6j)
k=1 k=0 \7=0
174 Power series solutions for linear equations 6.6

There are now two cases.

CASEI. M = 0. The recursion equations obtained by equating coefficients of like

powers of (tf — a) are

A odo = €Co, kd, = Aody +°-°+* + Apdo — €Cx.

Since M = 0, \y = Xo, and it follows that zero is a double eigenvalue of Ap. The
vector €g is already an eigenvector of Ay corresponding to zero. If Ao has an eigen-
vector independent of ¢o, take € = 0 and choose this eigenvector as do. Otherwise
choose € = 1. The equation A dy = Cg has a nonzero solution for dg by Corollary
3.22. The equations

(Ao = kT)d;, = —A,d;_4 eet DF Rg Ajo + EC;

are then solvable for each k > 1 since (Ay — kI)~! exists for each such k.

CASEII. M> 0. The recursion equations obtained by equating like powers of

(t — a) in Eq. (6.6)) are

kd, = Aodk + -+-> + Axdo for k=O 2 ees

Mdy = Apdu + °°: + Amdo — €Co,

(M +- k)duipy = A(duyp + = Ay 0g eC, for Die i

We write these in the form

(Ao — OF)do = 0,
(Ay — 1-T)di = —Aido,
(Ao —- 2-I)d> = —Aj,d, = Aodo,

(Ayo — MI)dy = —Ajdy_1 — ++: — Amdo + €Co, (6.6k)

(Ao = kI)d;, = —A,d;_ 1 eG Ard o == €Cy_—k, k = M+ si (6.61)

Choose an eigenvector of Bo corresponding to \»y for do. Then Aodo =

(Bo — AoI)do = 0. The vectors dy, ...,ds_1 may be found recursively since the
coefficient matrices (49 — kJ) are nonsingular for k = 1,...,M — 1.
The matrix Ag — MI = Bo — dol — (Ay — do) = Bo — AJ is singular. If
€ 18 properly chosen, however, the equation

(Ay — MI)dy = b + €€o, b= Ay

has a nonzero solution for dy. To see this, let x ~ 0 be a 2-vector, which is not
an eigenvector of Ao. By Theorem 3.23, (49 — MI)x = cdo, for some constant
c # 0, since do is an eigenvector of A corresponding to the eigenvalue zero. The
2-vector b can, by Theorem 3.15, be expressed in the form b = byey + dodo. If
 Proofs for the series representation theorems 175

we take dy = box/c, then

(Ao = MI)d yy = bodo =b-— b1Co,

and the equation (6.6k) is satisfied with the choice e = —b,.

After the coefficient dy is found, the d,’s for k > M+ 1 can be computed
from Eq. (6.61) since (Ag — kJ) is nonsingular for k > M + 1. Thus, there exists
a sequence d;,...,d;,... such that the series (6.6i) satisfies Eq. (6.6h) formally.
The proof that y;(t) = Sof ex(t — a)* converges is exactly the same as that
given for the analogous series in Theorem 6.2. Let us prove that

Y2(t) = ss di(t — a)"

k=0
converges. Let L satisfy0O<L<w-—a. Since f-, Az(t — a)* converges for
|t — a| < w — a, there is a constant Q such that |A,(t — a)*| < Q for |t — a| < L
and allk > 0. In particular, |4;,| < QL~* for allk > 0. By enlarging Q if necessary,
one can guarantee that |ec,| < QL~*™ for all k > 0 since (2.5 c(t — a)" con-
verges for |t — a| < w — aalso. As in the proof of Theorem 6.2, there is a constant
P such that
(M+ p)i(4o—-(M+ pI "| SP forall p> 1.
Now take norms in the equation

dtp a (Ao a, (M + pl {A dary p—1 oe te eee Am+pdo = ECy}

for p > 1 to obtain

LM*?°(M + p)idac+ol S POC + [dol + [dalZ + +> + [daryp—i| L477}.

This inequality has the same form as inequality (6.1j) in Lemma 6.1. Thus
3-1 dary p(t — a)”*” converges for |f — a| < L. This implies that ~~ di(¢ — a)”
converges for |t — al <w—a.
Finally, let us show that the solutions @; and ¢¢ are linearly independent over
a<t<w. If not, then there exists a constant 8 # 0, such that $;(t) = Bdo(t).
Since \; = Ao + M,

Bt — a)2tMy
1(1)= (t — a@)yo(t) + e(t — ays (d) In (t — aye.
We put this equation into the form

[8 + eln (t — a)\(t — a)"yi(t) = yal). (6.6m)

There are several cases. .
First suppose M > 1. If t— a+ in Eq. (6.6m), then yo(a) = dy = 0. This
is a contradiction.
If M = Oand e 0, then, as t > a+ in Eq. (6.6m), at least one component of
y2(t) must become unbounded. This is a contradiction since lim;_,« Yo(t) = y(a) = do
exists.
176 Power series solutions for linear equations 6.7

There remains the possibility that M = 0 and e = 0. When ¢ — a+, the equa-
tion By ;(a) = yo(a) results, that is, Bey = do. This is not possible since the algorithm
for computing y.(f) requires that ¢9 and do be linearly independent eigenvectors of
B(a). This completes the proof of Theorem 6.3.

EXERCISE 6.6

1. Write down the analogue of Eq. (6.6a) for

l|

X2 SNe

by making the substitution

[x1, xo]? = 2)?
yy, vel:
2. Let yi and y2 be continuous vector functions, defined for —2 < ¢t < +%,and assume
that y:(0) and y2(0) are Jinearly independent. Show that

gi =™yif) and a(t) = Pryo(t)

are linearly independent over 0 < t < +2 if
a) \1 = 3, Ao = 5, bAi =1+i% A2zs=1-i,
c)A\1 ~=i, he
= -i, d)Ar=T+i, Ag =1+i7.
3. Let A, = diag [1/k, 1/k?], kK> 1. Find a constant Q, independent of k, such that
[Ax(t — 1)*| < @ for |¢ — 1| < 4.
4. Let

Arle
—1 2

Find a constant P such that |k(Ao — kI)~1| < P for all k > 2.

6.7 BESSEL’S EQUATION

Trigonometry can be viewed as the study of solutions of the equation x” + a?x = 0,

where a > 0 is a constant (see Problem 13, Exercise 4.2). Similarly, the theory of
Bessel functions is the study of solutions of the equation

17x! + tx’ + (7? — v*)x = 0; (6.7a)

where y is a constant. This equation, which was solved for nonintegral 2y in Ex-
ample 2 of Section 6.5, is called Bessel’s equation of index v. \t hasa regular singular
point at f = 0. Thus the theory of Section 6.5 applies to it on the interval0 <1< +a.
For simplicity, we shall assume that v is real and non-negative.
There arise in the discussion of Bessel’s equation expressions of the form

ve + )@ £2): ° 3G ek by: (6.7b)

 Bessel’s equation 177

If v is a positive integer, the product (6.7b) can be written in factorial notation:

(v + k)!/@ — 1)!. When v is not an integer, the factorial notation is not defined.
A useful extension of it is provided, however, by the gamma function Tv). For
v > 0, one defines

Tv) = / tees dh (6.7c)

(0)

It is easy to see that (1) = 1. When the integral for T(v + 1) is evaluated by parts,
one finds that
Ty + 1) = 0TO). (6.74)
Thus T(z + 1) = n! for all positive integers n. The product (6.7b) is concisely
expressed in terms of the gamma function, for it equals lv + k + 1)/T(Q).
The identity (6.7d) makes it possible to tabulate the gamma function without
evaluating the integral (6.7c) for vy> 1. In the same spirit, one inductively defines
Tv) = (1/vy)T@ + 1) when vp is negative but is not an integer. It is known, for
example, that ['(4) = Vz. Thus ©) = 4P(Q) = 47 and P(—4) = —2T@) =
—2,/r. The graph of the gamma function is shown in Fig. 6.1.
T(r)

Figure 6.1

Turning now to Bessel’s equation, we observe again that the roots of the charac-
teristic equation are \; = v and \y = —». . .
We saw in Example 2 of Section 6.5 that if 2v is not equal to an integer, then
Bessel’s equation has two linearly independent solutions J, and J_«. Using the gamma
178 Power series solutions for linear equations 6.7

function, we display them as the series

Ce) C1), t ot
mo = (§)>» are een (5):
t\7” @ aay t oe
J_ft) = (5) Je MiG + 1) (5)
The graphs of J and J, are indicated in Fig. 6.2(a). If \y — Az = 2v is an integer,
then Corollary 6.3 applies, and there is a solution of the form

x=7o0) =F SS ant’ + eJ,(f) 10 ¢. (6.7e)

k=0

When » is an odd integral multiple of one-half, it happens that e = 0. On calculating

y(t) with an appropriate choice of do, one finds that ~o(t) = J_,(t). Thus the
general solution of Bessel’s equation has the form

x = constant - J,(f) + constant - J_,(t)

whenever v > 0 is not an integer. If vy> 0 is an integer, it is not hard to show that
J_,(t) = (—1)’J,(1), which implies that J, and J_, are linearly dependent.
Just as the trigonometric functions satisfy identities, so do the Bessel functions.
For example, J,41(¢) = 2v(J,()/t) — J,_1(t) for all »y X 0,1. Thus Bessel func-
tions of the first kind may be tabulated for all vy> 1 from the values of J,(t) and
J_,(t) with 0 < vy < 1. Ina problem below, the student will be asked to show that

Jyj2o(t) = V2/at sint and J_1)2(t) = V2/nt

cos t.

Using the identity above, one may express all the functions J,,4.1/;2(t) and J_m—1;2(0)
as finite algebraic combinations of sin f, cos f, and ¢. These are called the spherical
Bessel functions.

Figure 6.2
 Bessel’s equation 179

When 2y is an integer and y is not an odd integral multiple of one-half, then v

must itself be an integer. Corollary 6.3 now applies in all generality, and there is a
solution W2, independent of J,, with e 4 0. Substitution of the expression (6.7e)
into Bessel’s equation, yields

TES OVI MN SIH Ee

c=2
et lca tT ta EN
When J;(t) is replaced by its series representation, the result is

De Dy Cae (2m “e v) pon

m=0 mi(v + m)!2?"+"

= (1 = 2d + DO (Kk — 0)? = vd, + death

[=
(6.78)
Since the series on the left in Eq. (6.7f) has no odd power of f¢, neither has the series
on the right. Consequently, d; = 0 and

[(2m =e eS y)? a V*Wdom+1 ete Coma 4 = 0 for m = its

This implies that d2,,4, = O for all m > 1. We may therefore change the index of
summation for the right-hand series in Eq. (6.7f) to 7 = k/2. It is convenient to
change the index of summation on the left to 7 = m+ yp. Then

ee
j=v (Gj — vl j!2°? j=l
I ae a ae.
(6.72)
Consider first the case vy= 0 and recall from the proof of Theorem 6.3 that, for
this case, one may choose € = 1. We shall give to the constant do the value zero.
The resulting solution Wz is called the (Neumann) Bessel function of the second kind
of index zero and it is usually denoted by Y‘°. Equating coefficients of like powers
of ¢ in Eq. (6.7g), we find that

doy = yg) ~ 923-2 T

(Cie s}
9G |

Thus
dy = as d= 42 r oe <n = 4292 fa 5

1 1 1 oe 1 ewe *
de =o {-p(1- V5 soo| = pce tat y

De WOTRUY lenseae
 6.7
180 Power series solutions for linear equations

It follows that

¢ | aaa one Gy AY Gyre aie 0:

ya) = > Gly 2 NZ
Se)
For the case v = 1, Eq. (6.7g) yields the relations

| 2do(—1) (aie "QJ 1 2,

CS ea ea = py
The coefficient dy is undetermined. The choices dy) = —1 and dg = 3 give rise to
the solution
(ee =| j j 1 2j+1
YPO=y2() = — -5 > = ay f= =a t(5) Seyi (Oto
r=0
i= 0,
which is called the (Neumann) Bessel function of the second kind of index one. Series
representations of the Neumann functions Y” for vy = 2,3,... are obtained in a
similar way.
When vy is a non-negative integer, second solutions of Bessel’s equation other
than the Neumann functions are also of interest. The (Weber) Bessel functions of the
second kind Y, occur in applications. They are related to the functions Y” and
J, by the equations

¥) = 2[YW + (7 — n2)4,(0),
where

realm (14 p4+--4}-Ink) = 05772...

k>+0

is Euler’s constant, or by the equation

Y,@) = im (cot Br)Jg(t) — (csc Ba)J_p(x).

Graphs of Yo and Y, are indicated in Fig. 6.2(b).

Example. We see from their series expansions that Y(t) > —aw and Jo(t)—>
T(1) = 1 ast— 0+. Thus the only solutions of

t?x" + tx’ + t?x = 0 (6.7h)

which are bounded on an interval 0 < ¢ < € are scalar multiples of Jo. We made
use of this fact in Example 5 of Section 6.3 when we asserted that x = Jo(ar) is the
only solution of
r?x" + rx’ + r?v2x = 0 (6.7i)
which is bounded forO << r< 1. ||
 Bessel’s equation 181

EXERCISE 6.7

. Evaluate the integral (6.7c) to show that [(1) = 1.

. Prove that '(v + 1) = rI(v) for v > 0.
. Show that (vv +- k + 1)/TO) = pv + 1):--@ + 4) foryv > 0.
. Compute I'(—).
aA
NO
BW. Compute two linearly independent solutions of the equation

Fox at te (GF 3x = 90

by the method of Corollary 6.2.

6. Use the result of Problem 5 to write down the series for J1/3(¢) and J_1/3(2).
Je If vy> O is an integer, prove that J_,(t) = (—1)’J,(d.
8. Verify the identities
TKO) = GIO))/t —FaiO; vy > 0,
UG) tis (VI) 2 yet oe 0)
[Hint: Differentiate the products t~’J,(4) and 7’J,(f) and use the results in conjunction with
the series for J,(2).]
os Use the results of Problem 8 to derive the identities

Jr4iQ) = 20V,)/t) — 1),

O=ZCa10)= Ja).
10. Show that|, ?Js21(0) db = 12,1 dor T > 0;
it Show that Jijo(4) = /2/nt (sin 1) and J_1/2(t) = /2/rt (cos 2).
12: Compute J3/2(f) and J_3/2(2).
U3: Compute Y“?)(2).
14. Given that Y(#) and J,(1), v > 0, are linearly independent solutions of Bessel’s equa-
tion, show that Y,(A) is a solution which is linearly independent of J,(2).
15. Show that x = Jo(Ar) satisfies Eq. (6.71).
CHAPTER 7

QUALITATIVE BEHAVIOR OF SOLUTIONS

FOR LINEAR EQUATIONS

7.1 PRELIMINARIES

The theory developed in Chapter 4 is valid when the coefficient matrix A(f) in

x =24(0)% (LH)
or

x’ = A(t)x + b(t) (L)

is not constant. Thus, solutions to initial value problems for these equations exist,
but one usually cannot write them down concisely in terms of elementary functions.
Faced with this situation, an investigator must try to describe the behavior of solutions
without actually attempting to express the solutions as elementary functions.
We saw in Chapter 6 that power series techniques can be useful for studying linear
equations if the coefficients are analytic. Power series techniques have some dis-
advantages, however. The unique solution to the initial value problem

x” + /8y=0, x0)=1, ~@)=0, (7.1)

for example, does not have a power series expansion about the point t = 0 (see
Problem 9, Exercise 6.5). Even if a power series expansion for the solution to an
initial value problem is explicitly known, qualitative properties of the solution still
might be difficult to recognize. As an illustration, the reader might try to convince
himself that the unique solution to the initial value problem

x’ + x = 0; (0) =02 x Oj}aal

is periodic by studying only the power series

3
0 a
 Preliminaries 183

Similarly, it happens that all solutions of the equation x’ + (1 — e~')x = 0 are

bounded as t—> +, but it would be difficult to establish this result using the
techniques of Chapter 6.
In later chapters, methods of studying the behavior of solutions for the general
equation x’ = f(x, 7) will be discussed. Since linear systems are special cases
of it, many of these methods will apply to (L) and (LH). In this chapter, however,
the techniques presented are applicable primarily to linear systems. Specifically, we
pose three questions that are commonly asked about linear systems, and we make
essential use of linearity in answering them.
We assume throughout this chapter that the coefficient functions of (L) and (LH)
are continuous on an interval a < t < +o. In Sections 7.2 and 7.3 to follow we
shall take a = —a@. The questions to be studied are these:
1) Is there any solution @ of (L) or (LH) which has the property that @(t) — 0
as t— +? Do all solutions have this property?
ii) Under what conditions does (L) or (LH) have a bounded solution? Could
all solutions of either equation be bounded?
iii) Under what conditions does (L) or (LH) have a periodic solution? Could
either equation have more than one periodic solution?

EXERCISE 7.1
1. Prove Sturm’s comparison theorem: Let p and q be continuous real-valued functions on
a < t <w such that g(t) > p(t). Suppose that ¢ and yw satisfy

$7") + pO) =90 and w'() + aMO~C = 0.

If $(f1) = o(t2) = 0, but d(t) > O for a < f1 < t < te < w, then there is a point &,
ti < & < fg, such that ¥(&) = 0. [Hint: Note that $’(t1) > 0 and ¢’(t2) < 0.] Assume,
for contradiction, that ¥(7) > 0 for t1 < t < te and deduce that ¢’(to)W(t2) — o’ (ti W(t)
is positive to obtain a contradiction.
2. Prove Sturm’s separation theorem: Let p be a continuous real-valued function on
a<t<w. Suppose ¥; and we are linearly independent solutions of x” + p(x = 0.
Then y/; must have a zero between any two consecutive zeros of 2. [Hint: Take p() = g(a),
Vi) = (9, Yo = ¥(@ in Problem 1.]
3. Show that x = V/tJ,(0, v > 0, satisfies
fae
vt(1+45") 20, ip > 0,

4. Use Sturm’s theorems to deduce the oscillatory character of the Bessel functions depicted
in Fig. 6.2.
5. Show that there is a zero of Jo between every two positive zeros of J1.
6. a) Show that every solution of x’’ + (1 + e~‘)x = 0 has infinitely many zeros.
b) Let p be continuous ona < t < +. Abstract your study of (a) so as to obtain
a condition on p which guarantees that every solution of x” + p()x = 0 has infinitely
many zeros.
 Qualitative behavior of solutions for linear equations 1f?4
184

7. a) Show that no nontrivial solution y of x’ + (1/4t?)x = 0 could have infinitely

many zeros on an interval ¢ => fo.
b) Show that every solution py of x” + (k/4t?)x = 0 has infinitely many zeros if
he Vee? er 10:
8 Show that no nontrivial solution of x’’ + p(f)x = 0 can have more than one zero if
pi) < 9.
9. Could a nontrivial solution of the equation x’’ + x/t = 0, ¢ > 0, have infinitely
many zeros?
10. Let p be continuous on a < ¢ < + and let y be a solution of x” + p(t)x = 0.
Could y have infinitely many zeros in a bounded interval ty < ¢ < t2? Why?

7.2 THE HOMOGENEOUS EQUATION WITH CONSTANT COEFFICIENTS

Consider first the equation x’ = Ax, where A is a 1 X 1 matrix. Every solution is

of the form ¢(t) = e“‘c for a suitable constant c. The function ¢ will be constant,
hence bounded for —«# < t < +o, if either c = 0 or A = 0. We exclude these
cases from further discussion and write A = a + i@, where a? + B? 4 0. Then

JCD] = lel Je+4] = Iclews,

and it follows that |¢(t)| > 0 as t > +o if and only if a < 0. At the same time,
we see that |¢(z)| is bounded for —«1 < t< +o if and only ifa=0. Itisa
consequence of the equation
o(t ae 27/8) = cela tt) (tt 27/8)

= cele TB) to 2ra/B, 2ri

SUP
that ¢ is periodic if and only if a = 0. The only periods possible for ¢ are integral
multiples of 27/8.
In case A is ann X n matrix, the analysis proceeds in a similar way. When we
say that a solution @ is bounded, we mean that @ is bounded over the interval
=O KK =P ec

Theorem 7.1. Let the eigenvalues }; = a; + i8; (j = 1,...,n) of A have negative

real parts and let n > O be any number satisfying a; < —n for j= 1,...,n. Then
there is a constant M > 0 such that \e4'| < Me~" for all t > 0.
Proof. Let ¢ denote the vector with jth component equal to one and all other com-
ponents equal to zero. Then ¢,(¢) = e“‘e is the jth column of e4. By Theorem 3.9
of Section 3.7, however,
k 4
At = DI (ce, + eft --- + Se
o;(t) = ee
r=1
 The homogeneous equation with constant coefficients 185

Thus
k
e"1;(t)| < akDo Cleri| + leralé +--+ + leemelé"* ert! < e—*pr), > 0,
where k

P= au (emi elit eet |Ga|)e

is a polynomial in ¢ and —e = max 1<,;<z{a, + n}. But p,(de~*! attains
a maximum
M;; on the interval 0 < t < + wo. Since |e“"| < |@,(4)| + --- + |¢n(d], it follows
that |e“'| < Me~", where M = M,+---+ My. ||
Corollary 7.1. All the solutions of x' = Ax approach zero as t > +. if and only
if each eigenvalue of A has negative real part.
Proof. Let x = ¢(t) be a solution of x’ = Ax and suppose that each eigenvalue
Aj = a; + i8; has negative real part. By Theorem 3.9 of Section 3.7, there is a
constant vector ¢ such that (t) = e4‘’c. Let » be such that a; < —y < 0 for
omeletsi Lhen
l6| < le*4| - |e] < Mlcle™,
and it follows that |@(t)| — 0 as t—> +o.
Conversely, suppose that some eigenvalue \; = a; + i8; satisfies a; > 0, and
let p; be an eigenvector of A corresponding to \;. Then x = p,e*s is a solution
which does not approach zero as t— +o since |p,e*s"| = |p;le% > |p,| for
fos 02 ||
Example 1. When the equation x” + 3x’ — 4x = 0 is converted into the system
cog cated | 8Oey 1 Bs
y =493)\)10
one finds that its solutions are of the form

All solutions with ec. # 0 become unbounded as t— +o and all solutions with
C> = 0 approach zero as t— +o. Notice that, in terms of the scalar equation,
(| = |x()| + |x’()|. Hence the behavior of both x(t) and x’(t) is known as
to>+t+o. ||
Example 2. Every solution of the equation
Xj f 1 » 0) 0 0 0) Xi

Xo Ea or OO ee
peel Or 3 100.020) |xe oe
x4 in Oe he Sie a 22)
Xs i we i eee ae
x6 DO ay ay Sees
186 Qualitative behavior of solutions for linear equations 7.2

is a linear combination of the vectors

21 3h
¢i(t) = Die eae ose poe gay o3(t) = pse",
2
oa(t) = pase”, o5(t) = (Ps + ale, oo(t) = (>.+ ips + > ps)esis
where
Pp; = [1,i,0, 0, 0, 07”, po = [by—7,.0,.0,0, 01 -
D2 = 10,0, 16 0,.0, 01. ps = [0, 0, 0, 1, 0, 0)’,
Ps ll= 10)0;, 00, 120)7 nae 00, CORO
Any linear combination involving $1(f), ¢2(t), or $3(¢) nontrivially becomes un-
bounded as ft—> +o. Linear combinations which do not involve $1(¢), ¢o(t), or
$3(t) approach zero ast—> +o. ||
As one might expect from the scalar example at the beginning of this section,
questions (ii) and (iii) of Section 7.1 are essentially the same for constant-coefficient
systems. That is, the solutions of x’ = Ax which are bounded for —-w <t< +o
are linear combinations of the periodic solutions.
Theorem 7.2. Suppose the matrix A has a pure imaginary eigenvalue } = i8 # 0
corresponding to which there are k and no more than k linearly independent eigenvectors.
Then x’ = Ax has k and no more than k linearly independent solutions of period
2n/8. Every nonperiodic solution corresponding to \}= iB is unbounded as t— +a
or as t > —o.
Proof. If p is an eigenvector of A corresponding to \ = i8, then (t) = pe’! is a
solution of period 27/8 since
o(t ae 27/8) = pen ene cae pee?" sae, $(t).

Now suppose that

Pi,---5,DPk

form a set of linearly independent eigenvectors corresponding to \ = if. If there

are constants ci,...,c, such that c,p,e! + --- + c,p,e*% =0, then cip; +
“+> -+ cep, = 0. Since py,...,p, are linearly independent, cj = --- = c, = 0,
and it follows that the solutions

x = pe, @ X= spre
are linearly independent.
Let ¢ be any solution corresponding to \ = i8. Then, it follows from Theorem 3.9
that for some choice of constant vectors ¢;,...,€m, One may write

p(t) = (C1 + cof + +--+ + Ont™ Ye

where m is the multiplicity of \ = i8. The solution @ is unbounded as t > +o
(hence could not be periodic) unless cg = ++: = cm = 0. In that case, ¢,; must
be an eigenvector of A corresponding to \ = i8 in order that satisfy the differential
 The homogeneous equation with constant coefficients 187

equation. Thus (ft) is a linear combination of the solutions x = pe’! and the
proof is complete. ||
Corollary 7.2. All the solutions of x’ = Ax are bounded over the interval —w <
t < +o ifand only if
1) A has n linearly independent eigenvectors; and
ii) every eigenvalue of A has zero real part.
Proof. Suppose that p;,...,p, are linearly independent eigenvectors of A with
respectively corresponding imaginary eigenvalues. Then there are n and no more
than n linearly independent periodic solutions. Every other solution can be expressed
as a linear combination of these periodic solutions and is therefore bounded.
To prove the converse, suppose first that some eigenvalue \; = a; + i8; satisfies
a; ~ 0. If p; is an eigenvector corresponding to \;, then x = p;e**” is a solution
of x’ = Ax and |x| = |p,|e*’. Since a; ¥ 0, e%* > +0 either as t>-+o or
t— —o. Next suppose that A does not have n linearly independent eigenvectors
and Jet P = [p;,...,Pn] be such that PAP = diag [B,,..., B;] is in Jordan
form. Since A and diag [B,,..., B,] are similar, they have the same number of
linearly independent eigenvectors (see Theorem 3.20 of Section 3.12). Thus at least
one of the Jordan blocks By,..., B, is nota 1 X | matrix. Say
MeL! sere)
B; = 4
0 v
Since AP = P diag [B,,..., By], we have

eee AD AD oh a Ds Dees ne) lag nbs”. oe]

=i['o-5-AjPp Pi 7 ABs+1 - - -)-
Thus
(A — i,l)p; = 0 and (A — d,J)pj41 = By. (7.2b)
Equations (7.2b) imply, however, that x = (pjt + pj+1)e.#" satisfies x’ = Ax. This
solution is unbounded either as t > +o orast> —o. ||
Example 3. The solutions of
oA i ORS OLX
yl=|(—2 0 Oly (7.2c)
4 O20 O71 lz
are of the form
x(t) sin 2t cos 2t 0
y(t)| = ci, |cos2t} + cg] —sin 2t] + cg} 0];
z(t) 0 0 1

where C4, C9, C3 are arbitrary constants.

188 Qualitative behavior of solutions for linear equations Tppe

Every solution is bounded over the interval -~ <t< +x. No nontrivial

solution approaches zero as t—> +. In fact, each solution is constant-value d or
periodic. ||

Example 4. The matrix

pe SLAW
A. =Oue F1
O 20.7

has one eigenvalue 7 of multiplicity three and (up to scalar multiples) but one eigen-
vector [1, 0, 0]’ corresponding to it. One can check that

pi = [1, 0,0), “pe = [0,10], “and _p3 = [0,0, 11”

satisfy

(A — il)p2 = Pi and (A — i)ps = Po.

Thus
et tet 4176"

oi(t) =| 0}, oot) =e" 5 $3(t) = ie”

0 0 ev

form a fundamental solution set, and every solution is a linear combination of these
three. A solution is periodic if and only if it is a scalar multiple of ¢,. ||
Let us call a polynomial p(A) stable if and only if each of its zeros has negative
real part. In order to apply Corollary 7.1 to specific equations x’ = Ax, one must
have a way of deciding whether or not the characteristic polynomial of the equation
is stable. This is an involved, but much-studied, problem. An extensive discussion is
given in Marden [18], but a few remarks are in order here.
In the first place, it is not hard to see that any real polynomial has positive co-
efficients if it is stable. If = 2, then the converse of this statement is true. Ifn > 2,
then a polynomial may be unstable even though its coefficients are positive. In this
case, one may decide its stability by application of the following test.

Wall’s Criterion. Let the polynomial

p(X) = X” + ayh"7! 4+ +++ + an_id + an

have positive coefficients and let

ro(d) = ay"! + agr™—3 + asrn—5 4+ ---

denote the polynomial obtained from p(d) by deleting alternate terms beginning with \”.
The polynomial p(x) is stable if and only if the quotient ro(d)/p(d) has a continued
 The homogeneous equation with constant coefficients 189

fraction expansion of the form

ro) _ 1
PO) 1
ha aoe
1
ON
CaN gb (7.2d)
C3X + :

a 1
Cn\ + 0 :

where the coefficients c1,..., Cn are all positive.

The utility of the criterion for specific equations hinges upon the fact that a
continued fraction expansion for ro(\)/p(\) can be computed by performing at most
n polynomial divisions. The proof is beyond the scope of this text, but we shall
illustrate the application of the criterion.
To form a continued fraction expansion of rg(A)/p(), one divides p(d) by ro(\)
until a remainder of degree lower than ro(A) is obtained. Say p(d)/ro(A) =
biQ) + r1Q)/roQ). Then
ro(A) Ti 1

PO) 4.0) 4 riQ)

ro(A)

One now divides r,(\) into ro(d) to obtain ro(A)/r1(\) = be(A) + re(A)/riQ) and
continues in this way until the process stops at the mth (m < n) stage. Then

ro(A) jis 1

PQ) 1
by() +

b2(\) + :

at 1
rm).
Dad) + a

According to the criterion, p(A) will be stable if and only if m = n and

biQ) = Ci Nit", b2(A) = Cod, «++; bn(\) = Cnr, In) = 0,

where cj,..-, Cn are positive.

190 Qualitative behavior of solutions for linear equations 1h

Example 5. Consider the polynomial p(\) = \* + A* + 2d? + 2d + 4 with asso-

ciated polynomial ro(\) = A? + 2d. Since

we have c, = 1 and r,(\) = 4. Now

Thus cy = 4 and ra(d) = d?°.

Since the degree of r2(A) is greater than the degree of 7,(A), the expansion ter-
minates and
ro(A) Py 1

pr) pete 1 (7.2e)

ete
Faia
The expansion (7.2e) does not have the required form (7.2d). Hence p(A) is not stable.
(The zeros of p(\) are = —1 tiand\ = 4+ V7.) ||
Example 6. If p(\) = \* + 2d? + 3A? + 2d + 8, then 7o(A) = 2A? + 2d. Since
Weer rots caelee. eypeel hed
Is
tee
RO) eo a)
we have c; = 4 and r,(A) = 2d? + 8. Since
ro(A) —6r

Ce ar
— Xr ee

C2 = 1 and ro(A) = —6. Then

Li) Sy ee Nee
r(A) 3 rah)
which implies that c3 = —4 and r3(\) = 8. Finally
ro(A) _ _ 30
ES ON 7 4°
Thus
ro(A) os 1
P(X) '
ah Ae

N+ :

-p+—
 The homogeneous equation with constant coefficients 191

The expansion for ro(A)/p(A) has the required form, but cz and c4 are negative.
Thus p(\) is not stable. (Its zeros are \ = —3 + id\/7 and \ = 44 idV/7.) ||
Example 7. Consider the polynomial p(\) = \* + A + 2? + \ + 1 with ro(A) =
3 +d. Then
PQ) _ +1
OT
ge een)
Thus c; = 1 and r,(A) = 47 +1. Since ro(d)/r:(A) = 2, the expansion process
stops at the second stage with cy = | and ro(d) = 0, and

ro(A) oa 1
(7.2f)
pr) Gad icone

The polynomial p(\) is not stable since the continued fraction (7.2f) does not have
the form (7.2d). When the expansion process terminates in this way, one can use
the expansion to factor p(A). In fact, upon simplifying Eq. (7.2f), we have

pA) _ Aw+A41,
roa) _ d
Since we started with
PO) ie A Dh ed
ro(A) B+ dr
we conclude that

pA) = Q? +A+ DA? 41). |

EXERCISE 7.2
1. What is the largest number of linearly independent solutions of x’ = Ax which approach
zero as t—> +0 if

—1 1 Oo —1 s ( )
-1-1 0 0 Eyl [Link] 2 BON LOle
Uy tegen tied eS Has Vier me ea
Oo @ @ i 0 0-3 2

2. How do solutions of x’ = Ax behave as t > +2 if

Oe Bar Onn 1 0
=O OI Ow 22 UY 1).
aa act ena mat aed be tag ol hel ere ed bl
0 0O0-2 O 0 O-2 O

3. How many linearly independent periodic solutions can each of the systems in Problem 2
have?
 Qualitative behavior of solutions for linear equations Te?
192

4. Is every solution of x’ = Ax,

Ouse? 0 O
alae oS Ones
0 O 0 V2
0 0-v2 O

periodic? How many linearly independent periodic solutions are there?

5. Show that the boundary value problem

x1] =e 0™ On)
X5:| cm | Oa, aly Os
xo Or 0 e— 2 CO eee
Ben =i) 9 0, = 1|| x4

xi(+o)=0, x2 = (4) =1, x30) =1, x4(0) = 0

has a solution.
6. Which of the following equations have the property that every solution and its first
n — 1 derivatives approach zero as t— +0?
a) x) 4 x 4 x’ — x = 0. b) x8)" x’ + x = 0.
pal ae a alee a re —i"(0).
7. Suppose the polynomial \? + Bd + C has zeros \; and 2, where B and C are real
numbers. Show that a necessary and sufficient condition that \1, Az < 0 is that both
i) > Oeial C > ©,
8. How must the coefficients of the equation

x aix’ -aox -- asx = 0,

where a1, a2, and a3 are real, be related in order that the equation have a periodic solution?
9. Explicitly solve the equation

Gr) 4 x!” 4. 3x! + x! 4 2x = 0.

10. Which of the following equations have the property that every solution and its first
n — 1 derivatives approach zero as f— +0?
a) xv) 4 x! 4+ 2x" 4 2x’ + 4x = 0. BD) ir ee
c) xX” + x” + x’ + 6x = 0.

11. A large rectangular metal sheet of mass M rests on a horizontal frictionless surface and
is connected to a wall by a linear spring with stiffness coefficient k (Fig. 7.1).

Figure 7.1
 The homogeneous equation with periodic coefficients 193

A small metal block rests atop the sheet so that their centers-of-mass coincide. The frictional
force between the block and the sheet is proportional to the difference in their velocities.
Suppose the sheet is slowly pushed to the left and released with the spring under compression.
What will be the subsequent motion of the block? How wide must the sheet be in order
that the block not fall off?

7.3 THE HOMOGENEOUS EQUATION WITH PERIODIC COEFFICIENTS

If A is periodic on —x% < t < + with (minimal) period T > 0, the questions
of Section 7.1 can be answered in principle for the linear periodic system

x’ = A(t)x. (7.3a)
The study of this equation can be motivated by an ad hoc discussion for the case
that A(t) = a(t) is a scalar. In this case, every solution x = ¢(f) has the form

¢(t) = cexp | a(s) as|:

We take c = | and explicitly observe that

+7
/ a(s)ds = XT, aconstant,
t

for all ¢ since a(t) has period T. Thus

t (ere
g(t + T) = exp || a(s) as- exp || a(s) as= o(ter”.
0 t
The number } in this equation has a profound effect on the nature of the solution ¢.
Notice, for example, that ¢ is periodic with period 7 if and only if e’” = 1, that is,
if and only if \ = 2nai/T. The number } is called a characteristic exponent for the
equation x’ = a(f)x, and e*” is called the corresponding characteristic multiplier.
The notion of characteristic exponent can be used to represent the solution ¢ in
a particularly useful form. Define a new function p by p(t) = o(the®’. Then p
has period T since
pit ae lin) aa o(t es Teg te = die tes? = o(t)e™ = p(t).

Thus
g(t) = p(tye™. (7.3b)
The significance of this equation is that every nontrivial solution of x’ = a(t)x
can be represented as the product of an exponential function and a function of
period T. This representation allows one to study the behavior of all solutions on
the interval —~1 < t < +o. We noted above that ¢ has period T if and only if
\ = 2nri/T. Similarly, ¢(1) > 0 as t— +e and becomes unbounded as tf —x%
if and only if Red < 0.
 Qualitative behavior of solutions for linear equations Hess
194

It is possible that ¢ is periodic with a period other than T. In fact, suppose that
\ = i8 and that the ratio BT/2m is a rational number, say r/s, where r and s are
relatively prime integers and s > 0. Then P
o(t as sT) = pit a She. pee wa o(tye*”.

Since 8 = 2nr/sT, \sT = 2nri, and it follows that ¢ has period sT.
If \ = iB and if the ratio BT/2m is an irrational number, then ¢ is bounded for
—o <t< +o, but (t) does not approach zero ast > +o.
Example 1. The characteristic exponent \ of the equation

x
S sin ¢ cs
~ [2 + cos? f]1/2
iS
27
1 sin¢
Ve an cae
The integrand is not the derivative of an elementary function. It is, however, an odd
function of period 2x. Hence \ = 0. One sees from the representation (7.3b), that
every nontrivial solution
t
sin t
¢(t) = cexp| » Db cos? Ai a

has period 27. ||

When A(t) is ann X n matrix with n > 1, the fundamental theoretical tool in
the discussion of x’ = A(f)x is the representation theorem of Floquet.
Theorem 7.3 (Floquet). Let A be continuous on —w% < t < +0 with minimal
period T > 0. There is a fundamental matrix solution ® for the equation

x = A@s
which has the form

Bf) = Pe’,
where P is a matrix function having period T and B is a constant matrix in Jordan
canonical form.
Proof. If ® is a fundamental matrix solution, then W(t) = @(t + T) is a matrix
solution since W’(t) = ®(¢+ T) = At +T7)6(¢ + T) = A()V(t). By Abel’s
formula (4.5b),
t

det V(t) = det ®(¢ + T) = det B(7)- exp || TrA(s) as~ 0.

0
Thus W is in fact a fundamental matrix solution, and it follows from Theorem 4.7
that there is a nonsingular matrix C such that (¢ + T) = W(t) = BAC.
 The homogeneous equation with periodic coefficients 195

It is known from linear algebra that there is a matrix B such that C = e®7,

Define P(t) = &(t)e®". Then P is periodic with period T since

PO T= OC- Te FF ™ = aCe Bre
SO CC=Cn =| Oho. y=-7 (1).
By construction @(t) = P(t)e®*.
If B is not already in Jordan canonical form, let Q be a matrix such that O~!BOQ
isin Jordan form. Then @,(t) = 6(¢)Q is a fundamental matrix solution by Theorem
4.6. But
21 PO OOr50 (Pi) Ooo Oe Pine,
where P(t) = P(t)Q has period T and B,; = Q~'BQ is in Jordan form. ||
If & is a fundamental matrix solution for x’ = A(r)x and if @(¢ + T) = BNC,
then C is called a period-transformation matrix for the fundamental matrix solution ®.
For a given system, all period transformation matrices have the same eigenvalues. To
see this, let Y be a second fundamental matrix solution with period transformation
matrix D. By Theorem 4.7, there is a nonsingular matrix Q such that W(t) = ®(r)Q.
Since W(t + T) = V(t)D, ®(t + T) = B(ODQ~!. This means that C = QDOQ™},
whence C and D have the same eigenvalues. The eigenvalues of the period transforma-
tion matrices are called the characteristic multipliers of the system.
If wy,...,@n are the characteristic multipliers of the system x’ = A(t)x and
if C = e®, where C is a period transformation matrix, then the eigenvalues
h1,---,An Of B satisfy the relations

bn = eds? (hearts Seen): (7.30)

This may be established by showing that if \ is an eigenvalue of B, then e*” is an
eigenvalue of C. In fact,

det (C — ef) = det (e?" — e'T) = det (ee 7)F J) det oe

‘got 2%.
det (B — ADF-cet (@ = M) arAoi

ents of the equation x' = A(1)x.

The numbers \,,..., A» are called characteristic expon
Example 2. Consider the system
oa 0 —sin? t}] x},
; e Bay 0 id ee
By Floquet’s theorem, each solution is of the form

x(t)| _ |pil) Be K a |
Fal i Be Pale st. Gal
for some constant vector c.
 Qualitative behavior of solutions for linear equations 7.3
196

To evaluate \j, \e and Y, let S(t) = P(t)e?' be a fundamental matrix solution.

Since sin(—f) = —sin?,; it happens that V(r) = P(—te—®* is a fundamental
matrix solution also. Now (0) = P(O) = ¥(0), and it follows from Theorem 4.1
that (1) = V(r). In particular, P(r)e?™ = P(—m)e 8". But P(r) = P(—7). Thus
e238 — 7 and }, and Xz are integral multiples of 7, while Y = 0. Consequently the
system (7.3d) has a fundamental matrix solution of period 27 and all solutions are
periodic with that period. ||
The discussion in the example above can be modified slightly so as to yield a
theorem the proof of which we leave as an exercise.
Theorem 7.4. Let A be a continuous matrix function of period T for —-» <t<
+o and assume that A is odd, that is, A(—t) = —A(t). Then all solutions of

Kune LB)

are periodic with period T.

The next theorem, which generalizes the discussion of x’ = a(t)x above, shows
how the characteristic exponents of x’ = A(t)x affect the behavior of its solutions.
Theorem 7.5. Consider the system (LHP) x’ = A(t)x, where A is continuous and
periodic for —% <t < +o with (minimal) period T > 0. Let ®(t) = P(t)e®* be
a fundamental matrix solution in Floquet form, and assume that \ = a + i is an
eigenvalue of B of multiplicity m. Corresponding to ), let B have k and no more than
k linearly independent eigenvectors. Then

i) If a> 0, (LHP) has m, and no more than m, linearly independent solutions

corresponding to \ which approach zero as t > —« and become unbounded
HS 1 Sea Oop
ii) If a < 0, (LHP) has m, and no more than m, linearly independent solutions
corresponding to \ which approach zero as t > +o and become unbounded
GOK [L— Cc

iii) If a = 0, then (LHP) has k, and no more than k, linearly independent solutions
corresponding to \ which are bounded for -~ <t< +o.
iv) If a = 0 and if BT/2r is a rational number, then all linear combinations of the
solutions in item (ili) are periodic with common period qT, where q is a positive
integer.

Proof. We may assume that B is in Jordan canonical form and that ) is an eigenvalue
of B appearing in the first m columns. The information in the hypotheses allows us
to write down the matrices B and e* which are shown in Figs. 7.2 and 7.3, respectively.
The (x — m)X(n — m) submatrix R has the eigenvalues of B other than \ on its
diagonal, some distribution of zeros and ones above the diagonal, and zero entries
elsewhere.
We obtain solutions from the Floquet representation ®(f)e = P(t)e?“c by
choosing c in different ways.
 eB!
<a)

S
TSOS
I!

eeeete
[ieee

:
L

Sheree

|
K

S|

—
|
y
|
ie

a
atSon |
Etpe | FE e
wee
E
| = S PR =
| | =
: fy
50 ee
iil
| Ss 52 : a |
ee:
Sloa é
|

Figure 7.3
iy, | I a Pe| esate
— s ee SS
oF
a
aes teasaoe ete Sa CN, . —_—_—,
ase | | | =
Te OO ee ee Se Se tn ha ee Se)
| a x~<
| )
Y
— |
Nee eee
=. oe ah ae
| oO eves ney | ee

: |
: ati |
2
|
wm aa
za are
omen!
me, —
a oie —“—_
i | Ie
=

|
te s Re 2
=~ | S
A ah
age eae
ag
is)
x
The homogeneous equation with periodic coefficients

NS)
= oO
197
 Qualitative behavior of solutions for linear equations Th
198

To prove (i), set any component of ¢ from the first to the mth equal to one, and
set all other components equal to zero. Then the solution P(t)e®'c has the form
P(tyj(c, + Col = hss a5 et? !)e,

wherej is some integer between one and m and ¢;,...,¢; are constant vectors. If
a < 0, P(t)e?“c > 0 as t— +m since P(t) is bounded and tte!_»0ast—> +o
for integers » > 0. The solution P(t)e®‘c becomes unbounded as t > + since
P(t) is periodic and e’ + +m as t— —x. The proof of (il) is analogous.
To prove (iii), note that e?’ has exactly k columns corresponding to \ which do
not contain positive powers of ¢, say, pie’, poe’, ..., pee’. Each solution P(t)pje™"
is bounded for —~» < t < +o since |P(t)p;e*| = |P(t)p;| |e?"| = |P(p,|. It is
worth noting that the functions P(t)pje*’ (j = 1,...,k) form a maximal set of
linearly independent bounded solutions corresponding to the eigenvalue \. Thus,
any bounded solution which is linearly independent of these must correspond to a
different eigenvalue of B.
To prove (iv), consider one of the bounded solutions P(t)pje'"! Re Pree
and assume that 67/27 = p/q, where gq > O and p are relatively prime integers.
Then
P(t + qT )pje®
Te = P(t)pje’?7? = P(t)p;. ||
Corollary 7.5. A necessary and sufficient condition that every solution of (LHP)
approach zero as t—» + is that every characteristic exponent have negative real
part. Let e®” denote a transformation matrix for the equation. A necessary and sufficient
condition that every solution be bounded over intervals of the form —»n <t< +o
is that (i) every characteristic exponent have zero real part and (ii) B have n linearly
independent eigenvectors.

The proof of this corollary is very similar to that of Corollary 7.1, and it is left
to the exercises.
The last theorem shows the way in which the characteristic exponents (hence the
characteristic multipliers) of a linear periodic system determine the behavior of its
solutions. To utilize this theorem for the study of a specific system, however, one
must have some method of determining the characteristic multipliers. This is a
difficult problem and methods are known only in special cases. For example, Theorem
7.4 guarantees that all the characteristic multipliers of a periodic-coefficient equation
x’ = A(t)x equal one and that every period transformation matrix is diagonalizable
if —A(t) = A(—t). Even in the general case, however, Abel’s formula provides
some helpful information.
Theorem 7.6. Let u1,..., bn denote the characteristic multipliers of a periodic-
coefficient equation x’ = A(t)x. Then
vi
Pible ..» ply = EXD i Tr A(s) as Cie)

where T is the minimal positive period of A.

 -

The homogeneous equation with periodic coefficients 199

Proof. \f ® is the fundamental matrix solution which satisfies the initial condition
(0) = J and if C is a period transformation matrix for 6, then @(t + T) = (ZC.
Thus ®(T7) = #(0)C = C, and it follows that the eigenvalues of ®(T) are the char-
acteristic multipliers of the differential equation. But the determinant of a matrix
is the product of its eigenvalues. By Abel’s formula,
Tt

[io - ++ Mn = det O(T) = det (0) exp || Tr A(s) as|. |

0

Example 3. Consider the system

ae 4cost eS |x).
;] ee eran: eer ia i
The coefficient matrix has period 27. If uw, and ws denote the characteristic
multipliers of the system, then

Mike = exp | (—1 + cos 2) dt = e~*

0

by Theorem 7.6.
The subsequent analysis of the system is based upon the fact that one solution is
known: namely,
Sh pad,
sin=
2
o(t) = ze
é,'” Cos =
2

By Floquet’s theorem, the first component of ¢(t) must be of the form

cipi(tye™’ + co(vtpi(t) + po(t))e",

where p,(t) and po(t) have period 27, Y is either zero or one, and c, and cg are
suitable constants. That is,

ens sin 5= erpi(te + co(vtpi() + pr(t))e*".

This means that at least one of the numbers \; and go, say \4;, must have —4 for its
real part.
Now write \; = —4 + iB; and \2y = ag + iB. From Eq. (7.3e), it follows
that
e-?" = wipe = exp[—m + i281 + 2a + i2nBo],
where pw, and po are the corresponding characteristic multipliers. Thus, for some
integer k,
—2ne + 2rki = —a7 + 2raq + i27(B1 + Bo),

whence ag = —¥% also.

