Nonlinear Systems - Vydiasagar

Nonlinear Systems Analysis
M. VIDYASAGAR
Professor of Electrical Engineering

Concordia University
Montreal, Canada
Prentice-Hall, Inc. Englewood Cliffs , New Jersey 07632

Library of Congress Cataloging in Publication data
VlDYASAGAR, 1947-M
Nonlinear systems analysis.
(Prentice-Hall electrical engineering series)

Bibliography: p.
Includes index.
1. System analysis. 2. Differential equations,
nonlinear. I. Title.
QA402. V53 003 77-24379
ISBN 0-13-623280-9
© 1978 by Prentice-Hall, Inc., Englewood Cliffs, N.J. 07632
All rights reserved. No part of this book

may be reproduced in any form or
by any means without permission in writing
from the publisher.
Printed in the United States of America
10 9 8 7 6 5 4
Prentice-Hall International, Inc., London

Prentice-Hall of Australia Pty. Limited, Sydney
Prentice-Hall of Canada, Ltd., Toronto
Prentice-Hall of India Private Limited, New Delhi
Prentice-Hall of Japan, Inc., Tokyo
Prentice-Hall of Southeast Asia Pte. Ltd., Singapore
Whitehall Books Limited, Wellington ,
New Zealand
Contents
1
Preface, ix
Introduction, 1
1 .1 General considerations, 1
1 .2 Autonomy, equilibrium points, 5
Second-order systems, 10
2.1 Preliminaries, 10
2.2 Linear systems, 14
2.3 Nonlinear systems, 21
2.4 Periodic solutions and limit cycles, 20
2.5 Two analytical approximation methods, 41
Nonlinear differential equations, 50
3.1 Mathematical preliminaries, 51

3.2 Induced norms and matrix measures, 64
viii Contents
4
3.3 Contraction mapping theorem, 73
3.4 Nonlinear differential equations, 78
3.5 Solution estimates for linear equations, 88
Approximate analysis methods, 96
4.1 Describing functions, 97

4.2 Numerical solution techniques, 113
6
4.3 Singular perturbations, 124
Stability in the sense of Liapunov, 131
5.1 Basic definitions, 131

5.2 Liapunov's direct method, 147
5.3 Stability of linear systems, 166
5.4 Liapunov's indirect method, 186
5.5 The Lur'e problem, 198
5.6 Slowly varying systems, 218
Input-output stability, 224
6.1 Introduction to L^-spaces, 225

6.2 Input-output and state representation of systems, 231
6.3 Definitions of input-output stability, 233
6.4 Relationships between I/O stability and Liapunov stability, 240
6.5 Open-loop stability of linear 245
systems,
6.6 Linear time-invariant feedback systems, 260
6.7 Time-varying and/or nonlinear systems, 276
Index, 297
Preface
This book is intended as a text for a one-term or one-quarter course in
nonlinear systems, at either the first-year graduate or senior-graduate

level;it is almost self-contained and hence suitable for self-study. The
only prerequisite for using the book is a course in ordinary differential

equations. It is generally not necessary for the reader to have had a
course in linear systems, though it is perhaps helpful to have an under-
standing of the concept of the state of a system. The contents of the
engineers from all branches who are inter-
book should be of interest to
ested in the systems approach, as well as to applied mathematicians,
mathematical economists, biologists, et cetra. The results developed in
the bookare of a sufficiently general nature as to be applicable to all of
these disciplines. Most of the important techniques for the analysis of
nonlinear systems are covered in the book, though the coverage is by no
means encyclopaedic. One of the novel features of the book is a chapter
on input-output stability, presented at an elementary level for the first
time.
The first version of this book was written in 1973, while I was
visiting Subsequent drafts were classroom-tested at both
UCLA.
Concordia University and UCLA. In addition, portions of the book
were also used at Berkeley for one quarter. Generally, the classes con-
sisted of graduate students in both engineering and mathematics. This
experience revealed that the entire book can be covered in about fifty
classroom hours, while most of it can be covered in forty hours.
ix
X Preface
The book contains five chapters besides the introduction. Chapter

2 contains a discussion of various phase-plane techniques for the analy-
sis of second-order systems. In chapter 3, the reader is introduced to
some basic mathematical tools such as normed spaces, contraction
mapping theorem, etc; this is followed by statements and proofs of the
basic existence and uniqueness theorems for nonlinear differential
equations, and some useful solution estimates. Chapter 4 consists of an
introduction to several commonly used approximate analysis tech-
niques. Chapter 5 contains a thorough treatment of Liapunov stability,
including the Lur’e problem. Finally, chapter 6 comprises an elementary
discussion of input-output stability, including the Nyquist, circle, and
Popov criteria for feedback systems. There are numerous examples and
exercises throughout. Two appendixes close the book.
It is now my pleasure to acknowledge all those who helped me in
the writing of this book. I would like, first of all, to thank my wife
Shakunthala for her encouragement and complete moral support
throughout this project. Thanks are also due to Professor Charles A.
Desoer for his thorough review of the manuscript and numerous con-
structive comments, as well as to Professor E. I. Jury for class-testing
the manuscript and for several useful suggestions. I thank Professor
M. N. S. Swamy and Dean J. C. Callaghan, both of Concordia Univer-
sity, for providing excellent logistic support as well as for their moral
support. Professor A. V. Balakrishnan is to be thanked for making
possible my visit toUCLA, during which this project was started.
Finally, thanks to Veronica Markowitz and June Anderson for their
excellent typing.
Montreal M. VlDYASAGAR
Notes to the Reader
1. All items within each section of each chapter (equations, theorems,

examples, etc.) are numbered consecutively. A reference such as
“Theorem (17)” refers to the 17th item within the same section. If a
reference is made an item in another section, the full number

to is
given, e.g. example [5.1(13)] means example (13) in Sec. 5.1.

2. In some places, we write, e.g.
6 = tan
-1 —
*i
This means that 0 is the unique number in [0, 2 n) such that
sin ^ = (^rftir’ co ^ = wtxw ri
Thus tan -1 is a function of both variables x and x 2 and not just of x ,
the ratio x 2 /x Note that tan -1 is well-defined everywhere in R 2

1
.
except at (0, 0).

Nonlinear Systems Analysis
Introduction
1.1
GENERAL CONSIDERATIONS
Nonlinear physical systems, that is, systems that are not necessarily
linear, differ from two important respects:
linear systems in
1. Generally speaking, one can usually obtain closed-form
expressions for solutions of linear systems, whereas this is not
always possible in the case of nonlinear systems. More often,
one is forced to be content with obtaining sequences of approxi-
mating functions that converge to the true solution or with
generating estimates for the true solution. As a result, one may
not have a good “feel” for what makes a nonlinear system
“tick,” compared with a linear system.
2. The analysis of nonlinear systems generally involves mathe-
matics that is more advanced in concept and more messy in
detail than is the case with linear systems.
A mathematical model that describes a wide variety of physical
nonlinear systems is an rcth-order ordinary differential equation of the
type
-l
d ny(t) d n y(t)_
dt
n
=h t, y(t), y(t\ '
-
dr 1
u(t) t> 0
where t is the time parameter, u(-) is the input function (the terms
control function and forcing function are also used), and y(*) is the
. . .
2 Chap. 1 Introduction
output function (or response function ). If we define the auxiliary

functions
2 x^t) = y(t)
3 X 2 (t) = j)(0
4 X„(t ) = "- 1
dt
then the single nth-order equation (1) can be equivalently expressed as

a system of n first-order equations:
5 xft) =x 2 {t)
6 x 2 (t) =x 3 (/)
7 *„-i(0 = *„(0
8 xff) = Xj(0, * 2 (0> . .
. ,
w(0]
Finally, if we define ^-vector-valued functions x(«)* R + —* Rn and

f : R + X Rn X R—> R n by
9 x(0 = [xft), x 2 (t ), . . . ,
x„(0]'
10 f(t, x, w) [x 2 ,
X3 ,
• . . ,
/z(?, Xj, . . . ,
xni w)]
then the n first-order equations (5)-(8) can be combined into a first-
order vector differential equation, namely,
11 x(t) = f [t, x(t), i/(0], t >0

For the system described by (1), the « quantities Xj through x„ con-
stitute a set of state variables and the vector x constitutes a state vector.
,
Similarly, suppose a system with p inputs and k outputs is
described by a set of k ordinary differential equations of the form
12 = h Û î
mi_1>
J2OX
~dt^ • • • > (0> • • •
• • •
uft), • • • ? Wp(0]> 19 . . . 9 k
where uf *), . . . ,
are the input functions and yf *), . .
,y k (*)
are the output functions. As before, define
Sec. 1.1 General Considerations 3
14 x(0 = [x t (f) . . • xn (t)]'

15 u(0 = [u x
(t) . . . u p (t)]'
where we take m = Q 0, and define
16 n =m +1
. . . + mk
Then x(f) is a state vector for the system described by (12), and the
system of equations (12) can once again be equivalently represented by a
single first-order vector differential equation of the form
17 ±{t) =f [t, x(t), u(0], t >0

With this background in mind, we shall devote much of this book
by an equation of the form (17).
1
to the study of systems described
For (17) to truly represent a physical system, we would expect that,
corresponding to each input u(-),
1. one solution (existence).
(17) has at least
2. (17) has exactlyone solution (uniqueness).
(17) has exactly one solution that is defined over
3.
the entire
half-line [0, oo).
4. (17) has exactly one solution over [0, oo), and this solution
depends continuously on the initial condition x(0).

Statements 1-4 are progressively stronger. Unfortunately, without
some restrictions on the nature of the function f, none of these state-
ments may be true, as illustrated by the following examples.
18 Example. Consider the scalar differential equation
19 x(t) = —sign x(t), t> 0; x(0) =0

where the “sign” function is defined by
if x >0
20 sign x(t) =
if x <0
It is easy to verify that no continuously differentiable function *(•)
exists such that (19) is satisfied. Thus statement 1 does not hold for this
system.
22 *(0 = 2 x(t)’ t —° ; *(°) =0

This equation admits two solutions, namely,
iThe exception is Chap. 6, where we shall study distributed systems, e.g., sys-
tems containing time delays.

23 xft) = t
112
24 x 2 (t) = -t U2
Thus statement 1 is true, but 2 is false.
26 x(t ) = +x 1
2
(t), t> 0; x(0) =0
Then over the interval [0, 1), this equation has the unique solution
27 x(t) = tan t
but there is no continuously differentiable function x(-) defined over

allof [0, oo) such that (26) holds. Thus for this system statements 1
and 2 are true, but 3 fails.
It is therefore clear that the questions of existence and unique-
ness of solutions to (17), and their continuous dependence on the initial

condition, are very important. These questions are studied in Chap. 3.
In the last two examples, it was possible to derive the closed-form
solutions to the equations under study, because they were of an
extremely simple nature. However, in most cases, one cannot obtain
an exact solution to the differential equation describing the system
behavior. In such cases, one must be content either with generating
“approximate” solutions or with solution bounds which ,
tell us that
the solution at any time lies in a certain region of the state space.
Both of these are studied in Chaps. 3 and 4.
An important question is that of the well-behavedness in some ,
suitable sense, of the solutions to (17). This is usually called the question
of stability. Ideally, we would like to know whether or not the solutions
to (17) are well behaved without actually solving the system equations
(17). The stability question is studied in depth in Chaps. 5 and 6.
Finally, as a prelude to these more advanced subjects, we shall
study second-order systems in Chap. 2. As we shall see there, a “geo-
metric” approach to second-order systems yields much intuition and
insight.
Problem Determine whether or not each of the following differ-

1.1.
ential equationshas a unique solution over [0, oo), and if so, whether this
solution depends continuously on the initial condition.
(a) x(t) = [*(/)] 1/3 ; x(0) =0

(b) x(t) = -x 2 (t), x(0) = -1
if *(0>0j =0
x(0)
if x(t)< or
]
Sec. 1 2 Autonomy, Equilibrium Points 5
1.2
AUTONOMY, EQUILIBRIUM POINTS
In this section, we shall introduce two definitions that are frequently
used in the sequel. Before proceeding to these definitions, we shall clear
up one small point. Many of the definitions, theorems, etc., that follow
are stated for differential equations of the type
x(0 =f [t, x(/)]
Comparing with (17) of Sec. 1.1, 2 we see that in [1.1(17)] the depen-
(1)
dence of the right-hand side on an input u(*) is explicitly identified,
whereas this dependence, if any, is suppressed in (1). This might mislead
one into thinking that [1.1(17)] describes a “forced” system, whereas
(1) describes an “unforced” system. However, this is not necessarily
the case. In problems of system analysis as opposed to optimal control ,
problems, one is generally concerned with the behavior of a system of

the form [1.1(17)] under a fixed known input. Thus suppose that, in
[1.1(17)], u(-) is a known fixed function, and define fu R+ X R
n
> R
n
:
—
by
U(t, x) =f [t, x, u(0]
Then [1.1(17)] can be rewritten as
x(0 = fjt, x(t)\

which is of the form (1). Therefore (1) can represent either an “unforced”
system or a system with a fixed input.
We shall now introduce two concepts.
[ definition The system described by be autonomous

(1) is said to
if f(t, x) is independent of t and is said to be nonautonomous other-
wise.
[definition] A vector x 0 e R" is said to be an equilibrium

point at time t0 e R+ of (1) if
f(t, x0) = 0, V t > t0
x 0 is an equilibrium point of (1) at time t 0 then it is clear that

If ,
x0 is an equilibrium point of (1) at all times t x

also t0 Furthermore, > .
if (1) is autonomous, then x 0 e R is an equilibrium point of (1) at

n
some time if and only if it is an equilibrium point of (1) at all times.

Therefore we may speak of an equilibrium point of an autonomous
system without specifying the time.
2 Hereafter referred to as [1.1(17)].

(
The physical significance of an equilibrium point is as follows:

Suppose x 0 e R“ is an equilibrium point of (1) at time t Then, .
0
whenever t t >
t 0 the equation ,
x(t) =f [t, x(t)], t > ; x(t : ) = x0

has the unique solution
x(t) = x0 , V > t t1
Conversely, if an element x 0 e R" has the property that the unique
solution of (7) is given by (8) whenever t

l t 0 , then it follows by >
simple differentiation that x 0 satisfies (6), i.e., that x is an equilibrium
0
point of (1) at time t 0 Thus, in other words, x is an equilibrium point
.
0
of (1) at time t 0 if, should any solution x(-) of (1) assume the value x
0
some time /
at t0 ,
>
it then remains at that value x
0 for all t
.
t > x
.
The terms stationary point and singular point are also used in place of
equilibrium point.
Example. Consider the motion of a frictionless simple pendulum,

and let 6 denote the angle of the pendulum from the vertical. Then the
motion of the pendulum is described by
8(f) + j- sin d(t) =0

where g is the acceleration due to gravity and l is the length of the
pendulum. If we define
‘
0
( -9(ty
A
_*2 (0_ M.
then the dynamics of the system are described by the state variable
equations
x (ty
i x 2 (t)
_— i/l) sin X, (t)_
Notice first of all that the systemis autonomous. Next, we have that
*0 = [*io * 2 o]' is an equilibrium point of (12) 3 if and only if
•^20 === 0
sin x 10 =0
i.e., the set of equilibrium points of (12) is the set of points in R2 of the
form
(nn, 0), where n = 0, ±1, ±2, . . .
3
Notice that we need not specify the time because the system is autonomous.
)
Sec. 1.2 Autonomy, Equilibrium Points 7
Because we commonly identify two values of 9 that differ by a multiple

of 2n, this system has basically two equilibrium points, namely (0, 0)
and (0, n). Of these, the first equilibrium point corresponds to the
pendulum hanging straight down, while the second equilibrium point
corresponds to the pendulum being at rest pointing straight up. Of
course, we do not ever expect to find a real pendulum at rest pointing
straight up, because the slightest perturbation (such as wind drafts
present in the room) would knock the pendulum out of this equilibrium
position. This is intimately connected with the question of the stability
of an equilibrium point, which is studied in Chap. 5.
Example. Consider the one-dimensional motion of a particle in

a potential field. Let r denote the position of the particle, the mass m
of the particle, and p(r) the potential energy at r. We assume thatp(r
is a continuously differentiable function of r. The motion of the particle
is described by
_ f[r(t)l
m
m lf=r(r)
where /(/•) = dp(r)/dr denotes the force at r. To obtain a state variable

description, define
x(t)
Xi (0 A
|
X0"
x 2 (0- AO.
Then the state equations are
Xi (0 r x 2(t )
~Vf[xmim_
From (19), we see that the set of equilibrium points of this (autonomous)
system consists of all points of the form ( r 0 0), where fir]) 0. There-
,
=
fore this system is in an equilibrium state if the particle has zero velocity
and is at a position where the force is zero, i.e., if the potential energy
is stationary.
[definition] An equilibrium point x0 at time t0 of (1) is said
to be isolated if there exists a neighborhood N of x 0 in R" such

that N contains no equilibrium points at time t0 of (1) other
than x 0 .
Example. Both the equilibrium points of the system in Example

(9) are isolated. In the system of
Example (16), an equilibrium point
(/„, 0) is isolated if and only if r 0 is an isolated zero
of the function
/(.), i.e., if there exists a <5 > 0 such that f(r)

0 whenever 0 < ^
\r -r < 0 1
5.
|1
22 Example. Consider the linear vector differential equation

23 x(t) = A 0)x(», t >0
Clearly 0 is an equilibrium point of (23) at all times t 0 0. Suppose >
now that A(t„) is nonsingular for some t
0 This means that A(? 0 )x = 0 .
implies x = 0. In this case, 0 is the only equilibrium point at time t0

of (23) and is hence isolated.
24 fact Consider the system (1), and suppose x 0 is an equilibrium point at

time t0 of (1); i.e., suppose (6) holds. Suppose further that f(t 0 •) is con- ,
tinuously differentiable, and define
X)
25 A(? 0 ) =
dx x=Xo
If A(/ 0 ) is nonsingular, then x 0 is an isolated equilibrium point at time t0

of(l).
proof For each x = [x x . . . x„]' in Rn ,

define
/ n \ 1/2
26 IIx|| 2 = (e*?)
The real number x 2 is known as the Euclidean norm of the vector x. 4 If
1 1 1
A(7 0 ) is nonsingular, then there exists a positive constant c such that
27 l|AOo)x|| 2 > c||x|| 2, V X e Rn

Because f(/ 0 , •) is continuously differentiable, we can expand f(7 0 , x) in
the form
28 fOo, x) = f(7 0, x0) + A(/ 0 )(x - x + r(t 0
0) ,
x)
where the “remainder” term
29 lim
llx-x 0 2-»0 ||
M#
II
X Xq 1 2
=0
r(/ 0 , •) satisfies the condition
However, because x 0 is an equilibrium point at time t0 of (1), we

have f(/ 0 ,x 0 ) = 0; therefore,
30 f(/ 0 , x) = A(Y 0 )(x
- x0) + r(^ 0 , x)
Now, pick a number d > 0 such that
31
ll'xîl; ^ T whenever ||x - x 0 1| 2 < d
Such a choice for d is always possible in view of the limit condition (29).
Let N
be the neighborhood of x 0 defined by
32 N = {x G Rn :
II x - x0 || 2 < d]
4 A detailed discussion of norms, including the explanation for the subscript

2, is found in Chap. 3. For the present, it is enough to note that || x || 2 0 whenever >
x 0.
Sec. 1.2 Autonomy, Equilibrium Points 9
We shall show that N contains no equilibrium points at time t 0 of (1) other

than x 0 . By definition (20), this enough to show that x 0 is isolated.
is
Accordingly, suppose x and x N

x 0 we shall show that
e ;
f (t 0 , x) =/ 0. We have, whenever ||x — x 0 2 < d, that ||
||f(/o, x)|| 2 = ||A(/ 0 )(x - x + r(r 0 ,x)|| 2 0)
> |[A(f 0 )(x - x 2 - ||r(r 0 x)|| 2

0 )|| ,
>c||x — x0 1[ 2 - y||x x || 2
= yllx - x0 || 2
x^x
> 0 whenever 0
Hence, whenever xe N and x ^ x we have x)|| > 0, 0, ||f(/ 0 » 2 i.e.,
f x) ^ 0. Thus N contains no equilibrium points at time

(Vo, of (1) other
than x 0 and therefore x 0
,
is isolated. H
Problem 1.2. For the Volterra predator-prey equations
x x
= a Xi + bix x 2
x x
x2 = a 2 x 2 + b 2 X\X 2
(a) Show that (0, 0) is an equilibrium point.
(b) Show that (0, 0) is an isolated equilibrium point if and only if both
a x and a 2 are nonzero.
Problem 1.3. Consider the tunnel-diode circuit of Figure 1.1, where
id =vd — 3v$ + vj Af(vd )
FIG. 1.1
(a) Show that the voltage vd is governed by the equation
§ = —Gv d -f(vd)
(b) Find all the equilibrium points of this system, when (i) G= 0,
(ii) G = 0.1, (iii) G = 2.

:
Second-order systems
2.1
PRELIMINARIES
In this chapter, we shall study various specialized techniques that are
available for the analysis of second-order systems. In subsequent chap-
ters, we shall remove this restriction on the order of the system and
study some techniques of analysis that can be applied to systems of any
order. Obviously, the latter techniques are also applicable to second-
order systems. However, second-order systems occupy a special place
in the study of nonlinear systems, for many reasons. The most important
reason is that the solution trajectories of a second-order system can be
represented by curves in the plane. As a result, many of the nonlinear
systems concepts such as oscillations, vector fields, etc., have a simple
geometric interpretation, in the case of second-order systems. (All the
technical terms used above will be defined shortly.) For these and other
reasons, second-order systems, by themselves, have been the subject of
much research, and in this chapter we shall present some of the simpler
results that are available.
In general, a second-order system under study is represented by
two scalar differential equations
Xi(t) = /ifo X,(t), x 2 (t)\
Xl(t) =fz[t, *l(4 x 2 {t)]

A basic concept in the analysis of second-order systems is the so-called
:
Sec. 2.1 Preliminaries 11
state-plane plot. The state plane is the usual two-dimensional plane with
the horizontal axis labeled x and the vertical axis labeled x2 Suppose
1
.
[xjCO, x 2 (tj\, t e R+ denotes a solution of (1)— (2). Then a plot of x (t)

, t
versus x 2 (t), as t varies over R+ ,

is called a state-plane plot or state-plane
trajectory of the system (l)-(2). In such a plot, the time t is a parameter
that can be either explicitly displayed or omitted. In the special case
where (1) is of the form
3 jci (0 = *2(0
it is customary to refer to the state plane as the phase plane. Corre-
spondingly, in this case one also refers to phase-plane plots or phase-
plane trajectories. This special case arises quite commonly in practice.
In particular, if the system under study is governed by a scalar differ-
ential equation of second order of the form
4 y(t) = g[t, y(t), ;P0)]

then a natural choice for the state variables is to select
5 *i (0 = y{t)
6 x 2 (t) = y(t)
In this case, the system equation (4) is equivalent to the following two
first-order equations
7 Xj(0 =x 2 (t)
8 x 2 (t) = g[t, x t (t), x 2 (t)}

Phase-plane plots also have another useful feature, namely, that it is
easy to reconstruct the implicit parameter from a phase-plane plot. t
Suppose we are given a phase-plane plot, which we denote by Q. Sup-

pose we know that a particular point (x 10 x 20 ) on 6 corresponds to ,
time t 0 (see Fig. 2.1). Typically, t 0 might be the initial time and (x 10 x 20 ) ,
the initial state of the system. Now, if (x lf , x 2f) is another point on 6,

and if it is desired to determine the value of (say t f ) corresponding to
t
(x lf , x 2f), we proceed as follows: If x2 does not change sign along 6

12 Chap. 2 Second-Order Systems
between (x 10 x 20 ) and x lf x 2f), then

, (. ,
h— *0 + f *Ei
Je *2
where the integral in (9) is taken along 0. If x 2 does change sign along
0, then the integral in (9) has to be evaluated as the sum of several
integrals, one corresponding to each segment of 0 along which x does
2
not change sign (see Fig. 2.2). Note that, as x 2f > 0, the integral (9) —
becomes an improper integral. The proof of the relationship (9) is
easily obtained starting from (7) and is left as an exercise for the reader
(see Problem 2.1).
Another very important concept for autonomous second-order

systems is that of a vector field. Consider the autonomous system
*i(0 =/i[*i(0, * 2 (0]

* 2 (0 x 2 (t)]
For this system, it is possible to associate, with each vector (x l9 x 2 ), a
corresponding vector [fi (x 1 x 2 ),f2 (x l9 x 2 )]. The latter vector [f1 (x 1 x 2 ),
, ,
^ 2 )] is known as a vector field. This is made precise in the follow-

ing definition.
[definition] A vector field is a continuous function f: R 2 —> R 2 .
The direction of the vector field f at a point x e R2 is denoted by

0 f (x) and is defined by
Remarks: The definition of 0 (x)

t
is illustrated in Fig. 2.3.
Clearly 0 f (x) is undefined if f(x) == 0.
The utility of the vector field concept is immediately apparent

from (10)— (1 1). Suppose x = (x l9 x 2 ) is a point in R 2 then it is easy to ;
see from (10)— (1 1) that if 0 is a solution trajectory of (10)— (1 1) passing

Sec. 2. 1 Preliminaries 13
through x, then the vector f(x) is tangent to © at x. Hence, in principle

at least, it is possible to construct graphically the trajectories of (10)-
(11)by plotting the vector field f(x). Actually, the concept is very deep
and has many applications, only a few of which are touched upon in
thisbook. Furthermore, the concept of a vector field is applicable to
(autonomous) systems of any order; however, the geometrical visual-
ization is particularly simple for second-order systems. A reader inter-
ested in a deeper study of vector fields in differential equations may
refer to [1] Arnold. 1
Note that it is quite common to refer to f(x) as the velocity vector
field associated with the system of equations (10)— (1 1).
Our objective in this chapter is to present some ways of either

finding the state-plane trajectory of a given system to a reasonably high
degree of accuracy or determining some qualitative features of the state-
plane trajectory without too much work. Throughout this chapter, our
attention will be confined to autonomous systems, because, even though
the general concept of a state-plane trajectory is valid even for non-
autonomous systems, most of the significant results are applicable only
to autonomous systems. For example, in an autonomous system, we have
an oscillatory or periodic solution [i.e., the solution x(f) is periodic in /]
if and only if the corresponding solution trajectory is a closed curve in
R 2 A similar statement for nonautonomous systems is false in general.

.
Finally, in this chapter, we shall bypass the questions of existence

and uniqueness of solutions to (l)-(2) by assuming that the functions f t
and f2 in (1)— (2) are continuously differentiable. This ensures (as shown
in Chap. 3) that (l)-(2) have a unique solution at least locally in t.
That is, given the equations (l)-(2) together with some arbitrary initial
conditions
14 Xj(0) =x 10
15 x2 (0) = x 20
References section located at end of book, before Index. Reference number

is in brackets, followed by last name of author.
]
there exists a number T >

0 such that (l)-(2) together with the initial
conditions (14)— (15) have exactly one solution over [0, T]. In the case of
autonomous systems, some simple additional conditions on the vector
field f [basically stating that the system (1)— (2) “looks linear” for large
x 1 ,
x2] will ensure that in fact a unique solution exists over [0, oo).
Problem 2.1. Prove the relationship (9). Hint: Use (7) to write
x x
(t + At) = Xi(t) + At x 2 (t) + o(At)
Problem 2.2. Show that if 6 is a solution trajectory of (10)— (11) passing

through x, then the vector field f(x) is tangent to 6 at x. Hint: Express (10)
and (11) in difference approximation form as
x^t + At) = x (t) + x Atf^x^t), x 2 (t) + o(At)
x 2 (t + AO = x 2 (t) + AtAlx^t), x 2 (t)] + o(A0
Eliminate At as A* — > 0.
Problem 2.3. Does the function f: R 2 —> R 2 defined by
fl(x U X2 ) = * 2 + [1 “ (*1 + 1/2 ) 1
f2 (x 1 , x 2 ) = -Xi + [1 - (x\ + xl) 1/2 ]
constitute a vector field? Justify your answer. (Hint: Consider the behavior of
f near the origin.)
2.2
—
LINEAR SYSTEMS
We shall begin by studying linear systems, which are simpler to analyze
than nonlinear systems and yet provide much insight into the behavior
of nonlinear systems. The general form for a second-order autonomous

linear system is
(0 =a lx x x (f) + a 12 x 2 (t)
x2 (0 — @ 2 iX \(t) T" ct 22 x 2 (t)

together with the initial conditions
*l(0) = ^10
x2 (0) = x 20
or, in matrix notation,
x(0 = Ax(t)
:
Sec. 2.2 Linear Systems 15
To better understand the behavior of solutions to (5)-(6), it is
helpful to make a transformation of variables. Accordingly, let
7 z (0 -
where M is a constant nonsingular 2x2 matrix with real coefficients.
In terms of the transformed variables z, (5)-(6) become
8 zif) = M AMz(Y) -1
z(0) = M x
_1
9 0
It is known (see, for example, [4] Bellman) that by appropriately

choosing the matrix M, the matrix M AM can be made to have one of
_1
the following forms

1. Diagonal form. In this case,
Xl 0
10 M-‘AM=
0 2-2
where and X 2 are the real (and not necessarily distinct)

eigenvalues of A.
2. Jordan form. In this case,
X 1
11 Mr ‘AM =
0 A
where A is the repeated real eigenvalue of A.

3. Complex conjugate form. In this case,
a B
H
12 M-‘AM =
~B «
where a + jfi, a — j ft
are the complex conjugate eigenvalues of
A (and we choose ft < 0 to be definite).
We shall study each of these cases in detail.
Case 1 Diagonal form: In this case (8)-(9) assume the form
13 zft) = Aj zft)
14 z 2 (t) = X 2 z 2 (t )
15 z,00) = 2, 0
16 2 (0) = z2Q
2
which has the solution
17 2,(0 = 2,06*“
18 z2 (t) = 220 e*“
We may assume at this point that either or X 2 is nonzero, because if

k x and A 2 are both zero, then clearly z x (t) and z 2 (t) are constants, and
the state-plane plot (in the z x -z 2 plane) consists of the single point
(^lo, z 20 ). Thus suppose that ^
0. Then we can eliminate the pa-
rameter t from (17)— (18) to get
19 *2 Z 20
Equation (19) describes the state-plane trajectories of (13)— (16) in the

z 1 -z 2 plane. If k x and A 2 are of the same sign, the trajectories have the
characteristic shape shown in Fig. 2.4, but if k and 1 2 are of opposite
x
signs, the trajectories are as in Fig. 2.5. The arrowheads in Fig. 2.4
correspond to the case where X 2 < < 0; if k and A 2 are both posi-
x
tive, the direction of the arrowheads is reversed, and the

trajectories go
away from the origin as t increases. Similarly, the arrowheads in Fig.
2.5 correspond to the case A x <0< A2 . It should be emphasized that
the trajectories depicted in Figs. 2.4 and 2.5 are in the z x -z 2 coordinate
system; the corresponding trajectories in the x -x 2 coordinate
1
system,
although they will have the same general appearance as those in the
z 1 -z 2 coordinate system, will be a little distorted. This can be seen in
Figs. 2.6 and 2.7, where the trajectories in the x -x 2 coordinate system
1
are illustrated for, respectively, (1) and X 2 of the same sign and (2)
At and A 2 of the opposite sign. In the case where A t and A 2 are both of
the same sign, the equilibrium point (0, 0) is referred to as a stable node
if A and A 2 are both negative and as an unstable node if A and A are
x
2 x
both positive. The rationale is that if A and A 2 are both negative, the
1
trajectories of the system move toward the origin as t increases, whereas

if A and A 2 are both positive, the trajectories of the system move away
x
from the origin as t where A t and A 2 are of oppo-

increases. In the case
site signs, the equilibrium point
(0, 0) referred to as a saddle point
is ,
because, in such a case, if one were to make a three-dimensional plot

with x l9 x 2 and t (or z l9 z 2 and t) as the three axes, the resulting solu-
, ,
tion surface resembles a saddle.
Case 2 Jordan form: In this case (8)-(9) assume the form
20 zft) = Azft) + z 2 (t)
21 z2 (t) = Xz 2 (t)
22 z i(0) — îo
23 z (0) = z 20
2

24 zft) = z 10 e + z20 te
Ac AI
25 z2 (t ) = z 20 eu
Once again, can be eliminated from (24)— (25) the resulting expression
t ;
describing the trajectory is somewhat messy and is left as a problem.

The trajectories in the z x -z2 coordinate system, which are easily obtain-
ed from (24)—(25), are shown in Fig. 2.8 (for the case A <
0). The corre-
sponding trajectories in the x x -x 2 coordinate system are shown in Fig.

2.9. The equilibrium point (0, 0) is again referred to as a stable node if
A < 0 and as an unstable node if A >0.
Case 3 Complex conjugate form: In this case, (8)— (9) become
26 zft) = azjt) + fiz 2 (t)
27 z 2 (t) = -fizft) + ctz 2 (t)
28 zf0) = z 10
29 z (0) = z 20
2
To further simplify matters, let us introduce the polar coordinates
30 r= (zj + z 2 1/2
2)
31 6 = tan" ^ 1
where, as usual, 0 is chosen in the interval [0, iz] if z 2 is positive and in

[tt, 2 n] if z 2 is negative. Then (26)-(27) are transformed into
32 r(t) = ar(t)
33 0(0 = -fi
34 r(t) =r e 0
at
35 0(0 = -pt
4> 0
In the z x -z2 coordinate system, (34)— (35) represent an exponential

spiral. If a > 0, the spiral expands as t increases, while if a < 0, the
spiral shrinks as t increases; if a = 0, the trajectory is a circle. The
equilibrium point (0, 0) is referred to as an unstable focus if a > 0, as
a stable focus if a < 0, and as a center if a = 0.
The trajectories in the z 1 -z 2 coordinate system, corresponding to
each of these three cases, are depicted in Figs. 2.10, 2.11, and 2.12.
z2
FIG. 2.11
Table 2.1 summarizes the various kinds of equilibrium points for

second-order linear systems.
TABLE 2.1
Type of
Eigenvalues of A Equilibrium Point
Xu X2 real, Xx < 0, X2 <0 Stable node
Xu X 2 real, Xx > 0, X2 >0 Unstable node

X u X 2 real, X t X 2 <0 Saddle point
X u X 2 complex conjugates, Re X x >0 Unstable focus
Xu X 2 complex conjugates, Re Xi <0 Stable focus
X u X 2 imaginary Center
Problem 2.4. Eliminate t from (24)-(25), and obtain an expression for

the state-plane trajectory involving only z u z2 z 10 , and z 20 ,
.
Problem 2.5. Consider the electrical circuit shown in Fig. 2.13.

(a) Select the capacitor voltage x x and the inductor current x2 as the
state variables, and show that the network is described by the equations
x x (t)= — 2x (0 - x2 (t) +
1
2 v(t)
x2 (f) = xft) - x2 (t)

(b) Suppose v(t) = 0. Determine the nature of the equilibrium point
in i a
AAAAr- -AAAAr
0 1 F iz *\ 1 H
FIG. 2.13
Sec. 2.3 Nonlinear Systems 21
(0, 0), and find the matrix M that transforms the equations into the appro-
priate canonical form.
Problem 2.6. Suppose the resistor in Fig. 2.13 is replaced by a

general resistance R.
Write the state equations for the network, with v(t) = 0.
(a)
(b)For what values of R is the equilibrium point (0, 0) of this system

(i) a node, (ii) a focus, (iii) a saddle ?
Problem 2.7. For each of the A matrices below:

(a) Determine the matrix M
that transforms A into the appropriate
canonical form;
(b) Sketch the state-plane trajectories in both the z x -z 2 and the x x -x 2
coordinate system; and
(c) Classify the equilibrium point (0, 0) as to its type.
(i) (ii) A =
-r (iii) A=
’1
r
2_ _0 -i_
(iv) A= (v) A=
-r (vi) A =
"0 -r
o 2 -2
2.3
NONLINEAR SYSTEMS
In this section, we shall study four methods for obtaining the state-
plane trajectories of a second-order autonomous system of the type
1 (0 =/ [* 0 * (01
i i ( > 2
2 ± 2 (t) =f [x (t\x
1 1 1 (t)\
2.3.1 Linearization Method
The linearization method, as the name implies, consists of linear-

izing the given system in the neighborhood of an equilibrium point and
determining the behavior of the nonlinear system’s trajectories by study-
ing the resulting linear system. The power of the method lies in the fact
that, except for special cases to be specified later, the method yields
definitive results that are valid in some neighborhood of an equilibrium
point.
The method can be summarized as follows: Suppose that (0, 0) is
an equilibrium point of (l)-(2) 2 and that both fx and f2 are continuously
differentiable in some neighborhood of (0, 0). Define
2 There is no loss of generality in

assuming that (0, 0) is an equilibrium point
of (1)— (2), because if some other point
* 20 ) is an equilibrium point of (l)-(2),
(*io,
then we can always translate the coordinates x x and x 2 in such a way that (x x 0 * 20 ) ,
becomes the point (0,0) in the new coordinate system: More precisely, let jci =•
Xl — Xio X2,
= X2 — X 2 Q.
3 On ,
= §Il
dXj xi = 0,X2 = 0
a \\ a 12
4 A
- a l\ a 22_
Then, by Taylor’s theorem, we can expand fl and f2 in the form
5 /„(*„ x 2) =f t (0, 0) + a u Xi + a x + ri(x u x 12 2 2)
= a Xi xl + a 12 x + fiOi, x
2 2)
6 f2 {x u x 2 ) = a 21 x l + a 22 x + r (x u x
2 2 2)
where rt and r 2 are the remainder terms. [Note that we have used the
fact that /,(0, 0) =/ 2 ( 0, 0) = 0.] Now, associated with the nonlinear
system (l)-(2), define the linear system
7 = a n 4(0 + a £
4(0 12 2 (t)
8 4(0 = « 4(0 + a 22^ {t)

2i 2
Clearly = 0, g = 0 an equilibrium
2 is point of the system (7)-(8),
which is commonly known as the linearization of the system (1)— (2)
around the equilibrium point (0, 0). The linearization method is based
on the fact (proved in Chap. 5) that in most cases the trajectories of the
nonlinear system (1)— (2) have, in some suitably small neighborhood of
the origin, the same qualitative features as the trajectories of the linear
system (7)-(8). Table 2.2 summarizes the situation.
TABLE 2.2
Equilibrium Point £1 0, £ 2 = =0 Equilibrium Point Xi = 0, x2 =0

of the Linearized System (7)-(8) of the Nonlinearized System (7)-(2)
Stable node Stable node
Unstable node Unstable node

Saddle point Saddle point
Stable focus Stable focus
Unstable focus Unstable focus

Center ?
In other words, the matrix if A

does not have any eigenvalues with
zero real parts, then the trajectories of the nonlinear system (1)— (2) in
the vicinity of the equilibrium point x, 0, x 2 0 have the same = =
characteristic shape as the trajectories of the linearized system (7)— (8)
in the vicinity of the equilibrium point 0. c 2 0. The meaning of ^— 1 ~
the last entry in the table can be explained as follows If the equilibrium :
point £)1 = 0, f — 0 of the linearized system (7)— (8)

2 is a center, then the
linearized system exhibits perfect oscillations that neither grow nor
decay with time. Under these conditions, it is logical that, in the original
nonlinear system, the remainder terms r x {x u x 2 ) and r 2 (x u x 2 ) (which
are neglected in the linearization process) are the ones that actually
—
determine the trajectory behavior depending on the nature of these
remainder terms, the oscillations in the nonlinear system can either
grow or decay with time. This is why in this case studying the linearized
system provides no definitive answers about the nonlinear system.
9 Example. Consider the following second-order equation, gen-

erally known as Van der Pol’s equation:
10 y(t) — — y\t)\y(t) + y(t) = 0

ju[ 1
where fi > 0 a constant. By defining the state variables

is
11 Xl (t) = y(t)
12 x 2 (t) = y(t)
(10) is transformed into the pair of first-order equations
13 Xl (t) =x 2 (t)
14 x 2 (t) = —x x
(t) + jx[ 1 — x\(t)\x 2 (t)
The associated linear system is
15 Ut) = Ut)
16 Lit) = j(0 + n$ 2 (t)
The eigenvalues of the associated A matrix satisfy the characteristic
equation
17 A2 — //yl +1=0
Hence, for all positive values of ju<2, the roots of (17) are complex
with positive real parts, so that the equilibrium point Si = 0, £ 2 = 0 of
(15) -(16) is an unstable focus. Referring to Table 2.2, we see that the
equilibrium point x x = 0, x 2 = 0 of (13)-(14) is also an unstable focus.
The Van der Pol equation is studied further in a subsequent example.
2.3.2 Graphical Euler Method
The simple method presented below is quite well suited for con-
structing a single trajectory of (1)— (2), starting from a given initial
point, and is especially suitable for implementation on a digital com-
puter. This method actually amounts to a graphical interpretation of
the forward Euler numerical integration procedure and is an example of
i
a so-called first-order method. More refined numerical techniques,

which are applicable to nonautonomous systems of any order, are
discussed in Chap. 4.
For the system (l)-(2), we can write
18 x x
(t + At) = x x
(t) + A t-f [x x 1 1
(t), 2 (t )] + o(At)
19 x 2 (t + At) = x 2 (t) + At-f^x^t), x 2 {t)] + o(At)
Because we want ultimately to eliminate x (t) by x and

t, let us denote x x
x x
(t + At) by x 1 + Ax and 1 x 2 and x 2 + Ax 2 Then,
similarly define .
because the system (1) -(2) is autonomous we can eliminate At from

,
(1 8)-(l 9) to get
20 x2 + Ax 2 = x2 + JiKXu X Ax
2)
=X + 2 s(x u x 2 ) Ax,
where
fi(x i, x 2 )
21 (Yv
s{Xl v ^
’ X2)
=A 7 &^ 2)
The relationship (20) can be interpreted to mean that whenever a traj-

ectory of (1)— (2) passes through the point (x u x 2 ), its slope at that point
is given by .sfo, x 2 ). If f (x x 2 ) ^ 0,
1 1 ,
the quantity s(x u x 2 ) is well
defined. If fx (x u x 2 ) ^ 0 and f2 (x u x 2 ) = 0, then the tangent at (x l9 x 2 )
to the trajectory of ( 1 )— (2) passing through (x u x 2 ) is horizontal, while
if f (x x 2 )
l 1 ,
= 0 and f (x u x ^ 0, then the tangent
2 2) is vertical. Finally,
if fi( x u x2 ) = fi(x u x t) = 0, the quantity s(x u x 2 ) is undefined, but in
this case (x u x2 ) is an equilibrium point of (1)— (2), so that the trajectory
consists of the single point (x u x 2 ).
The numerical method consists therefore of starting with the
given initial point (x 10 x 20 ), calculating the slope ^(x 10 x 20 ), drawing a
, ,
“short” line segment through (* 10 x 20 ) with the slope s(x 10 x 20 ), picking , ,
the end of the line segment as the new starting point, and repeating the
procedure. There is only one small point to be cleared up. The quantity
s(X}, x 2 ) specifies the slope of the trajectory but does not specify its
direction. But this is easily determined by examining the signs of the
quantities x 2 ) and f2 (x u x 2 ). We can assume without loss of
f1 (x u
generality that not both f (x u x 2 ) and f2 (x u x 2 ) are zero, because if
x
fi( x u x 2 ) = fi( x u x i) = 0, then (x l3 x 2 ) is an equilibrium point of (1)-

(2), and no further analysis is necessary. Thus suppose f (x l9 x 2 ) 0, t
where i is either 1 or 2. If f {x x 2 ) is positive, then x is positive when

i l , t
the trajectory passes through the point (x u x 2 ). Because we are inter-

ested in the evolution of the solution to (1)— (2) as t increases from 0, it
follows that the direction of the trajectory at (x u x 2 ) should be so
chosen that x increases. Similarly, if f (x u x 2 ) is negative, the direction
t t
of the trajectory at (x l9 x 2 ) should be so chosen that x decreases. If t
both fx (x l9 x 2 ) and f2 (x l9 x 2 ) are nonzero, one will reach the same con-
clusions regarding the direction of the trajectory at (x l9 x 2 ) whether he
examines the sign of f1 (x 1 x 2 ) or that of f2 (x u x 2 ).
,
The situation is illustrated in Figs. 2.14 and 2.15. Suppose that

at some point (x l9 x 2 ) both f {x l9 x 2 ) and f2 (x l9 x 2 ) are positive. Then
x
the slope x 2 ) is also positive. To determine the direction of the

trajectory as it passes through (x u x 2 ), we observe that because
f {x l9 x 2) x
l9 x 2 ) corresponds to increasing
is positive, the trajectory direction at (x
x l9 as in Fig. 2.14. Alternatively, we could have reasoned that because

f2 (x u x 2 ) is positive, the trajectory direction at (x l9 x 2 ) corresponds to
increasing x 2 The conclusion is the same as before.
.
*2
j/h< /i <o
0
FIG. 2.15 I
Now suppose that at (x u x2 ) both f (x u x 2) and f2 (x l9 x 2 )

1 are
negative. Then x2 ) is still positive, so that the trajectory has a
positive slope as it passes through the point (x l9 x 2 ). However, the
direction of the trajectory isshown in Fig. 2.15.
as
In the special case where x 2 = x l9 i.e., when the state plane is
actually a phase plane, the considerations of the direction of the trajec-
tory can be somewhat simplified. In this case, (l)-(2) simplify to
22 x —= X 2
i
23 X2 =f (x „ x )
2 2
and the slope function s(x l9 x 2 ) assumes the form
24 s(x 1; x2 ) = '
x2
’ —
Thus the first thing to be noticed is that whenever x 2 = 0 we have
^(î, x 2 )
= oo unless /2 (x 1? x 2) = 0. Geometrically, this means that
whenever a trajectory of (22)-(23) meets the x r axis [at say (x 1? 0)] either
(x 1? 0) is an equilibrium point of the system (22)-(23) or else the tangent
to the trajectory of (22)-(23) at (x l9 0) is vertical. In addition, the direc-
tion of the trajectories of (22)-(23) can be determined by inspection:
Whenever the point (x l5 x2) lies in the first or second quadrants, we
have x 2 > 0, so that the direction of the trajectory of (22)-(23) should
be so chosen as to correspond to increasing x t . Similarly, whenever
(jc 15 x2) lies in the third or fourth quadrants, we have x 2 < 0, so the
direction of the trajectory at (x 1? x 2 ) corresponds to decreasing x If t
.
(x 1? x 2 ) lies on the Xj-axis, i.e., if x 2 = 0, then the direction of the traj-

ectory is vertical upward if /2 (x i, 0) > 0 and vertical downward if
/2 (x i, 0) < 0. Of course, if /2 (x l5 0) = 0, then (x l5 0) is an equilibrium
point of (22)-(23).
2.3.3 Isocline Method
The isocline method is a procedure for sketching trajectories that

is not particularly efficient if one wishes to sketch a single trajectory of
(l)-(2) but that is quite useful if one wants to sketch several trajectories
starting from several initial points, in order to obtain a understanding
of the overall behavior of the trajectories. In some cases, one is quickly
able to spot potential periodic trajectories using this method.
The procedure is as follows: The equation
25 i(x 1; x2) = = constant = c

MXj, x 2 ;
determines, for each value of the constant a curve in the x t -x 2 plane c,
along which the solution trajectories of (1)— (2) have the slope c. Thus,
whenever a solution trajectory of (1)—(2) crosses the curve defined by
s(x u x 2) = c,
must do so with a slope of c. The procedure is to plot
it
the curve j(x 1; x 2 ) = c in the x 2 -x 2 plane and along this curve draw
“short” line segments having the slope c. Such a curve is known as an
isocline.The directions of these line segments are determined as in Sec.
2.3.2. The procedure is repeated for sufficiently many values of the
constant c, so that the x -x 2 plane is filled with isoclines, and one is
1
then able to rapidly sketch the trajectories of (1)— (2) starting from any
initial point.
26 Example. Consider again the equations of the Van der Pol oscil-
lator, namely,
27 *i = *2
28 X2 = + /^(l - *l)*2
The “slope function” s(x u x 2 ) associated with this system of equations
29
is
*(*,. *,) - =a + ^
The curves x 2 ) = c corresponding to fi = 1 and for various
values of c are shown in Fig. 2.16. With the aid of this construction, one
can easily sketch the trajectory of the Van der Pol oscillator starting
from, say, xÔ) = —2, x 2 (0) =
Also one can see that there
3. may
possibly be a periodic solution as indicated by the bold line.
2.3.4 Vector Field Method
The vector field method is quite similar to the isocline method,

except that one plots the vector field itself rather than isoclines. Given
the system (1)— (2), one plots, at each point (x u x 2 ), the corresponding
vector [f1 (x u x 2 ),f2 (x l9 x 2 )] (of course with suitable scaling). In the end,
one obtains a vector field diagram that clearly indicates the nature of
the trajectories of the system. The procedure is illustrated in the follow-
ing example.
30 Example. Consider the system of equations

31 x 1
= —ax + x
bx x 2
1
32 x2 = —bx x 1 2
The system (31)— (32) is a very elementary model for the spreading of
disease within a population. Here x x
denotes the number of infected
people, while x 2 denotes the number of noninfected or “susceptible”
people. Equation (32) states that noninfected people become infected at
a rate proportional to x t x 2 which is a measure of the interaction be-
,
tween the two groups. The expression (31) for consists of two terms:
(1) —ax x which is the rate at which people die from disease or survive
,
and become forever immune, and (2) bx x 2 which is the rate at which
x ,
previously noninfected people become infected. Note that the system

(31)— (32) is a special case of the predator-prey equations of Example
[2.4(70)].
With regard any point of the form (0, x 2 ) is an equi-
to (31)— (32),
librium point; i.e., if initially is no infection, there will be none
there
thereafter. Hence the system exhibits a continuum of equilibrium points
on the x 2 -axis. To study the behavior of the solution trajectories, we
plot the velocity vector field of (31)— (32), with a = 2, b = 1 ;
the result
is shown in Fig. 2.17. Note that the velocity vector field is plotted for
all x l9 x 2 even though the first quadrant (i.e., x >0,x 2
, 0) is t > the
only practically relevant case.
From we can see the qualitative behavior of the solution
Fig. 2.17,
trajectories. Suppose (x 10 x 20 ) in the first quadrant is the initial condi-
,
tion. By examining the vector field, we see that as t * oo, xît) > 0 — —
FIG. 2.17
:
while x 2 (t) approaches a nonzero limit. In other words, eventually the

infection does die down. However, the larger the initial ratio x 10 /x 20 ,
the smaller the final ratio x 2 (oo)/x 20 .
Problem 2.8. Find all the equilibrium points of the Volterra predator-
prey equations
*i = —x + t
x x2
x
x2 = x2 — x x x2
Linearize the system around each of the equilibrium points, and determine,
if possible, the nature of each equilibrium point. ( Answer one center, one
saddle.)
Problem 2.9. Plot the isoclines for Rayleigh’s equation
*i = x2
*2 = —X + t 6 (x 2 —y)
for (a) e = 1, (b) e = 0.1
Problem 2.10. Plot the velocity vector field for the pendulum equation
xx = x2
x2 = — sin*!
Problem 2.11. Consider the nonlinear circuit in Fig. 2.18. Suppose the
voltage-current relationship of the nonlinear resistor is given by
vr = ij - 3i'i + 3 ir AfOr)
FIG. 2.18
(a) Select the capacitor voltage x and the inductor current x 2

x
as the
state variables, and show that the state equations are
x x = v — x — x2 x
*2 = *1 — fix 2 )
(b) With v = 0, calculate the equilibrium points of the system.
(c) Linearize the system around each of the equilibrium points, and
determine the nature of the trajectories around each point.
)
2.4
PERIODIC SOLUTIONS AND LIMIT CYCLES
2.4.1 Introduction
Some autonomous systems exhibit periodic solutions. For exam-

ple, consider a simple harmonic oscillator, which is described by the
linear equations
1 Xi(0 =x 2 (t
2 x2 (t) = -x x (t)
The solution of (l)-(2), subject to the initial conditions
3 *i(0) = x lQ
4 x2 (0) = x 20
is given by
5 x t (t) =r 0 cos(—t +0 0)
6 x 2 (t) =r o sin (— t +0 O)
where
7 r0 = (xf 0 + xl 0 )
1/2
8 0O = tan- 1 iaa
*10
Thus the solution of (l)-(2) is periodic regardless of what the initial
conditions are. Furthermore, the entire state plane is covered with

periodic solutions of (1)— (2), in the sense that given an arbitrary point
(x u x 2 ), one can always find a periodic solution of (1)— (2) passing
through it.
In contrast, consider the system of nonlinear equations
9 x t
= x + a x^f} — x\ — x\)
2
2
10 x2 = —x + a x (P — xf — x\)
x 2
2
where we have suppressed the dependence on t in the interests of brev-

ity. The reader can easily verify that the velocity vector field of the
system (9)— (10) is the sum 3 of two vector fields: (1) the velocity vector
field of the system (1)— (2) and (2) a radial vector field that is outgoing
for xf + x\ /? [Note that a vector

2
.
field f(x) is radial if the vector f(x) is always aligned with the vector x,
for all x.] Now, if we introduce the polar coordinates
3 The sum of two vector fields f(x) and g(x) is defined by
(f + g)(x) = f(x) + g(x)

.
Sec. 2.4 Periodic Solutions and Limit Cycles 31
11 r = (xf + xl) 1/2
0 = tan —
-1
12
x i
then the equations (9)— (10) are transformed into
13 r — —r
(X,r(fi
2 2
)
14 (j)
= —l
It can be easily verified that the solution of (13)— (14) is
15 =
+ c f-(1 0
2 ^ a ') 1/2
16 0(0 = 0 ~t O
where
Thus the system (9)— ( 1 0) has only one periodic solution, namely r 0 = /?,
i.e., x\ + x\ = /?
2
. Furthermore, whenever r0 ^ 0, all solutions of
(9)-(10)approach this periodic solution as t » oo. This example differs —
from the example of a simple harmonic oscillator in that the periodic
solution in the present case is isolated; i.e., there exists a neighborhood
of it that does not contain any other periodic solutions.
18 [ definition ] A limit cycle of [2.3(1 )]— [2 3 (2)] . is a periodic solu-

tion of [2.3(1 )]— [2 3 (2)] .
By convention, we do not regard an equilibrium point as a peri-

odic solution or as a limit cycle. Also, a simple consequence of defini-
tion (18) is that a limit cycle can be either isolated or nonisolated.
In the remainder of this section, we shall present some results
pertaining to limit cycles in nonlinear systems.
2.4.2 Bendixson's Theorem
19 Theorem Suppose D is a simply connected 4

domain in R2 such that
the quantity Vf(x) defined by
20 Vf(x) = X2 ) + *2)
4 A connected region can be thought of as a set that is in one piece, i.e., one in
which every two points in the set can be connected by a curve lying entirely within the
set. A set is simply connected if (1) it is connected and (2) its boundary is connected.
One can also think of a simply connected set as one that can be obtained by con-
tinuously deforming a circle. For example, an annular region is connected but not
simply connected.
.
is not identically zero over any subregion of D and does not change
sign in D. Then D contains no closed trajectories of [2.3(1 )]— [2 3(2)] .
21 Remarks: Theorem (19) gives conditions under which a region

does not contain a periodic solution and as such gives a sufficient
condition for the nonexistence of a periodic solution.
proof of theorem (19) Suppose J is a closed trajectory of [2.3(1)]—

Then at each point x = fc, x 2 ) e J, the velocity vector field f(x)
[2.3(2)].
= [f\(xu x 2 ),fi(xu x 2 )] is tangent to J. Let n(x) denote the outward nor-
mal to J at x. Then f(x)-n(x) = 0 for all x e J. In particular,
22 f (x) • n(x) dl =0
J
But by the divergence theorem,
23 J^f(x)*n(x) dl = JJVf(x) ds (= 0)
s
where S is the area enclosed by /. Therefore, for (23) to hold, we must have
either Vf (x) = 0 V x e S or (ii) Vf (x) changes sign over the region S.
(i)
But if S is a subset of D, neither can happen. Hence D contains no closed

trajectories of [2.3(1)]-[2.3(1)]. Note that if V/(x) is a polynomial in the
components of x, then one can systematically check whether or not V/(x)
changes sign over a given region; see [15] Jury.
24 Example. Consider the application of Theorem (19) to the linear

system of equations
25 Xi =a xl x x + 012*2
26 X2 == 021*1 ~ 022*2
or, in matrix form,
27 x = Ax
We know from Sec. 2.2 that a necessary and sufficient condition for the
system (25)-(26) to have periodic solutions is that the matrix A has two
nonzero imaginary eigenvalues. Because the eigenvalues of A satisfy the
characteristic equation
28 A2 — (a + 0
tl 2 2)A + (011022 01202 1 ) =0
it is clear that the system (25)-(26) has periodic solutions if and only if
29 0n + 02 2 = 0
30 011022 — 012021 >0

Equivalently, a necessary and sufficient condition to have no periodic
solution is that either (29) or (30) is violated.
Applying Theorem (19) to the present case, we get
31 Vf(x) = a + a 22
lx V x g R2
Hence we conclude that if a n + a 12 ^ 0, then no periodic solutions
exist, which is in accordance with the previous discussion.
32 Example. Consider the system of nonlinear equations
33 x x
= x + x x\
2 x
34 x2 = —X + x\xx 2
The linearization of this system around the equilibrium point (0, 0) is
35 x 1
=x 2
36 x2 = —x 1
which exhibits a continuum of periodic solutions. However, for the

nonlinear system,
37 Vf(x) = xf + x\ > 0 whenever (x l9 x2 ) ^ (0, 0)

Thus over every region in R 2
that does not consist of just a single
point, we have that Vf(x) is not identically zero and does not change
sign. Therefore this system has no periodic solutions anywhere in R2
(other than the trivial one x x
— x2 = 0).
38 Example. In applying Theorem (19), the assumption that D is a
simply connected region in crucial — it is not enough for D to be just
connected. To see this, consider the system (9)-(10), and let D be the
annular region
39 D= |(x 1? x 2 ): < x\ + x\ < 2/? 2
For this example, we have

40 Vf(x) = 2aj? — 4a(xJ + x$)
2
which is everywhere nonnegative on D. Yet D contains a periodic

solution. The reason, of course, that Theorem (19) is inapplicable here
is that D is not simply connected, even though it is connected.
2.4.3 Poincare-Bendixson Theorem
The Poincare-Bendixson theorem can be used to prove the

existence of a periodic solution, provided a region M satisfying certain
conditions can be found. The strength of this theorem is its generality
and its simple geometric interpretation. The weakness of the theorem is
the necessity of having to find the region M. We shall first introduce a

definition.
.
41 [definition] Let x(Y) be a solution trajectory of [2.3(1)]— [2.3(2)].

A point z g R2 is said to be a limit point of this trajectory if there
exists a sequence ( t]y
j° in R+ such that tn — > oo as n —
> oo and
x(t„) — z as n —»
> oo. The set of all limit points of a trajectory x(Y)
is called the limit set of the trajectory and is denoted by L.
42 Remarks : Basically, a limit point of the trajectory x(t) is a point

z which has the property that, as time progresses, the trajectory
passes arbitrarily close to z infinitely many times. We shall encoun-
ter limit points and limit sets again in Chap. 5.
43 Theorem (Poincare-Bendixson) Let
44 S = {x(Y), > 0} t
denote a trajectory in R2 of the system [2.3(1 )]— [2 3 (2)] .

,
and letL
denote its limit set. If L is contained in a closed bounded region M
in R 2
and if M contains no equilibrium points of [2. 3( 1)]— [2 3(2)], .
then either
(i) S is a periodic solution of [2.3(l)]-[2.3(2)], or
(ii) L is a periodic solution of [2.3(1 )]— [2 3 (2)] .
We shall omit the proof because it is beyond the scope of this
book.
45 Remarks: Roughly speaking, what Theorem (43) states is the fol-

lowing: Suppose we can find a closed bounded region in R 2 such M
that M
does not contain any equilibrium points of [2.3(1 )]— [2 3 (2)] .
and such that all limit points of some trajectory S are contained in
M. Then M
contains at least one periodic solution of [2.3(1)]—
[2.3(2)]. In practice, it isM contains
very difficult to verify that all
M closed, can
the limit points of a trajectory. However, because is it
be shown that some trajectory x(0

if eventually contained in M, is
i.e.,there a time < oo such that x(0 e M V > then

exists t0 t t0 ,
L contained in M. Thus the theorem comes down to

is we can this : If
find a closed bounded region M containing no equilibrum points

such that some trajectory eventually confined to M, then M con-
is
tains at least one periodic solution. Now, a sufficient condition for

a trajectory to be eventually confined to is that, at every point M
along the boundary of Af, the velocity vector field always points into
M. If this is the case, then any trajectory originating within must M
remain in M, and hence M
contains at least one periodic solution
trajectory. (This is depicted in Fig. 2.19.)
46 Example. Consider once again the system (9)-(10), and let M

be the annular region defined by
47 M = {(x u x 0.9 < x\ + x\ <

2 ): /?
2
l.lyff
2
}
Then M contains no equilibrium points of the system (9)— (10). Further-

more, a sketch of the velocity vector field for this system reveals that,
all along the boundary of M, the vector field always points into M, as
depicted in Fig. 2.19. Hence we can apply Theorem (43) and conclude
that M contains a periodic solution.
48 Example. In applying Theorem (43), the condition that M should
not contain any equilibrium points is very important. To see this, con-
sider the system
49 X l
= —X + x
l 2
50 x2 = —X — X
x 2
The velocity vector field for this system is sketched in Fig. 2.20. If we
chooseM to be the unit disk centered at the origin, then all along the
boundary of M the velocity vector field points into M. Hence all trajec-
*2
FIG. 2.20
tories originating within M

remain within M. The same conclusion can
be reached by analytical reasoning, because in polar coordinates, the
system (49)-(50) becomes
51 r = —r
52 0 = -1
53 r(t)= rê~ f
54 0(0 = 0o — t
However, M contains no (nontrivial) periodic solutions.
2.4.4 Index Theorems
The concept of index is a very powerful one, and the results given
below only scratch the surface of the many results that are available.
Unfortunately, the arguments involved in index theory are well beyond
the scope of this book. Hence we shall content ourselves with the few
results given below, most of which are given without proof. For further
results, see [18] Nemytskii and Stepanov.
The definition below introduces the concept of the index of a
vector field.
55 [definition] Suppose D is an open, simply connected subset of

R and suppose f\R —>R a vector field on R Suppose that
2
,
2 2
is
2
.
D contains only isolated equilibrium points of the system

56 ±(o = nm
Let / be a simple, closed, positively oriented Jordan curve in D
that does not pass through any equilibrium points of (56). Then
the index of the curve J with respect to the vector field f, denoted by
7f (/), is defined by
57 /,(/) = 2- dd,(x u x2 )
where 0 (x u x 2 ) t
is the direction of the vector field f at the point
Ol. x2 ).
58 Remarks : A positively oriented curve is one that is traversed in

a counterclockwise sense. The index of J with respect to f is noth-
ing but the net change in the direction of f, as x traverses around
J, divided by 2n. Clearly It (J) is always an integer.
59 [definition] Let p be an isolated equilibrium point of (56). Then

the index of p is defined as /£ (/), where J is any suitable Jordan
.
curve such that (i) p is interior to J and (ii) J contains no equilib-

rium points in its interior other than p.
We shall now state some facts without proof.

60 fact suppose J contains no equilibrium points of (56) in its interior. Then
h(J) = 0.
61 fact The index of a center, focus, and node is 1 and that of a saddle is
— 1. (This fact can be easily verified by a sketch of the velocity vector field
near a center, focus, node, and saddle.)
62 fact Suppose J contains in its interior a finite number of equilibrium

points of (56), say pu ... ,p„ (which are of necessity isolated). Then
63 It (J) = ± I (Pi ) t
i=i
64 fact Let f and g be two vectorfields on R 2 and let 0 and 9 g denote the, f
directions of f and g, respectively. Let J be a simple, closed, positively

oriented Jordan curve such that (i) \9 t — 9 g along / (i.e., the vector
\
<n
fields f and g are never in opposition on J ) and (ii) J does not pass through
any equilibrium point of (56) or of the system
65 x(/) = g[x(t )]
Then
66 It (J) = Tg (J)
67 fact Let J be a simple, closed, positively oriented trajectory of (56). Then
68 /,(/) = 1
(Note that the vector field f is always tangential to /.)
On the basis of these facts, we can state the following general

theorem.
69 Theorem Every closed trajectory of (56) contains at least one equi-

librium point of (56) in its interior. If the system (56) has only iso-
lated equilibrium points, then every closed positively oriented trajectory
of (56) contains in its interior a finite number of equilibrium points
such that the sum of their indices is 1
70 Example. We shall study in detail the Volterra predator-prey
equations
71 = —x + x t x
x2
72 X2 =X —Xx 2 t 2
We shall digress briefly to discuss the rationale behind the above model.
In (71)— (72), x x denotes the number of predators (foxes, say), and x
2
denotes the number of prey (rabbits). If x 2 = 0, (71) reduces to x x
= ~x u which implies that in the absence of prey the number of preda-
tors will dwindle exponentially to zero. If x ^ 0, (71) shows that x
2 x
contains an exponential growth term proportional to jc The situation .

2
is just the opposite in the case of the prey. If x == 0, x will grow expo-
x 2
nentially, while if x x ^ 0, x 2 contains an exponential decay term pro-
portional to x t
.
The velocity vector field for the predator-prey system is shown in

Fig. 2.21. Clearly there are two equilibrium points, namely (0,0) and
(1, 1). By linearizing the system (71)-(72) around each of these points,
we see that (0, 0) is a saddle, while (1, 1) is a center. Therefore, the index
of (0,0) is— 1, while that of (1, 1) is 1. Now, by Theorem (69), any
closed trajectory of the system (71)-(72) contains (1, 1) in its interior

and does not contain (0, 0) in its interior. Thus, by examining the index
alone, we can derive a great deal of qualitative information about the
(possible) closed trajectories of the system.
2.4.5 An Analytical Method
In this subsection, which could equally well have been entitled

“miscellaneous,” we shall present a technique for obtaining analytical
expressions for the closed trajectories of some nonlinear systems that
exhibit a continuum of periodic solutions. Rather than presenting a
general theorem, which would have to be rather weak because of all the
possible pathological cases, we shall illustrate the method by means of

two examples.
The basic idea of the method is as follows: Given the system
73 *i=/i(*i,*2)
74 x2 — fi(x ij * 2)
suppose we can find a continuously differentiable function V D —>

: R,
where D is some open subset of R such that ,
75 V(x l9 x 2 )A-g^'fi(x l9 x 2 ) + J^/ (*i>*2) = 0 2 V (x 1? x 2 ) e D
The function L(x 1? x 2 ) is known as the derivative of V along the trajec-

tories of (73)— (74), because if [xjfr), x 2 (t)] is a trajectory of (73)— (74),
then the time derivative of the function F[Xi(0, * 2 (0] is given by
x 2 (f)]. (The details of this argument can be found in Sec. 5.1.)
Now, suppose (x 10 x 20 ) e D and let e denote the solution trajectory
, ,
of (73)— (74) originating at (x 10 x 20 ). By (75), we see that the time

,
derivative of V[x x (t ), x 2 (f)] is zero along 6, so that Fjx^), x 2 (0] is con-

stant along 6. In fact
76 ^ 2 (0] = V{x^ x 20 ) V ^ >0

Let us now consider the set
77 5 = {(x l5 x 2 ): F(x 1? x 2 ) = K(x 10 x 20 )}

,
Then C is a subset of S. In particular, if S is a closed curve, we can
conclude (under some relatively mild additional assumptions) that e

itself is a closed trajectory of (73)— (74).
Of on the choice of the function V. If we

course, a lot depends
choose F(x 1? x 2 ) =x 2 ), then (75) is satisfied, but the set
1 for all (x 1?
S in (77) is the whole space R 2 and as a result no insight has been ,
gained using this particular V function. However, in some cases, by

properly choosing V, we can show that the family of sets
78 {(x l5 x 2 ): V(x u x 2 ) = c)
defines a continuum of closed trajectories of (73)-(74).
79 Example. We return to the predator-prey equations (71)— (72),
and try a function V of the form
80 V(x u x 2 ) = hfx + x ) h 2 (x 2)
where h and h 2 are to be adjusted so that

x (75) is satisfied. We have
81 L(xi, x2) = /zi(xj)( Xj + XjX + h 2) 2 (x 2 )(x 2 XjX 2 )
where the prime denotes differentiation with respect to the appropriate

argument. For (75) to hold, we must have
82 h\(x i )x l (x 2 — 1) + h' (x )x 2 2 2 (l
— Xj) = 0
which can be rearranged as
83 h\( Xl ) y*!- = h'2 (x 2 )

**-
1 X i
L X2
Because the right-hand side of (83) depends only on x 2 and the left-
hand side depends only on x u in fact both must equal a constant; in
other words,
84 h\(x 0 j-^1
1 ——
x
=c
1
85 h'2 (x 2 ) 7 -^— = c
1 — X2
where c is a real constant. It can be easily verified that the solution of
(84)-(85) is
86 h x (x x )
= c(ln x j
— Xj)
87 h 2 (x 2 ) = c( In x2 —x 2)
Hence (75) is satisfied if we choose

88 V(x l9 x 2 ) = (In x 1
— x + In x — x2
t 2 )
where we have dropped the arbitrary constant c without loss of gener-

ality. Now, for the above choice of F, any set of the form (78) is actually
a closed curve. Hence the family of curves defined by
89 In x t
— x + In x — x = const.
1 2 2
constitutes closed trajectories of the predator-prey system. Note that V

is defined only for x x > 0 x2
, > 0 ,
i.e., in the first quadrant.
90 Example. Consider the pendulum equation
91 Xj =x 2
92 x2 — — sin Xi
Let us once again choose V(x l9 x 2 ) to be of the form (80). Then, for
(75) to be satisfied, we must have
93 h\{x 1 )x 2 — h'2 {x 2 ) sin x x
=0
which implies, as in Example (79), that
94 ) = hMl = const. = c
sm Xi x2
Solving (94) gives
Sec. 2.5 Two Analytical Approximation Methods 41
95 î(-î)
— — C COS X 1
96 h 2 (x 2 ) = c^~
Hence the family of curves
97
^ — cos Xj = const.
constitute a continuum of closed trajectories for the simple pendulum.
98 Remarks Examples (79) and (90) illustrate how the method presented
:
here can sometimes yield good results. However, it should be clear that
(1) a function V(x u x 2 ) of the type (80) does not always work and (2)
even if it does, there is no guarantee that all closed trajectories of the
system are of the form (78). Despite these limitations, however, the
method is of some value, as indicated by the two examples.
Problem 2.12. Consider the system of equations
*1 = x2
*2 = -*(*i) - Kx 2 )
where g, h: R —* R are continuously differentiable functions. Using Bendix-

son’s theorem, show that this system has no periodic solutions if h'(£)^ 0
V f e R.
Problem 2.13. Sketch a vector field with exactly one node and one
saddle point. Show that it is not possible to continuously deform this vector
field in such a way that there is a periodic solution enclosing both the node and
the saddle point.
Problem 2.14. Using the method of Sec. 2.4.5, derive an expression for
the closed trajectories of the system
*1 = x2
x2 = —g(x i)
25
TWO ANALYTICAL APPROXIMATION METHODS
In this section, we shall describe two techniques for obtaining analytical
expressions that approximate the periodic solution of second-order
nonlinear differential equations. In contrast with the method presented
in Sec. 2.4.5 (which gives exact expressions, if it works), the two methods
presented here are only approximate. However, they have the advantage
of having a wide range of applicability and of enabling one to study the
so-called “slowly varying” oscillations. It should be emphasized that,
^ t t
depending on the particular problem to which they are applied, one

technique might work better than the other. Moreover, the two methods
presented here are only a portion of the many techniques that are
available, and the reader is referred to [6] Blaquiere for a more com-
plete account.
2.5.1 Krylov-Boguliubov Method
The Krylov-Boguliubov method is applicable to differential

equations of the type
1 y(0 + y 0 = Mf[y( 0, X0]

(
The class of equations of type (1) includes many of the commonly

encountered ones, such as the Van der Pol equation and the pendulum
equation. Note that, in (1), the angular velocity of the oscillations cor-
responding to n = 0 has been normalized to 1. Clearly this presents no
limitation and can always be achieved by scaling the variable t.
If fi= 0, the solution of (1) is of the form
2 y(t) = a sin (t + i)
<t
With this in mind, we assume that the solution of (1) (when n ^ 0) is
representable in the form
3 y(t) = a{t) sin + 0(0] [t
4 y(t) = a(t cos + 0(0]

) [t
where a( •) and 0(-) are “slowly varying”; i.e., a(t) and 0(0 are “small.”
Actually, if >’(•) is given by (3), we have
5 y(t) = a(0 sin + 0(0] + «(0 cos + 0(O][1 + 0(0]

[t [t
Hence, for (4) to be valid, we must have
6 a sin (t + 0) + a0 cos (t + 0) = 0
where we have suppressed the dependence of a and 0 on t in the interests
of brevity. Substituting for y and y from (3) and (4) into (1) gives
7 a cos (t + 0) — a0 sin + 0) = (f fif[a sin (f + 0), a cos + 0)] (t
Equations (6) and (7) represent two linear equations in the unknowns a
and 0. Solving for a and 0 gives
8 a — n cos (t + 0)/[o sin (t + 0), a cos (t + 0)]

9 0 = — sin ( + 0)*/[a sin + 0), a cos + 0)]
( (t
Ifwe want to find solutions of (1) of the form (3) where o(-) is
periodic, we impose an extra condition, namely

g(T) - g( 0) _ Q
T
or, equivalently,
11 dt = 0
T
where T is the period of «(•)• Unfortunately, (11) cannot be used
directly because the period Tis in general dependent on p and is hence
unknown. To get around this difficulty, we observe that «(•) goes
through one complete period as the phase 0 = t + 0(0 goes from 0 to
2n. Thus we change the variable of integration in (11) from t to 9; then
the limits of integration become 0 and 2n and ,
the integrand d(t)
becomes [using (8)]
12 d(t) — > p cos 9 •f{a sin 0, a cos 9)
Finally, we make the approximation
d9__df
13
2
n~ T
Equation (13) expresses the fact that as t varies by T, 9 varies by 2n.
Thus (11) becomes
2n
C
14
1
pcos 9-f(a sin 0, a cos 9) d9 =0

2^ J
Similarly, if we require 0 also to be periodic, this leads to the condition
15 sin 9-f(a sin 9, a cos 9)d9 =0

J
Equations (14) and (15) can be used in the following way. Suppose we
are interested in approximating the periodic solutions of (1) by func-
tions of the form
16 y(t) = a sin (1 + S)t
where a and <5 are now constants. In this case, (14) and (15) simplify to
r 2 tc
17 cos 9 •/(a sin 0, a cos 0) d9 =0
Jo
18 f sin 0 mf {a sin 0, a cos 0) d9 —0

Jo
Because a is a constant, f{a sin 0, a cos 0) is a periodic function of 0

with period 2n and hence can be expanded in a Fourier series. Hence
(17) and (18) state that in order for (16) to approximate (to the first
order in //) a periodic solution of (1), it is necessary for the first harmonic
of the periodic function f{a sin a cos 0) to be zero. This requirement is
0,
sometimes called the principle of harmonic balance. We shall encounter
1
the same reasoning again in Chap. 4 when we discuss the so-called

describing function method.
Note that ji does not appear in (17) and (18), because when
we study the periodic solutions of (1) we are in effect examining
the steady-state oscillations of (1), and ji does not affect the steady-
state solutions. However, fi is prominently present when we study the
so-called “slowly varying” or transient solutions of (1). For this pur-
pose, we make the approximations
19 d(t)
~ a(T - ) a(0)
T
20 0(0
where T is the period of the steady-state oscillations. However, as in

studying the steady-state oscillations, we have
21 = -L n cos 9 -f(a sin 9, a cos 9) dO

J
22
<t>(T) - m= _L j* _JL sin q .f( a sin e, a cos 0) d0
Hence the approximate equations describing the slowly varying oscilla-
tions of (1) are
2n
1 C
23 ^ = 2n j V cos 9-f(a sin 9, a cos 9) dQ
24 0 =^ ~ sin 9>f(a sin 9, a cos 9) d9

J
25 Example. We apply the Krylov-Boguliubov method to Van der
Pol’s equation, which can be written in the form
26 y(t) + y(t) = jiy(t)[ 1 - y\t)\

This is of the form (1) with
2i f(y y),
= y( - y 1
2
)
Hence
28 f(a sin 9, a cos 9) = a cos 9(1 — a 2 sin 2 9)
= (a — cos 9 -j-
-^jr cos 3 9
so clearly
29 f / cos 0-/Oz sin 0, a cos 0) d9 = -y ~

2^
t
30 ~ 6 -f(a 6 a cos 9) dd =0
^J sin sin ,
Thus the approximate equations governing slowly varying a and 0 are

given by (23) and (24) as
31
32 0 = 0
If we are interested in finding the steady-state periodic solutions of Van
der Pol’s equation, we set a = 0, 0 = 0, which gives a = 2. Hence, to
first order in //, the limit cycle of the Van der Pol oscillator is described
by
33 y(t) = 2 sin ( +0 O)
To get the slowly varying solution, we solve (31) and (32), which
results in
1/2
34 a(t) =2 1 ~]
_1 + cexp(— fit)]
35 0(0 = 00
where c is a constant determined by the initial conditions. Thus we see
that even though the parameter ji does not affect the steady-state solu-
tion, it does affect the rate at which the transient solution approaches
the steady-state solution.
2.5.2 Power Series Method
The power series method is applicable to autonomous second-

order differential equations containing a “small” parameter // and
consists of attempting to expand the solution of the given equation as a

power series in //. We procedure by means of an
shall illustrate the
example.
Consider the differential equation
36 y{t) + coly(t) + n\y(t)] = 0 3
together with the simplified initial conditions
37 j(0) =a 0
38 j(0) = 0
This equation represents the oscillations of a mass constrained by a
nonlinear spring. If jn > 0, the spring is said to be hard, whereas if
ju < 0, it is said to be soft.
Clearly, if ju = 0, the solution of (36) satisfying (37)—(38) is
39 y 0 (t) =a 0 cos co 0 t
If jj, ^ 0 but is “small,” we can attempt to express the solution of (36)-

(38) as a power series in //, in the form
40 y(t) =y 0 (t) + ny x (t) + ju
2
y 2 (t) + . . .
The idea is and equate the coefficients of all

to substitute (40) into (36)
powers of [i to zero. However,
done blindly, some of the y {t) if this is
t
may contain secular terms, i.e., unbounded functions of time. To see

this phenomenon, let us substitute (40) into (36) and set the coefficients
of all powers of pi to zero. This gives
41 Jo(0 + tOoj> (0 = 0 To(0)= ô(O) = 0

42 Ti(0 + coly\(t) + yl(t) = 0; ^(0) - 0, ^(0) = 0
Solving first of all fory 0 (-), we get
43 Jo(0 = cos co 0 t
Thisis to be expected, because ;; 0 (*) is the solution of (36)-(38) corre-
sponding to pi =
0. Now the equation for yj(.) becomes
44 y x
(t) + co 2
Q y x
{t) = ~yl{t) = -a\ cos 3 co 0 t
= —\al cos co 0 t — cos 3 g> 0 *
The solution of this equation is
3 /7 3 n3 n3
45 y {t) = * sin “ 32cof cos + 32^ cos 3
'
-st
The t sin a> 0 t term on the right-hand side of (45) is the secular term,
which arises because the forcing function of the equation for
y j
( )
contains a component of angular frequency co 0 , while the system itself
has a resonant frequency at o) 0 . Combining the above expression for
To(*) and /!(•), the approximate solution of (36)-(38), which is good to
the first order in n, is obtained as
46 y(t)~y 0 (t) + AFi(0

= a° ~ Si) cos ®°* ~ iS 1 sin °ht + Sf os 3c° ot
(
It is clear that this y{t) is an unbounded function of t and hence is an

unacceptable approximation.
The presence of the secular terms can be rationalized as follows:
If // = 0, the solution of (36) is periodic with angular frequency co 0 .
However, ^ 0,
the angular frequency of the periodic solution of
if pi
(36) is not necessarily co 0 but is different in general. On the other hand,

because ^(O [the so-called “generating solution” of the sequence of
.
functions j0( #
)> T i( #
)> • •
•] has an angular frequency co 0 , it is clear that
all functions j; f (*) consist of summations of terms of the type sin nco Q t
and cos nco 0 t where n is an integer. This attempt to express a function

,
whose angular frequency is not co 0 in terms of functions whose angular

frequency is co Q leads to secular terms. As an example, suppose <5 is
“small,” and let us express cos [(co 0 + S)t] as a power series in <5. This
leads to
S 2t 2
47 cos [(co 0 + S)t ] = cos co — Qt St sin co 0 t j- cos co 0 t . .
This power series converges uniformly in t as t varies over any finite
interval, but if the series is terminated after a finite number of terms, the
resulting finite summation contains secular terms. Moreover, the peri-
odicity and boundedness properties of the function cos [(co 0 + 5)t\ are
not at allapparent from the power series expansion in 5.
To alleviate this difficulty, we shall suppose that the solution X*)
of (36)— (38) is periodic with an angular frequency co, which itself is
expressed as a power series in //. In other words,
48 CO
2
= CO + 2
0 //f iOo) + /Ôo) + • • •
which can be rewritten as

49 col = CO — 2
ju£ i(«o) — — ...
In (48), we have explicitly identified the dependence of co on a 0 Sub- .
stituting (49) and (40) into (36) gives
50 To + ATi + • • • + C0 y - 2
0 a£iTo + MC0 yi
2
+
. . .+ juyl + . . . =0
Equating the coefficients of each power of ii to zero gives
51 y0 + co y = 0;
2
0
=
y 0 (0) a 0 y 0 ( 0)
,
=0
+ co y = —To + £i To; Ti(°) = Ti(°) = 0
2
52 y x x
Hence
53 To(0 = a cos o
54 y x
(t) + (o y = — al cos cot + ^ a cos cot
2
x
{t )
3
0
= —\a\ cos cot — \a\ cos 3 cot + £ x a 0 cos cot
For y x {^) not to contain any secular terms, the forcing function on the
right-hand side of (54) must not contain any terms of angular frequency
co. Hence we must have
55 f1 = %d 2
0
With this condition, the solution for is obtained as
56 JiO) = -^2 COS CO?

+ 2^5 cos 3©*
The overall solution of (36)— (38), which is accurate to the first order in
//, is
57 y{t) =a 0 cos cot — cos <ot + cos 3a>t
where
58 co
2 = col + \vtal
59 Example. Consider the simple pendulum equation
60 y + sin y =0
Equation (60) can be approximated by
61 y +y -lL = 0
which is of the form (36), with // =— Using the foregoing analysis,
we see that the frequency of oscillation of the simple pendulum is
related to the amplitude by
2 = -
62 O)
f
1
Problem 2.15. Apply the Krylov-Boguliubov method to Rayleigh’s

equation
Solve the same equation using the perturbation method, and show that both
methods give the same solution to the first order in ji.
Problem 2.16. Apply the perturbation method to the Van der Pol
equation
y +y= vy(i - y 2)
Problem 2.17. Apply the Krylov-Boguliubov method to the pendulum
equation
Show that the expression derived for the frequency of oscillation is the same
as (62).
Problem 2.18. Consider the second-order equation
y + y = nf(y, y)
XO) = 0; j>(0) = a 0
where the function y is continuously differentiable with respect to its two
arguments. Show that, to first order in ji , both the perturbation method and
the Krylov-Boguliubov method give the same results.
Nonlinear differential equations
In this chapter, we shall undertake a systematic study of nonlinear

ordinary differential equations (o.d.e.’s). As can be gathered from the
examples given in Chap. 1, a nonlinear differential equation can in

general exhibit very wild and unusual behavior. However, we shall
show in this chapter that, for a practically significant class of nonlinear
and uniqueness of solutions can be ascertained,
o.d.e.’s, the existence
as well as continuous dependence on initial conditions.

Except for very special cases, which are usually “cooked” in
advance, it is not possible to obtain a closed-form expression for the
solution of a nonlinear o.d.e.Hence it is necessary to devise alternative
methods for analyzing the behavior of the solution of a given nonlinear
o.d.e. without relying on being able to find a closed-form solution. In
this chapter, we shall put forward two techniques for doing this:
1. A technique, known as Picard’s iteration method, is given,
which generates a sequence of functions that converge to the

solution of a given equation. In this way, the solution of a given
equation can be computed to an arbitarily high degree of accu-
racy.
2. A method is given for obtaining bounds on the solution of a
given equation without actually solving the equation. Using
this method, it is possible to determine, at each instant of time,
a region in R within which the solution of the given equation
n
must lie. Such a method is useful for two reasons:

J :
Sec. 3. 1 Mathematical Preliminaries 51
(a) By obtaining bounds on the solution, one can draw con-

clusions on the qualitative behavior of the solution, and
the labor involved is considerably less than that needed to
actually solve the equation.
(b) The bounds obtained by this method serve as a check on
solutions obtained by alternative methods, e.g., a numerical
solution on a computer.
The study of nonlinear o.d.e.’s in general terms requires rather
sophisticated mathematical tools. For this reason, we shall develop
the necessary mathematical background in Secs. 3.1 and 3.2.
3 .1
MATHEMATICAL PRELIMINARIES
3.1 .1 Linear Vector Spaces
This subsection is devoted to an axiomatic development of linear

vector spaces, both real and complex. In most practical applications,
it is enough to deal with real vector spaces. However, it is sometimes
necessary to deal with complex vector spaces in order to make the theory
complete. For example, a polynomial of degree n has n zeroes only if
one counts complex zeroes.

We note also that it is possible to define a linear vector space
over an arbitrary field (e.g., the binary field, the field of rational func-
tions, etc.). However, we do not need these concepts in this book.
1 [definition A real linear vector space (resp. complex linear vector

space) is a set F, together with two operations + V X V — V, : >
termed addition, and R X V —* V (resp.

• : > Cx V — V), • :
termed scalar multiplication, such that the following axioms hold
2 (VI) x + y = y + x, \/ x,y e V (commutativity of

addition)
3 (V2) x + (y + z) = (x + y) + z 9 V x, y e V
(associativity of addition)
(V3) There is an element in V, denoted by 0 V (or 0 if V is

clear from the context), such that
4 X +§v =0V X + =
X, X v G V
(existence of additive identity)
(V4) For each x e V, there exists an element, denoted by

—x, such that
,
52 Chap. 3 Nonlinear Differential Equations
5 x + (— x) — 0 v (existence of additive inverse)
(V 5) For each rt ,
r 2 in R (resp. c x ,
c 2 in C) and each x e F,
6 *v(r2 -x) m {r r )-x x 2 [resp. c r (c 2 -x) = {c x c 2 )-x]
(V6) For each r in R (resp. c in C) and each x 1? x 2 e F,
1 r-{x x + x = r-x + r-x

2) t 2
[resp. c«(x + x = c-Xj + ox
i 2) 2]
(V7) For each r u r 2 in R (resp. c u c 2 in C) and each x e V 9
8 (''i + r 2 )'X =r x *x + r 2 -x
(V8) For each x g F,
9 1 • x =x
This axiomatic definition of a linear vector space is best illustrated
by several examples.
10 Example. The set R n consisting of all ordered ^-tuples of real

,
numbers, becomes a real linear vector space if we define addition and

scalar multiplication as follows: If x = (x 1? . .
. ,
x„) and y = (y u . . .
yn) are vectors in Rn and r is a real number, then
11 x + y = (*i +j„...,x + y J, n)
12 rx = (rx u rx„)
In other words, the sum of two ^-tuples is obtained by component-

wise addition, while the product of a real number and an n- tuple is
obtained by componentwise multiplication.
As a limiting case, it is interesting to note that jR 1 =R ,

the set
of real numbers, is itself a real linear vector space.
The complex linear vector space Cn ,
consisting of all ordered
^-tuples of complex numbers, is defined in a manner analogous to R n
.
13 Remarks: important to realize that whether a linear vector

It is
space is complex is determined not by the nature of the elements

real or
of the space but by whether the associated set of scalars is the set of real
numbers or the set of complex numbers. To bring out this distinction
more clearly, note that Cn can be made into both a real linear vector
space as well as a complex linear vector space.
14 Example. Let F[a ,

b] denote the set of all real-valued functions
defined over an interval [a, b ] in R. Therefore a typical element of F[a, b]
is a function /(•) mapping [a, b] into R. The set F[a ,
b] becomes a real
] ] ] ]
linear vector space if we make the following definitions: Let x(-) and
^(•) be elements of F\a b ]. Then the sum of x(-) and j(-), denoted by
,
(x + j)(*)> is the function whose value at t e [a, b] is given by x(t)

+ y(t). Symbolically,
15 (x + y)(t) = x(t) + y(t )
Similarly, the product of a real number r and a function *(•), denoted by

(rx)(«), is the function whose value at t e [a, b is rx(t). Symbolically,
16 (rx)(f) = rx(t)
In other words, the sum of two functions in F[a b\ ,
is obtained by point-
wise addition, and the product of a scalar and a function is obtained by
pointwise multiplication.
17 Example. The set Fn [a , b], consisting of all functions mapping an

interval [a, b] into Rn ,
is a real linear vector space if we make the follow-
ing definitions: Let x(*) and y(*) map [a, b\ into Rn ,
and let r be a real
number. Then, for all t e [a, b], set
18 (x + y)(0 = x(f) + y(0

19 (rx)(t) = rx(t)
20 Example. Let S denote the set of all complex- valued sequences

(*„)- i • Then S can be made into a real or complex linear vector space,
by appropriate choice of the associated set of scalars, if we define addi-
tion and scalar multiplication as follows: Let (x„)~ and (jjf be elements
of S and let a be a scalar (real or complex). Then the sum of (x„)r
,
and is defined as the sequence (xn + y n )~, and the product of a

0
and (xn )~ is defined as the sequence (ax„)
<
-
i
Because we can think of a sequence as a function mapping the set

of natural numbers into the set of real numbers, this example is basically
the same as Example (14).
21 [ definition A subset M of a linear vector space V is called a

subspace of V if
22 1. x+y e M
whenever x,y e M
23 2. ax g M
whenever x e and a M is a suitable scalar.
Loosely speaking, M is a subspace of V if it is a linear
vector space in its own right.
24 Example. Let F [a, b] be as in Example (14), let t 0 e [a, b], and

let Ft0 [a, b denote the subset of F[a, b consisting of all *(•) in F[a b ] ,
such that x(t 0 ) = 0. In other words, F t0 [a , b] consists of all those func-

tions in F[a b] that vanish at
,
t0 . Then F t0 [a, b] is a subspace of F[a ,
b\.
3.1.2 Normed Linear Spaces
The concept of a linear vector space is a very useful one, because

in that setting it is possible to define many of the standard engineering
concepts, such as a linear operator, linear independence, etc. It is also
possible to study the existence and uniqueness of solutions to linear
equations. However, the limitation is that there is no notion of distance
or proximity on a linear vector space. Hence it is not possible to discuss
concepts such as convergence and continuity. This is the motivation for
studying a normed linear space, which is basically a linear vector space
with a measure of the “length” of a vector.
25 [ definition] A normed linear space is an ordered pair (X, ||

•
||),
where X is a linear vector space and ||

•
||
is a real- valued function
on X (called the norm function), such that
26 (Nl) ||
jc ||
> 0, V * e X; ||*|| = 0 if and only if
x = Ox
27 (N2) ||a*|| = a V* |
|
• ||
jc||, e X, for all scalars a
28 (N3) ||* + y|| < ||*|| + ||y ||, Vx,yeX
(triangle inequality)
The norm on a normed linear space is a natural generaliza-

tion of the concept of the length of a vector in R2 or R 3
. Thus,
given a vector x in a normed linear space (
X , ||
•
||),
the nonnega-
tive number ||x|| can be thought of as the length of the vector *.
Similarly, giventwo vectors * and y in (X, ||),
||x y\\ can ||
• —
be thought of as the distance between x and y. With the aid of
this concept of distance or proximity, it is possible to define con-
vergence and continuity in a normed linear space setting.
29 [ definition] A
sequence (x„)7 in a normed linear space (X, || • ||)
is said to converge to an element * 0 e if ||*„ *0 0 as X — 1|
— >
n—> oo. Equivalently, (x„)7 converges to *0 if, for every e > 0,
there exists an integer N(e ) such that
30 1
1*„ —* 0 1|
<e whenever n > N(e).
This basic definition of convergence can be interpreted in many
ways. Clearly the sequence of vectors (*„)7 converges to * 0 if and only
if the sequence of nonnegative numbers (||*„ * 0 1|)7 converges to zero. —
Alternatively, let B(x 0 ,
e) denote the ball in X defined by
31 B(x o, e) = {* e X: ||* — * 0 II
< e}
Then the sequence (x„)T converges to *0 if, for any e, the ball B(x 0 ,
e)
contains all except a finite number of vectors in the sequence (*„)7-

[definition] Let {X, • x) and (T, • r) be two normed linear

|| \ \ || ||
spaces, and let / be a function mapping into Y. We say that X

/is continuous atx 0 g I if, for every e > 0, there exists a S(e, x 0 )
> 0 such that
1 1
f(x o) ~ f(y)\\r<£ whenever 1 1
x0 —y \ \ x < d(s, x 0)
/is continuous if it is continuous at x e X.

every Finally, /is uni-
formly continuous if (1) it is continuous and (2) for every e > 0

there exists a 8(e) > 0 such that
1 1
f(x) - f(y) Hr < £ whenever 1 1
x -y \ \ x < 5(e)
We shall postpone examples of these concepts until we have given
a few examples of normed linear spaces.
Remarks:
1. The concept of a continuous function from one normed linear
space to another is a natural extension of the concept of a
continuous real-valued function of a real variable; in a general
normed linear space setting the norm plays the same role as
the absolute value does in the set of real numbers. Similar
statements apply to uniform continuity.
2. The important difference between continuity and uniform
continuity is that in that latter case 8 depends only on s and
not on x or y.
A sequence (x„)~in a normed linear space (X, ||

•
||)
converges to
x 0 if \\xn — x Q ||
approaches zero as n — However, in many cases,
» oo.
we generate a sequence (xn)[ without knowing beforehand what its

limit is, or indeed whether it has a limit. (For example, this is the case
whenever we we
try to iteratively solve a nonlinear equation.) Thus
need a characterization of sequences that does not involve the (possibly
unknown) limit of the sequence. This is provided by the concept of a
Cauchy sequence.
[definition] A sequence (x„)f in a normed linear space ( X , ||

•
||)
is said to be a Cauchy sequence if, for every e > 0, there exists an

integer N(s) such that
\\xn - xJI < e whenever n,m> N(s)

Thus, a sequence (x„)T is convergent if its terms xn approach
arbitrarily closely a fixed vector x 0 while
,
it is Cauchy if its terms
approach each other arbitrarily closely as n > oo. The relationship —
between convergent sequences and Cauchy sequences is demonstrated
next.
38 fact Every convergent sequence in a normed linear space is also a Cauchy

sequence.
proof Suppose (x„)~ is a convergent sequence in a normed linear space

( X ,||), and denote its limit by x 0
||
• To prove that (;*:„)“ is also a Cauchy .
sequence, suppose fi > 0 is given; then pick an integer such that N

39 \\x„ —x 0 \\
< -^ whenever n>N
It is always possible to find such an integer N because converges to
Xq. Now, whenever n, m>N, we have, by the triangle inequality, that
40 1 1
Xn - Xm \\ < \\X„ - Xq + 1 1 1 1
Xm - Xq 1 1
< y + ~~ = S
where we have also used (39). Thus (x„)f is a Cauchy sequence. —

The converse of fact (38) is not true in general —a sequence can
be Cauchy without being convergent [see Example (82)]. However, some
normed linear spaces have the special property that every Cauchy
sequence in them is also convergent. This is clarified in the next defini-
tion.
41 [definition] A normed linear space (X, ||

•
||)
is said to be a
complete normed linear space, or a Banach space ,
if every Cauchy
sequence in converges (to an element of X). X
Banach spaces are important for two reasons: (1) If (X, ||
•
||)
is
a Banach space, then (as stated in the above definition) every Cauchy
sequence is convergent. (2) Even if ( X ||)
is not a Banach space, it , ||
•
can be made into a Banach space by adding some extra elements to X

Thus, in almost all applications, we can assume that we are dealing only
with Banach spaces.
We shall now give some examples of normed linear spaces.
42 Example. Consider the linear vector space Rn ,

together with the
function ||
•
H*,: R n — R defined by
»
43 1 1
x |
— max \x t \
1 <i<n
(The reason for the subscript oo will become clear later). The function
||
•
|
axioms (Nl) through (N3), as can be easily verified. In
|oo satisfies
fact, (Nl) and (N2) can be verified by inspection. To verify (N3),
suppose x fes (x l5 x„) and y Or, y„) are vectors in R We

.
n
.
. ,
= . .
. ,
.
know that, for each i, \x y \x y Therefore t + < + t \ t \ | t \.
44 ||x + y||„ = max \

x t+ y, < max(|x,.| + \ |y,.l)
i i
< max |
x , | + max | y t \
= ||x|U + ||y||„
,
Sec. 3.1 Mathematical Preliminaries 57
so that (N3) is satisfied. Thus the ordered pair (R n , ||

• WJ) is a normed
linear space. In fact, it can be shown that (R n , ]|
•
IU) is actually a
Banach space.
45 Example. Consider once again the linear vector space Rn ,

but
this time with the function ||
•
\\ t
: R — n
R defined by
46 ll*lli = i=i
SI*/l
Clearly ||
•
||j
also satisfies axioms (Nl) and (N2). To verify (N3),
suppose x, y e Rn . Then
47 ll
x + ylli = i=2 1* + *1 < i=l
2 (1*1 + 1*1)
1
= i=l
21*1 + 21*1
i=l
= 11*11. + llylli
Hence the ordered pair ( R n Hj) , ||
• is also a normed linear space and can
in fact be shown to be a Banach space.
48 Remarks: It is important to note that the normed linear space
R" 1*,) is a different entity from the normed linear space ( R

*
n
( || 1
) , ||
•
\\ t 9
even though the underlying linear vector space is the same in both cases,
(namely R n ), because even though the linear vector space on which the
two norm functions are defined is the same in both cases, the norm
functions themselves are different.
49 Example. Consider once again the linear vector space Rn ,
together with the function ||

•
\\ p : R n — R defined by
*
50 = ,jc
J’
ii*ii'
{§ 'iT
where p ranges between .
p becomes the norm
1 and oo. If /? = 1, || ||
function of Example (45), while if p oo 5 • becomes the norm func- = 1 1 | \

p
tion of Example (42) (hence the subscripts 1 and oo on the norms in
these examples). The function clearly satisfies axioms (Nl) and •

p || \\
(N2) and can be shown to satisfy (N3) whenever 1 oo. Thus <p <
the ordered pair (R n •
|[p)
is a normed linear space
, whenever
|| 1 p <
< oo. (Of course, for different values of
p, the corresponding normed
linear spaces are different.)
In particular, if p = 2, we have
[
rt
) 1/2
2
51 IMI 2 = {Sl*fl
}
which is generally known as the Euclidean norm or /2 - norm on R n

.
The Euclidean norm is a particular example of an inner product norm,

which is defined in Sec. 3.1.3.
Note that all the spaces (Rn , ||
• \\
p ) are Banach spaces.
>
R and C”. Normed linear spaces whose

Special Properties of
n
underlying vector space is R or C have some special properties, which n n
we shall now discuss. For ease of statement, we shall give the facts
only for Rn ,
but they are equally true for C”.
fact Let ||
•
|| a and be any two norms on
||
•
\\ fi
Rn . Then there exist finite
positive constants k t and k 2 such that
&i||x|| a < 1
1
X 1
1^
<k 2 ||x IU, V X e Rn
We shall not prove this fact. Instead, by way of illustration,

consider the two norms ||
•
||j
and ||
•
IU on Rn . Then it is easy to verify
that
l|x|U < HxIIj <«||x||„, V xeff

Similarly, we have
||x||oo < II x|| 2 <« 1/2
||x|U \/ X e R n
Fact (52) merely states that a similar relationship (known as topological

equivalence) exists between any two norms on Rn .
Now
us consider some consequences of fact (52). Let
let • a || ||
and - IU be two given norms on Rn let (x„)“ be a sequence in Rn and

|| ; ,
suppose x„ || —
x 0 a * 0 as n—> oo. Thus (xjr converges to x 0 in ||
—
(. Rn , ||
-
|| a );
i.e., (x B )“ converges to x0 in the sense of the norm ||
•
|| a.
Now, because (53) holds, we see that ||x„ —x 0 \\ fi < A: 2 || —x 0 ||«,
so that x„ —x 0 |U
— > 0 as n — oo. Hence (x„)7 also converges to x0
in the sense of the
||
norm ||
-
|U-
Conversely, suppose 1 1
x„ —x 0 1 U
— > 0
as n—+ oo. Then, because ||
x„ —x 0 JU < (l/î)||x — x M 0 |U ||x„ —x 0 || a
also approaches zero as n — > oo. Thus we see that (x„)~ converges to
x0 in the sense of ||
•
|| a if and only if it converges to x 0 in the sense of
IU This means that convergence in R is independent of the norm
n
•
||
chosen. Now, let us for the moment use the specific norm • IU to test ||
convergence on R n Then one can readily see that ||x„ XqU 0 .

— —
as n —
» oo if and only if, for each i between 1 and n , the component
sequence of real numbers (x^'OT converges to the real number x 0

l) (
.
Putting all these observations together, we can state the following.
fact Let ||-|| be any norm on Rn ,

let (xj'f be a sequence in Rn and ,
let x0 e Rn . Then ||x„ — x0 ]|

— > 0 as n — > oo if and only if each com-
ponent sequence converges to xjp, for / = 1, . . . ,

n.
Let us consider another implication of fact (52). Let x(.) be a

function mapping R into R In accordance with definition (32), we
n
.
would say that the function x(«) is continuous at t e R in the sense of
||
.
|| a if, whenever (Or is a sequence in a b] such that [< ,
tn —> t as n — oo,
1
we have that Thus if we view

||
x(/„) as — x(t) a — ||
> 0 as n — > oo. 40
a mapping from (. R
would appear that whether , |
•
|)
into ( Rn , ||
•
|| a ), it
or not x(*) is continuous at t e R depends on the choice of the norm

a on R
n
||
•
|[
However, in view of the above discussion, it follows that
.
the continuity of x(*) in one norm is precisely the same as the con-
tinuity of x(*) in another norm. In particular, using fact (52), we can
state the following.
57 fact be any norm on R n and let x(-) be a function mapping R

Let ||
•
|| ,
into R Then x(-) is a continuous function from ( R

n
.
|) into (R
n
||) if , |
•
, ||
•
and only if each of the component functions **(•) is a continuous function

on R.
Note that facts (52), (56), and (57) hold also for C”. But they
do not hold for an arbitrary normed linear space.
Further Examples of Normed Linear Spaces
58 Example. Let [a, b] be a bounded interval in R ,

and let C[a ,
b]
denote the set of all continuous real- valued functions over [a, b ]. Define
a function ||
•
|| c over C[a, b] as follows: If 40 £ C[a, b], then
59 II 40 lie
= max |
x(0
t(=[a,b ]
It is easy to verify that ||

•
||c satisfies all the axioms of a norm —axioms
(Nl) and (N2) can be verified by inspection. To verify (N3), let 40>
y(-) e C[a ,
b\. Then
60 ||x(.) + X-)||c = max t

|
x(t) + y(t)\ < max (|x(0l + MOD t
< max t
|
x(0 1
+ max y(t) = ||x(-)llc +
t
| \ II lie
where all maxima are taken over [a, b ]. Thus the ordered pair
( C[a b\, c) is a normed linear space. Notice that a sequence of
, ||
•
||
functions [x„(-)]r in C[a, b] converges to x 0 (-) in C[a b] in the sense of ,
|| c if and only if [x„(*)]r converges uniformly over [a, b\ to x 0 (0-

•
||
The norm c is usually known as the sup (for supremum) norm.

||
•
||
The space ( C[a b], c) is in fact a Banach space. , ||

•
||
61 Example. Let be a given norm on R n and let C n [a b] denote ||

•
|| , ,
the set of all continuous functions mapping [a, b] into R

n
where [a, b] ,
is a finite interval in R. Define the function • c C [a b]

n
> R as
1
|| ||
:
,
—
^he set Cn [a ,
b] is not to be confused with C (n) [a, 6], which is the set of all
continuous real-valued functions over [a, b] that are /z-times continuously differen-
tiable. We shall not have much occasion to use the latter symbol.
:
follows: Let x(-) e Cn [a, b]. Then
62 ||x(-)llc = max||x(OII
re [a, 6]
Once again, ||
•
|| c constitutes a norm on C"[a, b\, as can be readily veri-
fied.Axioms (Nl) and (N2) are easy. To test (N3), let x(.), y(.) e
C-[a, b]. Then
63 ||
x(.) + y(-)|| c = remax||x(0 + y(/)|| [a, 6]
< max re [a, b]

(|| x(t) || + || y(0 1|) (by the triangle
inequality on R)
n
< max re [a, b]

||
x(r) || + max ||y(f)||
re [a, 6]
= l|x(.)llc + lly(-)llc
The ordered pair ( C n
[a , b], ||
•
|| c) is in fact a Banach space.
In this example, it is essential to note the difference between
||
•
||
and ||
•
|| c
— ||
•
||
is a norm on Rn c is a norm on
,
while ||
\a , b\.
•
||
Cn
The former has an n - vector as its argument, while the latter has a vector-
valued function as its argument. When we study nonlinear differential
equations in Sec. 3.3 this distinction becomes crucial.
3.1.3 Inner Product Spaces
An a special type of normed linear space

inner product space is
in which, because of its special structure, it is possible to define geo-

metrically appealing concepts such as orthogonality, Fourier series, etc.
Axiomatically, an inner product space is defined as follows:
64 [definition] An inner product space is a linear vector space X with

associated field F= R
or C, together with a function ^ • , • )>
Xx X— > F, such that the following axioms are satisfied:
65 (II) <x, = <J, x> if F = R, <x, = <j, *>

y>
if F— C, V x,y G X
66 (12) <x, y + z> = O, y} + <*, *>, V x,y,z e X
67 (13) <x, a y> = a<x, y>, \/ x,y e X, Va g F
68 (14) <x, x> > 0, V x g X, <x, x> = 0
if and only if x = 0*
These four axioms present an abstraction of the familiar

notion of the scalar product or dot product in R or R .
2 3
The next result shows that an inner product space can be made
into a normed linear space in a natural way.
69 Theorem Given an inner product space X with inner product

<•, •>, define ||.||: X-+R by
70 ||x|| = <x, x> 1/2
, v X e X
Then ||
•
||
is a norm on X, so that the ordered pair (X, ||
•
||)
is a
normed linear space.
proof Clearly ||
•
||
satisfies (Nl) and (N2). To prove (N3), we need the
following extremely useful inequality, known as Schwarz’s inequality.
71 lemma Let x and y be arbitrary elements of an inner product space X

with inner product <•, •>. Then
72 |<*,y>|<|MHbll
73 |<x, y}\ = ||x||-||y|| if and only if ax + fly = 0* for some scalars a,
/?, not both zero.
proof of lemma (71) We shall only prove the case where is a real inner X
product space. The complex case is handled in much the same fashion.
Consider the function
74 /(a, /?) = ||ax + Py\\ 2 = <ax + py, ax + Py}

= a 2 ||x|| 2 + 2ay? <x,y> + P 2 \\y\\ 2
By (14), /(a, P)> 0 for all a, /? in R. Because / is a quadratic form in a
and P, it follows that /(a, /?) >0 V a, P e R if and only if the discrimi-
nant of the quadratic form is nonpositive, i.e.,
2
75 <x,y> <||x|| 2 .||y|| 2
Furthermore, suppose ax ^
+ Py 0* whenever either a or /? is nonzero.
Then/(a, p) > 0 whenever either a or /? is nonzero. This is the case if and
only if the discriminant of (74) is strictly negative, i.e.,
76 <x,y> 2
< ||x||
2
-||y||
2
Hence the lemma is proved. ^

PROOF, Continuation of, for Theorem (69) For all x, y in X, we have
77 ||x +y ||
2 = + y, x + y> = ||x|| + \\y + 2<x, y>
<x 2
||
2
< H^ll 2 + ||y|| 2 + 2||x||*||> (by Schwarz’s inequality) ;

||
= (MI +||y||) 2
Hence (N3) is satisfied, and ||
•
||
is a norm on X. —
] , , ,1
Theorem (69) shows that every inner product space can be made
into a normed linear space in a natural way. It is therefore possible to
speak of the completeness of an inner product space.
78 [ definition An inner product space that is complete (in the sense

of the norm induced by the inner product) is called a Hilbert space.
79 Example. Consider the linear vector space R n

,
together with the
function <(•, •>„: R X R n n — R > defined as follows: Let x = (x l9 . . .
x„) and y = (y u . .
. y n) belong to R\ Then
80 <X, y>„ =E
=
Xy
1 1
t t
It is routine to verify that the function <•, •>„ satisfies axioms (II)
through (14), so that the ordered pair(i£",<(-, -)„)is a real inner product
space. It is in fact a Hilbert space. The norm on Rn corresponding to
this inner product is
81 ||
x || = {£*?}
which is the Euclidean norm discussed earlier.
82 Example. Let C n [a, b] be the linear space of Example (61), and

define <•, *> c : C”[a, b] x C n [a, b\—>R as follows: If x(-) and y(*)
belong to C n [a, b\, then
83 <x(*)» y(*)>c = [
<x(0, y(t)}n dt
In this case, the ordered pair ( C n [a &],<-, -) c) is a rea l inner product

space. It is not, however, a Hilbert space. [Contrast with ( C n [a ,
b],
c), which is a Banach space.] The completion of ( C [a b] <•, -) c)

n
• ,
|| ||
is a space denoted by L\\a, b], which is the space of all square-integrable
Lebesgue-measurable functions. The inner product on L 2 [a, b ] is also

n
given by (83), except that the integral now has to be interpreted as a

Lebesgue integral (see also Chap. 6).
We shall close this section by giving two examples of continuous

functions.
84 fact Let (X, 1 1

•
1 1) be a normed linear space. Then the norm function 1 1
•
1
is uniformly continuous on X.
proof Use definition (32) of uniform continuity, and let <5(e) = 8, for
each 6. Because, for each x, y e X, we have
85 <11* -y\\
,
it follows that
86 11*11 - Ibll < 8 whenever \\x — y\\ <8

Thus 1 1
•
1
1 : X— R > is uniformly continuous on X.
87 corollary Suppose that X ( , ||

•
||) is a normed linear space and that
(x„)^° is a sequence in X converging to x 0 . Then the sequence of real numbers
(IK I l)f converges to [|x 0 ||.
88 fact Suppose (X, <• , •» is an inner product space. Then, for each y e X,
the function x t-> is uniformly continuous.
proof Use definition (32) of uniform continuity, and let S(8) = 8/\\ y\\,
8. Now, suppose ||x — z\\ < 5(8). By Schwarz’s inequality, we
for each
have
89 |<x, y> - <z, j»| = |<* - z,y>\ < \\x - z\\-\\y\\ < <5(fi)||y|| =8
Hence the function jc i-» y) is uniformly continuous. M
Problem 3.1.Show that the zero element of a linear vector space is

unique. [Hint: Assume that a linear vector space V has two zero elements,
Oj and 0 2 and use axiom ,
(V3).]
Problem 3.2. Show that, in a linear vector space, the additive inverse of
an element is unique.
Problem 3.3. Let S be the sequence space of Example (20), and let S r
be the subset of S consisting of sequences converging to r. For what values of
r is S r a subspace of *S?
Problem 3.4. Sketch the unit spheres, i.e., the sets
{x G R 1
1
x | |p
= 1}
for (a) p = 1, (b) p = 1.2, (c) p = 2, (d) p = 3, (e) p = co.
Problem 3.5. (a) Using the triangle inequality, show that
2 x, £ SI
=
i l
where m is any finite number, x f e Rn for i = 1, . .