 Qualitative behavior of solutions for linear equations Teo
200

From Theorem 7.5, we may conclude that every solution of the system (7.3f)
approaches zero ast—> +o. ||
Example 4. The equation x/’ + p(t)x = 0, where p has minimal period T > 0,
is called Hill’s equation. When converted into a system, it takes the form

| - rey 0] . (7.38)
The characteristic multipliers satisfy the equation
T
Mise = exp || (O + 0) as= |. (7.3h)
0
If corresponding characteristic exponents are

Ay = ay + 18 and Ag = a2 + iBa,
then
|e 1| =— eu and ord = e%27,

Thus exp [(a; + ag)’] = 1, and it follows that ay = —ay. Equation (7.3h) now
assumes the form

exp [(a@1 + 181)T] exp [(—a1 + i82)T] = I,

which implies that
exp [i(81 + 82)T] = 1.
Hence (8; + 62)T = 2rk for some integer k. If a; # 0, then \; + Xo, and the
system (7.3g) has two linearly independent solutions of the forms
pr(tye**e! and Doijennven we ee

where p(t) and po(t) have period 7. One of these solutions is unbounded as t> +0;
the other approaches zero. If a; = 0, then according to Floquet’s theorem there
are two linearly independent solutions having the forms

and (vtp4(t) —L Pol) je ere

pide’!

If Y = 0, each of these solutions is bounded. Each solution will also be periodic if,
in addition 6,7/2r is rational. If Y= 1, then 8; = By. Hence 8, = rk/T for some
integer k, and the solutions are

pile™7 — and (tp, () + pot) )e*™™7.

The first solution has period T if k is even and period 2T if k is odd. The second
solution is unbounded as t-—» +o. Theorem 7.5 thus makes the calculation of
only one characteristic exponent sufficient for the study of Hill’s equation. Much
research has been done on methods for calculating or estimating the one required
exponent. ||
 The homogeneous equation with periodic coefficients 201

Continuing the question of determining the characteristic multipliers for a periodic

system, it is well to make the observation that what is really needed is the value of a
fundamental matrix solution at two points t, and fo, exactly one period apart. In
fact, let @ be a fundamental matrix solution and let C be a period transformation
matrix. Suppose &(t;) and ®(t2) are explicitly known, where tg = t; + T. Then
P(t2) = B(t1)C, so C = &~'(t,)®(ty), and the characteristic multipliers can be
explicitly calculated.
A special case in which characteristic exponents are easily found is the case in
which

A(t)- [ A(s) ds = ihA(s) ds - A(t). (7.31)

In this case a fundamental matrix solution for x’ = A(f)x is simply

exp i A(s) as|?

and the characteristic exponents are (mod 27/) the eigenvalues of the matrix
T
/ A(s) ds.
0
The 2 X 2 matrices A(t) which satisfy the commutativity condition (7.3i) are
those of the form
iO CO) a),
a(f(t) — g()) g(t)
where a and # are any constants. This observation is not very useful in applications;
compare it with the scalar case at the beginning of this section, however.
Example 5. To find the characteristic multipliers of the system
/
xX COS fy;
y' = (1 + cos ty,
we note first that

sin f t cos f 1 _ [cost 1 Ee t |.

0 ¢-+sint|}| 0 1e--¥cos7
| Pl" 0 ts=costi] 0 t+ sint

Thus the characteristic exponents are (mod 27/) the eigenvalues of the matrix
27
shel oe 1 |ai ite i ‘I
2m Jo 0 1 + ccst Ont

These are \ = 0 and \ = 1. The corresponding multipliers are uy = 1 and uw = e.

Thus the given system has one solution of period 27 and a second, linearly inde-
pendent solution which becomes unbounded as t—> +o. ||
202 Qualitative behavior of solutions for linear equations 7.3

When studying a periodic-coefficient equation x’ = A(f)x, it is plausible to

suspect that the eigenvalues of A(t) might determine the behavior of solutions in a
simple way. There is a relation between the eigenvalues of A(t) and the characteristic
exponents of the system, but it is a difficult one. Suppose, for example, that ®(t) =
P(t)e®! is a fundamental matrix solution for x’ = A(t)x. Then
P'(phe®* -- P(t) Be?* = A(t)P(t)e?*,
and
P'()P 3) + PG)BE *@) = AG. (7.3))

It is clear from Eq. (7.3j) that the eigenvalues of A(t) do not coincide in general with
the characteristic exponents of the system, that is, with the eigenvalues of B. In
particular, all eigenvalues of A(t) can be negative for —w <t< +o and yet
there can be solutions of x’ = A(t)x which do not approach zero as t— +o.
Example 6. The equation

has for a solution vector

C4, sin
o(t) =
e!* cost
Z
The solution @ is unbounded as t > +o, yet \ = —+ is a double eigenvalue of the
coefficient matrix. ||

EXERCISE 7.3

1. Find the characteristic multiplier and a characteristic exponent for each of the equations
below. Describe the behavior of solutions as t > +.
a) x’ = (1 — 2sin dx. b) x’ = (—2 sin? Ax.
c) x’ = V1costx.
—

2. Give a necessary and sufficient condition in terms of a that x’ = a(1)x have only periodic
solutions.
3. Discuss the behavior of solutions as t > +00 for the equations
a) x’ = (cost + i7/3)x, b) x’ = (cost + im)x.
4. Solve the following systems by the method of integrating factors and express their
solutions in Floquet form. What are the characteristic exponents and multipliers?

D [oiae eave isk DL ho hacar mare

5. Use Floquet’s theorem to explicitly solve the following systems of equations:
“) * = |—sintcost 1-+ sin2¢][x
y —1—cos?¢ sintcost || y]’
 The homogeneous equation with periodic coefficients 203

b) Hf _| Gsin?d - 1 (3 sin¢cos tf) + 1] [x].

y (3 sin t cos t) — 1 Gcos*7) = | y
6. How can one be assured, without solving the equations, that system (b) of Problem 5
has at least one solution which is unbounded as t > +? Try to deduce the behavior of
solutions for system (a) in the same way.
7. Prove Corollary 7.5.
8. Show that the system of differential equations in Example 6 has one nontrivial solution
which approaches zero as t > +o.
In Problems 9 through 14, consider Hill’s equation
x'’ + p(t)x = 0 (H)
where p is real, periodic with minimal period T > 0, and continuous.
9. Suppose that the equation (H) has a nonreal characteristic multiplier u1. If wo denotes
the other characteristic multiplier, show that Z2 = uw; = e”® for some real 6. Show that
the equation has a fundamental solution set ¥1, Y2 = W1 such that Wi(t + T)=Wi(de®
and that all solutions are bounded for —~ < ft < +o.
10. A solution w of (H) is called normal if there is a number yu such that Y(t + T) = py(2).
Suppose the characteristic multipliers for (H) are real and neither equals +1. Show that
there are two linearly independent normal solutions.
11. Show that (H) has at least one normal solution even though » = +1 is a characteristic
multiplier.
12. Let Y be a normal solution of (H) which has no zero in the interval 0 < ¢ < T and
assume that fo p(t) dt > 0. Show that all solutions are bounded for —» <t < +o.
[Hint: Integrate by parts to deduce that fo (W/()/W())? ds < 0 and conclude that y is a
nonreal-valued solution. Then y is a normal solution. Complete the proof by showing
that the characteristic multipliers are nonreal roots of unity.]
13. Let Y be a normal solution of (H) which has a zero in the interval 0 < ¢ < T and
assume that fo p(‘) dt < 4/T. Show that every solution of (H) is bounded for —« <
t < +o. [Hint: Assume for contradiction that y is real-valued and let Y(a) = 0. There
exists a time 5b, a < b < a+ T such that ¥(d) = 0 and y(t) ¥ O for a < t < b [why?].
If |W| attains a maximum at £ in (a, 5), then there are points ¢) and f2,a < t1 < E<t2 <4,
such that the graph of || is concave on (f1, fz). Thus

1) VC a4 W'(ts) < W(t2) — ¥&)

ae
|
W(t) 2 ra ree
Establish the inequality

4 4 1 1 1 HO Ee)
<4 ct teal f= Th Pie fs
in) — Wt) Lf woes [| Oly
ear yee y.
t A
b
z-i 4
W’( s)| GS il|p(s)| ds = ; |p(s)| ds < =
<< / ws)

to obtain the contradiction. Next show that the characteristic multipliers are nonreal roots
of unity.]
204 Qualitative behavior of solutions for linear equations 7.4

14. Prove that every solution of (H) is bounded if

.
if
o<f p@)des ="
0 4

15. How do the solutions of x’’ + (? — cos? Ax = 0 behave ast > +2?

7.4 THE HOMOGENEOUS EQUATION WITH CONTINUOUS COEFFICIENTS

The problem of determining the qualitative behavior of solutions for x’ = A(t)x

when A is merely continuous on an interval a < ¢f < w can be much more difficult
than when 4 is also periodic. The theorems of this section apply, of course, to the
case of continuous periodic coefficients. We make no use, however, of Floquet’s
theorem.
Several definitions are necessary for the statement of the next theorem. Before
giving them, consider, as a motivating example, the function f defined by

aCe ¢ + ‘)sin 27x, BAe hs

which is graphed in Fig. 7.4. Even though lim, 4. f(x) does not exist, it is still
possible to describe the limiting behavior of f quantitatively. To do this, let x be
restricted to an interval t< x < +0, t¢> 1. Then the range of f for x > tis
the projection of the graph of f, x > t, onto the y-axis. No matter how t > 1 is
chosen in the example, this projection will be a closed, finite interval J(t). We denote
the upper endpoint of I(t) by M(t) = supz,; f(x) and the lower endpoint by m(t) =
infz,; f(x). The expressions “sup” and “inf” are abbreviations for the words
supremum and infimum, respectively. Notice that, if t; > t, then J(t;) is contained
in I(t), that is,
m(t) < m(t1) <0 < M(t) < M(0).
se = (2=> *)sin 27x

—_—
m(t) = m(t) a
aa
a
i
Uf

Figure 7.4
 -

The homogeneous equation with continuous coefficients 205

Since monotone functions which are bounded for 1 < t< +H approach limits
as t—> +a, lim, 4. m(t) and lim,_,,.. M(t) exist. We denote these limits by
lim sup f(x) and lim inf f(x).
r+ xr +0

They are called the /imit superior and limit inferior, respectively, of f as x > +a,
and they have the values +1 and —1, note.
Let us now define lim sup,_,,. g(x) for an arbitrary real-valued function on
an interval a < x < +o.
A set of real numbers S is said to be bounded above (below) if there is a number
M(m) such that x < M (x > m) for every number x in S. The completeness axiom
which is assumed for the real number system R! is that every nonempty set of real
numbers which is bounded above has a /east upper bound. The least upper bound
of a set S is called its supremum and is denoted by sup S. A dual form of the com-
pleteness axiom is that every nonempty set of real numbers which is bounded below
has a greatest lower bound. The greatest lower bound of a set S is called its infimum
and is denoted by inf S. If S is not bounded above (below), one writes sup S =
+o (infS = —o).
One says that a real-valued function is bounded above (below) if its range is
bounded above (below). A function which is bounded both above and below is
simply called bounded, and this is the sense in which we have previously used the term.
It follows immediately from the completeness axiom that if M is a bounded,
nonincreasing function on an interval a < t < +o, then lim; _,,. M(t) exists, the
value of the limit being the infimum of the range of M. A dual theorem for non-
decreasing functions is, of course, true.
Now assume that g is bounded above for a < x < +. For each t > a, then,
the set R; = {y: y = g(x) for some x > f¢} is bounded above. If M(t) = sup R:,
then M is nonincreasing. If M is bounded below, then lim;,,., M(t) exists.
Otherwise lim;,,. M(t) = —o. In either case, we define lim sup, 4. g(x) =
lim,;_,,. M(t). Ifg is unbounded above on the interval t < x < +, for arbitrarily
large t, we define lim supz_.4. g(x) = +o.
The notion of liminf,.,. g(x) is defined analogously, and we observe that
lim,_,.. (x) exists if and only if lim supz.4.« g(x) = lim inf,,.. g(x), where the
indicated limits are finite.
Theorem 7.7. Suppose that A is real-valued and continuous on an interval a <
t < +. and let m(t) and M(t) denote, respectively, the smallest and largest eigenvalues
of A(t) + A’(t). If, for some to in (a, +),
t

lim [ maa = 53
t+ J tg

then every solution of x' = A(t)x approaches zero as t—> +a. If, for some to in
(a, +0), t

lim sup | m(t) dt = +o,

t+-+0 to
 Qualitative behavior of solutions for linear equations 7.4
206

then every solution of x' = A(t)x is unbounded as t— +o.

Proof. If @ is any solution of x’ = A(t)x, let ||@(1)/| denote the euclidean norm

GQ) tae oa
Then ¢(t)> 0 as t— +o if and only if ||¢(t)|| 0 as t> +o. In matrix
notation
leA||? = o Oe.
Thus

aol = ¢' (61) + &(Od)

= $ (NADGD + MANO
¢ (H(A) + AO) 00.
Now A(t) + A?(ft) is a symmetric matrix. If m(t) and M(t) denote its smallest and
largest eigenvalues, then it is well known from linear algebra that

m(t)||x||? < |[x7(A@ + A7()x\| < MCO|[x||?

for any n-vector x. Thus
d
a [leO||? < MO|leOl!?, (7.4a)
and solving (7.4a) with the integrating factor exp is M(t) dt, we obtain
t

|4(O1 < lod exp|| mes) as). 0

(7.4b)
Similarly,

lotto exp |fms) a5]< oO

If
+0

M(s)ds = —~m,
to
then
lim ||¢()|| = 0.
t>-+0

If imsup 272. i, m(s)ds = +o, then @(t) is unbounded as t— +o. ||

Corollary 7.7. Let the hypothesis of Theorem 7.7 be satisfied. If
t

lim sup | M(t)dt < +o,

t+ to

then every solution of x' = A(t)x is bounded for ty < t < +o.
 *

The homogeneous equation with continuous coefficients 207

Proof. This is an immediate consequence of inequality (7.4b). ||

Example 1. Consider the system

, |—sintt—costt . 3
x
= D sill COS tal x
. (7.4c)
y —cos t sin? t —cos” t sin? t}Ly

The coefficient matrix is periodic, but the Floquet theory is difficult to apply.
Theorem 7.7 provides information about solutions much more readily. First one
solves the eigenvalue equation

=<in’t— cos*f— sin tcos® t'— cos ?sin®t

sin tcos? t — cos
ftsin? ¢ = Cosc Sin fk
= \7 4 \-esin-
feos’ 1 = 0
to find that

M(t) = —$ + glcos2¢|, m(t) = —}4 — |$c0s 21|.

The reader is asked to verify
t

lim exp | MS) ds=20

t>-+0 0

in the exercise below. It follows from this that every solution approaches zero as
to>+oa. ||
Example 2. Let us show that all solutions of the equation x’’ — (e~' — 1)x = 0
and their derivatives are bounded for0 < t << +m. If x’ = y, then

Hie =val)
The matrix A(t) + A7(f) is
OMen:
én 10

with M(t) = e~’ and m(t) = —e~‘. Since

t

lim exp | M(t) = limexp (1 — e~‘) = e,

tn” (0) t>-+0

Corollary 7.7 guarantees that |@(1)| and |’(t)| are bounded over 0 < ¢ < +o for
every solution x = ¢(t). ||
Example 3. Consider Hill’s equation x’’ + p(t)x = 0, where p is continuous on
—« <t< +o and has minimal period T > 0. When converted into a system,
 Qualitative behavior of solutions for linear equations 7.4
208

The matrix A(t) + A7’(f) has eigenvalues

M(t) = |L—p@|, m= —|! — pO),

t v

im | [1 — p(s)| ds = +a and tim | —|1 — p(s)|ds = —w

t—00 — 00

so Theorem 7.7 does not apply. The reader will recall from Example 4 of Section 7.3
that if Hill’s equation has one nontrivial solution ¢; such that lo1(2)| + |6;| - 0
as t— +o, then it has a second linearly independent solution ¢2 such that
|2(t)| + |63(t)| becomes unbounded as t—> +. Ina specific case, Hill’s equation
might have all its solutions bounded, but Theorem 7.7 and its corollaries do not
provide a sufficiently delicate test for the detection of this boundedness. _ ||
In many cases, it is difficult to apply Theorem 7.7 and its corollary because the
eigenvalue M(t) is difficult to compute. The theorem and its corollary remain true
if M(t) is replaced by

MC) = max JReai(+Dyaxl}, where AC) = [a0

Sa i#i
The proof, which makes use of the rectangular vector norm rather than the euclidean
norm, requires more mathematical background than is assumed here, and we omit it.
A well-studied special case of the system x’ = A(t)x is the case in which one can
decompose A(t) into a sum A(t) = L(t) + R(t), where L is either periodic or constant
and R(t) is “small” as t—> +o. In such a case, one would hope that the solutions
of x’ = A(t)x would behave (for ¢ sufficiently large) in the same way as the solutions
of x’ = L(t)x. For example, it is not implausible to suspect that all solutions of

wi (Ft 2)y ttt x= .0

approach zero as t—> +o since the solutions of x’’ + 2x’ + x = 0 do so.

Before stating and proving the standard theorems which apply to this special case,
let us explicitly make two important observations:
1. If Lis a constant matrix, then every fundamental matrix solution @ for x’ =Lx
has a trivial Floquet representation &(1) = P(t)e®', where P(t) is the identity matrix
and B=L.
2. If P is a continuous, nonsingular, periodic matrix function, then there isa
constant ¢ such that |P~'(t)| < c for =-« <t< +o. To verify this, one needs
only to examine closely an algorithm for computing P~'(f). If P;;(t) is the cofactor
of the element in the ith row and jth column of P(f), then the element in the ith row
and jth column of P~!(f) is
Pit) |
det P(t)
 -

The homogeneous equation with continuous coefficients 209

The numerator in P;,(t) is a sum of products of continuous, periodic functions.

Hence it is bounded. The value det P(t) is bounded away from zero since P is non-
singular and periodic. Therefore |P~ '(t)| is bounded.
One applies these observations to the equation

SSI Gaye

where L is either periodic or constant in the following way. Let &(t) = P(t)e®*
denote a fundamental matrix solution in (perhaps trivial) Floquet form. Suppose it
is known that all solutions of the equation approach zero as t > +o. Then all the
eigenvalues of B have negative real parts. Thus, there exist positive constants M,
and 7 such that |@(1)| < |P(t)|- My,e—™ for all t > 0. Further, there exist positive
constants M, and M3 such that |P(‘)| < Mz and |P~'(t)| < Ms for all t. Con-
sequently
|B(t)| < cye"”, Cc, = M,\Mo, (7.4d)
for all t > 0, and

Jee “(s)| = |P@e*te**P”

*(s)| (7.4e)
SEO Paes isin,
Co = M,M.2Ms3, for all s and t with t > s.
The estimates (7.4d) and (7.4e) are used to study x’ = (L(t) + R(t))x by way
of a modified variation-of-parameters algorithm.
Lemma 7.8. Suppose the coefficient matrix A(t) of the equation x' = A(t)x can
be written in the form A(t) = L(t) + R(t), where L and R are continuous on an interval
a<t< +o. Let ® be a fundamental matrix solution for x' = L(t)x and let tp > a
be arbitrary. If ® is any solution of x' = A(t)x, then

H(t) = BE "(t0)b(to) + O(Y) / b~"(s)R(s)o(s) ds (7.4f)

[ROT
lo oe
Proof. One performs the variation-of-parameters computation on the equation x’ =
L(t)x + R(t)x as though R(t)x were a nonhomogeneous term. This yields

&'()x’ — ©'OL@x = & “@ROx,

&(t)x’ — 1) 8/(Ne
(Nx = & (R(x,
db \(t) l| &—'(t)R(0)x,
&—*(t)x’ +
irae =
£ (@ (0x) = 6 1()R()x.

The conclusion then follows by integration. ||

 Qualitative behavior of solutions for linear equations 7.4
210

Theorem 7.8. Suppose the coefficient matrix A(t) of the equation x' = A(t)x can
be written in the form A(t) = L(t) + R(t), where
i) Land R are continuous fora <<t< +,
ii) either L is constant or L is periodic with minimal period T > 0,
iii) |R(t)| ~ 0 as t> +o,
iv) every solution of x’ = L(t)x approaches zero ast > +a.
Then every solution of x' = A(t)x approaches zero as t— +o.
Proof. Let ¢ be any solution of x’ = A(f)x and let &(7) = P(t)e®' denote a funda-
mental matrix solution for x’ = L(t)x in Floquet form. Let cg and 7 be defined as
in inequality (7.4e) and choose a constant cz > 0 so small that coec3 < 7. Since
|R()| +0 as t— +o, there is a positive to > a@ such that |R(t)| < cs for all
t > to. Express @ by means of (7.4f) and take norms of the resulting equation
to obtain ;

lp(t)| < e4|@(4)| + cats | e— "| 6(s)| ds, Pe he

to

where cy = |®~'(to)o(to)|. Applying the estimate (7.4d), we have

t

|o(t)\e”’ < C1C4 + cats | |p(s)|e”" ds, t > To.

to

By Gronwall’s inequality (Problem 9, Exercise 4.2),

lee” < cycge2"3"—

fOTrat alge ous

lim [6(O| < exeqe- lim eM! = 9, ||

t>-+0 t+ 00

Example 4. Let us write the equation

wrt (F— 2)x4 (et 4 Ie =0 (7.4g)

as a system
Balt 0 l x
| eeet
The coefficient matrix may be written as a sum L(t) + R(t) for t > 0:

On 0 0
 Pa

The homogeneous equation with continuous coefficients 211

All solutions of

>| s Oey
TA eal 24 [by
approach zero as t > + as does the matrix

0 0
=
Rie ba
te
t

Thus every solution ¢ of Eq. (7.4g) has the property that |¢(t)| + |¢’(t)| > 0 as
t—> +o. ||

Example 5. The system

Nes hetain eco

x} |—2— cost > + sin t||\x
we
provides an interesting illustration of the extent to which one can weaken the hy-
pothesis (iii) of Theorem 7.8. We saw, in Example 6 of Section 7.3, that the system
(7.4h) has a solution

o(t) = EE sine»
2,
e!/* cos Al
2

which is unbounded as t— +c. When the coefficient matrix is decomposed into

the sum
a me —cos? sin t
—3% -—}j sint cost

we see that the eigenvalues of the first term L(t) have negative real parts. Thus every
solution of

y r aoe
| Se eal
A|;

approaches zero as t > +. The second term R(t) is bounded and does not approach
zero as t—> «. This example shows that some generalization of hypothesis (iii) is
necessary for the truth of a result such as Theorem 7.8, and the mere boundedness of
|R(t)| is not enough. ||

Example 6. Consider the system

all ase Datel psa Sint x

ss ’ uF (7.4i)
y aoepenne nt —1 + 5cos+t y
 Qualitative behavior of solutions for linear equations 7.4
212

for t > 0. The coefficient matrix can be written in the form L(t) + R(t), where

dcost 2— 2sint
Henares —1-+ Zcost
and

1 Ee =
RG = 2s ;
ee Foe 0
D 2 —e-#
It is clear that |R(‘)| — 0 as t > +o, and it was shown in Example 3 of Section 7.3
that every solution of x’ = L(1t)x approaches zero as t—> +. By Theorem 7.8 so
does every solution of the system (7.41). ||
The preceding theorem has a companion which gives conditions under which
the solutions of x’ = A(t)x are merely bounded as t— +o.
Theorem 7.9. Suppose the coefficient matrix A(t) of the equation x' = A(t)x can
be written in the form A(t) = L(t) + R(t), where
i) L and R are continuous ona <t< +o,
ii) L is either constant or periodic with minimal period T > 0,
00

iii) / |R(s)| ds < +o for some ty > a,

to
iv) every solution of x' = L(t)x is bounded over the interval to < t < +a.
Then every solution of x’ = A(t)x is bounded over the interval tp < t < +o.

Proof. By Lemma 7.8, each solution @ of x’ = A(f)x satisfies the equation

H(t) = BNO "(to) (to) + H(A) i &~'(s)R(s)o(s) ds, (7.4))

where ®(t) = P(t)e®' is a fundamental matrix solution for x’ = L(t)x. By hypoth-
esis (iv), B(¢) is bounded over the interval tp < t << +o, for any tp > a. By
observation 2 above, @—'(t) is bounded over the same interval. Thus, for t > tos
t

lo] < cr + eo | [R(s)| |6(s)| ds, (7.4k)

where |®(t)b~*(to)P(to)| < cy and |&(t)}6—1(s)| < cp.
Now let ¥(t) = ¢1 + ¢2 Ji |R(s)| |o(s)| ds. Inequality (7.4k) becomes y/(t) <
C2|R(|¥(4), which can be solved with the integrating factor exp cs fi |R(s)| ds to
yield

lo(2)| < ¥(t) < c1 exp E if|R(s)| as]. (7.41)

Since lim;_,.o Ii |R(s)| ds < +, the proof is complete. ||
 -

The homogeneous equation with continuous coefficients 213

Example 7. All solutions of the equation

are bounded over every interval tg < t < +, where fo > Ois arbitrary. To verify
this assertion, consider the associated system

56 | 0 il ||||3% 0 I ia 0) Ise
= ] —

The solutions of

are of the form

@(t) = [p, sin wl, po cos wi],

where p, and py, are constant vectors. The norm of the last matrix in the system
(7.4m) is t~”, and fi, 1-2 dt = tj! for any to > 0. Thus all the hypotheses of
Theorem 7.9 are satisfied and the conclusion follows. _ ||
Theorems 7.8 and 7.9 do not generally apply to the equation x’ = (L(t) +
R(t))x when L is neither constant nor properly periodic. If & denotes a fundamental
matrix solution for x’ = L(f)x, then either of these conditions on L allows one to
estimate the norms of ®(f) and & '(t) because ®(f) has a Floquet representation.
Theorems similar to 7.8 and 7.9 can be proved when L is neither constant nor periodic
provided |®(f)| and |&~'(1)| can be estimated in some other way.
Let us suppose, for example, that L is neither constant nor periodic but that ®(t)
is known to be bounded on intervals of the form fp < t < +o. Since every element
of &—'(t) is a sum of products of elements of ®(t) divided by det ®(t), one can con-
clude that ~!(f) is bounded as t—> + provided det ®(t) is bounded away from
zero. By Abel’s formula,
t

det b(t) = det (to) exp [| eS) as|:

Thus, &~'(t) is bounded on to < t < +c if &(f) is bounded on that interval and if
t

lim int |Tr L(s) ds 4 —ce.

t— +00 to

Let us use this information to prove a third result.

214 Qualitative behavior of solutions for linear equations 7.4

Theorem 7.10. Suppose the coefficient matrix A(t) of the equation x' = A(t)x can
be written in the form A(t) = L(t) + R(t), where
i) L and R are continuous ona<t<+on,
t

ii) lim int | Tr L(s) ds # —~m, for some to > a,

t>+n to
+0

iii) |R(s)| ds < +o,

to
iv) every solution of x' = L(t)x is bounded over the interval th < t < +o.
Then every solution of x' = A(t)x is bounded over the interval th < t < + a.

Proof. If @ is a solution of x’ = A(t)x and if @ is a fundamental matrix solution

fOr X= keener

(1) = &(1)@~"(t0)6(t0) + wo | &~"(s)R(s)o(s) ds, (7.4n)

by Lemma 7.8. The hypotheses (ii) and (iv) imply that (1) and 67 '(1) are bounded
as t—> +o. Thus, for t > fo,

J6(] < ex + 2 | IRG)| Jo6)

t

This is the same inequality as (7.41) in Theorem 7.9. As there, it follows here that
+a

|o(4)| < cy exp E / |R(s)| as|

0

for tp <t< +o. ||

An example of Perron shows that some hypothesis of the type (ii) in Theorems
7.8, 7.9 and 7.10 is necessary if one is to formulate a similar theorem: The solutions of
aoape —ii @) x
| fe Reem onesie ili Beet

approach zero as t—» +o. The only bounded solution of

, 11

|
y
= | nee
exp[—o¢] sinIn¢
U . t>0
+ cosInr — 44]| y|’

is the trivial solution however.

Example 8. Consider the equation

= ea |p te (7.40)
y Bteee sar tme liliy
 a

The homogeneous equation with continuous coefficients 215

In order to apply Theorem 7.10, the coefficient matrix must be decomposed into a
sum L(t) + R(t), where
+o t

/ |R(s)| ds << +a and lim / Tr L(s) ds ¥ —o.

0 ta to
A plausible choice for R(t) is
me.
1+ f|-
0 0

Fe ia
If we take ty = O, then

is ” ds
I. | (s)| S * 1 ae 52 2

Since
=o 40
me ree sb
ie

lim exp |/ Tr L(s) as|= lim exp [e? — ie] =+o.

t+ 0 to+oa

We may therefore conclude that all solutions of the system (7.40) are bounded for
0 < t < +o as soon as we know the same for the solutions of

Fbepoceclbh
But this system can be solved explicitly for x. In fact, x(t) = cye~”. Thus y’ =
317y + 3c,t?, and it follows that y(t) = coe~ — cy. Since the solution [x(2), y(2)]"
is bounded over 0 < t < +o for any choice of the constants c, and cy, one may
be assured that all the solutions of the system (7.40) are also bounded over the same
interval. ||

EXERCISE 7.4

1. Verify that fo"? (—4 + 4 |cos 2¢|) dt = —@.

2. Let g be defined by
al t
SOft) = Pht
— cos —°

How do the solutions of x’ = (g’(1)/g(t))x behave as t—> +0?

3. How do solutions of the following equations behave as t > +0?
a) x” + (sin? Dx’ + x = 0. b) x” — ex’ +x = 0.
Oy gan + x = 0, ga 0.
d) x" +G()x'+x=0, dO <0, So lal dt < +.
e) x” +q()x' +x = 0, for” lad] — gO} dt < +.
 Qualitative behavior of solutions for linear equations 7.4
216

4. Discuss the behavior of solutions as t > + for the equation x’ = A(x in each of
the following cases.

1 ib
= Se:
by) See
a AG te
—t —l —t —1
=i) sin In ¢ —sin?t sin ‘e
= : A(t) = ;
740) BE Int =) | ore —sint —1

5. Prove Gronwall’s (generalized) inequality: Let u and v be non-negative continuous func-

tions on an interval a < t < wand let M denote a positive constant. Show that
t

ule) SoM i u(s)v(s) ds, b> 1;

to

implies that
t

u(t) < Mexp || u(s) as; ip = ie

6. Discuss the behavior as t > + of solutions for the following equations.

a) x (tv) Pia
nt Neal my
+ 3x" " + 2x0
!
+ ak
tsin~ a === Oy

Oe pales |b
7. Let p be a positive-valued, continuously differentiable function.
a) Show that the equation x’’ + p(x = 0, 0 < t < +, takes the form

yi tqsy+y=0, O<s <= fi? Vp)dr

if one defines s = {9 V/p(z) dr, x(t) = y(s), and q(s) = p'(1)/2p?/2(n).
b) If W’() + pV = 0 and r2() = ¥2() + Y’?2()/p(, show that
t

r°(0) exp ihm(r) dr < r'(t) < 7°) exp / M(r) dr

where M(t) = (p'(O| — p'()/2p() and m() = —(|p'(O| + p'(D)/2p00.

c) Show that every solution of x’’ + p(t)x = 0 is bounded if p is nondecreasing. Show
that every solution of the usual equivalent system is bounded if p is nondecreasing
and bounded above.
d) Discuss the behavior as t > + © of solutions for

t 1
x! Bx = 0, en ond en x” + tx = 0,

x
Pe d Ca OV ees
+(43)r=0,
Py
w+
Gig Sayeoes
(GEM*S) Wo
 Pd

The homogeneous equation with continuous coefficients Pd\G/

8. How do solutions of

' 1 1
x = = a 58
1+ £ Neen;

Gi 1 1
y y
5 Me a a
behave as t> +0?

9. a) Let pi = [1, 0]? and pe = [0,1]?. Show that the system

est oak (6
has two linearly independent solutions @; and @2 with the property that

lim (¢1()e—* — pi) = lim (¢2(t)e’ — po) = 0.

t— + +00 t+>-+00

b) Let L be a matrix with 7 linearly independent eigenvectors pi,..., pn and corre-

sponding eigenvalues \4,...,An. Let R be a continuous matrix function on
a<t< +o with
foo

/ |R(t)| dt < + for some 10. @.

i)

Show that there is a fundamental solution set @1,...,» for the system x’ =
(ZL + R(t))x such that (¢.(Ne* — p.) > 0 as t> +o.
10. a) Show that the equation x” + (w? + e~)x = 0, t > 0, w > 0 has a fundamental
solution set Yi, Y2 with the property that

lim |Wi(t) — sinwt| = lim |Y4(4) — cos at|

t—>+00 t— +00

lim |Wo2(t) — cos wt] = lim |Y4(t) + sin wt| = 0.

t—-b00 t>-+o

b) Conjecture, state, and prove a general theorem for the equation x’ + p(‘)x = 0
which yields the result of part (a) for p() = w? + e~.

11. Show that

t e
Sin” S sin f —cost
(= exo||eH :
7 & cos t sin ¢

is a fundamental matrix solution for

Relate this example to Theorem 7.9.

 Qualitative behavior of solutions for linear equations 7.5
218

12. Show that the system x’ = (L(t) + R(d))x, where

sin ¢
Ot t
I) = E i and R(t) =
sint |’
t

has all solutions bounded as t > +2. Note that all solutions of x’ = L{t)x are bounded
over —© <¢ < +o and that f°” |R(D| dt = +. Relate this example to Theorem 7.9.
13. How do solutions of the system

= tf}
behave as t > +0? Relate this example to Theorem 7.9.
14. a) Describe the behavior of solutions as t + for the system

ial: alls:
b) State and prove a theorem for x’ = A(‘)x,a < t < +, of which the result in (a)
is a special case.
c) How do solutions of the system in (a) behave as t > 0+?
15. Let o denote the supremum of a nonempty set S which is bounded above. Show that
there is a sequence of points ¢, in S such that tf, ~ o ask— +.

7.5 THE NONHOMOGENEOUS EQUATION

The questions asked in Section 7.1 can be answered for

x’ = A(t)x + b(t) (7.5a)
if the behavior of the solutions
x = Ah)x (7.5b)
is known. A basic tool in any such discussion is the variation-of-parameters formula

o(t) = &(Ac + (7) ‘A&—1(s)b(s) ds (7.5c)

which relates solutions ¢ of Eq. (7.5a) to a fundamental matrix solution for Eq. (7.5b).
Recall, in this connection, that if @, and @» are solutions of the nonhomogeneous
equation, then y = $; — 2 is a solution of the homogeneous equation.
Let us assume in what follows that A and b are continuous on an interval of the
forma < t < +o. Here, & will denote a fundamental matrix solution for Eq. (7.5b)
and ¢ will denote an arbitrary solution of Eq. (7.5a).
Theorem 7.11. Suppose every solution of the equation x' = A(t)x approaches zero
as t—> +o. If x' = A(t)x + b(t) has one solution which is bounded (unbounded,
 The nonhomogeneous equation 219

or approaches zero) as t—> +, then all its solutions are bounded (unbounded, or
approach zero) as t > +o.

Proof. Let ; and Wz be solutions of Eq. (7.5a). Then there is a solution ¢ of Eq.
(7.5b) such that ¥,(t) = Wo(t) + (2). If Wo is bounded as t—> +, the same is
true of ¥; since ¢(t) > 0 as t— +o. The parenthetical assertions are proved
analogously. ||
Corollary 7.11. Suppose every solution of x' = A(t)x approaches zeroast > +o.
If x’ = A(t)x + b(t) has one solution Y, which neither approaches zero nor is un-
bounded as t—> +, then every other solution 2 of x' = A(t)x + b(t) behaves in
the same way. Moreover, limt.40 (Wilt) — Y2(t)) = 0.
Theorem 7.12. Suppose every solution of x' = A(t)x either approaches zero or
becomes unbounded as t—> +a. If x = A(t)x + b(t) has a periodic solution 4,
then @ is its only periodic solution.
Proof. Suppose the equation had a second periodic solution y. Then ¢@ — y isa
solution of its complementary equation. The difference of two periodic solutions
is bounded. Thus it follows from the hypotheses that lim; ,,, [@() — Y()] = 0.
But this can happen only for w nonperiodic or Y(t) = ¢(t). ||
Example 1. Consider the scalar equation

x= = ax + b(t) (7.5d)

on the interval 0 < t< +o. We shall illustrate Theorems 7.11 and 7.12 by
choosing b(f) in different ways. The variation-of-parameters formula assumes the
form..(79 = 1)
t t t s

o(t) T exp|- | 1= ]-e+ exp|-f Lay). [exo [ 1ay].(6) ds

t

Leen
t
Lif
i da
ialaeibe
Here #(ft) == 1/f is a fundamental solution of the homogeneous equation

and

op(t) = oi sb(s) ds

is a particular solution of Eq. (7.5d). Since &(t) + 0 as t—> +a, every other solu-
tion of the equation is unbounded, bounded, or approaches zero with ¢,(f) as
t— +o. If, for example, b(s) = 1, then all solutions of Eq. (7.5d) are unbounded
as t—> +00 since ¢,(t) = t/2 — 1/2t. If b(s) = 1/s, then all solutions are bounded
220 Qualitative behavior of solutions for linear equations Ths)

as t—> + since ¢,(7) = L— 1/t. If bG)= 1/s*, then all solutions approach
zero as t—> +o since ¢,(t) = Int/t. ||
Problem. Exhibit a function 6 for which Eq. (7.5d) in Example 1 has a periodic
solution @. Find an explicit representation for any other solution y. Observe directly
that y could not be periodic.
The theorems and corollary above say nothing about the existence of bounded
or periodic solutions for x’ = A(t)x + b(t). Let us first examine it for bounded
solutions.
Theorem 7.13. Suppose the following hypotheses are satisfied:
i) A and b are continuous ona <t< +0;
ii) A is either constant or periodic with minimal period T > 0;
iii) b is bounded over some interval ty < t < +a;
iv) every solution of x’ = A(t)x approaches zero as t—> +o.
Then every solution of x’ = A(t)x + b(t) is bounded over the intervalty <<t< +o.

Proof. \f & is a fundamental matrix solution for the nonhomogeneous equation, then
the estimates (7.4d) and (7.4e) hold for &. Since any solution ¢ of x’ = A(t)x + b(d)
satisfies

o(t) = B(tc + 22) / &—!(s)b(s) ds, (7.5e)

we have
> : ; C= ent)
|o(t)| < Glew” -}- Co ea =e) ds < Ge + OSS Es t > to;
to U7]

for suitable constants c; and co. ||

Example 2. The equation x’ = (—(x/t) + 1) is an easy illustration that some form
of hypothesis (ii) is crucial in Theorem 7.13. Each of its solutions is of the form
c t 1
g(t) = F —f oy ay
for some constant c, and so becomes unbounded as t > +o. All the hypotheses of
Theorem 7.13 except (ii) are satisfied. If one thinks of —1/tasa1 X 1 matrix A(0),
t

lim / Tr A(s)ds = lim (—Int) = —o.

to +0 1 t>+o

We shall see in the next theorem that a trace condition can be used to replace
hypothesis (ii). ||
Theorem 7.14. Let the following hypothesis be satisfied:
1) A and b are continuous ona < t < +o;
 The nonhomogeneous equation 221

il) if A(t) is neither constant nor periodic, then

t

lim int|Tr A(s) ds # —~, for some to > a;

to+oa to
-+co

iii) / |b(s)| ds << +m;

t0

iv) every solution of x’ = A(t)x is bounded over the interval tp < t < +m.
Then every solution of x’ = A(t)x + b(t) is bounded over ty < t < +m.
Proof. One takes norms in the equation (7.5e) to obtain

lo(t)| < cy + c. | |b(s)| ds,

for some constants c; and cg. Note that hypothesis (ii) insures the boundedness of
je *(s)|- —||
Theorem 7.15. Let b be continuous and periodic with (minimal) period T > 0
and suppose that the constant matrix A has no eigenvalue of the form \ = 2nki/T,
where k is an integer. Then x’ = Ax + b(t) has a solution @ of period T. If A has
no pure imaginary eigenvalue, then @ is its only periodic solution.
Proof. Theorem 7.12 guarantees uniqueness when A has no pure imaginary eigen-
value.
The integer one is not an eigenvalue of e~“” A since no eigenvalue of A is of the
forth 2rki/T. Thus the matrix (e~47 — J) is nonsingular, and the algebraic equation
t

(e~47 — Tc = / e“%b(s) ds (7.5f)

0

has a unique solution for c. Then

t

o(t) = ee + e*! Hie “*b(s) ds

0

is the desired periodic solution of x’ = A(t)x + b(¢), for

iT
i LDS eA Te a acon f e4*b(s) ie
0

T 47
eee ee Ce iheA F— hs) ds + i e4A6—D ps) ds
0
¢

Sie ee ee et i e4"b(u) du
0
t

l| ee + cal e4"b(u) du = 9(t). ||

0
222 Qualitative behavior of solutions for linear equations Ths)

Corollary 7.15. The equation x' = Ax + b(t) will have at least one solution of
period T > Oif
7
i e “*h(s) ds = 0. (7.5)
0
Proof. If Eq. (7.5g) is satisfied, Eq. (7.5f) will have a solution for c even though
(e—47 — J) might be singular. ||
Theorem 7.15 remains true if A(t) is periodic with period T provided the eigenvalue
conditions are imposed on the characteristic exponents of x’ = A(t)x. The proof
will be left as an exercise.

EXERCISE 7.5

1. Explicitly compute all solutions of the system

5 een eee et le i)
I - al 0 AGIA
and describe their behavior as t — + in the following cases:

DON DAO = el ; eae eee

d)b(t) = sin, e)bw)= 147) ) om =144;
g) bh) = 1, h) ) = 6”, i)6) = St.
2. Consider the system in Problem 1 which has b(t) = sin t. Is there a periodic solution?
3. Describe the behavior of solutions as t > +. for the equation

x aL xl! ae 3x// JE 2x! a x= 1 ala at.

1++t
4. Describe the behavior of solutions as t+ + for the system

x |’ 4 —1 inelle|
;
sint |S ¥ 00s(¢)
= 1
=
t
a

y cost —IlJlLy 0
5. Either prove that the following proposition is true or give a counterexample to show
that it is not true: Let A and b be continuous fora < t < +0 and suppose A is constant
or has minimal period T > 0. Assume that every solution of x’ = A(x approaches zero
ast— +. If b(t) > c as t + +>, then every solution of x’ = A()x + b(‘) approaches
a limit as t— +0,
6. Relate the example

to Theorem 7.14.
 The nonhomogeneous equation 223

7. Either prove that the following proposition is true or give a counterexample to show
that it is not true: Let A and b be continuous for a < t < + and suppose that A is
constant or has minimal period T > 0. Assume that all solutions of x’ = A(x are bounded
for —%» <?t< +o. If there exists a constant ¢ such that for (b(s) — c) ds converges,
then every solution of x’ = A(Ax + b(t) is bounded as t > +”.
8. How are periods of periodic solutions for x’’ + w?x = sint, w > 0, related to the
common period of the coefficients?
9. Does the equation
x’ + (sinax’ + (1 + cos Ax = sin (20)

have a periodic solution? If so, what is its period?

10. Does the system

have a periodic solution?

PT -[- lb]+[% 4
Is there more than one periodic solution?
11. Under what conditions on a and w, if any, will the system

{= (a2, ai Z]+[3"
have a periodic solution? Is there more than one?
CHAPTER 8

THE EXISTENCE OF SOLUTIONS

8.1 PRELIMINARIES

The equations (1.4c)

eS,
Gm.

and (1.10c)
x? + e(x?— 1)xX’ +x =0
from Chapter | are special cases of the differential equation
KU nf (ae ee ee 1). (8.1a)
This equation is said to be of the nth order since the highest derivative involved is an
nth derivative, and it is called normal since x is given explicitly in terms of the
other symbols in the equation. Similar terminology is applied to systems of equations.
For example, the system (1.5c, d)

Q”’ — 2r’6!’ rl M72eG 2

=e Z ’ = — ré’

is a special case of the general system

Cay , (n—1) , (m— 1)
og sf (Xa Nee Mel Bi are toy © elke PD);

ye es a(x, x ny i a y, y’, ae poo t).

This is called a normal system of order n + m, and the individual equations are said
to be coupled.
Evidently, any finite number of normal ordinary differential equations of arbitrary
order can occur in a coupled system. In order to discuss the existence of solutions
efficiently, it is customary to observe that a judicious change of variables will reduce

224
 Open sets and closed sets 225

Eq. (8.1a) or an arbitrary normal system of (finitely many) coupled equations to the
standard form
x} = fi(1, sey Xn, t),

: (8.1b)
Remeeee Ln Xicue oy Xa b)-
Consider, for example, the equations (1.5c,d). If we define x; = 0, xo = 6’,
X3 = r, X4 = 1’, then they assume the form
,
Xi = Xe,
ae
X= —2X4X2/X3,
y
X3 fa X4,

Ke = MG) xs Axes:
The equation ([Link]) takes the form (8.1b) if one sets

Lie or See VOY aa ht: (8.1c)

In fact,
ie pee
x1 = Xa,
a
Xo = X3,

xe — ECCT, X25+++5Xn;s t).

The equations (8.1b) can be profitably studied from a geometrical point of view.
To implement the geometry, it is convenient to identify the space V,(@) of n-vectors
[x1,.-.,Xn]’ with the number space ©” consisting of n-tuples (x1,..., Xn) of
numbers. In a discussion involving only n-tuples of real numbers, the number space
corresponding to V,,(®) will be denoted by ®”. We use &” to stand for either ®”
or @”. Then the vector notation developed in Section 3.4 makes it possible to write
the system (8.1b) in the condensed form
x’ = i(x, 7) (S)

for theoretical discussion. We shall, under suitable hypotheses, study the existence
of solutions to (S) and handle in this way not only Eq. ([Link]), but a multitude of
normal systems.

8.2 OPEN SETS AND CLOSED SETS

The number space &? can be interpreted as the x; x2-plane in which analytic geometry
is traditionally done. Consider the sets of points
Be (fiyX2)s Xi ko) and =55(X, Xo)i. 1 > Xa)
in ®? (Figs. [Link] and 8.1b). The student has an intuitive notion of what one means
by the boundary of F and boundary of G. These consist of the points on the circle
x? + x2 = 1 and the points on the line x2 = x1, respectively. What is it, however,
 The existence of solutions 8.2
226

(a) Figure 8.1 (b)

that distinguishes boundary points from other points? It is intuitively reasonable to

call a point x = (x1, X2) a boundary point of F if there are points in F and points
not in F arbitrarily near x. The same remark holds for G.
To speak of a point y as being near a point x, one must have defined a notion of
distance. The identification of the vector spaces V,,(C) and V,,(R) with &” allows us
to use the rectangular norm to define a rectangular distance function on &”: we take
the distance from x to y to be the number |x — y| = |x; — yi] -°+ 4+ |[xXn—ynl-
The notion is illustrated in Fig. 8.2 for points x = (a,b) and y = (c,d) in @?.
There |x — y| = ja — c| + |b — dl.
Notice that |x — y| is the distance traveled in going from x to y by way of
z = (c, b) rather than in a straight line.
Given a point x in 8” and a number r > 0, one defines a rectangular neighborhood
of x of radius r as the set N(x;r) = {y: |x — y| < r}. In ®?, a rectangular neigh-
borhood is a square (Fig. 8.3). Even though it is not so easy to visualize a set S in
@” for n > 4, we simply define a point x to be a boundary point of S in &” if every
neighborhood of x contains a point in S and a point not in S. The boundary of S
is the set dS of boundary points of S.
A set in &” which contains none of its boundary points is called open, and one
which contains a// of its boundary points is called closed.
The whole space &” and the empty set @ are both open and closed. The set F,
in Fig. 8.1, is closed and G is open.

Figure 8.2
 Open sets and closed sets 294)

r2
a

2)

Figure 8.3

Suppose now that S # &” is a nonempty open set in 8”. If x is in S, then x is

not a boundary point of S since an open set contains none of its boundary points.
It must then be possible to find a radius r > 0 so small that N(x; r) contains only
points of S; otherwise x would be a boundary point of S. If, conversely, every point
of S can be surrounded by a neighborhood which lies entirely in S, no point of S
is a boundary point of S. Thus S is open. These remarks may be summarized by
saying that a nonempty set S is open if and only if each of its points can be surrounded
by a neighborhood lying entirely in the set S.
If S is any set of points in &”, then the set &” — S of points which are not in S
is called the complement of S. Since dS = 0(&" — S), one may conclude that the
complement of an open set is closed and the complement of a closed set is open.
A special property that makes an open set usefui in analysis is the property that
we established above: Every point of an open set can be surrounded by a neighbor-
hood lying entirely in the set. A special property of closed sets which makes them
useful has to do with sequences of points.
It is natural to say that a sequence {x;} of n-vectors x, = (X1n,..-5 Xnk) Con-
verses lord vector X = (X41, .--, xn) it limp7. X= x7 for j= 1,...,n. Since

[outer | ele wets one — en = [XE |

for each j = 1,...,n, an equivalent way of defining convergence is to say that
lim,_,4. Xz = x if and only if limz4. |x; — x| = 0. The latter definition em-
phasizes the geometrical aspects of convergence since one thinks of the distance
from the point x; to the point x as growing smaller with increasing k (Fig. 8.4).
A closed set may be characterized in terms of sequences by the following proposi-
tion: A nonempty set S in &” is closed if and only if every convergent sequence of points
x; of S has its limit x in S.
To prove the sufficiency assertion of the proposition, let x be a boundary point
of S. Then each neighborhood M(x; 1/k), k = 1,2,..., contains a point x; of S
and lim,_,4« |x, — x| = 0 by construction. If the convergence condition of the
proposition holds, then x is in S. Since x was an arbitrary boundary point, S contains
its boundary and is therefore closed.
228 The existence of solutions 8.2

Figure 8.4

Suppose, conversely, that S is closed and consider an arbitrary sequence of

points x; from S which converge to a limit x as k—» +a. We shall show that x
is not in &” — S; hence x isin S. Since Sis closed, &” — Sis open. Let y denote an
arbitrary point of &” — Sand surround y by a neighborhood N(y; r) which contains
only points of &”" — S. Then |y — x;| > r since each x; is S, hence not in 6” — S$
(see Fig. 8.5).

Figure 8.5

Letting k — +. in the inequality

TES NY xy awk | ees

we find that |y — x| > r > 0. This inequality implies that no point y in 6" — S
is the same as x. Hence x isin S.

EXERCISE 8.2
In Problems 1 through 10, there is given a set. Identify its boundary and tell whether it is
open, closed, neither open nor closed, or both open and closed.
1. {G15 %2): [x1 2x
Pie, (GRE
 Continuity of vector functions 229

33 Coie 2)? [xale<) l-and: |xo|_< 1).

4. {(x1, x2): x2e71 < 1}.
De Ai, x2)e a1 = 1/m, xo = 1/asm, n= 1,2, 02...
Gz i(ci.2; x8): xs < x + ol
DE SIDS =, Osh 91 2a).
82 {(Gc1, x2, X39, D2t + x3 4 xe < 77),
9. The rational real numbers.
10. The complement of the rational real numbers.
11. Let S be a nonempty set in ®”. A point x in ®” (not necessarily in S) is called a cluster
point for S if every neighborhood of x contains at least two points of S.
a) If S contains only finitely many points, could it have a cluster point?
b) If S contains infinitely many points, could it fail to have a cluster point?
c) Show that S is closed if and only if it contains all its cluster points.
d) Give an example of a set S all but one point of which is a cluster point for S.
12. Define the euclidean length of ann-vectorx = (x1,..., Xn) by ||x||. = a/x? eet xe
a) Show that n—1[x| < ||x||. < |x|.
b) Show that lim;_,4. |x — xx| = 0 if and only if lim,4. ||x — x:||e = 0.

8.3 CONTINUITY OF VECTOR FUNCTIONS

Up until now we have referred to a vector function as having a property (such as
continuity, differentiability, boundedness) if each of its components has that property.
One of the advantages of the vector concept, however, is that such properties may
be defined and used without reference to components. At the heart of the matter is
the fact that
|x; — vel S [x — yl = [xr — yal b> + [xn — yal (8.3)

fore" lean, Hawhenevel XK = (Ny... 0,.%,) and Y = (X4,.. 5 Yn) are m-vectors.
Let g denote a vector function defined on a set S in &”.
We shall now say that g is continuous at Xo in S if |g(x) — g(xo)| 0 as
[x — Xo|— 0. In view of the inequality (8.3), this definition is equivalent to our
previous one. As is usual, we say that g is continuous on S if it is continuous at each
point of S.
A set in &” is called unbounded if it contains points located at arbitrarily large
distances from the origin; otherwise it is called bounded. To show that a set T in
&” is bounded, one displays a constant M > 0 such that |x| < M for all points x
in T. One can show that a set T is unbounded by displaying, for each positive integer
k, a point x; in T such that |x;| > k.
We say that a function g is bounded if its range is bounded, that is, if there exists
a constant M > 0 such that |g(x)| < M for all x in the domain of g. Similarly, a
sequence {x;} is called bounded if the range of its values x; is bounded. Again we
comment that this is equivalent to the definition in terms of components.
 The existence of solutions 8.3
230

If g is continuous at a point xo in S and if |x, — Xo| 7 0 ask > +o, then

certainly |g(xz) — g(xo)| 2 0 ask > +. This phenomenon is called the preserva-
tion of convergence by continuous functions. It is very useful in proving propositions
involving continuous functions which are defined on closed sets. In order to apply
the notion of preservation of convergence in such a context, however, one must know
something about the existence of convergent sequences.
Bolzano-Weierstrass Theorem (Sequential Form). If {x;} is a bounded sequence
with values in 8", then there is a subsequence {xz,} of {Xx} which converges.
Example. The sequence of points x, = (k(—1)*/(k + 1),0), k > 0, in ®? does
not converge since x, = (k/(k + 1), 0) if k is even and x, = (—k/(k + 1), 0) if
k is odd. The sequence {x2;}, however, converges to the point (1,0) and {xox41}
converges to (—1,0). ||
As an application of the preceding remarks, we prove three propositions which
are needed later.
1. Assume that the vector function g is continuous and that its domain S in &” is both
closed and bounded. Then g must be a bounded function. Otherwise, there exists a
sequence of points x, in S such that |g(x;)| > kA for all integers kK> 1. By the
Bolzano-Weierstrass theorem, there exists a convergent subsequence {x;,} of {xz}
since S is bounded. If xo denotes lim;,4.. xx, then Xo is in S since S is closed.
Note now that

lg(xz,) — go)| = |g(xz,)| — lgo)| = ki — |g(ko)l.

Since k; > +00 as i> +o, |g(xx,) — 8(Xo0)| > +o also. This contradicts the
continuity of g at xo, hence on S.

2. Consider next a vector function defined and bounded over an interval a < t < b.
There is a sequence of points ty, in [a, b) such that t, > bask — +. and \imy_,40 (ti)
exists. To prove this, first select a sequence of points #/, of [a, b) such that 17, > b
asm-—> +o. There is, of course, no assurance that the sequence {#(Z/,)} converges.
Since @ is bounded, however, the sequence {@(7/,)' is bounded and therefore has a
convergent subsequence {9(f,,,)}. Now define 7, = ¢,, and the construction is
complete.
3. Suppose that @ and y are two vector functions which are continuous on the closed
and bounded interval [a,b] and assume that $(a) = ¥(a) but o(b) ¥ W(b). Then
there is a rightmost point t = c in [a, b] such that $(t) = Y(t). To prove this, we make
use of the notion of supremum defined in Section 7.4. Let c denote the supremum of
the set {f: a<¢ < band $(t) = y(t} and select a sequence of points ft, in [a, b]
such that 4, > cask — +o. Certainly a < c < b. Since @ and y are continuous
and (7) = (tz), one can let k + + and evaluate limits by substitution to obtain
o(c) = ¥(c). There is no point ¢ in (c, b) such that ¢(t) = (ft) since c was defined
as a Supremum, i.e., least upper bound.
 Connected sets 231

EXERCISE 8.3

1. Let g(x, y) = (x?y/(x? + y?), x? + y?) if (x, y) ¥ (0,0) and g(0,0) = (0,0). Is
g continuous at (0, 0)?
2. Define a sequence {x;} of points by the recursion scheme

> eee ee Xi cae LS, and X41 = Xo + xz/k, k > 1.

Is the sequence bounded or unbounded ?
3. Show that the members of the set in Problem 5, Exercise 8.2 may be taken as the values
of a sequence no term of which is repeated. Explain how to select convergent subsequences
of the sequence which you construct.
4. Define g(4) = Doe_1 k/(1 + k4(t — &)?) for 0 < t < +o. Isga bounded function?
5. Let o() = (sinz—!, cos¢—!) for —24 <t <0. Find a sequence {7} such that
t, > O— ask > +0 and limy_,4 (tx) exists.
6. Let g be a continuous, real-valued function defined on all of ®”. Show that the set of
points where g(x) > 0 is open. What can be said about the sets of points where g(x) > 0,
pixe= 10; 9(x) <0, g(x) < 02
7. Show that every nonempty closed set in®! which is bounded above contains a rightmost
point. Use this result and Problem 6 to reprove the third proposition above.

8.4 CONNECTED SETS

Frequently, it is necessary to speak of sets S in &" which, loosely speaking, are not
broken up into pieces. Continuous vector functions can be used to make the notion
precise. Let Cy denote a curve in &” parameterized by a continuous function @
defined on an interval [a, b]. If Cg does not intersect itself, that is, if @ is one-to-one,
then it is called a simple curve or arc. If o(a) = $(b) and $(t1) ¥ (2) for every
t, and fg in [a, 5), then Cg is called a simple closed curve or Jordan curve. These
notions are illustrated in Fig. 8.6. If @ is a differentiable function, then ¢’(f) is
called a tangent vector to Cg at the point (tf).

Are

Jordan curve

Figure 8.6
232 The existence of solutions 8.4

Figure 8.7

A set S in &” is called arcwise connected if any two points in S can be connected
by an arc which lies entirely in S. The set S = {(x1, xX): x} — x} > 1} in Fig. 8.7
is not arcwise connected. We shall call an open, arcwise connected set a domain.
The union of a domain and any subset of its boundary is called a region. Thus each
of the sets F and G in Fig. 8.1 is a region, but only G is a domain.
Arcwise connectedness is an adequate concept for describing the connectivity of
open sets. It is not a general enough notion for sets which are not open however. To
see this, consider the set C = A U B in ®?, where A is the closed segment |y| <1
of the, y-axis and. Biis'the sraph of y= sml/x%, 0.<ixm<1 (rigas.8)
This set is not arcwise connected, but intuitively one would certainly like to
consider it connected in some sense. In order to formulate an adequate definition of
connectedness, let us first say that a set S in &” is disconnected if it is possible to find
two closed sets F, and Fy, such that (i) S C F; U Fo, (ii) both K; = F; NS and
Ky = F2S are not empty, and (ili) Kj 9 Ko = ¢. One then calls S connected
if and only if it is not disconnected.

Figure 8.8
 Connected sets 233

According to this definition, the set C in Fig. 8.8 is connected, but we shall not
dwell upon the subject long enough to prove this. It is shown in analysis that every
arcwise connected set is connected. The set C above is an example which shows that
there are connected sets which are not arcwise connected. It can be shown, however,
that every open connected set is arcwise connected.
In Chapter 9, we shall encounter a closed and bounded set which is assumed to
be disconnected. We wish to establish that such a set is the union of two closed and
bounded subsets which are separated by a positive distance. If K; and K> are non-
empty sets in &”, we define the distance from K, to K as the greatest lower bound
of the set of distances |x — y| where x is in K, and y is in Ky (Fig. 8.9).

Figure 8.9

Suppose, then, that S is a closed and bounded set in &” which is disconnected
and let F, and F 2 be closed sets as in the definition of disconnectedness. The corre-
sponding sets K; and Ke are nonempty, closed, and nonintersecting by definition.
It is easy to see that each is bounded and that S = K; UK». Let d denote the
distance from K, to Ko.
Then, by Problem 15 in Exercise 7.4, one can find points x; in K,; and y; in Ky
such that lim; 4. |xz — yx| = d. Since each x; is in the bounded set Ky, there is
a subsequence {x;z,} of {x;} which converges, to X9 say. The point xo must be in
K, since K, is closed. If we consider the corresponding subsequence {yx,} of the
sequence {y;}, there is no assurance that it converges, but a subsequence {y;,.} of
{yx,} will converge to a point yo in Ky. Discarding some of the terms in {x;,,}, if
necessary, we form the corresponding sequence {xj,} which, of course, converges
to Xo also. Then [xo — yo| = lim;4o |x%, — yh,| = d. Since Xo is in Ky, Yo is
in Ky and since K; and Kg do not intersect, d > 0. This result is summarized as
follows: A closed and bounded set which is disconnected is the union of two nonempty,
closed and bounded sets which are separated by a positive distance.

EXERCISE 8.4
1. Explain why the set C in Fig. 8.8 could not be arcwise connected.
2. Compute the distance between the following sets in ®*:

Ky = {x: xi + xo +23 + x4/16 = 1},

Ko = er x1 = ae
234 The existence of solutions 8.5

3. Let f be a continuous real-valued function defined on the interval [a, 5].

a) Show that [a, b] is connected.
b) Show that the range of f is connected.

8.55 A GEOMETRICAL INTERPRETATION

The independent variable for the function f of the differential equation x’ = f(x, f)
is an ordered pair (x, f) where x is in &” and ¢ isin @'. The set of all such ordered
pairs is called 8” X @'. In order to define open set and convergence, it is necessary
to specify neighborhoods for &” X ®'. A neighborhood of the point (xo, fo) will
be a product rectangle, that is, a set of the form
R= {(x,
1): |x — xo| < a, |t — tol < 5},
where a > 0 and b > O (Fig. 8.10). Such a product rectangle is not one of the
rectangular neighborhoods defined for 8” in Section 8.2: it is, instead, a cylinder of
“altitude” 2b with a “base” consisting of an ordinary rectangular neighborhood in
&” and its boundary.
With this notion of neighborhood, one defines boundary for sets in &” X @!
and thereby gives meaning to the other concepts discussed in Sections 8.2, 8.3, and
8.4 for &”. In the following material the word neighborhood means a product rectangle
in &” X @' and a rectangular neighborhood in 8”.
Throughout this chapter, we consider as given the equation
Kee CXe (S)

where f is continuous on a domain D in &” X @'. By a solution of (S), we shall

U1

Figure 8.10
 A geometrical interpretation 235

mean a vector function ¢, defined on an interval Jy, such that

i) (¢(0), 2) is in © for all ¢ in Jy, and
ii) ¢(t) = f(¢(0, ¢) for all ¢ in Ig.
To study the equation (S) from a geometrical viewpoint, let us observe that the
function f defines a vector field on 9, the vector (f(x, 1), 1) being associated with
the point (x, #) in . Thus 9 is to be regarded as filled with vectors, each one of
which has the number one for its last component. Figure 8.11 illustrates the situation
for &” = ®”, where the vector (f(x, t), 1) is interpreted as a directed line segment
initiating from the point (x, £).

A Le \
AL |e
G4 tao f 7

7
Figure 8.11

We give the set D on which (S) is to be studied the name phase-time space, and
the projection © of D onto the x hyperplane will be called phase space. This is an
extended use of the term. It is traditionally used in connection with autonomous
systems (Chapter 9.) Note that @ will be domain since 9 is one.
Example 1. Consider the equation
<=
1—x
x
eee)
Phase-time space © can be chosen in several ways for this equation. It can be taken,
for example, as a subset of the xf-plane which lies above the line x = ¢/2 or asa
subset which lies below the same line. In either case, @ is in the x-axis. In Fig. 8.12,
© has been chosen to be the region beneath the line t = 2x, and the graph of the
unique solution @ which satisfies @(0) = 1 is shown. Notice that the graph of ¢
terminates at the boundary of D. ||
Example 2. If x’ = (tan #)/(1 — x”), D may be taken as any one of the rectangular
regions defined by x? # 1, t X kr/2, k an integer. ||
236 The existence of solutions 8.5

Figure 8.12

Let @ be a continuous vector function defined on an interval Jy. The set of all
points (@(f), t) is called the graph of @. Notice that the graph of @ can be regarded
as a curve. In this context, a solution @ of (S) traces a differentiable curve which
lies in @, and the graph of @ corresponds to a differentiable curve which lies in 9.
Example 3. The function @ defined by ¢(t) = (Asin (t — fto)), A cos (t — fo),
which is a general solution of

Pl = ler oll
traces a curve Cg (namely a circle of radius A) in the xy-plane. The graph of ¢ is
a helix about the f-axis. The tangent vector to Cy at (2) is ¢’(t) = (A cos (t — fo),
—A sin (t — fo)). The tangent vector to the graph of ¢ at (¢(2), t) is (A cos (t — fo),
—A sin (t — to), 1). See Fig. 8.13. ||

(¢’(,; 1)

$'(t)

71

Figure 8.13
 A geometrical interpretation 237

The problem of solving the differential equation (S) x’ = f(x, t) on a domain

D can be regarded as the problem of finding a differentiable function @ such that
the tangent vectors (¢’(1), 1) coincide with the vector field (f(x, ),1) on D. The
corresponding curve Cg is called an integral curve for the vector field.
When one asks for a solution @ of (S) which satisfies an initial condition of the
form $(to) = Xo, with xo and fo given, he is said to have posed an initial value
problem for (S). Geometrically, then, the initial value problem

x’ = f(x, 2), <e— XG when b= tf (IVP)

is the problem of finding an integral curve, which passes through the point (xo, fo)
in , for the vector field (f(x, 1), 1). Figure 8.14 is an illustration for 8” = 2.

7)

Figure 8.14

EXERCISE 8.5
1. Consider the differential equation x’ = x. Sketch the associated vector field in the
xt-plane and show several integral curves.
2. Repeat Problem 1 for the equation x’ = x”.
If x’ = f(x), where f is continuous on a domain @ in &", the assignment x — f(x) is
said to establish a direction field on @. The solution paths of x’ = f(x) are integral curves
of this direction field.
3. Sketch the direction field for each of the following systems and depict several typical
solution paths. Then solve the equations and compare your geometrical and analytical
results.
aN) ae! SS Be ah DG a= SX.
Cave —i ny. i—) iy, G) ieee ex
Care =e ey, = ey.
238 The existence of solutions 8.6

4. What regions might one plausibly designate as the phase-time space for the system
x/ def (x; 7) if

f(x, fh = (GPs 42? Se Gre ea) ee

where (x, ) is in R? X R!?
5. Consider the systems

a) x Se DS eae and
J x2 + y2 V/Vx2 + y2
b) , Ys / x
x = ——

V x2 + y2 J x2 + y2
Describe their direction fields near the point (0, 0) in R?.
6. Suppose the equation x’ = f(x, #) has a solution y such that Y(t) = constant. Describe
the graph of wy.
7. Suppose the equation x’ = f(x, 7) has a solution with minimal period T > 0. Describe
the graph of w.
8. Describe the graphs of solutions for x’ = y + t, y’ = —x. Depict a typical trajectory
in the xy-plane.

8.6 THE LIPSCHITZ CONDITION

Now let an initial value problem

x v= fine): X= Xo when t = (f5 (IVP)

be posed at a point (Xo, fo) in a domain ©. The student has seen, in Sections 4.2 and
4.3, how the method of successive approximations may be used to show that (IVP)
has a unique solution on (a, w) if f(x, t) = A(t)x + b(t) and D is the set &” X (a, w).
An equivalent integral equation for (IVP) is, of course,

X= Xo+ / f(x35) ds. (1)

In the linear case, f(x, t) — f(y, t) = A(t)(x — y). Thus

[fxP2) = 4G, Ol Ay), (8.6a)

where a <a <t<c <w and ||Al| is the maximum value of |A(f)| on [a, c].
If f(x, 4) is not necessarily of the form A(f)x + b(t), but nevertheless satisfies
an inequality such as

fx). —= typ b= Kix — yI, (8.6b)

where K is constant, then the proofs of Theorems 4.1 and 4.2 can be modified accord-
ingly so as to give an existence and uniqueness proof for solutions to (IVP). A
function f which satisfies an inequality of the form (8.6b) is said to satisfy a Lipschitz
condition. A function f which satisfies a Lipschitz condition for all (x, t) and (y, t) is
called globally lipschitzian. It is very restrictive to require of a function that it be
 The Lipschitz condition 239

globally lipschitzian. Consider, DOMED) a scalar function f such that f and df/ax
are continuous on a domain 9 in ®' X R!. Let R= {(x, 1): |x — xd <a,
|t — to| < b} bea product rectangle in D. Then by the mean value theorem,

G1) = f0,1) + 229 | = yy,

where @ is between x and y. If K denotes a bound for df/dx on R, it follows that

If, 0) —fO,0| < K- |x — »|

for (x, #) and (y, #) in R, but not necessarily elsewhere. Thus, a rather wide class of
scalar functions satisfy Lipschitz conditions if we impose the condition Jocally (on
neighborhoods) rather than globally. With this motivation, we make the following
definition.
Definition. If, for each rectangle R in a domain D of &" X @&', there exists a
constant K such that
(ft) SIG.) E= K|x — y| (8.6c)
for all (x, t) and (y, t) in R, then f is said to satisfy a (local) Lipschitz condition on D
or to be (locally) lipschitzian on 9.
When we subsequently refer to the general equation (S) x’ = f(x, 1), it is always
to be assumed that f is continuous and lipschitzian and that D is a domain.
As might be conjectured from the scalar case discussed above, f will always
satisfy a local Lipschitz condition on a domain D in ®” X @' if f and the partial
derivatives 0f/0x,,..., 0f/dx, are continuous on ©. To prove the conjecture, let
R be a given rectangle in © and let (x, f) and (y, £) be arbitrary points of R. For
each component f; of f, define

gi(s, t) = f,((1 — s)x + sy, t).

With ¢ fixed, apply the mean value theorem to g; on the interval 0 < s < 1 to
establish the existence of a number s; in (0, 1) such that

gl, = 20,1) + 8D
Og (Si, t

In terms of f;, we have

fil, ) = fx.) + Le
ZS
) vu 6i,
oy, 1) eee 4 2FMED
Ox,
Gy x), (8.64)
where 6; = (1 — s,)x + s;y. Now let M be a bound for the partial derivatives of f
on R. Then, by taking norms in Eq. (8.6d), we have

LA, ) — fis Ors Mix yl

and
f(y, t) — f&,)| <K-
where K = nM.
 8.6
240 The existence of solutions

One should not suppose from the preceeding discussion that it is necessary fora
function to be differentiable in order that it satisfy a Lipschitz condition. For example,
the function f(x, 1) = (sin 1)|x| satisfies

lf~o— fly, d| < 1-|x — y|

for all (x, ft) and (jy, ¢).

The function f(x, t) = /|x — Xo, on the other hand, cannot satisfy a Lipschitz
condition on any rectangle in R' X &' containing a segment of the line x = Xo.
Otherwise there would be a constant K such that

a Cre) a CN SW ale ar
for all x sufficiently near xo; then we would have

L>XO Vv |x = Xo|

which is not possible.

EXERCISE 8.6
he ete — COStA SINE?
iefind the maximum of |A(A| for 0 < t < 27.
; —sin ¢t cost
Show that the function f(x, )) = x/(1 + 1t?x?) satisfies a global Lipschitz condition.
Compute a Lipschitz constant for each of the following functions on the indicated region.
asf (xy) = sin xt ela, like:
b) f(x, 2) = Gixe, t+ xs, x2), |x| <a, |z < 6.
GC) fst) = te, x S100 =< Ha
d)- f(t) =1e-"x? sin Ax), x = 0. 1 x 1;
4. For what values of a > 0, does the function

0, Xen)

satisfy a Lipschitz condition on the interval [—1, 1]?

5. Let f be a real-valued function defined on an interval [a, b] and let xo in (a, b) be fixed.
One sometimes says that f satisfies a Lipschitz condition at the point xo if there is a constant
K such that |f(x) — f(xo)| < K |x — xo| for all x sufficiently near xo. If f(x) =
x cos 1/x, x ~ 0, and f(x) = 0, show that f satisfies a Lipschitz condition at x = 0 and
that f is not locally lipschitzian on the interval —1 < x < 1.
6. Let f be a continuous real-valued function defined on an interval [a,b]. Let to be in
(a, b) and suppose that f’(x) exists and is bounded for a < x < fo, to < x < b. Does ff
satisfy a local Lipschitz condition on [a, b]?
 The method of successive approximations 241

8.7 THE METHOD OF SUCCESSIVE APPROXIMATIONS

Consider now
x’ = f(x, 2), xX = Xp when t= 6 (IVP)

under the assumption that f is continuous and locally lipschitzian on a domain 9

in &" X ®’. Since f is not necessarily assumed to be of the form A(t)x + b(t), we
can only be assured that f satisfies a Lipschitz condition on each product rectangle
in D. The Lipschitz constant K will, in general, be different for different rectangles R.
Thus the existence proof to follow will be of a /ocal nature in that the solution ¢
of (IVP) which results will be guaranteed to exist on only a “small” interval
[to — €, to + €], € > O, around fo.
To begin the existence discussion, let us observe again that an equivalent integral
equation for (IVP) is
t

X = Xo + iff(xsis yids: (1)

to
To establish notation, let

NM = {(x, t): [x == Xo << a, \t = tol < b}

be contained in 9, let K be a Lipschitz constant for f on N, and let M > 0 be such

that |f(x, | < M for all (x,t) in N. We define « = min {b, a/M} and “shorten”
the neighborhood WN to

Re (NT): S| Ke Kl Ge (F 1G] Gey (8.7a)

If g is a continuous function such that g(t9)= xo and such that (g(f),t) is in
R for |t — to| < €, then the function A defined by

h(t) = Xo + iif(g(s),5)ds, |t— tol < (8.7b)

has the same properties (Fig. 8.15).
To verify the statement, note that
t

[h(t) — xo] < | |f(g(s), s)| ds < M-|t — to] < Me <a
0

for |t — to| < ¢. As a result, the successive approximations {@;} ;29 given by

$o(t) = Xo; 4102) = Xo + | f(¢,(s), s) ds (8.7c)

are well defined for |t — fo| < €. .

Notice also that if @ is a solution of (IVP), then by taking g=h = 4, one
sees that the graph of ¢ for |f — fo| < € must lie in R. If there were another solution
242s The existence
of solutions ~ ST

wh

Figure S18

of (IVP) on |¢ — fol < & say ¥ then

at

¥@) — o()| = | | WO. S) = {e@, Sy ds

Ive
at : a?

< || YOK 8) = He@, 9) es < K I vis) = @()/es)

. ® < a)

By Gronwall’s inequality,

) — O(2| < lolead — vege"! = 0

since (fo) = Xo = W(t Thus o(!) = ¥(¢) for jt — tol < & This result is stated
as the following theorem,
Theorem 8.1. Let f be continuous and locally lipschitzian on @ domain © ht
SX ®. [fe = min {b,@/M}, where @, db, M are as defined above, then (VP)
has no more than one solution on |t — tol < €
Next we show that the successive approximations (8.7e) converge on je-td) <6
Theorem 8.2. Under the hypotheses of Theorem 8.1, (IVP) has at least one solution
on |t — tol < €.
Proof. Let |t — to| < e and define vo) = @44() — o() for all > Then,
since @v41 = Xo + De v2 limyre @vai(2) = o(2) exists if and only
if the
Series }) a ¥)(¢) converges, But
:
42) = &@) — x = ik£(Xq, 8) as
 The method of successive approximations 243

and t

i419 = [ {f($; 1s). s)— f(¢As). s)} ds.

Thus

lt(O| < M-\t — to (8.7d)

arid
| t £ i

Wiss < K [.Iixss) — 6.43) a <4 [ YAS) ds] (8.7e)

for j> 1. Note that inequalities (8.74) and (8.7e) together imply that
lf 2
vA < x fm. Is — tol ds= oe ste
In general,

lt —to/?** M (Key* ; ;
WreO| < MK
G+ =KG+m J2° (8.78)
If M; denotes the expression on the right of inequality (8.7f), then

YM; = 2
jo K
— 0)
By Lemma 4.2, the series 5% 5¥,{t) converges to a limit $(1) + x, for 1 satisfying
lt — fol < €, and $(f) = lim; ..,.. Ad).
Since f is continuous, it is immediate that lim;_. f(¢A1). 1) = £(¢(1), 1) also.
Io show that @ satisfies the integral equation (I), then, it is enough to show that
£ t

lim [ f(oAs).5) ds = [ lim f(¢As). 5) ds.

Je S tp to J+

Fo do this, define £1) = f(9;:(0). 1) — f(¢A0). 1). Then

£ ft

/ f(s), s) ds = [ im flex. i(s). s) ds

: hy €
igri ag (8.7g)
= id(fixe, s) + j=—0
> EAs); ds.
but
g+1
leis S Klis — O10 = Kits MEDS s20.
 8.7
244 The existence of solutions

Since 329 ((Ke)i*1/(j + 1!) converges, it follows from Lemma 4.2 that

to
Seoa=d | soa
j=0

= >) {6542(1) — $3410} = 0 — 11

j=0
t

= ¢(t) — Xo — i)f(Xo, 5) ds.

to
By Eq. (8.7),

$(1)= Xo + | £(#(s), s) ds. ||

The method of proof for Theorem 8.2 is called the method of successive approxi-
mations with good reason. The function @y4 1 is an approximation to @ for
|t — to| < € in the sense that

l(t) — bv(2)| (1) < m=N

DLE n()|

< MMy EN =Big

yy KKO
m—N
— (Ke)”

eS a e (8.7h)
eM Koy. Ke aM ey
TK ENG any K WN!
Since (Ke)"/N! — 0 as N— +0, oy can be made to approximate arbitrarily well
on the interval |f — to| < € by taking WN sufficiently large. Moreover, inequality
(8.7h) provides a useful error estimate.
With regard to the role played by the Lipschitz condition, there are several
comments to be made. The first is that (IVP) has at least one solution when f is
merely continuous: the Lipschitz condition is not necessary merely to establish
existence. It is also not necessary that f satisfy a Lipschitz condition for solutions of
(IVP) to be unique. It is necessary, however, that f satisfy some condition in addition
to continuity. The problem
= 2/ ily 0) =
for example, has two distinct solutions x = 0 and x = t|t|. Note that the successive
approximations for this problem converge to the trivial solution x = 0.
Finally, let us observe that a solution to (IVP) may exist on a larger interval
than the interval |t — to| < €. Consider, for example, the problem
=l1+x°, %»=0 When f=0 (8.71)
with solution x = tant, —7/2 < t < 1/2. Here f(x, t) = 1+ x? does not vary
with ¢ and both f and df/dx are continuous on D = ®! X ®!. We determine the
 The method of successive approximations 245

largest value of € for which the proof of Theorem 8.2 is valid. Let

R= {(x, 1): |x| < a, |t| < 5}

be a rectangle about the point (0,0) in the xt-plane, and at the same time recall
that € = min {b, a/M} where |f(x,1)| < M on R. Since |f(x, | = |1 + x?| <
1 + a’, the smallest admissible value for M is M = 1 + a2. There is no constraint
on b so we assume that 6 > a/(1 + a”). Then e = a/(1 + a?), and its maximum
on the interval 0 < a < +. is found by setting de/da = (1 — a”)/(1 + a”)? = 0.
The maximum value e€ = $ occurs when a = 1. Theorems 8.1 and 8.2 guarantee
only that the successive approximations converge on |r| < 4 to a function y which
is a solution for problem (8.7i) and that there is no other solution on the same interval.
Note that y(t) = tans, |t| < 3.

EXERCISE 8.7
1. Compute several successive approximations for the solution to the initial value problem
x’ = —2tx, x(0) = 1. Compare this with the Maclaurin series for the solution. Repeat
the procedure for the solution to the initial value problem x’ = x?, x(0) = 1.
2. What is the largest interval of convergence as guaranteed by Theorem 8.2 for successive
approximations to the solution of x’ = x?, x(0) = 1?
3. Find all solutions to the initial value problem x’ = 2\/|x|, x(0) = 0. To which solution
do the successive approximations converge?
4. Let d:() = k?t(1 — O* for 0 < t < 1. Compute
t t

i! lim ¢x(s) ds and lim [ ox(s) ds

ty) k>+0 k++20

and compare their values.

5. If the solution to the initial value problem x’ = 1 + x?, x(0) = 1 is approximated
on the interval |r| < 4 by the 100th successive approximation ¢100(¢), how large can the
error be?
6. Let f(@) = 1/@ — 1), g@® = f¢, and

h(t) = iLfl e(s)) ds.

For what intervals |t — 2| < ¢ is A defined?
7. Show that the recursion scheme

Yrii(t) = A+ Bt - ih(t — s)¥,(s) ds, Yo(t)= A

will converge to a solution of the initial value problem x’ + x = 0, x = A, x’ = B when

t = O for certain values of ¢. For what values of t is convergence assured?
 8.8
246 The existence of solutions

8. Let A bea2 X 2 matrix in Jordan form, let xo be a given 2-vector, and let \ be a param-
eter. Under what circumstances will the sequence {x;} defined by
Xx41 = Xo + AAX:, k= 03

converge to a limit x? Assuming convergence, what can you say about the limit vector x?
9. Modify the proofs of Theorems 8.1 and 8.2 so that they apply to the scalar problem
OE = CRM Dy B= By i = xo when ¢ = fo without its being converted into a system
(see Problem 7).
10. Is there, in any reasonable sense, a solution to the initial value problem
pe / Ea
eex ees pe Ee
uv PS rein yy= Oe
WV x2 + y2 Vx2
+ y2

8.8 MAXIMAL SOLUTIONS

Observe that the functions ¢ and wy defined by ¢(t) = tant, —17/2 < t < 1/2 and
y(t) = tant, —4 < t < 4 are, strictly speaking, different solutions to the initial
value problem (8.7i). This sort of situation is, in a sense, artificial and is rather
inconvenient. The difficulty is avoided by introducing the notion of a maximal
solution to an initial value problem. Roughly speaking, a maximal solution is a
solution with a most inclusive interval of definition. In this sense, ¢ is a maximal
solution to the initial value problem above.
In order to give a more precise definition, let us write (#, Jg) for a solution @
of the differential equation
Xoo-e (Xt) (S)
with interval of definition Jy. We assume that (x, 7) is in a domain D in &” X @!.
Now it is entirely possible that a function & with an interval of definition J; may
satisfy (S) for every ¢ in J; and that (£(2), t) is not in © for all ¢ in J. For example,
f might be continuous and lipschitzian on a domain which properly includes the
domain 9 which we fixed at the outset. Such functions, however, were excluded from
consideration in the definition of the solution given above. Thus, if (@, J) is called
a solution to (S), it is understood that (#(2), t) is to be in © for all ¢ in Ig, and we
call Ig the interval of existence of @ (relative to D).
Definition. Let f be defined on a domain D in &" X @&’ and let (Xo, to) be in ®.
Assume that (IVP) x’ = f(x, t), x = Xo when t = to has at least one solution. A
solution (@, Ig) of (IVP) will be called a maximal solution (relative to D) if and only ips

i) ($(0), t) is in D for all t in Ig;

ii) whenever (y, Ty) is a solution of (IVP) with Y(t) = $(t) for every t in Ign ly,
then Ty Cc te

Lemma 8.3. Let @ and y be solutions of x' = f(x, t), x = x9 when t = to

and let Ig and Iy, denote their respective intervals of existence. If the hypotheses of
 Maximal solutions 247

Theorem 8.2 are satisfied, then $(t) = V(t) for t in the intersection Ig Q Ty of the
intervals Ig and Ty.

Proof. The intersection Ig M Jy is not empty since it contains an open interval about
to. Let us consider only the portions Bs and ly of Ig and Iy to the right of to. The
interval on which Theorem 8.2 applies is to be denoted by [fo —€, fo te]. If
either Ts or i is contained in [fo, to + €], there is nothing to prove. Therefore
assume that both Te and ly properly contain the interval [¢o, fo + €] and suppose
for contradiction that there is a point tf; > fy) + € in igs aly at which ¢(t;) #
Y(t). (The situation is illustrated in Fig. 8.16 for &" = ®?.)

(u(t), 4)

((t2), f2) = (g(t2), t2)

Figure 8.16

Since @ and wy are continuous, there must then exist a Jast point fg > fp t+ €
such that #(t,) = W(te). If the common value is denoted by a, then (a, fg) is in
® and both ¢ and y are solutions to the initial value problem x’ = f(x, ft), x(t2) = a.
Thus (ft) and y(t) are equal for all t sufficiently near (and particularly to the right
of) tg. This contradicts the existence of fy since it was supposed to be the Jast point ¢
at which g(t) = y(t). Thus f,; cannot exist either. The situation is analogous for
ato |||
With this groundwork having been laid, we are in a position to state a theorem
which underlies all our subsequent discussion.
Theorem 8.3 (Fundamental Existence and Uniqueness Theorem). Let f be contin-
uous and locally lipschitzian on a domain D in &" X &. If (Xo, to) is in D, then the
initial value problem
Xe tx, -2): x= "Xo when = TG (IVP)

has a unique maximal solution @. The interval of existence of @ (relative to D) is an

open interval T (Xo, to) < t < TCX OeLG):
 The existence of solutions 8.8
248

Proof. Let & denote the set of all solution functions to the problem (IVP). If y is
one such solution, denote its interval of existence by Jy. We construct a set Ig accord-
ing to the definition .
Ig = {t: t isin Jy for some y in $}.

Notice that Jg is an interval since each Jy is an interval containing fo. Now define a
function @ by the formula $(¢) = Y(#) if ris in Jy. Then ¢ is well-defined (Lemma 8.3),
and Jg is its interval of existence. The function ¢ is a solution to (IVP) by construc-
tion, and Jy contains Ly for every other solution y. Thus ¢ is a maximal solution.
The interval Jy must be an open interval, for if it had a first or a last point, the
solution @ could be extended (by the technique of Lemma 8.3) and @ would not be
a maximal solution. ||
A scalar version of the fundamental existence and uniqueness theorem is stated
as the following corollary.

Corollary 8.3. Let f be continuous and locally lipschitzian on a domain D in

&" X G! and let (cy, ...,5 Cny to) be a point of . Then there exists a unique maximal
solution y to the initial value problem
BO (KON ek ee es
he ye ee oe Soren, when b=:
The interval of existence of y (relative to D) is an open interval
CF Cia ts Cus Vo) hie Ta Cee Gear

Example. Consider the initial value problem x’ = 2tx?, x = x9 > O when t = fo.
If xo = 0, then the unique maximal solution is x = 0. If xo ¥ 0, then the unique
maximal solution x = ¢(f) satisfies

oa@e'Oi=21
for |t — fo| sufficiently small. Hence

for

Ts(Xoto) = —V Xo +16 <= Vag = taney

Notice that f(x, t) = 21x? is continuous and lipschitzian for all (Xeteyeeatiic
solution ¢ exists only for a finite time interval. ||
To interpret the fundamental existence and uniqueness theorems in terms of the
geometry introduced in Section 8.5, note first that (xo, fo) is just an arbitrary point
in the phase-time space. It is guaranteed that there is a curve, passing through
(Xo, fo) and defined on some sufficiently small interval |t — to| < €, which is an
 The estimation of escape times 249

integral curve of (f(x, A), 1). By Theorem 8.3, the corresponding solution can be
extended to a maximal solution @ with interval of existence 7~(xo, to) <t<
T*(Xo, to). One might plausibly guess that the graph of this solution @ extends
from boundary to boundary in phase-time space ®. We shall prove that this is
indeed the case in Theorem 8.4 below, but let us first summarize the geometrical
picture constructed.
Phase-time space D is to be regarded as composed of curves extending from
boundary to boundary. Each such curve is the graph of a maximal solution of
x’ = f(x, f). No curve can intersect itself or any other curve. If © is unbounded,
we shall think of it as having a boundary, portions of which are removed infinitely
far from the origin.
We shall subsequently write (Xo, fo, t) for that solution of (S) x’ = f(x, 4)
which satisfies the initial condition (Xo, fo, fo) = Xo and refer to ¢ as the maximal
solution function for (S) (relative to D). For fixed (Xo, to), the curve traced by
$(X0, fo, t), T(Xo, to) < t < T*(Xo, fo), in @, the projection ofD onto the x hyper-
plane, will be called the trajectory of ¢. The endpoints 7*(Xo, to) and 7 (Xo, fo)
of the interval of existence for (Xo, fo, 1) are called escape times. The example
above illustrates that (Xo, fo, 1) may have finite escape times even though f(x, f)
is continuous and locally lipschitzian for all (x, 1). Several standard results on the
estimation of escape times are given in the next section.
Finally, let us adopt the convention that the unqualified word so/ution is to be
interpreted in all later material as meaning maximal solution.

8.9 THE ESTIMATION OF ESCAPE TIMES

The next theorem is the basic result in the estimation of 7*(xo, fo) and 7 (Xo, fo)
for (Xo, tos t).

Theorem 8.4. The point (d(x 0; £0, 1), t) on the graph of ¢ approaches the boundary
of D or becomes unbounded as t > T* (Xo, to) and as t > T (Xo, to).

Proof. Write $(t) for (Xo, to, t) and suppose, for contradiction, that the point
(o(2), t) on the graph of @ neither becomes unbounded nor approaches the boundary
of D as t>7*t = 77(Xo, to). Then 7* is finite and |@(1)| is bounded for to <
t < 7*. Choose a sequence of times {t;} such that limy_,4. th = zt and such that
limp 4% ¢(,) =c exists. Then (c, 7*) is in 9, and there exists a unique solution
y of x’ = f(x, 1) on a nondegenerate interval |f — 7*| < € such that ¥(7*) = ¢.
We shall show that (1) = ¢(t) for ¢ in the interval J = [r+ — e,7*). This will
contradict the definition of @ as a maximal solution. Since
t

$(1) = $(te) + ii
k
£(4(s), s) ds
for t in J, it follows that

o(t) = e+ ibf((s), 5) ds.

 8.9
250 The existence of solutions

In view of the initial value problem satisfied by y,

t

W(t) =e+ a f(y(s), s) ds

for ¢in J. Thus

ls) — ¥@| < =i |f(o(s), s) — £(Y(s), s)| ds. (8.9a)

Now assume (without loss of generality) that € is so small the rectangle

RES (Kn) (Cex |ata a eae

is in © and that eK < 1, where K denotes a Lipschitz constant for f on R. From
inequality (8.9a), it follows that

le) — WO) < rel |e(s) — Ys)| ds < Ke-||¢ — y], (8.9b)
where ||@ — ¥|| = supser |o(s) — ¥(s)|. If |o(t) — Y(D| 4 O for every ¢ in J,
then eK > 1, a contradiction. ||

It is convenient to say the equation (S) x’ = f(x, f) has a cylindrical phase-time

space D if D = & X (a, w), where @ is a domain in the x-hyperplane and (a, w) is
an open interval. The remainder of this section is devoted to the following question:
If (S) has a cylindrical phase-time space and if @(Xo, fo, 2) denotes a maximal solution
of (S), under what conditions is T*(Xo, fo) = w? The results obtained will apply,
of course, to the analogous question for rT (Xo, fo).
Theorem 8.4 asserts that (Xo, fo, t) ceases to exist at 7‘ (Xo, to) by virtue of
(¢(Xo, fo, 1), 1) having approached the boundary of ©. There is a useful corollary
to this result.
Corollary 8.4. Let C be a closed and bounded set in D with (Xo, to) in its interior.
The point ($(Xo, to, t), t) on the graph of @ must leave C both as t > T*(Xo, to) and
ast—>T (Xo, to). (See Fig. 8.17.)

This corollary is evidently true whether or not Dis cylindrical. If © is cylindrical,

however, it is the basis for several analytical tests which determine whether or not
+

T' (Xo, to) = w.

Example 1. Consider the system

, x
x= Oo = MY = (1 — 2x2 + y2yt/2 p)
(8.9c)

VA sd Vo
Ci eee
where x” + y? > 0. For this example, we take @ as the xy-plane with the origin
deleted, a = —%, andw = +1. Thus 9 is that half of ®* lying below the plane
jel:
 The estimation of escape times 251

Figure 8.17

We examine the solution (x, y) beginning at a point (Xo, yo, fo) with 0 <
x% + y§ < land fy < 1 and show that 7*(xo, yo, fo) = 1 by constructing a closed
and bounded set C which (x(t), y(t), 1) must leave as t > 7*(Xxo,Yo, fo). Fix an
integer m > 1 so that tp < 1 — 2~” and let C denote the cylinder defined by the
equations x? + y? < 1,t9 —-1<t< 1—27”. Thecurve traced by (x(2), y(2), t)
initiates in C (Fig. 8.18) and the point must leave C before it ceases to exist.

Figure 8.18
 8.9
252 The existence of solutions

Since the point rises with increasing ¢, it passes either through the side of C or
through the top of C. The first situation is not possible. To see this, let r(t) Zz
[x2(2) + y2(2)]!/? denote the distance from the origin to the point (x(t), y(2)) in
the xy-plane. Then

Hp = 2OEOF ONO = Hy — (8.94)

Since r(to) < 1, r’(t) < 1 — (1/ — 9) = —(t/ — 8) as long as (x(), yO, #)
is in C. This means that r(t) decreases, and the graph of the solution terminates at
the top of C. Thus (x(t), y(t)) must exist on the entire intervaltp < ¢ << 1—2°”.
Since m can be taken as large as desired (thus raising the top of C), Tr? (XG. Vokes
et)
The technique of the example above is readily abstracted so as to provide a
general theorem which one may apply to other equations. The motivation is that
the point (x(t), y(t)) above remains in a closed and bounded subset (the unit disc)
of @ as long as it exists.
Theorem 8.5. Suppose that x’ = f(x, t) has a cylindrical phase-time space
D = 0 X (a,w). If o(Xo, to, t) remains in a closed and bounded subset B of ® for
to Sih< T* (Xo, to), then T* (Xo, to) = @

Proof. Suppose T+ < w and let B denote the closed and bounded subset of ® which
contains $(t) for tp < t < 7*. For sufficiently small e€> 0, the closed and bounded
set C = B X [to — €, T* + €] is in D and has (Xo, fo) in its interior, but (¢(2), t)
does not leave C as t+ 7+. This contradicts Corollary 8.4. Hence T* = w. ||
Notice that Theorem 8.5 makes it possible to study the escape times of solutions
by studying their trajectories in phase space rather than studying their graphs in
phase-time space. When possible, this procedure is more efficient. As an illustration,
let us reconsider Example 1. Phase space @ is the xy-plane with the origin deleted.
The trajectory starting at (xo, yo) # (0,0) at time fp < 1 is a curve in @ which
initiates inside the unit disc. We showed above that the distance r(t) from the origin
to the point (x(2), y(¢)) decreases as t increases. Thus the solution curve is contained
in the disc, and 7* (xo, yo, to) = 1 by Theorem 8.5.
To apply Theorem 8.5 in a specific case, one must show that a solution curve of
interest is contained in some judiciously chosen, closed and bounded subset B of @.
The first problem encountered in applying the result is thus the choice of B. When
® = &", the set B is frequently taken in practice to be a closed ball or a closed cube
because the condition that (Xo, fo, f) remain in a set of either type is that
|6(Xo, to, £)| be bounded for tp < t < Tt (Xo, ft). In such cases, the estimation of
escape times is thus reduced from a geometrical problem to an analytical problem.
Example 2. Let us show that all solutions to initial value problems for the equation

Ke xk xe = sine
 The estimation of escape times 253

exist on intervals of the form fg < t < +o. Let ¢ and y = ¢’ denote the solution
which satisfies the conditions (to) = x9 and ¥(to) = yo. Consider the function
V defined by
_¥O,#@
V(t) = 5) zi 4 ale ! .