.
m ,
and ||
•
||
is any
norm on R n
.
(b) Let Cn [a, b ] be as in Example (61). Using the Riemannian approxi-

mation for the integral, show that
4 4
Ja
f x(t) dt II
II
< Ja f 1 1
x(r) || dt
Problem 3.6. Prove Schwarz’s inequality for complex inner product

spaces. Show that if (*„)“ and (y„)~ are sequences in X converging to x 0 and
,
y0 ,
respectively, then the sequence of real numbers «a:„, converges to
<* 0 , J'o) (Hint: Write <x„, y„}
- <x, y) as
<x„, y„y - o, y„y + <x, y„y - <x, yy

and use Schwarz’s inequality.)
3.2
INDUCED NORMS AND MATRIX MEASURES
In this section,we shall introduce the concepts of the induced norm
of a matrix and the measure of a matrix. These concepts are vitally
needed in Sec. 3.5, where we shall study methods for estimating the
solutions to nonlinear differential equations without actually solving
them.
3.2.1 Induced Norms
LetC nxn (Rnxn ) denote all n X n matrices with complex the set of
nXn
(real) elements. Then C
can be made into a linear vector space if
addition and scalar multiplication are done componentwise. Moreover,
each element A e C nxn defines a corresponding linear mapping a from
Cn into Cny according to the rule
1 a(x) = Ax, V x e Cn
Conversely, every linear mapping P from C into itself can be associated
n
with an n X n matrix B, i.e., an element B of C nXn This can be shown as .
follows Let : e,- be the vector in C" which has all zero components, except
for the yth component, which is equal to 1. Now, p(ey ) is an element of
C because p maps C
n
,
n
into itself. Accordingly, let (fe l7 , i = 1
be the components of the vector p(e y), so that

2 P(e,) = [b lj9 b 2j , . .
. ,
bj A b'
Define b y as above, to be the element of
,
C n whose zth component is b tj .
Finally, if we form the n X n matrix
3 B= [b 1 |b 2 |...|bJ
then one can easily verify that
4 P(x) = Bx, V x g Cn
Therefore, there is a one-to-one correspondence between elements of
C nxn and linear mappings from C* into C”. 2 We do not in general dis-
2 The matrix B is known as a matrix representation of the mapping P with

respect to the basis (ei, sometimes known as the natural basis. It is
. .
.
e„), which is
possible for the same linear mapping to have different matrix representations with
respect to different bases. However, in this book we shall not explore such subtleties
of linear algebra.
Sec. 3.2 Induced Norms and Matrix Measures 65
tinguish between an element A in C nXn and the corresponding linear

mapping A from C n
However, this one-to-one correspondence
into C”.
is important for the notions of induced matrix norm and matrix
measure.
5 [definition] Let ||
•
||
be a given norm on C”. Then for each
matrix A e C nXn ,
the quantity ||
A || f
defined by
6 l|A||, = sup
X
sup || Ax|| = sup || Ax||
xÔ II |
| x|= 1 H x B<1
xGC”
is called the induced {matrix) norm of A corresponding to the vector

norm ||
•
||.
It should be noted that there are two distinct functions involved

in definition (5):One is the norm function mapping C" into R, ||
•
||
and the other is the induced norm function mapping C nxn into R. ||
•
||,
The induced norm of a matrix A (or, what is the same, the induced
norm of a linear mapping A) can be given a simple geometric inter-
pretation. Equation (6) shows that A is the least upper bound of the || || £
ratio Ax ||/|| x as x varies over C". In this sense, A

|| ||
can be thought || || f
of as the maximum “gain” of the mapping A. Alternatively, let B be

the unit ball in C"; i.e., let
7 B = {xe C n
: ||x|| < 1}
Now, suppose we distort B by replacing each x in B by Ax, i.e., by its

image under the mapping A. Then what results is the image of the set B
under the mapping A. In this setting, the induced norm of A, ||A|| f,
can be thought of as the radius of the smallest ball in C n that completely

covers the image of B under A.
Lemma (8) shows that the induced norm function actually
constitutes a valid norm on C nXn .
8 lemma For each norm on C n the corresponding induced norm ||

•
|| ,
function ||
maps C nXn into [0, oo), satisfies axioms (Nl) through (N3),
•
|| f
and is therefore a norm on C nxn .
proof Clearly, A 0 || || t
- > V A e C nXn and axioms (Nl) and (N2) can ,
be verified by inspection. To suppose A, B e C nXn Then

verify (N3), .
9 || A+B ||/
= sup 1 1
(A + B)x|| = sup || Ax + Bx||
11x11 =1 II
xH=1
< sup=
Ix!
[|
l
|
Ax + Bx || || ||] (by the triangle inequality on C”)
< lxsup 1
1 Ax 1 1
+ sup 1
1 Bx ||
= II All,- + ||B||,
H=1 llx||=l
Hence (N3) is also satisfied, and thus ||

• -
|| {
is a norm on C nxn .
bb
; | 1 1 >
In view of lemma (8), it is clear that, corresponding to each

norm on C”, there is an induced norm on C nXn However, the converse
.
is not true. Consider the function ||

•
\\ s
: C nxn — [0, oo) defined by
10 II All, = max K, I
i,j
x
Then one can easily verify that defines a norm on C" ". However, ||
- 1|,
no norm on C exists such that

n
||
is the corresponding induced
•
|| ||
• ||,
norm. Hereafter, when we say that is an induced norm on C

nxn
•
|| || f ,
we mean that there exists a norm on C such that

n
induces ||
•
|| ||
•
||
||
(Note that some authors use the term bound norm instead of
•
II;.
induced norm.)
In fact [3.1(52)], it is stated that any two norms on C n are topo-
logically equivalent. By the same token, any two norms on C nXn are
also equivalent. In particular, an induced norm is equivalent to another
“noninduced” norm. But a special property of induced norms is bought
out below.
11 lemma Let ||
•
|| f
- be an induced norm on C nxn corresponding to the
norm ||
*
||
on Cn . Then
12 1
1 AB 1 1/ < A 1
1
| h 1
1
B | | f, V A, Be C«*»
proof We have
13 1
1 AB |
|j* = sup || ABx ||
II
X =1 11
However, by definition,
14 ||
Ay || < A 1
1
|
h-
1 1 y 1 1, Vy g Cn
So, in particular,
15 ||
ABx || < 1
1
A | f 1
1 Bx 1
Similarly,
16 1 1
Bx 1 1 < 1
1
B 1
1,-
1 1
x 1
As a result, we get
17 ||
ABx 1 1 < A || ||;
||
B ||; |
|x ||
from which (11) follows immediately. —B
Thus induced norms have the special feature that they are
submultiplicative i.e., the induced norm of the product of two matrices
A and B is less than or equal to the product of the induced norms of
A and B. It can be readily veiified by example that the norm 1|, ||
-
defined in (10) does not have this property (and hence cannot be an
induced norm).
In general, if we are given a specific norm on C n (say, for instance,

the norm ||
•
\\ p defined in Example [3.1(49)]), it is not always easy to
find an explicit expression for the corresponding induced norm on
C nxn — the equations (6) serve more as definitions and do not always
result in explicit expressions. However, the induced matrix norms cor-
responding to the vector norms ||
•
W^, ||
•
|| l5 and ||
•
|| 2, respectively (as
defined in Examples [3.1(42)], [3.1(45)], and [3.1(49)], respectively),
are known and are displayed in Table 3.1.
TABLE 3.1
Norm on Cn Induced Norm on C nXn

|[x|Lo = max |x,-| 1
1
= max 2 a
A | b-oo (row sum) 1 ij 1
* j
llxlh =21*1 1A n1 = max

| 2 au
|
(column sum) 1 1
l|x|| 2 =(Slx,P) 1/2

All
II
= [A ma x‘(A*A)] 1/2 where
,
-2 ,
i
2max (A* A) = maximum eigenvalue of A*A
3.2.2 Matrix Measures
Let ||
• ||f
be an induced matrix norm on CnXn Then the correspond-
.
ing matrix measure 3

is the function pp C
nXn — R defined by
18 /!,.(
A) = lim 11
1 + eA ~ ll-
1
e->0 + S
From a purely mathematical point of view, the measure p t

(A) of
a matrix A can be thought of as the directional derivative of the norm
function ||
•
|| f ,
as evaluated at I in the direction A. However, the measure
function has several useful properties, as stated below.
19 Theorem Whenever ||
• || f
is an induced norm on C nXn ,
the cor-
responding measure //,(•) has the following properties:
(i) For each A e C" x ", the limit indicated in (18) exists and is
well defined.
(ii) -|| A||, < -/*,(- A) < ^.(A) < || A|L, V A € C"*"
(iii) /«,(aA)= a//, (A), V a > 0, V A e C n/n
(iv) max [/i,(A) — #,(— B), —/!,(— A) + /i,(B)] < /i,(A + B)
< Vi(A) + a,(B)
3 Some authors also use the term logarithmic derivative.
;
(v) /*<(•) is a convex function on C” x "; i.e.,
jU,[aA + (1 — a)B] < a/i,(A) + (1 — a)/i,(B),

V a e [0, 1], V A, B e C BX ”
(vi) — //,(— A) < Re X < /i,(A) whenever X is an eigenvalue of A.
proof
(i) The function ||
•
|| f
is a convex function on C nxn ;
i.e., it satisfies the
property
20 ||aA + (1 - a)B|U< a|| A||, + (1 - a)||B|| f>
V a G [0, 1], V A, B g C nxn

Because ||
•
|| t
- is a convex function, it can be shown ([11] Eggles-
ton) that it has a directional derivative at every point in C nXn in
every direction. However, we shall also give a constructive proof
that is applicable to the present case.
Let us define
= ll
I + fi A|li - 1
21 f(e A)
We first show that the function A) is nonincreasing as

s — 0+ ;
i.e.,
22 /(Si ;
A) < f(e 2 ;
A) whenever 0 < <e 2
By definition, we have
23 Ci/0>i l A) = 1
1 1 H~ Ci A ||/ 1
A
=ig< +*, >+( -g) i-i
i i i
<fi||I + e 2 A||,- + l — —1
02 02
= f02i (l|I + e 2 A||,--l)
= s f(s ;A) 1 2
Dividing both sides of the inequality (23) by yields (22). Also,

because
24 1 — eI|A|I, < [[I + eAH < 1 + f c|| A||,
it follows that
25 — A 1
1
1 1/ </(£; A) < A 1
1
| | f, V e> 0
Thus the function e i->/(s A) is nonincreasing and bounded below ;
as 8 decreases to zero and therefore has a well-defined limit as

8 — > 0+ .
(ii) The inequality (25) shows that

26 — II
A||; < /<i(A) < A 1 ! 1
1,-
Thus, to prove (ii), it only remains to show that
27 — M- A) < M A)
Now, clearly,
|I + 2eA ||I — 2eA

28 1 =||I+e(A-A)||,.< 1
1,-
,
Therefore
[[I + 2eA\\j — 1 HI — 2sA\\j — 1

29 — 2s ^ Is
Taking the limit as S —> 0 + yields
30 0 < MA) + JUi(-A)

which is clearly equivalent to (ii).
(iii) It is a simple consequence of the definition of f(s; A) that
31 /(£; aA) = a/(ae; A), v a> 0
Therefore,
32 Ht(aA) = lim f(s; aA) = a// (A), x V a> 0

e-»0 +
This proves the validity of (iii) in the case a > 0. If a = 0, (iii) fol-
lows trivially because the measure of the zero matrix is zero.
(iv) We first show that
33 JiiC A+ B) < // X (A) + jUi(B)
First, we have
34
||I + e(A + = ^(||! + 2eA + 1 + 2eB||, - 2)
^ ||I + 2eA||,- — 1 ||I + 2eB||, — 1
^ 28
+
,
2fi
Taking the limits in (34) as s —>0 +

proves (33) and establishes the
right-hand inequality in (iv). To prove the left-hand inequality in
(iv), we simply make use of (33). Clearly,
35 Mi (X) = MA + B - B) < MA + B) + M-B)

so that
36 M A) - M- B) < MA + B)
By symmetry, we must also have
37 jM/(B) - ///(-A) < //,(A + B)

The inequalities (36) and (37) together establish the left-hand
inequality in (iv).
(v) From (iii) and (iv), we have

>
Mil aA + (1 - a)B] < jUi(ccA) + // f [(l - a)B] by (iv)
= A) + (1 - a)//;(B) v a e [0,1]
by (iii)
This proves (v).
(vi) Let A be an eigenvalue of A, and let v be an associated eigenvector

such that 1 1
v ||
= 1 (where ||
•
||
is the vector norm that induces the
matrix norm ||
• ||,). Then
(I + aA)v = (1 + eA)v
Therefore,
III + cA||, - sup ||(I + eA)x||

11x11=1
>||(I + eA)v|| (because v is a specific unit vector)
— 1
1 4- cA |
This implies that
111 + fiAlU — 1 11 + CAI — 1
e — e
As £_ —
0 + the left-hand side of (41) converges to ///(A), while the
,
right-hand side of (41) (as can be easily verified) converges to Re A.
Thus, (41) reduces to the right-hand inequality in (vi). The proof of

the left-hand inequality in (vi) is similar and is left as an exercise. —
Summarizing the properties of the matrix measure function, we
see that although the measure of a matrix is a convex function as is a
norm, the similarity almost ends there. The measure can have positive
as well as negative values, whereas a norm can assume only nonnega-
tive values. The measure is sign-sensitive in that //,(— A) is in general
different from ///(A), whereas — A||, = A|| Because of these special || || f
.
properties, the measure function is useful in obtaining “tight” upper and

lower bounds on the norms of solutions of vector differential equations.
Theorem (19) lists only some of the many interesting properties
of the measure function. A more complete discussion can be found in
[10] Desoer and Yidyasagar.
nxn
In defining the measure of a matrix in C we have assumed ,
that the norm used in (18) is an induced norm. It is possible, given any
norm ||
on C nxn (induced or not), to define a corresponding measure
•
||
nxn
function //(•) mapping C into R. In this case, all the properties of
Theorem (19) still hold except for (vi). However, for the purpose of
estimating solutions of vector differential equations, only measure
functions corresponding to induced norms prove useful.
Given a particular vector norm on C", it is in general a very ||
•
||
difficult task to obtain an explicit expression for the corresponding

2
induced norm (as mentioned earlier), and therefore it is more difficult to

obtain an explicit expression for the corresponding measure function.
However, the measure functions corresponding to the norms • IU, ||
||
•
|| 1? and ||
•
|| 2 can be calculated explicitly and are displayed in Table
3.2.
TABLE 3.2
Norm on C n Matrix Measure on C nxn

||x||oo = max |
Xi |
A) = max {a + 2 |«,7
it I}
i i j*i
iixii. = i)i*/i
i=l
jUnW = max {ajj + "Z |«/y|}
j i*j
172
11*11* = (t k.-l
2
)
A2(A) = Amax [(A* + A)]/2
42 Example. Let
L-i -3J
Using the formulas given in Table 3.2, one can verify that
Mn(A) = -1; Mn(- A) - 4

/AzOV) = 2; A) = 3
îoo(A) = -1 ; /Loo(-A) = 4
This clearly shows that the measure of a given matrix depends on the
particular norm used to define the measure function. The eigenvalues of
A are
/1.2
1
- _a
-
2
±7,vr
+ —
If we use property (vi) of Theorem (19) to obtain bounds on the real
parts of the eigenvalues of A, we get
—4 < Re A 2 1( <— 1 using ju n and ju ioo measure

— 3<Re2, 12 <— using /i i2 measure
Thus for the particular matrix A studied in this example, the ji i2 measure
gives better bounds on the eigenvalues of A. However, this is very much
dependent on the specific matrix A, as shown by the following example.
43 Example. Let
0 1
The eigenvalues of A are 2 and 1, so that the smallest interval containing

the real parts of both eigenvalues of A is [1, 2]. Now, if we use the
measure functions ju iU fi i2 , and ju ioo to estimate this interval, we get
[1, 3] using // iOOJ
T3 13,11
L2
or approximately [0.793, 2.207] using ju i2 and [0, 2] using n Therefore ju, .
ju ioo gives an exact lower bound, ju n gives an exact upper bound, while
li i2 gives the interval of the smallest “width.”
In practice, to estimate the range of the real parts of the eigen-
values of a given matrix, one would utilize various measure functions
to obtain various intervals on the real line. Then, because property
(vi) of Theorem (19) is valid for all measure functions corresponding to
induced norms, one would take the intersection of all intervals in order
to obtain the best estimate. In Example (43), this intersection is [1, 2],
which happens to be an exact estimate.
Before we close this section, some comments are in order regard-
ing the use of the matrix measure in connection with norms on R n
and
nxn
the corresponding induced norms on R . First, it is easy to see that
every element in R nxn corresponds to a linear mapping from R n
into
itself and vice versa. Next, suppose ||
•
||
is a norm on R n
. Then, given
any A e RnXn we can define its induced norm A in a manner entirely
, || || f
analogous to definition (5) as well as a corresponding measure function

fi { A) in a manner analogous to (18). The question is, What properties
t
does this measure function have? Properties (i)-(v) of Theorem (19)

hold, and the proof for the complex case can be used as is. However,
it can be noted that in proving property (vi) of Theorem (19) (namely,
the bounds on the real parts of the eigenvalues of A) essential use was
made of the fact that A maps C n into itself, because 2 might be a complex
eigenvalue of A, with a corresponding eigenvector having some complex
components. In spite of this fact, however, it can be shown that, given
any norm on R n and any A e jR bx ", it is possible to “extend”
||
•
||
these, respectively, to a norm on C and to a mapping from C n into

n
itself. The details are omitted here, but the end result can be sum-
marized as follows.
Theorem Let ||
be a norm on R and let
•
||
R nXn —» R and
n
, ||
• -:
|| t
RnXn — > R be defined in a manner analogous with (5) and (18),

respectively. Then //,(•) satisfies properties (i) through (vi) of
Theorem (19).
Sec. 3.3 Contraction Mapping Theorem 73
Problem 3.7. Calculate ||A|| n , ||A|| /2 , and ||A|| /00 , when
A=
-2 1 o r
0) (ii) A
2 — 3_ -l o_
4 1
2~
(iii) A= 0 -2 1
_-l 1 — 3__
Problem 3.8. Calculate /* n (A), p i2 {A), // IOo(A), and // n (- A), // /2 (- A),
// f0 o(— A), for the three matrices above. Find an interval in the real line that
contains the real parts of all eigenvalues of A.
33
CONTRACTION MAPPING THEOREM
In this section, we shall state and prove a very important theorem,
which we shall use in Sec. 3.4 to derive the existence and uniqueness
of solutions to a class of nonlinear vector differential equations.
The theorem that we shall prove is generally known as the con-
traction mapping theorem (and sometimes as the Banach fixed point
theorem) and is usually given in two forms: the global version and the
local version. The local theorem assumes a weaker hypothesis than the
global theorem and obtains correspondingly weaker conclusions. We
shall first give the global version.
Note that, hereafter, we use mapping function , and operator
,
interchangeably.
3.3.1 Global Contractions
1 Theorem Let (X s ||
•
||)
be a Banach space, and let T : X X be
a mapping for which there exists a fixed constant p < 1 such that
2 \\Tx — Ty\\< p\\x — y\\, \? x, y e X

[where we write Tx instead of T(x) in the interests of brevity]. Then
(i) There exists exactly one x* e X such that Tx* = x*.
(ii) For any x e X, the sequence (x„)f in X defined by
converges to x*. Moreover,
4
R 1 1
Remarks: An operator T satisfying the condition (2) is known as

a contraction because the images of any two elements a and y are
,
closer together than a and y are. Furthermore, T is a global contrac-

tion because (2) holds for all a, y in the entire space X. An element
a* g such that Tx* = a* is called a fixed point of the operator
X
T because a* remains fixed when we apply T to X. Theorem (1)
,
asserts that every contraction has exactly one fixed point in X;

furthermore, this fixed point can be determined by taking any
arbitrary starting point x e X and repeatedly applying the operator
T to it; finally, (4) provides an estimate of the rate of convergence
of this sequence to the fixed point.
proof Let * e X be arbitrary. We first show that 007 is a Cauchy

sequence. For each w > 0, we have
5 1 1
X n+ 1
X„ 1 1 P\ %n
| %n- \ 1 1 ^ ^P n
1 1
*0 II
Now suppose m = n + r, r > 0. Then

r- 1
6 ll^m Aw ||
= H-Vh+j- AVj|| ^ S= / 0
l|A"n+i + l A n +,-||
< 1=0
S P n+i
‘11*1 -A 0 ||< S
1=0
p
n+i
\\Xt Xq 1
= x~-p\\
Xi ~
It is clear from (6) that ||

a„ — xm ||
can be made arbitrarily small by choos-
ing n large enough. Hence 007 is a Cauchy sequence, and because is X
complete, (07 converges (to an element of X). Let us denote this limit as
x*. Now, using definition [3.1(32)] of uniform continuity, one can easily
show from (2) that T is uniformly continuous. Therefore,
7 Tx* = T(lim x„) = lim Tx„ = lim xn+1 = x*

n~>°o tj->co n^°°
Hence a* remains invariant under T. To show that a* is the only element

of X satisfying (7), suppose that x e X and that Tx = x. Then, by (2),
8 ||
A* - All =|| 7** - Tx\\<p\\x* - All
which can hold only if ||a* — a|| = 0, i.e., a* = a (recall that p < 1).
Finally, to prove the estimate (4), consider the inequality (6). Because the
norm function ||: X is continuous, we have
||
•
9 A'
* - *„ll = II
(lim aJ - xn 1
m~* °o
= lim \\x m - x„\\ < ~~r|| ,

1
- *oll
m-'-o r
where we have used the fact that the right-hand side of (6) is independent
of m. —
Note that in general it is not possible to replace (2) by the weaker
condition
10 ||
Tx — Ty || < ||x — y\\ V x, y e X such that x y
It is easy to show that any mapping satisfying (10) has at most one
fixed point, but quite possibly it may not have any at all. As a simple
example, let X= R, and let /: R—+ R be defined by
11 /(x) = x + -y
— arc tan x
Then
12 f\x)
J v y
= — 1 1
1
p
+X
— *2 = -r—^ <
+X
1
^
1
1 f° r all x g R
Thus by the mean value theorem,
13 fix) - f(y ) =f'(<£)(x - y) for some { e [x, y]
and hence / satisfies (10). However, a quick sketch of / reveals that /

has no fixed points in R.
14 Example. Let /: R —> i? be a continuously differentiable func-

tion, and suppose
15 sup /'(*)!
Are/?
|
Ap< 1
Then Theorem is a unique number x* e 7? such

(1) tells us that there
that / (x*) = x* ;
number x* can be determined as the
furthermore, this
limit of the sequence (x )T obtained by choosing any arbitrary x 0 e R
f
as the starting point and then setting x /+1 =/(x ) V / > 0. The f
sequence of points so obtained is depicted in Fig. 3.1.
FIG. 3.1
3.3.2 Local Contractions
The application of Theorem (1) is limited by the fact that the opera-
tor T is required to satisfy (2) for all x, y in X. In other words, Thas to
be a global contraction. In Theorem (16), we examine the case where
T satisfies (2) only over some region M in X ,
i.e., the case where Tis
a local contraction, and we derive correspondingly weaker results.
16 Theorem Let (X, ||)

be a Banach space, let Mbea subset of X,
||
•
—
and let T: X >X. Suppose there exists a constant p 1 such that <
17 II
Tx — Ty\\ < p\\x — y\\, \/ x, y <e M
and suppose one can find an element x 0 e X such that the ball
B= -x Tx °y*°
18
{* G X: II X 0 II
< II
l
II
is entirely contained within M. Then

(i) T has exactly one fixed point in M (call it x*).
(ii) The sequence defined by
19 x, + i = Tx,, i>0
and x 0 is the element defined in (18), converges to x*. Further,
20 - **ll < P Tx 0 - *oll

11*1,
I _p \\
21 Remarks:
1. The significance of Theorem (16) lies in the fact that 7îs only
required to be a contraction over the set M, and not over all
of X. The price we pay is that the conclusions of Theorem (16)
are also weaker than those of Theorem (1).
2. Everything is contingent on finding a suitable element x0 in
M such that the B defined contained in M. In
ball in (18) is
means that we must be able to find an element x

effect, this 0
inM such that repeated applications of T to x a 0 result in

sequence that is entirely contained within M. Even if T satisfies
(17), it may not be possible to find such an x 0 For example,
.
let X = R, and let T be the function defined by

|2 if |*| <1
22
[0 if |
x |
> 1
If we choose M as the
interval [—1, 1], then Tis a contraction
on M. However, not possible to find an x 0 e
it is such that M
the ball B defined by (18) is contained within M. Accordingly,
T has no fixed point in M.
3. Suppose we do succeed in finding an x 0 e M such that the
hypothesis of Theorem (16) holds. Then the particular sequence
]
(x 0 ,
x 1? . .
.) converges to x* (the unique fixed point of Tin M).
However, if y is some other element of M, the sequence ( y Ty , ,
T 2 y, . .
.) may or may not converge to x*. In contrast, if T
is a global contraction, then the sequence (y, Ty, T 2 y, . .
.)
converges to x* for any arbitrary starting point y.

4. Theorem (16) does not rule out the possibility that T has
some fixed points outside it states only that T has exactly M—
one fixed point in M.
proof of Theorem (16) First, it is clear that Thas at most one fixed point
in M, because of (17). If x0 e M
chosen in such a way that the ball
is B
defined in (18) is contained within M, then clearly the sequence (xf )~ is
contained within M [apply the inequality (6) with n = 0]. Because the con-
traction condition (17) holds in B, one can show, as in the proof of theorem
(1), that (xt ) is a Cauchy sequence and therefore converges, say, to x*.
Because of the continuity of the norm function, x* e B. The rest of the
proof exactly follows that of Theorem (1). —
23 Example. Consider once again the case where X= R, and let
/: R— R > be continuously differentiable. Suppose
24 sup
*e[-l,i]
I
f'(x) |
A p< 1
and that there exists an x 0 e [—1, 1] such that the interval
= \f(Xp) l/Oo) — Xp
B \ + I
1 -p = [a, b
is a subset of [—1, 1]. Then Theorem (16) states that there is a unique
x* e [—1,1] such that /(x*) x* and that x* is the limit of the =

sequence {x 0 , / (x 0 ), /[/ (x 0 )], . . .). The situation is depicted in Fig. 3.2.
FIG. 3.2 y=m

X
We shall close this section by giving an alternative version of the

local contractionmapping theorem, which assumes a somewhat stronger
hypothesis than Theorem (16) but which is more convenient for later
applications.
25 Theorem Let X be a Banach space, and let B be a closed ball in

X, i.e., a set of the form
26 B= [x: \\x — z\\< r }
for some z E (.
, IHI) and some r Let P: X — X be
> an operator
satisfying the following conditions:
(i) P maps B into itself, i.e., Px e B whenever x e B.
(ii) There is a constant p < 1 such that
27 \\Px - Py\\< p\\x - y\\, \/x,yeB

Then
(i) P has exactly one fixed point in B (call it x*).
(ii) For any x 0 e B ,

the sequence (x„)f defined by
28 xn+1 = Px„, n> 0
converges to x*. Moreover,
29 ||JS„-X*||< -*oll
proof Obvious from Theorem (16). —

30 Remarks: The difference between Theorems (16) and (25) is that
in the present case P is assumed to map B into itself, which is a
stronger hypothesis than in Theorem (16). As a result, we can start
from any arbitrary point x 0 in B in order to compute x*.
Problem 3.8. Give a detailed proof of Theorem (25).
3.4
NONLINEAR DIFFERENTIAL EQUATIONS
3.4.1 Introduction
In this section, we shall derive some general and very useful

conditions which guarantee the existence and uniqueness of solutions
to the nonlinear differential equation
1 x(0 = i[t ,
x(/)], t > 0; x(0) =x 0
Sec. 3.4 Nonlinear Differential Equations 79
where x(t)eRn and f: R + x R n » R\ As shown in Chap. 1, the exist- —

ence and uniqueness of solutions to (1) is not guaranteed unless some
restrictions are placed on the nature of f. That is the subject of this
section. By a solution of (1) over an interval [0, 7], we mean an element
x(-) of C”[ 0, T] such that (i) x(-) is differentiable almost everywhere
{i.e., x(t) is e [0, T], with the
defined for all t possible exception of a
countable number of points] and (ii) equation (1) holds at all t where
x(t) is defined.
We shall first establish some conditions under which (1) has
exactly one solution over every finite interval [0, 8] for 8 sufficiently
small, i.e., conditions for local existence and uniqueness. We shall then
obtain stronger conditions for global existence and uniqueness, i.e.,
conditions under which (1) has exactly one solution over [0, oo).
One small point is to be cleared up before we proceed to the
theorems. First, if x(*) is a solution of (1) over [0, T], then x(*) also
satisfies
2 x(t) =x + 0 f f[r, x(t)] dr

Jo
On the other hand, if x(.) e Cn [ 0, T] satisfies (2), then clearly x(-)

is actually differentiable and satisfies (1). Thus (1) and (2) are equivalent
in the sense that every solution of (1) is also a solution of (2) and vice
versa.
3.4.2 Local Existence and Uniqueness
3 Theorem Suppose the function f in (1) is continuous in t and x

and satisfies the following conditions: There exist finite constants
7
T r A, k such that
, ,
4 ||f(7, x) - i(t, y) ||
<k\\x- y||, V x, y e B, \/te[0,T]
5 ||
f<>, x 0 )|| < h, V f e [0, T]
where B is a ball in R n
of the form
6 B= [x e R n
: || x — x0 1|
< r}
Then (1) has exactly one solution over [0, 8] whenever
7 hS exp (kS) <r

and
8 S ^ mia ( T’f’FTT?)
for some constant p < 1.
[ [
9 Remarks :
1. While following the proof of Theorem (3), it is important to

keep in mind the distinction between (which is a norm ||
•
||
on R n ), and c (which is a norm on C [ 0, 5]). Also, it should

n
•
|| ||
be noted that B is a ball in R n while S defined in (10) is a ball ,
in C n
[ 0, 8].
2. The condition (4) is known as a Lipschitz condition and the ,
constant k is known as a Lipschitz constant. Notice that we say

a Lipschitz constant, because if k is a Lipschitz constant for
the function f, so is any constant larger than k. Some authors
reserve the term Lipschitz constant for the smallest number
k for which (4) holds. Also, (4) is a local Lipschitz condition,
because it holds only for all x, y in some ball around x 0 and ,
for t e [0, T]. Accordingly, theorem (3) is a local existence and

uniqueness theorem, because it guarantees only existence and
uniqueness of solutions over a sufficiently small interval [0, 8].
Note that, given k ,

r, t, and /z, (7) and (8) can always be satis-
fied by choosing 8 sufficiently small.
proof of Theorem (3) By a slight abuse of notation, we use x 0 (*) to

denote the function in C n 0, [ 8] whose value is x0 for all t e [0, 8 ]. Suppose
8 satisfies (7) and (8), and let S be the ball in C n 0,[ (5] defined by
10 S = {x(-)e C”[0, 8]: ||x(-) —x 0 (-)|| c < r}

Let P denote the mapping of C n
[0, 5] into itself defined by
11 (Px)(0 = x + f[T, x(T)] dx V/e [0,

f 0
Clearly x(-) is a solution of (2) over [0, 8] if and only if (Px)(-) = x(-);
i.e., x(-) isa fixed point of the mapping P.

We first show that P is a contraction on S. Let x(-) and y(-) be
arbitrary elements of S ;
then x(0 and y (t) lie in the ball B, V t e [0, 8 ].
Thus
12 (Px)(r) - (Py)(t) = {f t, x(t)] - f[r, y(T)]} dx

J'
13 1 1
(Px)(t) - (Py)(t) 1 1
< J' 1 1
f t, x(r)] - f [r, y(T)] 1 1
dx
< J '/c||x(t) — y(T)

o
1 1
dx
< fa||x(0 - y(-)llc

< />l|x(-) - y(-)llc
because kt < kS < p by (8). Because the last term on the right-hand
side of (13) is independent of t, it follows that
14 ||(Px)(-) - (Py)(-)llc = sup \\(Px)(t) - (Py)(0||

re[0,<5]
<Pl|x(-) -y(-)llc
so that P is a contraction on S.
Next, we show that P maps S into itself. Let x(-) e S. Then
15 (Px)(t) — x 0 = f[T, x(t)] dx ‘
J0
= *
{ f[r, x(t)] - f(T, x 0) + f(r, x 0) } dx
Jo
- Xoll < {||f[T, X(T)] - f(T, X 0 )|| + l|f(T, Xo)|D dx
J‘
<kbd +hd<r
by (7). Hence
||(PxX-) — x 0 (-)llc = sup ||(PxXO — x 0 ||< r

fG[0,<5]
so that (Px)0) belongs to S.

Now, because P maps S intoitself and is a contraction on S, it has
exactly one fixed point in S,
by Theorem [3.3(25)]. Our objective, however,
is to show that P has exactly one fixed point in C n 0, <$], not just in S (the
[
point being that S is a proper subset of C n [ 0, 5]). Thus the proof is com-
plete if we show that any fixed point of P in C n [0, d] must in fact be in S .
Accordingly, suppose x(-) e C"[0, d] satisfies (2). Then we have
17 x(0 - x 0 = f [T, x(t)] dx

£
=J '
{f[T, x(T)] - f(T, x 0 ) + f(T, Xo)} dx
o
18 ||x(0 — x0 < 1 1
A| | x(t) — x(0)|| dx +ht
<hd + k\ |
x(t) — x(0) 1 1
dx
Applying the Bellman-Gronwall inequality (see Appendix I) to (18), we get
19 ||x(0 — x 0 < hd exp (kt) < hd exp (kd),

1| V t e [0, 5].
From (7) and (19), it S. Thus we have shown that any

follows that x(-) g
fixed point of P in C n 0, be in S. Because P has exactly one
[ 8] must in fact
fixed point in S, it follows that P has exactly one fixed point in C"[0, 8].
By the manner in which P is defined, we conclude that (2) has exactly one
solution over [0, cJ].
1BBB
The following result is actually a corollary to Theorem (3) but

is in aform that can be readily applied.
20 corollary Consider the differential equation (1). Suppose that in some

neighborhood of (0, x 0 ) the function f has continuous partial derivatives
with respect to its second argument (x) and continuous one-sided partial
derivatives with respect to its first argument ( t ). Then (1) has exactly one
solution over [0, 8], provided 8 is sufficiently small.
proof The differentiability properties assumed on f ensure that f
satisfies (4) and (5) for some set of finite constants r , t, k and,
h. M
Example. Consider the Van der Pol oscillator, which can be
described by the set of two first-order equations
*i= *2
X = — Xi —
2 1
— 1)*2
This set of equations is of the form (1), with f: R2 R2 defined by
f(x) = [x 2 —x - x li(x\ - l)x 2 ]'
Because each component of f is continuously differentiable, it follows

by corollary (20) that, starting from any arbitrary initial condition
[jcio x 20 ]', (22)— (23) has a unique solution over [0, 8] for some suffi-
ciently small 8.
3.4.3 Global Existence and Uniqueness
In this subsection, we shall show that (loosely speaking) if f

satisfies a global Lipschitz condition, then (1) has a unique solution over
[0, oo).
Theorem Suppose that for each Te [0, oo) there exist finite
constants kT and h T such that
||f(t, x) - f(t, y) ||
< kT ||
\
x -y ||, V x, y e Rn V , t e [0, T]
||f(t, x0 1|
< hT ,
V ? e [0, T]
Then (1) has exactly one solution over [0, T], VTe [0, oo).
proof We shall give two alternative proofs.
no. 1 Let T< kT and h T be finite constants such

oo be specified, and let
that (26)— (27) hold. Then the hypotheses of Theorem (3) are satisfied, with
r = t = oo. Thus, by Theorem (3), it follows that (1) has a unique solu-
tion over [0, 5 ] whenever 8 satisfies
for some constant p < 1. Suppose a positive number 8 satisfying (28) is

chosen. If T< 8, the theorem is proved, so suppose 5. Now (1) has T>
5]. Denote this solution by y ), and consider
(
a unique solution over [0, i *
|
the differential equation
29 *0) = fit?, x0)]; x(0) = yi((5)

where
30 fl (t, x) = f(/ + 5, x)
Thenf! also satisfies (26) and (27); therefore, once again by Theorem ( 3 ),
(29) has a unique solution over [0, S], where 8 same as before. Denote
is the
this solution by y 2 (*)- It is ea sy to verify that the function x 2 (-) defined by
yi(?>, o<t<S
31
.y i(t — 5), d < t<25
is the unique solution of (1) over the interval [0, 28]. Proceeding by induc-
tion, let x m (-) denote the unique solution of (1) over the interval [0, mS\
and consider the differential equation
32 x(0 = f At, x(r)]; x(0) = x m (md )
where
33 f mit, x) = fit + mS ,
x)
Let ym+ i(*) denote the unique solution of (32) over the interval [0, (5] (the
same 8 as before). Then the function x m+1 (>) defined by
x m(0j 0 < < m8
t
34 m+1
^ .
mS < < mS + t 8
is the unique solution of (1) over the interval [0, m8 + <5]. In this manner,
the unique solution can be extended to all of [0, T].
no. 2 Let T< oo be given, let P: Cn

[ 0, T] be given by (11),
[ 0, T ]
— * Cn
and let x 0 (*) denote (as before) the element of C”[0, T ] whose value is
x 0 V t e [0, T]. We show first that the sequence [P m x 0 (-)]m=i is a Cauchy
sequence in C n [ 0, T ] and that it converges to a solution of (2).
Let x m (* ) = (P m x 0 )(*)- Then we have, first,
35 xj(0 - x 0 (O = f f(r, x 0 )Jt

Jo
36 l|xj(0 -<£ X 0 (OII

Xo) ||f(T, II
dT < hT t
In general, for m > we have 1,
37 ||x m+I (0 - x m (OII< '||f[T, x m _n(T)J -iX [ ,

x m _i(t) dx ||
Jo
< k T J' ||x m (r) - x m _ 1 (r)|| dx
Substituting (36) into (37) and proceeding by induction, we get

t m
38 l|x m(0 - x m _i(OII < kq-'hr —
1 1
Thus for any integer p > 0 we have

p- 1
39 X m (/)|| < 2 1 1
Xm + * + 1 ) X„, +l-(^)||
*m+i + l
< s hT kr\m + i+ i)i
m+p fi
= =S
m+
i
Wtt
40 |x m+ p(-) - x m (-)llc = sup ||x m+J,0) - x m (OII
t€[0,T]
m+p T7 00
< =2m + i 1
h T kif 4t- <c 1
^ • i
2 M#-
—m+
1
/!
Now consider the sequence
41
As m— > sequence converges to ( h T /k T ) exp ( k T T ). Moreover, the

oo, this
1
the difference between 2^=0 h T kT~ (T li\) and (h T /k T)
i
last term in (40) is
exp (k r T ) and can therefore be made arbitrarily small by choosing m

sufficiently large. Hence [x m (-)lT is a Cauchy sequence in C [0, T], and
n
because CÔ, T] is a Banach space, [x m (-)]7 converges to a limit in C"[0, T].
Denote this limit by x*(-)«

1
Whenever Zj(») and z 2 (*) are two elements in C [ 0, 7 ], we have

n
42 (PZiYt) - (Pz 2 )(t) = {f[T, Z,(T)] - f[T, Z 2 (T)]} rfT

J'
43 ||(Pz,X0 - (PZ 2 XOII < l|f[T, Z,(T)] - f[T, Z 2 (t)]|| rfT

J'
<Arr r|| Zl (-) -z 2 (-)llc

44 >
||(/ Z 1 )(-) - (PZjXOIIc = SUP IKi’ZlXO - (-PZzXOII
/e[o,T]
<£jT||z,(-) - z 2 (.)llc
Because k T T is a finite constant, it follows that P is uniformly continuous

on C"[0, T], Hence, if [x m (-)]“ converges to x*(-), we have
45 (Px*X0 = lim (PxJ(-) = lim x m+1 (-) = x*(-)

m-*oo
Thus x*(-) is a solution of (2).
Next, to show that x*(-) is the only solution of (2), suppose y(-) also
satisfies (2). Then
46 y(/) - x*(/ ) = {f[T, y(/)] - f[T, x*(t)]} dz

J‘
47 ||y(0 - x*(r)|| < k £ ||y(r) - x*(r)|| dx

Applying the Bellman-Gronwall inequality to (47), we get
48 ||y(/)-x*(OI|=0 V te [0, T]
i.e., y(*) = x*(*). This shows that x*(-) is the unique solution of (2). —
49 Remarks :
m
1. The sequence [P x 0 (*)] that converges to the solution x*(-)
of (2) is known as the sequence of Picard’s iterations, and this

method of generating a solution to (2) is known as Picard’s
method. Actually, it is easy to show that Picard’s iterations
converge starting from any arbitrary starting function in
C"[0, T ]
and not just x 0 (-).
2. Note that some authors assume that f(t, 0) = 0. This assump-
tion, together with (4), implies (5), because then ||f(/, x 0 )]| <
k\\ x0 1|. However, in “forced” nonlinear systems, it is not neces-
sarily true that f(t, 0) = 0. The present development does not
require this assumption.
We shall next prove two theorems regarding the solutions of (2).

In effect, Theorem (25) states that (2) has a unique solution correspond-
ing to each initial condition. Theorem (50) shows that, at any given
time, there is exactly one solution trajectory of (2) passing through
each point in R n
. Theorem (58) shows that the solution of (2) depends
continuously on the initial condition.
50 Theorem Let f satisfy the hypotheses of Theorem (25). Then for

each z e R n
and each Te [0, oo) there exists exactly one element
z0 e Rn such that the unique solution over [0, T] of the differential
equation
51 *(/) = f|/,x(f)]; x(0) =z 0
satisfies
52 x(T) = z
proof Consider the equation
53 ±(0 = f,[7, x(f)]; x(0) -z

54 f,(/, x) = —f(T - t, x), V t e [0, T]
Then f, also satisfies the hypotheses of Theorem (25), so that (53) has a
unique solution over[0, T]. Denote this solution by y(-), and define z 0 =
y(T). Then one can easily verify that the functions y^-) defined by
55 yi(0 = y(T-t), V te[0, T]

1 ] x
satisfies (51) and also satisfies (52). To prove the uniqueness of the element
z 0 corresponding to a particular z, assume by way of contradiction that
there exist two functions >',(•) and y 2 (-) in C”[0, T] that satisfy (51) and
(52). Let y x (0) = zu y 2 (0)
= z2 . Then the functions y„(-) and y6 (-) defined
by
56 y a (t) = yi(r - t)
57 y b (t) = y (T - 2 t)
must both satisfy (53). However, because the solution to (53) is unique, it
follows that y^-) =y 2 (-)- Thus Zj

= z2 . m
58 Theorem Let f satisfy the hypotheses of Theorem (25). Let
Te [0, oo) be specified, and suppose x(-) and y(-) are two functions
in C"[0, T satisfying
59 = x(t)];
x(t) x(0) = x
f[t, 0
60 y(0 = f[^y(01; y(0) = y o
Then for each e > 0, there exists a T) > 0 such that c5(e,
61 x(*) - y(-)llc < e whenever ||x — y

II
< T) 0 0 1|
<5(e,
proof The functions x(-) and >(• ) also satisfy
62 x(t) = x0 + f[T, x(t)] dx

J*
63 y (?) = yo + £ f[T, y(T)] dx

Subtracting, we get
64 x(t) - y (t) =x - 0 y0 + £ {f[T, x(t)] - f[r, y(r)]} dx

65 1
x(t ) -y(t)||<||x 0 — yoll + A: r Jj|x(r) -y(x)\\dx
Applying the Bellman-Gronwall inequality to (65), we get
66 ||x(t) - y(/)||< ||x 0 - y 0 ||exp(/cr t)

Hence
67 1
1 x( • ) — y(*)llc < II 0 - yoll exp (kT T)
Thus given € > 0, (61) is satisfied if we choose d(€, T) = e/exp (kT T ). ^

Remarks: The results contained in Theorems (50) and (58) can be
given a simple geometric interpretation in terms of certain mappings
being continuous. Let 0:
n
-»
n
[ 0, R — C
T] be the mapping that asso-
ciates, with each initial condition x 0 in Rn ,
the corresponding unique
solution of (2). Then Theorem (58) states that 0 is uniformly continu-
x
ous on Rn . In the same vein, let y/: R n — > R n be the mapping that
associates, with each initial condition x0 e R n the value at time T
,
of the corresponding unique solution of (2). Then Theorem (50)

states that y/ is one-to-one [i.e., given y/(x), one can uniquely deter-
mine x] and onto (i.e., the range of y/ is all of R n ). Furthermore,
Theorem (58) shows that both y/ and its inverse map y/~ are con- x
tinuous.
69 x(0 = tanh x(t ) A f[x(t)\; x(0) = x0

Since the function x — tanh x
> is everywhere continuously differenti-
able, and since this derivative is everywhere bounded (in magnitude)
by 1, it is easy to verify that /(•) satisfies a global Lipschitz condition
of the form (26) with kT —

1 for all T (see also Problem 3.10 infra).
Also, for every x 0 there exists a finite h T such that (27) holds. Hence, by
,
Theorem (25), (69) has a unique solution over [0, oo) corresponding to
each x 0 moreover, this solution depends continuously on x 0
; .
70 Example. Consider the linear vector differential equation
71 x(t) = A(t)x(t); x(0) = x0
where A(«) is piecewise-continuous. Let ||

•
||
be a given norm on Rn .
Since A(.) is piecewise-continuous, for every finite T ,

there exists a
finite constant kT such that
72 II
A(0 II,
< kT , V / G [0, T]
Hence we have
73 ||A(r)x — A(0y||</cr ||x - y II, V x, y e R"; V t e [0, T]
74 II
A0)x 0 1|
< kT \\
x0 1|, V(e [0, T]
So (26) is satisfied with k T as above, and (27) is satisfied with h T =

kT ||
0 1|. Therefore, (71) has a unique solution over each finite interval
[0, T] corresponding to each initial condition x0 ;
moreover, this
solution depends continuously on x 0 .
In conclusion, in this section we have derived some conditions

that are sufficient to ensure that a given nonlinear vector differential
equation has a unique solution over some interval. It is easy to construct
counterexamples to show that the conditions derived here are by no
means necessary for the existence and uniqueness of solutions. For
instance, consider the scalar differential equation
x{t) = —x 2
; x(0) = 1
:
This equation has a unique solution over [0, oo) [namely, x{t) =
1 j(t + 1)], even though the function f(x ) = x 2
is not globally Lip-
schitz-continuous.
Problem 3.9. Let ||

•
|| fl
and ||
•
|| 6
be two given norms on Rn Show that
.
for each finite T there exists a finite kaT such that
||f(/, x) - f (/, y)|U < Mix - y |U V x, y e R*, V / e [0, T]
if and only if, for each finite T there exists a finite kbT such that
IlfO, X) - f o, y)!U < Mix - y|L V X, y e R\ V t e [0, T]
In other words, show that, in verifying whether a given function f satisfies
(26), the norm used on Rn is immaterial. [Hint: Use fact (3.1.52)].
Problem 3.10. (a) Let f:R + x R » R be continuous, and continuously —

differentiable in the second argument. Show that /satisfies (26) if and only if,
for each finite T there exists a finite kTf such that
dx
< kT , V x e R, V t e [0, T]
i.e., df(t, x)/dx is bounded independently of x over each finite interval [0, T].
(Hint: Use the mean value theorem).

(b) Let /: R+ x Rn — > Rn be continuous, and continuously differen-
tiable in the second argument. Show that f satisfies (26) if and only if, for each
finite T there exists a finite kT such that
dfi(t, x)
n
^ kT i
j V i,j, V X G Rn , V t e [0, T]
dxj
(Hint: Use the results of Problem 3.9, above).
(c) Determine whether the following functions satisfy a global Lipschitz

condition
(i) f(x) = [x] — x x 2 2xi — xl\ t
(ii) f(x) = [Xi exp (— x%) x 2 exp (— x?)]'
3.5
SOLUTION ESTIMATES FOR LINEAR EQUATIONS
In this section, we shall give a method for obtaining both upper and
lower bounds on the norm of the solution of a given differential equa-
tion. The Bellman-Gronwall inequality (Appendix I) does give an easily
applicable upper bound on the norm of the solution of a linear differ-
ential equation, and a similar inequality known as Langenhop’s inequal-
ityprovides a lower bound. However, both of these bounds suffer from
the deficiency of being sign-insensitive ; i.e., they give exactly the same
Sec. 3.5 Solution Estimates for Linear Equations 89
estimates for
x(Y) = A(/)x(/)
and
x(0 == — A(/)x(f)
They do so because both the Bellman-Gronwall inequality and Langen-
hop’s inequality utilize ||A(f)||, which is, of course, sign-insensitive.
In contrast, the method given here utilizes the concept of the matrix
measure, which is sign-sensitive. As a result, the bounds derived in this
section are always “tighter” than (or the same as) those given by the
Bellman-Gronwall and Langenhop’s inequalities.
Theorem Consider the differential equation
±(t) = A(t)x(t), t> 0

where x(t) e Rn
and A(-) is piecewise-continuous. 4
,
A(t) e RnXn ,
Let be a norm on R and let

||
•
||
and n
denote, respectively,
, ||
- ||, fjL t
the corresponding induced norm and matrix measure onf?nXn Then, .
whenever t tQ > >

0, we have
11 x(/o) II
exp - Hi- A(r)] c/r| <
{£
II
x(/) || < ||x(/ 0) ||
exp
{ £
^,[A(t)] JtJ
proof Clearly, the right-hand side of (2) is globally Lipschitz-continu-

ous in x, so that by Theorem [3.4(25)], (2) has a unique solution over all
1
bounded intervals of the form [0, 7 ]. To prove the inequalities (3), we
begin by observing that, from the integral form of (2),
x(t + 5) = x(t) + <5A(f)x(0 + o(<5), V <5 > 0
where o(<5) is used to denote an error term with the property that
italMM _o
<5 -+ 0
+ O
Rearranging (4), we get, successively,
x(t + d) = [I +
SA(t)]x(t) + o (5)
llxC + S) < III + <5A<y>||,-||x(OII + 0 (d)

II
|]x(f + S - ||x(f)|| < (III + JA(0IU - l)l|x(f)ll + 0 (d)

) II
4 By this we mean that, over each finite interval, A(-) is continuous at all
except a finite number of points; at each point of discontinuity, A(«) has well-defined
left and right limits and the norm of the difference of these limits is finite.
;
£
Chap. 3 Nonlinear Differential Equations

90
9 £||x(r)|| = lim
1|x(f
< ^[A(0]||x(0il
where d + /dt[ ] denotes the right-hand derivative. If we multiply both sides
of inequality (9) by the integrating factor
10 exp j—
The proof of the left-hand inequal-
we get the right-hand inequality in (3).
ity in (3) is entirely similar, starting with
11 x(f — 5) = x(t) — SA(t )x(t ) + o(<5)
The completion of the proof is left as an exercise. m

Theorem provides both upper and lower bounds for the norm
(1)
applying the
of the solution of the unforced linear equation (2). In
bounds it is important to remember that the norm being
used and the
(3),
measure must correspond to each other. The application of Theorem
(1) is illustrated by the
following examples.
12 Example. Consider equation (2), with n = 2 and
13 A(rt = r- 2 * ll
L-i -rj
14 x(0) = [l 0]'
calculate the measures and n,„ of A(f) as well as

First, we ji n , fi i2 ,
— A(0- This gives
15 y«„[A(0] = /UA(01 = ~t + 1
16 /f„[-A(f)] = /U-A(0] = 2t+l

17 finlMt)] =
18 /i, 2 [-A(0] =2 1
Thus, applying inequalities (3) with each of the above measures gives
19 exp (-t - t )
2
< |x t (0l + 1
x 2 (t) |
< exp (--t — -2 -
>
for ||
• iii»
20 exp (-* - t
2
) < |
xi(t) |, |
x 2 (t) |
< exp (-t 2 j
)
for ||
||oo, /tj-oo
21 exp (-t 2 ) < {|x,(0l + |x 2 2 1/2

2 (0I }
< exP -t2 for ||
•
Ik. J«<2
Thus the same two inequalities (3), when applied with different mea-
sures, yield different estimates for the vector x(t)-

By way of illustrating
vector x(l) is
the bounds obtained above, the regions to which the
confined by each of these bounds are shown in Figs. 3.3, 3.4, and 3.5.
91
|

92
22 Example. Consider equation (2) with n =2 and
23
24 x(0) = [l 2]'
Then the actual solution for x(r) is
25 x(0 = li~e
-‘ 2
- ê 5,2 ' !
However, if we calculate the various measures of A(t), we get
26 /*n[A(0] = -t\ = 51
27 MnWf)] - -2.97/; /i (2 [-A(r)] - 5.03r

28 aUA(/)] = -2/; /UA(01 = 6/
Thus the corresponding estimates for x(t) are as follows:
29 3 exp — < I *i(0 | + |

x 2 (t ) |
< 3 exp 2^ for||x(/)|| 1
30
2
a/3” exp (— 2.52r ) < {| x (0
;
2
+ |
x 2 (t) |
2 1/2
} < ^/T exp (-1.49/*)
for ||x(/)|| 2
2 exp (— 3t 2 ) < |
*,(/) |, |
x 2 (0 1
< 2 exp (-t 2
)
for ||
x(/> |U
These bounds are depicted for the case t = 0.5 in Figs. 3.6, 3.7, and
3.8.
1
Sec. 3.5 Solution Estimates for Linear Equations 93
We turn now to “forced” linear equations.
32 Theorem Consider the linear differential equation
33 ±{t) = A(t)x(t) + v(Y), V ^0

where x(t), v(Y) g Rn ,
A(t) g R nXn V t > 0, and v(«), A(-) are piece-
wise-continuous. Let ||
•
||
be a norm on R n
,
and let ||
• ||„
denote, respectively, the corresponding induced norm and matrix

measure on R nxn . Then, whenever t > > t0 0, we have
34 1 1
x(t ) ||
< exp £ a,[A(t)] rfrj
|
1 1
x(f 0 ) 1
exp /iJAfa)] <fc|||v(r)|| dr

+| {£
35 > exp {£ -A,[-A(t)] ||

x(0 ||
“1 exp ”A'[~A(5-)] *||!v(t)|| dr

{l
94
proof We first prove (35). As in the proof of Theorem (1), we have

36 x(t - 5) = x(t) - 8A(t)x(t) - S'/it) + o((5)
= [I - <5A(?)]x(r) — Sv(t) + o(S )

37
38
39
If
||x(/ - 15)
JM£)J1
So
< ||i - 5A(<)IW|x(t)|| + £||v(OII + o(S)
||
- \Ml -
^[||x(0||]> -M-A(0]||x(0ll
we
^)ll
> -1 III - ÂC
multiply both sides of (39) by the integrating factor

IKOII
<
)ll(||x(
0 -
|| ||
v(/)|| + ô
40 exp #[-A(t)] rfr}

{£
and regroup terms, we get
/t,[-A(r)] 4t}) > -II v(OII exp ^,[-A(t)] dx\

41 J(||x(/)|| exp {£ {£
Integration of (41) yields (35). Inequality (34) is proved similarly. —
It is clear that if v(t) = 0, the bounds (34) and (35) reduce to
those in (3). In (34) and (35), the first term on the right-hand side
while the
can be thought of as the effect of the initial condition x(f 0 ),
of the forcing function
second term can be thought of as the effect
v(0-
We close this subsection with an observation on systems of the
type (2), i.e., unforced linear systems. Suppose x(/ 0 ) ^0, so that
t, the extreme left-hand
side of (3) is
||
x(/ 0 ) ||
> 0. Then for all finite
while the extreme right-hand side is always
always strictly positive,
means that a system of type (2) cannot
finite. Physically speaking, this
have a finite settling time (i.e., it cannot go from a nonzero initial state
to a zero state in a finite amount of time), or a finite escape time [i.e.,

x(0 cannot “blow up” in a finite amount of time]. In contrast, a linear
discrete-time system described by
42 x* +1 = Akxk
settling time (though not finite escape time). For
can exhibit finite
example, if Ak is singular and x k belongs to the null space of A k then ,
Xi = 0Vi>k+l. As a more extreme example, if Ak = A V k,
initial state is reduced

where A is a nilpotent n X n matrix, then every
to zero in at most n sampling intervals.
Problem 3.11. Calculate upper and lower bounds for||x(0lli> IWOIU,

!|x(0IU given
Sea 3.5 Solution Estimates for Linear Equations 95
_* 2 (/)_ - —3 t jLx 2 (/)_ ’

_x 2 (0)_ — 1_
_
—1 "^(0" ~2~
pCi(0"j r-2 + 2 sin/ 1
~*i(0)
^ _*2(0- - 2 —3 — 2 cos t_ _x 2 (0- ’
-* 2 (0)- _ 1 _
Problem 3.12. Calculate upper and lower bounds for ||x(/)|| l5 ||x(0IU,
||x(0l|oo, given
“
~ii(0~
~ — + 2 sin
t t 0 ~*i(0~
" 1 ~
#
® -* 2 (0- - —1 —lt — cost__x 2 (t)_ _e~L ’

:
Approximate analysis methods
In this chapter, we shall present several methods for approximately ana-

lyzing a given nonlinear system. Because a closed-form analytic solution
of a nonlinear differential equation is usually impossible to obtain
(unless the equation has been “cooked” in advance), methods for carry-
ing out an approximate analysis are very useful in practice. The methods
presented here fall into three categories
1. Describing function methods consist of replacing a nonlinear
element within the system by a linear element and carrying out
further analysis. The utility of these methods is in predicting
the existence and stability of limit cycles, in predicting jump
resonance, etc.
2. Numerical solution methods are specifically aimed at carrying

out a numerical solution of a given nonlinear differential equa-
tion using a computer.
3. Singular perturbation methods are especially well suited for the
analysis of systems where the inclusion or exclusion of a par-
ticular component changes the order of the system. (For exam-
an amplifier, the inclusion of a stray capacitance in the
ple, in
system model increases the order of the dynamic model by
one.)
should be emphasized that the above three types of methods are
It
only some of the many varieties of techniques that are available for the
approximate analysis of nonlinear systems. Moreover, even with regard
Sec. 4. 1 Describing Functions 97
to the three subject areas mentioned above, the presentation here only
scratches the surface, and references are given, at appropriate places, to
works that treat the subjects more thoroughly.
4.1
DESCRIBING FUNCTIONS
4.1.1 Optimal Quasilinearization
Consider a nonlinear element N, which can be thought of as an

operator from C[0, oo) into itself, where C[0, oo) denotes the linear
space of continuous real-valued functions over [0, oo). In other words,

given any continuous function x(-) over [0, oo), the nonlinear element
N associates, with x(«), another function (Nx)(>) in C[0, oo). It natural is
to think of (7Vx)(-) as the output of N corresponding to the input *(•)•

A commonly encountered type of nonlinearity is the so-called memory-
less nonlinearity , where (Nx)(») is of the form
1 (Nx)(t) = n[t, x(7)]

where n(t *):R-^ R. The rationale for calling an operator of the
,
>
form (1) memoryless is that the output at time t [namely (7Vbc)(r)] depends
solely on the input value at the same time t [namely x(t)] and not on —
the past or future values of x(«).
The problem studied is the following: Suppose
in this subsection
that a particular function which might be called the
x 0 («) in C[0, oo),
reference signal is given. ,

The objective
is to approximate the function
( Nx 0 )( •) by means of another function (IFx 0 )(«) of the form
2 (JVxoXO = f
Wo (t — t)xq(t) dx
Note that (Wx Q )(>) can be interpreted as the output of a linear time-
invariant system with input x 0 (-) and impulse response w 0 (>y. Thus the
objective is to approximate the output of the nonlinear operator
[namely (iVx 0 )(«)] by means of the output of a linear system [namely
(JFxoX-)]. Furthermore, we would like the approximation to be the
“best possible,” in a sense to be made precise next. We would like to
choose the impulse response w 0 («) in such a way as to minimize the
error criterion
3 e(w) = lim ± f l(Nxo)(t) -

1 Jo
T-*°°
(Wx 0 )(t)] 2 dt
êe Sec. 6.1 for a definition of some of these terms.

98 Chap. 4 Approximate Analysis Methods
assuming, of course, that the indicated limit exists and is finite. The error
criterion e is recognized as a measure of the average mean squared devia-
tion between (Nx 0 )(-) and (Wx <,)(•)•
In the sequel, we make the following assumption regarding the
nonlinearity N: For each finite constant b there exists a corresponding ,
finite constant m Q (b) such that
4 |
(Nx)(t) |
<m Q (b) whenever x(-) e C[0, oo) and |
x(t) |
<b
Assumption (4) is not overly restrictive and places the
mathematical
manipulations that follow on a firm mathematical footing. For the same
2
reason, we restrict the impulse response w( ) by requiring that •
5 f |
w(7) dt |
< oo
Jo
These two assumptions, together with the restriction that the reference
input x 0 (-) is bounded, are enough to assure that all the integrals
encountered in the following arguments exist and are finite.
To determine the optimal choice for the impulse response w(-),
we use a standard variational argument. Assume that, for the problem
at hand, an optimum choice of w(-) exists, and denote it by w 0 (-)-
Suppose w(*) is any other impulse response; then, because we are as-
suming that w 0 («) is the best possible choice, we must have
6 e(w) > e(w 0)
or, in other words,
7 e(w ) — e(w 0 ) > 0, for all w(-)
Now, by definition,
8 e(w 0 ) — lim ~ f [(NxoXO — (ôôXOP dt

T~*°° 1 Jo
Hence
9 e(w) — e(w 0 ) = lim ^ f [(Nx 0 — Wx Q )\t)

r->oo 1 JQ
— (Nx — o WoXo)
2
(t)\ dt
= lim i, f' [(Dx a

r-oo 1JQ
m+ 2(Dx 0 )(t)-
(W0 x 0 - Nx 0 )(t)\ dt
where we define
10 (Dx 0 )(t) = (fVx 0 )(t) - (W x 0 0 )(t)
2 Assumption
(5) implies that the linear system represented by W is bounded-
input/bounded-output stable. See Sec. 6.3.
Next, for the right-hand side of (9) to be nonnegative for Dx 0 the

all ,
necessary and sufficient condition is that the linear term in Dx 0 should

be identically zero; i.e., we must have
11 [w V—0—W 0 {t
— r)]x 0 (r) ch
(W0 x 0 — Nx 0 )(t) dt = 0
Next, it is easy to verify that
12 f [w(t — t) — w 0 (t
— t)]x 0 (t) dr = f [w(r) —w 0 (t)]x 0 (7
— t) dr
Jo Jo
Substituting from (12) into (11) and interchanging the order of integra-
tion, we obtain
13 limyj [w(t) —w 0 (t)] x o (t — r)(fV o xo — Nx 0 )(t)dtjdr

=0
For (13) to hold for all functions w(«) — w 0 (*), we must have (after
interchanging limiting and integration with respect to r)
14 lim f x 0 (t — r ){W Xq — Nx q 0 )(t ) dt =0 V t >0

T ~*°° 1 Jr
Equation (14) can be rewritten more compactly. Define
15 Ko(T) = lim S f x 0 (t - z)OV 0 x 0 )(i) dt

1 Jr
16 </>n(z ) = lim ~
r- ~ i Jr
f x a (t — T)(Nx 0 )(t) dt
The function 0^ (*) is the cross-correlation function between the func-

o
tions x 0 (») and (^FoôXO? and similarly for N (>). With these definitions, (j)
(14) reduces simply to
17 V T >0
Equation (17) represents an important principle, namely: A
necessary and sufficient condition for an impulse response w 0 (») to be
optimal [in the sense of minimizing the error criterion in (3)] is that the
cross-correlation function between x 0 (*) and (fiE0 x 0 )(») is the same as
the cross-correlation function between x 0 (-) and (iVx 0 )(*)-
A few comments are in order regarding the above result: Given a
particular reference input x 0 (*), if some w 0 («) satisfies (14), this can be
taken to mean that a linear system with impulse response w 0 (*) is the
best possible linear approximation to the nonlinear operator N ,
with
the reference input x 0 (*)- However, (1) for a given x 0 («), there is in gen-
eral more than one w 0 (-) that satisfies (14), and (2) a function w 0 («) that
satisfies (14) for one choice of reference input x 0 (*) need not do so for
some other choice of x 0 (-). Keeping these points in mind, we refer to

an Y Wo(*) that satisfies (14) as an optimal quasilinearization of with N
reference input jc 0 (.).
In the preceding derivation, we assumed that the reference input

x 0 (*) is deterministic. However, it is also possible to carry out an opti-
mal quasilinearization of N with respect to random reference input.
Also, the above development can be extended to multi-input / multi-output
nonlinearities, withno essential changes. For a detailed discussion, see
[12] Gelb and Vander Velde.
4.1.2 Equivalent Linearization and Harmonic Balance
In this subsection, we shall specialize the results of Sec. 4.1.1 to

obtain the optimal quasilinearization of a memoryless time-invariant
nonlinearity in two special cases: (1) when the reference input is a con-
stant or bias signal and (2) when the reference input is sinusoidal.
Constant Reference Input
Suppose the reference input x 0 (*) is a constant, i.e.,
18 x 0 (t) = V t > 0
for some real number k. Suppose also that the nonlinearity N is memory-
less and time-invariant; i.e., suppose that
19 (Nx)(t) = n[x(t)\, V t > 0, V x(-) e C[0, oo]
where n : R— > R. Then the cross-correlation function (f> N {x) becomes

20 0*(t) = kn(k), V >0 t
Thus the optimality condition (17) becomes

21 = kn(k), V t >0
There are several possible solutions to (21). However, if k ^ 0, the
simplest choice for w 0 («) that satisfies (21) is
22 *>„(*) = ^<5(0
A system whose impulse response is given by (22) is nothing but a con-
stant gain of value [n(k)/k]. Thus we have shown the following.
Conclusion. If N is a memoryless time-invariant nonlinearity of
the form (19), then an optimal quasilinearization of N with respect to
a constant reference input k ^0 is a constant gain of value [n(k)/k].
The latter constant gain is sometimes referred to as the equivalent lin-
earization of the nonlinearity N
with respect to the constant reference
input k.
Sinusoidal Reference input
Next, suppose x 0 (.) is a sinusoidal function with period Injco and

amplitude a; i.e., suppose x 0 (-) is of the form
23 x 0 (0 = a sin cot
where we take a >
0 without loss of generality. In (23), we do not
include a phase angle [i.e., we do not choose x 0 (f) = a sin (cot + 0)]
because, as will be evident, 0 can be taken to be zero.
If N
is of the form (19), then (Nx 0 )(*) is a periodic function with
period Injco. Hence (Nx 0 )(>) can be expanded in a Fourier series, of the
form
24 (Nx 0 )(t) = cQ 2
+ i=i (c t cos icot +d
t
sin icot)
In general, there is no reason to assume that the d.c. bias term c Q in (24)
is zero. However, if the nonlinear element N, in addition to satisfying
(19), is also odd, i.e.,
25 n(— a) = —n(cj), V a e R
then not only is c 0 zero, but all the c-s in (24) are zero. In this case, (24)
simplifies to
26 (Nx 0 )(t) =2
i=i
d t sin icot
Next, when we apply the optimality condition (14), it is easy to

show that
27 (f> N(r) = lim -tf r a sin co(t — t) d

1
t sin cot dt, V t >0
i
r-,oo
JT
because of the orthogonality property of the family of functions
(sin icot)î, namely,
r*2n/co
28 sin icot sin yatf dt — 8u —

Jo 60
where <5 i7 is the Knonecker delta defined by
= if i =j
29 <5,7
li if i ^j
In other words, 0^(0 is the same as the cross-correlation between the
functions a sin cot and d t
sin cot. Thus (17) is satisfied if w 0 («) is chosen
such that
30 (W0 x 0 )(t) = d t
sin cot
In turn, (30) is satisfied if w 0 (*) is chosen as
31 w a (t) = A <5(0
i.e., if w 0 (-) corresponds to a constant gain of value dja.
Thus we have the

Conclusion: In the case of a sinusoidal reference input of the
form (23), the optimal quasilinearization is once again a constant gain.
An interpretation of the equivalent gain dfa is the following: If
the nonlinear element N is replaced by a constant gain of value y, the
resulting output of the constant gain element in response to an input
*o(0 is
32 ya sin cot
Comparing the function (32) with the output (26) of the nonlinear ele-
ment N we ,
see that if we choose y = dja ,
then the first harmonic of the
output (26) is exactly matched by (32). This is sometimes referred to as

the principle of harmonic balance.
Next, we shall give an alternative derivation of this principle —one
which, incidentally, removes the unnatural assumption (25). As shown
previously, if the reference input x 0 (») is chosen according to (23), and
if the operator N satisfies (19), then (2 Vjc 0 )( • ) is of the form (24). Now,
if we calculate (W0 x 0 )(»), keeping in mind that W 0 represents a linear
time-invariant system, then, in the steady state , (WôXO is of the form
33 (W0 x 0 )(t) = y t a sin cot + y 2 a cos cot
where
34 + jy = w 2 0 (jco)
[In other words, if we calculate the Laplace transform of w 0 (*) and

denote it by w 0 (-)> then y x
is the real part of w 0 (jco) and y 2 is the imagin-
ary part ofw 0 (jco).] Our objective, as before, is to choose w 0 (*) in such
a way that the function (33) is the best possible approximation, in the
mean least-squares sense, to the function (24). However, in view of the
orthogonality property (28), we have that (33) is the best possible approx-
imation to (24), provided
35 yx a = d\
36 y2 a — cx
The conditions can be satisfied if we choose the operator W 0 as
37 (Wa x)(t) = ^x(t) + ^x(t)

Such an operator W 0 is not, strictly speaking, within the class of opera-
tors satisfying (5). However, we choose to overlook this fact, and for the
purposes of sinusoidal steady-state analysis at the frequency co, we
treat W
0 as a complex gain of value (dfa) j{c 1 ja). The equations (35) +
and (36) once again state that in order to have the best possible linear
time-invariant approximation to the nonlinear element N, the first
i 1 .
harmonic of the nonlinear element output (Nx 0 )(-) should be exactly

matched by the linear element output (W0 x <>)(•)
38 [definition] Given a nonlinearity N satisfying (19), the complex

number
39 r,(a;o,) = [+j^
is called the equivalent gain of the nonlinearity N; the function

a rj(a; co) is called the describing function of N.
Remarks : At this point, we have designated rj as r\{a ; co) because,

in general, the describing function may depend on both the ampli-
tude and frequency of the reference input.
40 fact If the nonlinear characteristic «(•) is memoryless and time-invariant,

then tj(a;co) is independent of co.
proof Consider two reference inputs
41 Xq1 \t) = a sin (O x t
42 2)
*o (0 = a sin co 2 t
both of which have the same amplitude but different frequencies. If both
co and co 2 are nonzero, we can write
43 *Sf
,
(0 = 4 (^)
1,
i.e., can be obtained from Xol) (') by time

Jto
2)
(-) scaling. Because N is
memoryless, we have
44 (Nxrm =
Hence, if
45 (Nx 0v )(t)
(
= c0 + S cos iCO\t + dt sin iCO x t)
i=i
then
46 (Nx 02) )(0(

= c0 S
+ i=i cos 1^2 + di sin iC0 2 t)
This shows that
47 ll(a;C0 1 ) =t](a;C0 2 )
Because co x and co 2 are arbitrary, (47) shows that rj(a; co) is independent
of co. M
48 fact If «(•) is odd in addition to being memoryless and time-invariant,
then t](a) is a real number.
proof If N satisfies (25) and x 0 (-) is of the form (23), then (Nx 0 )(-) is
of the form (26). This shows that 7](a) is real.

—
49 Example. Consider the sign nonlinearity n 1 ( • ) shown in Fig. 4.1.
If we apply an input x 0 («) of the form (23) to this nonlinearity, the
resulting output is a square wave of amplitude 1, regardless of what a
FIG. 4.1
is (as long as a ^
0). The first harmonic of a square wave of amplitude
1 has an amplitude of 4/n, so that the describing function of this non-

linearity 77i(-) is given by
50 tjiia) = —
na
51 Example. Consider an element n( • ) which is piecewise-linear as ,
shown in Fig. 4.2. For a | |

< S, «(•) acts like a linear gain of value m u
whereas for a | |
> d, n(-) acts (for small perturbations in a) like a linear
gain of value m 2 . Therefore, if we apply a sinusoidal input of amplitude

a and
,
if a < 8, then the output is another sinusoid of amplitude m x
a,
so that
52 =m ri(a) 1
if 0 <a<8
However, if a > <5, the output of the nonlinearity is of the form shown
in Fig. 4.3. In this case, it can be verified through laborious but straight-
forward calculations that
Figure 4.2 depicts the nonlinearity n{>) in the case where m >m
1 2.
However, expression (53) for tj{a) is valid for any m u m 2 (See also .
Problem 4.5.) Thus (52) and (53) completely characterize the describing
function ;/(•)•
By selecting various specific values for m

we can now
x
and m2 ,
derive the expression for the describing functions of several commonly

encountered forms of nonlinearities. For example, if we let m = 0, we x
get the dead-zone nonlinearity w 2 (-) shown in Fig. 4.4. The corre-
sponding describing function tj 2 (>) is obtained from (52) and (53) as
0 <a<8
54 Ô) = Y'l
a
+ a \ a 2) J
„>
Similarly, if we let m 2 = 0, we get the limiter nonlinearity n 3 (>) shown
in Fig. 4.5. The describing function of n 3 {>) is
0 <a<3
55 i2
[
—m
n
1
1
_
•
sm
— i
1
<5
a
.
b
5
a \
i
<5
a 2)
2
\
1/2
1
a > 8
There is one function that occurs repeatedly in the above expres-

sions, namely,
for > x 1
56 /(*) A |-^-[sin 1
x + x(l - x 2 1/2
for 0 < x < 1
) ]
A sketch of f(x) is found in Fig. 4.6. With the above definition, tj 2 (a)
can be expressed compactly as
57 rj 2 (a) =m 2
while t] 3 (a) can be written as
58 rh(a) = m,/^j
While (54) and (57) give the precise expressions for q 2 (a), one can
readily derive the approximate shape of t] 2 (a) just by using common
sense. If the input amplitude a is less than <5, the output is zero, so that
rj 2 (a) is also zero. On the other hand, as a becomes larger, the effect of
the dead zone becomes smaller, so that fi 2 {a) is a monotonically in-
creasing function of a. As a becomes extremely large compared with 8 ,
the effect of the dead zone all but disappears, and n 2 (’) begins to look
like a linear element of gain m2 . Hence t] 2 (a) —m >
2 as a — * oo. The
characteristic shape of t] 3 (a) can be rationalized in the same way.
We shall close this subsection by deriving some bounds on the
describing function. Suppose the function «(•) in (19) satisfies a condi-
tion of the form
59 k x cr
2
< on(<j) < k a 2
2
, V o e R
for some real numbers k x and k 2 (We say that . n(>) lies in the sector
[k 1? k 2 ].) If, in addition, «(•) also satisfies (25) [i.e., «(•) is odd], we show
below that
k i
< 77 (a) <k 2, \/ a e R
In other words, if «(•) lies in the sector [k u k 2 ], its describing function
also lies between the limits k and k 2
{
.
This sector condition (59) has a simple graphical interpretation,

as shown in Fig. 4.7. It means that the graph of the nonlinearity lies
between two straight lines, having slopes k and k 2 respectively. Thus x ,
the above result states that if the graph of a nonlinear element lies
between two straight lines of slopes k and k 2 its describing function x ,
takes values only between k and k 2 1

.
The bound (60) is rather easy to prove. Because /?(•) is odd, 77 (a)
is real. Moreover,
1n!<o
=—
f*
n{a) n(a sin cot) sin cot dt

na Jo
=
na\
nâ S^ n ^ S^ n^ ^ (letting 9 = cot )
>
~— f k (a x sin 0)
2
dO [by (59)]
Jo
=k x
The upper bound in (60) is proved similarly.

For a generalization of this result, see Problem 4.6.
4.1.3 Existence of Periodic Solutions
In this section, we shall present one of the main applications of

describing functions, namely, predicting the existence of limit cycles
mind that the analysis based

(periodic solutions). It should be borne in
on describing functions is only approximate. For example, there are
instances where the describing function analysis predicts the existence
of periodic solutions, but the actual system does not exhibit any, and
other instances where the situation is reversed. Hence it is perhaps more
accurate to say that describing function analysis predicts the likelihood
of limit cycles. However, in spite of these shortcomings, describing
functions are widely used in practice, because of the ease of analyzing
nonlinear systems (albeit approximately) by using them.
Consider the nonlinear feedback system shown in Fig. 4.8. The
objective is to determine whether, in the absence of an input, it is pos-
FIG. 4.8
sible to have a nonzero periodic solution for x(-). If the two blocks
“N” and “G” are described by the input-output relationships
61 yit) = (Nx)(t) = n[x(t)\

62 z(t) = ( Gy)(t)
= f g(t - x)y(x) dx
Jq
then, with zero input, the closed-loop system is described by the in-
tegral equation
63 x(t) = — (GNx)(t) = — f g(t — x)n[x(x)] dx

Jo
The problem is to determine whether (63) has any nonzero solutions

for x(-).
We approach the problem by assuming that (63) has a periodic
solution of the form
64 x(t) = a sin cot
where a and to are to be determined. As pointed out in Sec. 4.1.2,
whenever *(•) is °f the form (64), (Nx)(>) is also periodic with period
2njco, and the fundamental harmonic of (Nx)(>) is
65 y x a sin cot + c ° s cot
where
66 tl(a;co) =y l +jy 2
is the describing function of «(-)• Next, let
61 i(jco) = g (co)
r +Mi(co)
be the Fourier transform of the impulse response g(>). Because (62)

describes a linear time-invariant operator, and because >>(•) * s periodic
with period 2n/co, it follows that in the steady state z(-) is also periodic
with period Injco and that the fundamental harmonic of z(*) is
68 ua sin cot + va cos cot

where
69 u + jv A g(jco)ij(a; co)
Next, of (63) to hold, the first harmonics of both sides of (3)
must be equal. Now, if x(-) is of the form (64), equating the first har-
monics of both sides of (63) gives
70 a sin cot = —(ua sin cot + va cos cot)

or, equivalently,
71 1 + u + jv = 0
This procedure is also referred to by some authors as the principle of
harmonic balance. Note that it is a very approximate procedure because
even if (63) has a periodic solution, it is highly unlikely that this periodic
solution a pure sinusoid of the form (64). However, the rationale is
is
that (64) is a good approximation to this periodic solution. (These

points are discussed in greater detail later on.)
The essence of the method for predicting the existence of periodic
solutions using describing functions can now be stated: Given a system
description of the form (63), formulate the harmonic balance equation
(71), and solve for a and co (note that u and v are functions of a and co).
Now, if a 0 and co 0 satisfy (71), i.e., if
72 1 + i(jco 0 )tj(a 0 ;co 0 ) =0

then we consider it likely that (63) has a periodic solution with an
amplitude of approximately a 0 and a frequency of approximately co 0 .
The heuristic justification given for this technique is that if (72)

holds, then x 0 (*) given by
9
73 x 0 (t) =a 0 sin co 0 t
is an exact solution of the equation
74 x (t) = -(PGNx)(t)
where P is the operator that associates, with each periodic function, its
first harmonic. Of course, (74) is not the same equation as (63), because
there is no operator P in (63). However, if the operator G is a low-pass
filter , i.e,. if G rapidly attenuates the higher harmonics of (Nx)(-), then
(so the reasoning goes) there is not much error in replacing (63) by
(74), and of course x 0 (*) solves (74).
In practice, the describing function technique for predicting

periodic solutions works well if (as stated above) \g(j^>)\ decreases
rapidly as co — > oo . It is possible to place the describing function tech-
nique on a firm mathematical foundation (see [5] Bergen and Franks);
however, this involves topics in topology that are well beyond the scope
of this book.
Special Case
In the case of odd nonlinearities n(>) that satisfy
75 on(c) >0 , V o e R
i.e., in the case of the so-called first and third quadrant nonlinearities , the
harmonic balance equation (72) is particularly easy to solve. In this
case, the describing function rj is real, nonnegative and is independent of
co. Thus (72) reduces to
76 1 + g (m)tl(a) + jg,(co) = 0
r
Separating (76) into real and imaginary parts gives
77 + g (co)ri(a) = 0
1 r
78 giico) = 0
Now, (78) shows that must be real.

if (a 0 , co 0 ) solves (76), then g(jco 0 )
Further, because //(«) 0 \/ a, (77) shows must >in fact be that g(jco 0 )
real and negative. Thus we first make a plot of Re g(jco) vs. Im g(jco) and
determine the frequencies at which g(jco) is real and negative. Suppose
there are m such frequencies, co u . . . , co m . At each such frequency co h
(77) can be solved to give
79 rj(a) = .
~ x
' g(jCOi)
Let a i \ (
. . . ,
a i] be the solutions of
{
(79), if any. Then we predict that
the system (63) has periodic solutions at frequencies co u ,
co m , with
amplitudes a[°, a fi corresponding to the frequency c»
. .
. ,
(
t
..
{
Two things are to be noted about the above procedure. (1) The
possible frequencies of oscillation co 1 , ... ,co m do not depend on the
nonlinearity «(•) in any way; rather, they are determined solely by the
transfer function g(-). (2) Once the possible frequencies of oscillation
co u .co m are determined, the possible amplitudes corresponding to
. . ,
each frequency are easily found from (79). Thus, in the case of odd non-
linearities, the harmonic balance equation (72) is effectively decoupled.
80 Example. Consider the system of Fig. 4.8, with
§(s) = s(s
+ lX-y + 2)
«(f) =£ 3
Then it is easy to verify that
and tj{a) is independent of co. Thus the harmonic balance equation (11)
becomes, in this case,
i 1 3a
©I i
^ jco(l+jco)(2+jco) 4
If we separate (81) into its real and imaginary parts, we get
+ gr(co)fj(a) =
1 0
82 g {co) = 0
t
can be solved for

First, (82) co in an entirely straightforward manner,
which gives
co = */2
Note that we ignore the solution co = —*f~2 because, in physical
terms, it is no different from the solution co = */2 Now, . substituting
co = into (77) gives [because gr (*/2) = —1/6]
i.e., a = js/TT. Hence, for this system, we expect to find a periodic

solution with an amplitude of approximately and an angular
velocity of approximately +J~2 radians per second.
Problem 4.1. Given two operators Ah and Ah of the type studied in

Sec. 4.1.1, we define their sum Ah + Ah to be the operator defined by
84 [(N, + N2 )x](t) = (. NtxYf ) + (N2 x)(t)
Let x 0 (-) be a reference input. Show that if Wi(-) isan optimal quasilineariza-
tion of Ah with respect to the reference input x 0 (*) and h> 2 (0 is an optimal
quasilinearization of N 2 with respect to x 0 (*)> then h>i(-) + w 2 (-) is an opti-

mal quasilinearization of Ah + N2 with respect to x 0 (*)-
Problem 4.2. Specialize the result of Problem 4.1 to describing func-

tions: Let n u n 2 \ R —>R be two given functions, and define their sum
( nl + n 2 ): R—>R by
85 («! + n 2 )(a) = nâ) + n 2 (tf)
Show that the describing function of n x + n2 is the sum of the describing

functions of n 1 and of n 2 .
Problem 4.3. Verify that the nonlinearity n(-) in Fig. 4.2 is the sum of
the nonlinearities « 2 (-) and az 3 (-) (in Fig. 4.4 and 4.5, respectively). Verify that
the describing function tj(>) is the sum of // 2 (-) and tJsO).
Problem 4.4. Using the results of Problems 4.2 and 4.3, derive an
expression for the describing function of the dead-zone limiter shown in
Fig. 4.9. ( Answer : tj 4 (a) = m[f(S 2 /a) — /(St/a)].)
Problem 4.5. Using the results of Problems 4.2 and 4.3, derive an
expression for the describing function of the piecewise-linear element shown
in Fig. 4.10. [Answer:rj 5 (a) (m 1 2 )f{d 1 la) ( =
2
—m
m^)f(S 2 la) 3 .] + m — +m
Generalize the answers to a piecewise-linear element with / different slopes.
Problem 4.6. Let «(•), «/(•), and «„(•) be odd functions such that
86 oni(o) < on(o) < on (o u ), V o e R

and let tj(-), denote the describing functions of n(-), «/(•), and
n u { •), respectively. Show that
87 rji(a) < tj(a) < t] u (a) 9 V a >0

Problem 4.7. Consider the system (63) with
88 g(s) = 5(3 + 1)
s 2 (s + 2) 2
Analyze the possible existence of periodic solutions in the cases where
(a) «(•) is the sign nonlinearity of Example (49),
(b) «(•) is the dead-zone nonlinearity of Example (51), with 5 = 1 and
m2 = 2, and
(c) n(-) is the five-segment nonlinearity of Fig. 4.10, with m = 2,
x
m2 = 1, m = 3 0, 5 1 = 1, and S 2 = 2.
Sec. 4.2 Numerical Solution Techniques 113
4.2
NUMERICAL SOLUTION TECHNIQUES
In this section, we shall discuss some methods for approximately solving
a given set of nonlinear vector differential equations. The techniques
presented here are particularly well suited for implementation on a
digital computer. Due to limitations of space, we shall discuss here only
the rationale of the various methods and omit the more detailed numer-
ical analytic aspects, e.g., an analysis of the errors created by the appro-
ximations. For an in-depth discussion of these points, see [8] Chua and
Lin.
Throughout this section, we are concerned with finding (approxi-
mately) the solution to the nonlinear vector differential equation
1 ±(t) =f [t, x(0], t > 0; x(0) = xj (given)
We assume that the function f{t •) is globally Lipschitz-continuous,

,
which assures (see Theorem [3.4(25)]) that (1) has a unique solution over
[0, oo). Let x*(?) denote the actual solution of (1) and let (?„)f be a
monotonically increasing sequence of real numbers with t n approaching
oo as n — > oo. Our objective is to find various expressions that approxi-
mate x*(t„) n 9
= 1,2,... . That is, instead of attempting to approximate
the function x*(f), which is defined for all values of t > 0, we attempt
to approximate the sequence of vectors [x*(f n )]“. This is in keeping with
themodern trend toward the increased use of digital computers. In
many instances, the numbers t n are equally spaced i.e., ,
2 tn = nh
- for some number h > 0, which is referred to as the step size.
4.2.1 Taylor Series Methods
As the name methods are based on

implies, the Taylor series
expanding x(«) in a Taylor about some point. Suppose we know
series
x*(^„) for some n\ i.e., we know the exact solution at time t n Let us .
denote x*(/„) by x*, in the interests of brevity. Knowing x*, we can

compute x*+1 by using a Taylor series, as follows:
3 X*+ 1 = X* + M*0„) + yX*((„) + . . .
The various derivative terms in (3) can now be calculated using the
differential equation (1). For instance,
4 x*(0 = f(f^.xj)
5 X*(/„) =
=f x(t„, x*)-f (t„, x*) +f ,(t„, X?)
(where the subscripts denote partial differentiation), and so on.

Let us now suppose of having at hand the actual
that, instead
solution x*, we have an approximation to xj, which we denote
available
by x„. With the aid of this information, we can construct xn+1 which ,
forms an approximation to x*+1 by a formula parallel to (3), namely,

,
6 x„ +l = x„ + hi(t,„ x„) + ~[fx (t„, x„)-f(r„, x„) + x„)] + • •
We can now make a further approximation by truncating the infinite

series in (6) after a number of terms. The number of terms in (6)
finite
that are retained determine the order of the Taylor series method if —
the series in (6) is truncated after the term involving h k where k is some ,
integer, then the Taylor series method is said to be of Mi order, or of

order k. For instance, the Taylor series method of first order (also
known as the forward Euler method) is given by the approximating
formula
7 x„ +1 = x„ + hi{t„, x„)
while the second-order Taylor series method is described by

2
h
8 X„ +1 = X„ + hf(t„, X„) + y[fx (?„, X„).fO„, X„) +f ,(t„, X„)]
Note that to initiate the Taylor series method(s), we use the

initial condition in (1) and set x 0 xj. =
While the Taylor series methods are a useful starting point for
understanding more sophisticated methods, they are not of much
computational use. The first-order method (7) is too inaccurate, as it
corresponds to simple linear extrapolation, while higher-order methods

require the calculation of a lot of partial derivatives. Also, if we are
dealing with vector differential equations and we use anything higher
than a second-order method, then some of the partial derivatives
involved (e.g.,^ x ) leave the realm of matrices and become higher-order
tensors. Furthermore, the computation of all these derivatives becomes
unwieldy.
4.2.2 Runge-Kutta Methods
The aforementioned problems are alleviated by the Runge-Kutta

methods, which duplicate the accuracy of the Taylor series methods but
do not require the calculation of partial derivatives. To bring out the

main idea of the Runge-Kutta methods, we begin by studying the
second-order Runge-Kutta method, which employs the approximation

formula
9 X„ +1 = x„ + /*{>! f(t„, x„) +a 2 f[t„ + ah, x„ + /?f(t„, x„)]}

In (9), the real numbers a l9 a 2 , a, and /} are to be chosen in such a way
that the right-hand side of (9) approximates the right-hand side of (8)
with an error of o(h 2 ). If a l9 a 2 , a, and p can be so chosen, then because
the right-hand side of (8) approximates the right-hand side of (6) with
an error of o (h 2 ), we have that formula (9) accurately reproduces the
first three terms of the Taylor series (6). This does not mean that for-
mulas (8) and (9) produce identical results. Rather, what it means is
that by employing the formula (9) instead of (6) we get an error of
o(h 2 ), just as if we had employed (8) instead of (6).
We turn now to the method of selecting a a 2 a, and p. Expand- l9 ,
+ aK x„ + pi\t x„)] in a Taylor series about (tn x„), we get

ing f [t n n9 ,
10 f[t n + ah, x„ + fii{t„, x = f x„) + ahf,(t„, x„)

n )] (t„,
+ fitx x„)-f x„) + o (h 2 ) (t n , (t„,
Hence
11 h{a t f(t„, x„) +a
+ ah, x„ + )Sf
2 f[tn x„)]}
= («i + a 2 )hf(t x„) + a ah n, x„) 2

2
f,(t n ,
+ a fihfjj„, x„).f() x„) + o (h )

2
2
Comparing the right-hand sides of (8) and (11), we see that they differ
only by o(h 2 ), provided
12 a + a2 —
l
1
13 a2 a = -j
14 a 2 fi =A
Because (12)— (14) represent three equations in four unknowns, we can
assign onenumber arbitrarily. Thus (12)—(14) can be rearranged as
15 = — 1 a2
16 a = J_
2a2
= h
17 fi
2a 7
provided a 2 ^
0. Therefore the general second-order Runge-Kutta
algorithm is described by
18 = x, + h {(1
n a 2 )f(t„, x„) + a 2 f \n + _A_
i
’
xn .
M (t n , x„y
2a 2
'
la 2
As a special case, by setting az = we get the so-called Hem's or

modified trapezoidal algorithm, which is described by
X„ +1 = x„ + h
{
f(
y n)
+ + h,\„ + hf(t„, x„)]J
Similarly, by setting a2 = 1, we get the modified Euler-Cauchy algori-
thm, which is described by
h_ hf(t n9 x„)
M
'
= x„ L n\
,
2 ’
Y" ,
'
It should be once again emphasized that formulas (19) and (20) do not
give identical results.However, if we start from the same x„, the x„ +1
2
given by (19) and that given by (20) would differ only by o (h ).
The most commonly used of the Runge-Kutta algorithms is one
3
of order 4, which is described by
x„ +1 = x„ + /zg 4 (*„, x„, h)
where
x„, h) = go + 2gi + + g3
g 4 (t„,
go = X„)
g, = r( + A, X. + If)
r,
gl = f(l. + X„ + If)
g3 = f + h,x + hg
(t n n 2)
Note that g 0 , g3,

are computed sequentially, beginning with g 0
. . .
.
Because the above algorithm is of order 4, it faithfully reproduces the

first five terms (including the constant term) in the Taylor
series expan-
sion for x„ +1 However,

. it is computationally not very efficient, because
for each value of n ,
we need to calculate four vectors g 0 ,
• • • > g 3 which
>
are never used again. In contrast, the multistep algorithms, to be dis-

cussed in the next subsection, do make use of previously computed
vectors.
Example. Consider the linear matrix differential equation
M(t) = AM(0; M(0) =I

where A e R” xb
is a given matrix. It is well known (see Appendix II)
3 This is only the most common of an infinite family of fourth-order Runge-

Kutta algorithms.
! A
that the solution of (28) is
29 M(0 = e Kt =2A 1
'
We shall now show that when applied to the differential equation (28)
the fourth-order Runge-Kutta algorithm exactly reproduces the first
five terms of the Taylor series (29).

From (29), we have
30 M(A) 2A'i
= i=o Z
Applying (2 1 )— (26) with f(t, x) = Ax, tN — 0, and x v = I gives

31 G =A 0
• I =A
32 g, = a(i + aA) = a + a*A
33 g = a[i +
2 (a + A A) AJ = A + A
2 .
2 A+A A 3
34 g = 3 a[i + (a +A 2 A+A 3
= A + A2h + A 3 A+a A 4
35 g4 = 2G t + 2G 2 +G 3
^ i
_ ^2 h h2 ,
/z
3
6 2 6 24
36 Xjv+i = I + AA + A 2 A + A A + A*A 3
Thus, clearly, contains the first five terms in the Taylor series (30).
Incidentally, this example illustrates that the Runge-Kutta algori-
thm is applicable not only to vector differential equations but also to
matrix differential equations.
In the practical implementation of the Runge-Kutta algorithms,

the following procedure sometimes works well. Given tn and x„ = x(t„),
1. Pick a step size h u and let 4+i — tn + hx .
2. Compute xî using a Runge-Kutta algorithm of the desired

order.
3. Halve the step size h x
by selecting h 2 = hJ2, = tn + /z
2,
= t„ + 2h 2 (=
tjS 2 Compute x<*\ and x£? 2 using the
same Runge-Kutta algorithm as in step 2.
4. Compare x^
2 with x^+V If their difference is less than some
preselected error bound, accept the value x^ and proceed to
step 1 ;
otherwise, repeat step 3.
1
4.2.3 Multistep Algorithms
The basic idea behind multistep algorithms is to approximate the

solution of a given differential equation (1) by a polynomial in t. This is
justified on the basis of the Weierstrass approximation theorem, which
states that any continuous function over a finite interval can be uni-
formly approximated to any desired degree of accuracy by a polynomial
of appropriate degree.
Specifically, given the differential equation (1) and the uniformly
spaced time intervals tn given by (2), we attempt to approximate x„ +1
by an expression of the type
k k
37 x„ +1 =E
= 1 0
ajXn-t E
+ /=- b,hfQn-i,^n-i)
where k is an integer and a h b are {

real numbers to be selected later.
Note that if b_ x ^ 0,
then (37) gives only an implicit formula for the
unknown quantity x„ +1 , because it appears on both sides of the equa-
tion. On the other hand, if b_ x = 0, then (37) gives an explicit formula
for x„ +1 . An algorithm of the type (37) is called a (k -f 1 )-step algorithm,
because the value of x at the (n + l)st step is given in terms of the values
of x at the previous k 1 steps (namely, x„_^, + x„). In a sense, one . . . ,
can think of formula (37) as arising from discretizing the expression
38 x(r) = + f f [t, x(t)] dx

Jo
The order of algorithm (37) is the degree of the highest-degree

polynomial for which (37) gives an exact expression for x„ +1 To see how .
this definition of order is used, we shall consider what is meant by a

second-order algorithm. Suppose 4 x(«) is a polynomial of degree 2,
i.e., that
39 x(t) = c0 + cf + c2t 2
Then
40 xn = Cq + c t nh +c 2n
2
h2
Also, (39) implies that
41 x{t) = c1 + 2c 2t
so that
42 f(t n x„),
= c + lcx 2 nh
4 To avoid notational clutter, we let ^ be a scalar. However, all arguments

remain valid in the case of vector x.
^ 1 ,
Substituting from (40) and (42) into (37) and choosing n 0 for con- =
venience, we obtain what is known as the exactness constraint for the
algorithm (37), namely,
43 Cq + cxh + c2h2 = X) a {c Q t + c t (—ih)

2
+c 2 (— h) 2 ]
i =0
X
+ i=- 1
hbi[c t + 2c 2 (—ih)]
Because we would like (43) to hold independently of the step size h, we

equate the coefficients of like powers of h. This gives
k
44 c0 =X
= 1 0
a tc 0
k k
45 Cl = XI —ia c i 1 + =X-l bfil
i =0 i
k k
46 c2 = i
1]
=0
i
2
GiC 2 + =2-1 —2ibiC 2
I
Once again, because we would like (43) to hold for all second-order
polynomials, (44)-(46) must hold for all c 0 , c l9 c 2 . This implies that
47 2= a t
= 1
1 0
48 XI — iî + XI b t
= \
i=-l i =-
49 X i2a + X- i
=
—216, = 1
1 =0 f 1
Thus on the 2k
(47)-(49) give three constraints 3 constants a 0 + ,
a k b_ u
,
b k and any choice of these constants such that (47)-(49)
. . . , ,
hold makes the corresponding algorithm (37) exact for second-degree

polynomials. In other words, whenever (47)-(49) hold, the correspond-
ing algorithm (37) is of second order.
It is clear that in order for (47)-(49) to hold we must have 2k + 3
> The second-order algorithm containing
3. the smallest number of
constants a v b is obtained by setting 2k + 3
,
t
= 3, i.e., k = 0. In this
case, (47)-(49) take the form
50 a0 = l
51 +b = 0 1
52 2b _ = x
1
which can be solved to give
53 ao = I? 1
=— , b2 =—
o io , ,
Hence a second-order numerical integration algorithm is given by the

formula
c. — x n+1 )
34 Xn + 1
Xn i“
I
Algorithm (54) is also known as the trapezoidal algorithm.

The derivation of the exactness constraints in the general case
proceeds in a manner entirely analogous to the above. Suppose x(») is
a polynomial of degree ra, i.e.,
55 x(t) =2 CjV
= j
Then
m
56 x(t) = 2 jc j
tJ l
= j
Therefore,
57 Xi = 2 c ji J h j
= j
m
58 f(t„ x,) = j=£ jcjp- 'h'~' 1
Substituting from (57) and (58) into (37) and taking n =0 for con-
venience, we obtain
59
m
£ Cjti = £=
km£ at Cj(—ihy
j= 0 1 0 7 =0
+ —£~l i
bt £= jcj(—ihy-
7 1
l
h
Equating the coefficients of h j on both sides and dividing the resulting

equation by Cj, we obtain
60 1 =2 tf/(— 0 J + S bJi—iy- 1
, , j = 1, . .
.
m
= i 0 = -i »
k
61 1 =2 ai (corresponding to j = 0)
i= 0
Note that in applying (60) we must take V = 1 when i = j = 0.

The set of equations (60)— (61) provide m + 1 equations con-
straining the 2k + 3 constants a Q a k 6_ l9 b k Thus, clearly, , . .
. , , . .
.
.
in order for algorithm (37) to be exact in the case of rath-degree poly-

nomials, i.e., for (37) to be of order ra, we must have
62 ra + < 2k +1 3
If equality holds in (62), then we can solve for the a-s and b- s uniquely.
Now, if ra is odd, one can always choose k such that
63 2k ~f~ 3 = ra -j- 2
,
and then assign = 0, so that (37) becomes an explicit formula for

*»+!•
By way of illustrating the above, suppose we want to derive a
third-order algorithm. Using (63), we choose k = 1 and b_ =0. The
x
corresponding equations (60) and (61) can be solved to yield

64 a0 = —4, a x
= 5, b0 = 4, bx = 2
Hence a possible explicit third-order algorithm is given by
65 x„ +1 = ~4x„ + 5x„_! + 4hf(t„, xn ) + 2hf(t„_ u x„_ t )
However, this algorithm is not very widely used.
Adams-Bashforth Algorithms
The Adams-Bashforth algorithm of order m is an explicit algori-

thm obtained by choosing
66 k =m— 1
67 a t
=0 for i = 1, . .
.
k
68 =0
Thus there are exactly k +2 constants to be determined, namely a 0 ,
b Q ... ,b k Clearly, (67) and (61) imply that

, .
69 #0 = 1
Thus, once we solve the k + 1 equations provided by (60), the algo-

rithm is determined. The formulas corresponding to the Adams-Bash-
forth algorithms of orders 1 through 4 are listed below:
70 x„ +1 = x„ + hi(t x„) n, (m = 1)
71 x„ +1 = x„ + h[jf(t„, x„) - if((„.!, X _ n j)] (m = 2)
72 x„ +I = x„ + x„) - f|f (?„_!, x„_,) + x„_ 2 )]
(m = 3)
73 X„ +I = X„ + X„) - XB -l) + 24 f(*«- 2 > x„- 2)
- x„_ 3 )] (/M = 4)
Adams-Mou/ton Algorithms
The Adams-Moulton algorithm of order m is an implicit algori-

thm obtained by choosing
74 k =m—2
75 at = 0, = i 1, . . . , k
As before, (75) and (61) together imply that
76 a0 = 1
, ,
The remaining unknowns b_ u .,b k can be determined using (60). . .
The formulas corresponding to the Adams-Moulton algorithms of

orders 1 though 4 are listed below:
77 X„ + , = x„ + hf(t„ +1 x„ +1 , ) (m = 1)
78 x„ + 1
= x„ + h[{f(t n+i x„ +1 ) + , x„)] (m = 2)
79 x„+i = x„ + h[^f(t„ +1 x„ +1 ) + , (t„, x„) - Tzf(t„-i, X„_,)]
(m = 3)
80 x„ + i = x„ + h[^f(t„ +1 x„ +1 ) , + x„) -
+ X„- 2 )] (m = 4)
Predictor-Corrector Algorithms
From the preceding discussion, we can see that a £-step Adams-

Bashforth algorithm is exact for polynomials of order k while a
,
A>step
Adams-Moulton algorithm is exact for polynomials of order k + 1.
However, the Adams-Moulton algorithm is implicit, in the sense that
knowledge of xn _ k9 x„ gives an implicit formula only for x„ +1 This
. .
. .
implicit formula can be symbolized as
81 x„ + 1 = y„ + x„ +I )
where y„ is a quantity that depends only on x„_ x„ and is hence fc ,

. .
.
known. Now, (81) can be solved iteratively as follows: Choose an initial

guess xn + l9 and set
{
82 x*?,” = y„ + hb-ii(t n+1 , x<«j), i >0

The iterations are terminated when xH+P |
— x%h |
is sufficiently small.
questions arise in connection with this procedure: (1) How
Two
does one choose the initial guess ? (2) Under what conditions does x^
the sequence of iterations generated by (82) converge? Tackling the
second question first, suppose f satisfies a global Lipschitz condition;
i.e., suppose there exists a finite constant M such that
83 ||f(?, x) — f(t, y)||<M||x —y ||, V t, x, y
Then, for sufficiently small values of h (in fact, whenever hb_ x
M< 1),
the right-hand side of (81) defines a contraction mapping of x B+1 Hence, .
by Theorem [3.3(1)], the sequence (x^lô converges to the unique solu-

tion of (81), regardless of the starting value x^V However, it is clear
that the convergence of (x^j)7 is accelerated if we choose x^\ to be
reasonably close to the solution of (81).
Accordingly, we choose x^\ to be the value generated by an
explicit k- step algorithm and then apply the iterative formula (82). This
2
is known as the predictor-corrector method. For example, a two-step

predictor-corrector algorithm can be defined as follows: Given x„,
x„_i, set
84 x<°+>, = x„ + h[\f(t„, x„) - n . u x„ - (t ! )] (predictor)
85 x<'+ 1
,)
= x„ + X' + + x„)] i) (corrector)
In the foregoing, (84) predicts the value of x n+1 using an explicit for-
mula, while (85) corrects this predicted value of x n+1 In practice, if the .
step size h is selected properly, relatively few applications of the correct-

ing formula (85) are enough to determine x n+1 to a high degree of
accuracy.
We shall conclude this section with a comparison of Runge-
Kutta and multistep algorithms. The main advantage of a multistep
algorithm is that there are no extraneous calculations, as in Runge-
Kutta methods. However, to apply a multistep algorithm, it is essential
for the time instances tn to be uniformly spaced. This is not necessary in
Runge-Kutta methods. Based on these considerations, it can be said
that Runge-Kutta methods are better if one expects the solution x(*)
of a given equation to vary very rapidly at some times and not to at
other times (as depicted in Fig. 4. 1 1). In such cases, it is possible to keep
adjusting the step size h as one goes along in order to always maintain
high accuracy. Otherwise, multistep algorithms are preferable.
/
/
/
x
/
/
t
/
I i
iii i i
>
FIG. 4.11 f
0 t\ 1 ^3 *4 t t t
5 6
The discussion in this section barely scratches the surface of a

very well-developed subject. For a more detailed treatment, see [19]
Ralston, and for an engineering viewpoint, see [8] Chua and Lin.
Problem 4.8. Show that in the case of a linear differential equation of

the form
4(0 = Ax(0 + y(0

the results given by a Taylor’s series method of order n and a Runge-Kutta

method of order n are identical.
Problem 4.9. Formulate the three-step and four-step predictor-cor-

rector algorithms.
43
SINGULAR PERTURBATIONS
In this section, we what is known as the problem of
shall briefly study
singular perturbations. The problem can be stated as follows: Suppose

we are given the system of nonlinear differential equations
xy) = f[x(0, y(01

ey(0 = g[x(0> y(01
where x(f) R", y (/) e
e ,
Rm
f: R" X > , Rm — Rn
and g: R" X -R Rm — m .
Note that for any value of e other than zero the system (l)-(2) consists
of n +m differential equations. However, if e =

0, the system (l)-(2)
consists of n differential equations and m algebraic equations, because

with 6 = 0, (2) reduces to
g[x(0, y(01 =0
Suppose it is possible to solve the m equations comprising (3) to obtain
an explicit expression for y (t) in terms of x(t), of the form
y(0 = h[x(/)]
where h: R" — Rm
* . Then (1) and (3) together reduce to
x(t) = f{x(/), b[x(/)]}

which a system of n differential equations.
is
Setting e —
0 in (2) is called a singular perturbation because it
completely changes the nature of (2), viz., from a differential equation
to an algebraic equation. The objective of the theory of singular pertur-
bations (stated in very simplified terms) is to examine the simplified
system (5) and from this to draw conclusions about the original system
(l)-(2) with 6^0.
Physically speaking, singular perturbations arise from an attempt
to approximate a high-order nonlinear system with another one of
5
lower order. The following example, due to [9] Desoer and Shensa,
illustrates this point.
Example. Consider the linear circuit shown in Fig. 4.12, and

suppose the capacitance of e represents a stray capacitance. In accor-
5 In terms of the number of differential equations describing it.

Sec. 4.3 Singular Perturbations 125
*1
1 H
FIG. 4.12
dance with standard practice, the inductor current x and the capacitor l
voltage x 2 are chosen as the state variables of the system. Using the
current and voltage laws of Kirchhoff, the dynamic equations of the
circuit in Fig. 4.12 can be written as
If we wish to neglect the stray capacitance, we set e = 0 in (8), which

gives
0 = x, -*»
r
x2 = rx x
Hence the new equation for x x

is obtained from (7) as
Xi = (~ 1 — r)x x
Now, it is clear from (11) that, as long as r > —l, the solution x of t
(11) decays exponentially to zero. However, returning to the original,

unsimplified system (7)-(8), we have
"
*r ’-1 -1 Xl
X2 _ _l/e — l/er _x 2 _
Letting A denote the 2x2
matrix in (12), one can easily verify that
if r <0 ,
then for sufficiently small values of the parameter e the matrix
A has at least one eigenvalue with positive real part, so that the original
system (7)-(8) is unstable in this case. 6 In other words, if r e (—1, 0),
then the simplified system is stable, but the unsimplified system is
unstable for sufficiently small values of e.

This example not only shows that singular perturbations arise in
connection with simplifying the dynamical model of a given system but
also demonstrates the need to proceed with care in the simplifying pro-
cess. Sometimes the simplified model can be stable, while the original
system is unstable.
6 See Sec. 5.3 for a full discussion of the stability of linear systems.
2 : 6
Even though the general singular perturbation problem has been

posed here for nonlinear systems, we shall content ourselves with two
results for linear systems, because we have not as yet developed the tools
needed to rigorously analyze the stability of nonlinear systems. A dis-
cussion of the singular perturbation problem for nonlinear systems can
be found in Sec. 5.4.
We shall now state concisely the two problems which we shall
address next
13 1. Given the system of linear equations
X- fA„ ^12 r x >i

14
_f*2_ 1< ^22_ i X 1—
N>
find the conditions on A n A 12 A 21 A 22

, , ,
which ensure that
there exists a positive number e 0 such that all eigenvalues of the
17 matrix
Ai Al2
15 A(e) = i
_A 2 i/6 ^22 /£_

have negative real parts, for all e in (0, e 0 ).
16 2. Find conditions on A n A 12 A 21 A 22
, , ,
which ensure that there
exists a positive number e 0 such that at least one eigenvalue of
A(e) has a positive real part, for all e in (0, e 0 ).
An answer to problem (13) is provided by the following theorem.
Theorem Suppose that all eigenvalues of A 22 and of A n —

A 12 A 22 A 21
have negative real parts. Then there exists a positive
number e Q such that whenever e belongs to (0, 0 ) all eigenvalues of
A(e) have negative real parts.
The proof depends strongly on the well-known fact (see, e.g.,

[22] Wilkinson) that the eigenvalues of a matrix depend continuously
on the elements of the matrix. So as to avoid any ambiguities we
shall state this fact precisely.
18 fact Let M(e) be a continuous, n x n matrix-valued function of €. Let

{2 10 , . . . , 2„ 0 } denote the (not necessarily distinct) eigenvalues of M(0),
and let {A u , ... 9
k ne ] denote the (not necessarily distinct) eigenvalues of
M(e). Then, for each r > 0, there exists an e 0 > 0 such that the inequality
19 1 /0 — Xi€ \
< r whenever 0 < € < e0
is satisfied, by renumbering the X i€ if necessary. 7
7 Note that when we solve the equation det [XI — M(e)\ = 0, the resulting n
solutions for X can be numbered in any order we choose.
,
proof of Theorem (1 7) In what follows, we omit the phrase “by renum-

bering necessary” in the interests of brevity. For the sake of definiteness,
if
suppose n e RnXn A 12 e RnXm ,A 21 e R mXn A 22 e R m xm The n

A , , . +m
eigenvalues of A(e) are the n zeros of the polynomial +m
A n -2l A12
20 A(6;2) = det
_ A 21 /e A 22 /6 — 2,1
Let {a 1# . .
. ,
a m} denote the (not necessarily distinct) eigenvalues of A 22 ,
and let , /?„} denote the (not necessarily distinct) eigenvalues of

A n — A 12 A 22 A 21 .
8
We claim that (1) n of the eigenvalues of A(6) are
“close” (in a sense to be made precise next) to {ft u . .
. /?„} and (2) the
remaining m eigenvalues of A(6) are “close” to {aje, . .
. ,
a m /e}. Spec-
ifically, we claim that, given any r > 0, there exists an € 0 > 0 such that
the inequalities
21 P <r,t\ i = 1, . .
. , n
whenever
22 i =n+ 1, +m 0 < € < 60
hold, where {2i(e), . .

. ,
An+m (e)} are the eigenvalues of A(e).
The proof of this claim proceeds as follows : We have

6A n 6A 12
23 eA(e) = 4B(e)
A 21 A 22 j
It is clear that the matrix B(e) is continuous in 6 and that
' 0 ”
0
24 B(0) =
_A 2 a 22 _
i
Because the n + m eigenvalues of B(0) are {0, . .

. , 0, 0C U . .
. ,
a m }, we have
by fact (18) that, given any r > 0, there exists an € x > 0 such that n
eigenvalues of B(e) are “close” to 0 and that the remaining m eigenvalues
ofB(e) are close to [Ct l9 a m }. Because A is an eigenvalue of A(e) if and
. .
. ,
only if eA is an eigenvalue of B(e), we can conclude that m of the eigen-

values of A(e), say { An+1 2„ +m }, satisfy , . .
. ,
25 \eAi — < r, i =n+ 1, ...,«+ m, whenever 0 < € <

or, equivalently,
26 At
<Xi-n
€
< € ’
/ =«+ 1, ...,« + m, whenever 0 < € < €x
The remaining n eigenvalues Au . .

. , An satisfy
27 1 62/ 1
< r, i = n ,
whenever 0 < 6 < €1
even though this is not of much importance.
8 Note that the hypotheses ensure that A2 \ exists.

A , , ,
Next, for any 6 ^ 0, we have that A is a zero of the characteristic

polynomial A(e; X) if and only if
An Al A 12
28 det a Ai(e; X) = 0
eXl
Note that the only change between (20) and (28) is that the second “row”
of matrices has been multiplied by e, which is permissible because 6^0.
Now, A i (6; A) is continuous in € at € = 0, and
29 A ff\ \
Ai(0; X)
1
= A
det
TAl1 ~ ^
A2 i A 2 2_
Noting that the determinant of the matrix
I
0“
30
— 22
!
A2 i
Ij
is 1, we have
"
An - Al Ai 2
r
I OF
31 det
A2 | A 22 _ A 22 A 2 .ji i
i
Ai — Ai 2 A 22 A 2 1 i < i -
det
0 A 2 2_
det (An — Ai 2 A 22 A 21 — Al) det A 22
Because det A 22 is a nonzero constant, we see that Ai(0; X) is a polynomial

of degree n whose zeros are {j?i, /?„}. Hence for € 0, n of the zeros . . . , ^
of A i (e; X) [and hence those of A(6; A)] are “close” to {ft u /?„}; i.e., . .
.
given any r > 0, there exists an e 2 > 0 such that
32 \Xtj — ftj\ <r , j = 1 , . .

.
n, whenever 0 < € < e2
where {7 1 , . .
. /„} is a subset of {1, . .
. , n + m}.
Now, if we pick the numbers r, € 1 ,
and e 2 carelessly, it is possible for the
same X to t satisfy both (26) and (32). However, suppose we choose r as
_ —min; [Re Oj
33 r
2
and choose e 0 such that
34 j- > \fit\, i = and

to
35 fo < min [6i, € 2]
where and e 2 are as in (26) and (32), respectively. Then whenever

0 < e < €0 ,
we have that m eigenvalues of A(e) satisfy (26) and that the
remaining n eigenvalues of A(6) satisfy (32).
We can now complete the proof. Because (26) and (32) account for all
n + m eigenvalues of A(e), and because all these eigenvalues have negative
real parts, we conclude that all eigenvalues of A(e) have negative real parts
whenever 0 < e < e0 .
H
36 Remarks: The proof of Theorem (17) actually gives a geometrical

insight into what happens to the eigenvalues of A(e) as e > 0. As —
e — » 0, n eigenvalues of A(^) approach the eigenvalues of A n —
A n A 22 A 21 , while the remaining m eigenvalues of A(^) approach the
eigenvalues of A 22 divided by , e. This is the import of the inequalities
(21) and (22). This geometrical insight enables us to provide an
answer to problem (16), which is given next.
37 Theorem If at least one of the eigenvalues of A 22 or at least one of

the eigenvalues of A n — A 12 A 22 A 21 1
has a positive real part, then
there exists an e 0 > 0 such that at least one eigenvalue of A(e) has a
positive real part whenever 0 < e < e 0.
proof Immediate from (21), (22), and remarks (36). M

38 Remarks: The behavior of the RLC network considered in Exam-
ple (6) can now be explained in terms of Theorem (37). In that exam-
ple, the eigenvalue a lt —
a l2 a 2 la 2X was negative, but a 21 itself was
positive. Accordingly, as predicted by Theorem (37), the system is
unstable for sufficiently small values of e.
In [9], Desoer and Shensa actually extend the results of Theorems

(17) and (37) to the case of linear systems containing both large and
small parameters, i.e., to systems described by
They study the behavior of system (39) as e —> 0 and fi —> oo. In this
case, the second equation in (39) becomes an algebraic equation, while
the third equation reduces to
40 x3 =0
The results proved by Desoer and Shensa are as follows: Define
41 B = An A 12 A 22 A 21
and let C be the matrix resulting from simplifying the set of equations
0 = AjjXjl T A 12 x + A x-
2 13 3
42 0 = A 21 Xj “T A 22 x + A 23 x 2 3
= A Xj + A 32 x + A 33 x
jli± 3 31 2 3
into the form

43 ju± 3 = Cx 3
Then we have the following.

:
44 Theorem Suppose all eigenvalues of A 22 ,

B, and C have negative
real parts. Then there exist e 0 > 0 and < oo such that all eigen-
values of the matrix
Ah A 12 Aj 3
45 A(6 ;
ju) = A 2 i/ 6 A 22 /f A 23 /6
_A 3 i//j a 32 fii A 33 /^_
have negative real parts whenever 0 <e< e0 and ju > ft 0 .
46 Theorem Suppose at least one eigenvalue of A 22 ,

or of B, or of C
has a positive real part. Then there exist e 0 0 and jli 0 oo such > <
that at least one eigenvalue of A(^) has a positive real part whenever
0 <e< eQ and n > jlc
0 .
The proof is left as an exercise.

Even though the study in this section is restricted to linear sys-
tems, the results derived here form the basis for the results given in Sec.
5.4 for singular perturbations of nonlinear systems.
Problem 4.10. Analyze the stability of the following system using

Theorems (17) and (37):
~
~x n ‘-I 1 r
* 12 = 2 -5 l *12
JtXll 0 -1 — 3_ _*2 _
Problem 4.11. Analyze the stability of the following system using

Theorems (44) and (46)
“
*11
“
1 Oil 2“
~*i f
*12 1-4 0-1 0 *12
€X 2 = 0-2-1 1 0 *2
fix 31 201-51 *31
_/**32_ 0 0-1 2 — 3_ _*32_
Problem Prove Theorems (44) and (46). [Hint: Proceed as in the

4.12.
proof of Theorem (17), and show that the eigenvalues of A(e) can be parti-
tioned into three sets: (a) those that approach the eigenvalues of B; (b) those
that approach 1/6 times the eigenvalues of A 22 ; and (c) those that approach
\\H times the eigenvalues of C.]

5
%J Stability in the sense of Liapunov
In this chapter, we shall study the concept of Liapunov stability, which

plays an important role in modern control theory. We saw earlier that
if a system is initially in an equilibrium state, it remains in it thereafter.
Liapunov stability is concerned with the trajectories of a system when the
initial state is near an equilibrium point. From an engineering point of
view, this is very important because external disturbances (such as noise,
component errors, etc.) are always present in a real system. The three
basic concepts of Liapunov theory are stability, asymptotic stability,
and global asymptotic stability. Roughly speaking, stability corresponds
to the system trajectories depending continuously on the initial state;
asymptotic stability corresponds to trajectories that start sufficiently
close to an equilibrium point actually converging to the equilibrium
state as t —
> oo and global asymptotic stability corresponds to every
;
trajectory approaching a unique equilibrium point as t oo. —

In this chapter, we shall formulate precise definitions of the various
concepts of Liapunov stability, and we shall state the basic theorems
that enable one to determine the stability status of a given system.
5.1
BASIC DEFINITIONS
In this section, we shall state the precise definitions of the various
concepts of Liapunov stability; in addition, we shall introduce the vari-
ous kinds of positive definite functions.
131
]
132 Chap. 5 Stability in the Sense of Liapunov
5.1.1 Definitions of Stability
Consider the vector differential equation
1 x(0 =f [t, x(0], t >0

where x(0 e R" and f: R + Y R" > R Throughout this chapter, we
n — .
shall assume that the function f is of such a nature that (1) has a unique
solution over [0, oo) corresponding to each initial condition for x(0)
and that this solution depends continuously on x(0). This is the case,
for example, if f satisfies a global Lipschitz condition (see Sec. 3.4).
Recall that an x„ e R" is said to be an equilibrium point of the system
(1) at time t 0 if
2 f(t, x 0) = 0. V t > t0
Throughout this chapter, we shall assume that the vector 0 is an equi-

librium point of the system (1). This assumption does not result in any
loss of generality, because if x 0 is an equilibrium point of (1) at time
t0 , then 0 is an equilibrium point at time f„ of the system
3 Hf) = f,[f, z(/)]
where
4 f ft, z) =f (t, z -x 0)
Also, there is a one-to-one correspondence between the solutions of

(1) and of (3). Thus, without loss of generality, we assume that
5 f(f, 0) = 0, M t> to
As discussed in Sec. 1.2, this means that if the system (1) is started
off in the initial state x(t 0 ) =0 ,

the resulting trajectory is x(t) =0
y t> t 0 However, an important practical consideration
.
is the follow-
ing: Suppose the initial state x(t 0 ) is not 0 but is “close” to it; what is
the nature of the resulting trajectory? The various concepts of Liapunov

stability are addressed to this question.
[definition] The equilibrium point 0 at time t0 of (1) is said to

6
be stable at time t 0 if, for each e > 0, there exists a S(t 0 , e) >0
such that
7 ||x(f 0 )ll < 5(t 0 ,

e)=^||x(0|| < e, V t>t 0
It is said to be uniformly stable over [r 0 , oo) if, for each e > 0, there
exists > 0 such that
a 5(e)
8 l|x(ti)|| < S(e), >t 0 ==H|x(OII < V t>ti

[i.e., the same 5(e) applies for all f,].
9 [definition The equilibrium point 0 at time /„ is unstable if it is
not stable at time tQ .