8
Differentiation yields

VD) = VOW + $°(O¢'D = HO sint — POV

<vsint< WO <VYOt)
Vor
2 g() , 1 = NG
Brae F 4

Thus V(t) < V(to)e?““—. Now suppose for contradiction that 7+ < +0. Then
V(t) < Vitoje?*— for to < t < 7, and it follows that |$(t)| + |y(1)| is bounded
over the same interval. By Theorem 8.5, 7 = w = +o. This is a contradiction.
Note that if x = ¢(¢) is the displacement of a particle moving with one degree of
freedom, then V(f) is its total energy at time f. ||
A slight variation of Theorem 8.5, which is useful in applications, can be given
when D = &” X (a, w).
Theorem 8.6. Assume that x' = f(x,t) has a cylindrical phase-time space
D = &" X (a, w) and that there is a continuous function m defined on to < t < w
such that |$(Xo, to, 2)| < m(t) for to < t < T*(Ko, to). Then T*(Xo, to) = o.
Proof. Suppose for contradiction that r+ < w. Then 77 is finite and m(t) is bounded
for to < t <7. It follows that (xo, to, f) is bounded over the same interval. By
Theorem 8.5, 77 = w. ||
Example 3. Consider the equation
6” + k(sin t)cos 6 = 0, (8.9e)

and let (0(t), 6’(t)) denote a solution initiating at time fo. Then |6’(‘)| < k for
ig SUK Tt (6(to), 6’(to), toe On the same interval,

—K(t — to) <0’ (@) — Oto) S KO — fo)

and
k 2 k 2
= 5 tt — to)” < A(t) — O(to) — O'(to)(t — to) S 5 = es

Thus

|a(t)| + [0’(2)| < |0(to)| + [6"(to)| + CK + [0’Ct0)|)I¢ — tol + TC — to)”.

By Theorem 8.6, 7*(6(to), 9’(to), to) = +o. ||
 8.9
254 The existence of solutions

It is occasionally helpful to know that Theorem 8.6 is immediately applicable if

f(x, t) is either uniformly bounded or globally lipschitzian.
In the first case, a solution @ which exists at fo satisfies an inequality |p(2)| <
constant (t — to) + constant, t = to.
In the second situation, there is a constant M > 0 such that |f(x, 1) — fQ, t)| <
M|x — y| for tp < t < +o and all x and y. Thus a solution @ with (to) = Xo
would satisfy

lo) — xol < i,{|f(xo, D| + Mle(t) — Xo} dt, (8.9f)

which implies (Exercise 8.9, Problem 4) that

t
lo()| < [xo] + FQ) +e” / MF(s)e~™° ds + ce™’, (8.9g)
0
where

EG) = i)|f(xo, s)| ds

0
and c is a constant.
Example 4. When converted to a system, Eq. (8.9e) becomes

y= ¢,
& —k(sin t) cos @.

Here f,(9, ¢, 1) = ¢ and fo(4, ¢, t) = —k(sin zt) cos 6. All partial derivatives of f;
andf» are uniformly bounded; consequently all solutions exist for —0 <t< +o. ||
The latter remarks do not, of course, apply to equations such as x’ = a+ x?
with polynomial nonlinearities. If an equation exhibits this type of nonlinearity,
then, in general, some solutions will exist for all time and some will not.

Example 5. In Section 2.1, the boundary value problem

LOPE = 05, FO) = 1FO) Oe AGE co) ld ae Sie)

of Blasius was cited as an example of the occurrence of nonlinear differential equa-
tions in physical problems. J. Serrin’s solution technique for this problem can be
used to illustrate the existence and uniqueness theory above.
First, we note that Corollary 8.3 guarantees that the initial value problem

XA Os x(0) = x O7= 0, x70)= 1, "="d/d (8.91)

has a unique solution x = y(t) which is defined on a forward interval 0 < t < rt.
Suppose it is known that 7* = +oo and that lim,,,.¥(‘) = \? exists, where
\ > 0. Then the boundary value problem (8.9) can be solved by defining f(n) =
¥(t)/X with t = 7/).
 The estimation of escape times 255

To show that 7* = +, one first proves that y(t) > 0 for 0 <t< rt.
Suppose, for contradiction, that there is a time fo in (0,7*) for which W/’"(to) = 0.
Then, if A = Y(to) and B = y’(fto), it is easy to verify that both ¢(f) = A +
B(t — to) and y(t) satisfy the initial value problem
Ke SS sec — 0, X(t) = A, x'(to) = B, x’’(to) = (0h

Consequently, ¢(t) = y(t) for 0 < t < r*. This is a contradiction since $’(0) = 1
and y’’(0) = 0.
Having established that y/’(t) > 0 for 0 < t < 7*, one concludes by integrating
that y’ and wy are positive and monotone increasing over the same interval. Using
the integrating factor exp ie ¥(s)ds on the equation y’’’(t) + y(Hy’’(t) = 0, one
can write
t
v(t) = exp Ei ¥(s) as| (8.9j)
Thus y” is decreasing and
ONTOS GNM ARae (8.9k)
When inequality (8.9k) is integrated twice, the inequalities
Ouea)eat OT et ee eee
result. By Theorem 8.6, 7 = +a. To show that lim;.,. ¥/(f) exists, it is only
necessary to show that y’ is bounded above since it is known to be monotone in-
creasing. Let c = y(1) and note that Y(t) > c for t > 1. Inequality (8.9j) implies
that
Ue) 66g eo
Integrating, we find that

vinsvay+ | Gade steal

1

Thus lim;_,. ¥/(t) exists and the proof is complete. ||

EXERCISE 8.9
1. Consider the system

a= Vee — yaa = ear xa):

a) Show that all solutions exist on intervals of the form to < t < +.
b) Show that all solutions exist on intervals of the form —2% <1 < fo.

2. Show that each solution x = u(f), y = v(t) of the system

p= yx, yo=xty?
which satisfies u(0) > 0, v(0) > 0 cannot exist on an interval of the form 0 < f < +.
 8.10
256 The existence of solutions

3. Use Theorem 8.6 and the Gronwall inequality (Problem 5, Exercise 7.4) to show that
every solution of the linear equation x’ = A(Z)x exists over each interval a < t < w of
continuity for A.
4. Verify that inequality (8.9f) implies inequality (8.9g). [Hint Argue as in the proof of
Gronwall’s inequality. ]
5. Suppose that f has real-valued components and that these components have continuous
first partial derivatives D;f; at every point of R” GR! (Gj = 1,...,n). Show that every
solution of x’ = f(x, A) is defined on an interval of the form fo < t < + © if there exists
a constant M such that |D;fi(x, )| < M for all (x, 4) and alli, j = 1,...,7.
6. Is the assertion of Problem 5 true if M is a continuous, positive function of ¢ for
—~a <t< +o?
7. Show that all solutions of the following equations exist on intervals of the form to <
IES eC,
a) x’ = yew 4+ 1, y = 1 xe b) x = V1 + y2, yy = V1 + x2.

8. Describe the solutions of

a) x) = 4, _y’ = —— and
Vx2 + y2 Vx2
+ y2

b) x’ = = Vy
Vx2 + y2 Vx2
+ y2
near the singular point (x, y) = (0,0). Relate your discussion to Theorems 8.5 and 8.6.
9. Let f be continuous and lipschitzian on &" X ®!. Suppose there exist continuous
functions a and b such that |f(x, | < a(d|x| + b(). What hypotheses on a and 6 will
guarantee that all solutions of x’ = f(x, f) exist on intervals of the form fo < t < +a?
10. Let f be continuous and lipschitzian on ®” X G!. Suppose there exists a real valued,
continuously differentiable function W, defined on ®”, such that W(x) — + as |x| >
+co and grad W(x): f(x, 4) < 0 for all (x, 4. Show that every solution of x’ = f(x, A
exists on an interval of the form to < t < +o.
11. Examine the intervals of existence of solutions for the equation

x + f(x)x’ + g(x) = e(),

where f and g are continuously differentiable and e is bounded and continuous for all f.
[Hint: Apply Problem 10 with W(x, x’) = x!” + f% g(s) du.]

8.10 DEPENDENCE ON INITIAL VALUES

Consider as above the system

x = f(x)4); (S)

where f is continuous and locally lipschitzian on a domain © in 8” X @1. For any

(Xo, fo)in D, the solution vector (x 9, fo, ¢) is defined on an open interval r(x 9, to) <
t < T"(Xo, fo). We want now to think of @ as a function of n + 2 variables: the
 Dependence on initial values 257

n-components of Xo, fo, and t. The domain of definition of @ is then

Q = U I(Xo, to),
(xo,to)eD

where J(Xo, to) is the open line segment in &” X ®! X @! defined by

I(Xo; to) = {(Ko, to, t): T (Ko, to) < t < T*(Xo, fo}.
Several questions of interest arise: (1) What are the properties of 2? (2) Is @ con-
tinuous on Q? (3) Does ¢ have continuous partial derivatives on Q with respect to
one or more of its variables? These questions are discussed below, but proofs are
omitted as they are better left to graduate courses in the subject.

Example 1. Consider the equation x’ = —t/x onthedomain x > 0,-» <t<t+a.

The maximal solution function is

Og = $(Xo; to, t) = x6 = iP sis 16

with 7 (Xo, to, t) = —V x2 + 12 and r*(xo, to, 1) = +V x2 + 12. The domain

of definition of ¢, which is illustrated in Fig. 8.19, is the region Q exterior to the cone
r= cA + 12. The set Q is an open, arcwise connected subset of @R® and ¢ is con-

t t

—V 2+ 6

Fig. 8.19 The exterior of the cone on the left is the domain of definition of $(xo, to, 1) =
[Link] b Ants #2. The vertical line segment indicates the interval of existence 7~ (x0, fo) <
t < 7+ (xo, fo). Phase-time space for the equation x’ = —f/x is the shaded half-plane
on the right. Notice the (¢(xo, to, 1), ))—for fixed (xo, to)—approaches the boundary of
phase-time space both as t > rt andast—7~.
258 The existence of solutions 8.10

tinuous on it. In fact,

0 Xo d Xo
ve (Xo, to, t) ca A/ Bie Pas 2 . (Xo, to, t) = A/a ee ae
0 rte ets 0 a eee
exist and are continuous on®. ||

In Example 1, (xo, fo, t) is explicitly known, and the properties of ¢ can be imme-
diately determined. The four questions asked at the beginning of this section can in
fact be answered without an explicit representation of (Xo, fo, ft).

Theorem 8.7. The domain of definition Q of the (maximal) solution function x =

(Xo, to, 4) of x’ = f(x, 2) is an open set and ¢ is continuous on Q.

To say that 2 is open is to say that no point of 2 is on the boundary of 2. Thus

if (Ko, Zo, 2) is in Q, then all points (xo, fo, f) sufficiently near it in 8” X ®? are also
in Q. As a consequence, $(Xo, fo, t) is defined. One may therefore rephrase the
openness assertion as follows: if (Xo, Zo, 2) is in Q, then there is a number r > 0
such that (Xo, fo, #) exists whenever |xo — Xo| + |to — Zo| + |t — 7] <r.
The openness of Q has a geometrical interpretation in phase-time space D. Con-
sider the graph in © of the solution x = (Xo, /o, t) for 79 < t < ¢ and let r be

g
[nn Ne
($(4o, to,t),1) »

Figure 8.20
 Dependence on initial values 259

chosen as above. If € < r is an arbitrary positive number, then the graph of x =

(Xo, fo, 4) remains in D for to < t < i —€ provided |xo — Xo| + |to — fol <
r —é. Roughly speaking, one may say that two solutions with sufficiently close
initial points may be continued for approximately the same closed interval in 7. See
Fig. 8.20 for an illustration of the situation when n = 2.

Figure 8.20 has been drawn in such a way that ($(xo, fo, £), t) does not appear
to be close to ($(Xo, Zo, 2), 7) since the purpose of the figure is to geometrically
illustrate a consequence of 2’s openness. Actually, these two points can be brought
as near each other as desired by merely bringing (Xo, fo, t) sufficiently near
(Xo, 0, 2). This is the geometrical interpretation of the continuity of @ which may be
phrased analytically as follows: given € > 0 and any point (Xo, fo, 7) in Q, there is
a 6 > O such that
|o(Xo, to, 1) — (Xo, fo, 2| < €
provided

xo — Xol + |to — Zo| + |t — Z| < 6.

Note that the continuity of@ does not guarantee that |@(Xo, fo, 1) — (Xo, to, D)| < €
for arbitrarily large tf no matter how small |xo — Xo| + |fo — Zo| might be.
The equation x’ = x/(1 — #t) with —o < x < +0 and —ow < t < +1 pro-
vides an illustration of the remarks above. Its maximal solution function is x =
(Xo, to, t) = xXo(1 — fo)(1 — 1). Phase-time space and graphs of x = (Xo, fo, t)
for several values of (xo, fo) are illustrated in Figure 8.21.

($(20, fo, ¢), 2)

(¢(x0, to, 0), 0)

(£0, fo)

Fig. 8.21 If the point (xo, fo) is sufficiently near the point (Xo, to) and if ¢ is sufficiently
near 7, then ((xo, fo, 1), f) is “near” (¢(Xo, Zo, 2), 2). Notice that |(xo, to, ) — (0, 0, 1}
is unbounded as t > + no matter how near (xo, fo) is to (0, 0).

Example 2. The knowledge that the solution of an ordinary differential equation

depends continuously on initial values has many applications in the theory of such
 The existence of solutions 8.10
260

equations. We illustrate by showing that the boundary value problem

pAddr eee, ena 0 X(0) 0 9x'(0) = On i) et (8.10a)

has a solution.
Consider the solution (7, f) to the initial value problem

alt ex!” = 0, xO) 10, oc 0) 058 (0) a 0.

We showed in Example 5, Section 8.9 that y(7, ¢) exists on the entire interval
0<t< +o. By Theorem 8.7, y’(7, 1) is a continuous function of Y for 0 <
Y < +o. We show that y’/(7, 1) < 1 for y > 0 sufficiently small and ¥’(7, 1) > 1
for Y > 0 sufficiently large. The intermediate value theorem then implies that there
is a Y = Yo such that (Vo, 1) = 1. Consequently, x = ¥(7o, t) is the solution
of the problem (8.10a).
We know from Example 5, Section 8.9 that 0 < y’’(7, t) < ¥ for all Y > 0 and
all 7> 0, This V4, ty <7¢ and vs) < 76-72) Wy <— then Gee
Suppose, on the other hand, that Y > 0 is taken so large that

6[1 — exp (—Y/6)] > 1.

We write the differential equation in the form y’’’(7, 1)/W’’(, t) = —wW(, t) and
integrate to find that

In (WY, t)/Y) = =i WY, s)ds > —vt?/6.

Thus y’(7, t) > Y exp (—7t?/6) and it follows that

1 Ay

AGEN) oe vf ET RS vf eat ds
0 0
=| rem aie ole
The existence of a solution to the problem (8.10a) then follows from the intermediate
value theorem in the manner indicated above. ||

Theorem 8.8. Consider the system x' = f(x, t), where f has continuous rth partial
derivatives with respect to its first n arguments on a domain D in ®” XK @!. The
solution function x = (Xo, to, t) may be differentiated r times, each differentia-
tion being with respect to any one of the n + 1 variables x, to, provided at most one
differentiation is with respect to ty. The partial derivatives, of order up to r inclusive, so
obtained are continuous on Q and each of these has a continuous first partial derivative
with respect to t.

This theorem is proved by verifying its truth for the case r = 1 and giving an
inductive argument. In connection with the proof, a useful representation for the
first partial derivatives of is shown to be valid. To state the result, we shall need
a definition.
 Dependence on initial values 261

Definition. Consider the solution x = (Xo, to, t) of x’ = f(x, t) with xo and to

fixed. Let Js(t) denote the Jacobian matrix whose i, jth element is Dj f (@(Xo, to, t), t).
The linear homogeneous differential system

y' = J(dy, TX, Vo) te Fal [Link])

is called the (first) variational equation based on ¢.
Theorem 8.9. For a fixed j = \,...,n, let p;(t) denote the first partial derivative
of ¢ with respect to its jth argument evaluated at (x0, t,t), that is, ;(t) = Dj@(Xo, to, 2).
Then f(t) is the unique solution to the initial value problem

y’ ma Je(Dy, y(t) = [Oe sees Leeson Os

where the jth component of y(to) is one and the other components are zero. Further,
Y = Dn+16(Xo, to, f) is the unique solution to the initial value problem

y’ = J(ty, Y(to) = f(Xo, fo).

This theorem is proved by justifying the following formal computation
a 5
V(t) = at D;o:(Xo, lo, t) = D; at i(Xo, to, t)

0
= 8X9) fild(Xo, to, t) t)

= aE, Dif (oxo, to, i); t) D joi(Xo; to, t)

= grad fi(¢(Xo, fo, )) - Wi),

where the ‘‘:”’ denotes euclidean inner product. For a specific equation, this computa-
tion is the starting point for the study of the partial derivatives with respect to initial
conditions. Their study is thereby reduced to the study of a linear homogeneous
equation, the first variational equation. The theory of Chapters 4, 6, and 7 then comes
into play.
Example 3. We illustrate Theorem 8.9 by applying it to a nonlinear equation. Spe-
cifically consider the initial value problem
x”wt +x’ ”
+) 25)
x’ +x see
t =
0, (8.10b)
Nea ke
= 1, ox = 0 awhen, 2 —0;

and let x = (a, 1, 0, 0, 1) denote its unique solution. We shall compute

D,W(0, 1, 0,0, 2). Notice first that ¥(0, 1,0, 0, 1) = t. We convert the equation in
(8.10b) into a system by defining x’ = y, y’ = z, 2 = —z-—1+y*—x+t.
If (a, B,Y,7, 1) denotes the solution for the system satisfying $(a, 6, Y,7,7) =
[a, 8, Y]’, then the variational equation based on ¢(0, 1, 0,0, 7) = [¢, 150} "is
Wi Ok Oe
|| = QO We UN aelys (8.10c)
w =i 3 he
 8.10
262 The existence of solutions

Now define g(t) = D,¢(0, 1, 0,0, 1). By Theorem 8.9, g is the unique solution of
Eq. (8.10c) which satisfies the initial condition g(0) = [1, 0, O]’. This initial value
problem was solved in Problem 10 of Exercise 4.6. The first; component gi(t) of
g(t) is :
er b=7 ) onl + vB
git) = —3e' 4 € = 4wey ae
But g,(t) = D,yv(0, 1, 0,0, 4). Thus the computation is complete. ||

EXERCISE 8.10
1. Compute the solution x = (xo, fo, ) for the equation x’ = t/x, x > 0. Describe
geometrically the domain 2 of ¢. Compute

5 #(%0, to, t), = oC, to, t), £ (x0,to, ¢).

2. If (xo, yo, to, t) denotes the solution of

x =y+x—x?-y%), y = —-x+ yl — x? — y?)

which satisfies (xo, yo, to, t) = (xo, yo), then @(0, 1,0, 7) = (sin?t, cos 7). Describe the
behavior of
0 fe)
a ae t) and aya t)
as f— +.

3. Let (a,B, Y, 2) denote the solution of x’”” + xx’’ = 0 which satisfies

0 Qo”
Va, B, MW 0) = a, 37 Vm Bs We 0) a B, 372 (Bs V- 0) aes

Describe the behavior of

) fe) te)
40 V0 {; 0, t), OB VC 0, il 0, t)’ ay vO 9431, 0,V5 1)
aa

ase — > oor

4. Let (xo, 0, 0, 7) denote the solution of

x = x — yp ox In V 72 eye,

y=x+y— ylnvx2 + y?
which satisfies (xo, 0, 0,0) = (xo, 0). Use Theorem
8.7 to deduce that there is a value
Xo such that the trajectory of (x, y) = (Xo, 0, 0, 4) is a simple, closed curve in the xy-plane.
(It will then follow from Theorem 9.5 in the next chapter that (x, y) = (Xo, 0, 0, 2) is
periodic.)
5. Let (u, fo, ) denote the solution of x’ = f(x, 4) which satisfies @(u, to, 0) = u, where
f is continuous and lipschitzian on a domain ® X (a, +0). Let xo in @ and to > @ be
given. Explain analytically and geometrically what is meant by the following statement:
“ is continuous in u at xo uniformly with respect to ¢ in the interval [to, +00),”
 CHAPTER 9

AUTONOMOUS SYSTEMS

9.1 PRELIMINARIES

The differential equation

x’ = f(x, 4) (S)

is called autonomous if the function f is a constant function of ¢ for each fixed x,

that is, if f does not explicitly depend on ¢. Similarly x = f(x, x’,...,x@7?, 2)
is called autonomous if f is a constant function of t. For example, the equations

Xe Xe) IN = 2, Xie — DT

are not autonomous, but the equations

x’ = x? b] a Se Xe eX

are autonomous.
We shall denote a general autonomous system, which we assume as given through-
out the chapter, with the symbols
x’ = f(x), (A)
and we shall assume throughout that f is continuous and lipschitzian on a given
domain @ in ®”. In the language of Section 8.5, @ is the phase space for the system
(A). Phase-time space D is a cylinder of the form ® X (—«#, +). The phase
space & is frequently called the Poincaré phase plane, when @ = ®?.
Nonautonomous differential equations occur in the physical world when a system
is subjected to external time-dependent influences. For instance, the angular dis-
placement 6 of the permanent magnet in an electric clock has as one mathematical
model the nonautonomous differential equation 6” + k(sin @)- sin (1207t) = 0, k
constant. The angular displacement 6 of the simple, frictionless pendulum, however,

263
264 Autonomous systems 9.2

satisfies the autonomous equation 6” + ksin@ = 0. There are no external time

dependent forces acting on the system.
We shall give the names orbit and solution path to a trajectory of an autonomous
system. The path of a constant solution will be called a critical point.

9.2 THE SHIFT FORMULAS

One normally studies the solution paths (orbits) of autonomous systems in phase
space rather than the graphs of solutions in phase-time space. To illustrate the dif-
ference in approach, consider the system x’ = y, y’ = —x of Example 3, Section 8.5.
The graph of (0, A, to, t) = (A sin (t — to), A cos (t — fo)) is a helix of radius A
about the f-axis. There is a different helix for each numerical value of fo. These are
illustrated in Fig. 9.1.

Figure 9.1

The projection of each helix onto the xy-plane is the circle x? + y? = A4?.
This circle is then a solution path corresponding to infinitely many solutions of the
differential equations. It is a geometric representation for each solution given by
$(0, A; 0,2) = (CA sini — fo), Acos (f — to)), no matter what the value of to
might be. The number fp is called a phase shift. Thus the circle x? + y? = A? is
the path of every solution which can be obtained from ¢(0, A, 0, 1) by introducing
a phase shift.
Let us consider two distinct solutions, say those given by $(0,A, fo, t) and
$(0, A, t1, t), where 4; > fo. We think of the path x? + y? = A? as being traced
out in phase space by each of two moving particles: p, with coordinates (0, A, fo, 2),
and q with coordinates (0, A, ¢,,1f). The particle p is at the point (0, 4) when
t = to and the particle q is at the point (0, A) when t = ft, > fo. See Fig. 9.2. A
moment’s reflection should convince the student that (0, A, t,, t+ ¢, — fo) =
 The shift formulas 265

Fig. 9.2 The point p is moving ahead of the point q by t; — fo units in time.

(0, A, fo, £1), which is just a way of saying that q is moving along the same path as
p but is 4; — fo seconds behind p. This phenomenon occurs for autonomous systems
in general.
Theorem 9.1 (The First Shift Formula). Let (xo, to, t) denote the solution of the
autonomous system x’ = f(x) which satisfies (Xo, to, to) = Xo. For any real num-
ber ty, (Xo, t1, t + ty — to) = o(Xo, to, t) as long as both solutions are defined.
Equivalently
$(Xo0, fo, 1) = (Xo, to + a, t+ @)
for all real numbers a and t such that both solutions exist.
Proof. Let &(t) = (Xo, t1,t + t1 — to) and let n(t) = O(Xo, to, 2). Then é’(f) =
f((t)) and y/(t) = f(n(4). But
E(to) = (Xo, f1, 11) = Xo = (Xo, fo, fo) = n(to).
Since the solutions of initial value problems are unique, &(t) = y(t). The equivalent
formula is obtained by setting a = ft; — fo. ||

Corollary [Link]. Initial value problems for the system x' = f(x) may always
be posed at time t = 0. In fact, $(Xo, to, t) = (Xo, 0, t — fo).

Proof. The solution vector &(t) = $(Xo, fo, t) for the initial value problem x’ =
f(x), X = Xo when ¢ = fo is related to the solution vector n(t) = (Xo, 9, 2) for
the problem x’ = f(x), x = Xo when f = fo by the formula &(¢) = n(t — fo). If
n(t) is known, then so is &(t). ||
Corollary 9.1b. If C is the path of the solution vector (Xo, 0, t), then it is also
the path of every solution vector $(Xo, to, t) produced by introducing a phase shift to.
Proof. (Xo, 9, t) = (Xo, to, t+ to). ||
The formula
(Xo, to, 1) = (Xo, fo + a, f+ @) (9.2a)

will be called the first shift formula.

 9.2
266 Autonomous systems

In view of Corollary 9.1b, it is sufficient to work with (Xo, 0, t) in order

to study the system x’-= f(x). We shall denote its interval of existence by
T (X09) < t < T*(Xo). The interval of existence for (Xo, fo, 4) is then 7 (Xo) +° to <
t <+(Xo) + to. In the notation of Section 8.7, then, 7*(xo, 0) = 7™*(Xo) and
TT (Xo, to) = Tt (Ko) + fo.
The second shift formula which we shall give below can be motivated, as was the
first, in terms of the motion of a hypothetical particle along a solution path in
phase space. Again, let (Xo, fo, t) denote the solution vector for x= te) wach
satisfies @(Xo, to, fo) = Xo. Suppose a particle p is at the point xo at time fo and
moves along the solution path through xo for a time interval of length a as in Fig. 9.3.

Figure 9.3

Then at time ¢; = to + a, the particle is at the point x; = (Xo, fo, t;). If the
motion of the particle continues for an additional time interval of length 8, then at
time ¢ = t} + B = to + a+ 8B, the location x of p can be described in two ways:
xX = (Xj, 7%), 2) and x = (Xo, fo, t). But x; = (Xo, fo, 1). Thus

(Xo, tos t) = $( (Xo, to, £1), 41, £). (9.2b)

We shall call formula (9.2b) the second shift formula. We prove that it holds in general.
Theorem 9.2 (The Second Shift Formula). Let $(Xo, to, t) denote the solution of
the autonomous equation x’ = f(x) which satisfies (Xo, to, to) = Xo. Then at any
times t, ty for which the indicated solutions are defined,

(Xo, to, 1) = $( (Xo, Lombidetts :),

Proof. Let &(t)= $(Xo, to, f) and let n(t) = o(¢(Xo, fo, t1), t1, 1). Then £’(1) =
f(E(t)) and n/(t) = f(n(t)). But £(t1) = $(Xo, to, t1) = 6( (Ko fo, t1)s 1, ii) =
n(t1) by the definition of @. By the uniqueness of solutions to initial value problems,
&(t) = m(t) for all ¢ for which they are defined. ||
Note that Theorem 9.2 is true also for nonautonomous equations x’ = f(x, 2).
 Phase portraits 267

EXERCISE 9.2
1. Solve the nonautonomous system x’ = y, y) = —x ++. Depict in phase-time space
the graph of the solution @ which satisfies @(0) = (0,1) and depict in phase space the
trajectory of @. Do the same for the solution W which satisfies ¥(0) = (1, 1) and compare
your results.
2. Sketch two trajectories of solutions to the system x’ = x, y’ = 2tx + y which cross
each other in the xy-plane. Could two distinct paths of solutions for the system x’ = x,
Va ydormthis:)
3. Interpret analytically and geometrically the first phase shift formula by explicitly solving
the autonomous system in Problem 2.
4. Let y denote the solution to the initial value problem x’ = x?, x = 1 when ¢ = 1.
Show that yw is the solution to some initial value problem x’ = x?, x = xo when t = 0
which is posed for initial time zero. What is the interval of existence of ¥? Relate the
example to Corollary 9.1a.
5. Let y denote the solution to the initial value problem x’ = 2¢x?, x = 1 when ¢ = 1.
Show that y is a solution to an initial value problem x’ = 21x”, x = xo when f = 0 even
though Corollary 9.1a does not apply.
6. Explicitly solve x’ = x? for o(xo, fo, 2) and $(x1, f1, 2). Then verify the shift formulas
(9.2a) and (9.2b) algebraically.
7. Explicitly compute all solutions of

x= —-x-yt+— oy’ = x - yy+ —*—

a/ x2 “f y2 \/ x2 he y?

which have the unit circle in the xy-plane as their path. Verify the shift formulas (9.2a) and
(9.2b) algebraically for these solutions.
8. Consider two identical linear oscillators of the type depicted in Fig. 1.3. One block is
set in motion by giving it unit initial velocity with the spring unstressed. The other block is
set in motion by stretching the spring one unit and releasing the block with no zero velocity.
What information about the future relative motion of the blocks is provided by the shift
formulas ?
9. Let x’ = f(x, ‘) have the property that there exists an a > 0 such that (xo, fo, 4)
satisfies (xo, to, 1) = (xo, to + a, t + a) for all (xo, to). What general property must f
have?
10. Suppose the identity (xo, to, ) = $(xo, to +a,t-+a) holds for every solution
vector (xo, fo, 2) of x’ = f(x, 4) and every a > 0. What general property must f have?
(Recall that f is assumed to be continuous.)

9.3 PHASE PORTRAITS

When one is confronted with the need to study a physical situation which is described
by an autonomous differential equation
Tix) (A)
it is frequently profitable to study the solution paths of the system qualitatively. This
268 Autonomous systems 9.3

Figure 9.4

is not the same procedure as solving the differential equations, for the emphasis is
upon geometry, and time dependence is a somewhat subordinate consideration.
A phase portrait for the differential equation (A) is a complete, qualitative
(perhaps pictorial) description of a// its solution paths. Phase portraits are most
easily given for the case n = 2 since they can be readily illustrated with drawings.
In higher dimensions, verbal descriptions are used.
Let us, for example, suppose that we wish to predict all possible motions of the
linear spring-magnet system in Fig. 9.4. We assume that the units are such that the
differential equation for the displacement x of the first magnet from its position when
the spring is unstretched is x” + x — (x — 2)? = 0.
It would be helpful if the student would stop at this point and conjecture on the
basis of his intuition what possible motions the left-hand magnet could undergo
were it displaced from equilibrium and given, perhaps, a nonzero initial velocity.
The first stop in an analytical study of the situation is to find the positions of
equilibrium. These correspond to constant solutions x = c. The numbers c must
satisfy the equation c — (c — 2)? = 0, that is,
¢c=1,.¢=6—V5)2 ~038, ¢©=6 + 5)/2= 267
We convert the equation into a system by setting
xX =y, y= —x + &— 2). (9.3a)
Phase space is chosen to be the region of the xy-plane to the left of the line x = 2.
The coordinates of the critical points (paths of the constant solutions) which are in
phase space are ((3 — /5)/2, 0) and (1, 0).
We obtain preliminary information about the solution paths by the method of
Section 2.8. The equation x’’x’ + xx’ — (x — 2)~?x’ = 0 is integrated, and we
find that each solution path must lie in the locus of one of the equations

ytx?4+2/(x —2=k,
where k is a constant and y = x’. The plotting of these equations for various values
of the constants k is basically a problem in the calculus. The labor involved is eased
somewhat by arranging the work in a definite order however.
i) Plot the critical points p = ((3 — »/5)/2,0) and q = (1,0).
li) Observe that the derivative dy/dx is given by the formula

3-75 3475
sp pepe tases al 2 Nx 2 )
ke SAG y =a ees)
 Phase portraits 269

Figure 9.5

except at critical points. Notice that the noncritical solution paths have vertical
tangents at places where they cross the x-axis. They have horizontal tangents at
places where they cross the lines x = (3 — +/5)/2 and x = 1.
iii) Find k for each critical point and plot the corresponding locus. At the
point (1, 0), for example, k = —1 and the equation ofinterest is y?> = —x(x — 1)?/
(x — 2). The graph of this equation is the heavy curve in Fig. 9.5. It is a tedious,
but straightforward, procedure to show that k = (5 — \/5)/2 at the point x =
(3 — /5)/2, y= 0. The graph of the corresponding equation y® + x? + 2/
(x — 2) = k contains only that point.

With this information, the preliminary sketch of Fig. 9.5 can be given.
If one thinks of each curve as being traced out parametrically by a point with co-
ordinates (x(t), y(t)) where x and y satisfy Eqs. (9.3a), then a time sense is assigned
to each curve and we have indicated it with arrowheads.
In order to use the depicted information to discuss the motion of the magnet, it
is necessary to make a few additional remarks about autonomous systems in general.

Theorem 9.3. No solution path of the equation x' = f(x) can intersect a different
solution path.
 Autonomous systems 9.3
270

Proof. Suppose two paths have a common point Xo. Let (Xo, 0, 2) and ¥(Xo, 0, 4)
denote solutions having these paths, where ¥(xo, 0,0) = Xo and (Xo, 0,0) = Xo.
Since $(Xo, 0,0) = ¥(Xo, 0, 0), it follows from the uniqueness of solutions to initial
value problems that (Xo, 0, t) = ¥(Xo, 0, 1). Hence the paths coincide. ||

Theorem 9.4. Let $(Xo, 0, t) denote the solution of the equation x’ = f(x) which
satisfies @(Xo, 0,0) = Xo. If (xo, 0, 4) approaches a point a in @ as t— T* (Xo),
then T*(X9) = +x and a is a critical point.
Proof. It is an immediate consequence of Theorem 8.2 that T'(Xo) = +o. To
show that a is a critical point, one assumes for contradiction that f(a) # 0. Since
¢ is a solution of x’ = f(x)

$(Xo, 0, t) = Xo Ets i £( (Xo, 0, s)) ds. (9.3b)

Since f(a) ~ 0 at least one component of the f(@(xo, 0, s)) does not approach
zero aS s—> +o. Thus, at least one component of the improper integral
fot* £(@(Xo, 0, s)) ds diverges as t— +a. This gives a contradiction when one
takes the limit on each side of Eq. (9.3b). ||
It follows from Proposition 9.3 that the heavy curve in Fig. 9.5 contains at least
four solution paths since q is a solution path the deletion of which would cut the
curve into three parts. Each of these three parts (A, B, C) is a path in its own right
by Proposition 9.3 since it contains no critical point. For the same reason, every
other curve in the figure is precisely one solution path.
The physical motion of the magnet can now be almost completely described.
There are two positions of equilibrium x = (3 — /5)/2 and x = 1. If the magnet
is very carefully displaced until x = 1, it will remain at rest. If it is perturbed slightly
to the right, it will be attracted to the fixed magnet and its velocity y = x’ becomes
(ideally) infinite as x — 2. If it is perturbed slightly to the left, it will move in that
direction until the spring is unstressed, then it will return. If the magnet is very
carefully displaced until x = (3 — »/5)/2, it will again remain at rest. Now if it
is slightly perturbed, it will undergo an oscillatory motion corresponding to the
simple, closed paths about p in Fig. 9.5. One is tempted to assert that the corre-
sponding solution functions are periodic. This is indeed true, but the statement
requires proof.
Theorem 9.5. Let C denote a path of the autonomous system x' = f(x). The
following statements are equivalent:
i) C intersects itself in at least one point.
ii) C is the path of a periodic solution.
ili) C is a simple, closed (Jordan) curve.
Proof. Suppose statement (i) is true. Let C intersect itself at the point xo in phase
space. Then we may regard C as the path of the solution @ which satisfies
 Phase portraits 271

$(Xo0, 0,0) = xo. Let T > 0 be the first time at which C intersects itself, that is,
the first positive time at which @(xo, 0,7) = xo. We shall show that

O(Kon0K tat) tl xq 2) ae fore Om 1/75 (xp) — Te

By the second shift formula, with t, = T,

$(Xo, 0, t i TP) = $($(Xo, 0, ip), ay. Ua p),

But $(Xo, 0,7) = Xo. Thus

(Xo, 0, t ot ibe) == $(Xo, Ls t ete i) a $(Xo, 0, t), (9.3c)

where the last equality stems from the first shift formula with a = —T.
Now suppose that 7*(xo) < +o, then as t—7*(xo) — T, $(Xo0,0,t+T)
becomes unbounded but (xo, 0, f) cannot become unbounded. This contradicts
Eq. (9.3c), and it follows that (xo, 0,t+ 7) = $(Xo, 0,4) for O< t< +o.
The argument for —«# < t < 0 is analogous.
Now assume that (Xo, 0, f) is periodic with (minimal) period T > 0. Then
C is certainly a closed curve. We must show that it is simple, that is, we must show
(Xo, 0, a2) = (Xo, 0, b) with O < a < b < T implies a = b. Suppose a < b and
define a = b — a. Writing the numerals one and two, respectively, above the appro-
priate equality symbols to indicate application of the first and second shift formulas,
compute
$(Xo, 0, + a) 2 $( (Xo, to, ), b, t+ a) = 6((Xo, fo, a), 6, t+ a)
||— $($(Xo, fo, a), 6 — a, t) = $(4(Ko, fo, @), a, t)
2 $(Xo, 0, 2).
This means that a is a period for (Xo, 0, ft). Since a < T, the minimality of T is
contradicted, and the assertion is verified.
If statement (iii) is true, statement (i) is clearly true. ||
The student should compare this theorem with Theorem 9.3. Broadly speaking,
one says that no solution path can intersect another, and should a path intersect
itself, then it is a simple closed curve. This is not true for nonautonomous systems.
Because of Theorem 9.5 and applications to celestial mechanics, a solution path which
is a Jordan curve is sometimes called a periodic orbit.

EXERCISE 9.3

i. A rigid electrical conductor is located at a distance of ZL units from a wall as illustrated

in Fig. 9.6. Attached by linear springs to the wall is a bus bar which carries a parallel,
similarly oriented, direct current.
a) Use Newton’s second law and the law of Biot-Savart to show that the displacement x
of the bus bar from the wall is given by a solution of a differential equation having
 9.3
272 Autonomous systems

Figure 9.6

the form
k
Ae 4 — =
of: ib = 7 us

where w and k are positive constants.

b) Predict the possible physical motions of the bus bar by sketching an appropriate
phase portrait. Is a periodic motion possible? In an actual situation is a periodic
motion likely ?
2. Consider a block having unit mass which moves on a horizontal frictionless surface
under the influence of a nonlinear spring. The restoring force is of magnitude g(x), where
x denotes the extension of the spring and xg(x) > 0 for x ¥ 0.
a) Show that the motion is periodic if fj g(t) dt > + as |x| > +o.
b) Suppose lim)z|,4 Jo g() dt < +. Could the motion be periodic? Does the
motion have to be periodic?
3. Consider the model for orbital motion given in Section 1.5.
a) If u = 1/r, show that

d°u 2
do2 + USK 3

where k is a constant.
b) Discuss the physical motion of the satellite by constructing an appropriate phase
portrait.
4. Construct a phase portrait for the equation mx’’ + kx’ + (mg/f)sin@ = 0 of the
simple, damped pendulum.
5. Consider the oscillator of Problem 2 under the assumptions that (1) there is friction
which is proportional to the velocity of the block and (2) that limjzi540 Jo @(@) dt <--o.
Describe the nonperiodic motions, if any, of the block.
 Divergence and Bendixson’s negative criterion 273

9.4 DIVERGENCE AND BENDIXSON’S NEGATIVE CRITERION

It is occasionally convenient to regard a phase portrait for an autonomous system as
though it were a depiction of a steady-state fluid flow. The solution paths can be
regarded as streamlines, and the arrowheads indicate the direction of flow. Critical
points can be regarded as sources or sinks or combinations. For example, the phase
portrait for the linear system x’ = x, y’ = —2y (see Fig. 9.7) can be regarded as
a picture of water flowing.

Figure 9.7

If one thinks of the triangle S in the picture as an oil slick, his intuition should
tell him that it will be carried along and distorted by the current. At a later time, it
will have become the triangle S(z). In more formal language, one says that the set S$
has been mapped onto the set S(t) by the flow mapping.

Definition. Let (xo, 0, t) denote the solution of the autonomous equation x’ =

f(x) which satisfies (Xo, 0,0) = Xo. For each fixed, positive t < T*(x9), the corre-
spondence Xo — (Xo, 0, t) is called the flow mapping.

Working with the flow mapping can be a bit tricky since T*(xo) can usually be
expected to vary with x9. For example, the autonomous system x’ = x” has the
real line —w < x < +o for its phase space and has ¢(xo, 0, t) = xo/(1 — Xof).
274 Autonomous systems 9.4

—f——_+——_+—___+—+> «
0 ye 2 3 4
Lash
Figure 9.8

For t = 1, the flow mapping carries the interval [1,2] onto the interval [%, 4]
(Fig. 9.8). For ¢ = 3, the flow mapping is not even defined for all xo in [1, 2].
The difficulty which occurs in the example occurs because the interval [1, 2]
is too long an interval for the mapping x9 — xo/(1 — 3xo/4) to be meaningful
everywhere on it.
Consider now two distinct points p and q on the path C of a solution x =
$(p, 0, 2) of (A) as shown in Fig. 9.9. Suppose that f has continuous first partial
derivatives on @. Then it is known that if one chooses two sufficiently small x — 1
dimensional neighborhoods P and Q (open in ®”—') with p and q as centers and
if C is not tangent to either P or Q at p or q, respectively, then the flow mapping
from P to Q is defined at every point Xo of P. In fact, one may choose Q so small
that P is mapped onto Q. The mapping is one-to-one, continuous, and has a con-
tinuous inverse. Geometrically, one thinks of a fibrous tube joining the points ofP
to the points of Q; it is called an open path tube.

Figure 9.9

It is of interest to know how the area of a set S changes as it is carried along by

the flow mapping. Let us confine our attention to two dimensions and consider an
autonomous system

x4 = fi(%1, x2), xb = fo(X1, X2), (9.4a)

where f; and f have continuous first partial derivatives on the plane R?. We suppose
that there is a closed and bounded region S with the property that every solution
X1 = $1(Uy, Ug, t), X2 = $2(U4, U2, t), which initiates at a point (uy, uz) in S at
t = 0, exists forO < t < +o. The functions ¢, and ¢2 being continuous, it follows
from the material in Section 8.3, that the image S(t) of S under the flow mapping
(ui, U2) > (X41, X2) is also a closed and bounded region for each fixed f.
Since no two solution paths may intersect each other, the correspondence
(ui, U2) > (X1, X2) is one-to-one; and the change-of-variables formula for double
 Divergence and Bendixson’s negative criterion 275

integrals implies that the

Area of S(t) = ie VaGy. xo) = i det Jo(u1, uy, t) d(uy, ue),

where

ifo(U1, 7 Ue, Didi(uy, U2, 0) Dodi (u4, Us, 4 ,

2, b)
t) = aeete t) Do
do(uua,y,t)
To compute the area of S(t), then, we must first examine the vectors

Vi) = (Didi, U2, t), Dydo(uy, Uo, t)) = D1 (uy, uo, t)

and

YH = (Doei(u1, U2, t), Dodo(uy, Uo, t)) = D2$(uy, ua, t),

where D; = 0/du; and Dy = 0/dug. Theorem 8.9 provides the needed information
since it guarantees that Y,; and yp» are solutions of the linear variational system
y’ = J,(t)y, where

Oe es X2,1) Dofi(%1,
x2, 4
Difax 1X ontews ova Xa, De ageae,
Further 1 (uj, v2, 0) = (1,0) and Po(uy, ve, 0) = (0,1). Thus, if Y = [Wy, Wo] is
the corresponding matrix solution of y = J;(t)y, then Abel’s formula (4.5b) implies
that
t t

det W(t) = det V(O) exp || Tr Je(s) as= exp || Tr J¢(s) as:
0 0
But Tr J-(s) = Dif 1(%1, X2, 8) + Dofo(x1,
X2, 5) = div f(x, x2, 5), and one may
therefore write:

Area of S(t) = i| exp fldiv f(o(u, 5)) ds|dus, us).

0

Example 1. Suppose S is the triangle in Fig. 9.7. The divergence of the coefficient
function is (0/dx)(x) + (0/dy)(—2y) = —1. The area of the triangle S(t) equals
the value of the integral

[| exp || (1) asd(uy, U2) = e+ II.d(u;, us) = e (Area of S). ||

Ue) Ss 0

We summarize the discussion above as a theorem.

Theorem 9.6. Let the function f = (f1, f2) in the autonomous system (9.4a)
have continuous first partial derivatives on a domain & in ®? and let S be a closed
and bounded region in ®. Suppose that $(Xo, 0, t) exists on the interval0 < t < +a
for each Xo in S and denote the image of S under the flow mapping Xo — $(Xo, 9, 1)
276 Autonomous systems 9.4

by S(t). Then the area of S(t) is given by

a exp || div f(¢(u, s)) asd(uy, U2),

0
where div f(x, t) = D,f1(x, t) + Dof 2, 0).

This theorem has a rather immediate application to the theory of nonlinear

oscillations. It yields a basic test, called Bendixson’s negative criterion, for the non-
existence of periodic solutions.
Theorem 9.7 (Bendixson’s Negative Criterion). Consider the systemx' = f(x) where
f has continuous first partial derivatives on a domain & in %”. If div f(x) ¥ 0 on ®,
then there is no (proper) periodic solution whose path is contained in ®.
Proof. Suppose, for contradiction, that div f(x) > 0 on @ and that x’ = f(x) has
a (proper) periodic solution. The corresponding path is then a simple, closed curve
which together with its interior forms a compact region S. It follows from Theorem
8.5 that any solution initiating at a point of S exists forO < t << +a. Thus S can
be mapped forward indefinitely by the flow mapping. Since no solution path can
cross another, the area A(t) at time ¢ of the image of S under the flow mapping
satisfies A(t) < A(O). By Theorem 9.6, however,

A(t) = if exp /.div f((u, 0, s) as)|du > A(0).

This is a contradiction. The analogous argument holds with t < 0 if div f(x) < 0
on ®. ||
For a second order equation x’ = f(x, x’, t) the divergence condition above is
that of (x, x',.2)/0x'’ = Oxon C;

Example 2. Consider the van der Pol equation x’ + e(x? — 1)x’ + x = 0, where
€ > Oisa parameter. The divergence condition is (1 — x”) # 0. Thus, if the equa-
tion has a periodic solution (it does, incidentally), the path of the solution must
intersect the lines x = +1 in the associated phase space x = x, x’ = Xo. ||
Example 3. The equation x’’ + x’ + f(x) = 0 could have no real periodic solution
in the xx’-plane since d(x’ + f(x))/ax’ = 1. ||
There is a generalization of the notion of exactness (Section 2.5) which is occa-
sionally helpful in plotting the solution paths of a two-dimensional autonomous
system
x’ = f(%;y),
(9.4b)
y’ = g(x, y).
Theorem 9.8. Suppose the coefficients f(x, y) and g(x, y) of Eq. (9.40) have con-
tinuous first partial derivatives on a domain & in @®. If the divergence (af (x, y))/ax +
(g(x, y))/dy is identically zero on @, then there is a function V defined on & such
 Divergence and Bendixson’s negative criterion PAY

that f(x, y) = dV(x, y)/dy and g(x, y) = —dV(x, y)/dx. Thus the solution paths in
® lie in the loci of the equations V(x, y) = constant.

Proof. If (xo, Yo) is in @, define

x

V(x, y) a ifI(x, B) dp ge / g(a, Yo) da.

Then 0V(x, y)/dy = f(x, y) and

gles en (zits
Ox S g(x, Vo)
y0 Ox ae
y
02(x, B
= = eet) a — g(%, Yo) = —8(%, y).
Yo

If x = ¢(t), y = Yd) is a solution of Eq. (9.4b) in @, then along its path

1) Oe Vv av aV AV_
amberoy S20k Oy Oy ee
Thus V(¢(2), ¥(t)) = constant. ||

Example 4. Consider the system

x! =4xy
+ dy, (9.4c)
y’/ = —4y?x — 4x.
The divergence of the coefficient functions equals zero on ®”. Thus there is a function
V such that V(x, y)/dy = 4x”y + 4y%. Integrating, we have
V(x, y) = 2x?y? + y* + A(x), (9.4d)
where / is an unknown function. The value A(x) is found by differentiating Eq. (9.4d)
and comparing the result with the first of the equations in (9.4c).
aV(x, y) A
sie eg 4xy~ 2 + h'(x).
/

Thus h’(x) = 4x°, and it follows that h(x) = x* + c, where c is a constant. The
solution paths for Eq. (9.4c) lie in the loci

V(x, y) = x4 + 2x?y? + y* + c = constant,

which consist of the point (0,0) and the family of circles described by LE =
r?, r > 0. Notice that the solution paths are precisely the same as those of the
linear system x’ = y, y’ = —X.

EXERCISE 9.4
1. Consider the system
ery (ee yn yee ee yx? y*).
 9.5
278 Autonomous systems

a) Show that there is a unique periodic orbit.

b) Show that each forward solution exists on every interval of the formfo < t < +.

c) Let S denote the region $ < Vx2 + y2 < 3, 0 < tan7'(y/x) < 7/6. Describe
the image S(27) of S under the flow mapping.

2. A block of mass m rests on a belt which moves with constant speed v > 0. The block
is attached to a wall by a linear spring having stiffness coefficient & as indicated in Fig. 9.10.
Assume that the frictional force on the block is proportional to its velocity relative to the
belt. Is a sustained, oscillatory motion possible?

ae

Figure 9.10

3. Consider a pendulum consisting of a rigid, massless rod of length ¢ which supports a

particle of mass m. Suppose that the frictional torque on the rod at the pivot is proportional
to the angular velocity of the pendulum. If a constant torque M is applied to the rod, the
equation for its angular displacement 0 is

mt?o” + ké’ + mgfsiné = M.

a) Can the differential equation have a periodic solution ?
b) Under what conditions on the coefficients does the pendulum have an equilibrium
position?
c) Can the pendulum undergo an oscillatory motion?
4. The equation mx’’ + g(x) = 0 is said to describe the motion of a particle of mass m
moving along a line under the influence of a conservative force. Convert the equations to a
system by setting y = mx’ and note that Theorem 9.8 applies with V interpreted as the
energy of the particle.
5. Construct phase portraits for the following systems:
a) x = 2y/x?, y= 3y2/x*, x <0.
b) x = 3(1/Vy), y = —$(01/V2), x > 0, y > 0.
ce) x = 27+ DE = y)G — 2p) = 2x*y, yy) = Qxy?
6. Explicitly solve the equations (9.4c). If x = u(t), y = v(/) is a solution satisfying the
initial conditions u(0) = c > 0 and v(0) = 0, compute its period. Note that all solutions
of the linear system x’ = y, y’ = —x (which has the same phase portrait as Eq. (9.4c))
have period 2r.

9.5 LIMIT SETS OF SOLUTION PATHS

We proved in Theorem 9.4 that if the solution x = $(Xo,0,1) of x’ = f(x)
approaches a limit as t—> 7 (xo), then 7*(xo) = +o and the limit is a constant
 Limit sets of solution paths 279

solution the path of which is a critical point. If (xo, 0, f) does not approach a limit
as t—>77(Xo), it is still desirable to have a way of describing its behavior as
t— T* (Xo).
Example 1. When the differential equations

x = —X — y+ x
———— > y=x-—yt y ,
V/x? + y2 Vx? + y2
are written in polar coordinates, they have the form 7’ = —r-+ 1, 6 = 1. The solu-
tion paths are then easily depicted in the xy-plane (see Fig. 9.11), for their equations
in polar coordinates are r = ce~® + 1, where c is an arbitrary constant.
The path corresponding to the value c = 1 is the unit circle about the origin.
All other paths are spirals which approach the circle as ¢, hence 6 increases. Consider
the ray from the origin which cuts the circle at the point p and the indicated spiraling
path at the points p, po,.... If this spiral is the path of the solution x = (Xo, 0, 4),
then there is a sequence {t;,} of times such that pz = (Xo, 0, tx), tf; ~ +a as
k—-+o, and limz. o(Xo, 0, t.) = p. Thus, even though lim;,4 (Xo, 9, t)
does not exist, it is possible to describe precisely the behavior of (Xo, 0, ¢) for ¢
sufficiently large. The unit circle is called the positive (w-) limit set of the solution
(Xo, 0, 2). ||
Definition. Let $(xXo,0, t) denote a solution of the autonomous equation x' =
f(x), where f is continuous and lipschitzian on a domain @ in &". A point p in &” is called

Figure 9.11
280 Autonomous systems 9.5

an w-limit point of the solution if and only if there is a sequence {tx} of times such
that limz.40 te = +o .and lim 40 (Xo, 9, te) = P.- The w-limit set 2 of
(Xo, 0, £) is the set of all its w-limit points. One defines a-limit points and an a-limit
set analogously using sequences {t,} approaching —n ask—> +a.
In Fig. 9.5, the critical point q is the w-limit set of the path A; it is both the a-
and w-limit sets of B: and it is the a-limit set of C. The a-limit set of A and the
w-limit of C are empty. The critical points p and q are the a- and w-limit sets of
themselves. Notice, however, that p is not a limit set of any path other than itself.
The paths of which D and E are typical have empty limit sets of both types. The
situation for the closed paths, of which F is typical, is analogous to the situation for
the critical point p: the entire path is its own a-limit set and w-limit set.
We have therefore seen by example that a closed path or critical point may be
a limit set of a path other than itself, but neither must necessarily be such. A closed
path which is a limit set of a path other than itself is called a Jimit cycle. It is in-
structive to see an example of a limit set which is neither a critical point nor a limit
cycle.

Example 2. Consider the differential system

sles) Fie walpha DRS sea AN CaaS 2

(9.5)
aoe We cas ca el NEA en eet ap)
The critical points 0, p, q have coordinates (0,0), (—1, 0), (1, 0), respectively. If
(x(t), y(2)) are coordinates of a point on any path except 0, then the point can be
located in the plane with polar coordinates r?(t) = x?(t) + y?(t) and 6(t) =
tan~!(y(t)/x(1)). It follows from Eq. (9.5) that

£P0=POU-PO) Law = -y¥O -— @@ - 1%

If a solution x,y satisfies x°(0) + y?(0) = 1, then r?(t) = 1 and 6’(1) = —y2(1) —
y*(). There are four possibilities for the path of such a solution: the critical points
p and q and either of the open semi-circular arcs joining them. These four paths
together comprise the w-limit set of every solution except the critical point 0. ||
For autonomous systems with phase space @ in the plane ®?, it can be proved
that every limit set is a critical point, a closed path, or a combination of solution
paths and critical points joined together (two ends of a noncritical path may be
joined to the same critical point). Limit sets for systems with higher dimensional
phase spaces can be much more complicated. There can be constructed, for example,
a system with phase space ®* which has a nonclosed path that is dense on a torus.
The next theorem gives several properties of limit sets which hold for autonomous
systems of arbitrary dimension.
Theorem 9.9. Let Q denote the w-limit set for the solution x = $(Xo,0, Log
the equation x' = f(x), where f is continuous and lipschitzian on a domain ©& in &".
 Limit sets of solution paths 281

The set Q is closed. If $(xo, 0, t) is bounded for all t > 0, then Q is not empty. It is
then connected and consists entirely of whole solution paths. The analogous assertions
hold for the a-limit set of ¢.
Proof. If Q or its boundary is empty, then Q is certainly closed. The remaining
possibility is that neither Q nor its boundary is empty. Let p be a point in the boundary
of Q. Then there is a sequence {p;} of points of 2 such that p, > pask—> +a.
Since each p; is an w-limit point, there is a time tf, > k corresponding to each p,
such that |@(xo, 0, t,) — px| < 1/k. Thus

ld(Xo, 0, tr) = P| — |p a Px| Ze 1/k,

and passing to the limit one finds that $(xo, 0, t,.) > pas k > +x. Thus p is in
Q and Q is therefore closed.
Now suppose that $(Xo, 0, f) is bounded for all > 0. This implies that 7(xo) =
+ by Theorem 8.5. Any unbounded sequence of times in the interval0 <t< +e
has a subsequence {t;} such that {¢(xo, 0, f;,)} converges by the Bolzano-Weierstrass
theorem. The limit is an w-limit point. Thus © is not empty.
To show that Q is bounded, let M be a constant such that |(xo, 0, )| < M for
all t > O and suppose that there is a point p in 2 which satisfies |p| > M+ 1. Let
{t,} be a sequence of times such that t,-— +o and px = $(Xo,0, t,) > p as
k—-+o. For all sufficiently large k, we have 4 > |p — p;| > |p| — |p| >
M+ 1-— M = 1, a contradiction, and 2 is bounded.
Suppose, for contradiction, that Q is not connected. Since it is closed and bounded,
it consists of two closed and bounded sets Q; and {22 which are separated by a distance
6 > 0 as indicated schematically in Fig. 9.12. If p is in @ we shall write dis (p, ;)
for the distance from p to Q; (i = 1,2). Now choose a distinct sequence of times
t, such that dis (o(Xo, 0, tor41), 21) < 5/3 and dis ($(Xo, 0, fox), 22) < 6/3 for
k= "12... Since the path of x =>°@(xo, 0, 7) 1s an_arc, there is, for each k = I,
a time 5%, for < Sze < tox41, such that dis (@(Xo, 0, sx), 2;) > 8/3 for each %%.
Since $(Xo, 0, t) is bounded for t > 0, there is a subsequence {7;} of {sx} such

(x0, 0, Tj)

6 075 <—

oo

( (xo, 0, ¢)
we Ea

Figure 9.12
 Autonomous systems 9.5
282

that (xo, 0, 7,;) approaches a limit p as j> +. But then p must be in either Q,
or Q» and must satisfy the condition dis (p, 2;) > 6/3 fori= 1, 2. This is impossible.
Now let p be an arbitrary point of Q and consider the point q = o(p, 0, 7) for
any fixed T > 0. We prove that 2 consists entirely of whole solution paths by showing
that q is in Q. This is done by exhibiting a sequence s, of times such that s; — +0
and q, = $(Xo, 0, 5;,) > qask—-+o. As a starting point, let {t,} be a sequence
such that 4 ~ +o and py = $(X0,0,%)—> p as kK— +m. Then define 5 =
t, + T. By the second shift formula,

= (Xo, 0, 51) = 6(G(Xo, 9, th), te, Sz);

and by the first shift formula

$( (Xo, 0, tk), tk, Sz) = $(4(xo, 0, th)s 0, Sk. ty,).

Thus q, = $(pz, 0, 7) and q; — o(p,0,7) = qask—-+o. Thus q is in Q, and

it follows that the path of any solution starting in 2 remains in Q. Since some solution
path passes through every point of Q, it is a union of such paths. ||

EXERCISE 9.5
1. Let x = u(, y = v(2) denote the solution of the system in Example 1 which satisfies
the initial conditions u(0) = 2, v(0) = 0. Specify a sequence {t,} of times such that
(u(t,,), U(tn)) > (1,0) as t> +0.

2. Ifxisapointin &” and S is a subset of &”, define the distance from x to S by dis (x, S) =
inf {x — y|: y isin S}. Suppose that Q is the w-limit set of a bounded solution @ of x’ =
f(x). Show that dis (¢(),2)— 0 as t> +o.
3. Let f be continuous and lipschitzian on”. Assume thatQ is the w-limit set of a bounded
solution @ of x’ = f(x).
a) Show that if there exists a continuously differentiable, real-valued function V on ®”
such that grad V(x) - f(x) < 0, where ‘-’”? denotes euclidean inner product, then
there is a constant c such that V(x) = ¢c for all x in Q.
b) Illustrate (a) by applying the result to the system in Example 1. Take Q to be the
unit circle in the xy-plane and explicitly specify V(x, y) and c.

4. Consider the frictionless linear oscillators depicted in Fig. 9.13. Let the equations for
the displacements x and z from equilibrium be x’’ + x = 0 and z” + \2z = 0, where
SS O,
a) Define what is meant physically by the phrase “the system undergoes a periodic
motion.”
b) Show that the system can undergo a nonperiodic motion for certain values of }.
c) Let (x, y, z, w) denote an arbitrary point of ®4. For all positive numbers a and b,
the set
S(a, b) = {(x, y, z, w): x? 42 y? = a?, 2? + w? = 5?)
 Limit sets in the plane 283

Figure 9.13

is a torus in ®*. Show that the path of any solution of the system

Y= yy — 7 z’ = w, w = —d2z (T)

which initiates on a torus 3(a, 5) in the phase space ®* must lie entirely on that
torus.
d) Show that if \? is rational, then every solution path on a torus 3(a, b) is a simple
closed curve. Describe the physical motion of the oscillating blocks in this case.
e) If \? is irrational, show that 3(a, b) is the w-limit of every solution which initiates
on it. Describe the physical motion of the oscillating blocks in this case.

9.6 LIMIT SETS IN THE PLANE

We confine our attention here to the autonomous system x’ = f(x) with f continuous
and lipschitzian on a domain @ in ®”. It will occasionally be convenient to dispense
with vector notation and write
aa = FLX, y), (9.6)

y’ = g(x, y).
When vector notation is used, $(xXo, 0, 7) will denote that solution (x, y) of
Eq. (9.6) which satisfies (x(0), y(0)) = Xo.
Examples were given above which show that a limit set of a solution of this system
can consist of an isolated critical point, an isolated closed path, or critical points
joined by nonclosed paths. The aim here is to show that a limit set for a solution of
the system (9.6) contains a closed path if and only if it does not contain a critical
point.
Theorem 9.10. Let p denote a noncritical point on the path C of a solution of the
system (9.6) and let N denote the normal line to C at p. Any path other than C which
crosses N sufficiently near p must cross in the same direction as C does.
 Autonomous systems 9.6
284

Proof. First observe that the slope of each path may be found at each noncritical
point in & by taking the ratio

dy SCGY)s
de O(X..¥)
Let p have coordinates (u,v). We may assume without loss of generality that the
normal line N is not horizontal. Since the function f/g is continuous at p, there is
a neighborhood of p on which |(f(x, y)/g(x, »)) — (fu, v)/g(u, v))| < 7/8. This
means that the tangent vectors to any two solution paths in the neighborhood could
make an angle of at most 45° (see Fig. 9.14). ||

Fig. 9.14 The angle between the indicated vectors is at most 45°.

Definition. Let p be anoncritical

point for the system (9.6). A path neighborhood
ofp is a neighborhood ofp whose boundary consists of precisely two segments of solution
paths and two normal line segments to the path through p (Fig. 9.15). The path through
p is called the axis of the path neighborhood. The two normal line segments in the
boundary of the path neighborhood are called its initial and terminal bases (relative to
the direction of its axis).

Theorem 9.11. Let C denote a closed path and let p denote an arbitrary point of C.
There is a path neighborhood which covers all of C except p, and the bases of the neigh-
borhood lie in the normal line N to C through p.

a
Figure 9.15
 Limit sets in the plane 285

Proof. Observe first that, by virtue of the continuity of the solution x = (Xo, fo, 1),
one can make the terminal bases of a path neighborhood as short as desired by
making the initial base sufficiently short. Since C is a closed and bounded set, it can
be covered with a finite number of path neighborhoods. The covering can be so
constructed that the initial portion of one path neighborhood is overlapped by the
terminal portion of its predecessor and so that the terminal base B of at least one path
neighborhood lies in N (Fig. 9.16a). The desired neighborhood is the set of Open
path segments which initiate on N and terminate in B (Fig. 9.16b). ||

(a) (b)
Figure 9.16

The neighborhood in Theorem 9.11 is called an open path ring. Because of the
continuity of@ with respect to Xo, the existence of an open path ring has the following
consequence: if a nonclosed path I initiates at a point q of N, then I can be made
to reintersect N arbitrarily close to p by bringing q sufficiently close to p.
Theorem 9.12. If the w-limit set Q for a solution of the system (9.6) contains a
closed path C, thenQ = C.
Proof. Suppose for contradiction that there is a point x9 in 2 — C. Since C isa
Jordan curve, Xo is either interior to C or exterior to C. For definiteness, assume Xo
is exterior to C.
There must exist at least one normal line N at a point p on C which contains the
point x9 [why?]. Choose a point q ¥ p on N so near p that the path I through q
is interior to an open path ring with bases in N. Then I reintersects N at point r
in the terminal base of the ring (Fig. 9.17). The arc of I from q to r together with
the segment of N from q tor forms a Jordan curve which separates C and x». Hence
Q is not connected, and Theorem 9.9 is contradicted. Therefore, 2 = C. ||
It is clear that if the w-limit set of a solution consists of precisely one closed path
C, then the w-limit set contains no critical point. The converse of this statement,
which in various forms is called the Poincaré-Bendixson theorem, is useful in the
study of nonlinear oscillations. To establish it, we need two preliminary results.
286 Autonomous systems 9.6

Figure 9.17

Theorem 9.13. Suppose that q is an w-limit point of a solution path C for the
system (9.6) but that it is not a critical point. There is a path neighborhood R of q
such that:
i) The axis of R consists entirely of w-limit points of C;
ii) R contains no w-limit points of C other than those on its axis.
Proof. Assertion (i) is an immediate consequence of Theorem 9.9 since an w-limit
set consists of whole solution paths.
Suppose assertion (ii) is not true. Then there exists a path neighborhood R
about q such that the normal WN to the axis at q contains both an w-limit point r of C
and a point Xo of C between q and r (Fig. 9.18). The path C can be described by the

Figure 9.18

solution x = $(Xo, 0, f), and there is a time ¢; > 0 such that x; = (Xo, 0, f,) is
on N and in R since q is an w-limit point.
It is not possible that x; = xo, for otherwise C is a closed path and both q andr
could not be w-limit points of it (Fig. 9.19a).
If x; is located on the same side of xo as q, then (Xo, 0, f) is bounded away
from r for ¢ > ft; and r could not be an w-limit point for @ (Fig. 9.19b).
If x, is located on the same side of xo as r, then (xo, 0, t) is bounded away
from q for ¢ > ¢; and q could not be an w-limit point for $ (Fig. 9:19¢):
A contradiction results in every case. Thus assertion (ii) is established. ||
 Limit sets in the plane 287

r
r

XxX]
4

(b) (c)

Figure 9.19

Theorem 9.14. If a path C of the system (9.6) contains one of its own w-limit
points, then C is either a critical point or a closed path.

Proof. Suppose C contains an w-limit point q of itself. If q is a critical point, then

C is precisely q by the uniqueness of solutions to initial value problems. If q is not a
critical point, there is a time tf; > 0 such that x; = ¢(q, 0, £) is on the normal line
N to C at q (Fig. 9.20). If x; #q, then ¢(q,0, t) is bounded away from q for
t > t, and q could not be an w-limit point of C. ||
N N

x1

Figure 9.20

Theorem 9.15 (Poincaré-Bendixson). If the w-limit set Q of a solution of the system

(9.6) is nonempty and contains no critical point, then Q consists of a closed path C.

Proof. Let p be inQ and consider the path C of the solution x = ¢(p, 0, /). Certainly
C is in @. We shall show that C is a closed path. Theorem 9.12 then implies that
Q = C.
Let q denote an w-limit point of C and note that q, being in , is not a critical
point. By Theorem 9.13, there is a path neighborhood of q which contains no «-limit
point of C except those on its axis. But then the axis, which contains q, must be a
segment of C, and it follows from Theorem 9.14 that C is a closed path. |
288 Autonomous systems 9.6

Example 1. The system

= ay
V2)
ye ye
has a (proper) periodic solution. To see this, write the system in form
/
x = x(1 — y? — x’),

Yaa C= a. =)
and consider the compact region G between the concentric closed curves

x ty= s “and x ye =3:

Typical tangent vectors to the solution paths crossing the boundary of G have been
sketched in Fig. 9.21.
A solution path I initiating at a point p of G can never leave G since any solution
path crossing the boundary of G, crosses inward. The solution path I must have an
w-limit point since it is bounded, and this w-limit point must lie in G since G is a
closed set. There is no critical point in G. In fact, the only critical point for the system
is the origin of coordinates. By the Poincaré-Bendixson theorem, the w-limit set of T
is a closed path, and the associated solutions are periodic by Theorem 9.5. ||
In view of the discussion above, one can make the following qualitative assertions
about a nonempty limit set of the system (9.6).
1) If it contains a closed path, it is precisely that path.
i1) If it contains a critical point, it does not contain a closed path.
ili) If it does not contain a critical point, it consists of precisely one closed path.
We have seen by example that a limit set can consist of precisely one critical point

ote
or it can consist of two critical points approached by nonclosed paths. In general, a

Se RS
Figure 9.21
 Limit sets in the plane 289

limit set of the system (9.6) which does not contain a closed path, consists of a finite
number of critical points to which adhere at most a countable infinity of nonclosed
paths.
The method of proof for Theorems 9.13 and 9.14 can be applied to show that
paths in the plane which approach limit cycles do so by spiraling. This observation
leads to a classification of limit cycles. A limit cycle is called
1) Stable (unstable) if it is the w-limit set (a-limit set) of a path in its interior and
a path in its exterior, and
li) Semistable if it is the w-limit (a-limit) set of a path in its exterior and the
a-limit (w-limit) set in its interior. See Fig. 9.22(a), (b), (c).

(a) (b) (c)

Figure 9,22

Bendixson’s negative criterion (Theorem 9.7) is an important test for the non-
existence of periodic solutions. We shall conclude the discussion here with another
nonexistence test: A closed path for the system (9.6) must contain a critical point
in its interior. Thus, for example, the second order equation x’’ = f(x, x’) cannot
have a real periodic solution unless it also has a constant solution. Before proving
the result, let us consider an example illustrating a feature of the proof.
Example 2. The differential system

x = EN (Xs) y’ = Xe Vik Xa)

where f(x, y) = (x? + y”)? sin (x? + y”)~' for (x, y) ¥ (0,0) and f(0, 0) = 0,
takes the form

r(t) = (1) sin (ais): © a(t) = |

in polar coordinates. Thus every circle x? + y? = 1/nz, n a positive integer, is a
closed path. The origin is the only critical point. Each closed path is the a-limit set
of the spirals (Fig. 9.23) immediately inside it and the w-limit set of the spirals
immediately outside it.
The example illustrates that a closed path may have a nested sequence of closed
paths in its interior. ||
 9.6
290 Autonomous systems

Figure 9.23

Theorem 9.16. A closed path of the system (9.6) must have a critical point in its
interior.

Proof. Let Co be a closed path of Eq. (9.6) and suppose the assertion is false. Let
I be a path inside Cp. Then I has a nonempty a-limit set and a nonempty w-limit set.
The path Cp cannot serve as both limit sets. By the Poincaré-Bendixson theorem
there is another closed path C, in the interior of Co. If C, is any closed path in Co,
let B, denote the union of C, with its interior and let S denote the collection con-
sisting of all the B,’s. A chain in S is a subcollection P of the B,’s which is ordered
by set inclusion; that is, if both B, and Bg are in P, then either B, C Bg or By D Bg.
A chain Q in Sis called maximal if there is no other chain in S which properly contains
all the elements of Q. It is a fundamental assumption of analysis (called the
maximality principle) that there exists a maximal chain in S. Let Q be such a maximal
chain. Since the elements of Q are nested, closed and bounded, it follows from the
nested set theorem that there is a point p, common to all the Q,’s in Q. Then the
path [ through p has nonempty a- and w-limit sets 4 and Q which are contained in
every Q, in Q. Neither A nor Q contains a critical point since By does not contain
one. Thus, both A and Q are closed paths in the interiors of all the Q,’s. This con-
tradicts the maximality of the chain Q. ||

Example 3. The equation x’” + x? + 1 = 0 has no real periodic solution since it

has no real constant solution. _ ||

Example 4. The system x’ = x? + y? — 1, y’ = x — y? could have no periodic

solution with a path contained wholly in the interior of the disc x? + y? < 1. ||
 Critical points in the plane 291

EXERCISE 9.6
1. Consider the system of Example 1, Section 9.5. Construct a neighborhood of the point
(0, —1) on which the tangent vectors to any two solution paths make an angle of at most 1°.
2. Consider the system of Example 1, Section 9.5. Construct an open path ring with its
bases centered at (0, —1) on the negative y-axis. Find the ratio of the lengths of its initial
and terminal bases.
3. Let C denote a simple, closed differentiable curve in the plane and let p denote a point
exterior to C. Show that at least one normal line to C passes through p.
4. Discuss the existence of periodic solutions for the following systems of equations.
a) = ey 1, oy = 2x b) x = xe y-y = yer 4 x:
5. Let a(x, y) = ci and h(x, y) = ce define two continuously differentiable Jordan curves
C,; and C2, respectively, and suppose that C2 is in the interior of C;. Assume that

For 9)5.
7)
ae 9)+ BG 9)Fale »)<0
at every point on C; and

FG, y)5 Bl,¥)+ BCsVF. BOsy)> 0

at every point of Co. Show that the system x’ = f(x,y), y’ = g(x, y) has at least one
periodic solution.
6. Discuss the existence of periodic solutions for the system
x =1—x? 4 y?, yo
= 2x),
7. Refer to Problem 4 of Exercise 9.5. Could a nonperiodic orbit of the system (T) contain
one of its own w-limit points? Compare with Theorem 9.14.
8. Let f and g be continuously differentiable with g bounded. Give, in terms of f and g,
two conditions which are necessary for the existence of a periodic solution for

ote lx! ict e(x)o= M,

where M is a given constant.
9. Does the system x’ = x? + y? — 1, y’ = 2xy have (proper) periodic solutions?
10. Apply the result of Problem 5 to the equation x” + (x? — 1I)x’ + x34 x = 0.

9.7 CRITICAL POINTS IN THE PLANE

In Section 9.6, limit sets for planar, autonomous systems

x’ = f(x, y), (97a)

y’ = g(x,y)
were discussed. Two specific types of limit sets were of particular importance: limit
cycles and critical points. If a closed path is a limit cycle, i.e., the limit set of a path
other than itself, then other paths must approach it spirally. If a critical point is the
limit set of a path other than itself, then other paths can approach it in a variety of
292 Autonomous systems 9.7

ways. In this section, there are listed the various modes of approach. We assume
that the origin is an isolated critical point for the system (9.7a) and that f and g are
continuous and lipschitzian on a domain & in ®” containing the origin.
If every neighborhood of the origin contains a closed path, then it is called a
rotation point. The origin in Fig. 9.23 and the point p in Fig. 9.5 are rotation points.
A point (such as p above) for which there is a neighborhood containing only periodic
solutions is a special type of rotation point called a center (see Fig. 9.24a).
If there is a neighborhood WN of the origin with the property that every solution
initiating in N approaches the origin as t—> +» (tf — —w), then the origin is called
an attractor (repeller). An attractor with the property that all nearby solutions spiral
about it is called a focus (Fig. 9.24b). If the origin is an attractor and if every line
segment through the origin is tangent to some path as it approaches the origin, then
it is called a proper node (Fig. 9.24c). If the origin is an attractor and if there is one
line segment through the origin which is tangent at the origin to all approaching paths,
then the origin is called an improper node (Fig. 9.24d, e).
The origin in Fig. 9.24(e) is an example of a saddle point. A saddle point is a
critical point which is approached by only a finite number of other paths ast > +a.
Note that a saddle point is neither a rotation point nor an attractor (repeller).
The critical points described above (center, focus, nodes, saddle) are called
elementary critical points. If the origin is an attractor (repeller), these points are
further designated as stable (unstable).

y
ccs
ee

(a) A center (c) A proper node

FYING
(d) An improper node (e) An improper node (f) A saddle point

Figure 9.24
 Critical points in the plane 293

Elementary critical points are of special interest because the system (9.7a) possesses
such critical points if it is linear.
Theorem 9.17. Consider the system

(9.7b)
Gh i fi 4|
a : . ,
where A = E ‘|is a real, constant, nonsingular matrix and let \, and \» denote

the eigenvalues of A.
i) If \; and Xz are complex conjugates, say \; = a + iB with B ¥ 0, then the
origin is a center if « = 0 and a stable (unstable) focus if a < 0 (a > 0).
ii) If the matrix A does not have two linearly independent eigenvectors, then the
origin is a stable (unstable) improper node if the common value \ of \, and \»
is negative (positive).
iii) If the matrix A has real eigenvalues \, and \» and two corresponding linearly
independent eigenvectors, then: the origin is a stable (unstable) proper node if
Ay = Ag < OQ y = Ag > O); the origin is a stable (unstable) improper node
if \, and Xz are negative (positive) and unequal; and the origin is a saddle point
Tee)
Proof. (i) The eigenvectors p; and py» will be complex conjugates, say po = py,
since hy = 4; = a — i8. The matrix P = $[p; + po, —ip: + ipe] is real and
nonsingular, and the system takes the form

HAE: IB 79
under the change of variables [x, y]” = P[u,v]’. The nontrivial solution paths of
the system (9.7c) can be quickly described with polar coordinates r? = u? + v?
and @ = tan! (v/u), for the system becomes (r”)’ = 2ar”, 6’ = —8, where r(t) =
roe*’ and 6(t) = 89 — Bt. If a = 0, the paths are circles. If a ¥ 0, the paths are
spirals. In the xy-plane, the paths of the system (9.7b) have the same general features.
The curves are distorted, however. To see this, notice that the images of the u-axis
and v-axis under the transformation [x, y]" = Pfu, v]” are generally skewed lines
in the xy-plane (Fig. 9.25).

Hlecalld
ha
y 0

yx >

oe c@ nh fo) se 0)

Figure 9.25
 9.7
294 Autonomous systems

ii) Under a linear variable change [x, y]” = Plu, v]", the system (9.7b) becomes
a Jordan system

BiG
alba oleh: | 619
9.7d
The solution paths in the wv-plane are the images of the curves

u(t) = (c1 + cot)e, v(t) = cze™

which are depicted in Fig. 9.26 for \ > 0. The tangents to the paths tend to the
horizontal as (u, v) > 0 since
dv uv’ dv Co 0
di it MU lo (Gece
as (u,v) 0. Figure 9.26 depicts the paths of the system (9.7d) in the uv-plane
v
y

Figure 9.26

and the corresponding paths of the system (9.7b) in the xy-plane. Notice that the
line of common tangency is the u-axis in the wv-plane and is along the line of eigen-
vectors of A in the xy-plane.
iii) If A has two linearly independent eigenvectors, then proceeding as before,
we put the system (9.7b) into the Jordan form

u |’ = 4 0 u
A OEE [Link] Hi
and find that u(t) = cye*, v(t) = coe2'. The description of the solution paths is
now straight-forward, and the details are left to the exercises. ||
 Critical points in the plane 295

A critical point of a nonlinear system may, but does not have to, be an elementary
critical point. Commonly considered are the systems of the form

x’ = f(x, y) = ax + by + P(x, y),

(9.7e)
y’ = g(x,y) = cx + dy + QG,y),
where ad — be # 0. Then if P(x, y) and Q(x, y) have sufficiently small values near
the origin it is plausible that the paths of the system (9.7e) will resemble the paths of
the linear system (9.7b) near the origin. This is sometimes the case. For completeness,
we State a pertinent result as a theorem without proof.
Theorem 9.18. Suppose the system (9.7e) satisfies the hypothesis ad — bc ¥ 0
and
P(x, y)| + |OG y)I _ 9
Izi+lyl+0 [lx] + |y|]}i*<
for some € > 0. Then the origin is a node, focus, or saddle point for the system (9.7e)
if it is a node, focus, or saddle point, respectively, for the linear approximation (9.7b).
If the origin is a center for (9.7b), it may be either a center or focus for (9.7e).
Example 1. The origin is an unstable focus for the system

x =xty4tx?4+y?,
VY =x yxy. © ||
Example 2. The origin is an unstable node for the system
x’ = sinx + e” — 1,
y = e+ cos xy — 2,
for by Taylor’s theorem
x’ = x + P(x, y),
y =y+
QC, y),
where
x? xy?

EO) = Azeri D ae

and
2 xy?

OF ite se el
Example 3. The origin is a center for the linear approximation to the system

eer ee ey (9.78)
Vee Vie oy)
The origin, however, is an unstable focus for the nonlinear system. To see this,
observe that its solution paths are described in polar coordinates by the equations
(r2y = —2r*, 6’ = 1. Thus r? = 73/(1 + 2tro) and @ = 0) + 1#. ||
 9.7
296 Autonomous systems

the
In the case of a nonlinear system, to which Theorem 9.18 does not apply,
behavior of solution paths near an isolated critical point can be much more com-
plicated. The behavior has, however, been thoroughly studied. Suppose, for example,
a Jordan curve C is drawn as indicated by the dashed curve in Fig. 9.27. Then,
inside C, the solution paths or

xi = Q(e— Den y = =2x 1)? = 267: y*)(x — 1) + 8x (9.78)

are as indicated.

Figure 9.27

The heavy lines of the figure depict certain path segments which connect the
curve C to the critical point. Such a path segment is called a base path. The interior
of C is thus cut into nonlinear sectors I, H, Il], IV, V, VI. The sectors I, II, VI are
characterized by the fact that they contain only two base paths (oppositely sensed)
and no other path in the sector approaches the critical point. Such a sector is called a
hyperbolic sector. The sectors III and V consist entirely of base solutions. A sector
of this type is called parabolic. The sector IV is characterized by the fact that its
boundary consists of one path, and all the paths inside the sector approach and
recede from the critical point. This type of sector is called elliptic.
If an arbitrary isolated critical point of a planar autonomous system is not a
rotation point, then it can be shown that the surrounding paths form sectors about
the point. There are a finite number of elliptic and hyperbolic sectors. The remaining
sectors are parabolic. If the number of elliptic sectors is even (odd), the number of
hyperbolic sectors is even (odd).
 Phototropic platyhelminthes 297

9.8 PHOTOTROPIC PLATYHELMINTHES

An example of an autonomous differential system which illustrates much of the

theory in this chapter occurs in the problem of the phototropic platyhelminthes. A
scientist has a sample of a liquid which contains several species of lightseeking flat-
worms. He wishes to separate the species so that they may be studied. Each species
is known to swim at a different speed. To isolate and extract the worms of one
particular species, with characteristic swimming speed v, he places the fluid in a
cylindrical glass jar of radius R. The jar is made to rotate near a light source with a
constant angular speed w > v/R as shown in Fig. 9.28. The flatworms swim toward
the light against the rotation of the fluid, and the scientist hopes that they will cluster
at a point in the tank so that they can be extracted by dipping the fluid at that point.
The question of interest is whether such a point exists and, if so, where is it? If
coordinates are constructed as in Fig. 9.29, then the paths of the flatworms are the
solution paths of the autonomous differential system
R-x
x’ / = —wy+v
ORE SP ae
y’
= wx
y (9.8)
AR —xP+ ye

_ Light
Wiz-

Ld Turntable W

2D Ww

Figure 9.28

y
A

Figure 9.29
 9.8
298 Autonomous systems

We shall analyze these solution paths in detail as an application of the theory presented
in this chapter.
Phase space for the system consists of the xy-plane with the point (R, 0) deleted.
Consider an arbitrary nonconstant solution vector ¢(t) = (x(t), y(t)) initiating
in the interior G of the circle x? + y2 = R?. If r?(t) = x?() + y7(0, it is easy to
check that r’(t) < O if and only if the point ¢(t) lies outside the circle C with equation
y2 + (x — (R/2)? = R*/4. See Fig. 9.30. Thus $(f) cannot leave G.
y

>r
(R, 0)

Figure 9.30

It is not possible that o(t) — (R,0) as t— +, but this might happen as ¢

approaches some finite time 7+. In this case, the w-limit set 2 of @ will be empty.
Let us assume then that ¢(t) does not approach (R, 0).
Since ¢(¢) remains in G, its w-limit set is not empty by Theorem 9.9. The diver-
gence of the coefficient functions is —v/d, where d = ./(x — R)2 + y2. By Ben-
dixson’s negative criterion (Theorem 9.7), there is no periodic solution in G. Thus
Q contains a critical point by Theorem 9.15.
The system (9.8) has but one critical point (X, p). It lies at the intersection of the
line x = 1/R(v/w)? and the circle C in the first quadrant and has coordinates

We should like to be able to assert that 2 contains only the point (%, $). This will
be the case if (%, ) is a nonrotation point to which there does not adhere an elliptic
sector.
Since the coefficient functions have continuous partial derivatives of all orders
at (x, }), Theorem 9.18 may be used to study the solution paths in the vicinity of the
critical point. The matrix of the linear approximation is the Jacobian of the co-
efficient functions evaluated at the point (X, }) namely
 The equations of van der Pol and Liénard 299

Se Lag yr = R)
ii ae an ca

came ae bo
Vit) (x — R)?

The characteristic polynomial of J is \? + v\/d + w?, where dQo> is the value of d

at (x, }). The eigenvalues of J are

wysy R2a? - v\2

r — 4/5— 4 |. —— ni 2 — _ .

2d a D a ss w

By Theorem 9.18, (%, ¥) will be either a stable node or a stable focus.

Summarizing, any solution initiating in G will either approach the point (R, 0)
in a finite time or else approach the point (x, $) as t— +~, and all paths passing
sufficiently near (X, #) must behave in the latter manner.

EXERCISE 9.8

1. Derive the differential equations (9.8).

In Problems 2 through 9, let @() = (x(d), y(A) denote a solution of the system (9.8).
Let r?(t) = x?(t) + y*(A and (1) = tan—! (p()/x()).
2. Show that r’(¢) < 0 if and only if (Z) is outside the circle C with equation

y? + (x — R/2)? = R2/4.
3. Show that $(f) cannot approach (R, 0) as t— +0.
4. Set up polar coordinates with the point (R, 0) as the pole and study the behavior of
solution paths near (R, 0).
5. Compute the divergence of the coefficient functions for (9.8).
6. Compute the coordinates of the critical point for (9.8).
7. Compute the matrix J and its eigenvalues.
8. On the basis of physical intuition would you expect the critical point for (9.8) to be a
node or a focus for large values of w? Check your conjecture by examining the eigenvalues
of J.
9, Suppose the speed w of the turntable could be adjusted so that w = v/R. Could the
flatworms with characteristic speed v be separated from the other flatworms in the tank?
10. Describe the physical operation of the separation device for very large values of angular
speed w.

9.9 THE EQUATIONS OF VAN DER POL AND LIENARD

In Section 1.10, we studied a nonlinear electronic oscillator circuit and showed that
its operation is described by the solutions of van der Pol’s equation
x" + 1 — x”)x' + x = 0, ee 0. (9.9a)
300 Autonomous systems : 9.9

We asserted that this equation has a unique periodic solution path towards which
all other nontrivial paths tend with increasing time. We shall now establish this,
not only for Eq. (9.9a), but also for the more general equation

x" + f(x)x’ +x =0 (9.9b)

of Liénard.

Theorem 9.19. Let the coefficient f(x) in Liénard’s equation (9.9b) denote a con-
tinuously differentiable, even function defined for ~n <x < +m. Suppose there
isana > O such that f(x)(a — \x|) < O for x # aand such that [Rios a6, dx = +o.
Then Eq. (9.9b) has a unique (nontrivial) periodic orbit toward which every nonconstant
solution tends as t-> +a.

Proof. Convert Eq. (9.9b) to the equivalent system

x =y — Flr), y = -x, (9.9c)

where F(x) = fo f(s) ds.

In order to set up a context in which to study the system efficiently, let us make
the following observations:
i) The origin is the only critical point.
ii) One can show by the method of Example 2, Section 8.9, that all solutions
exist for —-»m <t< +m.
ili) The slope of each noncritical path is given by the ratio

Lf pee Ze, (9.9d)

dx F(x)-—y
Thus a noncritical path can have a horizontal tangent at a point (x, }) if
and only if x = 0. Such a path can have a vertical tangent at a point (X, §)
if and only if the point is on the curve C defined by the equation y = F(x).
iv) The function F is odd, has a negative minimum at x = a, has a unique positive
zero b > a, increases monotonically for x > a, and approaches + as x
does. See Fig. 9.31.

Having made these preliminary observations, let us now proceed with the proof.
Let C denote that part of the curve y = F(x) which lies in the right half-plane and
let U and L denote the portions of the right half-plane which lie above and below C
respectively.
It follows from Eq. (9.9d), that any path in U has a negative slope and any path
in L has a positive slope. Thus, if (0, yo) is an arbitrary point on the positive y-axis,
a particle following a path segment I'(yo) through y) must enter U and then move
downward and to the right. If I'(yo) did not intersect C, then a particle following it
would have to monotonically approach a critical point on C or in U. Since there is
no such critical point, I'(y) crosses C (with a vertical tangent) at some point (x2, yo).
 The equations of yan der Pol and Liénard 301

Figure 9.31

Conversely, if one chooses a point (x2, y2) on C and follows the path through
(X2, ¥2) counterclockwise into U, then he must reach the positive y-axis at some
point (0, yo). To see this, let (¢,~) be the solution of Eq. (9.9c) which satisfies
$(0) = x2, ¥(0) = yo. Suppose that, for —x < t < 0, (4(0), ¥(d)) is in U. Then
$(t) is bounded below by zero. Thus lim, ,_. (ft) exists. Since (4(t), ¥(1)) cannot
approach the critical point (0, 0) from inside U, lim;_,_. Y(t) = +~«. This implies
that the tangent to the curve traced by (¢(f), ¥(t)) becomes vertical as t— —~.
Equation (9.9d), on the other hand, implies that the tangent becomes horizontal.
Since this is a contradiction, the curve must cross the positive y-axis at some point
(0, Vo).
Together, the assertions of the last two paragraphs imply that the points of C
and the positive y-axis are put into a one-to-one correspondence by the flow mapping.
By a dual argument, one shows that the points of C and the negative y-axis are in a
similar correspondence. Thus given yo > 0, there is a path segment I'(j9) in the
right half-plane which initiates at (0, yo), crosses C at a point (x2, y2), and intersects
the negative y-axis at a point (0, —y4). Moreover, any two of the quantities yo,
2, ¥4 are monotone increasing functions of the third, and all three approach zero or
become infinite together.
The proof of the theorem hinges upon our comparing the distances from the
origin to each of the points (0, yp) and (0, —y4). A convenient method of comparison
consists of observing the algebraic sign of the function J(vo) = 407 — ya). The
attendant difficulty is that the functional dependence of y4 on Yo is not explicitly
known. Note, however, that /(j9) is given by a line integral

Io) = | (x dx + ydy).
T(yo)

Let j, denote the unique value of yo for which x2 = 6. For 0 < yo < Jo,
the path I'(y’9) may be parameterized with the variable x, say x = Y(y), —Y4 <y<yo-
302 Autonomous systems 9.9

Then

ox) = — | HO) dy> 0.

For yo > Jo, P(¥o) intersects the line x = b at a point (6, y;) in U and at a
point (b, y3) in L. This permits the parameterization of (yo) in three steps: y = g(x)
for0 < x < bandy > 0;x = AQ) forys < y < yi;and y = k(x)forO0 <x <b
and y < 0. Then
b Y1
xF(x) b
x|FOO|
Kyo) = i FeV Gee [ TARO) y a i Fe—
3 e)k@)
The first and third integrals are positive and strictly decrease as yo increases. For
y3 <y < yi, hy) = x > b. Thus F(h(y)) > 0. Since F(x) increases strictly to
+o as x > b does, the second integral is positive and increases strictly to +0 as
Yo 2 Vo does.
Thus J(yo) > 0 for 0 < yo < Po. For yo > Po, IKo) is strictly decreasing and
approaches —« as yp ~ +o. By the intermediate value theorem there is unique
value y = c such that/J(y>) > [Link] yo > c [Wo) = 0 for yo. = cy and 17a) 0
for yore, (Fig. 9:32);

>x

Figure 9.32

If (¢, y) is a solution for the system (9.9c), it is easy to check that (—¢, —yv)
also is. The paths are therefore symmetric with respect to the origin. This implies
that I'(c) is a segment of a closed path which is the w-limit set of every noncritical
path. ||
 The equations of van der Pol and Liénard 303

EXERCISE 9.9
1. Generalize Theorem 9.19 so that it applies to an equation of the form

ee A fi(X)X 2(x). = 0;
2. On the basis of physical intuition, for which of the following equations might one
plausibly conjecture the existence of a periodic solution?
a) x" 4- |x|x’ + x = 0.
b) x” + (|x| — 1)x’ + x? = 0

ert (21) vp emo

ee ee )) ee eX Cm)

3. Two identical conveyor belts with common speed v > 0 feed into a slot as indicated in
Fig. 9.33. A square, flat block with edgelength ¢ and mass m moves on top of the belts.

ae Figure 9.33

The magnitude of the force of friction between the block and either belt is proportional to
the product of the speed of the block relative to the belt and the area of the block in contact
with the belt.
a) Suppose the block is oriented so that an edge is parallel to the belts. Describe its
motion.
b) Suppose the block is oriented so that a diagonal is parallel to the edge of the belt.
Describe its motion.
CHAPTER 10

STABILITY

10.1 DEFINITIONS

In order to work within a fixed context, we shall consider in this chapter only the
equation
x = [0 7) (S)

with f(x, £) continuous and lipschitzian on a cylindrical region D = @ X [8, +o],

where @ is a domain in ®”. We allow the possibility that (S) is autonomous. In this
case, we write
x tat Cx) (A)
when it is desired to call specific attention to autonomy. Attention is restricted here
to questions associated with the behavior of solutions for (S) as t > +0. Analogous
considerations can be made for the case t— —o.
The term “stable” was applied to certain critical points and limit cycles in
Chapter 9. The use of this word in the theory of differential equations stems from
language used to discuss physical situations. If, for example, the grid voltage V,
in the circuit of Fig. 1.9 of Section 1.10 is set equal to zero, then V,(t) = 0 is a solu-
tion of Eq. (1.10b). The plate current then maintains the constant value J,(t) = y(0),
and the circuit operates statically, its operation corresponding to the critical point
x = 0, x’ = 0 of van der Pol’s equation. Physically, this mode of operation cannot
be maintained, for stray electrons will impinge upon the grid and V, will vary slightly.
Then it will begin to fluctuate with increasing amplitudes; that is, the corresponding
solution path of van der Pol’s equation will spiral outward from the origin. Thus
the static operation of the circuit is, in a physical sense, unstable.
On the other hand, the circuit is intended to be a source of alternating current
having a definite frequency. Whenever activated, say by a telegraph key in series
with the cathode, the circuit is required to oscillate. If the telegraph key is pressed

304
 Definitions 305

at a random time, the voltage V, may correspond to any one of the spiral solution
paths in the phase space for van der Pol’s equation. In a well-constructed circuit,
however, the spiral paths wind about the limit cycle very rapidly. Thus, from a
physical point of view, V, appears always to correspond to the limit cycle.
In this sense the oscillator is stable.
The stability phenomena discussed above illustrate the mathematical terms
“stability (instability) of equilibrium” and “orbital stability” (a limit cycle is some-
times called a closed orbit because of applications in celestial mechanics). These are
special cases of a more general type of stability: the stability of an invariant set.
Definition. A set M in @ is called (positively) invariant with respect to (S) x’ =
f(x, £) if each solution $ of (S) has the property that $(to) in M implies $(t) is in M
Or ON Lato = tb.
Notice that any solution path of an autonomous system is an invariant set. In
the plane, a domain bounded by paths of an autonomous system is also an invariant
set. Periodic orbits for autonomous systems and critical points for (S) are invariant
sets of particular interest. They are closed and bounded. Recall that any limit set
for a bounded solution of an autonomous system is a closed, bounded, invariant set
by Theorem 9.9.
If M is a set and x a point in ®”, we shall take the distance from x to M to be
dis (x, M) = inf {|x — y|: y isin M}.
Definition. Let M be invariant with respect to (S) x’ = f(x, t). Then M is called
i) Stable if, given to > B and € > 0, there is a 6 (which may depend on both
to and €) such that whenever a solution @ of (S) satisfies dis (¢(to, M) < 6,
then $(t) exists and dis (¢(t), M) < € for tp) <t< +m. (Fig. 10./a.)
ii) Asymptotically stable if it is stable and if, given t; = 8, there exists a 6, such
that lim;_,4. dis (@(4), M) = 0 whenever |$(t1)| < 41.
In the case of an autonomous system, one sometimes attaches the phrase “‘in the
sense of Poincaré” to the definitions in (1) and (ii).

Fig. 10.1 (a) A stable invariant set. (b) An asymptotically stable invariant set.
 10.1
306 Stability

of
The notion of stability of an invariant set is occasionally too general a notion
portion of the positive x-axis to the right of
stability. Note, for example, that the
(1,0) is an invariant set M for the system x’ = x, y = yy (Pig 02).

><

Figure 10.2

As an invariant set, M is asymptotically stable. Consider, however, the solution

x(t) = e’, y(t) = 0 which has M for its path. A different solution with path I ini-
tiating at a point (a, b) in the right half plane at ¢ = O will have the form ¢(t) = ae’,
V(t) = be’. The difference |x(t) — ¢(f)| + |y() — ¥()| = |1 — ale’ + |ble~* be-
comes unbounded as ¢ > +o unless a = 1.
Assume, specifically, that 0 << a< 1 and b> 0. We interpret the solutions
(¢, ¥) and (x, y) in terms of the motions of points p and q which are located at the
points (a, b) and (1, 0), respectively, at time t = 0. The points p and q follow the
paths T' and M, respectively. But p moves more slowly than q, and the distance
between them becomes infinite as t does. To describe such behavior, the notion of
stability of solution functions is introduced.