Sec. 5. 1 Basic Definitions 133
According to definition (6), the equilibrium point 0 at time t0
is stable at time norm x(r) of the

t 0 if, given that we do not want the || ||
solution of (1) to exceed a prespecified positive number e we are then ,
able to determine an a priori bound S(t 0 e) on the norm of the initial ,
condition ||
x(r 0 ) ||
such that any solution trajectory of (1) starting at
time tQ from an initial state lying in the ball of radius 8(t 0 e) always ,
lies in the ball of radius e at all times t>t 0 . Another way of stating
this is as follows : Arbitrarily small perturbations (about the equilibrium
state 0) of the initial state x(£ 0 ) result in arbitrarily small perturbations
of the corresponding solution trajectories [of (1)].
To give yet another interpretation, let Cn [t Q , oo) denote the linear
space of continuous ^-vector- valued functions on [* 0 ,
oo), and let
BC”[t 0 oo) denote the subspace of
, C n
[t 0 , oo) consisting of bounded
continuous functions. If we define the norm
I|x(.)|| t = sup II *(011, V x(.) e BCn [t 0 ,
oo)
then BCn [t 0 , oo), together with the norm ||

• \\ s ,
actually becomes a
Banach mapping
space. Let us define a T tQ from R n
into Cn [t 0 , oo) in
the following manner Given any v e R n :
, let T to y
[which is an element of
Cn [t 0 , °°)] be the solution of (1) corresponding to the initial condition
x (/o) =v - In other words, Tt0 maps initial conditions into the corre-
sponding solution trajectories of (1). Now, (5) implies that T maps the
t0
element 0 of Rn into the zero function in Cn [t 0 ,

oo). According to
definition (6), the equilibrium point 0 at time t0 is stable at time tQ
if the following two conditions hold:

1. There is some constant c > 0 such that the image, under Tto ,
of the ball i? = {v:||v|| < c} is actually contained in

BCn [t 0 , oo).
2. The restricted map T B —> BCn [t 0

t0 :
,
oo) is continuous at
0 g R n
.
Another small point needs to be clarified. In (6), is any norm ||

•
||
on R Because all norms on R n are topologically equivalent (see fact

n
.
[3.1(52)]), one can see that the stability status of an equilibrium point
does not depend on the particular norm used to verify (6).
Once the notion of stability is understood, it is easy to understand
what uniform stability means. According to definition (6), the equilib-
rium point 0 is stable at time t 0 if, for each e > 0, a corresponding
8 can be found such that (6) holds. In general, this 8 depends on both
e and / 0 However, if a 8 can be found that depends only on e and not
.
on the initial time / 0 then the equilibrium point 0 is uniformly stable.

,
Finally, let us turn to a discussion of instability. According to

definition (9), instability is merely the absence of stability. It is unfor-
tunate that the term instability leads some to visualize a situation
where some trajectory of the system “blows up” in the sense that x(/) || ||
— > oo as
t —
> oo. While this is one way in which instability can occur,
by no means the only way. Stability of the equilibrium point 0

it is
means that, given any e > 0, one can always find a corresponding <5
such that (7) holds. Therefore, the equilibrium point 0 is unstable if,
for some e > 0, no 8 > 0 can be found such that (7) holds. Physically
speaking, the equilibrium point 0 is unstable if there is some ball B e
of radius e centered at 0 such that for every 8 0, no matter how small, >
there is a nonzero initial state x(/ 0 ) in Bs such that the trajectory starting
at x(f 0 ) eventually leaves Be This and only this is the definition of insta-
.
bility. It may be that some trajectories starting within Be actually

“blow up” with ||x(f)|| — » oo as t — » oo. However, this is not necessary
for instability. This point is illustrated in the following.
11 Example. Consider the Van der Pol oscillator, described by
12 X 1
= —X 2
13 x2 = —x + 1 (1 — x\)x 2
which is also studied in Chap. 2. The point x = x 2 = 0 is an equi- 1
librium point of this system, and as such, the solution trajectory starting
from Xj(0) =x 2 (0)
=0 is given by x (t) = x 2 (t)
t
0 = V t > 0. How-
ever, solution trajectories starting from any nonzero initial state all
approach the limit cycle, as shown in Fig. 5.1. Now, let us study the
stability status of the equilibrium point (0, 0) using definitions (6) and
(9). If we choose 6=1, say, then B = e {(x u x 2 ): (pc\ + xl) l/1
< 1} lies
entirely within the region enclosed by the limit cycle. Therefore all
trajectories starting from a nonzero initial state within B e eventually

leave B e.
Thus if we choose e = 1, no 8 > 0 can be found such that
(7) is satisfied. Accordingly, the equilibrium point (0, 0) is unstable.
> t
Note that all trajectories of the system, no matter what the initial state
of the system is, are bounded, and none “blows up,” so the system is
well behaved in this sense.

In definition (6), stability is defined at an instant of time t 09 and
uniform stability is defined over an interval However, if t 0

[f 0 ,
oo).
is clear from the context, we can just speak of stability and uniform
stability. Some authors assume t 0 = 0, on the rationale that if t 0 0, ^
one can always “translate” the time variable.
We turn now to asymptotic stability and global asymptotic
stability.
14 [ definition] The equilibrium point 0 at time t 0 is asymptotically

stable at time t0 if (1) it is stable at time t0, and (2) there exists a
number (J^) > 0 such that
15 l|x(f 0 )ll < <5 1 (/ 0 )=H|x(0|| —^0 as t — ^ oo
It is uniformly asymptotically stable over [i 0 ,

oo) if (1) it is uniformly
stable over (Y 0 ,
oo), and (2) there exists a number 8 X > 0 such that
16 11x00 II <$U t l >*o=>||x0)|| — as t —ôo
Moreover, the convergence is uniform with respect to t{ .
1
According to definition (14), 0 is an asymptotically stable

equilibrium point at time t0 if it is stable at time t 0 and in addition
all trajectories starting from an initial state x(r 0 ) sufficiently close to
0 actually approach 0 as t — > oo. In this case, we can think of the set
BSl {to) defined by
17 B dlit0) ={xe R n
:
||
x ||
<S t (t Q )}
as a region of attraction for the equilibrium point 0 in the sense that ,
all trajectories starting at time t0 from an initial state within B dl{to)
eventually converge to 0 Notice, however, that (15) does not imply .
that any trajectory starting within BMo) is confined to B for all

Sl{to)
t > : It is quite possible that some trajectory starts at time t0 from an

initial state within Bsdto) but leaves Bsdto) at some later time. All that
(15) implies is that any such trajectory will ultimately return to B Sl{to)
after a finite amount of time, 2 will thereafter be confined to B Sl{to) ,
and will approach 0 as t — 00 . The equilibrium point 0 is uniformly

asymptotically stable over [0, 00 ) if it is uniformly stable over [0, 00 )
lr
The last statement can be made more precise, as follows: Given any e > 0,
there exists a T(f) < 00 such that
||x(/0|| < <5i => 1

1 x(r 1 + OH < € whenever t > T{e)
2 During this finite period of time, the trajectory can leave and enter Bsm
any number of times.
and if one can find a nontrivial ball of attraction that is independent of

initial time.
A word of caution in applying definition (14). One might be

tempted to think that condition 2 of definition (14) implies the stability
at time t0 of the equilibrium point 0 and that condition 1 of definition
(14) is therefore superfluous. However, a simple example shows that
this is not the case. Consider a system whose trajectories take the
form shown in Fig. 5.2, where all trajectories starting from nonzero
initial points within the disk of radius 8 X first touch the curve <3 before
converging to the origin. This system satisfies condition 2 of definition
However, condition 1 is violated, because 0 is an unstable equilib-
(14).
rium point. This is most easily seen by applying definition (6) with
e = S u as shown in Fig. 5.2; then no 5 > 0 can be found such that
(7) holds.
We shall now give one last definition, which corresponds to the

strongest kind of stability that we shall study in this book.
[definition] The equilibrium point 0 at time t0 is globally

asymptotically stable if x(t) — > 0 as t — » oo [regardless of what
x(* 0 ) is].
Thus global asymptotic stability results if all trajectories of the

system converge to the equilibrium point 0 as t —> oo, i.e., if the sphere
of attraction is the entire state space Rn .
In talking about and asymptotic stability,

stability, instability,
one should be very careful to associate these properties with a specific

equilibrium point of the system, because in general a system can have
more than one equilibrium point, each of which has its own stability
status. However, if 0 is a globally asymptotically stable equilibrium
point at time tQ of a particular system, then it is clearly the only equi-

librium point at time t0 (see Problem 5.1). Hence, one can say, without
fear of confusion, that the system is globally asymptotically stable as
well as that the equilibrium point 0 is globally asymptotically stable.
The distinction is that stability, asymptotic stability, and instability
are basically local concepts, dealing with the trajectories of the system
in the vicinity of an equilibrium point, whereas global asymptotic sta-
bility, as the name implies, is a global concept, having to do with the
behavior of all the trajectories of the system.
19 Example. The traditional example for illustrating the various

concepts of stability is that of a ball rolling on a curved table top under
the influence of gravity. To keep matters simple, we assume a one-
dimensional motion, with the table top being frictionless, or “with fric-
tion” as the example demands. It is clear that the profile of the table
top, within a scale factor, corresponds exactly to the profile of the
potential energy of the ball.
As our first example, we shall consider the case of the table top
being absolutely flat, as depicted in Fig. 5.3. In this case, r = 0, r = 0
O
^ > Position r
FIG. 5.3 r = 0
isan equilibrium point, as is r = r 09 r = 0 for any r 0 because no matter ,
where we initially place the ball, it does not move. If we define the norm
of the state vector x = [r r]' as
20 ||
x ||
= (r
2
+r 2
)
1/2
then it is clear that ||

x(t) = x(0) for all > 0. Hence the condition
|| || ||
t
(7) holds for every e > 0 if we choose S = e. Accordingly, the equilib-

rium point x = 0 is stable, by definition (6). Every other equilibrium
point [r 0 0]' is also stable.
Now, suppose the table top is as shown
in Fig. 5.4 and is friction-
less. Then once again r = 0, r =0
an equilibrium state. Now, is
suppose the ball is displaced to the dotted position and is let go (with
zero initial velocity). In the absence of friction, the ball will roll down
through the equilibrium position and up the other side to the same height
h and back again, as there is a sustained oscillation resulting from a

continuous exchange between potential energy and kinetic energy.
Tn this case, once again, x = 0 is a stable equilibrium point.
Now consider the situation in Fig. 5.4, and suppose that there
is a frictional force proportional to r. In this case, after the ball is
displaced to the dotted position and released, it will roll back and forth,
but because of the energy dissipation due to friction, the amplitude of
the oscillations gradually decreases and approaches 0 as t — » oo. In
this case, the equilibrium point 0 is asymptotically stable and is indeed

globally asymptotically stable.
Suppose now that the table top is shaped as shown in Fig. 5.5,
and suppose there is a frictional force proportional to r. Then the
equilibrium point at r — r u r = 0 asymptotically stable, the equilib-
is
rium point r =r 2, r = 0 unstable, while the equilibrium point at

is
r =r 3 ,
r =0 is stable. The details are left as an exercise.
21 Example. To illustrate the distinction between stability and

uniform stability, consider the scalar differential equation
22 x(t) = (6t sin t — 2t)x(t)

The solution to (22) is given by
23 x{t) = x(t Q) exp {6 sin t — 6t cos t — t

2 — 6 sin t0 + 6t 0 cos t0 + tQ
The system an equilibrium point at *

(22) has 0. We shall show that =
this equilibrium point is stable at any time t0 0 but is not uniformly >
stable over [0, oo). First, let t 0 0 be any fixed initial time. Then >
24
*(0 — exp [6 sin — t 6t cos t — t
2 — 6 sin t0 + 6t 0 cos t0 + tl }
*(*o) I
Now, if t — t0 > 6, the quantity on the right-hand side of (24) is

bounded above by exp [12 + T(6 — T)], where T — t — t 0 Because it .
is a continuous function of time, it is also bounded over the interval
[t 0 , t 0 + 6]. Hence if we define
25 c(t 0 ) = sup exp {6 sin — 6t cos — t t t

2 — 6 sin + 6r t0 0 cos t0 + ô}
t>to
we know that c(t 0 ) is a finite number, for any fixed t0 . Thus, given any
x n
e > 0, if we choose <5(6, t 0 ) = e/c(t 0 ), then (7) is satisfied, showing

that x — 0 is a stable equilibrium point for all times t0 . On the other
hand, if we choose t0 = 2nn, then (23) yields
26 x[(2 + 1 )n] = x{2mi) exp {(4 n + 1)(6 — 71 ) 71 }
This shows that
27 c(2nn) > exp {{An + 1)(6 — 71 ) 71 }
Hence c(t 0 ) is unbounded as a function of t0 . Thus, given 6 > 0, it is
not possible to choose a single <5(6), independent of the initial time,
such that (8) holds. Therefore the equilibrium point x = 0 is not uni-
formly stable over [0, 00 ).
We shall close this subsection by showing that, in the case of 3
autonomous and periodic systems, stability is equivalent to uniform

stability and that asymptotic stability is equivalent to uniform asympto-
tic stability. Actually, it is enough to prove these statements for the
case of periodic systems, because an autonomous system can be thought

of as a periodic system with arbitrary period. The following preliminary
result is useful in its own right.
28 lemma Suppose the equilibrium point 0 at time t 0 of the system (1) is

stable atsome time t > t 0 Then 0 is also a stable equilibrium point
x
. at all
times r e h 0 1 ]. , 1
proof Let s(t, x0 , t) denote the solution of (1) together with the initial
condition
29 x(T) = x0
as evaluated at the time t. The fact that the equilibrium point 0 is stable at
time 11 means > 0, there exists a
that, for every 6 <5 > 0 such that
30 ||
0 1|
< 5=>||s(/,x 0 ,fi)|| < € V t>t > x
Let r e ho, L], and define
31 m{r) = sup ||s(/i 9 x0 , t)||

!lxo!l<r
By assumption, m{r) is continuous in r, and m{0) = 0. Therefore one can

find an r0 > 0 such that
32 m{r) < <5 whenever r < r0
Finally, suppose ||x 0 1|

< r0 ;
then (32) and (31) imply that
33 ||s(/ 1 ,Xo,T)|| < S

Next, (33) and (30) together imply that
34 ||s[/, s(fi,x 0 ,T), fJH <e, V t>t y
3
The remainder of Sec. 5.1.1 may be skipped without loss of continuity.
1 1 ] 1 1
However, it is easy to show that
35 s [t, s (t u x0 , T), *i] = s 0, x 0, T)
(see Problem 5.3). Therefore we conclude that
36 ||s (t, x0 , t) II
< 6, \f t> T whenever ||x 0 1|
< r0
Here we have made use of the obvious fact that 3 < €. Because the above
argument can be repeated for any e > 0, it follows that 0 is a stable
equilibrium point at time T. mm
37 Theorem Consider the system (1) and suppose f satisfies
38 i(t , x) = f(/ + T, x), \/ xe R ,\/ t> 0 n
for some positive number T. (If the system is autonomous, T is arbi-

trary.) Under these conditions, the following two statements are
equivalent:
(i) The equilibrium point 0 of the system (l) 4 is stable at some
time tQ > 0.
(ii) The equilibrium point 0 of the system (1) is uniformly stable
over the interval [0, oo).
proof Clearly (ii) implies (i), so it remains only to show that (i)
implies (ii). Accordingly, assume (i) holds. Because of the periodicity of f,
it follows readily that
39 s (t, x0 , t 0) = s(* + nT, x0 ,

t0 + nT)
for all nonnegative integers n. Therefore, if (i) holds, then 0 is a stable
equilibrium point at all times t0 + nT, n > 0. Now, applying lemma (28)
shows that 0 is a stable equilibrium point at all times T e [0, T]. Define
40 /*(x 0 , t) = sup ||
s(t + y, x0 ,
t) ||
y>0
Because 0 is a stable equilibrium point all times T e [0, T], the function
jh(xq, t) is continuous at x 0 = 0, for each T (see Problem 5.4). There-
fore, if
41 //(x 0 ) A sup pi(x 0 ,t)

tE[0,T
then //(•) is continuous at x 0 =0. Thus, given any € > 0, we can find a
<5 > 0 such that
42 ||xo|| < 3 => f](x 0 ) <6

But (42), together with the definition of means that
43 1
x0 1
< <5, T E [0, T] => 1
s(t + y, x 0 , t) 1
<6 vy>0
4 Note that, because of the periodicity of the function f, we have that if 0 is
an equilibrium point at some time, it is an equilibrium point it all times.

Finally, because of the periodicity of f, (43) can be broadened to
44 ||x 0 || < <5=H|s(t + y,x 0 ,T)|| < e, vr>0,vy>0

Because this argument can be repeated for any e > 0, (44) establishes that
the equilibrium point 0 is uniformly stable over the interval [0, oo). B—
A result similar to Theorem (37) can be proved for asymptotic

stability as well.
45 Theorem Consider the system (1), and suppose f satisfies
46 i(t , x) = i{t + T, x), V x e R V n

,
t >0
for some positive number T. Under these conditions, the following
two statements are equivalent:
(i) The equilibrium point 0 of the system (1) is asymptotically
stable at some time t 0 0. >

(ii) The equilibrium point 0 of the system (1) is uniformly asymp-
totically stable over the interval [0, oo).
The proof is omitted here in the interests of brevity and can be

found in [13] Hahn (see also Problem 5.5).
5.1.2 Definite and Locally Definite Functions
In this subsection, we shall present some concepts, such as positive

definite functions, that are central to the development of Liapunov
stability theory, as well as the notion of the derivative along a trajec-
tory.
47 [definition] A continuous function V: R + x Rn —> R is said to

be a locally positive definite function (l.p.d.f.) if there exists a
continuous nondecreasing function a : R—+ R such that
48 (i) a(0) = 0, cc(p) > 0 whenever p > 0
49 (ii) V(t, 0) = 0, V*>0
50 (iii) V(t, x) > a(||x||), V *>0, and for all x belonging to
some ball
51 B r
= {x: ||x|| < r}, r> 0
V is said to be a positive definite function (p.d.f.) if (50) holds for all

x e Rn and in addition a (p) — oo as p — > oo.
The kinds of functions a satisfying (48) above arise quite fre-

quently in Liapunov theory —so much so that it is desirable to give a
name to this class of functions.
]
52 [ definition A continuous function a R — R is said to belong to

: >
class K if
(i) a(*) is nondecreasing,
(ii) oc(0) = 0, and
(iii) cc(p) > 0 whenever p > 0.
Note that some authors replace (i) above by the more stringent
requirement that <%(•) is strictly increasing. It turns out, however,
that both definitions are equally good in proving stability theorems.

Given a continuous function V: R + X Rn — > R, it is rather diffi-
cult to determine, using definition (47), whether or not V is a p.d.f.

or an l.p.d.f. The main source of difficulty is the need to exhibit the
function a(-)- Lemmas (53) and (61) give an equivalent characterization
of l.p.d.f.’s and p.d.f.’s and have the advantage that the conditions
given therein are more readily verifiable than those in definition (47).
53 lemma A continuous function W: R n — > R is an l.p.d.f. if and only if
54 (i) W(0) = 0
55 (ii) W(x) >0, VxÔ belonging to some ball B r, r > 0 [where
the ball B r is defined in (51)].
W is a p.d.f. if and only if
56 (iii) W(0) =0
57 (iv) W(x) >0, V x 7* 0
58 (v) W(x) > oo — as ||x|| — » oo, uniformly in x.
proof We shall first consider l.p.d.f.’s. Suppose W

is an l.p.d.f. in the
sense of definition (47). Then clearly and (ii) above hold. To prove the
(i)
converse, suppose (i) and (ii) above hold, and define
59 a (p) = inf W(x)

p<M<r
Then clearly oc(0) = 0, a is continuous, and a(-) is nondecreasing, because
as p infimum is taken over a smaller region. Also, suppose
increases, the
p > 0; then the infimum in (59) is also positive, because is a continuous W

function. 5 Therefore OC(-) is a function of class K. Because W(x) a(||x||) >
V x e B r, W is an l.p.d.f. in the sense of definition (47).
In the case of p.d.f.’s, the necessity of conditions (iii), (iv), and (v) is
immediate from definition (47). To prove the sufficiency, define
60 0C(p) = inf W(x)

p<l|x||
The rest of the proof is the same as in the case of l.p.d.f.’s and is left as an
exercise. wm
5
The proof of this statement involves the so-called compactness property of
the unit ball in R n and can be found in any standard text on real analysis; see, e.g.,
[20] Royden.
61 lemma A Rn — R is an l.p.d.f. if and

continuous function V: R+ x >
only if there exists anR such that

l.p.d.f. W: Rn — >
62 V(t, x) > fV(x), V t > 0, V x e B r > 0 r,
where B is a ball in R A continuous function V: R+ x Rn — R is a p.d.f.

r
n
. >
if and only if there exists a p.d.f. W\ R — R such that

n
>
63 V(t, x) > W(x\ V t > 0, V x e Rn
proof We shall give the proof only for l.p.d.f.’s,

because the proof for
p.d.f.’s is entirely parallel. Suppose is an l.p.d.f. and that (62) holds. W
Then it is easy to verify that V is an l.p.d.f. in the sense of definition (47).
Conversely, suppose V is an l.p.d.f. in the sense of definition (47), and let
a(-) be the function of class K such that (50) holds. Then fV(x) = a(||x||)
is an l.p.d.f. such that (62) holds. M
64 Remarks :
1. The conditions given in lemma (53) are easier to verify than

those in definition (47).
2. If W(x) is a polynomial in the components of x, then one can
systematically check, in a finite number of operations, whether
or not (55) is satisfied; see [15] Jury for details.
3. Lemma shows that a continuous function of t and x is
(61)
an and only if it dominates, at each instant of time
l.p.d.f. if
and over some ball in R n an l.p.d.f. of x alone. Similarly, a

,
continuous function of t and x is a p.d.f. if and only if it domi-

nates, for all t and x, a p.d.f. of x alone.
4. Suppose W(x) = x'Mx, where is a real symmetric n X n M
matrix. Then it is easy to show that W is a positive definite
function if and only if M is a positive definite matrix (see also
Problem 5.7).
65 [definition] A continuous function V: R+ X R n —R > is said to

be decrescent if there exists a function /?(•) belonging to class K
such that
66 V(t, x) < jS(||x||), v t > 0, v X g B„ r > 0
where B r
is a ball in R n
.
In other words, a function V: R+ X R — Rn

-> is decrescent if and
only if, for each p in some interval (0, r), we have
67 sup sup V(t, x) <C oo
Ilx||0
68 Example. The function
Wi(x u x 2) = x\ + x\
is a simple example of a p.d.f. Clearly W1

satisfies (56) and (57); it can
be seen to satisfy (58) also, because
^(x) = ||x|P
if we take ||
• || to be the Euclidean norm on R2 .
Example. The function

V x
{t, x u x 2) = (t + l)(x? + xl)
is a p.d.f., because it dominates the time-invariant p.d.f. W x
. However,
it is not decrescent, because for each (x u x 2) (0, 0), ^ V (t, x u x 2 )
x
is
unbounded as t increases.

V2 (t, x u x 2 ) = e~‘(x\ + x\)
is not a p.d.f., because no p.d.f. W: R exists such that (63) holds.
This can be seen from the fact that, for each (x l5 V(t, x u x 2 ) 0 x 2 ), —
-*•
as t —
> oo This example shows that it is not
.
possible to weaken (63)
to the statement
V(t, x) > 0, V t> 0, V x ^0

The present function V2 is decrescent.

W (x u x = x\ + sin
2 2)
2
x2
is an 1. p.d.f., though it is not a p.d.f. Note that (1) W 2 ( 0, 0) = 0, and
(2) W
2 (x 1 ,x 2 )> 0
whenever (x„ x 2 ) =£ (0, 0) and \x 2 \<n/2. This
is enough to ensure that 2 is an l.p.d.f.
However, W
2 is not a p.d.f., W
because it vanishes at points other than (0, 0), e.g., (0, n).
We by introducing the concept of the

shall close this subsection
Suppose V
derivative of a function V(t, x) along the trajectories of (1).
has continuous partial derivatives in all of its arguments, and let VF
denote the gradient of V with respect to x. Suppose x(.) is a solution of
(1), and consider the function t h* V[t, x(t)]. This function is differen-
tiable with respect to t, and indeed we have
±V[t, x(f)] = ~[t, x(0] + VF|), x(t)]f[t, x(01

We use the symbol V[t, x(t)] to denote the right-hand side of (71). This
choice of symbols is justified because
V[t, x(/)] = V[t 0 , x(/ 0 )] + f'

J t0
F[t, x(t)] dx
whenever x(-) is a solution of (1). This leads to the following definition.

74 [definition] Let V: R x Rn — > R be continuously differentiable

with respect to all its arguments, and let VV denote the gradient
of V with respect to x. Then V: R + x R — R is defined by
n
>
75 V(t, x) = ^-(t, x) + V V(t, x)f (/, x)
and is called the derivative of V along the trajectories of (7).
76 Remarks :
1. Note that V depends not only on Fbut also on the system
(1). If we keep the same V but change the system (1), the
resulting V will in general be different.

2. The quantity V(t 0 , x 0 ) can be interpreted as follows:
Suppose a solution trajectory x(.) of (1) passess through
x0 at time t0 . Then, at the instant t0, the rate of change of
V[t, x(t)] is V(t Q x 0 ). ,
3. Note that if V is independent of t and if the system (1) is

autonomous, then V is also independent of t.
Problem 5.1. Show that if 0 is a globally asymptotically stable equilib-

rium point of (1) at time t0, then it is the only equilibrium point of (1) at
time tQ .
Problem 5.2. The purpose of this problem is to generalize Example (21)

by developing a whole class of linear systems with equilibrium points that are
stable but not uniformly stable. Consider the linear scalar system
77 x(t) = a(t)x(t) 9 t> 0

where «(•) is a continuous function.
(a) Verify that the general solution of (77) is
x(t) = x(t 0 ) exp
(b) Show, using definition (6), that x == 0 is a stable equilibrium point

at time t 0 if and only if
sup f a(t) dr A m(t < 0) oo

f>fo Jfo
Show that, in this case, (7) is satisfied with
<50o. e) = <7 exp [

m{ t 0 ) ]
(c) Show that, if m(t 0 ) is unbounded as a function of t0, then 0 is not

uniformly stable over [0, oo).
(d) Construct several functions a(f such that m(t 0 ) is finite for each t0
but /w(-) is an unbounded function.
\
-(2/i + 3) h
FIG. 5.6
Problem 5.3. Let s(/, x 0 / 0 ) be as defined in the proof of lemma

, (28).
Show, using the uniqueness of solutions to (1), that
78 s|[t, s o (/ 1 , x0 ,
t o), t t ] = s (t, x0 ,
t 0 ), V t > > tx t0
Give a physical interpretation of (78).
Problem 5.4. Let ju(x 0 t) be defined as in

, (40). Show that the stability
at time T of the equilibrium point x0 = 0 is equivalent to the continuity, at
x0 = 0, of ju(x Q ,
t). [Hint: Use (7).]
Problem 5.5. Suppose the system (1) is autonomous. Show that

79 s (t, Xq, to) — s (t + T, Xq, to + T), V / > to, V xG JRP, V T^> 0
Using (79), prove Theorem (45) for autonomous systems.
Problem 5.6. Complete the proofs of lemmas (53) and (61).
Problem 5.7. Suppose V: R+ X R n — R is defined by
V(t, x) = x'M(t)x
where M(f) is a real symmetric n x n matrix, and t h-> M(0 is continuous.
Find the necessary and sufficient conditions for V to be a p.d.f. [Answer:
inf,>o AminMW > 0, where 2 min denotes the smallest eigenvalue.]
Problem 5.8. Determine whether or not each of the following func-

tions is(a) an l.p.d.f., (b) a p.d.f., or (c) a decrescent function.
(i) W(x u x 2 ) = x\ + x 2
(ii) W(x 1} x 2 ) = x\ + x\
(iii) W(x u x2 ) = (x + x\) 2 t
(iv) V(t, x u * 2 ) = t(x + x\)
(v) V(t, x u x 2 ) = (x\ + x 2 )l(t + 1)
(vi) W{x u x 2 ) = sin 2 fa + x2 ) + sin 2 (xj — x2 )

Sec. 5.2 Liapunov's Direct Method 147
5.2
LIAPUNOV'S DIRECT METHOD
The idea behind the various Liapunov theorems on stability, asymptotic
stability, and instability is as follows: Consider a system which is
“isolated” in the sense that there are no external forces acting on the
system. Equation [5.1(1)] is a suitable model for such a system because
no input is explicitly identified on the right-hand side of [5. 1(1)]. Suppose
that one can identify the equilibrium states of the system and that 0 is
one of the equilibrium states (or possibly the only equilibrium state).
Suppose it is also possible to define, in some suitable manner, the total
energy of the system, which is a function having the property that it is
zero at the origin and positive everywhere else. (In other words, the
energy of the system has its global minimum at 0.) Now suppose the
system, which is originally at the equilibrium state 0, is perturbed into
a new nonzero initial state (where the energy level is positive, by assump-
system dynamics are such that the energy of the system is
tion). If the
nonincreasing with time, then the energy level of the system never
increases beyond its initial positive value. Depending on the nature of
the energy function, this may be enough to imply the stability of the
equilibrium point 0. On the other hand, if the system dynamics are such
that the energy of the system is monotonically decreasing with time and
the energy eventually reduces to zero, it may be possible to conclude
the asymptotic stability of the equilibrium state 0, under suitable
additional assumptions. The theorems of Liapunov, as well as their
generalizations proved by later researchers, cast these ideas into a
mathematically precise form. In addition, they offer the flexibility of
allowing one to choose the energy function corresponding to a particular
system.
In this section, we shall present the three basic kinds of theorems
stemming from the so-called method of Liapunov, namely:
direct
stability theorems, asymptotic stability theorems, and instability theo-
rems. These are illustrated by several examples, some of which are of a
dynamics nature. These examples serve to show that, in some cases,
the total energy stored within the system is a logical choice for the
Liapunov function 6 of the system.
5.2.1 Stability Theorems
Theorem 1 is the basic stability theorem of Liapunov’s direct

method.
6 This term is defined later.

Theorem The equilibrium point 0 at time t0 of [5.1(1)] is stable
if there exists a continuously differentiable l.p.d.f. V such that
V(t, x) < 0, V t > t 09 V x e Br for some ball B r
proof Because V is an l.p.d.f., we have that
V(t, x) >a (|
|
x 1 1), \/ t>t 0 V ,
x.e B s for some ball B s
—
where a R > R is a function of class K. To show that 0 is a stable equilib-
:
rium point at time / 0 ,we must show that, given any € > 0, we can find a
S(t 0 €) > 0 such that [5.1(7)] holds. Accordingly, given e > 0, let
,
=
min [e, r, s} and pick 8 > 0 such that
9
/?(f 0 , S) = sup V{t 0 x) ,

< a (fj)
Hx||<5
Such a 8 can always be found, because a(ei) > 0 and fi(t 0 , 5) — » 0 as

<5 — We now show that [5.1(7)] holds if 8(t 0
> 0. ,
chosen equal to this
e) is
8. Suppose||x(^ 0 )|| < S. Then V[t 0 x(^ 0 )] < ,

S) < a(6 t ). But because
V(t, x) < 0 whenever ||x|| < 8 we have
,
7
V[t, x(0] < V[to, x(r 0 )] < a(6j), V > to /
Now, because
a(||x(0ll) < V[t, x(0]
(5) and (6) together imply that
a(l|x(0ID < a(6i)
Because a is a nondecreasing function, (7) implies that
1 1
x(r) 1 1
<6j <€, V t>t 0
Hence [5.1(7)] holds and 0 is a stable equilibrium point at time f0 - mm
Theorem The equilibrium point 0 at time t0 is uniformly stable

over [t 0 , oo) if there exists a continuously differentiable, decrescent
l.p.d.f. V such that
V(t, x) < 0, V > t to, V x e Br for some ball B r
proof Because V is decrescent, the function
f}{8) = sup sup V(t, x)

|jx || <5 t>to
is nondecreasing and satisfies, for some d > 0, the condition
m < co whenever 0
Thus, letting €i = min {e, r, s, d], we can always pick 8 > 0 such that
<5<d
m < a(60. One now proceeds exactly as in the proof of Theorem (1) to
7 Note that S < €\ < r.

show that [5.1(8)] holds with this choice of <5. The details are left as an
exercise. —
13 Remarks :
1. Theorem (1) states that if we can find a continuously differen-

tiable l.p.d.f. V such that its derivative along the trajectories
of [5.1(1)] is always nonpositive, then the equilibrium point
0 is stable at time tQ . Theorem (9) shows that in order to con-
clude the uniform stability of the equilibrium point 0 over the
interval [? 0 enough to add the assumption that V
,
oo), it is
is also decrescent. It should be noted thatTheorems (1) and

(9) provide only sufficient conditions for stability and uniform
stability. [Actually, one can prove the converse of Theorem
(9). This is discussed later on.]

2. The function V is commonly known as a Liapunov function or
a Liapunov function candidate. The term Liapunov function is
is a source of great confusion, and we shall attempt to avoid

it in the following manner Suppose,: for example, that we are
attempting to show that 0 is a stable equilibrium point by apply-
ing Theorem (1). Then we shall refer to a function V as a Lia-
punov function candidate (L.f.c.) if it satisfies the requirements
imposed on V in the hypothesis of Theorem (1), i.e., if V is
continuously differentiable and V is an l.p.d.f. If, for a parti-
cular system [5.1(1)], the conditions imposed on V are also
satisfied, then V is referred to as a Liapunov function. The
rationale behind this convention is as follows: Theorems

(1) and (9) are sufficient conditions for certain stability prop-
erties. To apply them to a particular system, it is a fairly simple
matter to find a function V satisfying the requirements on V.

At this stage, V is a Liapunov function candidate. Now, for the
particular system under study and for the particular choice
V may or may not be met. If the require-
of V, the conditions on
ments on V are met, then definite conclusions can be drawn,
and V then becomes a Liapunov function. On the other hand,
if the requirements on V are not met, because the theorems
are only sufficient conditions, no definite conclusions can be
drawn, and one has to start again with another Liapunov
function candidate. The examples that follow illustrate this
usage.
14 Example. Consider the simple pendulum, which is described by
15 0 -j- sin 9 —0
:
after suitable normalization. The state variable representation of this

system is
16 x i
—X 2
17 x2 — sin Xj
Now, the total energy of the pendulum is the sum of the potential and
kinetic energies, which is
18 F(x l5 x 2 ) = (1 — cos x ) + t
where the term represents the potential energy and the second
first
represents the kinetic energy. One can readily verify that V is an l.p.d.f.
and is continuously differentiable, so that V is a Liapunov function
candidate for applying Theorem (1).
Next, we have
19 F(x l9 x 2 ) = sin XjXi +xx 2 2
— sin Xj(x 2 ) + x (— sin Xj) = 0

2
Therefore V also satisfies the requirements of Theorem (1). Hence V is

a Liapunov function, and the equilibrium point 0 is stable by Theorem
(1). Actually, because the system is autonomous, 0 is a uniformly stable
equilibrium point (see Theorem [5.1(37)].)
20 Example. Consider a mass constrained by a nonlinear spring,

with friction also present. Such a system can be represented in state
variable form by
21 Xj =x 2
22 x2 = -f(x - g(x,) 2)
where f(x 2 ) represents the friction and g(x,) represents the restoring
force of the spring. It is assumed that / and g satisfy the following
conditions
(i) / andg are continuous.
(ii) Whenever a belongs to some interval [— <7 0 ,
we have
23 of(a) >0, V o g [— <7 0 ,

<7 0 ]
24 < 7 g(cr) >0, \/ a e [— cr 0 ,

<7 0 ], cr ^0
It is easy to see that Example (14) is a special case of Example (20)
with f(p) =0 and g(p) = sin 7
< .
The energy stored in the system consists of the kinetic energy

of the mass plus the potential energy stored in the spring. Thus the
Liapunov function candidate is chosen as
^
25 V(x u x 2 ) = da
^+ J*
g(<r)
Note that the energy dissipated through friction is not “stored” in the
system (because it is not recoverable) and is therefore not included in
V. Note also that V as defined in (25) is a natural generalization of that
defined in (18). Now, by virtue of the assumptions on g(*), it is easy
to show that V is continuously differentiable and is an l.p.d.f., and
is hence a suitable Liapunov function candidate for applying
Theorem (1).
Next, we have
26 V(x u x 2 ) = x x + g(x )x
2 2 1 1
= x [—f(x - g(x
2 2) t )] + g(x 1
)x 2
= -x f(x 2 2)
By assumption (23) on/(.), it follows that
27 V(x u x 2 ) <0 whenever x 2 | |

<a
Hence V also meets the requirements of Theorem (1). Therefore V is
a Liapunov function for the system (21)— (22), and 0 is a (uniformly)
stable equilibrium point.
28 Example. Consider the system described by
29 y(t) + y(t) + (2 + sin 0^(0 = 0
or, in state variable form,
30 X x
—X 2
31 x2 = —x — 2 (2 + sin f)*i
Equation (29) an example of what is known as the damped Mathieu

is
equation. In this case, there is no physical intuition readily available
to guide us in the choice of V. Thus (after possibly a great deal of trial

and error), we might be led to the Liapunov function candidate
32 V(t Xi 9 , x2) = xj + z o'"!

2
sm -
f
t
-f-
Then V is continuously differentiable; moreover, V dominates the p.d.f.

33 Wi(v l5 x2) = x\ -]
—
and is dominated by the p.d.f.
34 W (xi 2 , x2) = x\ + x\
Hence Fis a and thus a suitable Liapunov function

p.d.f., decrescent,
candidate for applying Theorem (9). We have
cos
35 V(t,x u x 2 ) = -Xl
t
+L 9v
2. V
x vx, +2+
4- jjfil •)
J
(2 + sin 02 si nt
cos t
~~ X2
2
(2 + sinf)
2 + 2x^2
+ 2 +^Sn [-)<i-<2 + sin ' ) *‘ 1
, + 2(2 + sin t)
.2
1
cos t
+ sin tf (2
~~ _ Xl 4 + 2 sin + cos
2 f t
(2 + sin 0
2
< 0, V t>. 0, V X u x2
Thus the requirements on V in Theorem (9) are also met. Hence V is a

Liapunov function for the system (30)—(31), and 0 is a uniformly stable
equilibrium point over the interval [0, oo).
36 Example. of the main applications of Liapunov theory is in

One
obtaining stability conditions involving the parameters of the system
under study. As an illustration, consider the system
37 y(t) + p(t)yU) + e~‘y(t) = 0

which can be represented in state variable form as
38 X\ = x2
39 x2 = —p{t)x — e~’x,
2
The objective is to find some conditions on the function p(-) that

ensure the stability at time 0 of the equilibrium point 0 For this pur- .
pose, let us choose
40 V(t, x u x2) = x\ + e x\ l
Because V is continuously differentiable and dominates the p.d.f.
41 W(x u x 2 ) = xj + xl
we see that V is a suitable Liapunov function candidate for applying
Theorem (1). However, because V is not decrescent, it is not a suitable
Liapunov function candidate for applying Theorem (9). Hence, using
this particular F-function, we cannot hope to establish uniform
stabil-
ity by applying Theorem (9).
Differentiating F, we get
42 V(t, x u x2 ) = e*xl + 2xj(x + 2e'x 2) 2 [—/?(r)x 2 — £?"%]
— &x\[ — 2p(t) + 1]
Hence we see that V is always nonpositive provided
43 P(t) > y, V/>0

Thus the equilibrium point 0 is stable at time 0 provided (43) holds.
It should be emphasized that by employing a different Liapunov

function candidate, we might be able to obtain entirely different
stability conditions involving /?(•).
44 Remarks: The definition of stability in the sense of Liapunov is a

qualitative one, in the sense that given an e > 0, we are only required
to find some 5 > 0 such that [5.1(7)] holds — or, to put it another way,
we are only required to demonstrate the existence of a suitable 8.
In the same way, Theorems (1) and (9) are also qualitative in the
sense that they provide some conditions under which the existence of
a suitable 8 can be concluded. However, in principle at least, the
conditions fi(t 0 , 8) < oc(e) [if Theorem (1) is being applied] or ft (8)
< oc(e) [if Theorem (9) is being applied] can be used to actually deter-
mine a suitable value of 8 corresponding to a given e. But in practice
this is rather messy.
5.2.2 Asymptotic Stability Theorems
In this subsection, we shall present some theorems that give

sufficient conditions for asymptotic stability and global asymptotic
stability.
45 Theorem The equilibrium point 0 at time t0 of [5.1(1)] is uniformly

asymptotically stable over the interval [t Q ,
oo) if there exists a con-
tinuously differentiable decrescent l.p.d.f. V such that —V is an
l.p.d.f.
proof If — V is an l.p.d.f., then clearly V satisfies the hypothesis of

Theorem (9), so that 0 is a uniformly stable equilibrium point over [t 0 , oo).
Thus, according to definition [5.1(14)], it is only necessary to prove the
uniform convergence, to 0, of solution trajectories starting sufficiently
close to 0. More precisely, we have to prove the existence of a Si >0 such
that, for each e > 0, there is a T(e) < oo such that
46 1 1
s(t j + t, x 0 Oil
,
<€ whenever t > T(e\ [|x 0 ||
< S u and V tx > t0
*1
The hypotheses on V and V imply that there are functions a(-), /?(•),
and y(0 belonging to class K such that
47 «(||x||)< V(t,x)< Allxll) w
V t
_ .
t o, V x e B r for some ball B r

< — y(||x|i)
.
48 V(t, x)
Now, given € > 0, define positive constants S u S 2 and Tby

,
49 P(Si) < «0)

50 P(5 2 ) < min {a(e), fi(d ,)}
51 II
Now we claim that

52 I|s(f 2 ,x 0 ,fi)|| <s 2 for some t2 e [t uh + T]
whenever ||x 0 1|
< 8 1 and t1 >t 0
To prove (52), assume by way of contradiction that
53 l|s(f, X0 , *i)|| >S 2, V t e [tuh + T]
Then
54 0 < a(<5 2 ) < V\t\ + J 7
, s(î + T, x0 , fi)] by (47)
rti+T .
— V(ti, x ) + 0 |
V[r, s(t, x0 , ti)] dr
Jt i
< p(5,) - Ty(8 2) by (47) and (53)
= Aî) - «oo by (51)
<0 by (49)
which isa contradiction. This shows that (52) is true.

To complete the proof, suppose t t\ + T. Then >
55 a(||s (t, x0 , fi)||) < V[t, s (t, x0 ,
1 1 )] by (47)
< V[t 2 , s(^2, X0 , *i)]
by the nonpositivity of V. [Note that t 2 is defined in (52).] Finally,
56 V[t 2 , s (t 2 , x0 , ti)] < j3[\\s(t2 , x0 , *i)||] < P(S 2) by (52)
Now, (55) and (56) together imply that
57 a[| s(r,I
x0 , *i)||] < p(d < a(6) 2) by (50)
The inequality (57) establishes (46). wm
58 Theorem The equilibrium point 0 at time t0 of [5.1(1)] is globally
asymptotically stable if there exists a continuously differentiable

decrescent p.d.f. V such that
59 V(t, x) < — y(Hx||X V t > t0 , V x g Rn

where y is a function belonging to class K.
*
proof Same as that of Theorem (45), except that one can take r = oo
in (47). m
Remarks :
1. Note that, in Theorem (58), there is no requirement that

— V should be a p.d.f., because there is no requirement that
y(p)
— > oo as p — oo.
2. In Theorems (45) and (58), V is required to be decrescent, but

— V need not be decrescent.
3. In view of lemma [5.1(53)], condition (59) above can be
replaced by an equivalent condition, as follows: There exists
a continuous function W\ R" * R such that —
60 V(t, x) < — W(x), V t > t0, V x e R"
where
61 W(0) = 0; W(x) > 0, VxÔ

63 Xi = x^xj + x\ — 1) — x 2
64 X2 = X + X (X + A ~ 1)
1 l \
and try the Liapunov function candidate
65 V(x u x 2 ) = x\ + x\
Then V is a p.d.f. Further,
66 V(x u x 2 ) = 2(x\ + xl)(x\ + xl— 1)
Thus — V is an l.p.d.f. over the ball {(*„ x 2 ): x\ + x\ 1]. Hence by <

Theorem (45), it follows that 0 is a (uniformly) asymptotically stable
equilibrium point of this system. However, because — V is not an l.p.d.f.
over all of R", Theorem (58) does not apply, and so we cannot conclude
global asymptotic stability. [Of course, since Theorem (58) is only a
sufficient condition, the fact that Theorem (58) is inapplicable does
not
mean that the system is not globally asymptotically stable. However,

for this particular system, a quick sketch of the vector field associated
with equations (62) and (63) reveals that not all trajectories do con-
verge to 0.]
67 Example. Consider now the system of equations
68 xft) = —xft) — e- 2, x 2 {t)
69 x 2 (t) = xft) —x 2 (t)
and choose the Liapunov function candidate
70 V(t, x u x2) = x\ + (1 + e~ 2,
)xl
It is easy to verify that V is a decrescent p.d.f. Moreover,
71 V(t x l5 x 2 )
,
= —2[x\ + x x2 t + x\{ 1 + 2^“ 2f )]
Hence V satisfies (60) with
72 W(x u x 2 ) = 2{x\ +x t
x2 + x\)
Because W satisfies (61), it follows that 0 is a globally asymptotically
stable equilibrium point.
The rest of this subsection is devoted to what is known as

LaSalle’s Theorem. LaSalle’s theorem is more powerful than Theorems
(45) and (58) because it enables one to conclude asymptotic stability
even in cases where — V is not an 1. p.d.f. However, it applies only to
autonomous or periodic systems.
73 [definition] A set S in R n
is said to be the positive limit set of
a trajectory s(«, x 0 ,
/0) of [5.1(1)] if, for every y e S, there exists
a sequence (/„)7 such that tn —+ oo and s(t„, x0 ,
t 0) — > y.
Note that we have previously encountered the concept of the

positive limit set in Chap. 2, in connection with Poincare’s theorems
on limit cycles. Also, note that if 0 is a globally asymptotically stable
equilibrium point of [5.1(1)], then the positive limit set of every trajec-
tory of [5.1(1)] is the single element set { 0 }.
74 [definition] A set M in R n
is said to be an invariant set of the
system [5.1(1)] if, whenever y e M and > t0 0, we have
75 s (t, y, t0) e M, V t >h

In other words, M
is an invariant set of [5.1(1)] if every
trajectory that startsfrom an initial point in stays within at M M

all future times. For example, any equilibrium point of [5.1(1)]
is an invariant set. If [5.1(1)] is autonomous, every trajectory of
[5.1(1)] is a invariant set.
76 lemma If s(-, x0 , *o) is a bounded trajectory of [5.1(1)], then its posi-

tive limit set S is closed and bounded. Furthermore, s (t, x0 ,
t 0) approaches
S as t — > oo; i.e.,
77 inf ||s(f, x0 , /o) — y 1 1

— * 0 as t — > oo
yes
The proof of this lemma is relatively easy but is omitted here,

because it uses a few concepts from real analysis not heretofore intro-
duced.
>
78 lemma Suppose the system [5.1(1)] is autonomous or periodic, and

let S be the positive limit set of any trajectory of [5.1(1)]. Then S is an
invariant set of [5.1(1)].
proof We shall give the proof only in the case of autonomous systems.
Suppose S is the positive limit set of the trajectory s(-, x 0 t 0 ) of [5.1(1)] ,
and that y e S. Let t x >

0 be arbitrary; we must show that s (t, y ,t x ) e S
for all / >
t x Now, by the definition of S, y e S implies that there exists
.
a sequence (/„)? such that t n > oo and s(/„, x 0 t 0 ) > y as n » oo. — ,

— —
Because the solutions of [5.1(1)] are (assumed to be) continuously depen-
dent on the initial conditions, we have, for each fixed t>t u
79 s (t, y, f 0 = lim s[/, s (t n , x0 ,

t 0 ), * i]
n—>°°
However, because the system is autonomous, we have
80 s [t, s (/„, x0 , to), 1 1 ] = s(£ + tn —t + ! to, x0 , to)
(see Problem 5.5). Now, as n — oo, tn — » oo, so that the right-hand side
of (80) converges to an element of S, by lemma (76). Hence s(r, y, 1 1 ) e S
for all t > tx . tm
81 lemma Suppose the system [5.1(1)] is autonomous. Let V: Rn — R be

continuously differentiable, and suppose that for some c > 0, the set
82 Q = {xe c R” V(x) : < c]

is bounded. Suppose that V is bounded below over the set Q c and that
V(x) <0 VxeO c. Let S denote the subset of Q c defined by
83 S = {xe Q c : V(x) = 0}
and let M
be the largest invariant set of [5.1(1)] contained in S. Then,
whenever x 0 e £2 C? the solution s(t, x 0 0) 8 of [5.1(1)] approaches as ,
M
t — * oo.
proof Suppose x 0 e Q c. Clearly, F[s(^, x0 , 0)] is nonincreasing as a

function of t, because V(x) <0 V x e Q This shows, by the definition c .
of Q c, that s(t, x0 , 0) e Q V/> 0. Also, F[s x 0)] has a definite

c, (t, 0,
limit as t — * oo, because it is a nonincreasing function that is bounded

below; let c 0 denote this limit. Let L denote the limit set of the trajectory
s(* , 0). Because V[s (t, x 0 0)]
x0 ,
> c 0 as t > oo, it follows that V(y) =
,
— —
c0 y e L. Because L is an invariant set of [5.1(1)], we have that F(y) = 0
V
V y e L, so that L is a subset of S. Because is the largest invariant sub- M
set in S, L is also a subset of M. By lemma (76), s(f, x 0 0) approaches L ,
as t — > oo, whence we conclude that S (t x 0

} , 0) approaches Mas / — » oo.
8 Note that we can take the initial time to be 0 without loss of generality,
because the system is autonomous.
»
With the aid of lemma (81), we can state and prove a sufficient
condition for asymptotic stability that is weaker than Theorem (45).
Theorem Suppose the system [5.1(1)] is autonomous. Let V: R

n —
R be a continuously differentiable l.p.d.f., and suppose that for some
c > 0, the set
n c = {xGR n V(x)<c} :
is bounded. Suppose V is bounded below on that F(x) <0 VxE

£2 C ,
and that the set
S = {xE£l c : V(x) = 0}
contains no trajectories of [5.1(1)] other than the trivial trajectory
x(t) = 0. Then the equilibrium point 0 of [5.1(1)] is asymptotically
stable.
proof By Theorem (1), 0 is a stable equilibrium point. By

lemma (81), whenever x 0 E £l c ,
the corresponding solution S (t, x 0 0) ,
approaches the largest invariant set contained in S which by hypoth- ,
esis is {0}. Since V is continuous, contains some ball B r Hence .
[5.1(16)1 is satisfied with d l = r, which shows that 0 is an asymptot-

ically stable equilibrium point of [5.1(1)].
Remarks: One should be careful in applying this theorem: The

set S consists of all points in R n
where V(x) = 0 and V(x) < c ,
and
as such may contain points outside the ball B r
.
The results of Theorem (84) can be readily extended to encompass

global asymptotic stability.
Theorem Suppose the system [5.1(1)] is autonomous. Suppose

V: Rn — > R is a continuously differentiable p.d.f., V(x) 0 V x e <
Rn ,
and the set
S = {x e Rn : V(x) = 0}
contains no nontrivial trajectories of [5.1(1)]. Then 0 is a globally
asymptotically stable equilibrium point of [5.1(1)].
proof Let c be any constant in the proof of Theorem (84). H

Up now, we have only considered autonomous systems. We
to
shall now state, without proof, La Salle’s theorem for periodic systems.
Theorem Suppose the system [5.1(1)] is periodic; i.e., suppose
f(t, x) =f + (t T, x), V t, V X g R n
,
for some T> 0
Suppose V: R + X Rn — > R is a continuously differentiable p.d.f.,

with
90 V(t, x) = V(t + T, x), y X e Rn V t>0

,
Define
91 5 = {x g Rn : V(t, x) = 0, V t > 0}
Suppose V(t, x) < 0, V/>0, V x g Rn and that contains no
,
nontrivial trajectories of [5.1(1)]. Under these conditions, the equi-

librium point 0 is globally asymptotically stable.
92 Example. Consider once again the system of Example (20),
namely,
93 x 1
= —X 2
94 x2 = —f(x - g(*i) 2)
where / and g now satisfy the following conditions:

(i) / and g are continuous.
(ii) /( 0) = g( 0) a* 0; cr/(a) > 0, og{o) > 0 whenever o ^ 0.

(iii)
Jo ^(0 d\-+oo as | (7 1
— > oo.
We once again choose, as the Liapunov function candidate, the total
energy of the system, namely,
95 V( Xl ,x 2 ) = £ + £g(Od{
Conditions (i)— (iii) above ensure that V is a continuously differentiable
p.d.f., so that V is a suitable Liapunov function candidate for applying
Theorem (87). Calculating V, we get
96 K(x l5 x 2 ) = -x 2 f(x 2 )
Thus F(xj, x 2 ) < 0, V (xi, x 2) g i^
2
. Moreover,
97 S = {(x u x2) G R 2
: V(x u x 2 ) = 0} = {(x u x 2) G i?
2
: x2 - 0}
because of (ii) above. If any trajectory of (93)-(94) is contained in S ,

it
must have x 2 (0 = 0. This implies that (i) xx = constant =x 10 ,

by (93),
and (ii) x 2 (t) = 0. Now, (94) gives
98 -/(* 2 )-g(* 10 ) = 0, i.e., -g(x 10 ) = 0
because x 2 = 0. By (ii) above, this shows that x 10 = 0. In other words,

the only trajectory that lies entirely within S is the trivial trajectory
xx
=
x2 = 0. Thus all the conditions of Theorem (87) are satisfied,
whence 0 is a globally asymptotically stable equilibrium point.
99 Example. Consider once again the damped Mathieu equation of

Example (28), namely,
x
Chap. 5 Stability in the Sense of Liapunov

160
100 Xi =
101 x2 =— 2 + (2 + sin 0*1
This system is clearly periodic with period 2 n. Let us choose V as in
(32), i.e.,
xi
102 V(t, x t,
x 2) = xj + 2 sin t
-f-
Then a continuously differentiable p.d.f. and is periodic with

V is
period In. Therefore, V is a suitable Liapunov function candidate for

applying Theorem (88). Calculating V, we get
103
tV/.
V(t, x u x 2)
n
= -x .. 2
a
4 + 2 sin + cos t t
(2 + s in0
'
2
[see (35)]. Thus V< 0 for all t, (x„ x 2 ). Moreover, the set S defined
in (91) is given by
104 S = {(x u x 2 ): x2 = 0}
As in the previous example, if a trajectory of (100)-(101) is entirely
contained in S, it must have x 2 (t) 0, which in turn implies that x, =
is a constant [by (100)], and that
105 (2 + sin t)xj =0

Keeping in mind that x, is a constant, (105) implies that Xj 0. Hence =
the only trajectory entirely contained within S is the trivial trajectory
xx =
x2 =
0. Because all the hypotheses of Theorem (88) are met, it
follows that (0, 0) is a globally asymptotically stable equilibrium point.
5.2.3 Instability Theorems
In the two previous subsections, we presented sufficient conditions

for stability and for asymptotic stability. In this final subsection,
we
shall give some sufficient conditions for instability.
Theorem The equilibrium point 0 at time t0 of [5.1(1)] is unstable
if there exists a continuously differentiable decrescent function

V: R+ x R" — * R such that (i) V is an l.p.d.f., and (ii) V(t, 0) = 0,
and there exist points x arbitrarily close to 0 such that V(t 0 x) ,

> 0.
proof To demonstrate that 0 is an unstable equilibrium point, we

must
show that, for some e > 0, no 5 > 0 exists such that [5.1(7)] holds. Suppose
Vis an l.p.d.f. over some ball Br with r> 0. Because V is decrescent,
there
,
is a function /? belonging to class K such that

107 V(t, x) < J?(||x||), V t > t0 , V x G Bs for some ball B s
We shall show that, if we let € = min [r, s], then no matter how small
x
we choose 8 > 0, we can always find a corresponding x such that 0
|| 0 < 8 and ||s(f, x

1|
eventually equals or exceeds €. Thus given
0 , f 0 )||
any 8 > 0, pick x such that ||x < 8, and

0 V(t x ) > 0. In the (i) 0 ||
(ii) 0, 0
interests of brevity, let x(/) denote s (t, x0 ,

? 0 )* Now, as long as x(t) remains
within B we have
e, V[t, x(0] > 0, which shows that
108 V[t, x(/)] > V(t 0, x 0) > 0
The inequalities (107) and (108) together imply that ||x(/)|| is bounded
away from zero, which in turn implies, by the positive definiteness of V,
that V[t, x(01 is also bounded away from zero. Hence, within a finite
amount of time, V[t x(/)] will exceed the value /?(6), which means that
,
x(t) exceeds €. Thus 0 is an unstable equilibrium point.

[| 1 1 m
Note that, in contrast with previous theorems, the function V
in Theorem (106) can assume positive or negative values. Also, the
inequality (107), i.e., the requirement that V is a decrescent function
places no restrictions on the behavior of V(t, x) when it assumes nega-
tive values.
110 x t
=x —x +x
t 2 t
x2
111 X2 = —x — x\ 2

112 V(x l9 x 2 ) = ( 2x 1 — x2 ) 2 — x\
Even though V assumes both positive and negative values, it has the
property that it assumes positive values arbitrarily close to the origin.
Hence it is a suitable Liapunov function candidate for applying Theorem
(106). Differentiating V, we get
113 V(x l9 x 2 ) = 2(2x — x )(2x — x — 2x

t 2 1 2) 2 x2
= [(2xj — x + xj]( + x 2)
2
1 2)
Thus V is an l.p.d.f. over the ball B _ 5 for any 8 e (0, 1). Hence Fis a l
Liapunov function, and 0 is an unstable equilibrium point, by Theorem

(106).
114 Remarks: Some authors prove a less efficient version of Theorem

(106) by showing that 0 is a unstable equilibrium point if one can find
a function V such that both V and V are l.p.d.f.’s. Actually, it can be
shown that if one can find such a function V, then 0 is what is known
as a completely unstable equilibrium point; i.e., there exists an e >0
such that every trajectory x(«) other than the trivial trajectory x(t) =0
satisfies ||x(f)|| > e for some Such a hopeless situation rarely occurs
t.
in practice, and as a result, a theorem of this type is of very little use.

x
Theorems (115) and (126) give alternative sufficient conditions

for instability. They also have certain other advantages, which are
discussed later.
115 Theorem The equilibrium point 0 at time t„ of [5.1(1)] is unstable

if there exists a continuously differentiable decrescent function
V: R+ X R" —* R such that (i) V(t 0 0) ,
= 0, and V(t 0 x) assumes
,
positive values arbitrarily close to the origin, and (ii) V(t, x) is of the
form
116 V(t, x) = XV(t, x) + Vi(t, x)

where X > 0 is a constant, and V R x
: X Rn — > R satisfies the condi-
tion
117 V x {t, x) > 0, V t > t0 , V x e B r

for some ball B r
proof show that if we choose e = r, then [5.1(7)] cannot be

We
satisfied forany choice of 5 > 0. Thus, given 8 > 0, choose x 0 such that
(i) |0 < 8, and (ii) V(t 0 x 0 ) > 0. Let x(-) denote the resulting solution
]
] I ,
trajectory s(-, x 0 / 0 ) of [5.1(1)]. Because, whenever ||x(t)||

,
r, <
118 ^{V[t, x(/)]] = XV[t, x(l)] + V x
[t, x(f)] > XV[t, x(/)]
it follows that
u V[tMm> 0
119 ^-t {e-
Hence
120 V[t, x(t)] > V(t 0 , x 0 )
Hence V[t, x(f)] increases without bound, and therefore ||x(t)|| must even-
tually equal e. This shows that 0 is an unstable equilibrium point. H
122 *1 = *i + 2x + XiXi 2
123 x2 = 2x + x — x\xx 2 2
and let
124 V(x u x 2 ) = x\ — x\
Then Visa suitable Liapunov function candidate for applying Theorem
(115). Calculating V, we get
125 V(x u x 2 ) = — 2x 2 x 2 = 2x\ — 2x\ + Ax\x\
= 2V{x x ,
x2 ) + Ax\x\
Because 4x\x\ >0 V (x u x2 ) e R2 ,
it follows by Theorem (115)
that 0 is an unstable equilibrium point. Note that neither V nor V is

an l.p.d.f. in this example.
In Theorems (106) and (115), the function V is required to satisfy

certain conditions at all points belonging to some neighborhood of
the origin in R n In Theorem (126), the condition on V is only required
.
to hold in some region for which the origin is a boundary point (and
not an interior point). Theoretically, however, Theorems (115) and
(126) are equivalent, as shown later. Theorem (126) is generally known
as Cetaev’s theorem.
126 Theorem The equilibrium point 0 at time t 0 of [5.1(1)] is unstable

ifthe following conditions hold: There exist a continuously differ-
entiable function V: R + x R n —
R and a closed set containing Q
0 in its interior such that
(i) There is an open subset Qj of Q containing 0 on its boundary,
127 (ii) V(t, x) > 0, V t>t v 0, X e Q 1
128 V(t, x) = 0, V t > t0 V ,

X G (JQj\ n
boundary of
(the Q 1
in £2)
(iii) V(t, x) is bounded above in Q, uniformly in t, and
129 (iv) V(t, x) > y(|| x||), V t>t 0 ,y xe Q,
where y is a function belonging to class K.
proof Pictorially, the situation can be represented as in Fig. 5.7. In

view of the assumptions on V and V, it is clear that along any trajectory
starting in V increases indefinitely and cannot cross theset(?Qj\ Q.
Because V is bounded above on f2, the trajectory must eventually reach
the boundary of £2, regardless of its starting point in This means .
that 0 is an unstable equilibrium point. wm
FIG. 5.7
131 x i
= -f- x2 + x\
132 ±2 = *i — x2 +xx t 2

133 V(x u x 2 ) =x 1
x2
Let Q be the set
134 Q = {(jcj, x 2 ): x\ + x\ < 1}
and let Ofi be the subset
135 = {(x u x 2) e Q: x 1 > 0, x2 > 0}

Then conditions (i)-(iii) of Theorem (126) are satisfied. To verify condi-
tion (iv), we differentiate V to get
136 V(x u x 2 ) = x (x — x + x x + x (2x + x + xj)

1 1 2 x 2) 2 1 2
= x\ + x\ + 2x x + 2x\x >0, V (*i, x x 2 2 2) E £2i/{0}
Hence by Theorem (126), (0, 0) is an unstable equilibrium point.
5.2.4 Concluding Remarks
In this section, we have presented various theorems in Liapunov

stability theory. The key feature of all these theorems is that they enable
one to draw conclusions about the stability status of an equilibrium
point, without solving the system equationsby constructing a suitable ,
Liapunov function. The favorable aspects of these theorems are

1. There is no need to solve the system equations (as mentioned
above), and
2. The Liapunov function F, especially in the theorems on stabil-
ity and asymptotic stability, has an intuitively appealing
geometric interpretation as the energy of the system.

In fact, in many physical problems, the total energy stored within the
system is the logical candidate for a Liapunov function. This is brought
out in Examples (14) and (20).
The unfavorable aspects of the theorems given here are (1) they
present only sufficient conditions for various forms of stability, so
that if a particular Liapunov function candidate V fails to meet the
requirements on F, no conclusions can be drawn, and the testing process
must be begun once again, with a different Liapunov function candidate.
(2) In a general system of nonlinear equations, there is no systematic
procedure for generating Liapunov functions. Both of these are serious
drawbacks and have prevented the Liapunov theory from being more
widely used than it is.
In this context, it is worthwhile to note two points: First, Liapu-

nov presented his original theorems only as sufficient conditions.
However, later researchers have shown that the existence of a Liapunov
function is also a necessary condition for various forms of stability.
In particular, the existence of a function V satisfying the hypotheses of
Theorems (9), (45), (115), and (126) is a necessary as well as sufficient
condition for the equilibrium point 0 to have the stability status men-
tioned in those theorems. [This means, in particular, that Theorems
(115) and (126) are equivalent.] However, knowing that a Liapunov
function exists still does not help one to find it. As a result, the so-called
converse theorems, which establish the existence of a Liapunov function,
are mainly of theoretical interest.
The second point is that, even though in general it is a very
task to find a Liapunov function for a given system, there are
difficult
two special cases where the choice of a Liapunov function is relatively

easy, namely: (1) linear systems, and (2) “weakly” nonlinear systems.
Linear systems are thoroughly explored in Sec. 5.3, while weakly
nonlinear systems are studied in Sec. 5.4. In his original work, Liapunov
also derived a method whereby the stability of a nonlinear system can
be ascertained by linearizing the nonlinear system about an equilib-
rium point and studying the resulting linear system. This method is
known as Liapunov’s method and is covered in Sec. 5.4.
first or indirect
Many authors present the indirect method before the direct method,
but we prefer to derive the indirect method as a consequence of the direct
method and of the Liapunov theory for linear systems.
Problem 5.9. A phase-locked loop in communication networks can be

described by
y(t) + [a + b cos y(t)]y(0 + c sin y(t) = 0
(a) Transform this equation into state variable form by choosing x x =

y, *2 = y.
(b) Suppose > 0. Using the Liapunov function candidate
c
V(x 1 ,
x2 ) = c(l — cos Xi) + x\\2
show that (0, 0) is a stable equilibrium point if a > b > 0, and that (0, 0) is
an asymptotically stable equilibrium point if a > b > 0. (Hint: In the latter

case, use LaSalle’s theorem.)
Problem 5.10. Consider the differential equation
m+f[y(t)]Xt)+g[y(t)] =0
(a) Transform this equation into state variable form by choosing x t =
y, *2 =X
(b) Suppose that, in some interval [— cr 0 , <7 0 ], we have

erg (a) > 0 whenever o e [— <7 0 ,
<7 0 L o ^0
Generalize the Liapunov function candidate of Problem 5.9 to the present
system. Show that (0, 0) is a stable equilibrium point if f(p) 0 >
V o e [— (J 0 tfo], and that (0, 0) is an asymptotically stable equilibrium
,
point if f(a) >0 V <7 e [— <7 0 ,

<7 0 ].
Problem 5.11. Consider the system
Xi = x t + 2x\
x2 = 2x x 2 x + x\
Using the Liapunov function candidate
V(x u x 2 ) = x\ — x\
show that (0, 0) is an unstable equilibrium point.
Problem 5.12. Consider the system
xt = x\ — x 1 x2
x2 = — x\ — 2x t x 2
using the Liapunov function candidate
V(x u x 2 ) = x 1 (x 1 - x2 )
and Theorem (115), show that (0, 0) is an unstable equilibrium point.
5.3
STABILITY OF LINEAR SYSTEMS
In this section, we shall study the Liapunov stability of systems described
by linear vector differential equations. The results presented here not
only enable us to obtain necessary and sufficient conditions for the
stability of linear systems but also pave the way to deriving Liapunov’s
indirect method (which consists of determining the stability of a non-
linear system by linearization).
5.3.1 Stability and the State Transition Matrix
Consider a system described by the linear vector differential

equation
x(0 = A(t)x(t), t >0

The system autonomous if A(?) is constant as a function of time;
(1) is
otherwise it is nonautonomous. It is clear that 0 is always an equilibrium
point of the system (1), at any time t 0 0. Further, 0 is an isolated >

equilibrium point at time t Q if A(/) is nonsingular for some t t0 > .
1
S ec. 5.3 S tability of Linear S ystems 167
The general solution of (1) is given by (see Appendix II)
2 x(0 = t 0 )x(t 0 )
where <!>(•, •) is the state transition matrix associated with A(«) and
is the unique solution of the equation
4 O(t 0 , to) =I
With the aid of this explicit characterization for the solutions of (1),
it is possible to derive some useful conditions for the stability of the
equilibrium point 0. Because these conditions involve the state transi-
tion matrix <D, they are not of much computational value, because it is
in general impossible to derive an analytical expression for . Never-

theless, they are of conceptual value, enabling one to understand the
mechanisms of stability in linear systems.
5 Theorem The equilibrium point 0 at time tQ is stable at time tQ
ifand only if 9
6 sup
t>to
1 1
t0) 1 1,
A m(t < oo 0)
proof
(i) “If” Suppose (6) is true, and let e 0 be specified. If we define >
S(e, t 0 ) as e/m(t 0 ), then, whenever ||x(f 0 )|| <5(6, t 0 ), we have <
7 1 1
x(/) 1|
= ||
<D (t, t 0 )x(t 0 ) 1 1
< 1 1
<D (/, f 0 )IHI x(f 0 ) 1
<m(t 0 )-d(€,t 0 ) =€
so that [5.1(7)] is satisfied. This shows that the equilibrium point 0
is stable at time t0 .
“
(ii) Only If” Suppose 0(- / 0 ) is an unbounded
(6) is false, so that 1
1
, 1 1/
function. We shall show

an unstable equilibrium point. that 0 is
Accordingly, let € > 0 be any given number; we show that, no

matter how small we choose 8 > 0, it is always possible to find an
x(f 0 ) with xô) < 8 such that ||x(/)|| € for some t t Q Thus
|| II
> > .
suppose 8 > 0 is fixed. Because ||<D(«, / 0 )||, is an unbounded func-

tion, there exists a t>t 0 such that
8 l|3>(Mo)ll,>|^
where <5i < <5 is some positive number. Next, select v to be a vector
of norm 1 such that
9 ll$(Mo)v|| HI$(Mo)lli
9 Recall that ||
• ||,- denotes the induced norm of a matrix.
*
This is possible in view of the definition of the induced matrix norm

(definition [3.2(5)]). Finally, let x(t 0 ) = St\. Then ||x(f 0 )|| < 5;
moreover,
10 l|x(01i = ||®(Mo)x(fo)ll = ll<Mi>0,fo)v||

= ||®(Mo)ll*> e
<5 1
Hence the equilibrium point 0 is unstable. H

11 Theorem With regard to the system (1), the following three condi-
tions are equivalent:
(i) The equilibrium point 0 is asymptotically stable at time t 0 .
(ii) The equilibrium point 0 is globally asymptotically stable

at time t0 .
12 (iii) lim,_ ||<D(7, t 0 ) ||,

= 0.
proof (ii) => (i) Obvious.
(iii) Suppose (12) holds. Then, because <!>(•, f 0 ) ||* is continuous
(ii) 1
1
—
and approaches 0 as t > oo, it follows that 1®(- / 0 ) k is bounded over the 1 , I
interval [r 0 oo). Hence, by Theorem (5), the equilibrium point 0 is stable

,
at time t 0 Furthermore, for arbitrary x(r 0 ) e R

n
. we have ,
l|x«l|<l|0(Mo)IH|x(f 0 )l| — >0 as / — * co
so that x(t )
— » 0 as t — oo. This shows that the equilibrium point 0 is
globally asymptotically stable at time t0.
(i) => (iii) Suppose (iii) is false. Then at least one component of
<D(f, t 0 ), say 0 l7 (/, t 0 ), does not approach zero as t — > oo. Let x(/ 0 ) = dej,
where e is the yth elementary unit vector (i.e., e y has all zero components,
except for the yth component, which is 1). Then, regardless of how small we
make 5, it is clear that the /th component of the corresponding solution
x(-) does not approach 0 as t — > oo. Thus, if (iii) is false, then there are
trajectories starting from initial conditions arbitrarily close to 0 that do
not converge to 0. Thus (i) is also false. mi
13 Theorem With regard to the system (1), the following two state-
ments are equivalent.

(i) The equilibrium point 0 is stable at time t0 .
(ii) The equilibrium point 0 is stable at all times tx > t0 .
proof (ii) => (i) Obvious.

(i) => (ii) Suppose (i) is true. By Theorem (5), this implies that
14 sup||®(f,f 0 )l|f < 00

t>t 0
Now, suppose to>t x . By the semigroup property of the state transition

matrix (see Appendix II), we have
15 •(/,/i) -Q&toymtuto)]- 1
S ec. 5.3 S tabi/ity of Linear Systems 169
Hence
ll®& *i)ll« < ll®0> MIHIflKn, ?o)]

-1
|l/
However, because t 1 and t 0 arc finite numbers, |[[®(f,, _1

f 0 )] II; is also a
finite number. This fact, together with (14), shows that
sup
t>t i
|| ©ten) || f < oo
Hence the equilibrium point 0 is stable at time tx . Because this argument

can be repeated for any t u (ii) is established.
m,
Remarks: Theorem (13) does not say that stability of time t0 is

equivalent to uniform stability over [t Q
| ,
oo). Rather, Theorem (13)
isto be contrasted with lemma [5.1(28)]. The latter shows that, for
any system, stability of the equilibrium point 0 at some time t
0
implies stability at all times prior to t0 . In contrast, Theorem (13)
shows that, in the special case of linear systems, stability at time tQ
implies stability at all times after t0 . The key point in the proof of
Theorem _1
(13) is that || [<!>(£, £ 0 )] IL-
is finite. In a general nonlinear
system, the transition map x 0 i-> s(/
1? x0 ,
* 0 ) is bounded, but (1) it
may not have an inverse, and (2) even if it has an inverse, the inverse
may not be bounded. Thus Theorem (13) is a special result for linear
systems. 10
The necessary and sufficient conditions for uniform stability

are now given.
Theorem The equilibrium point 0 is uniformly stable over the

interval [0, oo) if and only if
sup m(t 0 ) = sup sup ||
®(f, t0) \\i^m 0 < oo
to>0 t0 >0 t>t 0
Remarks Recall, from Theorem (5), that a necessary and sufficient

:
condition for the equilibrium point 0 to be stable at time t is that

0
m(t 0 ) is a finite number. From Theorem (13), if m(t 0 ) is finite for
some value of t 0 it is finite for all values of t 0 However, it is still
, .
possible for m(t 0 ) to be an unbounded function of t0 . See Problem

5.2.
prodf
(i) “If” Suppose m 0 is finite; then, for any e > 0, [5.1(8)] is satisfied
with 5 =
€lm 0 .
10 Theorem
(13) also holds in the case of nonlinear systems of the form ±(t)
= X (0L where f is continuous in t and Lipschitz-continuous in x; this is because
in this case also the map x 0 s(*i, x0 ,
tQ ) has a bounded inverse; see Theorem
[3.4(50)].
>
(ii) “ Only If” Suppose m(t 0 ) is unbounded as a function of t 0 Then at .
least one component of <!>(•, •), say 0 O (-, •), has the property that
sup |
(f>ij(t, f 0 )l is unbounded as a function of t0
t>to
Let x 0 = ey-, the yth elementary vector. Then, because of (22), the
ratio ||®(f, f 0 )xo||/||x 0 ||
cannot be bounded independently of t0.
Thus the equilibrium point 0 is not uniformly stable over [0, oo). —
Theorem The equilibrium point 0 of (1) is uniformly asymptotically
stable over [0, oo) if and only if
sup sup ||®(7, t 0 )\\t< OO
fo>0 t>t o
ll®(Mo)lli—>0 as t — oo, uniformly in t0
Remarks: The condition (25) can be equivalently expressed as

follows: For each e > 0, there exists a T(e) < oo such that
\m, f 0 )ii, <€, v / > + t(€), vt >o

t0 0
proof
” then the equilibrium point 0
(i) “ If By Theorem (19), if (24) holds,
is uniformly stable over [0, oo). Similarly, if (25) holds, the ratio
1 1
cD(f, 1 0 )x 0 1 1/| |
xo 1 1
approaches zero as t — oo, uniformly in 1 0 and
x 0 Thus both conditions of definition

. [5.1(14)] are satisfied, whence
0 is uniformly asymptotically stable over the interval [0, oo).
(ii) “ Only If
”
This part of the proof is left as an exercise. M
Theorem (27) below gives an alternative equivalent condition for
uniform asymptotic stability.
Theorem The equilibrium point 0 of (1) is uniformly asymptotically

and only if there exist positive
stable over the interval [0, oo) if
constants m and A such that
||®(*, OIL* < me-*-", V t > t0 , V t0
Remarks authors refer to a system satisfying (28) as being

: Some
exponentially stable because, if (28) holds, then the solutions of (1)
,
decay exponentially. The import of Theorem (27) is thus to show that

exponential stability is equivalent to uniform asymptotic stability
over [0, oo).
proof
(i) “If” Suppose (28) holds. Then clearly (24) and (25) also hold, so
that the equilibrium point 0 is uniformly asymptotically stable.
(ii) “ Only If” Suppose (24) and (25) hold. Then there exist m and T
such that
r
Sec. 5.3 Stability of Linear Systems 171
29 \\®(t,ti)\\i<m 0 , V t>t u V ti
30 ||®(Mi)llf< Y> V‘>ti+T, Vf,
In particular, (30) implies that
31 ||0(ri + r, rOII,- < 4"’ V*i
Now, given any t0 and any t > *o> pick k such that tQ + kT < < to
t
+ kT -f T. Then we have
32 <D(, } /o ) = 0(/, to + kT)-Q(t 0 + kT ,

f0 + ftr - J)
1
. . . <D (/ 0 + 7 , ô)
Hence
33 ||0(f, r 0 )ll( < to + kT)\\rh\mo + jT, to + jT — T)||,-
where the empty product is taken as 1. Substituting from (29) and

(31) in (33), we get
<m -2~ k = (2m 0 )2~ (k+v < 2m -2- { , - , >),T

- '
34 ||0(r, r 0 )||i 0 0
Hence (28) is satisfied, if we choose

35 m — 2m 0
A _
, log 2
36 rj-
In conclusion, in this subsection we have presented several results

that relate the state transition matrix of a linear system to its stability
properties. These results are not of much use for testing purposes,
because finding the state transition matrix of a linear system is in general
a formidable task. However, some conceptual insight is provided by
these results. For example, Theorem (27), which shows that exponential
stability is equivalent to uniform asymptotic stability, is not very obvious
on the surface.
5.3.2 Autonomous Systems
Throughout this subsection, we shall restrict ourselves to linear
autonomous systems, i.e., systems described by
37 x(/) = A x(t)
In this special case, Liapunov theory is very complete, as demonstrated
by the following series of theorems.
38 Theorem The equilibrium point 0 of (37) is (globally) asymptotically

stable if and only if all eigenvalues of A have negative real parts.
The equilibrium point 0 of (37) is stable if and only if all eigenvalues

of A
have nonpositive real parts and every eigenvalue of A having
a zero real part is a simple zero of the minimal polynomial of A.
proof From Appendix II, the state transition matrix 0(r, t0) of the
system (37) is given by
39 <D(f, t0) = exp [A(/ - f 0 )l
where exp[A(-)] is the matrix exponential of A. Furthermore, exp (AO

can be expressed as
r mi
40 exp (AO = S S tJ ~ 1
exp (A/O^/A)
i=l 7=1
where r is the number of distinct eigenvalues of A, A 1} . . . ,

A r are the
(distinct) eigenvalues of A, and m t is the multiplicity of A,- as a zero of
the minimal polynomial of A, and ptj are the interpolating polynomials.
The stated conditions for stability and asymptotic stability now follow
readily by applying Theorems (5) and (11). B—
Thus, in the case of linear time-invariant systems of the form (37),

the stability status of the equilibrium point 0 can be determined by
studying the eigenvalues of A. However, it is possible to formulate an
entirely different approach to the problem, namely by the use of quad-
ratic Liapunov functions. This theory is of interest in itself and is more-
over useful in the study of the first method of Liapunov (the indirect
method).
Given the system (37), the idea is to choose a Liapunov function
candidate of the form
41 V(x) = x'Px
where P is a real symmetric matrix. Then V is given by
42 V(x) = x'Px + x'Px = x'(A'P + PA)x

= — x'Qx
where
43 A'P + PA = -Q
Equation (43) is commonly known as the Liapunov matrix equation.
By means of this equation, it is possible to study the stability properties
of the equilibrium point 0 of (37). For example, if a pair of matrices
(P, Q) satisfying (43) can be found such that both P and Q are positive
definite, then both V and — V are positive definite functions. Hence, by
Theorem equilibrium point 0 of (37) is globally asymptot-
[5.2(58)], the
ically stable. Alternatively, if a pair (P, Q) can be found such that
Q
is positive definite and P has at least one negative eigenvalue, then —V
is a positive definite function, and — V assumes positive values arbi-
trarily close to the origin. Hence, by Theorem [5.2(106)], 0 is an unstable

equilibrium point of (37).
This, then, is the rationale behind studying equation (43). There
are clearly two possible ways in which (43) can be tackled: (1) Given
a particular matrix A, one can pick a particular matrix P and study
the properties of the resulting matrix Q, or (2) given A, one can pick
Q and study the resulting P. We shall pursue the latter approach,
for two reasons —
one pragmatic and the other philosophical. The prag-
matic reason is that the second approach is the one for which the theory
is better developed. On a more philosophical level, we can reason as
follows: Given a matrix A, we presumably do not know ahead of time

whether 0 is a stable or unstable equilibrium point. If we pick P and
study the resulting Q, we would be obliged (because of our available
stability theorems) to make an a priori guess as to the stability status
of 0 (we would pick P to be positive definite if we thought 0 were asymp-
totically stable and P to have some negative eigenvalues if we thought
0 were unstable). On the other hand, if we were to pick Q, we could
noncommittally choose Q to be positive definite and study the resulting
P If P were to be positive definite, 0 would be asymptotically stable
:
(by Theorem [5.2(58)]), while if P were to have some negative eigen-

values, 0 would be unstable (by Theorem [5.2(106]).
One difficulty with selecting Q and trying to find the correspond-
ing P is that, depending on A, (43) may not have a unique solution
for P. Theorem (44) gives necessary and sufficient conditions for (43)
nxn
to have a unique solution for P corresponding to every Q e R .
44 Theorem Let A e
and let [X l9 RnXn
2„} be the (not necessarily
. . . ,
distinct) eigenvaluesof A. Then (43) has a unique solution for P

corresponding to every Q e R nXn if and only if
45 Xt + XfÔ, V Uj
(where * denotes complex conjugation).
The proof of this theorem is not difficult but requires some con-
cepts from linear algebra not heretofore covered, and the reader is
therefore referred to [7] Chen for the proof.
Using Theorem (44), we can state the following corollary.
46 corollary A necessary condition for all eigenvalues of A to have nega-

tive real parts is that (43) has a unique solution for P corresponding to
each Q e R^ 71
.
proof If all eigenvalues of A have negative real parts, then (45) is
satisfied. _
The following lemma provides an alternative characterization of

the solutions of (43).
47 lemma Let A be a matrix whose eigenvalues all have negative real

parts. Then the unique solution for R of the equation
48 A R + RA = -M
where M is a given real symmetric matrix, is
49 R= r eA '
1
dt
Jo
proof Consider the linear matrix differential equation
50 S(0 = A'S(0 + S(f)A + M; S(0) =0

Because the right-hand side of (50) is affine in S, and is thus globally
Lipschitz-continuous in S, it follows from Theorem [3.4(25)] that (50) has
exactly one solution. We shall now show that this solution is in fact given
by
r e A r'Me Az
'
51 S it) ) dr
Jo
Differentiating both sides of (51), we get
52 S(0 = e A ‘Me At
'
On the other hand,
53 A'S(0 + S(t)A = f A'e A T Me Ax dr

'
+ [
eA
'
T
Me ATA dr
Jo Jo
= f (A'e A
'
r
Me Ar + eA
'
T
MAe Ar ) dr
£
^(e A,T Me AT ) dr = e
Aft
Me At - M
/
= s(0 M
so that (50) is satisfied.
Now, because all eigenvalues of A have negative real parts, we see from
(52) that S(0 — » 0 as t —* co. Also, as t — > oo, the integral in (51) remains
well defined and in fact approaches R as given in (49). Thus, as t — > oo,
in the steady state, (50) reduces to
54 0 = AR + RA + M
which is equivalent to (48). Hence (49) provides a solution to (48). By
Theorem (44), it is in fact the unique solution. ^
We can now state one of the main results for the Liapunov matrix
equation.
Sec. 5.3 S tability of Linear S ystems 175
55 Theorem Given a matrix A e R nXn , the following three statements

are equivalent:
(i) All eigenvalues of A have negative real parts.
(ii) There exists some positive definite matrix Q e Rnxn such
that (43) has a unique solution for P, and this solution is
positive definite.
(iii) For every positive definite matrix Q e Rnxn , (43) has a
unique solution for P, and this solution is positive definite.
proof (iii) => (ii) Obvious.

(ii) => (ii) is true for some particular matrix Q. Then we
(i) Suppose
can apply Theorem [5.2(58)] with the Liapunov function V(x) = x'Px to
conclude that 0 is an asymptotically stable equilibrium of (37). Thus, by
Theorem (38), it follows that all eigenvalues of A have negative real parts.
(i) => (iii) Suppose all eigenvalues of A have negative real parts. Then
by lemma (47), corresponding to each Q e Rnxn (43) has the unique solu- ,
tion for P given by
56 P = f°° e A>t Qe At dt
Jo
It only remains to show that P is positive definite whenever Q is positive

definite. Thus suppose Q is positive definite, and express Q in the form
Q= M'M, where M
e Rnxn is nonsingular. We show that P is positive
definite by showing that
57 x'Px >0, V x e Rn
58 = 0 => x = 0
x'Px
With Q = M'M, P becomes
P = eA
'
59 f
r
M'Me Ax dr
Jo
Thus, for any x e Rn we have

,
60 x'Px = f x'e A
'
r
M'M^ AT x c/t = f \\Me Ar x\\l dr >0
Jo Jo
where ||
•
|| 2 denotes the Euclidean norm. Also, (60) shows that if x'Px = 0,
then
61 Me Atx = 0, V t>0
If we take t = 0 in (61), we get Mx = 0, which implies that x = 0 because
M is nonsingular. Thus P is positive definite and (i) => (iii). H
62 Remarks :
1. Theorem (55) is very important in that it enables one to un-

ambiguously determine whether or not 0 is an asymptotically
stable equilibrium point, in the following manner: Given
A g Rnxn ,
pick Q g R nxn to be any positive definite matrix
(a logical choice is Q=I
or some other diagonal matrix),
and solve (43) for P. Suppose that (43) has no solution or that
it has more than one solution (i.e., it does not have a unique
solution); then 0 is not asymptotically stable. Suppose that

(43) does have a unique solution for P and that this solution
is not positive definite; then once again 0 is not asymptotically
stable. On the other hand, if this unique solution for P is posi-
tive definite, then 0 is an asymptotically stable equilibrium
point of (37).
2. Theorem (55) states that whenever Q is positive definite (and
all eigenvalues of A have negative real parts) the corresponding
P is also positive definite. It does not say that whenever P
is positive definite the corresponding Q is positive definite.
The latter statement is false in general.
3. As a computational device, Theorem (55) can be compared to
Routh’s criterion, which enables one to determine whether or
not the zeroes of a given polynomial all have negative real
parts without actually computing the zeroes. In thesame way,
Theorem one to determine whether or not the
(55) enables
eigenvalues of a given matrix all have negative real parts
without actually computing the eigenvalues. If A is an n X n
matrix, then (43) represents a system of n(n + l)/2 equations
in the same number of unknowns (note that P and Q are sym-
metric). Thus, for small values of n, there seems to be no advan-
tage to using the Liapunov matrix equations as a testing device.
However, for intermediate values of n (say 6 <n< 15),
Theorem (55) provides a useful tool. See [15] Jury for alternative
computational techniques for testing whether or not all the
eigenvalues of a given matrix have negative real parts.
11
63 Example . Let
and let
= P i Pi
P
.Pi Pi
1
!To avoid getting lost in an avalanche of numbers, we shall illustrate the use
of the Liapunov matrix equation only for second-order systems, even though for
n — 2 this is the hard way to do it.
be the matrix of unknowns. Then (43) becomes

—2p 1
—p 2 - 2pi—Pi —2pi — Pi Pi +P2
. Pi + Pi Pi+Pi - 2 Pi.
— Pi Pi +P 3 _
“
"-1 0
0 -1
which translates into the three equations
— 4Pi~2p = -1 2
Pi — Pi~ Ps = 0
2Pi + 2/7 = —1 3
The solutions of these equations are
Pi = Pi = —2
Thus
- 1/2
P=
3/2
Even though (43) has a unique solution in this case, the resulting matrix
P is clearly not positive definite. Hence not all eigenvalues of A have
negative real parts.
64 Example. Let
- "
'-2 -3“ 2 0 Pi Pi
A= , Q= P=
1 2_ _0 2_ . Pi P3_
Then (43) reduces to the three equations
- 4Pi + 2p 2 = -2
— 2>Pi + Pi — 0
— 6p 2 + 4/7 = —2 3
Adding 4 times the second equation and —3 times the first equation,
we get
—6p 2 + 4/7 = 3 6
which is clearly inconsistent with the third equation. Thus (43) has no
solution in this case. Thus we conclude that not all eigenvalues of A
have negative real parts.
65 Example. Let
n -n n on
2
Then (43) has a unique solution for P, given by
42 -i r
-11 5
Because P is positive definite, the conclusion is that all eigenvalues of

A have negative real parts.
Theorem (55) shows that if the equilibrium point 0 of (37) is
asymptotically stable, then this fact can be ascertained by choosing a

quadratic Liapunov function and applying Theorem [5.2(58)]. The
following result, which is stated without proof, enables us to conclude
that if the equilibrium point 0 of (37) is unstable because some eigen-
value of A has a positive real part, 12 then this fact can also be ascer-
tained by choosing a quadratic Liapunov function and applying Theo-
rem [5.2(106)].
66 Theorem Consider (43), and suppose (45) is satisfied, so that (43)

has a unique solution for P
corresponding to each Q. Under these
conditions, if Q is positive definite, then P has as many negative
eigenvalues as there are eigenvalues of A with positive real part.
proof See [21] Taussky. M

Remarks :
1. Note that the hypotheses of Theorem (66) rule out A having

any eigenvalues with zero real part.
2. The significance of Theorem (66) lies in that it enables one to
prove the instability of (37) using a quadratic Liapunov func-
tion. Specifically, suppose the matrix A satisfies the hypotheses
of Theorem (66) and that V is defined by
67 V(x) = — x'Px
Then
68 V(x) = x'Qx
Moreover, V and V assumes positive values arbi-
is a p.d.f.,
by Theorem (66), P has at least one

trarily close to 0, because,
negative eigenvalue. Hence, by Theorem [5.2(106)], the equi-
librium point 0 of (37) is unstable. Moreover, it turns out that
a similar argument can be used in Liapunov’s indirect method.
1
The equilibrium point 0 of (37) can be unstable in another way, namely,
all eigenvalues of A have nonpositive real parts, but some eigenvalue of A having a
zero real part is a multiple zero of the minimal polynomial of A.
5.3.3 Nonautonomous Systems
In the case of linear time-varying systems characterized by (1),

the stability status of the equilibrium point 0 can be ascertained, in
theory at least, by studying the state transition matrix. This is detailed
in Sec. 5.3.1. The purpose of the present subsection is twofold:
1. To present some simple sufficient conditions for stability,
asymptotic stability, and instability based on the concept of
the matrix measure, and
2. To prove the existence of a quadratic Liapunov function for
uniformly asymptotically stable linear systems.
We shall begin by studying some stability conditions based on the
matrix measure. The following result proves useful for this purpose.
lemma With regard to (1), the following inequalities hold:
[[' -Ad-A(T)]*} £||®(/,/ 0 )||,
< exp A-[A(t)] rfrj

{£
is any norm on R and
n
where ||
•
|| ||,, ^(-),
respectively, denote the cor-
||
•
responding induced matrix norm and matrix measure on R nxn .
proof Immediate from Theorem [3.5(1)]. m

Many simple stability conditions can be derived from lemma
(69). The proofs are left as exercises and are straightforward applica-
tions of results in Sec. 5.3.1.
lemma The equilibrium point 0 of (1) is stable at time t0 if there

exists a finite constant m(t 0 ) such that
f A([A(t>] dx <C m(t 0 ), V t tr\

Jtn
0 is a uniformly stable equilibrium point over [f 0 , oo) if there exists a finite

constant m 0 such that
J‘
A[A(t)] dx <w 0, V t>t „ V ti >t 0
lemma The equilibrium point 0 of (1 ) is asymptotically stable at time
P jU,[A(t)] .
-co as t —> oo
Jta
0 is uniformly asymptotically stable over [t Q , oo) if, for every M> 0, there
exists a T< oo such that
rh+t
/*t[A(T)] dr < — Af, V t>T, V ti > \
Jti
>
lemma The equilibrium point 0 of (1) is unstable at time / 0 if
f
Jto
— fii[— A(x)] dx — > oo as t — oo
Remarks:
1. As we saw and 3.5, the measure of a matrix is
in Secs. 3.2
strongly dependent on the vector norm on Rn that is used to
define the matrix measure. In lemmas (71), (74), and (77), the
indicated conclusions follow if the appropriate conditions hold
for some matrix measure //,(•)• Thus there is a great deal of
flexibility in applying lemmas (71), (74), and (77), because one
is free to choose the matrix measure ///•)> and a wise user will
exploit this flexibility to the fullest. The examples that follow

illustrate this point.
2. The lemmas above provide only sufficient conditions

three
that areby no means necessary. However, they do have the
advantage that they are easy to verify and do not require the
determination of the state transition matrix.
3. Lemma (77) is of rather dubious value because its hypothesis
actually ensures the complete instability of the equilibrium
point 0; i.e., every nontrivial trajectory “blows up.”
Example. Consider the system (1) with
“cos t —sin f
A(0 =
_sin t cos t
and apply the measure /z i2 . In this particular case

”2 cos t 0
A(0 + A'(0 0 2 cos t
Because = 2max (A + A')/2 (see Sec. 3.2), we have

PnlMf)] = cos t
f A(t)] dx = sin — sin

ji i2 [ <2, V > Vt t0 t t0, t0 >0
Jto
Hence, by lemma (71), the equilibrium point 0 is uniformly stable over

the interval [0, oo).
Example. Consider the system (1), with
-2 + 3e-'~
-3
and choose the matrix measure ju n ( •). It is easy to see that, as t —
-> oo,
A (t) approaches a constant matrix A(oo) and that ^ n [A(°o)] =— 1.

: 0
Because the matrix measure is a continuous function, it follows that

HiAMt)] — > -1 as t — » oo. Hence (one can show that) (76) is satisfied,
so that, by lemma (74), the equilibrium point 0 is uniformly asymptot-

ically stable over [0, oo).
Note that if we had employed

the matrix measure ju io (.) in this
example, lemma (74) would have been inapplicable. Instead, we would
have had to settle for lemma (71) and a conclusion of uniform stability
over [0, oo) (verify this statement).
The Existence of Quadratic Liapunov Functions
In Sec. 5.3.2 it is shown that if the system (37) is asymptotically

stable, then a quadratic Liapunov function exists for this system.
We shall now prove a similar result for uniformly asymptotically
stable systems of the form (1). This result is derived from using the fol-
lowing lemmas.
82 lemma Suppose Q(*)‘ R+ — > RnXn is continuous and bounded and

that the equilibrium point 0 of (1) is uniformly asymptotically stable over
[0, oo). Then, for each t > 0, the matrix
83 P(r) = J
0'(t, 0Q(t)(t, t) dz
is well defined; further, P (t) is bounded as a function of t.
proof The hypothesis of uniform asymptotic stability implies [because

of Theorem (27)] that there exist positive constants m and A such that
84 ||€>(t, OIL- < me~ x{r ~ t

\ V T > t
The bound (84), together with the boundedness of Q(«), proves the
lemma. _
85 lemma Suppose that, in addition to the hypotheses of lemma (82),
we also have the following
(i) Q(0 is symmetric and positive definite for each t 0; moreover, >
there exists a positive constant a such that
86 ax'x < x'Q(0x, V x G Rn V ,

t >0
(ii) The matrix A(0 in (1) is bounded with respect to t; i.e., there exists
a finite constant m 0 such that
87 ||A(0IL-2 <m 0i v t> 0

(where ||
•
|| /2 denotes the /2 -induced matrix norm). Under these
conditions, the matrix P(/) given by (83) is positive definite for each
t> 0; moreover, there exists a positive constant j} such that
88 fix'x < xT(0x, V x e R V >0 n

, t
proof Let x e Rn ,
and consider the triple product x'P(/)x. From (83),
we get
89 x''(t, 7)Q(?)0(t, t)x dr
s'(t, x, t)Q(t)s(r, x, t) dr
where s(t, x, t) denotes (as before) the solution of (1), as evaluated at time
r, corresponding to the initial condition x at time t. Now, (86) and (89)
together imply that
90 xV(t)x >aJ ||
s(t, x, t) 1 1| dr
where ||
•
|| 2 is the Euclidean norm on Rn Now, by Theorem
. [3.5(1)], we
have
91 ||s(t, x, OII 2 > ll

x lk exp
j-J
jU i2 [-A(0)] tfflj
> ||z|| a exp

{-J'
||A(0)|| i2 rftfj by (ii) of [3.2 (19)]
>||x|| 2 exp[-w 0 (T — 0] by (87)
Substituting from (91) into (90) gives
92 x'P(/)x >a f x'xe~ 2mo{r

~ t)
dr
xx
2m o
The inequality (88) now follows by taking /? = a/2m 0 .
93 Theorem Suppose Q(-) and A(«) satisfy the hypotheses of lemmas

(82)and (85). The the uniform asymptotic stability of the equilibrium
point 0 of (1) can be established by selecting the Liapunov function
94 V(t, x) = x'P(r)x
where P(-) is given by (83), and applying Theorem [5.2(45)].
proof With V(t, x) defined by (94), we have

95 V(t, x) = x'[P(t) + A'(OPO) + P(/)A(0]x
It is easily verified by simple differentiation of (83) that
96 P (t) = -A'(/)P(0 - P«A(0 - Q(0

Hence
97 V(t, x) = — x'Q(/)x
Thus the functions V and V satisfy the various hypotheses of Theorem
[5.2(45)]. _
Remarks: Theorem (93) can be thought of as the extension, to the

nonautonomous case, of Theorem (55). In short, Theorem (93) states
that if the equilibrium point 0 of (1) is uniformly asymptotically
stable over [0, oo), then this fact can be deduced by applying Theorem
[5.2(45)] with a quadratic Liapunov function. However, whereas
Theorem (55) also provides a useful testing procedure for asymptotic
Theorem (93) does not have any value as a testing procedure.
stability,
The utility of Theorem (93) lies in the fact that it provides the basis
for Liapunov’s indirect method (linearization), as applied to non-
autonomous systems.
Some Specialized Results
We shall close this subsection by presenting a few specialized

results for nonautonomous systems.
First, suppose that the matrix A(t) in (1) is periodic in t. In this
case, we know
(by Theorems [5.1(57)] and [5.1(45)]) that the stability
at time 0 of the equilibrium point 0 is equivalent to the uniform stability
of 0 over [0, oo) and that the asymptotic stability at time 0 of 0 is equi-
valent to the uniform asymptotic stability of 0 over [0, oo). Further, it
is possible to test for the stability and/or asymptotic stability of 0 by

examining a constant matrix. This is brought out in the following.
98 lemma Suppose A(-) in (1) satisfies
99 A(/+ T) = A (0, V t >0

for some T > 0. Then the corresponding state transition matrix ®(f, t 0)
can be expressed as
100 •(/, t 0) = V(t, t0) exp [M(t - t Q )]
where
101 *¥(t + T, t 0) = ¥(*, t 0 ), V t >0

and M is a constant n x n matrix.
proof Define
102 R= O(/ 0 +T ; to)
and choose M such that

103 R- exp (M T)
We claim that *F(r, t 0 ) defined by
104 'VO, to) = <1>(t, to) exp [— M(/ - ?o)]

—
satisfies (101). To see this, we proceed as follows:
105 ¥(? + T, t 0) = ®(/ + T, f„)-exp [-M0 + T - f 0 )l
= + T, to + T) &(to + T ) t
,
10 • exp ( M.T)
•exp [-M {t - / 0 )1
However, because of the periodicity of A(*), we have
106 + T,t 0 + T) = t Q)
Hence (105) simplifies to
107 (/ + T ,
t 0) = ®(f, /0 )-exp [— M(f - t 0 )] = V( t, / 0 )
108 This establishes (101). M

Once this representation (100) for the state transition matrix is
obtained, the results of Sec. 5.3.1 can be used to obtain necessary and
sufficient conditions for stability. These are given next.
Theorem Consider the system (1), and suppose A(«) satisfies (99).
Then the equilibrium point 0 of (1) is uniformly stable over [0, oo)
ifand only if all eigenvalues of have nonpositive real parts, and M
any eigenvalue of M
having a zero real part is a simple zero of the
minimal polynomial of M. The equilibrium point 0 of (1) is uniformly
asymptotically stable over [0, oo) if and only if all eigenvalues of M
have negative real parts.
We turn now to so-called slowly varying systems. Consider the

differential equation (1), and suppose that for each fixed t the eigen-
values of the matrix A(t) all have negative real parts. One might ask
whether this is enough to ensure the asymptotic stability of the equilib-
rium point 0. The answer, unfortunately, is no, as brought out in the
following.
109 Example. Consider the system ( 1 ), with

— + a cos
1
2
1 1 — a sin cos t t
A(0 =
-
1— a sin cos t t — + a sin
1
2
1 _
Then it is easy to verify that

1
cos t sin t
0) =
sin t '
cos t
Now, the eigenvalues of A (t) are in fact independent of t and satisfy
the characteristic equation
P + (2 - a)k + (2 - a) = 0
Thus, if 1 <a< 2, the eigenvalues of A(?) have negative real parts for
all t , and in fact the eigenvalues of A(>) are bounded away from the
imaginary axis; yet, the equilibrium point 0 is unstable.
The above example shows that even if the eigenvalues of A(Y)

all have negative real parts for each t the equilibrium point 0 may ,
still be unstable. On the other hand, it is possible to prove the following

result.
110 Theorem Consider the system (1); suppose that A(-) is continu-
ously differentiable and that the eigenvalues of A(t) all have negative
real parts for all t > 0. Under these conditions, the equilibrium
point 0 of (1) is uniformly asymptotically stable over [0, oo), provided
the quantity
111 sup
t> o
||
A(f)||/ Av
is sufficiently small.
Because the quantity v can be thought to provide a measure of the

rate of variation of A(-), the above result states that the equilibrium
point 0 of (1) is uniformly asymptotically stable over [0, oo), provided
(1) the eigenvalues of A(/) all have negative real parts for each t >0
and (2) the rate of variation of A(-) is sufficiently small.
Because Theorem (110) is a special case of a general result for
nonlinear systems, we shall omit the proof. For the general result, see
Sec. 5.6.
Problem 5.13. Using the Liapunov matrix equation, determine whether

each of the matrices below is Hurwitz (i.e., whether all of its eigenvalues have
negative real parts).
(a) A= (b) A= (c) A

Problem 5.14. Consider an RLC network that does not contain any
capacitor loops or inductor cutsets. Such a network can be described in state
variable form by choosing the capacitor voltages and inductor currents as the
state variables. Specifically, if x c denotes the vector of capacitor voltages and
x L denotes the vector of inductor currents, then the state equations are of the
form
±c — ~C °
- -1
Rl2 Xc
-X L _ .
_0 L_ R22- Lx l _
where C is the (diagonal positive definite) matrix of capacitance values; L

is the (positive definite) matrix of inductance values; R n ,R 12 R 22 are ,
matrices arising from the resistive subnetwork; and R n R 22 are symmetric.

,
(a) Show that the total energy stored in the capacitors and inductors is
given by
V = yXcCx c + yXiLx L
(b) Using this V as a Liapunov function candidate, show that (0, 0) is
a stable equilibrium point if R n and R 22 are both positive semidefinite, and
that (0, 0) is an asymptotically stable equilibrium point if Rn and R 22 are

both positive definite.
Problem 5.15. Using the results of Sec. 5.3.3, determine whether 0 is

a (uniformly) stable or (uniformly) asymptotically stable equilibrium point,
for each of the systems below.
*i (0 —2 + sin t
2 1
(a)
-*2 ( 0 - cos / — 1_
“
"*i or "It - t
2
1 - t 3 +n ~*i( 0
(b) *2(0 = t
2 - 1
3
0 *2(0
-•*3(0- „ 2 5 4 — t_ -*s(0-
Problem A system of the form [5.1(1)] is said
5.16. to be bounded at
time t 0 if for every S > 0 there exists an e(d, t 0 ) such that
x(/ 0
|| < <5=>||x(0|| < €(t 0
) || V t> t 0 , <5)
It is said to be uniformly bounded over [r 0 , oo) if for every <5 > 0 there exists
an 6(d) > 0 such that
||x(*i)ll < <>, h >*o =H|x(f)ll < € (ft V t>t x
(a) Show that for a linear system of the form (1), the equilibrium point
0 is bounded at line t0 if and only if it is stable at time t0. [Hint : Use Theorem
(5).]
(b) Show that for a linear system of the form (1), the equilibrium point
0 is uniformly bounded over [/ 0 , oo) if and only if it is uniformly stable over
[to, oo). [Hint: Use Theorem (19).]
5.4
LIAPUNOV'S INDIRECT METHOD
In this section, we shall combine the results of the earlier sections to
obtain one of the most useful results in stability theory, namely Lia-
punov’s indirect method (also known as Liapunov’s first method). The
great value of this method lies in the fact that, under certain conditions,
it enables one to draw conclusions about a nonlinear system by studying
the behavior of a linear system.
We shall begin by studying the concept of linearizing a nonlinear
system around an equilibrium point. Consider first the autonomous
system
Sec. 5.4 Liapunov's Indirect Method 187
1 i(0 = f WO]
Suppose f(0) = 0, so that 0 is an equilibrium point of the system (1),
and suppose also that f is continuously differentiable. Define
2
a ran
i.e., A denotes the Jacobian matrix of f evaluated at x — 0. By the

definition of the Jacobian, we have that if we write
3 fj(x) = f(x) — Ax
then
4 lim JliWIUo
||x||-»0 II
X ||
where all norms are /2 -norms. Alternatively, we can say that
5 f(x) = Ax + f i(x)
is the Taylor’s series expansion of f(.) around the point x = 0 [recall
that f(0) = 0]. With this notation, we refer to the system
6 z(t) = Az (t)
as the linearization or the linearized system, around the equilibrium

point 0, of the system (1).
The development for nonautonomous systems is similar. Given
the nonautonomous system
7 ±{t) = f [t, x(t)]
suppose that
8 f(t, 0) = 0, v/>o
and that f(7, •) is continuously differentiable. Define
10 f i(t, X) = f(/, x) — A(f)x

Now, by the definition of the Jacobian, for each fixed t > 0, we have
11
[lf,(t,x)H _n
However, it may or may not be true that
f
12 lim sup l!
ifr*)H
||X||
=Q
l|x|H0 f>0
In other words, the convergence in (1 1) may or may not be uniform in t.
Provided (12) holds we refer to the system,
13 m= a(/)z(o
0 o
as the linearized system (or linearization), around the equilibrium point

0, of the system (7).
14 Example. Consider the system
15 *i(0 = — *i(f) + tx l
16 x 2 (t) = x t) - xt ( 2 (t)
In this case, f{t, •) is continuously differentiable, and
_i
17 A(/)=r °i v<>
_
L i — iJ
However, the remainder term f\{t, x) is given by
18 f 1 (t,x) = [txl 0]'
so that the uniformity condition (12) does not hold. Accordingly, we do

not refer to the system
19 i,0) = — Zj(0
20 i 2 (0 = z ~ z (0i(t) 2
as a linearization of the system (1 5)— (1 6).
Theorem (21) is the main stability result of Liapunov’s indirect

method. Because there is nothing to be gained by assuming autonomy,
we shall state the result for nonautonomous systems.
21 Theorem Consider the (nonautonomous) system
22 x(0 = f[t, x(0]
Suppose that
23 f(/, 0) - 0
and that f(t, •) is continuously differentiable. Define
24
25 i x (t, x) =f (t, x) — A(f)x

and assume that
26 lim sup IMijOU = o

t>
Hx ||->0 11*11
27 A(0 is bounded
Under these conditions, if the equilibrium point 0 of the system
28 z(0 = A(t)z(t)
is uniformly asymptotically stable over [0, oo), then the equilibrium
point 0 of the system (22) is also uniformly asymptotically stable

over [0, oo).
proof Because A(-) is bounded and the equilibrium point 0 of (28) is

uniformly asymptotically stable over [0, oo), it follows from lemma
[5.3(85)] that the matrix
29 P(0 = <J>'(T, /)<D(t, t) dx

J
is well defined for all r > 0, is positive definite for all t > 0, and satisfies,
for some positive OC and /?, the inequalities,
30 ax'x < x'P(0x < /?x'x, V x G Rn V ,

t >0
Hence the function
31 V(t, x) = x'P(/)x
is a decrescent p.d.f. Calculating V for the system (22), we get
32 V(t x) ,
= xT(/)x + f'(t, x)P(r)x + x'P(0f(*, x)
= xTP(/) + A'(0P(0 + P(0A(01x + 2x'P(0fi(^, x)
However, from (29), it can easily be shown that
33 P(0 + A'(t)P(t) + P(0A(0 = -I

Hence
34 V(t, x) = — x'x + 2x'P(r)f 1 (^, x)
Because (26) holds, pick a number r > 0 such that
35 ||fi(f, x)||^^j||x||, V x with x < || ||

r, Vt>0
Then (35) and (30) together imply that
36 1
2x'P(/)fi(*, x) |
< V t >0 whenever 1 1
x 1 1
< r
Therefore
37 V(t, x) <— V t >0 whenever ||x|| < r
Thus all the hypotheses for Theorem [5.2(45)] are satisfied, and we con-
clude that the equilibrium point 0 of (22) is uniformly asymptotically
stable over [0, oo). ma
38 corollary Consider the autonomous system
39 ±(f) = f [x(01
Suppose that f(0) = 0 and that f is continuously differentiable. Define
Under these conditions, the equilibrium point 0 of (39) is (uniformly)

asymptotically stable if all eigenvalues of A have negative real parts.
To prove the instability counterpart to Theorem (21), we have to

assume that the linearized system is autonomous, even if the original
nonlinear system is not.
41 Theorem Consider the system
42 x(0 = f[t, x(*)]
Suppose that (23) holds and that f(t, •) is continuously differentiable.

Suppose in addition that
43 A(?) = x) =A 0 (a constant matrix), V i >0

L UX Jx = 0
and that (26) holds. Under these conditions, the equilibrium point 0
of (42) is unstable if at least one eigenvalue of A 0 has a positive real
part.
proof We shall give the proof only for the case where the matrix A 0
satisfies the condition [5.3(45)]. The proof for the general case can be
obtained from the one given below by using continuity arguments.
Because A 0 is assumed to satisfy [5.3(45)], and has at least one eigen-
value with a positive real part, we have from Theorem [5.3(66)] that the
equation
44 A 0 P + PA 0 = I
has a unique solution for P and that this matrix P has at least one positive
eigenvalue.Now, by arguments entirely analogous to those used in the
proof of Theorem (21), it can be shown that if we choose
45 F(x) = x'Px
then V(t, x) is an l.p.d.f., so that, by Theorem [5.2(106)], the equilibrium
point 0 of (42) is unstable. The details are left as an exercise. H
Remarks: Theorems (21) and (41) are very useful because they
enable one to draw conclusions about the stability status of the equi-
librium point 0 of a given nonlinear system by examining a linear
system. The advantages of these results are self-evident. Some of the
limitations of these results are the following: (1) The conclusions
based on linearization are purely local in nature; to study global
asymptotic stability, it is still necessary to resort to Liapunov’s direct
method. (2) In the case where the linearized system is autonomous, if
some eigenvalues of A have zero real parts and the remainder have
negative real parts, linearization techniques are inconclusive, because
this case falls outside the scope of both Theorems (21) and (41).
(This can be rationalized as follows: In the case described above, one
can say that the linearized system is on the verge of stability or, if —
one is a pessimist, that it is on the verge of instability. Thus it stands
to reason that the stability status of the equilibrium point 0 is actu-
allydetermined by the higher-order terms, which are being neglected
in the linearization.) (3) In the case where the linearized system is
nonautonomous, if the equilibrium point 0 is asymptotically stable
but not uniformly asymptotically stable over [0, oo), the linearization
technique is once again inconclusive. It can be shown by means of
examples that the assumption of uniform asymptotic stability in
Theorem (21) is indispensable. Note that corollary (38) is the basis
of the linearization technique discussed in Sec. 2.3.
46 Example. Consider the Van der Pol oscillator
47 x 1
=x 2
48 x2 = ju( ~ xj)x — x
1
2 t
The linearization of this system around (0, 0) is
If > 0, the eigenvalues of this matrix A both have positive real parts.
Hence, by Theorem (41), the equilibrium point (0, 0) of the system
(47)-(48) is unstable.
50 Example. The purpose of this example is to show that inequality

(35) can sometimes be used to obtain an estimate for the domain of
attraction of the asymptotically stable equilibrium point 0 .
Consider the system
51 Xi = —X + X + x 2 X x2
1
52 X2 = —x + x\ t
The linearization of this system around (0, 0) is
If we let A denote the matrix in (53), we find that the eigenvalues of A

are (— 1 ± hence, by Theorem (21), we know that (0, 0) is an
asymptotically stable equilibrium point of (51)— (52).
To obtain an estimate for the domain of attraction r[cf. (35)], we
first solve the equation
54 A'P + PA = —I
3 3
for P. It can easily be verified that

"
— 1/ 2
55
3/2_
Hence fl in (30) can be chosen as the largest eigenvalue of P, which is
(5 + a/5~)/4, or approximately 1.81. To satisfy (35), we must choose r

in such a way that
56 ||fi(x)||<^||x|| whenever ||
x ||
< r
where p < 1 is some number. Note that we chose p = f in (35) for the
purposes of proving the theorem, but here it is desirable to choose p as
close to 1 as possible in order to get the least conservative value for r.
In the present case, we have

1/2
57 iif.fr) ii Or \x\ ± x!) l* 2 |<||x||
11*11 fr? + *I )
1/2
Hence, to satisfy (56), we can choose r as close as possible to 1/2/? or

approximately 0.27. In other words, what we have shown is that every
trajectory starting from an initial condition whose norm is less than
0.27 approaches 0 as t — > oo.
An Application Feedback Stabilization of

:
Nonlinear Control Systems 1
As an application of Liapunov’s indirect method in the design of

nonlinear control systems, we shall consider the following problem:
Given an autonomous system described by
58 ±(t) = f [x(0, u(/)]

where f : Rn x R m —» R n ,
the objective is to find a feedback control law,
i.e., a relationship of the form
59 u(0 = g[x(0]
in such a way that the equilibrium point 0 of the closed-loop system
60 ±{t) = f[x(0, g(x(0)]

is asymptotically stable.
To give a solution to this problem based on Liapunov’s indirect
method, we make the following assumptions:
1
This application requires some familiarity with linear systems and the linear
regulator problem in optimal control theory.
Sec. 5.4 Liapunov’s Indirect Method 193
1. f(0, 0) - 0.
2. f is continuously differentiable.
3. If we define
A= <7f(x, u) ~[
_ dx Jx= 0} n = 0
~
= df (x, u) ~|
B
- <?U _Lc=0,u = 0
then we have
rank [B| AB| . . . |
A" -1 B] =n
Assumption (3) above has a very simple interpretation In much :
the same way that (6) is called the linearization of (1) around the point
x = 0, one can think of the system
z(t) = Az(t) + Bv(/)
as the linearization of the system (58) around the point x =0 ,

u =0 .
Thus, assumption (3) states that the linearized system (65) is controlla-
ble. For such systems, the following result is well known (see, e.g.,
[7] Chen).
fact If (64) holds, then there exists an n x n matrix K such that all eigen-
values of A — BK have negative real parts.
Thus fact (66) tells us that if we choose the feedback control law
v(0 = — Kz (t)
in system (65), it is possible to choose the matrix K in such a way that

the equilibrium point z = 0 of the closed-loop system
z{t) = (A - BK)z(0
is It is worth noting that fact (66) is not just of
asymptotically stable.
theoretical interestand that there are actually systematic procedures for
finding a suitable K, given A and B.
One such procedure comes from the linear regulator problem in
optimal control theory (see, e.g., [2] Athans and Falb) and can be sum-
marized as follows Given A, B that satisfy (64), pick positive definite
:
matrices P e RnXn and Q e and solve the so-called Riccati

equation
-P - A M - MA + MBQ _1 B'M = 0
for the unknown matrix M e Rnxn . It can be shown that, under the
stated assumptions, (69) has a unique solution for M, which is sym-
metric and positive definite. Moreover, if we choose

70 K = Q B'M _1
then all eigenvalues of A — BK have negative real parts. The last state-
ment can be proved using Liapunov theory, and we shall now do so.
71 fact Suppose that A, B satisfy (64), that P e Rnxn and Q e R mXm are
positive definite, that M satisfies (69), and that K is defined by (70). Then
all eigenvalues of A — BK have negative real parts.
proof We make use, without proof, of the fact that the solution M
of (69) is positive definite. We have
72 A — BK = A — BQ _1 BM
Therefore
73 (A - BK) M + M(A - BK) = [A- BQ _1 B'M]'M

+ M[A - BQ _1 B'M]
= -P - MBQ -1 B'M
where we use (69) in the last step. Now, by assumption, P is positive defi-
and
nite, MBQ _1 B'M is at least positive semidefinite, so that the right-
hand side of (73) is A — BK
negative definite. Therefore, the matrix
satisfies A — BK playing the
the Liapunov matrix equation [5.3(43)], with
role of A (in [5.3(43)]), M playing the role of P, and P + MBQ B'M play-
_1
ing the role of Q. Hence, by Theorem [5.3(55)], all eigenvalues of A — BK

have negative real parts. M
Up to now, we have studied only the background material dealing
with the stabilization of the linear system (65). We shall now present the
application of Liapunov’s indirect method, which is contained in
Theorem (74). Basically, Theorem (74) shows that once a stabilizing
feedback gain has been found for the linearized system (65), we have
already found a way to stabilize the original nonlinear system (58).
74 Theorem Consider the autonomous system
75 ±{t) = f [x(0, u(0]

where f: Rn X R m — R »
n
. Suppose that f is continuously differen-
tiable and that f(0, 0) = 0. Define
df~
76 A= (?X_ x=0>u =o
df~
77 B =
.ÛJx=0,u = 0
Suppose there exists a matrix K e i?mXB such that all eigenvalues of

A — BK have negative real parts. Under these conditions, if we
choose
:
78 u(0 = — Kx(r)
in (75), then 0 is an asymptotically stable equilibrium point of the
resulting system
79 x(t) = f[x(t), — Kx(t)]

proof Let g: R" - > Rn be the function defined by
80 g(x) = f(x, — Kx)

Then the system (79) can be represented as
81 x(r) = g[x(/)]
Next, we observe that
Because, by assumption, all eigenvalues of A - BK have negative real

parts, it follows from corollary (38) that 0 is an asymptotically stable
equilibrium point of (79). tm
Thus Theorem (74) provides a powerful design tool for nonlinear

control systems.The technique based on Theorem (74) can be split into
three parts
1. Linearize the given nonlinear system.
2. Find a stabilizing control law for the linearized system.
3. Implement the same control law for the original nonlinear
system.
83 Example. Consider a second-order system described by
Xi = 3*! + x\ — sat (2x + u 2 )
x2 = sin x — x + u t 2
where the sat function is defined by
satf a =
\a, M< 1
\
(sign
.
tr, 1
cr |
> 1
The linearization of this system around x = 0, u = 0 is
The A matrix above has one eigenvalue with a positive real part. How-
ever, the linearized system can be made stable by applying the control
law
v = [4
Hence the original nonlinear system can be stabilized by choosing

u — 4xj — x 2
An Application : Singular Perturbations
In Sec. we derived some results for singularly perturbed

4.3,
linear systems. Inwhat follows, we shall use Liapunov’s indirect method
in conjunction with the results of Sec. 4.3 to derive some results for
singularly perturbed nonlinear systems. The proof of Theorem (84) is
omitted, because it follows readily from earlier results.
84 Theorem Consider the system
85 x(f)= f[x(», y(/)]

86 eyO) = g[x(0, y(0]
where f: R n x R m — Rn » , g: Rn X R m —* Rm are continuously dif-
ferentiable and satisfy
87 f(0, 0)=0
88 g(0,0) = 0
Define
df
89
~dx ( 0, 0)
90 _ <5f
<?y ( 0, 0)
91 _
<?x ( 0, 0)
92 — o*g
dy ( 0, 0)
and suppose A 22 is nonsingular. Under these conditions, there exists

a continuously differentiable function h: R n — *
m such that, in some
R
neighborhood of (0, 0),
93 v = h(x)
is a solution of
94 g(x, y) =0
Moreover, we have that
(i) If all eigenvalues ofand of A n — A 12 A 22 A 21 have nega-
A 22 1
tive real parts, then there is an e 0 such that (0, 0) is an asymp-

totically stable equilibrium point of (85)— (86) whenever
0 <e<e 0.
(ii) one eigenvalue of A 22 or ofA tl — A 12 A 22 A 21 has a

If at least 1
positive real part, then there is an e 0 such that (0, 0) is an

unstable equilibrium point of (85)— (86) whenever 0 < e e0 < .
Remarks : The existence of the function h follows from the implicit

function theorem.
96 = x (x + y — 1) + x\ + y
1 1
97 x = x (x\ — 5) + x (x — 3) + y(x + 1)
2 t 2 1 2
98 ey = x (x - 2) - x (x - 1) - y(y + 1)
1 1 2 1
For this system, one can easily verify that
A2 — i [ 2 1]
^22 — [“l]
An ^ 12 -^ 22^21
Since all the eigenvalues of A 22 and of M

have negative real parts, we
conclude that there exists an e 0 >
0 such that (0, 0, 0) is an asympto-
tically stable equilibrium point of the system (96)-(98) whenever
0 < < € €0 .
Problem 5.17. For the system of Problems 5.9-5.12, apply Liapunov’s

indirectmethod, and determine whether or not it yields conclusive results.
Problem 5.18. Consider a feedback system described by
x(t) = A x(t) + bu(t)

y(t) = c'x(0
«(0 = — 00 (0)
Suppose that the matrix A— bkc' is Hurwitz (i.e., has all eigenvalues with
negative real parts) whenever k belongs to the interval [0, Ar 0 ]. Show that 0 is
an asymptotically stable equilibrium point whenever 0(-) is a continuously
differentiable function such that
0 < [d<f>{o)ldo]a =Q < k 0

4
5 5
THE LUR'E PROBLEM
In this section, we of an important class of
shall study the stability
control systems, namely, feedback systems whose forward path contains
a linear time-invariant subsystem and whose feedback path contains a
memoryless (possibly time-varying) nonlinearity. Such a system is
shown in Fig. 5.8 and is described by the equations
1 x(0 = Ax(t) + b £(t), t> 0
2 o{t) = c 'x(0 + d((f)
3 m =
g(s)= c(sl - A)- b + d
l
In the above, A e Rnxn ;

b, c, x e Rn ;
d
, g, £, are all scalars; and the
function 0(*, •) is continuous in both its arguments. 14 study of sys- A
tems of the form (l)-(3) is important for at least two reasons:
1. Many physical systems can be naturally interpreted as con-
sisting of a linear part and a nonlinear part so that the descrip- ,
tion (l)-(3) is reasonably general.