Definition. Let y be a solution of (S) x’ = f(x, t) which exists forB <t< +a.
Then wy is called
i) Stable if, given to > B and € > 0, there is a 6 such that whenever a solution
¢ of (S) satisfies |\@(to) — W(to)| < 6, then $(t) exists and satisfies

I¢@) —¥@)| < efor tg ee

(Fig. 10.3a).
ii) Asymptotically stable if it is stable and, given t, > 8, there exists a 6, such
that |(t) — ¥(t)| > 0 as t— +x whenever ¢ is a solution of (S) that
satisfies \p(t1) — W(ty)| < 41.
 Definitions 307

ii) Uniformly stable if, given € > 0, there is a 6 > 0 such that whenever
a solu-
tion @ of (S) satisfies |(to) — W(to)| < 6 for any to = B, then $(t)
exists
and satisfies |(t) — ¥(t)| < € for to <t < +m (Fig. 10.3b).
iv) Uniformly asymptotically stable if it is uniformly stable and if there is
a 8,
with the following property: Given €, > 0, there exists a T > 0 such that
lo(t) — ¥(2)| < €, for all t > T + t1 whenever ¢ is a solution of (S) which
satisfies |p(t1) — ¥(t1)| < 61 for any ty > B.

>e

Pe

(b)

Figure 10.3

Stability as defined in items (i) and (ii) is sometimes designated as being “‘in the
sense of Lyapunov.” In case (S) is an autonomous equation and wy is a constant
solution, items (iii) and (iv) are equivalent to items (i) and (il) respectively. Note
that uniform asymptotic stability of a solution implies its asymptotic stability which,
in turn, implies its stability.
If a solution or invariant set is stable and if every solution approaches it as
t— + a, then it is called globally asymptotically stable. If an invariant set or solu-
tion is not stable, then it is called unstable. It is worth pointing out that a solution y
 10.1
308 Stability

of (S) may be unstable even though every other solution approaches it in the sense
that |@(1) — ¥())| 2 [Link] t+. Such a solution is sometimes called quasi-
asymptotically stable (Fig. 10.4). See Problem 10, Exercise. 10.1 for a specific example
of this phenomenon.

><

SS
A

Fig. 10.4 A phase portrait for a system with a quasi-asymptotically stable solution x = 0,
y = 0. An analytical example is given in Problem 10 below.

The trajectory of a constant solution of (S) is a critical point. The stability of a

constant solution is thus equivalent to the stability (as an invariant set) of its path,
and the two notions may be identified. The unified concept is then called the stability
of equilibrium, the terminology stemming from classical dynamics.
The study of the stability of a given solution y of (S) x’ = f(x, #) is reduced to
the study of the stability of a critical point as follows. Set x = y + y(Z) in (S) to
obtain
y’ = gy, 4), (10. 1a)
where g(y, 2) = f(y + ¥(d, t) — f(¥(d, t). Then y = 0 is a solution of Eq. (10.1a),
and it will be stable or unstable in the same manner that wy is for (S). If, further, f
has first partial derivatives on @, then one may write g(y, t) = J(f)y + h(y, 1), where
J(2) is the Jacobian matrix D; f (yd), t) and h(y, t) = f(y + va), t) — f(v(a), t) —
J(ay. If the partial derivatives D;f;(x, ft) are continuous on @, it follows from the
mean value theorem that |h(y, 1)|/|y| — 0 as |y| — 0 for each fixed t. We abbreviate
this limiting procedure by writing h(y, t) = o(ly|) as ly| > 0. Subsequently, then,
we shall consider the equation

x = AMx+ hd), B<t<+o, h(0,H=0, h(x, sf = o(|x|) as |x|—0.

(10.1b)
This equation is called quasi-linear or perturbed-linear under the stated hypothesis.
Note that when the reduction procedure above is applied to the linear equation
x’ = A(z)x + b(z), the resulting quasi-linear equation has h(x, ft) = 0, that is, it is
the linear equation x’ = A(t)x. Thus the stability properties of any solution to a
 Definitions 309

nonhomogeneous linear equation are precisely the stability properties of the zero
solution to its complementary equation. It follows also that every nonzero solution
of x’ = A(t)x exhibits precisely the same stability properties as the zero solution.
Thus stability considerations for any solution of a linear equation are reduced to
Stability considerations for its zero solution or the zero solution of its complementary
equation. One may therefore speak of the linear equations or systems, rather than
their solutions, as being stable or unstable.

EXERCISE 10.1

1. Refer to Fig. 9.11 (Example 1, Section 9.5). List the subsets of the xy-plane that are
invariant with respect to the system of the example.
2. Prove that a stable limit cycle for a two-dimensional autonomous system must be an
asymptotically stable invariant set.
3. Prove that a stable node or focus for a two-dimensional autonomous system must be
an asymptotically stable invariant set.
4. Show that a rotation point, hence a center, for a two-dimensional autonomous system
is a stable invariant set.
5. No solution of the equation x’ = —t/x is stable. Why?
6. Show that the zero solution of x’ = (2 — Ax/t, t > 1, is uniformly asymptotically
stable. Compute a 6; and, given €; > 0, compute a 7 such that a solution y with |W(r1)| <
61 for some ¢; > 1 satisfies |¥(t)| < €1 for all t > T 4+ £1.
7. Let g(t) = x1 1/01 + k4(t — kK)?). Show that the zero solution to the system
x’ = g’(t)x/g(X) is stable, but not asymptotically stable.
8. For each of the following equations find a quasi-linear equation based on the given
solution.
a) x’ =y+xl — x? — y”), yk = —x + yl — x? — y?), x = sint, y = cost.
b)sx) nee = 1x. x = 0, x4 = x’ = 0:
9. Suppose one wanted to deduce the stability of the nontrivial periodic solution x = p(t)
of van der Pol’s equation by studying the stability of equilibrium for the associated quasi-
linear equation. What is the primary difficulty in actually implementing the procedure?
10. a) Show that the critical point (1, 0) is a quasi-asymptotically stable solution of the
system
3 2
3) = ae = 28
x’ =x-—yt+ es
ee sis ee
2 2 3
; acta KEV.
een AS a
Vx2
+ y?
b) Why could a quasi-asymptotically stable cruising attitude be an undesirable feature
in an aircraft?
11. Prove that a constant solution to an autonomous system is uniformly (asymptotically)
stable if and only if it is (asymptotically) stable.
 Stability 10.2
310

10.2 THE STABILITY OF LINEAR SYSTEMS

Most of the results obtained in Chapter 7 are in fact statements about the stability of
linear systems. Here, we shall discuss these results in terms of the formal definitions
of stability. .
Theorem 10.1 below can be motivated by consideration of the phase portrait for
the equation x” + w?x = 0, w > 0 (Fig. 10.5). One thinks of the elliptical paths

=
A

Figure 10.5

as the orbits of hypothetical particles which were all on the y-axis at time ¢ = 0.
To say that the zero solution is stable is to say that a particle which is once sufficiently
near the origin will never leave a previously designated neighborhood of it. Let us
suppose now that the curve labeled C in Fig. 10.5 is the path of the solution x = y(f),
y = W'(0), and consider the path labeled fr. Since the equation x’ + w?x = 0 is
linear, there is a scaling factor u > O such that I is the path of the solution x =
uy(t), y = wy’(t). In fact, every nontrivial path in the plane is related to every other
one in this way. Thus the boundedness of solutions is equivalent to the stability of
the zero solution for this particular equation. We show below that a similar assertion
is true for linear equations in general. A second implication of Theorem 10.1 is that
linear systems do not exhibit the pathological behavior illustrated in Figure 10.4 or in
Problem 10 of Exercise 10.1.
Theorem 10.1. The equation (LH) x’ = A(t)x, t > 6 is stable if and only if all
its solutions are bounded on intervals of the form to <t< +m, to > 6. It is
(globally) asymptotically stable if and only if all its solutions approach zero ast > +a.
Proof. Let & denote a fundamental matrix solution for (LH) and let to > 6 be given.
Suppose first that every solution is bounded for tg < t < +o. Then there
is a constant M such that |®(t)| < M for to <t < +a. Let e > 0 be given and
choose 6 = €/M|®~'(to)|. If @ is a solution satisfying |@(to)| < 4, then

[e()| = |) "(t0)d(to)| < |&C| - |&7!(£0)| - |@(t0)| < €

for to < t < +o. Thus (LH) is stable.
Conversely, suppose that (LH) is stable and let € > 0 be given. There exists a
5 > O such that |¥(to)| < 6 implies |¥(| < € for to < t < +a whenever y is a
 The stability of linear systems 311

solution of (LH). Let ¢ denote any nonzero solution and define ¥(t) = $(t)6/|@(to)].
Then |W(to)| < 6, and it follows that |y(A)| < € for tp <t< +o. This means
that |@(t)| < € |@(to)|/6 for to < t < +, that is, o(t) is bounded over the in-
dicated interval.
Next suppose that every solution approaches zero as t > +. Then, for each
to = B, each solution is bounded over intervals of the form tp < t < +o. Thus
(LH) is stable by the second paragraph of this proof. By definition, then, it is
asymptotically stable and globally so.
Conversely, suppose that (LH) is asymptotically stable. Then it is stable and,
given 1, there exists a 6; > O such that |y(t,)| < 6; implies Y(t) ~ 0 ast> +a
for each solution y. Let @ denote an arbitrary nonzero solution of the system
and define ¥(t) = 6:0(t)/|o(t1)|.. Then |¥(t,)| < 6,. Consequently y(t) > 0 as
t— +o and ¢(f) must do likewise. ||
Corollary [Link]. The equation x' = Ax is asymptotically stable if and only if
each eigenvalue of A has negative real part. The equation is stable, if and only if
i) every eigenvalue of A has nonpositive real part, and
ii) A has m linearly independent eigenvectors corresponding to each eigenvalue
with zero real part and multiplicity m.
Proof. The assertions are immediate consequences of Theorem 10.1, Corollary 7.2,
and Corollary 7.1. ||
Corollary 10.1b. The equation x' = A(t)x, where A has minimal period T > 0,
is asymptotically stable if and only if every characteristic exponent has negative
real part. Let e®” denote a period transformation matrix. Then the equation is stable,
but not asymptotically stable, ifand only if
i) every characteristic exponent has zero real part and
ii) B has m linearly independent eigenvectors corresponding to each eigenvalue with
zero real part and multiplicity m.
Proof. The assertions are immediate consequences of Theorem 10.1 and Corol-
lanys/25o81
The next theorem gives necessary and sufficient conditions that a general linear
equation be uniformly stable or uniformly asymptotically stable.

Theorem 10.2. The equation (LH) x’ = A(t)x is uniformly stable if and only if
every fundamental matrix solution & has the property that |\b(t)@~ '(t,)| is uniformly
bounded for all t, and t in the interval [8, +x). The equation (LH) is uniformly
asymptotically stable if and only if there are constants M > 0 and » > O such that
\b(t)}b—1(t,)| < Me~"'— for all t and ty satisfying BS ti St< +z.

Proof. Let denote a fundamental matrix solution of (LH).

Suppose first that there is a constant M = M(@) such that |B()e~'(t1)| < M
forB <t; <t< +a. Lete > 0 be given and choose 6 = e/M. If @is a solution
312 Stability 10.2

such that |$(t,)| < 6 for some t, > 8, then |6(1)| = |L(& *(t1)o(t1)| < Mig) < €
for all t > t,. Thus (LH) is uniformly stable.
Conversely, suppose that (LH) is uniformly stable. Let € > 0 be given and let
6 > 0 be as in the definition of uniform stability. Fix an arbitrary ¢; > 6 and let
V(t) = [¥i(0),..., Wn(t)]" denote that solution of (LH) which satisfies ¥,(t1) = 6/2,
¥(t1) = Oif 7 #i (1 <i<n). By definition of uniform stability |¥()| < € for
all t > f,. Since ¥(t) = &(t)}6~ 1(t,)¥(t,), it follows that € > |W(t)| = M(t, t1) 6/2
for all t > t,;, where M,(t, t,) is the norm of the ith column of (ft) '(t,). Thus
1(t,)| < 2ne/6 for all t > fy.
|\&(1)@—
To prove the second assertion, assume that there are positive constants M =
M(8) and » = 7(8) such that |6()}6~1(t,)| < Me"? for all ¢, and ¢ satisfying
B<t, <t< +a. The zero solution is then uniformly stable by the preceding
paragraph. Let @ be any solution of (LH) which satisfies |@(t,)| < 1/M = 6, at
any time ¢; > 6 and let €, > O be given. If t > T+ t1, where T= —(ln€;)/n,
then |@(1)| = |®()b—1(t:)d(t,)| < e7™—'_ < el" 4 = €;. Thus, the zero solution
is uniformly asymptotically stable.
Conversely, suppose the zero solution is uniformly asymptotically stable. Then
it is uniformly stable, and there exists a constant N = N(@) such that |@(1)@~ '(t,)| <
N for all ¢ and ¢, satisfying B < tj < t << +a. Let 6, > 0 be as in the definition
of uniform asymptotic stability and choose an €; > Oso small that 2ne€,;/6; =X < 1.
Then let 7 be as in the definition of uniform asymptotic stability also.
To show that |®()~ '(t;)| is in fact dominated by an exponential function, we
consider again a special solution y satisfying ¥;(t;) = 6,/2 and y,(t,) = 0 if
JeriSn) Then, tori

Milt, 1) = [B(OH = WO! < a1

where M,(t,t,) denotes the norm of the ith column of &(t)@7'(t,). Thus
|\B(t)}b—*(¢,)| < > for all ¢, and ¢ satisfying ¢ — t; > 7T, t; > 8. Now write
t— t; = kT +r, where k is a positive integer and 0 <r < T. Then

|B()b-*(41)| = [OC + AT + NO7*(14)|

k

lA le, + YO") TI le + m7 + No (t, + (m — 1I)T + 7)

m=1
xe Nv ~ Ne? ‘(t-4-7 Ind = Me»)

where M = Ne” and 7» = —T~'In . ||

Uniform stability is nor a refinement of the usual stability concept for the equation
x’ = Ax with constant coefficients or the equation x’ = A(f)x with periodic
coefficients.
Corollary 10.2a. The equation x’ = Ax is uniformly (asymptotically) stable if
and only if it is (asymptotically) stable.
 The stability of linear systems 313

Proof. Uniform (asymptotic) stability implies (asymptotic) stability by definition.

Conversely, assume first that x’ = Ax is asymptotically stable. If © is a fundamental
matrix solution, then ®(t) = e4‘C for some matrix C. Since the equation is asymp-
totically stable, there are positive constants M and n such that |®(1)| < Me~",
t> 0. Thus |6()6~'(¢,)| = |e4—'?| < Me™), and it follows that the equa-
tion is uniformly asymptotically stable. If x’ = Ax is merely stable, the same con-
siderations apply with 7 = 0. Then |®(1)@~ '(r)| is merely bounded, and it follows
that the equation is uniformly stable. _ ||
Corollary 10.2b. The equation x' = A(t)x, where A has minimal period T > 0,
is uniformly (asymptotically) stable if and only if it is (asymptotically) stable.

The proof is analogous to that of Corollary 10.2a and is left to the exercises.
Example. Let a(t) denote the sawtooth function depicted in Fig. 10.6. The function

a(t)
A

2m ---—> |

/ 4(m + 1)

Figure 10.6

a(t) has the property that ApS a(t)dt = —a. The differential equation x’ = a(t)x
is asymptotically stable since

o(t) = exp iffa(r) ar|

is a fundamental solution. It is not uniformly stable since

Ho =ex0 i a(r) ar|

is not uniformly bounded over all intervals of the form 0 <4; <tf< +a. To
see this, take ¢; = 2m, t = 2m +1. Then

ot) _ Fiber A cova aa lee as m—>-+o. ||

$(t1)
One can construct other examples of nonuniformly stable equations with the aid
of the functions tsinIn ¢ and fcosIn¢. They behave in somewhat the same way
as the sawtooth function above.
314 Stability 10.2

The notion of uniform stability is pertinent to Theorems 7.8, 7.9, and 7.10. The
student will be asked to reformulate these theorems in terms of uniform stability in
the exercises below.
Theorem 7.7 can be used in conjunction with Theorem 10.2 to test the general
linear homogeneous system x’ = A(t)x for uniform stability. Specifically, we have
the following theorem.
Theorem 10.3. Let A be a continuous n X n matrix function on the interval
B <t< +o and let M(t) denote the largest of the eigenvalues of A(t) + AL(Do Ife
M(t) < 0 for t > 6, then x’ = A(t)x is uniformly stable. If there exists a constant
n > 0 such that M(t) < —7n < 0 for t > 8, then the equation is uniformly asymp-
totically stable.
Proof. Let @ be an arbitrary solution and suppose that M(t) < —n for ¢t = 8B,
where 7 > O is a constant. Let € > 0 be given and choose positive 6<e. By
Theorem 7.7,
t

HsColl < lloten|iexe|

—4fae as|< Cad] |exp[Bole — 10)
ie
for BED Gest eee (10.2)
If |o(t1)| < 6, then it follows that |¢()| < € for 4; <t< +o. Thus the
equation is uniformly stable.
Suppose now that 7 > 0. Choose 6; = 1 and lete, > 0 be given. If |¢(t,)| < 1,
then it follows from inequality (10.2) that |¢(f)| < €; for all t—1t, > T=
—(2/n)Iné,. Therefore the equation is uniformly asymptotically stable. ||

EXERCISE 10.2

1. For what values of the real parameter a is the zero solution of

x’ i = {sinIn¢ + cosIn¢ — 2a}x

asymptotically stable, but not uniformly so?
2. Prove Corollary 10.2b.
3. Prove Theorem 7.9 with hypothesis (ii) deleted and hypothesis (iv) changed to the
following: x’ = L(x is uniformly stable.
4. Consider a linear equation x’ = L(1)x, where L is continuous on an intervala < t <
+oo, Suppose it is stable and
t

lim int | Tr L(s) ds % —~, to > a.

toto Jt

Must it be uniformly stable? Could a stable equation with

t

lim inf Tr L(s) ds = —o@

t—-+00 to

be uniformly stable? Relate your answers to Theorem 7.10.

 The stability of equilibrium for quasi-linear systems 315

5. Discuss the stability of the systems

» 3] -4[7 SI) »[s] -2[7) SIe]+E]

6. Show that the zero solution of an equation of the form x” + p(t) = 0, where p is
continuous ona < t < +, could not be asymptotically stable. Generalize the result to
nth order scalar equations.
7. Discuss the stability of the following equations.
a) xe wemminc tar —-(), b) x’ —e tx’ +x = 0.
c) x’ + x7 + (1 — (V1 + 212)/t*)x =

8. Let p and q be continuous functions on a < tf < +. Give conditions on p and q

which will guarantee that the equation x’’ + q(x’ + p(x is uniformly stable.
9. Give conditions on constant-coefficient and periodic-coefficient equations x’ = Ax and
x’ = A(t)x which will imply instability.

10.3 THE STABILITY OF EQUILIBRIUM FOR QUASI-LINEAR SYSTEMS

In elementary physics, the equation of motion

6’ + (g/£) sin 6 = 6” + (g/f)(6 — 07/6 + ---) =0
of the simple pendulum is often approximated by the linear equation 6”’ + (g/£)0 = 0
under the stipulation that |6| be sufficiently small. This is a specific case in which a
quasi-linear system is approximated by its linear part.
Under the hypothesis h(x, t) = o(|x|) as |x| > 0, one might feel that the zero
solution of a quasi-linear equation
x’ = A(t)x + h(x, 2) (10.3a)
should be stable whenever the linear approximation x’ = A(ft)x is stable. This is
not generally true. The zero solution of the nonlinear equation (9.7f) in Example 3,
Section 9.7 is unstable (Fig. 9.24b). The linear approximation is stable. As the
following example of Perron shows, even the asymptotic stability of its linear approx-
imation does not imply that the zero solution of a quasi-linear equation is stable.
Example (Perron). A quasi-linear system with unstable equilibrium and asymptotically
stable linear approximation.
We consider the system

eS saat (10.3b)
y’ = [sin Int + cos Int — 2a]y + x’, Bt
2 Sig <=) + e=8)/2, “7 > 0:
One easily finds that all solutions are given by
at
p= Cier t
y(t) = i
ee ae = at (oy + of2 | e —s $s sin
sin Ins
ds):
0
316 Stability 10.3

We first underestimate the integral

t

F(t) = / fds f(s) = —ssinIns.

0

To do this, notice first that if 0 < u< vu < ¢, we have

ro > | e?) ds.

Next we find an interval (u,v) on which f(s) is increasing. Set o = Ins. Then
f(s) = —(sino + cos a), and sketching the graphs of cos o and —sin o (Fig. 10.7),
we see that f is increasing for

se+ 2nr < Ins < usA Dine -OL Ol cen ee ee AU:

Figure 10.7

Let t, = e?"*+*/? and u, = t,e-*, Un = tre

°* 4, n > 1. Then
Un

FG yee / ef ds > (un — unje™” = (vn — un)e™*.

uU n

Now look at the sequence {y(t,)}. We have

(tn) = ein sin In t,—2at, lca é | C ef) as

_lool + ci(e—8"!4 — et, exp [tne

> e296 7}
= —|cnle7 4 Cie "4 =e" )\iexp [7b ee = 2a),
Then, regardless of the choice of cy and the choice of c; ¥ 0, we have

lim sup y(t) = too.

n—-0

Thus the quasi-linear system has a critical point (0,0) which is unstable.
 The stability of equilibrium for quasi-linear systems 317

However, the linear approximation

nt, 10.3
y’ = [sinIn¢ + cosIn¢ — 2a]yp cee
has a fundamental matrix solution

eat 0
0 ef sin Int—2at

which approaches zero as t—» +. Thus the origin is globally asymptotically

stable for the linear approximation (10.3c). ||
The reader will recall from Problem 1 of Exercise 10.2 that the system (10.3c)
is not uniformly stable. The next theorem shows that uniform asymptotic stability
of the linear approximation carries over to a quasi-linear system with an added
hypothesis on h(x, f).
Theorem 10.4. Consider the system

x’ = A(t)x + h(x, 1), Bit< +o, (10.3d)

where h(0, t) = 0 and h(x, t) = o(|x|) as |x| > 0 uniformly with respect to t for
B<t< +o. [If the zero solution of x' = A(t)x is uniformly asymptotically stable,
then so is the zero solution of Eq. (10.3d).
Proof. Let & denote a fundamental matrix solution of x’ = A(t)x. Since this equa-
tion is uniformly asymptotically stable, there are constants M > 1 and 7 > 0 such
that |&()é~ 1(t,)| < Me”? for all tf; and fg satisfying B < ty; <t< +o.
Since h(x, t) = o(|x|) as |x| + 0 uniformly with respect to ¢ > 8, there exists
a 6, such that |h(x, 2)| < n|x|/2M for 8 < t < +o whenever |x| < 6;. Define

Now let @ denote a solution of Eq. (10.3d) which satisfies |@(t¢;)| < 6 for some
t; > 6. Then |@()| < 6; on some maximal forward interval 4; <¢<T<
7*(Xo;fo). Since

ld) = a a,J0(4) + 00)| BORO), 5)

it follows that :

aC] < Mlpcayien™—” + 2 fe" —Ig¢0)| ds

for tj <t< T. Consequently,
t

loMle"— < Mle(s)| + a er’ [g(s)| ds

for the same values of ¢. By the Gronwall inequality (Problem 5, Exercise 7.4),
"
|o(t)| < Mid(tyle2”, (10.3e)
318 Stability 10.4

for ft; < t < T. Since M|d(t;)| < Mé = 464, inequality (10.3e) implies (by Theorem
8.6) that T = TT (Xo, fo) = +o. Thus $(¢) exists for t; < t < +0 and inequality
(10.3e) holds over the same interval. This inequality implies that the zero solution
of the quasi-linear equation is uniformly asymptotically stable. The details are left
to the exercises. ||

EXERCISE 10.3
1. Explicitly compute a general solution for the system (10.3b).
2. Is the zero solution of the system (10.3b) unstable for any value of a > (1 + e~”)/2?
Compare your result with your answer to Problem 1, Exercise 10.2.
3. Prove that inequality (10.3e) implies the conclusion of Theorem 10.4.
4. Let g be defined by
= k
0O= yma ee 129
= > > () .

Is the origin a uniformly asymptotically stable critical point for the system

MS SX yy 2);
Neate? Naa oe aaa |
5. Abstract the analysis for Problem 4 and obtain a theorem similar to Theorem 10.4.

10.4 THE STABILITY OF PERIODIC SOLUTIONS

We consider here the stability of periodic solutions for the equation

xii 2) (S)
under the assumption that there exists a T > 0 such that f(x, t+ 7) =f (x, d,
—%» <t< +a. Notice that if (S) is autonomous, then f(x, t+ T) = f(x, 2) for
all T > 0. As a matter of completeness, we state an existence theorem for an
important special case of (S).
Theorem 10.5. Consider

x’ = f(x, t) = Ax + h(x, f) (10.4a)

where A has a stable characteristic polynomial, h is bounded, and h has minimal period
T > 0 for each fixed x # 0.
i) If Eq. (10.4a) has no constant solution, then it has a periodic solution with
minimal period kT, where k is a positive integer.
ii) If h(x, t) is of the form h(x, t) = q(x) + p(t), then Eq. (10.4a) has a solution
with minimal period T > 0.
The proof of this theorem is beyond the scope of this book, and we omit it.
 The stability of periodic solutions 319

Example 1. The system

x’ = yt sin(x+ DO),
y = -x—yt+cos(x+9),
has a solution of period 27k, where k > 1 is an integer. ||
Example 2. The system

= y+ pas + sin t,

Dee A Tsaire
oe rece:Sic COS Js
has a solution of period 27. ||
We assume for the rest of this section that the equations to be considered have
(nontrivial) periodic solutions and consider the stability of these solutions.
Theorem 10.6. Suppose that the partial derivatives D,f, k = 1,...,n, are con-
tinuous on D = P X (—xwx, +), where & is a domain in &", and let the equation
x’ = f(x, £) have a solution @ with minimal period w = kT, where k > | is an integer.
If the linear variational system

y =J (Dy, (10.4b)
where J,(t) = (D; f:(o(2), t)), is asymptotically stable, then so is the solution 9.
Proof. In view of the normalization remarks at the end of Section 10.1, the theorem
can be proved by demonstrating the uniform asymptotic stability of the zero solution
to the system
y= JDy + hey, 2), (10.4c)
where
h(y, t) = f(o() + y, t) — S(O, t) — Jey.
First let us show that h(y, 4) = o(ly|) as |y| — 0 uniformly for —%2 <t< +o.
To do this, let C be a closed and bounded subset of ® with the trajectory of @ in
its interior (Fig. 10.8).

o() +y

Figure 10.8
 Stability 10.4
320

Since ¢ is periodic, it follows that ¢(¢) + y is interior to C for — GS t <b

provided |y| is sufficiently small. For j= 1,...,n, it follows from the mean value
theorem that
nly.) = D5 DiflEO, 1) — DifiloO, OV
—!

where £;(t) is on the line segment from ¢(f) to ¢(t) + y. Since the partial derivatives
D, f ;(x, t) are uniformly continuous and periodic on the product set C X (—o, +),
it follows that there exists a positive function M(y) such that M(y) — 0 as |y| — 0
and |h(y, t)| < M(y)y, for all y with |y| sufficiently small, uniformly with respect
LONE:
The system (10.4b) is periodic and asymptotically stable. It is therefore uniformly
stable by Corollary 10.2b. The zero solution of (10.4c) is then uniformly asymptoti-
cally stable by Theorem 10.4. ||
Example 3. The system
xo = = 1-203 xe 2 cosi#ti3 sin f-Eecos? &
; . (10.4d)
y = —-y+y?sint
+ y?x
has a periodic solution x = sint, y = 0.
The variational equations based upon this solution are
ai . a teesin f eH {| (10.4e)
As defined in Theorem 10.3, M(t) = —1. Thus the system (10.4e) is (uniformly)
asymptotically stable. By Theorem 10.6, so is the solution x = sin t, y = O of the
system (10.4d). ||
The last theorem does not apply to autonomous systems with periodic solutions,
for the corresponding variational equations will always have a periodic solution and
cannot be asymptotically stable. Suppose, in fact, that y is a periodic solution of
x’ = f(x); then y = y’(¢) is a periodic solution of

y’ = J,(t)y, I(t) = (Df).

The proof of this remark is left to the exercises. There is an analog of Theorem 10.6
for autonomous equations. Since the proof is notationally involved, we shall give
it for the case n = 3 only.
Theorem 10.7. Suppose that the partial derivatives D,f, k = 1,...,n, are con-
tinuous on a domain & in &” and let the equation (A) x’ = f(x) have a periodic solution
Y with minimal period w > 0. If the linear variational equation

y’ = Je(d)y, (10.4f)
where J-(t) = (D;f iW(D)), has n — 1 characteristic exponents with negative real
parts, then the path T of p is asymptotically stable. If Eq. (10.4f) has one characteristic
exponent with positive real part, then T is unstable.
 The stability of periodic solutions 321

Proof. (n = 3). Write

(Xo, V0. Z05 t) = (¢1(xo, V0. Z05 t), $2(Xo, V0. Z0s t), $3(Xo, V0 Z0; t))

for the solution of (A) which satisfies $(xo, yo, Zo, t) = (Xo, Yo, Zo). Then y(t) =
(0, 0,0, t). We may assume without loss of generality that the coordinates are
chosen in such a way that ¥(0) = 0 and y/(0) = f(0) is the unit vector along the
positive x-axis (Fig. 10.9).

Figure 10.9

It follows from Theorem 8.7 that (0, yo, Zo, f) exists for |yo| + |zo| + |t — w| <
€, say, since (0, 0,0, w) exists. Consider the equation $,(0, yo, Zo, t) = 0. The
values yo = Zo = 0, ¢ = ware solutions by construction. Since (0/dt) $,(0, 0, 0, w)
= (0/dt) $,(0,0,0,0) = 1, it follows from the implicit function theorem that
there exists a continuously differentiable function T defined for |yo| + |zo| < €1 < ¢,
say, such that
$1(0, yo, Zo, To, Z0)) = 0.
The geometrical significance of these remarks is that any solution vector
(0, yo, Zo, ¢) initiating on the yz-plane with |yo| + |zo| sufficiently small will re-
intersect the yz-plane at a time 7()o, Zo), not very different from w. Thus

ie Vos Zo) =, (¢2(0, V0. Z0s TWV0; Zo))s $3(0, V0. Z05 T(0; Zo)))

maps a neighborhood of the origin in the yz-plane onto another such neighborhood.
By the mean value theorem,

F(Vo, Zo) a Jr(O, 0, *°|

ae R(xo, Yo), (10.4g)

where

T(0, O)s= D2$2(0, 0, 0, w) D362(0, 0, 0, a

#0, = |9.45(0, 0,0, 0) Ds63(0, 0, 0, )

and |R(xo, ¥o)|/(|xo] + |vol) 2 0 as |x| + [vo] > 0.
322 Stability 10.4

By Theorem 8.9,
D,$,(0, 0, O; t) D2 1(0, 0, 0, t) D3 1(0, 0, 0, t)

Y(t) = |DiG2(0, 0,0, 1) Dod2(0, 0,0, 1) DsH2(0, 9, 9, 1)

D,¢3(0, 0, 0, t) D2$3(0, 0, 0, t) D363(0, 0, 0, t)

is a fundamental solution matrix of the periodic, linear variational equation

y = Je(d)y
and satisfies Y(0) = J. Thus Y(w) is a period transformation matrix. By Problem 6
of Exercise 10.4 and the choice of coordinates, y = &(t) = y(t) is the unique periodic
solution of the variational equation which satisfies &(0) = [1,0, 0} sen Butey =
D, (0, 0, 0, ¢) is a solution which satisfies the same initial condition. Thus the first
column of Y(w) is [1,0,0]’. Since, by hypothesis, the variational system has two
characteristic exponents with negative real parts, it follows that the eigenvalues
1, Ag Of Jpr(0,0)—which is the lower right 2 X 2 submatrix of Y(w)—satisfy
[A] < land |\y| < 1. We assume without loss of generality that |\o| < |\i|. Now
let P be a 2 X 2 matrix such that

=F a NS
P= J;(070)P' = E a

with

[hal < [Ae] + ¥ < 1. Introducing new coordinates [u,v]" = P~*[y,z]’ in the
yz-plane, one finds that (10.4g) takes the form

d Yiju
G(uo, Vo) = ine ef[3 + S(uo, Vo),

where G(uo, Vo) = P~'F(0, Zo) and S(uo, Vo) = P~'R(Xo,

Yo). Then

IG@o, o)| < [Aa] + [Mol + [7] +[vol + |Aal [vol + [S(o, vo)|

IA (|u| + eo) | ce Piel |

ats [wol + [vol

Since |S(uo, 09)|/([uol + |vo|) > 0 as |uo| + |vo| — 0, there exists a neighborhood
N of the origin in the yz-plane such that every point p in N with wv-coordinates
(Uo, Vo) is carried by F into a point q with w-coordinates G(uo, Vo) where |G(uo, vo)| <
k(|uo| + |vol) and k < 1. Thus, the path of every solution (0, yo, Zo, t) which
initiates on the yz-plane sufficiently near the origin reintersects the plane at a point
F(yo, Zo) which is strictly nearer the origin in terms of the wv-coordinates. This
implies that the orbit of y is asymptotically stable. The instability assertion is left
to the exercises. ||
Example 4. The equation
x! + (x? + x’? — 1x’ +x =0 (10.4h)
 The stability of periodic solutions 323

has a periodic solution x = sin ¢. The system of variational equations based upon
this solution is
1 i |
Uo 0
— 1 — cos 5 “|. (10.41)
H i ge = sii2/)
Let us write uw, and yo for the characteristic multipliers of Eq. (10.4i). Since the
system (10.41) has at least one periodic solution (u = cost, v = —sin f), at least
one characteristic multiplier has unit modulus. Say |w,| = 1. By Theorem 7.6,

Mise = exp || (—1 — cos 22) a =e

0
Thus |w2| = e~* < 1. By Theorem 10.7, the unit circle in ®? is a stable invariant
set for the system x’ = y, y’ = —x + y — x?y — y? equivalent to Eq. (10.4h).

EXERCISE 10.4

1. Let (x0, fo, t) denote the solution of

x’ = —x + sin(x+ A

which satisfies (xo, to, 0) = xo. Consider the function y = F(xo) = (xo, to, to + 27)
on an appropriate interval —R < xo < Rand use the intermediate value theorem to show
that there is a value Xo such that x9 = ¢(xo, fo, fo + 27). Then deduce the existence of
a periodic solution.
2. Carry out the analysis of Problem 1 for the equation x’ = —x + sinx + sint.
3. Is the solution x = cos 2t, y = sin¢ of the equations

x’ = —2x + dy? + 2cos2t — 2 sin2t — |Z

y! = —2y + 4x? — $cos?2t + 2sint + cost
stable?
4. The equations
x! = —4x 4+ x7y + xz,
y=-4yt+z2—-1+y24 224+ 4sinzs,
z= —y— 4z+ x*y 4+ 4cost
have a periodic solution with x = 0. Is it stable?
5. Discuss the stability of the periodic solutions of

SS y’ = (cosy + x?.

6. Let w denote a (proper) periodic solution of x’ = f(x), where f is continuously dif-

ferentiable. Show that y = W’(d) is a periodic solution of

y =J(dy, Je) = (Dif OD),

which satisfies y = f(¥(0)) when ¢ = 0.
324 Stability 10.5

7. The system

Xe yy ae ey
has a periodic solution x = sint, y = cost. Is the path of this solution stable?
8. The system x’ = x — xyt — x°, y’ = —yx* — y® has a periodic solution with path
x4 + y+ = 1. Show that the path is asymptotically stable.
9. Let f and g have continuous first partial derivatives and assume that the system v=
f(x,y), ¥’ = g(x, y) has a periodic solution (x, y) = ¢(f) = (u(, v(t). Show that the
orbit of @ is asymptotically stable if Di f(x, y) + Dog(x, y) < 0 at all points (x, y) on the
orbit. Try to apply this result to van der Pol’s equation.
10. Prove the instability assertion of Theorem 10.7.

10.5 THE DIRECT METHOD OF LYAPUNOV

In Sections 10.2 and 10.3, we discussed the stability of equilibrium for a quasi-linear
equation
x = Ax +b, 7) (10.5a)

by comparing its solutions with those of the linear approximation x’ = A(#)x. It

was pointed out at the beginning of Section 10.2, that this method is not always
useful. For some quasi-linear systems, the stability or instability of equilibrium is
determined primarily by the nonlinear terms, and the systems must be studied in a
way that takes nonlinearity into account. Lyapunov’s direct (or second) method is
useful for the study of such systems. Before formalizing the discussion, we motivate
the technique by applying Lyapunov’s method to the study of a specific system.

Example 1. The origin is an isolated critical point for the system

x= Nb 2) = ae yey) ee (10.5b)
The associated linear approximation is the system x’ = — x, y’ = 0, and it gives no
information about the stability or instability of equilibrium for the nonlinear system.
Consider, a priori, the function V(x, y) = (x? + y?)/2. It is an example of
what we shall later call a Lyapunov function. The stability of equilibrium for the
system (10.5b) can be deduced with its aid.
Let us, as indicated in Fig. 10.10, realize in ®*, the surface M defined by the
equation z = V(x, y), regarding the xy-plane as the phase space for the system
(10.5b).
The surface M is a circular paraboloid with its lowest point at the origin. Now
let o(t) = (u(t), v(t)) denote a nonconstant solution of the system (10.5b) and
denote its path by C. For each t > 0, the number V(o(0) is the length of the vertical
line segment L(t) from the point $(t) on C to the surface M.
We shall show that DV(¢(t)) = dV(@(1))/dt < 0 for t > 0, provided |$(0)|
is sufficiently small. This will imply that the length of L(t) decreases strictly with
passing time. Because of the geometrical features of M, the point $(f) will then
 The direct method of Lyapunoy 325

$(t) = (ud), oY)

Figure 10.10

move toward the origin in such a way that the asymptotic stability of the origin
can be deduced.
Note that it is not necessary to solve the differential equations in order to compute
DV(¢(t)). In fact (suppressing the argument ¢ for notation convenience), we have
DV(@) = uu’ + vv! = —(u? 4 v*) + 2(u? + vu + v)?.
By the triangle inequality,

2(u* + vu + 0)? < 2G + v*)(\ul + |e))?.

Thus
DV(o) < (u? + v*)[2(\u| + |v|)? — 11.
Notice now that V(x1, 1) > V(xo, yo) if and only if |x,| + [yal > [xo] + |yol-
Let € > 0 be given and assume without loss of generality that € < 1/\/2. Let a
positive 6 < € be chosen so small that V(x, y) < € for all points (x, y) with
|x| + |y| < 6. If |@(0)| < 6, then
DV(¢()) < (v*() + u?2()[26? — 1] < 0
on some interval 0 < ¢ < ty < 7*(¢(0)). Thus, if O<s<1t< t1, V(d)) <
V($(s)), and it follows that |@(2)| < |d(s)| < 6 < €. Hence 4; = 7*(¢(0)). Then
it follows from Theorem 8.5 that 7*(¢(0)) = +o, the required closed and bounded
set of the theorem being the square |x| + |y| < 6. Thus ¢(f) exists and |@(f)| < e€
for 0 < t < +o provided |¢(0)| < 6. This means that the origin is stable.
To establish asymptotic stability, we show that ¢(/) > 0 as t— + if |@(0)| < 6.
Since DV(¢(t)) < 0 and V(¢(t)) > 0, lim: 4. V(¢() = L > 0 exists. Suppose
L > 0. Then there exists a tf; > to such that V(#(t)) > L/2 for ty < t < +o.
This implies that there is a constant 7 > 0 such that v4(¢) + u(t) > 7 for ¢ in the
same interval. Then
DV(6(2)) < (28? — 1) < 0
326 Stability 10.5

for all ¢ > t, and it follows that V(¢(t)) ~ —x as t— +, a contradiction.

Thus lim;4 |@(O|2 < 2s so (W2() + v7(1) = 4limy44. V(@@) = L = 0. ||

The general technique which was used to deduce stability in Example 1 is called
the direct method of Lyapunov. We shall give Lyapunov’s basic stability theorems for
the equation
x’ = {Go 7), (S)

where f is continuous and lipschitzian on a region ® X [6, +~) in R” X @!, under

the assumption that the origin x = 0 is an isolated critical point for (S) in ®. The
other results of this section will be given only for autonomous equations (A) x’ = f(x).

Definition. A real-valued function W is said to be positive (negative) definite on a

neighborhood G of the origin if W(x) > 0 (W(x) < 0) for all x # 0 in G and
W(0) = 0. A function V defined on a cylinder G X [8, +), where G is a neighbor-
hood of x = 0, is called positive (negative) definite if V(0, 7) = 0 for t > 6 and
there exists a positive (negative) definite function W on G such that W(x) < V(&, 2)
(V(x, t) < W(x)) for all (x, t) in G X [8, +a). If Wx) > 0 (W(X) < 0) on G,
W is called positive (negative) semidefinite. If V(x, t) > 0 (V(x, 1) < 0) on G X
[8, +»), W is called positive (negative) semidefinite.

Example 2
i) The rectangular and euclidean vector norms |:| and ||-|| are positive definite
on &”.
ii) The function W defined

W(x, y) = —x? — y+ 20? + yp) yp)?

is negative definite on the neighborhood |x| + |y| < 1//2.
lii) The function V defined by V(x, y, 4) = x? + y® + e~ is positive definite
on ®? X [8, +) for every B.
iv) The function W defined by W(x, y) = (x? — l)y? is negative semidefinite
on the strip |x| < 1, —o < y < +00.
v) The function V defined by V(x, y) = (x? + y”)e~ is not positive definite
on any cylinder G X [8,+.0). ||

We are now in position to formally define the Lyapunov functions.

Definition. A continuous real-valued function V is called a Lyapunov function

for the equation (S) x’ = f(x, ft) at the origin if:
i) there is a cylinder G X [8, +), where G is a neighborhood of the origin, on
which V is positive definite, and
ii) when @ is a solution of (S) with (to) in G, then V(9(0), t) does not increase
with increasing t > ty > B.
 The direct method of Lyapunoy 327

Although the definition has been phrased for nonautonomous systems, it is an

intuitive aid to think of V(¢(0), t) as the decreasing length of a line segment, as in
Example 1. In this case, the surfaces M; defined by z = V(x, 1), t > B, can be
thought of as fluctuating in time but always lying above the surface M defined by
z = W(x), where W is positive definite and W(x) < V(x, 1). See Fig. 10.11.

Figure 10.11

Lyapunov’s theorem which we shall prove below states that the origin is a stable
critical point for (S) x’ = f(x, 2) if there exists a Lyapunov function for (S) at the
origin. The theorem does not tell how one might go about constructing a Lyapunov
function. If, however, one has in hand a continuous, positive definite function V
which is suspected of being a Lyapunov function at the origin for (S), it is desirable
to have analytic criteria for verifying item (11) of the definition above.
If V is a continuously differentiable, positive definite function on the cylinder
G X [6, +~), we define the derivative of V with respect to the equation (S) to be

DV(x, 1) = DiVOx, fis 1) + ++ + DVOX, Of, 1) + £VOX, 0.

If @ is a solution of (S), then

£V(6(0), 1) = DV(9(0, 1).

Thus one may say that a continuously differentiable, positive definite function V
is a Lyapunov function at the origin if DV is negative semidefinite on G X [8, +).
Notice that it is not necessary to solve the equation (S) in order to compute
DV (x, t)—this is one of the advantages of Lyapunov’s method.
328 Stability 10.5

Example 3. The total energy of the frictionless oscillator described by the equations
x =y, yy = —g), (10.5c)
xg(x) > 0, |x| < A, is given by
2 zx

Vix, y) = os + ifg(u) du.

The function Vis positive definite on the strip |x] < A, -~ < y< +m. Further,

DV(x, y) = yy + x'g(x) = —yes) + yalx) = 0.

Thus V is a Lyapunov function for the system (10.5c) and, by Lyapunov’s theorem
which we mentioned above, equilibrium is a stable state for the oscillator. ||
Theorem 10.8 (Lyapunoy). If x’ = f(x, t) has a Lyapunov function V at the
origin, then the origin is stable.

Proof. Let W be positive definite and such that W(x) < V(x, t) on G X [8, +).
Let to > Bande > Obe given and assume without loss of generality that the neighbor-
hood |x| < eisinG. Let\ = min),;—. W(x). Since Vis continuous on G X [8, +)
and V(0, t) = 0, there isa 6 = 6(€, to) < € such that Vp = V(Xo, to) < if |xo| < 6.
Along the trajectory of (Xo, fo, 4) in G, we have DV($(Xo, to, 1)) <0)» This
V($(Xo, to, t)) < d as long as $(Xo, fo, £) is in G. This implies that 7*(Xo, fo) =
+o and |6(Xo, to, t)| < € for all t > to. Otherwise, there is a first time ¢, at which
|o(Xo, to; t1)| = ¢, [hen

V((Xo, fo; £1), £1) < Vo < ¥ < W((Ko, to, 11)) < V(G(Xo, fo, £1), t1);
which is a contradiction. ||
Example 4. Consider again the system (10.5b) of Example 1. Clearly V is positive
definite on ®?. Since
DAC eS) (bebe Sa
DV is negative definite (hence semidefinite) on |x| + |y| < 1/\/2. By Theorem 10.8,
the origin is stable. ||
Example 5. If a system x’ = f(x, f) has a first integral V which is positive definite
on a neighborhood of the origin, then V serves as a Lyapunov function since
DV(x) = 0 by definition of first integral. ||
Now suppose that the solutions of (S) x’ = f(x, #) describe the state of some
physical mechanism and that the state x = 0 corresponds to the desired mode of
operation for the mechanism. It is never possible to start the device in the desired
mode of operation because of inherent mechanical inaccuracies. If the state x = 0
is stable, however, then the mechanism will operate properly provided its initial state
is sufficiently close to x = 0. From a physical point of view, the asymptotic stability
 The direct method of Lyapunoy 329

of the initial state is more desirable than its mere stability since, with passing time,
the operation of the device would tend to become optimal.
The asymptotic stability of equilibrium can be studied by Lyapunov’s method;
additional conditions must, however, be imposed on the Lyapunov functions. One
might expect, on the basis of Example 1, that the origin would be asymptotically
stable for (S) if there existed a Lyapunov function V for (S) with DV negative definite
on a neighborhood of the origin. This speculation is correct if (S) is autonomous or
if f(x, 4) is bounded on cylinders G X [8, +~) where G is a neighborhood of the
origin. It is not true in general as the following example of Massera shows.
Example 6 (Massera). The function g defined by
|
g(t) = > ] ate = k)2

is continuous and has a continuous derivative g’ given by

Pam hh.« = 2k (ak)
g(t) Dy fl + k4@¢ — k)2P2
k=1

since the indicated series converge uniformly on bounded subintervals of ®!. It is

easy to see that g(m) > 1 for each integer m > 1 and g(t) < 2+ 0%, k-*=M
for —«# <t< +o. Consider the equation
sg’),.
x =
ey) * Sams
10.5d

A general solution is
x(t) = g(tec,
where c is an arbitrary constant. Since g(m) > 1 for each integer m > 1, the solution
x(t) = 0 could not be asymptotically stable. Let us define

Vx, = aa| SF ie 2(u) au.