2. Many comprehensive results are available concerning the
stability of such systems (of which only a few are given in this
book).
way on
All the results given in this section depend in an essential
a theorem knownKalman-Yacubovitch lemma. Actually, the
as the
result used here is a generalization of this lemma due to Lefschetz. The
proof of this particular theorem is not given in this book, because it
utilizes several concepts not heretofore introduced. For an early treat-
ment of this result, see [16] Lefschetz. A thorough development of the

results of this section can be found in [17] Narendra and Taylor.
1
We also assume, as in the rest of this chapter, that $(• , •) is of such a nature
that the system (1)— (3) has a unique solution corresponding to each initial condition
x(0) and that this solution depends continuously on x(0). This is the case, for exam-
ple, if 0(* , •) satisfies a global Lipschitz condition.
;
Sec. 5.5 The Lur'e Problem 199
5.5.1 Problem Statement
In this subsection, we shall state the problem under study more

precisely and discuss two conjectures, due to Aizerman and Kalman,
respectively.
The problem dealt with in this section is as follows : Given a sys-
tem described by (1)— (3), suppose
(Al) All eigenvalues of A have negative real parts (in this case A
is known as a Hurwitz matrix), or
(A2) A has a simple eigenvalue of zero, and the rest of its eigen-
values have negative real parts.
(A3) The pair (A, b) is controllable i.e.,
rank [b| Ab| . .

.
|
A n_1
b] =n
(A4) The pair (A, c) is observable; i.e.,
rank [c |
A'c |
. . .
|
(A')"
_1
c] =n
(A5) The nonlinearity 0(., .) satisfies
0(f, 0) = 0, V t >0
<J0(/, a) > 0, Va e R, V t >0
[The condition (7) is sometimes stated as “0(f, .) is a first and third-
quadrant nonlinearity,” because (7) ensures that the graph of cn->
(f)(t, o) lies in the first and third quadrants, for all t > 0.]
Subject to the above assumptions, the problem is to find condi-
tions on A, b, c, d which ensure that the equilibrium point x 0 of the =
system (l)-(3) is globally asymptotically stable whenever any non-
0 is
linearity satisfying (A5).
In contrast with the systems studied in Sec. 5.2, we are concerned
at present not with a particular system but with an entire class of sys-
tems, because we have not made any assumptions on
0 other than that
it satisfies (A5). For this reason, the problem under study is sometimes
referred to as an absolute stability problem. It is also known as the Lur'e

problem after the Soviet scientist A. I. Lur’e.
,
It should be mentioned here that assumptions (A3) and (A4) do
not result in any loss of generality, because given any proper rational
function h(s), one can always find A, b, c, d such that
h(s) = c\sl - A) -1 b + d
and such that (A3), (A4) hold. See [7] Chen for full details.
In some cases, (7) is replaced by the following more stringent
condition: There exist constants k 2 '>k such that l
>0
k a2x
<a <f)(t, o) <ko 2
2
, V o e R, V * >0
Because of its importance, (9) is given a name.
10 [definition] The function 0(., •) is said to belong to the sector

[k t , k 2 ], or to lie in the sector [k u k2\ ,
if (9) holds.
Aizerman's Conjecture
In 1949, the Soviet scientist M.A. Aizerman made the following

tempting conjecture: Suppose that d = 0, that 0 belongs to the sector
[ki
k 2 \, and that for each k e [k u k 2 ],
,
— the matrix A be 'k is Hurwitz.
Then the equlibrium point 0 of the nonlinear system (l)-(3) is globally
The basis for this conjecture is the following reasoning: Suppose
that d = 0 and that 0 belongs to the sector [k u k 2 \. If we replace 0 by a
linear element of gain k the system (1)— (3) reduces to the linear time-
,
invariant system
11 ±(t) = [A — bc'/:]x(/)
The hypothesis is that the system (11) is globally asymptotically stable

forall k e \k k 2 ], and we would like
t ,
to conclude that the original
nonlinear system is also globally asymptotically stable. 15 If it were true,
this conjecture would have been extremely useful, because would have it
allowed one to deduce the stability of a nonlinear system by studying

only linear systems. Unfortunately, several counterexamples have
shown that this conjecture is false in general. (See, e.g., [23] Willems:
also see Theorem [6.7(42)] for a related result.)
Kalman's Conjecture
In 1957, R. E. Kalman made the following conjecture, which

strengthens the hypothesis of Aizerman’s conjecture. Suppose the non-
linear function 0 satisfies
12 ki < <k A, V o e R, V t >0

and suppose the matrix A— be
f
k is Hurwitz for all k e [k 3 , k 4 ]; then
the nonlinear system (l)-(3) is globally asymptotically stable.
The hypothesis of Kalman’s conjecture is stronger than that of
Aizerman’s conjecture, because in general k 2 x
and k 2 k 4 There- <k < .
fore the matrix A—

assumed to be Hurwitz for a larger range of
be 'k is
values of k. Also, by Problem 5.18, the hypotheses of Kalman’s con-

jecture insure that 0 is an asymptotically stable equilibrium point
15 As pointed out in Sec. 5.1, one can unambiguously say that the system is
globally asymptotically stable, rather than that the equilibrium point 0 is globally
6
whenever 0(-, .) satisfies (12). In spite of this, however, Kalman’s con-

jecture is also false.
Even though both Aizerman’s and Kalman’s conjectures were
ultimatelyshown to be false, they nevertheless served a very useful pur-
pose, because they created a great deal of interest in the Lur’e problem
and were thus indirectly responsible for the many results that are avail-
able today.
In the remainder of this section, we shall present three kinds of
stability criteria for the Lur’e problem. These criteria are arranged
according to the complexity of the Liapunov function employed in
deriving them. That is, the circle criterion (Sec. 5.5.2) is based on the
simplest Liapunov function, the Lur’e-Postnikov criterion (Sec. 5.5.3)
is based on one that is rather more complicated, while Popov’s criterion
(Sec. 5.5.4) employs the most complicated Liapunov function of the

three.
The proofs of these criteria make use of the following generaliza-
tion of the Kalman-Yacubovitch lemma, which is due to Lefschetz.
13 Theorem Given a Hurwitz matrix A, a vector b such that the pair

(A, b) controllable, a real vector v, scalars y
is 0 and e 0, and a > >
positive definite matrix Q, there exist a positive definite matrix P
and a vector q that satisfy the equations
14 A'P + PA = — qq' — eQ
15 Pb — v = y q 1/2
if and only if e is small enough, and the scalar function
16 h(s) A y + 2v'(sl - A) _1
b
16
satisfies
17 Re h(jco) >0, V co e R
proof See [16] Lefschetz [pp. 115-18].
The use of this theorem in constructing Liapunov functions

becomes clearer in Secs. 5.5.2-5.5.4.
5.5.2 Circle Criterion
Given the nonlinear system (l)-(3), suppose that the matrix A is

Hurwitz [so that (Al) is satisfied] and that (A3)-(A5) hold. The circle
1
Although we have not introduced the concept here, we mention that, with
the assumptions on A, (17) is equivalent to the requirement that h(s) is a strictly
positive real function.
criterion arises out of attempting to deduce the global asymptotic sta-
bility of the system by employing a Liapunov function of the form
18 F(x) = x'Px
Note that the nonlinear characteristic does not appear explicitly in <f>
the Liapunov function (13). Accordingly, this Y-function is referred to

as a common Liapunov function, because the same Liapunov function
is used with all nonlinearities This is in contrast with the Liapunov
<f).
functions used in Secs. 5.5.3 and 5.5.4.

Throughout this subsection we shall make the following basic
assumptions'.
to the sector k] for some finite positive k, and
1. 0 belongs [0,
2. 1 + kd 0.
If, $ belongs to some general sector [a, fi],

in a given problem,
one can always employ what are known as loop transformations to
ensure that the transformed </> belongs to the sector [0, /? — a]. This is
discussed later. Assumption 2 requires a little more explanation. Sup-
pose 1 + kd < 0. Then one can always find a number k 0 e [0, k] such
that 1 + k d = 0. Suppose we now choose (p) = k o. Then
0 <f> 0 0 <j> 0
belongs to the class of nonlinearities under consideration (namely, (f> 0

belongs to the sector [0, /c]). Now, absolute stability requires that the
system (l)-(3) be stable with <j) =
0„- However, in the present case (2)
becomes
19 = c'x + d£, — c'x — k do

a 0
20 (1 + k d)a
= c'x Q
However, because + k d = 0, (20) 1 0

can hold only if x = 0, which
means that the feedback system is degenerate. The situation can also be
explained using transfer functions. Define
21 gfs) = c(.sT — A) -1 b +d
Then g is the open-loop transfer function. If we now apply a feedback
gain of k 0 ,
the closed-loop transfer function becomes
22 A tot
gA ’
=
1 +
M
k 0 g(s)
Now, because f is a rational function of s, so is gc (s). If we let |

s |
— * °o,
we get
„ . . i?(°°) = d
23 gA°°) = i fc 0 g(wj r+M
However, note that 1 + k 0 d = 0 (and that d 0). Hence g(oo) = oo, ^
which means that g c is not a proper rational function and so represents
an unstable system (see Sec. 6.3).
The conclusion, therefore, is that if cj) belongs to the sector [0, k],
then 1 + kd > 0 is a necessary condition for absolute stability, so that

we can safely assume that it holds.
Calculating the derivative of V along the solutions of (1)— (3), we
get
24 F(x) = x Px + x'Px = x'(A'P + PA)x + 2x'Pb£

We would like to derive conditions on A, b, c, d, and k that ensure that
— Fis an l.p.d.f. over all of R". To do this, we proceed as follows: First,
it is easy to verify from (2) that
25 k(x' c + (1 + kd)( 2 - ((ka + () = 0

Therefore, we may substract this quantity from V without affecting its
value. This gives
26 F(x) = x'(A'P + PA)x + 2£x'(Pb — %kc)

- (l + kd)( 2 + ((ka + 0
Now, because <j) belongs to the sector [0, k], we have
27 0 < a<f)(t, a) < ka 2
, Vo e R
which can be rearranged as
28 0 < G
< k, V o 0
This inequality (28) means (loosely speaking) that <j)(t, a) and ka —

<p(t, a) always have the same sign; more precisely,
29 , a)[ka — </)(t, a )] > 0, Vo e R

In other words,
30 ((la7 +0<0
Hence the last term in V is nonpositive. We now make use of Theorem
(13) in order to complete the square, so as to make the rest of the terms
in V negative. Specifically, apply Theorem (13), with Q any positive
definite matrix, A, b as at present, y 1 kd, and v Ax/2. Then = + =
Theorem (13) assures us that if (17) holds (the implications of which are
discussed later on), then there exist a positive definite matrix P and a
vector q such that
31 A'P + PA = — qq' — eQ
32 Pb — Â:c = (1 + Ax/) q 1/2
Thus V becomes
33 F(x) = — ex'Qx — x'qq'x + 2£x'(l + kd) q — (1 + kd)( l/2 2
+ ((ka + 0
= —ex'Qx — [x'q — (1 + kd) (] + ((ka + 0 1/2 2
< -ex'Qx
where we make use of (30) and the obvious fact that
34 [x'q - (1 + kdy /2
£]
2
>0
Thus, clearly, V is a p.d.f., and —V is a p.d.f., whence by Theorem
[5.2(58)] it follows that the system (l)-(3) is globally asymptotically
stable.
Now, let us come back to condition (17), which, as has just been
demonstrated, is a sufficient condition for global asymptotic stability.
In the present case, we have

35 h(s) = (14- kd) + kc\sl - A) -1 b - + kg(s) 1
where
36 g(s) A c(sl — A) -1 b + d
is the transfer function of the open-loop system (l)-(2). Therefore, a
sufficient condition for the global asymptotic stability of the system
(l)-(3) is
37 Re [1 + kg(jco)] > 0, V co e R
This is stated formally as a fact.
38 fact If (f)
belongs to the sector [0, k] and 1 + kd > 0, then (37) is a suf-
ficient condition for the absolute stability of the system (l)-(3).
If </) belongs to the sector [0, oo), then the corresponding absolute
stability criterion can be obtained from (37) by dividing through by k
and then letting k — > oo.
39 fact If </) belongs to the sector [0, oo) and d > 0, then a sufficient condi-
tion for absolute stability is
40 Re g(jco) >0, V co e R
Actually, it should be clear that we can drop the phrase “and
1 + kd > 0” in the statement of fact (38), because if (37) holds, then

automatically
41 1 + kd = Re [1 + kg(oo)] > 0
Similar remarks apply to fact (39).

If (/) belongs to some sector [a, /?], then the system (l)-(3) can be
made to satisfy the hypothesis of fact (38) by loop transformations as ,
shown in Fig. 5.9. In this case, the transformed forward path transfer
function is
42 &0) = +gO)
ag(s) 1
Sec. 5.5 The Lure Problem 205
S t (s)
I I l
I J
FIG. 5.9 <t>

t
while the transformed nonlinearity is
43 = <f>(a) — aer
Now, belongs to the sector [0, /}
— a]. However, before we can
apply fact (38), we must ensure that the transfer function g (s) can be t
represented in the form
44 g (s)
t
= c£s\ - A,)->b, + d,
where A t
is a Hurwitz matrix. Now, if the pair (A„ b,) is controllable
and the pair (A c,) is observable, then the eigenvalues of A, are the
f,
same as the poles of g (s). In turn, if a 0, then the poles of g (s) are
t ^ t
the same as the zeroes of 1 + ag{s). Therefore, to apply fact (38) to the
transformed system (42)-(43), we must first ensure that 1 + ag{s) has
no zeroes in the closed right half-plane.
The last condition can be readily verified using the well-known
Nyquist criterion. Suppose (l)-(2) is a realization of the untransformed
transfer function g(s), and suppose (1) A has no eigenvalues on the
imaginary axis, and (2) A has v eigenvalues with positive real part. Then
gCs ) has no poles on the imaginary axis and has exactly v poles in the
1
open right half-plane. In this context, the Nyquist criterion states that
1 + agC?) has no zeroes in the closed right half-plane if and only if the
Nyquist plot of g(jco) does not intersect the point — 1/a + j0 and encir-
cles it in the counterclockwise sense exactly v times.
We are now ready to extend fact (38) to the general case of a non-
linearity in the sector [a, fi].
Based on the foregoing discussion, a suf-
ficient condition for the absolute stability of the system (1)— (3) can be
given as follows.
45 fact Suppose A in (1) has no eigenvalues with zero real part and has v
eigenvalues with positive real part ; suppose 0 in (3) belongs to the sector
[a, /}], a ^ 0. Then a sufficient condition for absolute stability is
46 The Nyquist plot of g(jco) does not intersect the point — 1/a + j0
and encircles it v times in the counterclockwise sense.
47 Re[l+(/? — «)
r
^g>Lj >0’ VCoeR
The conditions (46) and (47) can be made simpler. However, to do

this, it is necessary to separate the cases a < 0 and a > 0.
Case 1 a > 0 Clearly (47) can be rewritten as

:
48 Re \ \
Ll +
Mm
± ag(ja>)J >0 , VcoeR
Suppose
49 g(jco) = u+jv
Then (48) becomes
50
(l + pu)(l + au) + a^ > R
(1 + aw) + v
2 2
If (46) holds, the denominator of (50) is never zero, so that (50) can be
simplified to
51 (1 + /?w)(l + aw) + apv 2

>0, V co e R
52 (u —^w H-
-J- ~t~ 0, V co e R
The condition (52) can be given a simple geometric interpretation. Let
D(a, P) be the disk in the complex plane that is centered on the real
axis and passes through — 1/a + jO and — 1 //? + j 0. This disk is shown
in Fig. 5.10. The inequality (52) states that the Nyquist plot of g(jco)
never enters the disk.
FIG. 5.10
The encirclement condition (46) can also be slightly simplified.

If the Nyquist plot of g(jco) does not enter the disk D( a, p) and encircles
the point —1/a +y0 exactly v times, then the plot must encircle the
disk Z)(a, P) v times and vice versa.
Combining these results, we get the following.
53 fact Suppose A in (1) has no eigenvalues with zero real part and has v
eigenvalues with positive real part; suppose 0 in (3) belongs to the sector
[a, fi], with a > 0. Then a sufficient condition for absolute stability is
54 The Nyquist plot of g(jco) does not enter the disk Z>(a, /?) and
encircles it in the counterclockwise sense exactly v times.
Case 2 a a <0<p : In this case, it is clear, first, that a necessary

condition for absolute stability is that A is Hurwitz. This is easily seen
by setting 0 (<t) = 0, which is an admissible nonlinearity. Thus (46) and
(47) become, respectively,
55 The Nyquist plot of g(jco) neither encircles nor intersects the point
~ l/ a +/0.
56 Re
H+
MM
H + ag(jco)_\ >0, ] V oe R
ft
Now (56) can be reduced to (51), as before. However, in going from

(51) to (52), we divide by a/?, which is now Hence the ine-
negative.
quality is reversed, and (56) becomes
57 (u + I)(« + +v < 2
0, V a) e R
j)
Thus the geometrical interpretation is that the Nyquist plot of g(jco)
lies in the interior of the disk T>(a, p). Moreover, if this is the case, (55)
is automatically satisfied. Thus we have
58 fact Suppose A is Hurwitz, and suppose 0 belongs to the sector [a, /?],
with a < 0 < p. Then a sufficient condition for absolute stability is
59 The Nyquist plot of g(jco) lies in the interior of the disk D( a, P).
Case 2b a < p < 0 In this case, replace g by — g, a by — p,

:
p
by —a, and apply fact (53).
The large number of criteria are now summarized in the following
theorem, which is generally referred to as the circle criterion.
60 Theorem Suppose A has no eigenvalues on the imaginary axis and

has v eigenvalues with positive real parts; suppose
0 belongs to
the sector [a, /?]. Then a sufficient condition for absolute stability is
one of the following, as appropriate:
(0 If 0 <
a <
p, the Nyquist plot of g(jco) does not enter the
disk Z>(a, P) and encircles it v times in the counterclockwise
sense.
, a
(ii) 7/0 =a
v)
fi
(iii) If a, < 0 < p ,

the Nyquist plot of g(jco) lies in the interior of
the disk D(a, ft).
(iv) If a <P< 0, replace g by — g, a by — /?, /? by —a, and

apply (i).
62 Example. We apply the circle criterion to study the stability of

the damped Mathieu equation
63 y + 2 fly + O + a — q cos co
2 2
0 t)y =0
It is easily verified that (63) can be put in the form (l)-(3) with
1
64 d= 0
^+a 2 ]’
65 0(cr, t) = oq cos co 0t
Our objective is to derive conditions on //, a, q and

,
co 0 that ensure the
global asymptotic stability of the system (63). We assume that ju, a > 0,
so that A is Hurwitz and that q > 0. Next, note that
66 z(s) - S2 + 2ns + u
2
+ 2
67 (f)
belongs to the sector [—q,q]
Therefore, case (iii) of Theorem (60) applies here, with a q and =—

f = q. Here the disk D(oc, ft) is the disk centered at the origin with
radius 1 jq. Plotting g(jco), we note (see Fig. 5.11) that g(jco) attains its
maximum modulus at co =
) an d that this maximum mod-
=fc(fl
2 — M
2
ulus is 1/(2 jua). Thus the condition that the Nyquist plot of g(jco) lies in
the interior of the disk D( a, fi) reduces to
68 J_<±
Ilia q
Sec. 5.5 The Lure Problem 209
or, equivalently,
69 q < 2 pa
Thus (69) is a sufficient condition for the global asymptotic stability of
(63). Furthermore, because the circle criterion is a sufficient condition
for absolute stability, (69) actually implies the global asymptotic sta-
bility of all systems of the form
70 y + 2 py + ( ju
2
+a 2
)y - qf(f 9 y)
=0
whenever /is any nonlinearity belonging to the sector [—1, 1].
71 Remarks: We shall conclude this subsection with a discussion of

Theorem (60) (the circle criterion). One of the most appealing aspects
of this result
is its simple geometric interpretation, which is reminiscent
of the Nyquist criterion. 17 Indeed, it is easy to see that if a /?, then =

the critical disk D(p, /?) shrinks to the critical point —1/a, and the circle
criterion reduces to the sufficiency portion of the Nyquist criterion.
Another appealing aspect of the circle criterion is that one needs

only to plot the transfer function g(jco). This means that if we think of
the forward path in Fig. 5.8 as a black box then to apply , Theorem (60)
we need only to determine the transfer function of the black box and
that thereis no need to construct a state variable realization of the black
box. Because the transfer function can usually be determined experi-

mentally, based on input-output measurements, this feature greatly
facilitates the use of the circle criterion.
In Chap. 6, we shall encounter another circle criterion which for
all intents and purposes looks the same as Theorem (60). However, the
one in Chap. 6 is applicable to a broader class of systems, e.g., dis-
tributed systems, systems with time delays, because the method of
derivation in Chap. 6 is different. On the other hand, the present
Theorem (60) is strictly limited to systems described by ordinary differ-
ential equations of finite order.
Finally, it is worth noting that the some circle criterion is in
sense also a necessary condition for absolute stability. The appropriate
setting for showing this is the functional analytic framework of Chap. 6,
and a precise statement of this result can be found in Theorem [6.7(40)].
5.5.3 Lur'e-Postnikov Criterion
In this subsection, we shall once again consider the system (1)-

(3). Throughout, we assume that A is Hurwitz, d = 0, and that 0 is a
time-invariant nonlinearity that belongs to the sector [0, oo).
17 For a precise statement of this criterion, see Theorem [6.6(58)].

To prove the global asymptotic stability of the system (l)-(3), we

employ a Liapunov function of the form
72 V(x) = x'Px +
where
73 «(ff)A[>0«
Jo
P isa positive definite matrix, and /? 0. The function (72) is called a >
Liapunov function of the Lur’e-Postnikov type. Because <j) belongs to
the sector [0, oo), it is clear that 0>(<7) is always nonnegative, so that V
is a p.d.f.
Calculating the derivative of V along the solutions of (l)-(3),

we get
74 = x'(A'P + PA)x + 2£x'Pb + P<f>(o)d

v(x)
= x'(A'P + PA)x + 2 £x'Pb - fi&c'Ax + c'bf)

where we use the facts that £ = — 0(c) and
75 a = c'x = c'Ab + c'b<j;
Defining
76 v = Pb — -^-A'c
(74) can be rewritten as
77 F(x) = x'(A'P + PA)x + 2£x'v - fic'bZ

2
Let us choose P to be the solution of the equation
78 A'P + PA = -Q
where Q is a positive definite matrix of our choice. Then
F(x) = — x'Qx + 2<j;x'v — J?c'b
2
79 <^
= -(x - Q' vf)'Q(x - Q-‘vO - 0?c'b - v'Q-'v )^ 2

1
We claim that if
80 fic'b
— v'Q
-1
v > 0
then a negative definite quadratic form in x and £. This is easily

V is
verified: Suppose (80) holds; then it is clear that V is at least negative
semidefinite in x and Therefore it suffices to show that
81 v = 0 == x = 0 , £ = 0
Accordingly, suppose V= 0; this implies, because of the positive

definiteness of Q and (80), that
82 x —Q - 1
v<j; =0
83 1 = 0
V
Clearly, (82) and (83) together establish (81). Next, because V is a nega-
tive definite quadratic form in x and there exist positive constants
and a 2 such that
84 V(x) < — ocjx'x — oc 2 £
2
Hence
85 V(x) < — OGiX'x
which shows that — is a p.d.f. and establishes the global asymptotic
stability of (l)-(3).
The preceding discussion is summarized in the following.
86 Theorem Suppose that 0 is a time-invariant nonlinearity belonging

to the sector [0, oo) and that d— 0. Let Q be a positive definite
matrix, and let P be the corresponding solution of (78). Define Y by
(76). Then (80) is a sufficient condition for the absolute stability of
the system (l)-(3).
87 Remarks: Theorem (86) is applicable to a more limited class of

systems than Theorem (60), because Theorem (86) d to be requires
zero and 0 to be time-invariant; neither of these needed in Theorem is
(60). Moreover, for a given A, b, c, whether or not (80) holds depends

strongly on the choice of Q. This is a major drawback of this cri-
terion, because there is no a priori method for choosing
Q intelligent-
ly. At this stage, the Lur’e-Postnikov criterion is mainly of historical
interest.
5.5.4 The Popov Criterion
In this subsection, we shall derive a well-known stability criterion

for autonomous systems. This criterion was first derived by the Ruma-
nian scientist V. M. Popov, who considered a class of systems slightly
different from (l)-(3). Specifically, the class of systems studied by
Popov are those of the form
88 ±{t) = Ax(0 — b0[cr(O]
89 <r(0 = c'x(0 + d£(t)
90 fa) = -0[<7(O]
Clearly, the difference is between (90) and (3). By recasting (88)-(90)
in the form
i
x(0“ o —
x(0“ ~b~
91
+
-0 o_ m. _ 1 _
X (0
92 o(t) = \ c'
d]
7
'
L£(0J
93 u(t) = -0HO1
We see that the present system is of the form (l)-(3), where the A
matrix satisfies assumption (A2) instead of (Al).
Throughout this subsection, we shall make the following Assump-
tion: 0 belongs to the sector (0, oo), and d 0. ^
The following result shows that d > 0 is actually a necessary
condition for absolute stability.
94 fact Assuming that d ^0 and that 0 belongs to the sector (0, oo), a
necessary condition for the absolute stability of the system is d > 0.
proof (outline of) system (88)— (90) is absolutely stable, then

If the
it is globally asymptotically stable for all 0 of the form
95 0(0*) = eo
for all € > 0. With this choice of 0, the system becomes
KOI _ ["A - ebc' —ebdl rx(0-

96 ~~
£(oJ L -fc' -ed J L£(0-
which can be rewritten as
±(01 ["A — 6bc' —€bd~\ rx(0"

97
_p£(t)_ _ c' -d JLttO-
where p = 1/6. The system (97) is similar to those studied in Sec. 4.3. If
we neglect the terms involving 6 on the right-hand side of (97), the system
is qualitatively similar to
±(01 _ I” A 0 1 r x(0”
98
ju£(t)_ c' — d_ _£(t)_
Suppose now by way of contradiction that d < 0. Then, by letting 6 — > 0

(i.e., p— * oo), we get the simplified system
99
OITA 0 1 rx(0
p£,(i)_ _-c' -d] L«0
100 x(0 =o
101 Z(t) = <?(0) exp
In deriving (100), we use the fact that A is Hurwitz and is therefore non-
singular. Clearly, by (101), the system (99) is unstable. Hence, by Theorem

[4.3(44)], the system (98) is unstable for all sufficiently large values of ft.
Finally, by continuity, the system (97) is unstable for all sufficiently small
values of €. This shows that d > 0 is a necessary condition for absolute
stability.
To study the stability of the system (88)-(90), Popov uses a Lia-

punov function of the form
102 V= x'Px + OCjf
2
+ /?d>(<7 )
where defined in (73), P is a positive definite matrix, and a 1?

 is are
fi
nonnegative constants, not both of which are zero. Because d > 0, V
can be rewritten as
103 V= x'Px + a(<7 — c'x) 2 + fl<b(a)

where a = ctjd 2
. Actually, Popov begins with a more general Liapunov
function of the form
104 V= x'Px + a (<r — c'x) 2 + /?$(cr) + £v'x

He then shows that if V is a p.d.f. and is a Liapunov function for all
systems of the form (88)-(90), then v must be zero. Thus, in effect, there
is no loss of generality in studying F-functions of the form
(103) in-
stead of (104). The details can be found in [16] Lefschetz [pp. 98-99].
Calculating the derivative of F along the trajectories of (88)-
(90), we get
105 V= + 2a (a — c'x)d£ +
2x'Px
= x'(A'P + PA)x - 2x'Pb0 - 2ad(a - c'x)0
+ c'Ax — c'b 0 — d<j>)
where 0 is shorthand for 0 (<j). Now, V can be rewritten as

106 V= x'(A'P + PA)x - 2x'(Pb - v)0 - y0
2 - 2aâ0
where
107 v A Â'c + a dc
108 y A + d) jff(c'b
We are now in a position to state Popov’s criterion precisely.

109 Theorem The system (88)— (90) with d =£ 0 is absolutely stable for
all nonlinearities 0 belonging to the sector (0, oo) if there exists a
nonnegative constant r such that
110 Re [(1 + jrco)g(jco)] >0, V co e R
where
111 g(j) = -^ + c'(jI - A) -1 b

,
proof We prove the theorem by once again using the Kalman-

Yacubovitch lemma. Notice, first, that (110) is equivalent to
112 rd + Re [(1 + jrCO)c'(jCO\ — A) _1 b] >0, V CO
We now replace (112) by

113 Pd + Re [(2<x,d + jcoP)c'(jcoI - A)" b] >0, 1
V co
where a > 0 and p > 0 are chosen such that
We next show that if (1 1 3) holds for a particular set of values a and /?

then the corresponding V given by (103) is a Liapunov function for the
system (88)-(90), provided of course that P is chosen properly. Specifically,
Theorem (13) assures us that if Q is any positive definite matrix and
115 y + 2 Re \'(jcol — A) _1 b >0, V co e R

then there exist a positive definite matrix P and a vector q such that
116 A P + PA = — 6Q - qq'
117 Pb - v = y^ 2 q
We shall return to condition (115) later on. Now, if (115) holds, then V
becomes
118 V = — ex'Qx — (x'q) 2 — 2x'q}> 1/2 0 — y0 2 — ladofy
= — 6xQx — (x'q + y 1/2 2 — 20Ldo(j)
(j))
Now, because P is positive definite and a > 0, V is a p.d.f. in x and

Next, it is clear from (118) that V is always nonpositive [recall that
ff0(<7) >0, V cr, because 0 belongs to the sector (0, oo)]. Thus, to show
that — V is an 1. p.d.f. over all of R n ,
it is enough to show that V< 0 when-
ever either x or £ is nonzero. If x ^ o, V < 0 because Q is positive definite.
On the other hand, if x =
0 and £ ^
0, then a ^
0; this shows that
<70(<r) > 0, because 0 belongs to the sector (0, oo). Because a d 0, it >
follows that V< 0 in this case also. Hence — V is an l.p.d.f. over all of Rn ,
and the system (88)-(90) is globally asymptotically stable.

We see, therefore, that (115) is a sufficient condition for absolute stabil-
ity. Substituting for v and y from (107) and (108), we get
119 P(c'b + Re (Pc'A + 2adc')(jcol - A)" b > 0

+ d) J
120 Pd +. Re pc'[l + A(jcoI - A) _1 ]b + 2ocdRc c'(jcoI - A)"^ > 0

121 Pd + Re [j co Pc'(j col - A) _1 b] + 2adRc c'(jcol - A) -1 b > 0
122 Pd + Re [(2a d + jcoP)c'(jcol - A) _1 b] >0, V co e R
In going from (120) to (121), we use the obvious identity
123 I + A(jcoI — A)" 1

= jco(jcol — A)" 1
Finally, we see that (122) is the same as (113). This proves the theorem. M
As it stands, Theorem
(109) shows that (1 10) is a sufficient condi-
tion for the existence of aLiapunov function of the form (103) for the
system (88)-(90). However, Popov shows that (110) is also a necessary
condition for the existence of a Liapunov function of the form (103).
This should not be misinterpreted to mean that (1 10) is a necessary and
sufficient condition for the absolute stability of the system (88)-(90).
Rather, (110) is a necessary and sufficient condition for the existence of
a Liapunov function of the type (103).
Geometric Interpretation
Like the circle criterion, the Popov criterion can also be given a
geometric interpretation. Suppose we plot Re g(jco)\s. colmg(jco), with
co as a parameter. Note that we only need to do the plotting for co 0, >
because both Re g(jco) and co Im g(jco) are even functions of co and no ,
new information is Such a plot is known

gained by plotting for co < 0.
as a Popov plot in contrast with a Nyquist plot, which is a plot of
,
Re g(jco) vs. Im g(jco). Now, the condition (110) simply states that there
exists a nonnegative number r such that the Popov plot of g lies to the
right of a straight line of slope 1 jr passing through the origin. If r 0, =
this straight line is taken as the vertical axis of the Popov plot. The
situation is depicted in Fig. 5.12.
Nonlinearity in a Sector (0, k)
Suppose now that we are interested in the absolute stability of the

system (88)-(90) for all nonlinearities 0 belonging to the sector (0, k),
where k isa given finite number. Because the class of allowed nonlineari-
ties is smaller than before, one would expect to find a weaker sufficient
condition for absolute stability than (110). And this is indeed what one
finds.
124 Theorem A sufficient condition for the system (88)-(90) to be

absolutely stable for all nonlinearities 0 belonging to the sector (0, k)
is that there exists a nonnegative number r such that
125 Re [(1 + jcor)g(jco)] + -^ > 0, V co e R
proof The present problem is reduced to the problem covered by

Theorem (109) by means of loop transformations. Suppose 0 belongs to
the sector (0, k). Then the system (88)-(90) can be transformed to that in
Fig. 5.13, where the new forward transfer function is
126 £i(s) = g(s) + -j
g ,(.v)
L
FIG. 5.13 01
while the new feedback nonlinearity is
127 & =0 • (:1 -x '

f)
Note that, because 0 belongs to the sector (0, k), the inverse in (127) is
well defined. 0j belongs to the sector (0, oo). Therefore, by Theorem

Now,
(109), a sufficient condition for the absolute stability of this system is that
there exists a nonnegative constant r such that
128 Re [(1 +jcor)g 1

(jco)] >0, V CO e R
It is easy to see that (128) and (125) are equivalent. bb
Remarks: In the case where 0 belongs to the sector (0, k), we use
the sufficient condition (125), which is weaker than (110). Clearly,
as k —> oo ? (125) reduces to (110). The geometric interpretation of

(125) is similar to that of (110): (125) states that the Popov plot of g
lies to the right of a straight line of slope l/r passing through the
point —1/k+jO.
129 Example. Consider a system of the form (88)-(90), with
130 - 5(JTT?
In this case
la
R'^-frrl 5 5
)
2
132 co Im g(ja>) = — 1
(1
The Popov plot of g is shown in Fig. 5.14. It is clear from this figure
that (125) can never be satisfied unless — l/k < — i.e., k < 2. On the
other hand, if k < 2, (125) can always be satisfied by a suitable choice
of r. Hence this system is absolutely stable for all nonlinearities in the
sector (0, 2).
5.5.5 Concluding Remarks
As is the case with the circle criterion, one of the main advantages
of Popov’s criterion is that it is applied directly to the transfer function
g(^) and is thus independent of the state variable realization (88)-(90).
In fact, there is no need to construct a state variable realization for a
given g(s).
In Chap. 6, we shall encounter another Popov criterion, which is
derived by using functional analytic methods. In contrast with the

circle criterion(where the functional analytic methods yield a true gen-

eralization of the Liapunov methods), the Popov criterion of Chap. 6
differs in some details from the present one.
The noteworthy thing about Popov’s criterion is the presence of
the term 1 + jcor, which is known as a multiplier. Other, more general
multipliers are studied in [17] Narendra and Taylor.
Problem 5.19. Consider a feedback system with
i(s) =
(s + l)(s + 2)(s + 3)
Using the circle criterion, determine the largest range [OC, /?] such that the
system is absolutely stable for all (possibly time-varying) nonlinearities <f>
in
the sector [a, ft].
Problem 5.20. Consider a feedback system with
^= (s - l)(s + l) 2
Using the circle criterion, determine the largest range [OC, /?] such that the sys-
tem is absolutely stable for all nonlinearities in the sector [OC, /?].
Problem 5.21. Given a system of the form (1)— (3) with
apply the Lur’e-Postnikov criterion with (a) Q = I, (b)
Compare.
Problem 5.22. Consider a system of the form (88)— (90), with
—
s(s -f 1)
Using the Popov criterion, show that this system is absolutely stable for all
time-invariant nonlinearities 0 belonging to the sector (0, k), where k

is any
finite number. Verify also that (110) does not hold in this
particular case.
'
5.6
SLOWLY VARYING SYSTEMS
In this section, we shall prove a very general result on the stability of
so-called slowly varying systems. The approach given here is
due to
[3] Barman.
Consider a general nonautonomous system described by
1 ±(t) =f [t, x(0L t >0

s :
Sec. 5.6 Slowly Varying Systems 219
where f : R x Rn — *•
R". If we now consider the autonomous system
2 x(t) = i[p, x(t)], t >0

where p a particular nonnegative number, then we can think of (2)
is
as a particular case of the system (1), with its time dependence “frozen”
at time p. This concept can be clarified by considering linear systems.
Suppose we are given the nonautonomous system
3 x(t) = A(t)x(t)
Then, for any p > 0, the autonomous system
4 x(t) = A(p)x(t)
can be thought of as describing the system (3), frozen at time p.
In this section, we shall address ourselves to the following ques-
tion: Suppose that, for each p > 0, 0 is a globally asymptotically stable

equilibrium point of the frozen system (2); under what additional condi-
tions is it possible to conclude that 0 is a globally asymptotically stable
equilibrium point of the original system (1)? The answer turns out to
be that the additional requirements are (1) the global asymptotic sta-
bility of 0 for each of the frozen systems is in some sense uniform with
respect to p, and (2) the rate of variation of f with respect to t is suffi-
ciently small (hence the name slowly varying systems).

To streamline the presentation of the results, we shall employ the
following notation: As in Sec. 5.1, s(f, x, t 0) denotes the solution of (1)
as evaluated at time t, starting from the initial state x at time t0 . Simi-
larly, s p (t, x, t 0 ) denotes the solution of (2) as evaluated at time t,
starting from the initial state x at time t 0 The symbol is used to . D

denote partial differentiation. For instance
5 D i(t, x) =
t
and so on. Throughout, denotes the / 2 -norm. ||

• ||
The main result for nonlinear systems is given next.
6 Theorem Suppose the function f: R X Rn — R >

n
satisfies the fol-
lowing conditions
7 (i) 0) = 0,\?t>0.
(ii) f is continuously differentiable.
(iii) f is globally Lipschitz-continuous ;
i.e., there exists a finite
constant k such that
8 ||f(7, x) — f(t, y) ||
< &||x — y ||, V t > 0, V x, y e R*
Suppose the family of frozen systems (2) satisfies
9 II p (t, x, t 0 )\\ < m\\x\\ exp [-a - )l (t t0
V > V > t t0 , *0 o, v x g Rn v P
,
>o
for some finite positive constants m and a. Under these conditions,

there is a number <5, which depends only on k, m and a, such that ,
the equilibrium point 0 of (1) is globally asymptotically stable

whenever
10 \\D t
f(t, x) ||
< <5||x||, \/ t >0, \/ x g Rn
In fact, 8 can be so chosen that the solution trajectories of (1) satisfy
11 II s(/, X, Oil < ™il|x|| exp [~p{t - O],

V > t t0 , V t0 > 0, V x e Rn
proof The function s p (t, x, t 0 ) satisfies the integral equation
12 s P it, x, t 0) = x + f l[p, s p (t, x, t a )]

Jto
dr, V t > to
Because the frozen system (2) is autonomous, there is no loss of informa-
tion in studying only s p (t, x, 0). From (12) we have
13 s p {t, x, 0) =x+
JQ
f [p, s^Ct, x, 0)] dx, V t >0
14 ||f[i7,s i,(T,x,0)]||> -&||s p (T,x,0)||> -km II X II , V T>0

Substituting from (14) in (13), we see that
15 1 1
s p {t, x, 0) 1 1
> whenever t e [o,
Next, we differentiate both sides of (13) with respect to p to get

1
16 D p s p (t, x, 0) = f \Df[p, s p {x, x, 0)] • DpSpir, x, 0)

Jo
+ D pi[p, s p (t, x, 0)]} dx

Now, because f satisfies the Lipschitz condition (8), can be shown that
it
the induced /2 - norm of the Jacobian matrix Dî{t, x) always less than or
is
equal to k. Moreover, Dpf(p, x) satisfies (10) (where 8 is to be specified
later). Utilizing these facts in (16), we get
17 1 1
DpSpit, X, 0) 1 1
<J ‘
[*1 1
D p s p {t, X, 0)|| + <5|| s p (r, x, 0) ||] dr
0
< 8J w||x|| exp (— «t) dr + k\\DpS p (r, x, 0) ]| dr
<^\\x\\ + j\\\DpSpir, x, 0)|| dr
Applying the Bellman-Gronwall inequality (Appendix I) to (17), we get
18 II
-DpSjO, x, 0)|| < ||x||^ exp kt
Next, we construct suitable Liapunov function candidates for the

family of frozen systems (2) as well as for the original system (1). Choose
y > 0 such that
19
and choose
20 Vp (x) = \\s p (r, x, 0)]r dr

Jq
= [s P (r, x, 0 )s^(t, x, 0)p dr

Jq
as a Liapunov function candidate for the frozen system (2). Note that
Vp (x) is purely a “conceptual” Liapunov function candidate, because
determining Vp (x) requires explicit knowledge of the solution trajectories

s p (r, x, 0). From (15) and (9), it is easy to show that Vp is both decrescent
and a p.d.f. In fact, from (15) and (9), we have
21
To calculate the derivative of

IWI
2
^
Vp (-)
V/>>0,
along the solutions of
VxeR-
(2), we observe
that
22 V(x) = lim
n
o+
Now
23 Vp [sp (h, x, 0)] = 1
s p (h, x, 0), 0] |
p' dx
Jo 1
However, because the system (2) is autonomous, we have
24 s P [T, s p (h, x, 0), 0] = s p (r + h, x, 0)

Therefore
2' 2'
25 Vp [sp (h, x, 0)] = s P (x + h, x, 0)| dx = ||s„(t, x, 0)|| dx
J“ || I
J”
26 Vp [s{h, x, 0)] - Vp (x) = -J*||s,(r, x, 0)||

2 >’
dx
27 Vp (x) = -||x|| 2 >’
because s p (h, x, 0) — * x as h — >• 0. Thus Vp satisfies all the hypotheses of

Theorem [5.2(58)] and a Liapunov function for the frozen system (2).
is
We shall now complete the proof of the theorem by demonstrating that

the function V(t, x) defined by
28 V(t, x) = Vp (x)\ p=t = ||s,(T, x, 0)11^ dr

£
is a Liapunov function for the original system (1), provided 8 is sufficiently
small. First, because (21) holds, we see that V is both decrescent and a
p.d.f. Next, calculating the derivative of V along the trajectories of (1), we
get
29 V(t, x) = D x V(t, x)-f (f, x) +D t V(t, x)
= -||x|| *? + D t V(t,x)
c 1
where we use (27). Now,

<y— 1 > dr
30 D,V{t, x) = J” y\ s r (T,
I
X, 0) | p - [s r (T, X, 0)]'-As,(T, x, 0)
Hence
31 |
D, V(t, x) |
< J“ y\ s t (r, |
x, 0) 1
2 » 1 1
s,(t, x, 0) 1 1
•
1 1
As f (r, x, 0) 1 1
dr
Substituting from (9) and (18) in (31), we get
32 |
D,V(t, x)|
<J
r°°
y||x|| >7n
2 2 >'- 1
exp [-(2 y - l)at] • —
8m
exp kr dr
= ||x|p<$ f ^-exp {[k — (2 y — l)a]t} dr

Jo
From the manner in which y was chosen [namely, (19)], we have

33 k - (2 y
- l)a < 0
Hence the integral on the right-hand side of (32) is finite; let us denote it
by M. Clearly M depends only on m, k ,

and a. Now if
34 <5*M < 1
it is clear from (29) that — V is a p.d.f. Hence V satisfies all the hypotheses
of Theorem [5.2(58)]. Therefore we conclude that V is a Liapunov func-
tion for the system (1) and that 0 is a globally asymptotically stable equi-
librium point of (1).
To prove that the solution trajectories of (1) actually satisfy (11), we

proceed as follows: From (21), (29), and (34), we see that there exist finite
positive constants c l9 c 2 , and c 3 such that
35 Cl \x\\
|
2 r<V(t,x)<c 2 \\x\\ 2 y
36 V(t, x) <• — 3 || x ||
2y
Therefore, along solution trajectories of (1), we have
37 -jj
V\t ,
s (t, x, /o)] < -c 3 ||s(r, x, r 0 )||
2y
< =c~- V[t, s 2

(t, x, t 0 )]
From (37), we get, after using the integrating factor exp [— c 3 (t — t Q )/c 2 ],
~ C3 ~"~
38 V[t, s(t, x, f 0 )] < V(t, x) exp
[
'
c2 ]
Using the bounds (35) in (38) gives
39 Ci||s0, x, / 0 )IP y <c 2 ||x||

2 >'
exp
[
Zl£
^ Zl£
^]
One can now obtain (11) from (39) by routine manipulations. ^

40 Remarks : What Theorem (6) shows is that there is a 8 such that
whenever (10) holds 0 is a globally asymptotically stable equilibrium
point of (1). In other words, if 0 is a globally asymptotically stable

equilibrium point of each of the frozen systems (2) and if the rate of
variation of / with respect to t is sufficiently slow, then 0 is also a
globally asymptotically stable equilibrium point of the original sys-
tem (1).
The specialization of Theorem (6) to linear systems is perhaps

more easy to visualize.
41 corollary Consider the differential equation

42 ±(t) = A(t)x(t)
Suppose
(i) A(-) is continuously differentiable.
(ii) A(-) is bounded.
(iii) There exists a positive number a such that, for each t > 0, the
eigenvalues of A(0 all have real parts less than or equal to —a.
Under these conditions, there exists a number 5 > 0 such that 0 is a
uniformly asymptotically stable equilibrium point of (42) whenever
43 l|A(0llz < S, Vt> 0

where 11*11, denotes the induced /2 -norm.
e
Input-output stability
In this chapter, we some basic results concerning input-output

present
stability. This subject is of more recent origin than Liapunov stability,
and was pioneered during the 1960’s by Sandberg and Zames (see [21]-
[23] Sandberg and [27, 28] Zames). There exists a large
amount of litera-
ture on this subject, and a thorough treatment can be found in [10]
Desoer and Vidyasagar and [26] Willems.
Before proceeding to the study of input-output stability it is ,
necessary to reconcile the input-output approach to system analysis

with the state variable methods. The preceding four chapters are
predicated on the system under study being described by a set of differ-
ential equations that describe the time evolution of the system state
variables. In contrast, the systems encountered in this chapter are
assumed to be described by an input-output mapping that assigns, to
each input a corresponding output. In view of this seeming disparity,

,
it is important to realize that an input-output representation

and a
state representation are two different ways of looking at the same
—
system the two types of representation are used because they each
give a different kind of insight into how the system works. It is now
known that not only does there exist a close relationship between the
input-output representation and the state representation of a given
system but that there also exists a very close relationship between the
kinds of stability results that one can get using the input-output approach
and the state variable approach.
224
1
Sec. 6. 1 Introduction to L p -Spaces and Their Extensions 225
At one may well ask, why not simply use one of the
this stage,
approaches —why
both? The answer is that while the two
use
approaches are related they do not give identical results. Because, in
,
analyzing a system, it is advantageous to have as many answers as we

can, it is good to have, at our disposal, both the approaches, each
yielding its own set of insights and information.
This chapter is organized as follows: In Sec. 6.1 we shall give a
very brief introduction to L^-spaces and their extensions and introduce
the concepts of truncations and causality. In Sec. 6.2 we shall explore
the relationship between the input-output representation and the state
representation of a system. In Sec. 6.3 we shall present the basic
definitions of input-output stability. In Sec. 6.4 we shall give some of
the relationships between input-output stability and Liapunov stability.
In Sec. 6.5 we shall discuss the open-loop stability of linear systems,
while Sec. 6.6 contains some results on linear time-invariant feedback
systems. The material in these two sections serves as a stepping stone
for Sec. 6.7, where we shall state several criteria for the stability of
nonlinear time- varying feedback systems.
6-1
INTRODUCTION TO ^-SPACES AND
THEIR EXTENSIONS
In this section, we shall give a brief introduction to L p -spaces and their
extensions, and to causality. We do not require any deep results from

Lebesgue theory other than the completeness of the L p -spaces.
6.1.1 /.^-Spaces
We begin with a pedestrian introduction to Lebesgue measure; a thor-
ough treatment can be found in [20] Roy den. Given an interval E = (a, b ) in
the real number system R ,
define r(E) = b — a. Clearly every subset of R
is contained in a countable union of open intervals. Given any subset A
of R define
,
H*(A) = inf | X T(Et): AC (J E t

•
[ i= 1 i=
where the Efs are all open intervals. Note that ijl*(A) is always nonnegative
and may be infinite for some subsets of R. The function u* is called an j
outer measure. A subset S of R is measurable if
M%4) = ii*(A nS) + fi*(A n Sc ) for all subsets ,4 of R

c
where S denotes the complement of S. If S is measurable, n*(S) is called
the Lebesgue measure of S. All open and closed sets are measurable, and
a countable union of measurable sets is again measurable. A function
/: R—+R is measurable if the set
1
f~ (A) = {x GR: f(x) GA} is mea-
226 Chap. 6 Input-Output Stability
surable whenever A is an open subset of R. In particular, a continuous

1
function / measurable because f~ (A) is open whenever^, is open. The
is
definition of the Lebesgue integral is more involved, but the Lebesgue

integral of / reduces to the Riemann integral if / is continuous everywhere
except on a set of measure zero.
1 [definition] For all p e [1, oo), we label as Lp [ 0,

oo) the set
consisting of all measurable functions /(•).' [0, oo) — > R such that
2 r \f(t)Y dt < oo
Jo
The label LJfi, oo) denotes the set of all measurable functions
/(•)• [0> °°) — R that are essentially bounded on 1
[0, oo).
Thus, for 1 < p < oo, Lp [ 0, oo) denotes the set of measurable
functions whose pth powers are absolutely integrable over [0, oo),
whereas ZÔ, oo) denotes the set of essentially bounded measurable
functions.
3 Example. The function f(t ) = e~ at

, a > 0, belongs to Lp [ 0,
oo)
for all p. The function g(t) = l/(t + 1) belongs to Lp [ 0,
oo) for all
p > 1 but not to L x [ 0, oo). The function
4 m= L*
1/2
(i +
i
log 0-
12 /p
where p e [1, oo), belongs to the set Lp [ 0, oo) but does not belong to
Lq [ 0,
oo) for any p.
5 fact Whenever p e [1, oo], the set Lp [ 0, oo) is a real linear vector space,
in the sense of definition [3.1(1)].
Thus, whenever /(•),#(•) belong to Z^fO, oo), we have that (/ + #)(•)
also belongs to L p [0, oo), as does (&/)(•), for all real numbers a. This fact
follows from what is known as Minkowski’s inequality (see [20] Royden).
6 [definition] For p e [1, oo), the function LJL 0, oo)

[0, oo) is defined by
7 11 /( 011 ,= T
-Jo
i/(o i'*
l/p
The function ||
• ||„: T„[0, oo) — > [0, oo) is defined by 2
8 11 /(*) II ~ = ess. sup. |/(t)|

re [0, oo)
^‘Essentially bounded” means “bounded except on a set of measure zero.”