Since

g(t) << M?+ / g(u) du,

V(x, t) > x”. Thus V is positive definite. Moreover DV(x, t) = =x This.) as
a Lyapunov function for Eq. (10.5d) with DV negative definite, yet the origin is not
asymptotically stable. ||
Although the conclusion of the classical asymptotic stability theorem of Lyapunov
is that the origin is asymptotically stable, the hypotheses of the theorem actually
imply that the origin is uniformly asymptotically stable. We shall give the stronger
result here at the expense of a more complicated proof, but we shall still call the
theorem Lyapunov’s asymptotic stability theorem.
330 Stability 10.5

Definition. A positive definite function V which is defined on a cylinder G X

[8, t+-0), where G is a neighborhood of the origin, is called decrescent if there exists a
positive definite function U on G such that V(x, t) < U(x) for all (x, t)inG X [8, + co).
Note that V is decrescent if and only if V(x, 1) ~ 0 as x > 0 uniformly with
respect to ¢ in the interval [6, +).
We leave it as an exercise for the student to modify the proof of Theorem 10.8
so as to verify the next theorem and proceed directly to Theorem 10.10.
Theorem 10.9. If x' = £(x, t) has a decrescent Lyapunov function at the origin,
then the origin is uniformly stable.
Theorem 10.10. If x' = f(x, t) has a decrescent Lyapunov function V at the
origin with DV negative definite, then the origin is uniformly asymptotically stable.
Proof. Let € > 0 be given. Since x = 0 is uniformly stable, there is a 6 = 6(€),
0< 6<e, such that |@(Xo, to, 0| < €, to > 8, implies |@(Xo, fo, 2)| < € for
to <t< +m. Now let €;,0 < €; < 6, be given. Again since x = 0 is uniformly
stable, there is an 7 = n(€1), O < n < €1, such that |@(xo, fo, t1)| <7, t1 > B,
implies |(Xo, to, | < €; for ty <t< +a. Now take 6, in the definition of
uniform asymptotic stability to be the 6 defined above. We shall show that there
exists a T = T(€,) > O such that |xo| < 6, and fo > 6 implies that there exists a
t1, to < ty < to + T, with the property that |@(Xo, to, t1)| < . The existence of
such a f, implies that |(Xo, fo, t)| < €, for all ¢ > t,, hence for all tf> tp) + T
(Fig. 10.12).
Let W,, W2, and W3 be positive definite functions such that W(x) < V(x, 1) <
W(x), DV(x, t) < —W3(x) for all (x, ) in G X [6, +~).

Figure 10.12
 The direct method of Lyapunoy 331

Define
i)) A= min
eee W i(x), ii
il) = pee W(x),

ili) y= min Ws3(x), iv) T = (wu — X/2)/v.

NS |x| Se
Now suppose for contradiction that 7 < |xo| < 6, but that 1 < |$(Xo, fo, D| < €
TOLslo Os to Le hen

dX < W1(o(Xo, fo, t)) < V(o(Xo, to, 2) < Vixo, to) — v(t — to)
Se Xo) at lo) (Lt)
and, at ¢= to + T, we have \ < uw — vyT = X/2, a contradiction. Thus there
exists a ¢1, fo S t1 < to + T, such that |(Xo, fo, t1)| < », and the proof is
complete. ||
There are dual theorems of the Lyapunov type which give criteria for the in-
stability of equilibrium. We shall, however, leave these to the exercises and present
the instability theorem of Cetaev.
For the rest of this section, we shall consider only an autonomous equation
x’ = f(x) (A)
on a neighborhood G of the origin.
Theorem 10.11 (Cetaev). Let U be a domain in G and assume that the origin is a
boundary point of U. Suppose that there exists a continuously differentiable function
V on G such that V(x) > 0 and DV(x) > 0 on U but such that V(x) = 0 on that
part of the boundary of U which lies in G. Then the origin is unstable.
Proof. Let € > 0 be given so small that the region |x| < € is in G (Fig. 10.13) and
let x9 be a point of U with |xo| < ¢. Suppose that the path of $(Xo, fo, 4) is in U

Figure 10.13
332 Stability 10.5

and that |6(xo, fo, 1)| < € for fo < t< +m. Since V(x) > Oand DV($(x0; to, 2) >
0, V(o(Xo, to, 1) > Vixo) > 0 for to < t< +m. Thus, $(Xo, fo, £) is bounded
away from the boundary of U in the region |x| < ¢€ [why?]. Then DV($(Xo; to, 2) >
m > 0, for some mand allt > fo. It follows that

V(o(Xo, to, t)) = V(X0) + | aSa— V(xo) —+- mt — to). (10.5e)

As t— +o, the right side of inequality (10.5e) becomes unbounded. This is im-
possible since V is bounded over the region |x| < €. Thus (Xo, fo, ¢) leaves the
region |x| < € at a finite time ¢,, and the origin is therefore unstable. ||
Example 7. Consider the system
X= yxy, ey = eee (10.5f)
The origin is a center for the linear approximation; thus the stability or instability of
equilibrium is determined by the nonlinear terms. In Cetaev’s theorem, let V(x, y) =
y? — x?, let U denote the domain defined by the inequality |x| < y < 1, and let
G denote the square |x| < 1, |y| < 1 (Fig. 10.14). Then V(x, y) = 0 if y = |x|,
—Jee Xl, Vesey)> Oton Une
DV(x, y) = 2y? + 2x3y > 2p? — 2xl2p > 2y3C1 — y) > O xeon, 0
The origin is therefore unstable. Moreover, any solution path initiating in U must
leave U by crossing the line segment y = 1, -l<x<l. ||
Cetaev’s theorem for nonautonomous systems is given in Hahn [14].
As a practical matter it is not always enough to know only that an equilibrium
position for a physical system is asymptotically stable. One frequently wants to know
the size of the perturbations from which the system can regain equilibrium. Suppose,
for example, a certain attitude for an aircraft is known to be asymptotically stable.
One might ask the following question: If the nose is bumped downward by turbulence,
how many degrees of pitch can be tolerated without the pilot’s intervention to prevent
a dive?
Definition. Let x = 0 be an isolated, asymptotically stable critical point for the
autonomous equation x’ = f(x), where f is continuous and lipschitzian on a domain

><

> x

Figure 10.14
 The direct method of Lyapunoy 333

® in BR". The set S of all points xo in © such that (Xo, to, t) ~0 as t> +x is
called the region of asymptotic stability for the origin.
Example 8. The origin is an isolated, asymptotically stable critical point for the
system (10.5b) of Example 1. If @(¢) = (u(t), v(t)) is a solution vector with
|(0)| < 1/\/2, then |¢(t)| > 0 as t—> +m. Thus the region of asymptotic stability
includes the domain |x| + |y| < 1//2. ||
Notice that one can never hope to make an assessment of the size of a region of
asymptotic stability by Theorem 10.4 or similar results. Linear systems, when asymp-
totically stable at all, are globally asymptotically stable. Nonlinear systems, on the
other hand, can display equilibria that are asymptotically stable, but only locally so.
A basic result on the extent of the region of asymptotic stability is LaSalle’s theorem
which is discussed below. It can be motivated by consideration of the system
YS Sib erate BR Paull Sree (10.5g)
Va aCe abs nO ore Sas
If G = ®”, then V(x, y) = x* + y? defines a Lyapunov function V at the origin
since the system (10.5g) takes the form

f= —ra-ry, Fo=p?+@- 17 (10.5h)

in polar coordinates. A phase portrait for the system (10.5g) is shown in Fig. 10.15.
Notice that the points p, 0, q with respective coordinates (—1, 0), (0, 0), (1, 0)
are critical points. The unit circle is the w-limit set of every path in its exterior and
the a-limit set of every path in its interior. It is an invariant set consisting of four
invariant subsets: p, q, the path C, and the path D (Fig. 10.15). The critical point 0
is an invariant set as in every closed disc x* + y? < R*®, R > 0, and the open
annulus Ou x7 y- =<).

Figure 10.15
334 Stability 10.5

It is important to observe that for (x, y) in the invariant sets which are also limit
sets (the origin and the unit disc), DV(x, y) = 0. We shall use these observations to
illustrate LaSalle’s theorem before proving it.

Theorem 10.12 (LaSalle). Let B be a closed and bounded region which is invariant.
Suppose V is a continuously differentiable function with the property that DV(x) < 0
if x is in B and let Z denote the set of all points in B where DV(x) = 0. IfM denotes
the union of all the invariant sets in Z, then every solution of x' = f(x) in B approaches
Miasct Poe
To apply LaSalle’s theorem to the system (10.5g), let B denote any disc x” + y* <
R?, R > 1, and let V(x, y) = x? + py”. The set Z consists of the origin and the
unit circle. The set M is, in this case, the same as the set Z. As the conclusion of the
theorem states, any solution with x?(0) + y?(0) < R® approaches M as t—> +o.
Notice that those solutions with x?(0) + y?(0) > 1 actually approach the unit
circle, and those with x7(0) + y?(0) < 1 approach the origin. Thus the theorem
does not say which component of M is approached. The theorem is used to estimate
the region of asymptotic stability for the origin by choosing B, hence R, in such a
way that M consists precisely of the origin. That is, one chooses B to be any disc
x? + yp? < R*,0 < R < 1, and concludes that any solution in B approaches the
origin as t—> +a. Thus the disc x? + y”? < 1 is the region of asymptotic stability.
Proof (of Theorem 10.12). Let x9 be in B. Then x(t) = $(Xo, fo, f) is in B for all
t > to since B is invariant. The w-limit set 2 of @ (which is closed, nonempty, and
connected) is contained in B since B is closed.
We shall show that V is constant-valued on Q. Let p and q be points of 2 and
let {s;,} and {tf} be sequences such that x(s,) — p and x(t) > q as k — w. Since
DV(x) < 0 on B, the quantity V(x(t)) cannot increase as t—> ++». The function
V, being continuous on the closed and bounded region B, is bounded below. Thus
V(x(t)) approaches a limit L as t—> ++«. Then
V(p) = lim V(x(s,)) = lim V(x) = lim V(x(@e)) = V¢Q).
Consequently V(x) =c, say, for x in Q.
We may conclude from the last line that Q is in Z since DV(x) = 0 for x in Q.
Since Q is invariant and M is the union of all invariant sets in Z, it is also true that
Q isin M. This implies that dis (x@: M) —Oast— +o. Otherwise, there exists
ane > 0 anda sequence {t;,} such that t, > +0 ask — +. and dis (x(t), M) >
€. The sequence {x(v;,)} must contain a subsequence which converges to a point p
of B since B is a closed and bounded region. But then p is an w-limit point of @
which is not in M. This contradicts the inclusion of Qin M. ||
Example 9. If ¢ is replaced by —t in Liénard’s equation (9.9b), the phase portraits
for the resulting equation
x” — f(d)x' +x = 0 (10.5i)
 The direct method of Lyapunoy 335

are the same as those for the unmodified equation except that the time sense of the
paths is reversed. We assume, as in Theorem 9.19, that f is continuously differentiable
and even, that there exists an a > 0 such that f(x)(a — |x|) < 0 for x ¥ a, and
that ica” f(x) dx = +a. With respect to the equivalent system

x= ytk(x), 9 = —x, Fe) = [ f(u) du, (10.5j)

the origin is now an asymptotically stable critical point, and the limit cycle is unstable.
An estimate of the magnitude of the oscillation corresponding to the limit cycle
for Liénard’s equation can be made by estimating the region of asymptotic stability
since this region is the interior of the limit cycle.
We apply LaSalle’s theorem with V(x, y) = (x? + y?)/2. Note that DV(x, y) =
BUX) ee SE (SEC LEIZ-49532)=D (Xa y)ueOnit, |x| > b DViGe y).= Oat |x|. = bor
|x| = Oand DV(x,y) < Oif 0 < |x| < b. If we take the disc W/x2 + y2< R< 5b
for the compact region B of the theorem, then Z consists of the points on the y-axis
with —R < y < R. The only invariant set in Z is the set M consisting of the origin
only since x’ = y = O if and only if y = 0. Since every solution initiating in B
must approach the origin as t—» +, B is interior to the region of asymptotic
stability. Thus the limit cycle must lie in the region x? + y? > b”. For van der Pol’s
equation, F(x) = e(x — x°/3). Thus b = /3, and the limit cycle lies outside the
circle x? + yp? = 3. ||
Example 10 (LaSalle). Consider the equation x’ + ax’ + bx + x” = 0 where a
and 6 are positive constants. An equivalent system is
x v= jy, y’ / = —ay — bx — x’. (10.5k)
It follows from Theorems 9.18 and 10.4 that the origin is an asymptotically stable
focus and the point (0, —b) is a saddle point. This information, in conjunction with
the directions of the tangent vectors to the solution paths allows one to sketch the
qualitative phase portrait of Fig. 10.16.

Figure 10.16
AMG Stability 10.8

The shaded region 4 between the patha C and 2 ta the region of aaymptotic
stability, The problem here ia to estimate Ui quaaiiaiel, Tone thinks of the system
(105k) as deseribing the motion of a particle oscillating with fietion under the
influence of a nonlinear force, it is natural to study the problem with the ald of the
energy
V(x,y) = y?/2 + bx®/2 + x*/3,
The derivative DV is given by PV(yy) = —ap®. 1 A ik any compact invariant set
which has the origin in its interior but doea not contain the port (4,0), then 7
is the portion of the veaxia whieh lies in A, and M eonsixts only of the origin, Ry
LaSalle’s theorem, A is inside the region of aaymptotio stability, To estimate 8, then,
we construct a compact, invariant set # whieh does net contain (4,0) As a start,
we require that any point (vy, )) in A satay VS =a &® A Oe Be A Then
any path initiating on the hall-line v¥= =@, » & 0 will subsequently move to the
right and never reorass it (Pig, 10,17),

\ (Ay ») = sare?

y= ae eA)

Piguee 10.17

We construct a line in the third quadrant through (3,0) whieh solution paths
must cpass in the positive ydirection, For y < OQ and =& < x < Q the slopes of
solution paths satisfy
dy
/ Ae — b+
= NX —
ax or tay
Thus any path lying above the line p = —a@(y +> 8) in the Uhind Quadrant can never
enoss it there, Since DV(y,») < 0, AO path ean crass a level curve riny) = eat
the energy, The level curve V(x, ») = 4478” through the polat x = Q » = <a@
passes around the origin and intersects the line v= =@ at a point iy the soond
quadrant, The region # consisting of points (x,»)such that > =@»y > <e(e a)
Vy) S 34°8" is thus a compact, invariant set of the requived type, By LaSalle’s
theorem, 8 is contained in the region of agveptotic stability,
~
 The direct method of Lyapunoy 337

Example 11. The region of asymptotic stability for the equation

x!’ + 2\x|x’ + 2xe*? = 0 (10.51)

is the entire xy-plane. To see this, let V(x, y) = 4y? — e~**. Then DV(x, y) =
~y*|x| < 0, where x‘ = y. Thus the sets V(x, y) < c, c > —1, are closed, in-
variant sets. They are not compact, however, since they are unbounded for c > 0
(Fig. 10.18).
y

V(z, DD Uae

Figure 10.18

Let (xo, Yo) be any point in the plane and fix c > 0 so large that V(x9, yo) < c.
Further, choose a > »/2(c-+ 1) so large that (xo, yo) lies in the compact region B
defined by the inequalities V(x, y) < c, |y + F(x)| < a, F(x) = fp 2\u| du. The
path through (xo, yo) cannot cross the curve V(x, y) = ¢ since DV(x,y) < 0. If
U(x, y) = (vy + F(x))”, then DU(x, y) = —4(y + F(x))(xe~*). Thus the path
through (x9, yo) cannot cross the curves |y + F(x)| = a either. Therefore B is
invariant. The set Z of LaSalle’s theorem consists of those segments of the axes
lying in B, and the only invariant set M in Z is the origin. The path through (xo, yo)
approaches the origin, and it follows that the origin is globally asymptotically stable
since (x9, Yo) was arbitrary. ||

EXERCISE 10.5
1. Let A denote ann X n matrix with n linearly independent eigenvectors and consider the
system x’ = Ax. Under what further conditions on the matrix A is there a Lyapunov
function of the form V(x) = x7Bx?
2, Show that
t 1 2 2 rey,
V(x, x', 1) = (++) NSE 2? |bea

is a Lyapunov function for the equation x” + (¢ + I)x/f = 0,1 > 1.

 3. Abstract the analysis of Problem 2 and construct, under suitable hypotheses om 2, &
Lyapunov function for the equation x” + pix = 0.
4. Examine the possibility of using the absolute value function V@) = ix} as a Lyapeaov
function for a scalar, autonomous equation x’ = #(, —2@& <x < +e, with F(~@ = 0
if and only
if x = 0.
5. Prove the first instability theorem of Lyapunov: Let V denote a positive definite, de
crescent function on the cylinder G X [8,+=). If DV.) > 0 for all &D BR GX
[8, +), them the zero solution of x’ = fix, 2) is unstable.
6. Examine the stability of the equation x” + m/(1 + 2) = Qr> 1.
7. Abstract the analysis in Problem 6 and formulate a theorem for x” + pax = O,
t > B.
8. Prove Theorem 10.9.
9. Interpret Theorem 7.7, in the light of Lyapunov’s method,
10. Consider the autonomous system
3 3
2> sin
ox
», Ye
?
eg ¥

Use the euclidean norm as a Lyapunov function and show that the origin is asymptotically
stable.
11. Consider the autonomous equation x’ = f(x), where f is continuously differentiable
and f(x) = 0 if and only if x = 0. Let J¢{x) denote the Jacobian matrix (D)/i(x)) and
assume that there is a constant 7 > 0 such that each eigenvalue \y of J7(@ + J) satisiies
Ne < —y < 0 for K = 1,...,m. Use the ecuciidean norm as a Lyapunov function to
show that the origm is asymptotically stable.
12. Prove the second imstability theorem of Lyapunov: Let V denote a continuously dif
ferentiable function with the property that, given « > Q, there exists a point (Xa, fe) such
that |xo| < « and Vag dD > 0 for tr>m TH DV&,D = AVR) + WEA where
X > 0 is a constant and W is positive semidefinite, then the zero solution of x’ = f(x, a
is unstable
13. Deduce that the zero solution of the equation x” + (1 + + — ©)x = 0 is unstable
with the aid of the function V(x, x") = xx". [Himr> Reason as in Theorem 10.11, but take
nonautonomy into account]
14. Consider the equation x” + f(x, x’) + g{~) = 0, where f and g are continuously dift
ferentiable, yf (x,y) > 0 for y = 0 and xg(x) > 0 for x = Q Show that the origin is an
asymptotically stable critical point.
15. Prove that if one imposes the condition {)” g{(s)
ds + + as |x} + + in Problem 14,
then the origin is globally asymptotically stable.
16. Reconsider Problem 14 under the assumption that f(x, y) = ya(x, »), where A(x, ») >
7 > 0 for some constant y and all (x,y). Show that the origin is gledaily asymptotically
stable.
17. Estimate the region of asymptotic stability for the equation
F ee =
x"
++ xe + 2xe = O
 The direct method of Lyapunoy 339

18. Estimate the region of asymptotic

stability for the zero solution of the syste
m
x = —xX + 3xy? + x3. y = —y® — 2x2y 4 yd 4 Bx y?:
19. Prove the asymptotic stability theor
em of Lyapunov: If x’ = f(x, t) has a
Lyapunov function V decrescent
at the origin with DV negative definite, then
the origin is asymptotically
Stable. [Hint: Pattern your argument after
Example 1.]
mS
FURTHER READING

AT AN INTRODUCTORY LEVEL
1, Boyce, EF, B, and R, C, DiPrima, Elementary Differential Equations and Boundary Value
Problems, New York: John Wiley, 1965.
2, Brauer, F, and J, A, Nohel, Ordinary Differential Equations, A First Course, New York:
W. A, Benjamin, 1967,
3, Kreider, D, L., R. G, Kuller, and D. R. Ostberg, Elementary Differential Equations.
Reading, Mass.; Addison-Wesley, 1964.

AT AN ADVANCED LEVEL
A, Birkhoff, G., and G, C, Rota, Ordinary Differential Equations, Boston: Ginn, 1962.
5, Hartman, P., Ordinary Differential Equations, New York: John Wiley, 1964.
6, Hurewiez, W,, Lectures on Ordinary Differential Equations, Cambridge, Mass.: M_LT.
Press, 1958,
7, Lefschetz, S., Differential Equations: Geometric Theory, New York: Interscience, 1957.
8, Nemytskii, V. V., and V. V. Stepanov, Qualitative Theory of Differential Equations,
Princeton, N, J.: Princeton Univ. Press, 1960.

ON APPLICATIONS
9, Minorsky, N., Nonlinear Oscillations, New York: Van Nostrand, 1962.

ON PARTIAL DIFFERENTIAL EQUATIONS

10, Weinberger, H. F., A First Course in Partial Differential Equations, New, York:
Blaisdell, 1967,

341
342 Further reading

ON METHODS OF SOLUTION
11. Kamke, E., Differentialgleichungen Lésungsmethoden and Lésungen, Ann Arbor, Mich.:
J. W. Edwards, 1945.

ON STABILITY
12. Cesari, L., Asymptotic Behavior and Stability Problems in Ordinary Differential Equations,
New York: Academic Press, 1963.
13. Coppel, W. A., Stability and Asymptotic Behavior of Differential Equations, Boston:
D. C. Heath, 1963.
14. Hahn, W., Theory and Application of Lyapunov’s Direct Methods, Englewood Cliffs,
N. J.: Prentice-Hall, 1963.
15. LaSalle, J., and S. Lefschetz, Stability by Liapunov’s Direct Method, with Applications,
New York: Academic Press, 1961.

ON LINEAR ALGEBRA
16. Franklin, J. N., Matrix Theory, Englewood Cliffs, N. J.: Prentice-Hall, 1968.
17. Halmos, P. R., Finite Dimensional Vector Spaces, New York: Van Nostrand, 1958.

ON THE ZEROS OF POLYNOMIALS

18. Marden, Morris, The Geometry of the Zeros of a Polynomial in a Complex Variable,
New York: A.M.S., 1948.
HINTS AND ANSWERS
FOR SELECTED EXERCISES
eS
HINTS AND ANSWERS
FOR SELECTED EXERCISES

Exercise 1.1
[Link]=2+4+4142 bx=2477-44+4
c) x = sint d) x = $In({1 + x|/|1 — x|) + 2
e) x =/tan-!4-4nnd+%+3y
2. vo = —240/7 ft/sec
3,0 = —V/v2 + 2g(Ro — r), where vo and Ro denote initial velocity and altitude.
4. —88 ft/sec”, 440 Ib 5. 316.8 hp

Exercise 1.2

1. Let c denote an arbitrary constant.

2) ee Ce lo) 32 = (2 c) x = ce”?
G)eca—ner e) x = c(sec¢ + tan f) D)iea—resint
Se Cl Leetat)/ (1) 11/2 h) cll te Orr)?
2.-a) e® = 2981 b) 375 million
3. $822 rather than $791, an increase of 0.392% per yr
4. Thirty billion dollars 5. 192 hr
6. More than 10,000 ft?/min 7. 115,000 yr old

Exercise 1.3

1. Let c denote any constant.

a) x = ce”! — 3/2 b) x = ce?! + 31e?!
c) x = ce*t — 3e—7'/4 d) x = ce”! — t/2 — 1/4
2. 280 ft/sec, 8.75 sec 3. 17.5 1n 10 = 40.3 sec

343
344 Hints and answers for selected exercises

. 0.98 cm/sec, t = 0.002 In 10 ~ 0.0046 sec

» Diese = 2PM See, trom = mass - In 10
. The heavier one
. The hot water takes In (9/98)/In (9/28) ~ 2.1 times as long to reach 32° F.
WN. $57.83/mo compared with $25.00/mo
DOA
oN

Exercise 1.4

1. 11,111 m/sec
2. Yes, if vo > (2Gm,/ro)'/?, then r’ > k/r!/?, where k = V/2Gm,.. Thus r(t) >
[ro*/* + 3k(t — t0)/2]?/* > + as t> +o.
3. 200 km
4. Use (1.4d).
5. v(t) = vo + k In (m(0)/m()). The exhaust velocity should be as high as possible,
and as much of the rocket’s mass as possible should be fuel.

Exercise 1.6

1x TCOSIOL
2. x = w! sin wt. One might strike the block while it is at rest in the equilibrium position,
or one might compress the spring and release it, letting t = 0 denote the first time at which
the block passes from left to right through equilibrium.
3. Both functions describe motions of the block which initiate in the same way. If the
motions were different, then the physical hypothesis that like causes produce like effects
would be violated. Thus the motions are the same and the functions must coincide.
4. Sete, = a/2\/B and c; = 1 in (1.6b). Pull the block one unit to the right and release it.
5. 2r/\/4w2 — a2 sec

Exercise 1.7

1. mw” = kol(y — w), may” = —koly — w)

2. mw" = —kiw + koly — w), moy” = —kely — w) — kay
3. Mx” = —ky + 0)’ — x’), my” = —d(y’ — x’), where x and y are the displacements
of the plate and block and 2 is the coefficient of friction.

Exercise 1.8

1. Ki) = (1 — e-P/2)E/R
2. I(t) = Ee-#®C/R
3. I(t) = (.02)e—599% sin 5000¢
4. I(t) = EVC/L sin (t/\/LC), frequency = 1/2r/LC
 Hints and answers for selected exercises 345

Exercise 1.9

1. Let t+ 0+ in (1.9a), noting that y(0) = E/R.

2. RLCx” + Lx' + Rx = 0
3. Let 8 = —R-*C-2 + 4L-1C—! and a = —1/RC.
Xe a le Si (4/6.5)2). cifianbe 0,
Ax = 3(RL = R-!C1)x 3. RE—1y, 9’ = —RL-1x = 2RL—1ly, x(0) = —E/2R,
yO) = 0

Exercise 1.10

1. a > RC/M
2. I(t) = CV;(1). Show that the derivative of a periodic function is periodic.
3. Wind a third coil on the same form as the other two coils. The voltage across the third
coil will be proportional to V,.

Nee ete(5 1) sy”

ds

Exercise 2.1

1. (a), (6), (d), (©), (g)

2. (a), (c), (e)
3. —8(2g)1/2
4. r(t) = 4t — 7

Exercise 2.2

2. a) The disc x? + 1? < 1

b) The hyperbolic region x? — 1? < 1
c) The cylinder x? + y2?<1, -~ <t< +o
d) The wedge-shaped solid |)| < x, ¢ > 0
3. The projection of the locus onto the xy-plane is the locus of y = 1 + x?.

Exercise 2.3. Let a, b, c, A, B, C denote arbitrary real constants unless otherwise designated.
1. a) x = +1 and x = sin(t+oc) b) x = tan {tan—*
7-1 c)
c) x = 0 and x = (#2/9 + c) d) x = cexp (-1/(1 + 9)
2 R—a 2 2
Phy DE = oe
i5a2 aoe
Be (3R° + 4Ra + 8 8a’)
3. T = R2\/2gh/ga, where R = radius of tank, h = height, and a = radius of vertical
pipe.
4. x = C~!(-1 + cosh (Cd)
346 Hints and answers for selected exercises

5. x = 3400(1 — 3.4/(2400kr + 3.4)), k = 3.4c/(2400)(3400 — c). It would take an

infinitely long time—theoretically—for the reaction to go to completion. How long would
it take to produce 40 kg ten times?
6. W() = a — Va?
— 12, 0 < |t| < a, semicircles
7. Let Vo denote the initial velocity of the boat and let & denote the coefficient of friction.
a) x’ = (—ket/M + V6)1/. The boat stops when t = MV6/ke.
b) x’ =-Voe *™ c) x’ = 1/(ekt/M + Vo)!

Exercise 2.4
s 1/2
2 =] Uu
1. (r’)
(r’)” — 2Gm,
Gm.r =c,t=
Cut ae a Se rs7)
to + ib&& du

2.a) x? 4+ =€c b) x +sinx

— t?7 =c
ejizvi=f (ix) dee di O0/)4 — kex=* =e
e) (x’)}? — 2cosx =c f) (x)? +. (Qkx?2*?/@-4 e?)) =e
3), 32 = 1)
4. x = cye*' + coe~*, cy and cg arbitrary

Exercise 2.5

loa) xe tr = b) tx +t =€
ec) x*+x+rf=c d) x?¢ + 42/3
=c
e) Not exact f) x7/2+ 4x4 #4/4=c
g)x+sinxt=c h) Not exact
i) x8/3 +xt4+-x4+7/34+t=c j) Not exact

Exercise 2.6

1.x = 4-* In (e/h)

2. a) x or xe b) e* c) te* or tx d) et e) sinx f) te™

Exercise 2.7. Let a and c denote real constants.

x= QF 1)?
. xX = tant
nk =COS, (tel 72)
= x°/3 + x7/2 +x
~x=a-—(t+oInijt+cl+(t+o,t+e>0
x’27/2—x-l=¢
~ x’ + In|x’ —1] = x7/2+¢
. x'?/2 + fi f(u)du = c
. x’ sinx’ + cosx’ + x2/2=c
 Hints and answers for selected exercises 347

10. x3 + x2/2=c
11. In|jx’ + 1] 4+- @’ + 1)7!4 x2/2=c
12eAine? <4) ine —11) $8 3e2720are
fin 1) ee ete
14. In|x’? — 1] + x? =¢
15. x4/4 + x’ + In|x’ — 1| I fey
16. x?/3 + cx +a=2t
17. x’ —In|x’ +1, + x4*/4=c
Ise — e.-x*/4 = c

Exercise 2.8

1. Plot y? + x? + x4/2 = 2E.

2. Plot y? + 3x4/3/2 = 2E.
3. Plot y2 — e-*? = 2E.

Exercise 3.1

1. x = Ae**/3 — Be**/2 4+ Ce!

= — Ae?! + Be!
a Ae2t

= —(3At + A)e2*/9 — Be?!/3 + Ce-5!

= Ate** + Be?!
= Ae**
= Ae?*/9 + Bre®' + Ce
= — Ae?'/3 + Be
|— Ae2*

= —Ate?t + Beit + Cet

= —Ae*t + Ber!
a =
tae
See
IS)
tat
Sie
ad
SS Ae
= (Ar?/2 + Bt + Che!
l| (At + B)e*

Ces
Ne = Ae?t

Exercise 3.2

PARTE M1 bal ae ET /2, 0]? 2. (37, 7, 2x)”

39 (9.93; Gl" Aol 2 Stole
5. [14, 34, 24]” 6. [0, 0, 0]”
I 3e = By) Sesh, = 7 See eee
348 Hints and answers for selected exercises

9, x = Ae?!/3 — Be3!/2-+ Cet, y = Bet — Ae*!, z= Ae?!

10. Same as 9
11. x = Ce3! + Bre?! + Ae, y = Be?! — Ae*', z = Ae”!

Exercise 3.3

0 0) —9 2 —3i
ib ||3! 7, \\Ml 3h. 6 4.}1 + 2i
3 1 3 i

De PEAY) i @ 4 6 8 -9
SO 4% 2 6 ib 7.{—1-—2 6
0 @ 3 iL @ 02 20: Fes.

i —1+4i 2 —-3i 0)Oil 0 0 O

Sean) —i 14 2i C1 @ @ LOMO ORO
0 0 i OOO OPFOR 0
xi Deen ONE: a6 BD Bye FEI Be
ti. el] = ||O SB Al sy 125) sy oe easy,
Zz (0) Oh WAL zZ Qi P| \\|$7
xe lV G11 @i2 a13|) x1
13. |xo] =|@21 a22 a23)\\xe
x3 a31 a32 a33\| x3
14. a) x = Ce®* — (B/3 + A/9)e?* — Ate?*/3 b) x = Ce* + Bre
y = Be** + Ate? = Be
= Ae*! z= Aew*
oy 0) 2 iV Nie AV)
54/0) 2) @ Gs ||) 72 Ie NOY 2 4
0 0 9 oo © 2 OL Oe
18. x = Ae 19. x = Ae?** + Bre?! 20. x = (A + Bt-+ Crt?/2)e*
y = Be! y = Be*# y = Bet + Cre”
z = Ce! z = Ce! z = Cert

Exercise 3.4
_ [t?e2¢ ab te?!, 2te2* aL e*t, 2e2¢]?

(2x + y, 3x7, 17
. [tan-1¢, 4, n@Q + V1
4+ 2)"
2 10,00]F
» fet seh Jey”
. Any vector of the form [cos ¢+ csin+¢, —sin t]’, where c is arbitrary
5 (byes sree
. Any vector of the form [cosh ¢ + c sinh ¢, sinh ¢]”, where c is arbitrary
© ‘ [Se7'/4 =— e3t/4, eftyr
Nn
=
NHN
WwW
fh
CONN

a toe. [21324 + et, 2e2* — et, et]?

 Hints and answers for selected exercises 349

Exercise 3.5

i @ 1 i &
ia) i aE E |+ E le

10 0 0.4 2 6 4
DOs LO DEaD 7a. | li omesel72
OF004 3 at Les
2t 0 eit 0 1 t
Deal |
a) |0 A b) ki ne c) e a ke 1

oye 0) it @): @)
Gyreuen BS) CPHl@d ib © i) CeO il %
@ @) i QO @ il

[2 Pot?/2 Log B00 PE ae Whore

geO 1 4 hy ert| 0 tO. 0 gen
|Unwae 0
lo Ow et C=O lee} Oe Toea
0) Oe iI OOO
(IN yA LV GAG) ee 2 ee 6
s wt Ont cp 0 wt O° tf 12/2
PO 0 i. 0 ies omon cre oe
@) © 0 1 0 0 O 1

3. a) x = Ae*!, y = Bett, z = Cet

b) x = Ae®! + Bred, y = Be®!, z = Cet
c) x = Ae? y = Be** + Cre**, z = Ce*t

De (4 Br Cr )e7! yy = (B-. Cie”, z = Cor

Exercise 3.6

1.8 23 ea 4a
Spgs O89 O, 3), 35.30) AZyPne 5, 45 Ph 2
9. 6,3, —1 10. 1, 2 + 3i, 2 — 3i

Exercise 3.7

1. For j = 1, 2, 3,4, let c; denote an arbitrary constant.

a) x1 = (c1 + Scot)e*, x2 = coe*!
By = Cicer sete exe (91 = C9)e°t/9 — ente""/3
c) x1 = (e1 + c2/4)e?! + cate’, x2 = (c1 — c2/4)e?* + cate!
d) x1 = (c1 + c2)e* + cote*', x2 = (—e1 + ca)e* — cote
e) x, and xg are the same as in part (b); x3 = c3e?
f) x1-= 2c1e** + cee + cse—*
xo to = cje2* + (1 + dDege* + 1 — dese
x3 = (-2 4+ Deve + (—2 — dceze—*
350 Hints and answers for selected exercises

g) x1 = (c1 — co + teie®!, x2 = (—2c1 + 3e2 — 2cane* — c3e°!

x3 = (—2¢1/3 + °c2/3 — 2c2t/3)e3* + 2c3e%'/3
h) x1 = —2cye2* + (1 — 2iese?* + (—0? + 4+ 3/2)e3e7'
xe = 3cje2¢ + (—1 + 3n)c2e2* + (3142/2 — t — 5/2)c3e”!
x3 = —cye2t — tege?* + (—1?/2 + 3/2)e3e7*
. a) Use l’Hospital’s rule.
b) If and only if each eigenvalue of A has negative real part
Wee ert 2 3 Den 2
x =1— 72/2, y = e-7t, z = —e-4/2 + 3e—V2 4/2
~x = (cr + cathe—@/?
3m, 4m to the left of center
. w= cye* + coe—* + cze* + cae—*
x = cre — coie~* + c33ie?* — c43ie—3"
y = 3cyet*/2 + 3c2e—*#/2 — c3e3**/2 — cae—3/2
z = 3czie/2 — 3ceie~**/2 + 3c3ie3"/2 — 3c4ie—3**/2
w(t) = (—Scos 1/6 — (cos 3f)/2, y(t) = (—Scos t)/4 + (cos 32)/4; 2m sec

, a) x= 1501e 098 yp = (0.6:— 300De7 9%

b) m= Q(e—500¢/3 — e— 15002) 80, y= (—e 5008/3 aie Qe—15004) /g

¢) x = de—125003 sin (12502/3), y = se + 7°9/*[cos (12507/3) — sin (22501/3)]

. If the coordinates of Problem 1, Exercise 1.7 are used, w = cos? ¢, y = sin? ¢. The
length of the spring is L — cos 2¢ at each instant f¢.

w= BOSE T 1k SID E51 COS I =e

eye

ee = 1
4 +4sin
= $cos¢ — :4cosV3t
t ++ —<si
574 sin v3 t

A sum of two periodic functions is not periodic unless the ratio of the periods is a rational
number.
Periodic motions occur when w(0) = +y(0), w’(0) = +y’(0).

Exercise 3.8. For j = 1, 2, 3, 4, let c; denote an arbitrary constant.

ib x = cye—* + coe3! D. x = (c1 + cathe
x = c1sin (V3 1) + ce cos (V3 2) 4. x = cye~* + cote
x = (cr + cot + c3t?)et 6. x = (c1 + cot + c3t?)e*
es el Se vr— COSIZ
nwx
oN = (c1 + cet) sint + (c3 + e4f) cost
xX =c1 + coe" + czte—*#
x = c¢c1 + (c2 + cole”? + c4e—3!
x = cyeV3! + eoe—V3 + cet cos 2t + caet sin 2t
 Hints and answers for selected exercises 351

13. x = cye—3* + cote—3! + cze# + cye-it

14. x = c1 + coet*/® 4+ eze—V/5
15. x = cye—** + cote—2* + c3e—3# +. eget
16. Let a = —R/2L and 6? = (R2C2 — 4L0)/4L7C2.
I(t) = Ee~**(e5t — e—8)/28R if B XO, HO) =I he fs} = ©.
17. t = \/L/g cosh~! (2), where g is the acceleration of gravity.
18. Solve the initial value problem x” = 2gx/L — g, x(0) = 2, x’(0) = 0.

Exercise 3.9. Let $(t) = ci sint + c2 cos ¢, where c, and ce are arbitrary constants.
l1.x=¢()+1 2.x=¢o)
+1
3.x=¢o() +1+1 4 x=¢)4+14+2t4+7
5. X = g(t) + e'/2 6. x = o(t) + e#/5
Tepe OC) e(t —) 1)/2 8. x = ¢() — 24 #2 +4 3e!/2
9. x = o(t) — (sin 21)/3 10. x = g(t) — 3¢t(cos £)/2
11. x = g(t) — c(sin 21)/3 — 4(cos 21)/9 12. x = ¢(f) + t — tcos 1/2
13. x = d(4) + (et sint — 2e cos £)/5 14. x = (4 — (e?! sin 3t + 3e7! cos 32)/40
15. x = o(t) + e[14 — 102) cos t + (St — 2) sin 1]/25
16. x = ¢() + ¢ + [e’sint — 2e' cos t]/5
17. x = cye~* + cote + cze—* 4+ t2e-*/6
18. x = cye~* + cate + cze—** + (42 — £?)e—*/18
19. x = cye~* + cote? + cz3e—*! + (#2 — t?)e~*/6 + Te—*/9
20. x = ph/w + (a — ph/w) cos (t\/ wg/ph)
One could determine the period 7 of oscillation and the radius r by observation. Then
ph = wgT?/4nr?. The weight of the buoy is mr?hpg.
21. T= c1 exp (—R/2L + RV14L/R2C/2L)
— + ce exp (—R/2L — RvV/1 — 4L/R2C/2L)
wEo
sin (wt + 9),
+ V(Ra® + (C71 — La
where tang = (C-! — Lw?)/Rw. The light burns most brightly when L = 1/w?C.
22. The solution becomes unbounded as t > +.
23. w = Vk/m

Exercise 3.10

1. Linearly independent 2. Linearly dependent 3. Linearly independent

4. Linearly independent 5. Linearly dependent 6. Linearly independent

Exercise 3.11. Let & denote any constant.

1. [—3k/2, 3k, k]* 2. No solution
3. Any vector [x1, x2, x3]" such that 4x; = 1 — 9x2 — 13x3
352 Hints and answers for selected exercises

Ake shoe 214 5. [1 — 2k, —1 + 3k, —k]”

Tee
6. [k, —2k, —2k/3]” Te 2 ens
—§8 —3 3
{-—1, 50 ee) ee
Sa ese 9.14 9 13 102 4-i3e8
E 3 i :6 8 yaa san

Exercise 3.12. Let a, 6, c denote any nonzero constants.

3. 8. xa{1 Ulee l Dles bl”
Ae allel ie Olle
5. 2 + 3%, all, i]; 2— 3%, df, —i]?
6. 6, afl, —2, —2]7; 3, b[—1,3,4]7; 9, c[0, —1, 217
7. 30, a[—1,2,—2]7;_ 3, ci{[—1, 3, 4]? + co[0, —1, 2]” for any c1, ce with c? + c} ¥ 0
8. 2, e116, —7, 3]? + coll, 1, —1]? for any ci, ce with c? + 3 #0
eal One Lie
10. 3, afl, —2, —2]7; 6, b[0, —1,2]7
1 2, ae OF rebel ee cl ee
12a[te ree ee ae a
13. Three through seven and eleven

Exercise 3.13

1. The answers are not unique.

a 5 ee
3. A has distinct eigenvalues. Its Jordan forms are diagonal matrices.
x = Ae?* + Ber + Ce7*/20, y = Bett + Ce?t/4, z= 'Ce%*
3 i ©
Ab ALS JD) = f 3 0} and letP denote any nonsingular 3 X 3 matrix. Define A = PDP}.
0 @ 7
5. The answers are not unique.
Oaoant 2 1 1 1-1 0
fo 0 oI l+i 1-i c)}—-2 3-1
CSO) 0 -2+i -2-i —-2 2 2
—2 1 38
d) 1 =|
-1 0 2

Exercise 4.2

1. It is necessary and sufficient that a and b not lie in different quadrants.

2. It is a hexagon.
 Hints and answers for selected exercises 353

3. |A|- |x| = 6(1 + 7) for (a), 12 for (b), and (14 + /2)(2 + mw) for (c).
4. Hint: A continuous real valued function on a closed and bounded interval attains its
maximum on that interval.
8. a) x(t) < x(O)e?+ b) x() < x(O)e”
c) x(t) < x(0) exp (1 — cos A) d) x(t) < x(0) exp JO a(s) ds
9. Hint: Set x(t) = A + Bf u(s) ds. Then proceed as in Problem 8.
10. (4) = —e-‘ cos (¢ — 1/4), T = 37/4, no.
11. y@) = (e=" — es3t427)7/3
12. Set ¢ = T + 1 in answer 11 to obtain e~7(1 — e—)/3.
13. Hint: In each identity, let d(7) denote the left side and Y(#) the right. Show that x =
o(t) and x = Y(*) are both solutions to some one initial value problem.

Exercise 4.3

1. a) 1,1 + at, 1 + a2t?/2; e b) 1 + F(t) + F2()/2; exp (F(t)

c) 0, —t, —t — 7/2; 1 — e~ ral) besea We
5 He [ eat 1 += 27 4-27? \; Bl
11 Te 3 37 97772 e3!
iyed de lo 722° cosh7
1 1+ 4: 1+ 4¢ + 1547/2). [QQ + de
8) Hi i dtl ie= eee | e3!
h) aI: 3 aE a : SE ORE ze ; pe a8 oe
4 4+ 241 4+ 24¢ + 5422]? |4e3¢ + 12163
2. a) M, = 2/k? b) M, = 2/k? c) My = 3 + 2k)/2*+!
d) M, = (1/2* + 1/2?* + 1/ek?)
3. Let M; = |ao|/k!
4. o(f) = et — 1. It satisfies several; one is x’ = x + 1, x(0) = 0.
6. x = t. No, Theorem 4.2 says nothing about points of discontinuity.

Exercise 4.4

15a) x. = 2(et — 1) b) x = (e” — 1)/2

c) x = (1 + Det x= —2
e) x = e(1 + sind f) x = cosz(1 — In (cos A)
g) x = (t — 7/4)(sec t + tan“)
h) x = Ge” — 1)/2, +> 0, x = (e-* + 1)/2, 4 <0
i) x = (et + sint — cos t)/2 )) = 1 — e?

2. x = cexp (—cos/), c#0

3. x = cexp (t/2 — (sin 2/)/4)
4. x = —(sint + cos 1/2. If and ¢ are solutions, then ¥(1) — Y(t) = ce’. They could
not both be periodic unless c = 0.
354 Hints and answers for selected exercises

5. Hint: This is really a question of deciding whether or not there is a constant w # 0

such that
d t+w t
— & i e° sin’ sds — / e’ sin’ s | = 0.
dt 0 0

7. b(t) = b(t + 2m) and {37 b(De** dt = 0

9. x = ce?! + e?!{§e—8 b(s) ds, c constant
10. All solutions approach zero as tf +.
11. All solutions are asymptotic to x = (sint — cos f)/2 ast +.
12 Yes
13. There are times ¢, where t, > ++» ask — +, such that I’(t,) = 0. When I’(t) =
0, U(r) = Eo(sin wt,)/R(2 — 3-). A bound for all such J(t,)’s is Eo/2R.

Exercise 4.5

1. e 3! e 2S Cm ea
3. ae we Ave fe fe
Sab COS, ea Sint Ch 2 Siig CrreeSic, Sins weor

| 8 e2t Stet 9 e' cos ¢ e'sin ¢

|
et ett
|
—e-2t 3 ett : 0 z e Qt | easee' sin
e ¢ ; e' t cost

10. efi, 1, 1}7, eC PV21, =1 4/2, 3 — 2V/2, e191," —1 47, 3

hut: t= 12) ta ta lind
13. sin Una), cos (In 7) 14. et cost + e'sin t)/5, e~*
15. sec r, sint + ¢sect 16. e, te
17. t2evt, r—levi
18. (1 —9)/Q4+9, 4470 —)4+40 -—d)iInd — )J//d+90
19. sint-*< cost 20. sin t?, cos £? 21. Use Abel’s formula.

Exercise 4.6

1a 2e—t — e—3t 2. ee 2e 7 4 Set

Soe e—3t 4 Ate—t 4.x= (t+ Pe
Dee Ca COSt
OOxX e 'sint + 2e~‘cost+ 2sint — 2cost
f fate e-2t ate ett, y= —e—2t ae ett

Sa 2e** + Ste2!, y = e#
92x 3e' sin ¢ + 2e' cost, y = 3e' cost — 2e! sin ¢
10. u (—2e! + (3 + 2V2el—1+VDE + (3 — 2W2el-1-VD/4, v = Layee Ud

cB heee'g 37-1 — ¢-# 12. x =2-* eet

IS 3 sinInt¢ + 2cos Int 14. x = 2e' cost + e‘sint
 Hints and answers for selected exercises 355

15. x =sint+ (1+ dsect 16. x = 4te”

17. x = fev! — t—levi 1I8.x=(1-—A)/1+9
19. x = sin t—! + 4r-? cos 17!
20. All solutions to the differential equation are of the form c, cos (t2) + ce sin (¢?). The
first initial value problem has no solution. The second problem has infinitely many solutions
x = cos (¢?) + c2 sin (t?), ce arbitrary.
JE BY 8
if Ss |
JA E |: = the ©)

23. x = Ee-** cosh tV3, y = Ee~*'sinhrV/3, at

t = 3-1/2 tanh—! (4/3/2) = 12-1/2 In 7 + 44/3)
24. w = nr, n a nonzero integer
Ap) tS () Bal yy SC

Exercise 4.7

1. x = (sin¢ + (2 — Acosr)/2
2.x = sint+¢sint-+ (cos/Incos¢, |t| < 1/2
82x = 277 — 272 + 7472
AY x = (cost + +-"1-! cos 1—1)/2
5. x = ef + e74 y = e?! — et
6. x = e — t/3 — 1/9, y = —1/3 — 1/9 + 10e%*/9
Cn er
8 x=t+14+AnNnd+9)y=14+014+)mn0+9, |e <1
9 x=1t+intzy=r—t? +¢lnt,z=t+tInt
10. x = 4/3 — t, y= 2/3 4+ 2¢—-—1,2=t7
+2
11. x = —sint+cost—1, + <0; x = —sint—cost+1, ¢+>0
1D se] Sine
Sie WSSs ae x = (1 —7/2)sint+
cost, t>7/2

Exercise 4.8

1. x = cye?! + coe**, y = —c2e”!

2. x = 3c1e24 + ce", y = —2cye** + coe?!
3. x = c1e8! — 2cve', y = cye®8! + S5c2e!
4.x =c,cost+ cesint, y = —c1sint+ c2cost
5. x = e*“(c, cost + cosin’), y = e*'(—c1 sint + c2 cos 2)
6. x = (4e-* — e—*4)/3
easly
8.x = V3e-? sin (1/3 t/2)
9. x = (et — cos V2t + (sin V2 d)/V/2)/3
356 Hints and answers for selected exercises

10. x = —e3! + e8t, y = 3e3t — e% — 2e6t, z = (e3! 4+ 2e9% — 2¢€%)/3

11. x = €3% — 2e3t y = '—2¢e30t 4 ge3t, z = 2¢e3% 4 23!
12. Let x and y denote the displacements of the plate and block respectively. The eigen-
solutions are obtained by setting equal to zero any three, but not four, of the constants
A, B, C, Din x = —Ae~*+-+ B — Ce-8', y= —2e7*! + Cer 4D.
13. Set any three, but not four, of the constants A, B, C, D in

w= Aett + Bet + Cert ait De-3*,

My l| 3Aett/2 + 3Be-*/2 — Cett/2 — De—3*/2

equal to zero.
14. Set any three, but not four, of the constants A, B, C, D in

w= A+ Bt+ Ce? — Ce~%#t,

y =A+ Bt — Cert — Ce-*t

equal to zero.
15. Set any three, but not four, of the constants A, B, C, D in

Aiet — Bie~* + CieV3** — Cie V3,

y = Aie* — Bie~* — CieV3#t + Die—V3it
equal to zero.
16. a) [cos 7, —sin 7]”, [sin z, cos z]”
b) e'[cos ¢, —sin ¢]”, e'[sin t, cos ¢]?
c) e®[1, 1)", e?[¢ + 1, ¢ — 1]?
d) e7*[1, 1, 2]7, ev", —14-+/20])7,. 2-24, =1— 4/7, 0
sin ¢ COs f d+ Asin t (1 + acost
ae cos t et t ae (1 Te) cos ¢ —( + 2)sin t
0 0 sin ¢ ‘ cos ¢
0 0) cos ¢ —sin ¢
1 t ie
17./1 1+¢f foeOr het
12+7¢ 24+4¢+4+2

Exercise 5.1

1. a) [x’(0) + xO)A + 4]/A + 3)A + 1)

b) Lx’) + x) + 2)]/A + 1)?
c) [x'"(0) + A + 5)x'O) + A? + 5A + 7x(0)I/A3 + 5rA2 + TA + 3)
d) [x’’(0) + A + 3)x’O) + A? + 3d + 3)x(0)I/A3 + 32 + 3A + 1)
e) [x’(0) + A + 2)x(O)J/Q? + 2d + 2)
f) [x'"(O) + QO + 2)x'"(O) + CA? + 2d + 3)x’(0) + (A3 + 2A2 + 3d + 2)x(0)]/
QA 23 32 ED)
 Hints and answers for selected exercises 357

roe Sa= a1 plunge 23 les

3 \ — 1]] yO)

mees
My ad
—9]7o*
fees:
ee
bo11210)
|e)
Dee Fes
7z74+A—3 A+1 = 1 |[ uO
j) QAP +d? — 314+ 1972 —1 27 +r A FT] vO)
—h 3\— 1 2 |] w(O)
2. a) 1/r2 61/021) Oa 1
d) 3/(A? — 2d + 10) e) 1/(2 — 6d+ 10) f) (A — 2)/Q2 — 4\4+ 13)
g) (A — 3)/(? — 6 + 13) h) 1/AQ2— 2A-—2) i) IMA— 2
Di Qe 2/O= 3)00= 1), «ky Or D/A = 3+ 1) 2 2/XA7 + 4)
m) 2/(\ — 4)(A? — 84+ 20) n) 1/(1 — A)? 0) 2e~(sinh bd)/X
p) emer + e-®% — 2)/Br2 q) 1/M1 — e>)
3. Assume the contrary and use |’Hospital’s rule.
4. If |f()| < Me™ for 0 < t < +0, then |L&f(A)| < M/( — 7») for Red > 7.
5. Set u = at in fo'? f(ade—!
dt.
7. w—1 tan—!(r/d) = 1—! (r/2 — tan—! (/n))

Exercise 5.2

il, A) Cesily b) e~‘ cos ¢ c) e ‘3cost — sinf)

d) e?' sinh ¢ e) 2e3! — et fjo2-* sin 27
g) 2cos 2t — 27! sin 2r h) (3e2* + 5e-24)/4 1) e ‘(1 — cos f)
j) e ‘11 + 2 cos f) k) (sinh ¢ — sin f)/2 1) (tsin 4)/2
m) e¢ Re eovt = Sin iE
1, i) es b) n!/( — ct!

c) SsA + iw)” “A — iv)"

Teroenerte.

Wy)arn 22 (\ + iw)” "A — iw)*(-1)"

e) —6e(A2 + 2 — e — e)/A4
b) x= Bier a2e>' Saxe 4te-— en
Sy A)ex Seen re,
é) x = e cost
d) x = te“{l + 2)
f) x = e-*sint + 2e* cost + 2sint — 2 cost
g) x = e~7# + eff, y = ett — e-#! h) x = Ste*! + 2e74 y =e
iN) 2% 3e' sin t + 2e' cost, y = —2e'sint + 3e' cost
358 Hints and answers for selected exercises

j) u = (—2et + BF 2 Del +VB + (3 — 2/2)e—1-V?!)/4, v = u’, w= Ul

4. a) te—?# b) 6—!¢ sin 3¢ c) —tcos 3t
d) e‘ cos 2r e) efcos 2t — 2-1etsin2¢ f) —2712? sin? t

Exercise 5.3

1. a) A(t — 1)e—¢— sin (t — 1) b) A(t — 2)e—“— cos (t — 2)

c) A(t — 1)2—! sin (2t — 2) d) A(t — 2)2—}[sinh (¢ — 2) — sin (¢ — 2)]
e) A(t — Dil + 4+ (@ — Ze] f) A(t — 10)5~!e—2(¢—-19 sin 5(¢ — 10)
2. hd — e)/r
3. —m2e"/\ + m(1 — e”)/d2
4. a) 1/(? + 1) b) (1 + A2)—! coth (7/2)
C)aHt/ NC em) d) m/d\2 — h/de"t’™ — 1)
5. a) B has period land b(f) = e 4, O<7t <I.
b) 5 has period 2 and b(t) = 2—1t, 0<t <2.
c)) byhas period
2 and’}(7)i =. 180 << b(t) — Ot le ee
G)rbrhas periods2;and’)(1) ist 1s One —aleyD(7) — Oiler an pean
e) b has period 4 and b(t) = Ay(t) — 2H3(t) + As(f) for 0 < t < 4.
f) Let f() have period
4 and satisfy f() = 1for0 < ¢t < 2, f(t) = —l1for2 < t < 4.
Then f(A) = \~! tanh (A). The required function b is given by

t—1

b@) = He — Deo / f(s) ds.

0

6. a) (1 — e—34)/3 b) 1 — cost c) t— sint d) 1 — cost — $rsint

e) H(t — 3)[1 — cos (t — 3) — 4(¢ — 3) sin (¢ — 3)]
f) The inverse f has period 4 and satisfies f(t) = t for 0 < t < 4.
g)1—te*—e h) 3 — ¢t — 3e~—* — 2te~*

Exercise 5.4

1. For parts (a) through (j), there is given a particular solution of the equation.
a)x=1 b)x =f c) x=1-+% dx=1+254+ 72 e) x = e/2
f) x = e'(t — 1)/2 g) x = —24 2? + 3et/2 h) x = —(sin 29/3
i) x = (e'sint — 2e* cos f)/5 J) x =t+ (et sint — 2e' cos 1/5
2. There is a solution of period T which is given by

— T/2
e
XG =) eee
i or ° EF OLS 2

tet?
50) I N by “Toh OES 120s ye
 Hints and answers for selected exercises 359

3. Let x denote the displacement of the block from the position on top of the cart that it
occupies when the spring is unstressed. Then x = 1 — cost for 0 <4<1 and x =
—cos f 4- cos (¢ — 1) fon t= 1.

RE a)
4. £LV(A) =
L Q2 + w2)Q2 + RC + R/L)
coth (Am/2w)
+ R(RCX(0) — drAy(0))/” + ROA + R/L)
Exercise 6.2

1. a) crepe = (ce + kK + Merz )/(K + DK +2), k > 0,

x= f+ 72/6 + 4/12 + 712/120 + ---.
b) cee = —(ce + KK + Deri Vk + DK+2, k>0,
x = (t— 1) — (@t — 1939/6
+ (@t — 194/12— (t — 199/244 ---.
2. a) It is infinite. b) One
3. Use the ratio test on the recursion formula for Problem 1(b).
4. x = o%_o cet", where co and c, are arbitrary, and
a) co = 0, cry = —bex_i/(k + 2)k +1), k = 1,
b) chpo = — CK = bc Kit 2K + 1), ke 0,
C) C2 = °** = Cm41 = 0, Cramg2 = —(ak + Dlee/(kK + m+ 2k +m + I),
k>0.
5. Let co and c; be arbitrary. For |t| < 1,

a ae (1Sere ey cont) +c (:Sy ete ey cam tst™*) :

where
Clm = —C2m—2[1 — (2m — 2)(2m — 3)]/(2m)(2m — 1)
and
Com41 = —Com—1[1 — (2m — 1)Qm — 2)]/2m + 1m).

6. Solutions are of the form x = >-?-9 cxt*, where cry2 = —Y—A)O+K + 1)ex/
Geek Dp If k =v, then ci42 = 0. Thus cyi2, = 0 for all m = 1, 2,°° ;
SU aS 9 MR RC heclr che Sal A GD I
7. Represent e’ by its Maclaurin series. Then x = >°¥-0 cxt*, where co = ce = 0,
eq = 1, and cro = —[1/K — 1)! co/(k — 2)! + oe Hepa /1N/K + 2K + YD,
le S21,
8. In each part, an arbitrary matrix solution is given by X = > 20 Crt, where Co is
arbitrary, Cy = ACo, (kK + 1)Ci41 = AC, + BCy_1, k => 1, and A, B are as given below.

a)A=0, B= i At X = Yo%29 02"B"Co/2"m!

0 0 Ona _{o 0 =[P4]
vy 4=|i Alt =|; | )4=[i de 6
Ceti 7. B=|) ql
360 Hints and answers for selected exercises

|
0 1
9. X = o_o B™Cot?”/2"m!, Co arbitrary and B = F 0
1 O TiS)
a)

10. x = fe exp (—s?) ds = t — 13/6 + 15/40 — 17/336 +---,

y= = 1/240) 10/300 eee
Exercise 6.4. Let c; and cz denote constants.

1a) x= 1 '*G Ficoll

b) x = cit + cot Int
c) x = cf tV5/2 4 ep¢l—y5)/2
d) x = Vt(c1 cos (3/2 In V1) + c2 sin (\/3/2 In V1)
é) x = (c1 cos (372 In V/t) + c2 sin (3/2 In 1))/Vt
1
2A) Xe ah a8ere + c2| Jo
—-1-—/2
b)ix =e |e + iei

Rea t sinInt a tcos Int

tcos Int —?f sinInt
= sin In ¢ t~! cos Int
d)ixe—=ie
—1 cos In¢ —t—! sinInt
a -[sin \/3/2 Int -| cos 3/2 Int
ear irvf /3/2 =]
In cack vi| Se /3/2 In |

Exercise 6.5

1.a)x=¢+4+ 20? — 23 + £4/3 — £5/12 + 19/60 4---,

y=t+t?
— 4/64 15/12 — 19/40 +---,

x = 277— 15 + 74/3 — 79/12 + 16/60 + ><,

y = t? — ¢*/6 + £°/12 — 19/40 4---
b) x = 14 #7 + 8123/3 + 524 + 10825/15 + 1506/8 + ---,
y = —2 + 2213/3 + 2t4 + 5425/15 + 6016/8 + ---,

x = 27 + 8f3/3 + 544 + 10825/15 + 15026/8 + ---,

y = 2123/3 + 2r4 + 5425/15 + 6076/8 4+ -
2. a) 07 — 5’\+1; A’=4,4
b) X2 3: \ = 4/3
c) AZ +1; ’\=+i
Gy) Xe ees
e) 2 - d= 0,0
 Hints and answers for selected exercises 361

Pee AE ON AS ern St
oe ha N= as
h) A7+ 2 7 Awe One
eee | eee
xX = f(2) ko ext® and x = g(t) 2-9 det*, where co = do = 1 and
a) f(t) = 117, cry = —6cx/[6(K + 3/2)(k + 1/2) + (k + 3/2) + 1];
g(t) = 119, cher = —6cx/[6(K + 4/3)(k + 1/3) + (k + 4/3) + 11.
b) f(t) = £93, cy, = (—1)*/ + 2V3)(4 — 4V3)- ++(K? = 2kV3);
g(t) = 1-8, cy = (—1)*/01 — 2/3)4 — 4/3) --- (KR? — 2kV/3).
c) f@) = t, cry = —ex/(K + 10k +1 4 21);
a) = las Gy, = Oy

D/O = 01, cepa = —ex/(K + DK + 1 + 28;

BO) ht a= Cy.
e) x = b1(t) = Dio (—1)*t*/(k!)?_ and
x= —2>02., (—-Dtfd + °°- + 1//)/(k)? + 10 Int
f) x= $i) = 1! Diteo (—D*/()? and
x= —2>02. (-D'0 +++ 4+ 1//)/k)? +: $12 Ine
g) x = Vi and x = vVitlnt
h) x = $1(2)= ee (— DF /kVK + 1)! and x = 17! >0%_9 dkt* + 61(0 Int, where
do = —1, dq, is arbitrary, and di41 = —[d, + (—1)*(2k + IkM(K + 1)N/k(K + 0D,
Vero be
i) x = bi() = VED Keo HH/kMk + VD! and x = (Co det)/Vi + 61(0)Int, where
do = 1, dy is arbitrary, and 4k(k — 1)d, = 4d,_1 — 4/k\(k — 2)!, k => 2.

. It is sufficient to observe that sin t/t = 1 + ¢7/3!+--- for ¢ ¥ 0.

5.x = 14+ 0210+ be -k + D6 - 04/2K?
5 @)) TEC DN = Th, ae = 4 ae =
For )\ 1 = 312 yee ptt?
Oe

b) Let Cy = (kK? + 2k " 1 et co = [i, 117, c. = Cur, > 1. x =

t > %_o ext® and x = 1~*)°%-9 ext* are two linearly independent solutions.
OT ee A fe
alt Cle at) 1 Deus a a CeO, acai eee
ay

Cyen—2, k > 0. Then x = >°%_, ext* is one solution. To obtain a second linearly
independent solution x = 11/3 D0f_9 dy 7*, let

do = Ul, — 1)... diz= 0,

Sut Rye
D, = — (3k?— 2k)—} Ve a d; = Dydy_2, k > 2.
362 Hints and answers for selected exercises

= Poa. eile "|

is
0 t

» [Fs (AY, Deo dit! + tel(n du,

re) ae?

where u = [1,0]7, do = [0,1]", kd, = —u/k! +|

10
ak
diss, bee

QP, Ore ser

8. a) [t, hae [13/2, ia i
b) pitt Po c.t* and Lea aie 9 Cit", where (Oy = Wied ee

== (k*
(E2 + 2ik) 1
“)—1 lee
cy mee c.-1 for k> LE

c) Let u=[1,0]7. x = @1()= 2? Db fu/k! and x = g2()= 0? Dekeo dit"

—1
$1(1) Int, where do = [0, 1]” and k?d, = —u/(k — 1)! + Is : dy-i, k IV

9. Set s = 11/3 and x() = y(s) to obtain sy” — 2y’ + 9s!°y = 0, y(0) = 0, y’(0)
xX = Po C1imt 11/3, where co = 1 and c1im = —9¢11~@—1)/11m(Q11m — 3), m > 1.
10. Solutions of the transformed system are of the form c1s?@i(s) + c2sp2(s) +
c2Ys?@1(s) Ins, where ci, c2 are constants and 1, $2 are analytic at s = 0. These
solutions approach zero as s does; solutions of the original equation approach zero as
(> o

Exercise 6.6

1. y=
yo i 1-1
ge (B)
—
_ 1/2) yr,
y, BO -|
a
0 a,
aie

9
4 . a) Let t> 04+ in cyt? yi() 4+ coyo( = 0.
b) Let t— 04 in e1f?'yi() + coye(t) = 0.
c) Note that lim;_,o+ t* does not exist.
d) Let > 0+ in cyyi() + ceyo(t) = 0.
|Ax(t — 1)]| < 214.) < 2-1 4-2) <1
pS Map h) e | 1 2/(k + 1)|. |K(A — kI)“}| < 34/5
2/(k + 1) 1

Exercise 6.7

. Hint: Integrate by parts and use l’Hospital’s rule.

. 4/7/3
tn x = (0/23 DE 9
(—D*G/2)*/kIT(k
+ 4/3) and
x = (¢/2)—¥3 VF (—1)74(¢ + 2)7*/KIT(K + 2/3)
. Recall that lim,.; [(@)| = + for every negative integer k.
an . Hint: Prove that dJ,(Or7]/dt = rJ,_1(d and that dUJ,()t—’//dt = —1-”. v41(t).
10. Use Problem 8.
 Hints and answers for selected exercises 363

12. J3/2(t) = V2r/t (t—! sin t — cos 2)

J_3/2(1) = —V/2r/t (t~! cost + sind
BPO = 3-2 — 5 — Dokee daxt?*—- + Jo(t) Int, where ds = 5 and

2(—1)'-2(2k — 2)
4k(k — 2)d2; = —do,_2 +
(k — 2)'k122%=D ©
14. Hint: Show that r?x’” + rx’ + (r?\ — n?)x = 0 becomes 12y” + ty’ + (22 — n?)y =
0 if t = Vdr and y(t) = x(v).

Exercise 7.1

. Another hint: Differentiate u(t) = o()Y/(1) — Y()¢d’(1) and integrate wv’ from f; to fo.
. Use Problem 3.
. Use Sturm’s comparison theorem.
5 10) yA Gy Se G BS Uy or erp ee AEE)
. a) x = (ec; + colndr}/? is a general solution.
Ss
On
“ION

b) x = (ci cos In r*/? + c2 sin In r¢/?)t!/? is a general solution.

8. Use Sturm’s comparison theorem and the equation x’”” = 0.
9. Every solution x = y/(7) will have infinitely many zeros. Consider, in plane polar
coordinates, the curve x = Y(t), y = W’(t) and show that

ee ha, 13
ee (42) oe
10. Supposey did. Let {t;,} be a sequence of points such that Y(t.) = O and lim; 4, t. = ty
exists. By continuity, ¥(t,,.) = 0. Use Rolle’s theorem to find a sequence s, such that
wW'(s,) = 0 and s, > ft, as k— +0. Thus deduce that y(t.) = 0. Hence y(t)= 0.

Exercise 7.2

1. a) Three b) One
2-3. a) There are four linearly independent periodic solutions of period 7.
b) There are precisely two linearly independent solutions of period 7. All other
solutions are unbounded.
4. There are four linearly independent periodic solutions: two have period 7 and two have
period \/27. Not every solution is periodic since a sum of functions with irrationally
related periods is not periodic.
5. x1 = —e~ 24/2 + (1 + V/2)e—V¥*#/2 R
xg = 1 —e-24/2, x3 = e724, x4 = —e7-74/2 4 e-V7t/2
6. Only (c)
8. a2 > 0, a3 = a2a\
9. x = ci sint + cecost + e~/2(e3 cos(V7 t/2) + c4 sin (V7 t/2))
10. Only (b)
364 Hints and answers for selected, exercises

11. Let x and y denote the displacements of the sheet and block respectively from equilibrium.
The equations of motion are Mx” = —kx + a(y’ — x’), my” = —a(y’ — x’). Note that
\ = Ois an eigenvalue corresponding to [x, x’, y, XY 7 = [0,0, —1, Q}?. By Wall's criterion,
all the other eigenvalues have negative real parts. Now deduce that x — 0.x" - Q»— — 1,
y’—0as t—-+o. The plate must be two units wide.

Exercise 7.3

1. Characteristic exponents are (a) 1, (6) —1, (©) 8v 2/r.

2. There exists a T > 0 such that fo a(s)ds = 0.
3. a) Periodic b) Bounded, but not periodic

4. a) (tf) =a Re
1 F 0}|e 1 b) (A) _ ie
jen" ‘ Of‘|2 x

sin ft fsint — cost fe2*sint —e-‘cost

ae ec NT te t tcost + sin i] tO eS cost e-‘sint
6. Use Theorems 7.5 and 7.6.
7. Use Theorems 7.5 and 7.6.
14. Use Problems 9 through 13 as lemmas.
15. Apply Problem 13.

Exercise 7.4
1. The integrand is periodic.
2. They are bounded but do not approach zero as f—> +c:c. Make use of uniform con-
vergence to approximate g by asum > ;_, k—? cos (tk—") with error less than 4—-* F_, &-?.
Assume, for contradiction, that g(t) ~ 0 as t— +2 and consider fy = 2+ MN? for sof
ficiently large M.
3. They are all bounded (Corollary 7.7).
4. Use Theorem 7.7. a) They are at least bounded. b}d) They approach zero.
6. a) Use Theorem 7.8 and Wall’s criterion. b) Use Theorems 7.7 and 7.8.
7. d) They are all bounded.
0 0
8. Take
ake R(t)
R(t) = E Va + capand apply ; Theorem 7.10.

9. Pattern your argument after the proof of Theorem 7.9.

10. Consider the system x’ = y, Y¥ = —(w? + e~*})x and follow the proof of Theorem 7.9.

11. Using the decomposition L(r) -| ’ | and R(‘) = t-*(sin? AN, ome sees that
hypothesis (iii) is violated and the conclusion of the theorem does not hold either.
12. A fundamental matrix solution is
t

exp|| sin eae]-| a 1).

0 —1 9
The conclusion of the theorem holds even though hypothesis (ii) is violated.
 Hints and answers for selected exercises 365

if aie sin t —cost|. :

(1) ree a fundamental solution matrix. Even though hy-
pothesis (iii) is violated, a stronger result than the conclusion holds: every solution approaches
zero aS t— +0,
14. Introduce a new independent variable s = 1—?.
15. Consider the intervals o — k-1 < t < o, k > 1.

Exercise 7.5

1. A general solution for the homogeneous equation is x = cyt! — cot, y = cota}.

A particular solution of the nonhomogeneous equation is y = 0 and
a) x = 17/3 b)x=t—1+277!In[t+1|
C)ive—s7/ 2 d) x = ¢-!sin
¢t — cost
e) x =2#/2+4+1 f) x= t/2 +171 In|z7]
9) x= 22/3 h) x = ¢t-!In¢
i) x = —t-! cost
. No. Compare the problem with Theorem 7.15.
. Use Theorem 7.13 and Wall’s criterion.
N
RW. Use Theorems 7.7 and 7.13.
5. The proposition is true. Let « > 0 be given, and fix fo so large that |b() — ¢| < e
for to < t < +x. Nowuse the variation-of-parameters formula and Eqs. (7.4d) and (7.4e).
6. All the solutions approach zero as t ~ + even though hypothesis (ii) is not satisfied.
7. If is a pure imaginary eigenvalue of multiplicity m, then there are m linearly independent
eigenvectors corresponding to it.
8. If w = 1, the solutions are unbounded. If w is irrational, the solutions are bounded
and the only periodic solution is x = (w* — 1)~!sin+¢. If w is rational, say w = p/q ¥ 1,
where p and q are relatively prime, then there is one solution of period 27 and the other
solutions have minimal period 27p.
9. Yes. x = sint. Note that the right side has period 7, not 27 as has the periodic
solution.
10. x = —3~!sin2r, y = —2-! — 67! cos 2r is a particular solution of period 7. All
other solutions have period 27.
11. There will be a unique periodic solution if w is rational. Solve the first equation ex-
plicitly and apply Theorem 7.15 to the second equation with x eliminated.

Exercise 8.2

1. |x;| + 2|x2| = 1. Open 2. #. Open and closed

3. |xy| = 1 and |x2| = 1. Neither AD xe 1 Open
5. The boundary consists of the given set, the origin, and points with coordinates (m—1, 0)
and (0,~!), where m and n are non-negative integers. Neither
6. x3 = + + ree Closed 7. The boundary is ®!. Neither
366 Hints and answers for selected exercises

8. xt + xd + 2g = 27. Open 9. ®!. Neither

10. R!. Neither
11. a) No b) Let S be the set of integers.
c) Consider sequences converging to cluster points.
d) A closed interval and one point not in the interval

Exercise 8.3
1. Yes. Note that x2/(x? + y?) < 1 for all (x, y) ¥ (0, 0).
2. Yes. Let a, denote the jth component of x,;. Then 0 < azi1 — a, < (k — 1)7}
(ax — ax—1) for k > 2. Thus 0 < agai — a < (a2 — a1)/(k — 1)! and az41 < a2 +
Fs (a2 — ai)/(r — U!.
3. When plotted in the plane, the points form a rectangular array. List the points in
diagonal order beginning with (1, 1). Any subsequence consisting of points arranged hori-
zontally or vertically will converge. There are also subsequences converging to (0, 0).
4. No. g(m) > m for every integer m > 0.
ee — — eer a — ee ee
6. The boundary of {x: g(x) > 0} is {x: g(x) = O}.
7. Construct a sequence of points from the set which converges to the least upper bound o
of the set. Then deduce that ca is in the set.

Exercise 8.4

1. Let p be a point A and let q = (1/27, 0). If there were an arc from q to p, it would
be given by y = sinx—! for 0 < x < 1/2 and it would be true that (0, sinx—!) — p
as x > 0+.
7 gg
3. a) If it were not connected, it would be the union of two closed and bounded sets
separated by a positive distance. Now use the definition of interval.
b) Assume that the range is disconnected and deduce that [a, 5] is disconnected.

Exercise 8.5

4. Consider unit spheres with centers (0,0,0) and (0,0,1). These spheres separate
®? X R! into four domains.
5. In plane polar coordinates the systems have the form (a) r’ = 4, & = O and (b)r’ = 0,

6. It is a line or line segment parallel to the ¢-axis.

7. It is a helix (not necessarily circular) with axis parallel to the r-axis.
8. Points (x,y, 4) on the graph satisfy (« — 1)? + (y +)? = 1.
Exercise 8.6
1. 2/2

2. af/dx = (1 — t°x?)/(1 + 1°x?)? is bounded for all (x, f) since —3 < a/(1 + a?) < 4
for any real a.
 Hints and answers for selected exercises 367

6 fa) ie b) 1 + 2a + 3a? yal d) 3

AS Oe oe
5. Use the result of Problem 4.
6. Yes. Use the mean value theorem.

Exercise 8.7
11,1 — 7,1 —#? + £4/2!; e-? = > 2) (—2)*/k!}
Lite,lt+r+74+P73; Q—-)=PEot
2€=¢%
pe I be any interval containing ¢ = 0. Define x = 0 for in Jand x = #4 for ¢ not
in J.
4. dx(s) ~0Oask>+oua, O0<s<1
foes)ds
> lasko+to, 0<t<1
5. e/100!
GSE)

7. Hint: Yi@) = A + Bt — 27/2, Peri) — WO = —SJO@ — Se) — Pa-19] ds,

k > 1. Consider )’f-0 Wet1@) — ¥x().
8. When and only when the eigenvalues \j, Ag satisfy |\Xy] < 1, |AXo| < 1
9. Hint: Consider x = xo + xo(t — 1) + Si, (t — s) f(x, x’, s) ds.
10. x = mt, y = nt, m? +n? =1,t> 0. Use polar coordinates.

Exercise 8.9
1. Let p = x? + y? and note that p’ = 29(1 — p). Now use Theorem 8.5 taking discs
for B.
2. The trajectory of any solution initiating in the first quadrant is contained in the first
quadrant. Note that u/(f) > u?() and vo'(t) > v(a).
4. Let u(t) = fo M\6(s) — xo] ds.
5. Show that f is globally lipschitzian.
6. Yes. Modify your argument for Problem 4.
7,. a) The right sides are globally bounded.
b) The right sides are globally lipschitzian.
8. Use polar coordinates. The examples indicate that Theorem 8.6 is a stronger result
than Theorem 8.5.
9. Continuity is sufficient. Modify your solution of Problem 6.
10. Let @ denote a solution. Show that W(@(z)) is nonincreasing; hence @ is bounded,
11. Pattern your argument after Example 2.

Exercise 8.10
- € 2 SS)
1. $(xo, to, = bo + ¢? — ali”?
2. They are at least bounded by Theorem 7.7. 87,
Alternatively, seek a solution e*'[sin
cos 81]? of the variational equations and proceed as in Example 3, Section aot
368 Hints and amswers far seleegadl exewomes

® 1 @
5 Te
imarom ct ma 0 i}
®@ 0 2
4 Use polar coordinaies.
5. Given « > Q there & a 3 > O such that ko — a < 3 bmplies
gixo, to, — dfa,m,2|
<< Rrahl +> wm

Exercise
9.2

1. of) = G&D. YO = (ont + 4 —Sar+ DL The waiectony of @ does net Interact

itself. The trajectory of w intersects Meelf and that of @
2. Solve the equations exphatly. No
3 ofxo, vo, to, D = (xge™2, woe + Bayle — we™™e)
4. oxo, to, 2 = xa/(l — xo — mh OG, 1, 9 = 46, QD
5. oxo, to, 2 = xo/(l — x — QD), 60, 1, 9 = 46, QD
6 See4.
7.x = os ¢ — my = Sal —w®
& At any ume, the powtion and the velooity of the frst lock will be the same as the
position and velocity of the second block 7/2 woonds before.
9. fix, t+)
= fx, 2 for al x and %
10. fs a constant fenction of

Exervise
9.3

1. bb Ye. Yes
2. a) Show that »* + 2 fq 2 a = ¢ defines a Simple dosed curve fee all ¢ > Q@
b) Yes No. Comsder g(x) = xe,
3. b) The critical poimt » = 0 corresponds to escape; w = £7? conrespends to a Giowlar
orbit. Plot > = 2ku*/3 — x? + ¢ and consider eparatnives first,
4. Nowee that there are critical points which correspond t an inverted position for the

indicating swinging only.

>. Use the energy W(x, ») = ¥*/2 + Jp e@ adeto Show that all solution paths approach
the ongm.

Exercise 9.4

1. a) x = sale — mm yy = eos (¢ — ®) >) Use Theorem

85,
c) Let a = x* + y* and @ = tan™* »/x. Nowe that e = 1/(2 + 2ven/(@A —
=£
9) and
2 No. The block tends to equilibrium at the position x = aw, &, where
a is the coefiiciant
of friction between the block and the belt.
 Hints and answers for selected exercises 369

3. a) By Theorem 9.7, no b) If and only if M/(mfg) < 1

c) Yes. Assume that M/mfg > 1 and note that
t

pote = On hf e*(e + 1 — sin @(s) ds),

where a = k/m(? and e > 0. Now show that @—> +2 ast—> to.
5. Use Theorem 9.8.
6. u(t) = ccos (4c7A), v(t) = csin (4c71). The period is 7/2c?.

Exercise 9.5

NS en = Dip
2. If not, then there is a bounded sequence of points @(t;,) such that dis (@(t;),2) > € > 0
for some e. Now use the Bolzano-Weierstrass theorem and Theorem 9.9.
3. a) Show that lim,_,,, V(@()) exists; let p and q denote distinct points of Q; and
consider sequences {s,} and {f,} such that @(s,) — p and @(t,) — q. Thus deduce
that V(p) = V(q).
b) Take V(x, y) = x? + y?.
4. a) (x — z) and (x’ — 2’) are periodic. b) Let A be irrational.

Exercise 9.6
1. In polar coordinates, dy/dx = [1 — r) + rcot 6]/[1 — r)coté — r].
2. Use polar coordinates. The ratio is e?”.
3. Let x = (2), 0 < ¢ < 2 denote a parameterization of C. Then

f = (9 — p)- (6 — P)
is a differentiable real-valued function which a minimum at fg in [0, 27]. Thus

0 = f’(to) = $'(to): (G0) — P).

4. a) Use Theorem 9.16. b) Use Theorem 9.7.
5. Pattern your argument after Example 1.
6. There is one critical point (1,0). Any periodic orbit would have to encircle it. This is
impossible since the x-axis consists of three entire solution paths.
7. Yes. If X is irrational, then every point on 3(a, b) is a limit point for every solution
path initiating on 3(a, 5).
8. The equations g(x) = M — f(0) and f’(x) = 0 must have real solutions.
9. No. Consider tangent vectors to paths on the lines x = 0 and y = +1.
10. Use circles for the Jordan curves.

Exercise 9.8
3. The condition wR > v implies that y’ is bounded away from zero near (R, 0).
4. With r2 = (x — R)? + y?, 6 = tan—! (9/(x« — R)), the differential equations become
r = —v+oRsin#, rf’ = rw + wReos 8.
370 Hints and answers for selected exercises

8. A large value of w will give rise to a focus.

9. As w > v/R, the critical point approaches the light source.
10. If w > v/R, all the lightworms are swept to the wall of the jar.

Exercise 9.9

1. Assume that xg(x) > 0 for x ¥ 0 and fo'” g(x) dx = +2.

2. b), c), d) Roughly speaking, the coefficient of x’ should have different algebraic signs
for large and small values of x, and the last term in the equation should act as a restoring
force to some extent.
3. Let x denote the displacement of the center of mass from the slot. An equation of
motion is mx’ + al?x’ + av(A,(x) — Al(x)) = 0, where @ is the coefficient of friction
and Af¢(x) and A,(x) are the areas in contact with the left and right belts respectively.

Exercise 10.1
1. The annulus 0 < r < 1, the circle r = 1, the region r > 1, any solution path, any
region between two distinct solution paths.
2. Construct open path tubes around the limit cycle.
3. What really needs to be verified here is stability. A node or focus is quasi-asymptotically
stable by definition.
4. Use the Jordan curve theorem and Theorem 9.3.
5. All solutions have nonfinite future escape times.
6. Note that |Y(O| < W(t) I[ — 11)? + lle~"-@ < 4\W(t)\e?". Choose 6; = 1
and T = 2\In (€1/4)|.
7. To show stability note that |g(d| < rai ( + &4)-1 + 24+ wie dd +k),
where N is the greatest integer in k. To disprove asymptotic stability, note that g(t) > 1
whenever f¢ is an integer.

8. a) In the notation of Eq. (10.1b), A() = ae ec zs ea

—1-— sin2t —1W— cos2r
b) x” — px’ + x + h(t, x’) = 0
9. Let x = $(t), y = ¥(@ denote the periodic solution of interest. In the notation of
Eg. (0.1b), A@ = paren: Hi. Additional quantitative information is
necessary.
10. a) Use polar coordinates. b) Disturbances due to turbulence could be amplified.
11. Assume that x = 0 is the critical point of interest. Choose 6’s for (xo, 0, ) and apply
the first shift formula for autonomous equations.

Exercise 10.2

1. The equation is asymptotically stable if and only if a > 4. Use the mean value theorem
to estimate ¢sinIn¢ — ssinIns and apply Theorem 10.2 to conclude that uniform asymp-
totic stability holds for a > 1//2. Next let s, = exp [2rn + 1/4] and t, = exp [27m +
n—1 + 7/4] and show that the asymptotic stability is not uniform for $a <0/x/2.
 Hints and answers for selected exercises 371

. Use Lemma 7.8, Theorem 10.2, and Gronwall’s inequality.

. Yes, by Theorem 10.2. Yes; let L(t) = [—1].
. By Theorem 10.2, they are asymptotically stable but not uniformly stable.
W
fk. Use Abel’s
Nun formula (4.5c).
7. They are at least uniformly stable by Theorem 10.3. One can show that (c) is uniformly
asymptotically stable by using Theorem 10.2 and the method of proof for Theorem 7.8.
8. One such condition is that 7, [-—q) + V@2(s) + (1 — p(s))?] ds converge. One can
also use Theorems 7.8, 7.9, and 7.10.
9. If the matrix A has an eigenvalue with positive real part, or if it has a pure imaginary
eigenvalue \ of multiplicity m > 1 and there are not m linearly independent eigenvectors
corresponding to \, then x’ = Ax will be unstable. Use Floquet’s theorem to obtain a
similar result for the periodic equation.

Exercise 10.3

4. Pattern your argument after the proof of Theorem 10.4 with A(x, y, 1) = g(Ao(V/x2 + y2).
Use Gronwall’s inequality (Problem 5, Exercise 7.4) with v(/) = g(#), noting that [o°” g(t) dt
converges even though g is unbounded for 0 < ¢t < +.
5. Replace the hypothesis h(x, 7) = o(|x|) by h(x, 4) = g(‘)b(x), where b(x) = o(|x|) and
0” |g(t)| dt converges.

Exercise 10.4

1. Note that |@(xo, to, d| < (|xo| + e2 — le—“—o). Thus, if |xo| < 1, then |F(xo)| < 1
also. By uniqueness of solutions for initial value problems, @(Xo0, fo, 1) = $(Xo, fo, t + 27).
3. Use Theorems 7.7 and 10.7.
4. The solution is x = 0, y = sin?¢, z = cos¢. It is asymptotically stable by Theorems 7.7
and 10.7.
5. It is stable, but not asymptotically stable.
7. Use Theorems 7.6 and 10.7 as in Example 4.
8. Use the hint for Problem 7.
9. Use the hint for Problem 7. One sees from van der Pol’s equation that the condition
is sufficient for asymptotic stability, but it is not necessary.

Exercise 10.5
1. Hint: Let B be symmetric. Then it is necessary that all eigenvalues of B be positive and
that all eigenvalues of A7B’ + BA be nonpositive.
3. Take V(x, x’, 1) = (p(x? + x’? )e?®, assuming that p is positive-valued and non-
increasing.
4. Use right- and left-hand derivatives.
6. It is unstable.
7. Let p be differentiable, positive-valued, and strictly increasing. Then x’’ + p(‘)x = 0
is unstable.
372 Hints and answers for selected exercises

9. The theorem uses the euclidean norm for a Lyapunov function.

11. Make a Taylor expansion for f and estimate an appropriate quadratic form as in Theorem
Ustle
14. Let V(x, y) = y?/2 + f§ g(w) du and note that V(x, y) = ¢ defines a simple closed
curve for sufficiently small c > 0.
15. This problem differs from Problem 14 in that V(x, y) = c defines a simple closed curve
for all c > 0.
16. Argue as in Example 11.
17. Argue as in Example 11.
18. The disc x? + y2 <1. Use V(x, y) = x*/4 4+ y?/2.
INDEX
 INDEX

Abel’s formula, 105 Bolzano-Weierstrass theorem, 230

for characteristic multipliers, 198 Boundary of a set, 226
relation to the flow mapping, 275 Bounded set, 229
Acceleration, centripetal, 11
Coriolis, 11
due to gravity, 2 Capacitor, 16
Amplification factor, 20 Cauchy product of series, 149
Analytic function, 148 Center, 292
Annihilation method, 73-76 Cetaev’s instability theorem, 331
Arc, 231 Characteristic equation and polynomial
Arcwise connected sets, 231—232 for a matrix, 65
Asymptotic behavior of solutions for equation for a linear differential equation,
linear homogeneous equations with Tal
constant coefficients, 184-188 exponents and multipliers, 193, 195
linear homogeneous equations with values, 65
continuous coefficients, 204-215 Closed sets, 226-228
linear homogeneous equations with Cofactor, 61
periodic coefficients, 193-202 Complement of a set, 227
linear nonhomogeneous equations, Complementary equation, 90
218-222 Completeness axiom, 205
Attractor, 292 Connected set, 232-233
Autonomous equation, 203 Convergence criteria for vector series, 96,
150
Basis, 79 Convolution theorem, 141
Bendixson’s negative criterion, 276 Critical points, 42, 264
Bessel functions of the first kind, 166, 178 classification for planar linear systems,
of the second kind, 179-180 293-294
spherical, 178 for planar quasi-linear systems,
Bidiagonal matrix, 54 295-296
Blasius’ boundary value problem, 23, 254 Cylindrical phase-time space, 250-253
374 Index

Damped linear oscillator equation, 13, 67 Flow mapping, 273

Decrescent Lyapunov function, 330 Focus, 292
Degrees of freedom, 14 Forcing function, 73
Dependence of solutions on initial values, Friction, 3-4 —
256-262 Fundamental existence and uniqueness
continuity, 258-260 theorem, 247-248
differentiability, 260-262 matrix solution, 104
Derivative with respect to a differential solutions for the first order homogeneous
equation, 327 system, 104
Determinants, 60-63 for homogeneous equations with
Diagonalization of matrices, 83 constant coefficients, 121-125
Direction field, 237 for nth order homogeneous equations,
Distance between sets, 233 107
Divergence conditions, 273-277
Domain, 231 Gamma function, 171
Gating function, 142
Eigenvalue, m-fold, multiplicity, 65 General solution for first order
Eigenvector, 82 homogeneous systems, 63-68
Electrical current, 15 for nth order homogeneous equations, 72
Elementary critical points, 292 Geometry of x’ = f(x, ft), 234-237
Equivalent integral equations for initial Gronwall’s inequality, 94, 216
value problems, 92, 241
Hamiltonian system, 37
Escape times, 249-255
Hard spring, 43
Estimates for matrix solutions of linear
Heaviside unit step function, 137
periodic systems, 208-209
Helmholtz’s equation, 158
Euclidean length of a vector, 229
Hermites’ equation, 154
Euler’s equation, 161
Hills’ equation, 200, 203-204
Exact equations, 33-36
Homogeneous system of linear, algebraic
Existence and uniqueness
equations, 80
of periodic solutions for the equations of
van der Pol and Liénard, 299-302 Identity matrix, 50
linear homogeneous equations, 186, 196 Imaginary part of a complex number, 127
linear nonhomogeneous equations, Improper node, 292
2IQ, 221-222 Inductor, 16
planar autonomous systems, 270 Infimum, 204
quasi-linear equations, 318 Initial value problems, 3, 24, 91-93, 237
of solutions to initial value problems, 25, Integral curve, 237
26,92, 97, 2422247, 248 Integrating factors, 6, 7, 37, 45, 99
Exponential matrix, 54—60 Interpolation function, 133
estimate of the rectangular norm, 184 Interval of analyticity, 150
Exponential order, 128 Invariant set, 305
Extension of solutions, 249-255
Inverse Laplace transform, 135
Irregular singular points, 157
First integral, 30
First shift formula, 265 Jordan (block) matrix, 54
First variational system, 261 canonical form, 64, 84-89
Floquet’s theorem, 194 curve, 231
 Index 375

Kirchhoff’s current law, 18 fundamental solution sets, 121-125

voltage law, 17 general solutions, 70-72
reduction to first order systems, 70
Laplace’s theorem on computing Linear homogeneous systems with constant
determinants, 61 coefficients,
Laplace transform, 127 asymptotic behavior of solutions,
of a derivative, 130 184-188
of a general solution for a first order classification of critical points, 293-294
system, 131 existence of periodic solutions, 186
of a general solution for an nth order fundamental solution sets, 121-125
equation, 132 general solution, 64
of an integral, 140 matrix notation, 49, 52
of a periodic function, 139 reduction of nth order equations to, 70
Lasalle’s theorem, 334 stability, 311-312
Legendre’s equation, 170 Linear homogeneous systems with periodic
Lerch’s theorem, 135 coefficients,
Liénard’s equation, 300 asymptotic behavior of solutions, 193-202
Limit cycles, 280 existence of periodic solutions, 196
stability classification of, 289 fundamental solution matrices in
Limit inferior, 205 Floquet form, 194
Limit sets, 278-282 stability of, 311, 313
in the plane, 283-290 Linear nonhomogeneous equations, 90
Limit superior, 205 asymptotic behavior of solutions, 218-221
Linear combination of functions, 73 existence of periodic solutions, 221-222
of vectors, 78 general solutions, 116
Linear (in)dependence of solutions for method of annihilation, 73-76, 120
differential equations, 104 stability, 308-309
of vectors, 77 variation of parameters, 117-120
Linear friction, 3 Linear oscillators, 12, 14
Linear homogeneous equations with Linear spring, 12
asymptotic behavior of solutions, 204-215 Linear variational equation (system), 275,
continuous coefficients, 90 8208322
existence of solutions to initial value Lipschitz condition, 238—240
problems, 92 Lyapunov, direct method of, 324-337
fundamental solution matrices and sets, function, 324
104, 107 stability in the sense of, 307
general solutions, 112 stability theorems of, 328, 330, 339
initial value problems, 91
linear (in)dependence of solutions, 104 Massera’s example of an equation with
linearity of the solution space, 110-113 nonasymptotically stable equilibrium
reduction of order using Abel’s formula, and a Lyapunoy function with
108 negative definite derivative, 329
stability, 310-314 Matrix,
uniqueness of solutions for initial value diagonal, 83 .
problems, 92 eigenvalues of a, 65
Linear homogeneous nth order equations functions, 52
with constant coefficients, inverse of a, 50, 81
376 Index

multiplication, 49 Poincaré phase plane, 263

nonsingular, 50 Positive definite functions, 326
notation for systems of differential Product rectangle in &" & @1, 234
equations, 49, 52 Proper node, 292
principal diagonal of, 50
singular, 50 Qualitative behavior of solutions for linear
transpose of a, 49 equations, 182-223
Maximal solutions, 246-247 Quasi-asymptotically stable solution, 302
Maximality principle, 290 example in closed form, 309
Minor, 61 Quasi-linear equations, 308
Mutual inductance, 20 existence of periodic solutions for, 318
stability of equilibrium, 315-318
Neighborhood in &" « ®!, 234 stability of periodic orbits, 320
Newtonian gravitational constant, 2 stability of periodic solutions, 319
Newton’s law of gravitation, 2
second law of motion, 2 Radius of convergence of series, 149-150
Nonelementary critical points, 296 Ramp function, 142
Nonexistence of periodic solutions, 276, 290 Real part of a complex number, 127
Nonhomogeneous system of linear Rectangular distance function, 226
algebraic equations, 80-81 Rectangular neighborhood in &”, 226
differential equations, 90 Rectangular norm, 92
Normal differential equations, 22, 224 Reduction of an nth order equation to a
first order system, 235
Open sets, 226-227 Reduction of order for second order
Orbit, 264 equations, 39-40
Orbital motion, 10-12 Region, 231
Order of a differential equation, 22, 224 Regular singular points, 157
Ordinary points for a differential equation, Resistor, 16
150 Rocket, 8-10
Oscillator, electronic, 19-21, 304 Rotation point, 292
hard spring, 43
linear, 12, 14 Saddle point, 292
soft spring, 41 Scalar, 47
Second shift formula, 266
Parallel RLC circuit, 18-19 Separable equation, 27-29
Particular solutions, 101, 115-120 Separation of variables, 5
Period transformation matrix, 195 Series representation of solutions, 147-181
Periodic orbit, 271 around ordinary points, 150-156
Perron’s example of a system with unstable around regular singular points, 161-169
equilibrium but asymptotically proofs of theorems, 171-176
stable linear approximation, 315 Series RLC circuit, 15-17
Phase portraits, 42, 267-271 Similar matrices, 83
shift, 264 Simple (closed) curve, 231
space, 42, 235 Simply coupled systems, 44—46
-time space, 235 Soft spring, 41
Picard, method of, 95, 241-245 Solution, 24, 44, 91, 235
Poincaré-Bendixson theorem, 287 explicit, 30
 Index 377

implicit, 33 Successive approximations, 95, 241-245

maximal, 246 Superposition principle, 75, 120
path, 264 Supremum, 204
Spanning set, 79
Stability, 304-339
Triangular matrix, 62
asymptotic, for invariant sets, 305
for solution functions, 306
determination of, from linear Undamped linear oscillator equation, 13,
approximations, 317 5, 32
global asymptotic, 307 Uniqueness theorems for
in the sense of Lyapunov, 307 periodic solutions of linear equations,
in the sense of Poincaré, 305 186-187, 196, 198, 219
of equilibrium, 305 solutions to initial value problems, 25,
of equilibrium for quasi-linear systems, 26, 92, 242, 247-248
315-318
of invariant sets, 305 Vacuum tube characteristics, 20
of linear systems, 310-314 Van der Pol’s equation, 21, 299-302, 304
of periodic solutions, 318-322 Variation of parameters, 101, 117-118, 209
of solution functions, 306 Vector functions, 52, 229-230
quasi-asymptotic, 308 Vectors, 46-47
region of asymptotic, 332-337
uniform (asymptotic), 307
Wall’s continued fraction criterion, 188-191
Stable elementary critical points, 292
Wronskian, 104
polynomials, 188
Stiffness coefficient, 12
Sturm’s comparison and separation Zero matrix, 49
theorems, 183 vector, 47

ABCDE7987654321
 seat) SUMS Lett Qaags
ral => ron EES = OF

—)
=
ae |
oa i) > D

= ab @ (60 aaa
= JOC Bowe i
= ter sy :
 ) why AN A}
aA
 Ah

Uf