The reader need not worry about the distinction between “essentially bounded” and
“bounded.”
2 For “ess. (= supremum), one can
sup.” essential safely read “supremum.”
Sec. 6. 1 Introduction to L p -Spaces and Their Extensions 227
Definition (6) introduces functions •

|p for 1 || <p < oo, that
map the linear spaceLp [ 0, oo) into the interval [0, oo). Note that, by
virtue of definition (1), the right-hand sides of (7) and (8) are well
defined and finite.
9 fact (Minkowski's inequality) Let p E [1, oo], and let /(•), g(')
e Lp [ 0, oo). Then
10 \\(f+ g)(-)\\ P <\\f(-)\\ P + \\g(-)\U
Thus fact (9) shows that p satisfies the triangle inequality. ||
•
\\
Because axioms (Nl) and (N2) of definition [3.1(25)] are easily verified,
we see that the pair (Lp [ 0, oo), ||
. \\ p ) is a normed linear space, for each
P e [1, oo].
11 fact For each p e [1, oo], the normed linear space (L p [0, oo), ||
•
|p is
complete and hence a Banach space. For p

is = 2, the norm ||
•
|| 2 cor-
responds to the inner product
12 </(•),*(•)>» = \j{t)g{t)dt
Thus (L 2 [0, oo), <•, «> 2 ) is a complete inner product space, i.e., a Hilbert
space.
Fact (11) brings out the main reason for dealing with L^-spaces
in studying input-output stability. We could instead study the set
Cp 0, [
oo), which consists of all continuous functions /(•) satisfying
(2). Then Cp 0,
is a subset of L 0, oo), and the pair (C [ 0, oo),
p [
oo)
p [
||
•
IP normed linear space. However, it is not complete.
is also a
Hence, if we require a Banach space as the setting for the stability
theory, it is better to work with the spaces Lp [ 0, oo). Note that
(IJ0, oo), |p is the completion of the normed linear space
|| .
(CJ0, oo), || . |p.
13 fact (Holder' s inequality) Let p, q E [1, oo], and suppose
(Note that we take p = 1 if q = oo and vice versa.) Suppose/(-) e L p [0, oo)

and g(*) e L q [0, oo). Then the function
15 h(t) A f(t)g(t)
belongs to LfiO, oo). Moreover
16
Jo
"
i m i
•
i
sit) dt i
< [ J“ i m \> dtj*
[ J“
i
g(t) i« dt
1/9
]
The inequality (16) can be concisely expressed as
17 !l/(-k(-)lli^ll/(-)IUk(-)ll 9
/
6.1.2 Extended /.^-Spaces
Once the concept of Z^-spaces is understood, we can define

extended L^-spaces. First, we require the notion of truncated functions.
18 [definition] Let x(*): [0, oo) — > R be measurable. Then for all
Te [0, oo), the function xT {>) defined by
19
0 <t<T
T< t
is called the truncation of x(-) to the interval [0, T].
20 [definition] The set of all measurable functions /(•): [0, °°) > —
R ,
such that /r («) e Lp [ 0 ,
oo) for all T, is denoted by L pe [0 oo) ,
and is called the extension of Lp [ 0,

oo).
Thus the set Lpe [ 0,

oo) (which is also sometimes called the
extended Lp -space) consists of all measurable functions /(•) which
have the property that all truncations fT (» ) of /(•) belong to L p [ 0, oo),
although /(•) itself may or may not belong to Lp [ 0,
oo).
21 Example. Let /(•) be defined by
22 m= t
Then for each T< oo, the truncated function r (-) is the triangular
pulse function defined by
23
T< t
It is clear that, for each finite value of T, the function /r (-) belongs to
all thespaces ZJ0, oo). Hence the original function /(•) belongs to
each of the extended spaces L pe [0, oo), for each p e [1, oo]. However,
/(•) does not belong to any of the unextended spaces Lp [0, oo).
The relationships between the unextended spaces Lp [0, oo) and
the extended spaces Lpe [ 0,
oo) are further brought out in the following.
24 lemma For each p e [1, oo], the set L pe [0, oo) is a real linear vector
space. If p e [1, oo] and /(•) e L pe [0, co), then
(0 ||/H*)lh> isa nondecreasing function of T.
(ii) /(•) E L p [0, oo) (the unextended space) if and only if there exists a
finite constant m such that
25 ||/r(*)llp < V T< oo
In this case
26 11/(011, = lim \\M-)\\p

T-* oo
]] ,
Sec. 6. 1 Introduction to L p - Spaces and Their Extensions 229
is almost self-evident and is left as an exercise.

The proof
Thus, in summary, the extended space Lpe [ 0, oo) is a linear
space that contains the unextended space Lp [ 0, oo) as a subset. Notice,
however, that while L p 0,
[
oo) is a normed linear space, Lpc [0, oo) is
only a linear vector space and cannot be made into a normed linear
space.
27 [ definition The symbol L"[0, oo) denotes the set of all n-tuples
f = [/i . . . where e Lp [0, oo) for i — 1,
f, n. L pe [0,
n
. . .
oo)
is defined similarly. The norm on L"[ 0, oo) is defined by

l/1
28 l|f(-)ll,A[gll/XOII5]
norm of a vector-valued function f(-) is the

In other words, the
square root of the of the squares of the norms of the component
sum
functions /(•)• This choice is to some extent arbitrary but has the
advantage that, with this definition, L 2 [ 0, oo) is a Hilbert space.
n
Remarks on Notation: Hereafter, we shall use Lp to denote

Lp [ 0, oo); Lpe ,
L”, and L npe are defined similarly.
6.1.3 Causality
We shall close this section by introducing the notion of causality.

If we think of a mapping A as representing a system and of ( Af )( •)
as the output of the system corresponding to the input/(•), then a causal
system is one where the value of the output at any time / depends only
on the values of the input up to time t. This is made precise next.
29 [ definition A mapping A\L pe -^> L™ is said to be causal if n

e
(Af) T = AfT) T V T < oo, V / e L pe n

30 ( ,
An alternative formulation of causality is provided by the

following.
31 lemma Let A:L npe > L pe Then A is causal — . in the sense of definition
n
(29) if and only if, whenever /, g e L pe and
32 fT = gr for some T< oo
we have
33 (Af)r = (Ag)r
proof
u n
(i) If Suppose A satisfies (32) and (33); we must show that A sat-
isfies (30). Accordingly, let / L pe and T < oo be arbitrary. Then
e
clearly (f) T = ifT)T, so that by (32) and (33) we have

34 (4/)r = (Afr) T
n
Because (34) holds for every T and for every / e L pe , (30) is
established, and A is causal in the sense of definition (29).

(ii) “ Only If” Suppose A satisfies (30). Let L npe and suppose fge ,
fr = gr for some T; we must show that (33) holds. Now, by (30),

we have
35 (Af) T = (AfT T )
36 (Ag)r — (Agr)T
Moreover, because fT = gT , (35) and (36) together imply that
37 ( Af) T = (Ag) T
which is the same as (33). H
Definition (29) and lemma (31) provide two alternative (but
entirely equivalent) interpretations of causality. Definition (29) states
that amapping A is causal if any truncation of Af to an interval [0, T]
depends only on the corresponding truncation fT of /. To put it
another way, the values of Af over the interval [0, T] depend only on
the values of / over [0, T Lemma
]. (31) states that A is causal if and
only if, whenever / and g are equal over an interval [0, T], we have that
Af and Ag are also equal over [0, T].
Problem 6.1. Determine whether each of the functions below belongs

to any of the unextended spaces L p and to any of the extended spaces L pe .
(a) fit) = te~ f
(b) f(t) = tan t
(c) fit) = 1/(1 +

1/2
t)
(d) fit) =
fO, if t =0
(e) /(/) = l/t 1/2
,
if 0 <f< 1
|
[l It, if 1 <t
Problem 6.2. Prove lemma (24).
Problem 6.3. Suppose H : L pe — » L pe is one of the form
(Hf)(t) = hit, T)/(T) dx
Show that H is causal if and only if
hit, t) = 0 whenever t <T

Problem 6.4. Suppose A L pe
: — > Lpe is of the form
dr
(Af )(t) = f a(t, X)n[f(x)}
Sec. 6.2 Input-Output and State Representation of Systems 231
where n: R —>R is not identically zero. Show that A is causal if and only if
a(t, t) = 0 whenever t < r
6.2
INPUT-OUTPUT AND STATE REPRESENTATION
OF SYSTEMS
In this section, we shall give a brief introduction to some results on
constructing a state representation of a given input-output relation-
ship. The purpose in presenting these results is to once again emphasize
that the input-output representation and the state representation are
two different ways of describing the same system.
Given is a system described in state variable form by the set of
equations
1 x(t)=f [t, x(V), u.(*)] ;

x(0) =x 0 (fixed)
2 y(0 = g [t, x(0, u (t)]
Suppose (1) has a unique solution for x(.), for each u(-) belonging
to a prespecified class of inputs. [This is the case, for example, if f is
continuous and satisfies a global Lipschitz condition in x for each fixed

ii(*)-] Then, once we know the x(-) corresponding to a particular input
u(-), we can use (2) to calculate the output y(-). Thus, conceptually at
least, it is possible to construct the input-output relationship of a sys-
tem given the state representation of the system.
For this reason, in this section we shall concentrate on the problem
of constructing a state representation given the input-output relation-
ship. There are many results available on this subject, and the brief
discussion below presents only the simpler results.
6.2.1 Linear Time-Invariant Systems
Suppose that a system has the input-output relationship
3 y(t) = f
H(f — t)u(t) dr
Jo
In words, the output y(«) corresponding to the input u(.) is given

by the convolution of u(*) and another function H(-)- The function
H(-) is characteristic of the system under study and is usually known
as the impulse response of the system. Assuming that the functions u(-),
y(-), and H(-) are all Laplace-transformable, one can express (3) in
the equivalent form
4 y(s) = &(s)Q(s)
where a caret denotes the Laplace transform.

Our objective is to find matrices A, B, C, D, if possible, such

that the system of equations
x(t) = Ax(t) + Bu(t) ;

x(0) =0
y(0 = Cx(t) + Du(?)
has the input-output relationship (3) [or, equivalently, (4)]. In this
case, (5)-(6) is a state representation of the input-output relationship
(3). The conditions under which this can be done, as well as some
properties of the resulting system of equations (5)-(6), are given next.
fact Given a Laplace-transformable distribution 3 H(-), one can find

A, B, C, D
such that the system of equations (5)— (6) has the input-output
relationship (3)if and only if H
(s) is a proper rational matrix in s i.e., ;
every element of the matrix H

(s) is a proper rational function of s.
fact Given a proper rational matrix H(Y), one can always find A, B, C, D
such that
HO) - COI - A) _1 B + D, Vs
and such that
rank [B AB . . . A” -1 B] =n
(A')”“ C ]=«
/
1
rank [C' A'C' ...
where n is the order of the matrix A. In this case, the set of eigenvalues of
A is the same as the set of poles of HO); i-e., a complex number k is an
eigenvalue of A if and only if H(*) has a pole at k.
The above results can be found in any standard book on linear

systems, e.g., [7] Chen.
In summary, facts (7) and (8) state that if HO) is a proper rational
function of then one can always construct a system of equations of
s,
the form (5)-(6) which has the input-output relationship (3). Further,
we can ensure that the system (5)-(6) is controllable and observable
and that the set of eigenvalues of A coincides with the set of poles of
HO).
6.2.2 Linear Time-Varying Systems
Consider now a linear system which has the input-output rela-

tionship
y(0 = f H(7,
Jo
t)u(t) dx
3
Note that if D ^ 0, then the impulse response of the system (5)— (6) contains
an impulse distribution at t = 0.
Sec. 6.3 Definitions of Input-Output Stability 233
We would like, if possible, to construct a system of equations of the

form 4
13 ±(t) = A(t)x(t) + B(f)n(f); x(0) = 0
14 y(t) = C(t)x(t)
such that the system of equations (13)— (14) has the input-output rela-
tionship (12).The conditions under which this can be done are given
next.
15 fact Given H(-, •), one can find A(«), B(*), C(«) such that the system of
equations (13)— (14) has the input-output relationship (12) if and only if
H (t, t) can be expressed in the form
16 H(/, T) = *(*)(T)
proof See [7] Chen.
6.3
DEFINITIONS OF INPUT-OUTPUT STABILITY
In this section, we shall introduce the basic definitions of input-output
stability.
1 [definition] Let A: Lnpe — > We say that the mapping A

L™e .
(or the system represented by the mapping A) is Lp -stable if

(1) Af g L p whenever f e L”, and (2) there exist finite constants
k b such that 5
,
2 JIAfflU <k\\f\\, + b, Vf eLp

Definition (1) can be interpreted as follows: Suppose that, from
an input-output point of view, a system with n inputs and m outputs
can be represented by the mapping A: L npe > L™e That is, given an — .
input f g Lnpe ,
the output of the system is Af g L pe Now, we say
.
that the system is Lp -stable if, whenever the input f belongs to the
unextended space L p the resulting output Af belongs to L p and more-
,
over the norm of the output Af is no larger than k times the norm of the
input f plus the offset constant b.
For example, if we put p = oo, then the concept of L^-stability

becomes what is commonly referred to as BIBO (bounded-input/
bounded-output) stability; namely, a system is L^-stable if, whenever
4 Note that, in no term of the form D(/)u(0- This is in the interests

(14), there is
of simplicity.
5 Note that we use ||
•
\\ p to denote the norms on L" as well as Lp .
) ^ „
/e
the input (is bounded), the output Af e L (is bounded) and
moreover (2) holds.

Note that some authors use the adjective “L^-stable” to describe
a system A that satisfies only condition 1 of definition (1) (i.e., inputs in
Lp produce outputs in Lp but with no norm restriction).
3 Example. Consider the mapping A defined by
4 (Af)(t) = (
,-«<»-*>/(T
) dx
Jo
where a > 0 is a given constant, and suppose we wish to study the

Loo-stability of this system. We first establish that A L :
ooe — > L^, so that
A is in the class of mappings covered by definition (1). Accordingly,
suppose / e Loo,. Then for each finite T, we have that \\fT \\oo is finite
(see definition [6.1(20)]. Hence, for each finite L, there exists a finite
constant mT such that
6
5 \f{t)\<m T a.e., V t e [0, T]
To show that Af e L ,,, let g denote Af. Then
6 git = (
dr
Hence, for t <T, we have
7 ['\e-^\.\f{r)\dr
|g(0l< Jo
< f mT |
e~ a{t
~ r)
1
dr = mT ^ ~~ e
a
—
Jo
— a
a.e.
The inequality (7) shows that g( •) is bounded over [0, T]. Because this
reasoning is valid for every finite L, it follows that g(-) g L^, i.e.,
that A Lê
.
>
Looe •
Next, we show that the mapping A is L^-stable in the sense of

definition (1). Suppose fe L^; then there exists a finite constant m
such that
8 \f(t)\<m a.e., V *> 0
By using exactly the same reasoning as before, we can show that 7
9 \g(t)\<^’
a
Vt> 0
6 “a.e.” stands for “almost everywhere,” which means “except on a set of mea-
sure zero.”
7 Hereafter we shall drop the phrase “almost everywhere,” it being implicitly
understood.
\ ^
Hence (2) is satisfied with k = 1/a, b = 0, which shows that v4 is
Loo-stable.
10 Example. To illustrate the difference between conditions 1
and 2 of definition (1), let A be the mapping defined by

11 (40(0 =
First, it is easy to show that A L ooe: —> because if
12 \f(t)\<m, V/g[0J]
then
13 \(Af)(t)\<m\ V * e [0, T]
Also, by the same reasoning, it is clear that Af e L whenever / e Loo.

Hence A satisfies condition 1 of definition (1). However, it is clear that
no k b exist such that (2) is satisfied, because the function
,
2
xhx
cannot be bounded by a straight line of the form x h* kx + b. This is
illustrated in Fig. 6.1.
14 Example. Consider the system whose input-output relationship

is
15 (Af)(t) = je‘~ Tf(t) dT
This mapping A also maps into itself. To see this, let / e L^.
Then, for each finite L, there exists a finite constant mT such that
16 \f(t)\<m T , V f e [0, T]
Thus, whenever t e [0, T], we have
17 | ( Af)(t ) |
< m T Jo f |
e
{t ~ r)
|
dr = m T(er — 1)
Now, for each finite T the right-hand side of (17) is a finite number.
Hence Af g L^.
On the other hand, A is not Loo-stable, as can be seen by setting
f(t) = 1, V t. Then /(•) e L^, but
18 (Af)(t) = f e {t
~ T)
dr = e* — l
Jo
^
which does not belong to L (although it does belong to Looe). Hence

there is at least one input in L whose corresponding output does not
belong to L which shows that A is not Z^-stable.
Remarks: Example (14) illustrates the advantages of setting up

the input-output stability problem in extended Z^-spaces. If we deal
exclusively with Z^-stable systems, then we can represent such a system
as a mapping from Lp (the unextended space) into itself, rather than
from Lpe into itself. However, if we are interested in studying unstable
systems (for example, the feedback stability of systems containing
unstable subsystems), then we must have a way of mathematically
describing such a system. This is accomplished by treating such a sys-
tem as a mapping from Lpe into itself.
We shall now turn to the definitions of feedback stability. Consider

the feedback system shown in Fig. 6.2, where the various quantities
u 1? u 2 e 1? e 2 y l5 y 2 are
, ,
all members of some extended space L” e and the ,
operators G t
and G map L pe
2 into itself. The equations describing this
feedback system are
19 e, =u -y x 2
20 e = u + yi
2 2
21 y, = G^
22 y2 = G2 e 2
The idea is that G t
and G2 represent two given subsystems which are
interconnected in the manner shown, u t and u 2 represent the (known)
inputs, e 1and e 2 are the (unknown) “errors,” and y and y 2 are the 1
(unknown) outputs.
In analyzing the system described by (1 9)— (22), two distinct
types of questions arise. It is clear that we can eliminate y l5 y 2 from
(19)-(22) and rewrite the system equations as
23 ej = Uj — G e 2 2
24 e 2 = u + Gjej 2
Alternatively, we can eliminate e l9 e 2 and rewrite the system equations

as
25 y, = GÛ! - y 2)
26 y2 = G (u + y,) 2 2
The system descriptions (19)— (22), (23)-(24), and (25)-(26) are all
equivalent. With regard to this system, the first type of question that
one can ask is the following: Given inputs u and u 2 in L pe
n
does there 1 ,
exist a setof solutions for e l5 e 2 y 1? y 2 in L npe such that the system

,
equations are satisfied, and if so, is this set of solutions unique? This is
usually called the existence and uniqueness question. The second type
of question takes the following form: Given inputs u 1? u 2 in L”, assum-
ing that solutions in L npe exist for e 1? e 2 y 1? y 2 do these solutions actually
, ,
belong to L np ? This is called the stability question. The reason for making
the distinction between the two types of questions is that, in practice,
different types of conditions arise in answering the two types of ques-
tions. For instance, existence and uniqueness of solutions can be assured
under very mild conditions [usually some type of global Lipschitz
conditions; see Theorem (27).] On the other hand, stability requires
stronger conditions, which are the main subject of this chapter.
Accordingly, we shall state below a theorem regarding the exist-
ence and uniqueness of solutions to ( 1 9)— (22) and then proceed to a
study of stability conditions. Theorem (27) is not the most general of its
kind, but it is easy to prove and is adequate for most situations.
27 Theorem Consider the system described by (19)-(22), and suppose

the operators G t and G 2 are of the form
28 (G.xXO = JoP GO, T)n,[T, x(t)] dr

29 (G 2 x)0) = n x(t)] 2 [r,
where G(«, •) is continuous, and the continuous functions n 1? n 2 :
R+ x R — R
n
>
n
(i) Are unbiased, i.e.,
30 n t
(t, 0) = 0 V >0 t
31 n 2 (^, 0) = 0 V >0 t
(ii) Satisfy global Lipschitz conditions; i.e., there exist finite con-
stants k and k 2 such
x
that
32 II n^, x) -n 1
(r,y)||<fc 1 ||x- y||,
V/>0, V x, y e Rn
33 \\n 2 (t,x) - n2 (f,y)||<fc 2 ||x- y||,
V t > 0, V x, y g Rn
Under these conditions, and G map L pe

2
n
into itself.
Further, given any u 1? u 2 e Lnpe ,

there exists exactly one set
of e l5 e 2 y 1? y 2
, ,
each belonging to Lnpe ,
such that (19)-(22)
are satisfied.
The proof is omitted here, because although it is quite straight-

forward, it is also rather messy. The interested reader is referred to
[10] Desoer and Vidyasagar [Chapter III], where a more general version
of this theorem is proved. The basis of the proof is the observation that,
under the stated assumptions, the set of equations (23)-(24) [which is
equivalent to (19)-(22)] reduces to a nonlinear integral equation of the
Yolterra type. Basically, Theorem (27) assures us that existence and
uniqueness are guaranteed provided G t
is a causal integral operator
(possibly nonlinear) G 2 is a (causal) memoryless nonlinear operator.
and
Actually, the roles of G and G in Theorem (27) can be interchanged,
t 2
and the conclusions still hold. In addition, even if both G and G 2 t
are integral operators of the form

and uniqueness still (28), existence
follow. The key is that at least one of the operators G and G 2 is an x
integral operator and as such introduces an element of smoothing in

the closed loop. The situation that is not covered by Theorem (27) is
the case where both Gj and G 2 are memoryless. In this case, existence
and uniqueness can fail in some seemingly innocent systems. See [10]
Desoer and Vidyasagar [Chapter I].
With this cursory treatment of existence and uniqueness results,
we shall turn now to the definitions of stability for the system described
by (19)-(22).
34 [definition] The system represented by (25)-(26) is said to be

Lp -stable if, whenever u 1? u 2 belong to L”, we have that any
corresponding y r y 2 in ,
Lpe
n
that satisfy (25)-(26) actually belong
to Lp n
,
and in addition, there exist finite constants k and b such
that
35 ||y 1 ||,<A:(||u 1 + ||u || i, 2 ||,) +A

36 ||y 2 IU < ^(IlUi IU + u II 2 IU) + b
whenever u l9 u 2 y l5 y 2 ,
satisfy (25)-(26).
Remarks:
1. Notice that, according to definition
(34), the question of the
Lp -stability
of the system (25)-(26) is quite divorced from
the question of existence and uniqueness of solutions to
(25)-(26). If u 1? u 2 belong to Z”, then in order for the system
(25)-(26) to be Z^-stable, one of two alternatives must occur:
(i) either there do not exist any y l5 y 2 both belonging to
,
Lpe ,
such that (25)-(26) are satisfied, or (ii) there do exist
y l9 y 2 in L pe
n
such that (25)-(26) are satisfied, and all such
solutions in fact belong to the unextended space L np and
satisfy (35)— (36). This, and only this, is what is meant by
Instability. There no a priori assumption that, corre-
is
sponding to a given set of inputs u l5 u 2 in L", there actually

n
exist y 1? y 2 in L pe that satisfy (25)-(26). However, if such
y 1? y 2 do exist, then they must actually belong to L p and

n
satisfy (35)— (36).

2. Definition (34) mentions only y t y 2 and does not say any- ,
thing about e 1} e 2 However, the requirements on e l5 e 2 are

.
implicit in definition (34). First, recall that the system

descriptions (19)-(22) and (25)-(26) are equivalent and
that L np is a linear space. Hence, if u l9 u 2 y l5 y 2 satisfy (19)
,
and (20), we see that e l5 e 2 also belong to L”. Next, if y i9 y 2

satisfy (35)— (36), then clearly, from (19)-(20), we have
37 He, H, < (k + 1X11.1,11, + ||HalU) + b
38 ||e 2 \\ p < (k + 1 )(| |
lit II* + ll^ll*) + b
Hence e l5 e 2 satisfy conditions similar to (35)— (36). Thus,
even though definition (34) does not explicitly mention e x
and e 2 the conditions imposed on yj and y 2 imply that e

, t
and e 2 satisfy similar conditions.

Actually, in definition (34), instead of imposing con-
ditionson y 1? y 2 we could have imposed similar conditions
,
on and the resulting definition would be entirely

e 1? e 2 ,
equivalent to definition (34), because (19) and (20) can be

rewritten as
39 yi =e -u 2 2
40 y2 = Uj - ej
Hence, if u 1? u 2 e l9 e 2 belong to
,
L pn and if e 1? e 2 satisfy
conditions such as (37)— (38), then y l9 y 2 also belong to L np
and satisfy conditions similar to (35)— (36).
3. Comparing definitions (1) and (34), one may feel that there
are two distinct concepts of Instability, one for open-loop
stability and one for feedback stability. However, this is
not the case. It can be shown that if we take the system
equations (25)-(26) as defining a relation between the input
pair (u 1? u 2 ) and the output pair (y l5 y 2 ), then definition (34)
requires the same properties of this relation as definition
(1) does of the operator A.

4. In definition (34), there is no requirement that the subsystem
operators t
G
and G 2 be themselves Z^-stable.
) ^ )
Because we have not as yet presented any means of testing for

stability, other than solving the system equations by brute force, the
following example is of necessity rather simple.
41 Example. Consider a system described by (19)-(22), with
~ r)
42 ( G x
x){t = f
Jo
e~ a{t x{x) dr
43 (G 2 x)(t) = kx(t)
with k a being given constants. Let u 2 (t

, =0 for simplicity. The
subsystem represented by G x
has a transfer function of \/(s + a), while
G2 represents a constant feedback gain of k. Thus, using elementary
control theory, we see that the closed-loop transfer function is
lf(s +a+k which corresponds to an impulse response of e~ {a+k)t

), .
Thus, corresponding to an input uf-) in there exists a unique set

of solutions to (19)-(22), 8 given by
~ T)
44 y i(0 = f e~ (a+k)(t u 1 (z) dx
Jo
45 y 2 (t) = kyft)
First, we see from Examples (3) and (14) that y u y 2 e L„ e whenever
u1 e L^. Next, based on the same examples, it follows that^ 1? y 2 e
whenever u e L provided a + k > 0. Thus the system is Z^-stable,
x
in the sense of definition (34), if a + k > 0. On the other hand, if a + k

< 0, then one can find inputs in [e.g., u (t) 1] such that the cor- x
=
responding y u y 2 do not belong to L^. Thus the system is Z^-unstable
if a k 0.
64
RELATIONSHIPS BETWEEN I/O STABILITY AND
LIAPUNOV STABILITY
In some cases, input-output stability methods can be used to establish
Liapunov stability. In this section, we shall present some results that
state such results precisely.
1 Theorem Consider a system described by the vector differential

equation
2 ±(t) = Ax(t) — l[t, x(?)] ;

x(0) =x 0
8 The existence and uniqueness of solutions can also be established using

Theorem (27).
Sec. 6.4 Relationships Between I/O Stability and Liapunov Stability 241
where all eigenvalues of the n X n matrix A have negative real parts,

and the mapping /: R + x R n » R n is continuous. Associated with —
system (2), define a corresponding nonlinear feedback system (shown
in Fig. 6.3) described by
3 e(r) = u(/) — f e Mt
~ r]
y(t) dr
Jo
4 y (t) =f [t, e(0]
With the stated hypotheses, if the feedback system (3)-(4) is L 2-

stable in the sense of definition [6.3(34)], then the equilibrium point
0 of system (2) is globally asymptotically stable.
Remarks: Theorem (1) states that, under certain conditions, the

global asymptotic stability of system (2) can be established by prov-
ing the Instability of the associated feedback system (3)-(4). As
yet, we have not given any methods for proving Instability — this
will be done later.
proof We wish to show that, for all x 0 the corresponding solution

,
x(*) of (2) has the property that x(/) —> 0 as t —

> oo. We begin by express-
ing (2) as the equivalent nonlinear integral equation
~ r)
5 x{t) = e At x 0 — f e A{t { [t, x(t)] dr
Jo
Next, we observe that if we let
6 u (t) = e At x 0
in the system (3)-(4) (i.e., the system of Fig. 6.3), then the “error” e(-) is
governed by (5). Now, because all eigenvalues of A have negative real

parts, it is clear that the function t h-» ||e A 'x 0 is dominated by a decaying
1|
exponential; i.e., there exist constants m> 0, a > 0 such that
7 ||e
Ar
x0 1|
< me~ at
, V f >0
where ||
•
||
denotes the Euclidean norm. Thus the particular input u(*)
defined by (6) belongs to L n2 Therefore, if the system (3)-(4) is I 2 -stable
.
(in the sense of definition [6.3(34)]), then both e(*) and y(-) belong to L\.
Let us define
~ r)
8 z (t) = e Mt y(r) dr
[ - >>
Then, by (3),
9 e(0 = u(0 - z (t)

As mentioned previously, e(-) in this case is the solution x(-) of (2). Also,
because of (7), we have that u(0 — > 0 as t — » oo. Thus if we can show that
z (t) —
» 0 as t —
> oo, it will follow that e(t) > 0 as — t — > oo, and the global
asymptotic stability of the equilibrium point 0 of (2) will be established.
We shall now show that z (t) — 0 as t — > oo, and this will conclude the
proof of the theorem (readers disinterested in the details can skip to the
next paragraph).
We have y(-) e L\ and
~ T)
10 zif) = e Mt y(x) dr
rt/2 rt
= e Mt
~ x)
y(r) dr + e Mt
~ x)
y(r) dr
JO Jt/2
f e ATy(t — T)dx+ Jt/2

= Jt/2 ‘~
f e A( t)
y(z) dx
after changing the dummy variable of integration in the first integral. Thus
by the triangle inequality, we get
11 l|z(0ll< fJkAT IU|y(f

Jt/2
t)|| dr r |k A(r -
+ Jt/2 T)
IH|y(T)||^T
where || •
|| f
denotes the induced /2 -norm on RnXn (see Sec. 3.2). Next, by
Schwarz’s inequality on L 2 we ,
get
' 2
12 ||z(0||
< [J (||e
At
||i)
2
dxY •
||y(f - r)||
2
dxj'
( 2 [JJ 2
+ 1/2
(||eA< '' i>
||i)2
1/2
dx
T •
[£ 2
||y(T)| 2
1/2
'
dx
T
< [J “ (lk A 1l0 ^]
;
2
•
[J~
||y(T)||
2
rfr]
U1 " 1/2
eAt d *]
+ [J0
°°(H lli)
2 •
[J(/ 2
||y(T)|prfr]
Now we use the following well-known result from real analysis: If

/(•) e L2 ,
then
13 j~f 2(x)dx >0 as t — * oo
This result, coupled with the fact that both Ar and

||e ||* ||y(-)ll belong to
T2 ,
implies that the right-hand side of (12) approaches zero as t — » oo.
This shows that z(t) — » 0 as t — oo.

B—
In some instances, the nonlinearity in the differential equation

(2) enters only via a scalar nonlinear function. In this case, the global
asymptotic stability of (2) can be established by studying the Instability
of a single-input/single-output system.
s
Sec. 6.4 Relationships Between I/O Stability and Liapunov Stability 243
14 Theorem Consider a system described by the vector differential

equation
15 x(0 = Ax(t) — b0|>, c'x(01 ; x(0) =x 0
where ail eigenvalues of the n X n matrix A have negative real

parts, (j) : R+ x R — R > is continuous, and the pair (A, c) is observ-
able, i.e.,
16 rank [c' |
A'c' |
. . . |(A')”
_1
c'] =n
Associated with the system (15), define the single-input/single-
output feedback system (shown in Fig. 6.4) described by
17 e(t) = u(t) — f c'e

A( '- T)
by(T) dx
Jo
18 y(t) = <ji[t, e(t)]
With the stated hypotheses, if the system (17)— (1 8) is Z, 2 -stable in the

sense of definition [6.3(34)], then the equilibrium point 0 of (15) is
globally asymptotically stable.
Remarks: Note that (15) is the same as the system studied in

Sec. 5.5 in connection with the Lur’e problem.
proof The proof closely parallels that of Theorem (1). First, (15) can
be expressed as an equivalent nonlinear integral equation
19 x(/) = 6?
Af
x0 — 1
e A(
'~ T)
b0[T, c'x(t)] dr
Jo
Therefore
20 c'x(0 = c'e A 'x 0 — P c'e A(f ~ T)

b0[T, c'x(t)] dr
J0
If we define
21
—
II o **
^ 'w'
22 u(t) = c'e A 'x 0
then (20) becomes
23 e(t) = u(t) — c'e A(r

~ T)
b0[T, e(r)] dr
> > >
which is the same as (17)-(18). Because w(*) e I 2 if the system (17)— (18)
,
is Testable, it follows that e I 2 Next, as in the proof of Theorem

.
(1), y(-) E I2 implies that z(t) — > 0 as t —

> oo and that e(t) >0 as —
t — oo. However, because e(t) is a scalar and x(0 is a vector, we need to
do some more work to show that x(f) — > 0 as t — > oo. If we define
v(0 = b>»(0 = b <f)[t, e(t)]
then the system description (15) becomes
i(t) = Ax(/) — v(/)

e(t) = c'x(r)
Now, using the facts that the pair (A, c) is observable and that e(t) —» 0,
v(0 — > 0 as t — oo, we can show by standard arguments from linear sys-
tem theory that x(f) — 0 as t — oo. —
Theorems (1) and (14) provide a glimpse of how one can prove
Liapunov stability results using the input-output approach. Specifically,
any criterion for the Instability of the type of system shown in Figs.
6.3 and 6.4 also yields a corresponding result for the global asymptotic
stability of systems of the type (2) and (15). The advantage of using
input-output techniques to tackle Liapunov stability problems is that
some of the stability criteria are better motivated in the input-output
framework. However, the disadvantage is that input-output techniques
yield global asymptotic stability or nothing at all. It is difficult to
estimate a “region of attraction” using input-output techniques.
Problem 6.5. Prove the following generalization of Theorem (14):

Consider a system described by the vector differential equation
±(0 = Ax(0 + bv(t); x(0) =x 0
o(t) = c'x(0 + dv(t)
Vit) = ~(f>[t, (7(0]
where all eigenvalues of A have negative real parts; 0: R+ x R — R is con-

>
tinuous; and the pair (A, c) is observable. Suppose 0(-, •) belongs" to the
sector [a, JJ], i.e.
a<7 2
< 0(7, a) < P<7 2
, V t > 0, V g e R
c'(5l — A) -1 b + d
FIG. 6.5
Sec. 6. 5 Open - L oop S tability of Linear S ystems 245
and suppose 1 + kd ^ 0, V k e [a, /?]. Under these conditions, if the system

shown in Fig. 6.5 is L 2 -stable, then 0 is a globally asymptotically stable equi-
librium point of the system (27)— (29) {Hint: Show that, for all x 0 we have
,
y e L2 ,
w(t) — dy(t) — > 0 as t —* co and hence w(t) — d(j)[t, —w(t)] — 0 as
t — > co. Thus conclude that w(t) — > 0 and hence x(t) — » 0 as t — » oo.}
65
OPEN-LOOP STABILITY OF LINEAR SYSTEMS
Before attempting to study the stability of interconnected systems such
as [6.3(19)-(22)], it is helpful to first obtain conditions under which the
operators Gi and G2 represent L p -stable subsystems. In this section, we
shall concentrate on linear systems for the most part and obtain necessary
and sufficient conditions for a linear system to be L p -stable. The term
open-loop stability refers to the fact that we are studying the subsystems
Gi and G 2 rather than the , overall closed-loop system, which is described
by [6.3(1 9)— (22)1 -
6.5.1 Time-Invariant Systems
Throughout most of this subsection, we shall study single-input/

single-output systems. Consider a single-input/single-output time-
invariant system, which is characterized by a scalar transfer function
h{s). Suppose that, in addition, h(s) is a rational function of s. It is well
known that such a system is Z^-stable (BIBO stable) if and only if
1. h(s ) is a proper rational function of s, and

2. All poles of h(-) have negative real parts.
However, the situation is more complicated if h(s ) is not rational.
Such a situation arises whenever h(>) is the transfer function of a distri-
buted system, such as an RC transmission line (integrated circuit) or
an LC transmission line (power line) or if h(-) represents a simple delay,
etc. In what follows, we shall give precise conditions under which a
scalar h(s) (rational or irrational) represents an Z^-stable system. These

conditions illustrate one of the chief advantages of the input-output
approach to stability, namely, that it places lumped systems [rational
h(s)} and distributed systems [irrational h(s)] in a unified framework.
(This is much harder to achieve using Liapunov theory.) Then, once the
single-input/single-output case is thoroughly analyzed, the results for
the multi-input/multi-output case follow easily.
To do this, we introduce the sets Cfc and Basically, (as shown
later) a is the set of stable impulse responses, and & is the set of stable
transfer functions. The precise definitions are given next.
/
1 [definition] The symbol a denotes the set of generalized func-

tions (distributions) of the form
fO, t < 0
2 f(t) = J oo
t> 0
U= 0
where <?(•) denotes the unit delta distribution, tt are nonnegative
constants, fl (.) is a measurable function, and, further,
3
1
S
= 0
l/d < 00
4 [\m\dt<™
Jo
The norm ||
•
|| a of a distribution /(•) in 0, is defined by
5 ll/(-)lla =L
= 1 0
I/I + Jof \m\dt
The convolution of two distributions /(•) and g(*) in a, denoted
by f *g, is defined by
6 (/ * g)(0 =[/(* — r)g(r) dr = f f(r)g(t - t) dr

Jo Jo
Thus, the set a consists of all distributions that vanish for t < 0,
and for t > 0 consist of a sum of delayed impulses and a measurable
function, with the additional property that the weights of the impulses
form an absolutely summable sequence and the measurable function
is absolutely integrable. One can think of Ct as the space L augmented x
by delayed impulses.
Note that, in computing the convolution of two distributions,
one should take
7 d(t — t t) * 8{t — tj ) = 8{t — tt — tj)

8 S(t - r,.) *fjt) =f - a (t t,)
In other words, the convolution of two delayed unit impulses with

delays of t t and t j9 respectively, is another delayed unit impulse of delay
tt + tj , and the convolution of a delayed unit impulse 8{t — t]) and
a measurable function fa (t) is the measurable function fa (t — t[). Thus,
given two elements /(•) and g(«) in a, of the form
i =0
10 g(t) = S giS(t - t!
g)
) + g (t)a
Sec. 6.5 Open-Loop Stability of Linear Systems 247
their convolution is given by
11 (/ * g)(t) = tt AgAt -
i =0 j= 0
t\
f) - & y)
+ £figjLt-tn
i=0
+ t=0
£g,f.(t-tP)
+ j0
f/aO - PgaiPdT
12 Example. The function

13 /i(0 -
belongs to d, whenever a > 0. The distribution
14 f2 ( t) =2
= 1 0 (z +
1
^
1)
— zT), T >0 a given number
which is a sequence of evenly spaced delayed impulses, belongs to d

because the sequence [l/(z + 2
l) ]r=o is absolutely summable. However,
the distribution
15 f (t) = ^ _L_ S(t — zT), T> 0 a given number

3
/=o + z 1
does not belong to d because the sequence [l/(z + 1)]q is not absolutely
summable. The distribution
16 /4 W = <5(0 +
belongs to d and has a norm of 2.
17 Remarks :
1. Note that L 1
is a subset of d; further, if /(•) e L ls then
18 ll/(*) lla
= 11/(0 Hi
2. If /(•) and g(-) are elements of d, and at least one of them

is in Li (i.e., does not contain any impulses), then /* g does not
contain any impulses. This is clear from (11).
As mentioned previously, the set d can be interpreted as the

set of stable impulse responses; i.e., a system with impulse response
/z(.) is Z^-stable for all p if and only if h(-) e d. To prove this important
result, we shall first derive some useful properties of d. The set d is
an example of what is known as a Banach algebra — this is implied by the

properties given below.
19 lemma The set d, together with the function ||

•
|| a and the convolu-
tion *, has the following properties:
(i) | H
la is a norm on d, and & is complete under this norm.
(ii) The convolution * is distributive; i.e.,
20 /* O + h) = f*g +/* h, V f,g,h g a
21 ( f+g)*h=f*h+g*h , V f, g, h G d
(iii) The convolution * is commutative; i.e.,
22 f*g =g*f, V f,g g d

(iv) Whenever/, g g d, we have that f*ge d ,
and in fact
23 \\f*g\\a<\\f\\a-\\g\\a
(v) <5(«) is the unit element of <$; i.e.,
24 5*f = f* 8 =f V/e d
(vi) Cfc has no divisors of the zero element; i.e.,
25 f g * = 0=>/ = 0 or ^ = 0
proof (<outline of)

(i) It is easy to verify that ||
-
|| a is a norm on d. The completeness of
d under this norm is more difficult and is stated here without
proof.
(ii) and (iii) are obvious.
(iv) Suppose /(0 and g ( •) are of the form (9) and (10), respectively.
Then using the rules of convolution defined above,
26 (/**xo =
i
±
=
£= figAt - p -
0 j 0
t 1</>)
S figait -
+ 1=0 t\
f)
) £ g,ut -
+ 1=0 1!*>)
+ f'/M - T)ga(r) dr
The first term on the right-hand side represents the distributional

part of /* g, while the last three terms represent the measurable
part of f*g. To calculate ||/*^|| a we take each of the terms ,
separately. First,
27
i
S
= 0 7=0
SI/,Mftl = (SI/«l)-(S
V =
'*’= 0 / 7 0
k,l)
/
Next
28
f“
Jo
I
I 1
S
=0
fiUt - tl
f)
) dt<±
= 1 0
\ft \
r
J0
| g/f
- /P)l dt
Similarly,
29
f Ii. *- (,§
i*i) •
*
I [j;
u
Finally,
f° (t ~ dx dt
— Jo I
“ T) H I
dr dt
Jo I Jo I Jo
I
= ~ J~\fa(f -T)\-\ga (r)\dt dr

J
(interchanging the order
of integration)
= |£«(t)| fJd - t)| dr

|o |
Putting all these inequalities together proves (iv).

(v) is obvious.
(vi) The proof is beyond the scope of this book. —
Suppose /(•) e Ofc. Then, whenever Re s > 0, the integral
M= /{tie- dt =£ + Us)
converges and is well defined. Therefore, all elements of a are Laplace-
transformable, and the region of convergence of the Laplace transform
includes the closed right half-plane
C+ = {s: Re s > 0}
With this background, we define the set &.
[definition] The symbol & denotes the set of all functions/: C+

— » C that are Laplace transforms of elements of CL
Thus, according to definition (33),/(.) e & is just another way

of saying that the inverse Laplace transform of /(•) belongs to Ct.
When we deal with feedback systems, the symbol d comes in handy to
keep the symbolism from proliferating.
fact Suppose /(s) is a rational function of 5. Then /(•) e A if and only if

(i) / is proper, and
(ii) All poles of / have negative real parts.
proof If (i) and (ii) hold, it is clear that /(•) [the inverse Laplace
transform of/(*)l consists of (i) a measurable function that is bounded by
a decaying exponential, and (ii) possibly an impulse at t = 0. Hence
/(•) e i.e.,/(-) e &. On the other hand, if (i) does not hold, then /(•)
contains higher-order impulses and hence does not belong to a, while if
(ii) does not hold, then the measurable part of
/(•) is not absolutely
integrable.
Having defined a, we can define its extension a e, in exactly the

same way that we define Lpe from Lp .
35 [definition] The set consisting of all generalized functions

/(•) which have the property that all truncations /r (.) of /(•)
belong to G. for all finite T is denoted by (Z e and is called the
extension of a.
The set a, has some very useful properties. Most physical systems,
even “unstable” systems, have impulse responses that belong to
=
a e [e.g., h(t) e1]. Moreover, it can be shown that if /;(•) is any dis-
tribution that vanishes for t < 0, then the operator H defined by
36 (Hf )(t) = f* h(t t)/(t) dx
d0
maps Lpe into Lpe for all p e [1, oo] if and only if /?(-) e GL e 9 In other .
words, the class of linear time-invariant systems that interest us

(namely those that map L pe into itself for all p) is exactly the same as
the class of systems that have impulse responses in a e . Thus systems
whose impulse responses lie in a e are the most general (linear time-
invariant) systems that we need to consider.
We shall now prove the main results of this subsection.
37 Theorem Consider the operator H defined by (36), where h(-) e

d e . Then the following four statements are equivalent:
(i) His L r stable.
(ii) H is L^-stable.
(iii) H is Testable for all p e [1, oo].
(iv) /*(•) e a.
Remarks: Theorem (37) brings out fully the importance of the

set a. According to this theorem, a system with impulse response
(i)Tj -stable, or
K') is (ii) L^-stable, or(iii)Z^-stablefor all p e [1, oo]
if and only if the impulse h{-) belongs to the set a. This justifies the
statement made earlier that a is the set of stable impulse responses
and that & is the set of stable transfer functions.
proof We shall first prove that (iv) implies each of (i), (ii), and (iii).
Accordingly, suppose h(-) e 6L

(iv) (i) If f(-) e L u then /(•) e a, and in fact
38 ll/Olla = ||/(’)lli
Hence, by (iv) of lemma (19 ), h*f e Ct, and

39 P*/l|a <PI|a -|l/IU = ll^ll a ||/lli-
9 The proof of this statement is very similar to that of Theorem (37) and is
therefore left as an exercise.
<
S ec. 6. 5 Open -Loop S tability of Linear Systems 251
Next, because /(•) contains no impulses, neither does h*f, which means
that h *f e L u and
40 P*/IU = P*/lli
41 P*/lli<£PMI/lli
which shows that is L j -stable. H
(iv) => (ii) Suppose that/(-) e L„ and that h( •
) is of the form
42 hit) = i—0
S h,5{t - t t) + ha (t)
Then
43 (h *f)(t) = S hi f(t - t,) + f ha (t - t)/(t) dx

= J 0 J0
44 \(h *f)(t)\ S \h,\'\f(t - f,)| + f \ha (t

— t)|-|/(t)| dx
1 =0 J0
< ess. sup. 1/(01 S= |A/I+ r \h (x)\dx

a
t 1 0 Jo
= ll/(-)IU-P(-)ll«
Because (44) holds for (almost) all ?, we see that h *f e L„ and that
45 P*/IU<Plla-||/ll~
This shows that H is L„ -stable.
(iv) This part of the proof is rather involved and can be skipped
=> (iii)
without loss of continuity. Because we have already established (i) and (ii),
we shall concentrate on Z^-stability for 1 p < oo. Suppose /(•) G L < p .
We firstshow that h*f& L p Accordingly, suppose h(-) is of the form .
(42); then h* /is of the form (43). Now, if /(•) e L p then clearly the ,
function
46 8 i(t) A2 hj(t - t,)

= 1 0
belongs to L p and ,
47 piii P <ii/ip(| pi) o
Thus we concentrate on the integral
48 ha (t - x)f(x) dx A g 2 (t)
jo
Let q = pj(p — 1) so that (l/q) + (1/p) = 1. Then
49 1
82(f) I
<£ I
K(t - t) I
*|/(t) j
dx
=£ I
K(t - x)\'/p\ha (t - xY'f\f(x)\dx
< |J 1
Kit - T)1 rfrj
/9
• ha (t - t) * f(x) p dx
17 '"
0 [Jj I | \
J
by Holder’s inequality (fact [6.1(13)]). Hence

9
50 1*2(0 1* < [J
'
\ha(t - T)| drj' 1

ha(t - T)|-|/(T)P dx
o JJ
< [J“ \h (X)\ dx] a

P/Q
J0
‘
m- T)|.|/(T)P dx
Now, if /(•) e Lp ,
then the function r i—
> |/(t) p \
e L { . Thus, using the
estimate (41), we get
9
51
J~
|* 2 ( 0 I
P dt < [J“ \h (x)\ dx^' a
•
f~
\ha(x)\ dx •
Jo |
f(x)\> dx
= \h a(x)\ dx]” \m\>dx

[J“ J”
because (p/q) + 1 = p. Raising both sides of (51) to the power 1 Ip gives
52 ||*2 II, < J”|«T)|rfT-||/||,

Because h*f=g + g (47) and x 2, (52) show that h*fGL p . Moreover,
53 P*/ll ,<||*ill + |l*2ll,

i i,
< ll/ll J£ PI + f Mr)\dx

= ii/iIp-pii«
This shows that H is L^-stable for all p.
Up to now, we have shown that (iv) implies (i), (ii), and (iii). We shall
now prove the converse, namely, that (i), (ii), or (iii) each implies (iv).
(i) => (iv) Suppose (i) is true. Because H is linear, (i) implies that H
is continuous. Now, because can be approximated every element in &
arbitrarily closely (in the sense of distributions) by an element of L u this
implies that iTmaps (2 into itself. Let f(t) = 8{t) then h * f e (2, because
,
\
iTmaps (2 into itself. However, h * d = h, which shows that h e (L. Thus

(i) implies (iv).
(ii) => (iv) This is a special case of a result contained in Theorem (70).
(iii) => (iv) Suppose (iii) is true, i.e., that H is L p -stable for all
p e [1, oo]. Then, in particular, H is L 1

-stable. As shown above, this
implies (iv).
54 Theorem Consider the operator H defined by (36), where h(-) e (2.
Then H is L 2 -stable, and

55 \\h*f\\ 2 <m.\\f\\ 2
where
56 m= sup h(jco) | |
Remarks: The main purpose of Theorem (54) is to prove the

bound (55). For an arbitrary p e [1, oo], we have the inequality
(53). However, for p — 2, we can establish the tighter bound (55)

(note that m <Pii a ).
proof Let g = h*f Then

57 g(jOi) = h(j(0)f(j03)
Using Parseval’s equality, we get
58 lkl!I = ^ f \i(jco)\ 2 dco=± i

f \h(jco)\ 2 \f(jco)\ 2 dco
•7—00 J — OO
<^J l/(/®)l
2
^=« 2
ll/ll!
Taking the square root of both sides of (58) establishes (55). —

We shall summarize the results of this subsection up to now:
1. We have introduced the sets d and &.
2. We have shown that a system with impulse response /*(•) is
(i) Testable, (ii) Z^-stable, (iii) Z^-stable for all p e [1, oo] if
and only if /?(.) e d.
3. We have established the useful bounds (53) and (55). These
show that, given a system with impulse response h(>) e d 9
thel^-norm of the output h */is no larger than h a times the || ||
Lp-norm of the input /; in the particular case where / e L 2 ,
a still tighter bound can be obtained.
Multi-Input/ Multi- Output Systems
We shall turn now and outputs.

to systems with multiple inputs
The from Theorem (37).
results in this case follow very easily
Consider a system with n inputs and m outputs, with the input/
output relationship
59 (Hf)(0 = f HIt - T)f(T) dr

Jo
where the impulse response matrix H(.) e a™*"; 10 i*e., the entries of
the m X n matrix of distributions H all belong to d e In analogy with .
the scalar (single-input/single-output) case, such types of operators H

are the most general linear time-invariant operators that we need to con-
sider.
The criteria for the Instability of systems of the form (59) are
given next. Note that Q mXn denotes
J the set of all mx n matrices of
distributions in d.
1
°To avoid confusion, we write H for the operator and H(«) for the associated
impulse response.
. ,
60 Theorem Consider an operator H of the form (59), where H(«) e

x
ft™ ”. Under these conditions, the following statements are equivalent
(i) H is L r stable.
(ii) H is Loo-stable.
(iii) H is L^-stable for all p e [1, oo].
(iv) H(.) e a mxn .
The proof is left as an exercise, because it is entirely parallel

to that of Theorem (37).
Basically, Theorem (60) states that an «-input/m-output system
with the impulse response matrix H(.) is (i) L r stable, (ii) Loo-stable,
(iii) L^-stable for all p e [1, oo] if and only if each component of
H(-) belongs to ft.
In proving Theorem (37), we obtained the useful bound (53). A

similar bound can be obtained for multivariable systems also and is
given below without proof. In this context, it is worthwhile to recall the

definition of the norm on Lnp (definition [6.1(27)].
61 Theorem Consider an operator of the form (59), where H(.) e

ft/nx„ xhen, whenever f(.) e L” for some p e [1, oo] we have
62 I
|Hf ||j, < «i ||f IU
where
63 «i = l|Mj ||( 2
||
•
|| i2 indicates the /2 -induced matrix norm on Rmxn and ,
M t
is the
mX n matrix whose ijth entry is ||/z l7 (.)|| a If p = 2, then
.
64 ||Hf || 2 <a 2 ||f || 2
where
65 a2 = ||M 2 ||, 2
and M 2 is the mx n matrix whose zyth entry is given by
66 (M 2 ) i7 = sup |
hij(jco) |,
i=l,..., rn,j = 1, . . n
CO
To explain in words, the bounding constants and a 2 are cal-
culated as follows: Take the m x n matrix H(-) and replace each

element by its norm in ft. The resulting matrix is M 1? and is the
/ 2 -induced 2 M
easy to verify that the bounds (62) and
norm of . It is
(64) reduce to (53) and (55), respectively, in the case of scalar systems.
6.5.2 Time-Varying Systems
In the previous subsection, we studied linear time-invariant opera-

tors of the form (36). The class of operators studied in this subsection
Sec. 6. 5 Open-Loop S tability of Linear Systems 255
represents a natural generalization of those of the form (36). Speci-

fically, we shall consider operators of the form
67 (Gf)(t) =S gi(t)f(t — t,) + [

gjt, t)/(t) dx
Actually, because fit) —0 whenever t < 0, we can rewrite (67) as
68 (Gf)(t)= X)
iei(t)
gi(t)f(t — t,)+ [
Jo
ga(t, x)f (t) dx
where
69 I(t) = {/: t t < /}
In other words, (68) is obtained from (67) by taking the summation only
over those indices i such that t t t. <
Theorem (70) gives necessary and sufficient conditions for an
operator G of the form (68) to be Z^-stable.
70 Theorem Consider an operator G of the form (67), where
71 t
E
iei(t)
1
g,{t) |
€ Lme
72 x i->
ga(t, x) e Lu Vt>0
73 t t-y f | ga (t, r) dx |
G L me
Jo
Then G maps L me into itself. Further, G is Testable if and only if
74 sup
t
I 2
\iei(t)
| gi (t) |
+ Jof |
g„(t, x) dx)
|
A c„ < oo
)
proof We leave it as a problem to show that if (71)— (73) hold, then G

maps Looe into itself.
To show that (74) is a necessary and sufficient condition forToo-stability,
we shall first tackle the sufficiency part, which is simpler to prove.
-
(i) “T/ ” Suppose that (74) holds and that /(•) e Then
75 1(00(01 = I S - ti) + [gait, T)/(T) dx

h'€J(r) J0
<{ s/(0 \m\+

b*e
[\ga(t,T)\dx) -11/(011-
Jo j
<^||/(-)IU
Because the right-hand side of (75) is independent of t, we see that
(00(0 G Too and that
76 11(00(011°° < Coo ||/(0IU
Hence G is Too-stable.
(ii) “Ow/y If' We shall show the contrapositive, namely, that if (74)
does not hold, then (GO ||oo/||/||oo can be made arbitrarily large by
||
suitably selecting /(0- Accordingly, suppose the function
77 r{t)A Z
iei(t)
k*(0l+ [
Jo
\ga (t, T)\dx
isunbounded. Let k be an arbitrary constant; we show that a func-

tion /*(•) e Loo can be found such that !/*(•) IU = 1 and G/* ||oo | ||
> k. The first step is to pick some € > 0 and to choose 9 k such that
r(9 k ) >
k + e. This is always possible because r(-) is unbounded.
Next, pick intervals S of the form t
78 = Wk — ti —G i, 6k — ti + Gil, i E I(6 k)
such that
79 2
i&1 ^) JrGSt
f \ga{9 k , T)|rfT<4-
1
Clearly, this can always be accomplished by choosing the numbers

G t
sufficiently small. Finally, let /*(•) be the function defined by
80 m =
(
l
sign gi{9k),
sign ga (6 k ,t),
if t
if t
e
i
<5,
U
teHM
$t
Then /*(•) e Loo and ||A(")oo|| = 1. Moreover, we have

9,
81 (Gfk)(9 k ) = S gi (0 k )fk (6 k - u) + [ gJ0k r)A(r) dr
,
iEl(dk) Jo
2 gi(Qk)fk{0k - ti) + 2 f ga(.9 k , t)A(t) dr

i&i (8k)
i£l(0 k) ieitfk) J T edi
+I ga{0k, T)A(T) dr
TG[O,0 t]
xt U Si
iei(d k )
Thus, rearranging, we get

0k
82 0Gfk )(9k)> 2
*£/<»*)
\
gi (6k)\+
Jo
[ \ga (e k ,r)\dr
- 2 f \ga(0 k ,r)\dr
iei(O k) J T edi
- 2 \ga(0k, t)V\fk (r)\dr

f
ieiiflk) Jxedi
However, because | fk (r) |

< 1 for all t, the last inequality can be
simplified to
83 (Gfk)(6 k ) > ( 2
(iene k )
gi(9 k )+ C\ga(9k,r)\dr]
Jo )
-(2 2 f \ga(9 k ,r)\dr]

ienOt)( J t£ «, J
k -\- 6 — € =k
because the first term in the braces is not less than k + € by the way
we chose 9 k ,
while the second term in the braces is not more than €
by (79). Finally,
84 \\Gfk\\ = ess. sup. [(<%)« I

> \(Gfk)(0k > k )\
B
Because this process can be repeated for any arbitrary k we see ,
that G
not Loo-stable. This shows that (74)
is is a necessary condi-
tion for G to be Loo-stable. —
The next theorem gives necessary and sufficient conditions for
G to be Lj-stable.
85 Theorem Consider an operator of the form (68), where
86 2
iei(t)
| g (t
t + t *) |
G L oe 0
87 t i-> ga (t, t) G Lu Vt>0

88 T h* ga (t, T) dt E
J | |
Under these conditions, G maps L le into itself. Further, G is Lr

stable if and only if
89 sup
r
{£
U=0
\gfy + 01 + f
Jt
I gaif, t)| dt) Acôo
)
proof We leave it as an exercise to show that if (86)-(88) hold, then G

maps L le into itself.
We shall first prove the sufficiency of (89) for L r-stability.

(i) “If” Suppose that (89) holds and that /(•) G L Then x .
90 rmm
Jo
dt < f s
Jo fe/(r)
\ g,(t)\-\f{t-td\dt
+ f Jo
Jo
i)\-\m\dTdt
=s
= i 0
r \g (f)v\ Kt -
Jo
t t t)\dt
+ \~\ga(t,T)l\m\dtdT
<£ f"kl(T + OM/(T)l*

=0 Jo f
+ J
\ 8a(J ’ dt dZ
Jo
< sup r £
=
i
gi(x + /,)i
T Ll 0
+ {Jj/(r)|rfr}
= Cl||/(0lll
This shows that G is Li -stable.
(ii) “ Only If” Suppose G is Testable. Because G is linear, this implies
that G is continuous on L and
x hence that G maps Ct into Ct. Let
f(t ) = 5{t — t), where t > 0 is a given number. Then
91 ( Gf)(t )
= 2 Si(f)5{t —T— tt) + ga (t, t)
iei(t)
= 2 £i(T + t t )S(t - X - ti ) + £»(>, t)

f=0
By assumption, (<?/)(•) e Ct; i.e.,
92 S ki(T + *i)l + f l&>(*» 1

dt < co
t = 0 JT
where we use the fact that ga (t, t) 0 for t T. Now, the In- = <
stability of G
not only implies that (92) holds for each fixed T but
also that ||G/(-)||a is bounded independently of r; i.e.,
93 sup ( S |^(T + G)l + f l&iO, 01 dfl < oo

T [i= o JT J
which establishes (89). H

Finally, we shall show that if G of (68) is both L r stable and
Loo-stable, then it is Z^-stable for all p g [1, oo].
94 Theorem Suppose an operator G of the form (68) satisfies both (74)

and (89). Then G is Z^-stable for all p <= [ 1, oo]. Further, whenever
/(•) g Lp we have ,
95 \\Gf\\ p ll/ll,
where <7 = /?/(/? — 1) is the conjugate index of p.
proof We shall give the proof only in the case where g t = 0 for all i.
The proof in the general case is similar.

If we assume that £7 = 0 for all /, then (74) and (89) reduce, respectively,
to
96 sup f I ga {t, T)| dT A Coo < 00

t Jo
97 sup f I ga {t, T)| dt A Cl <00

T JT
Suppose /(•) g L p We . can assume that 1 < < 00, because (95) is
clearly true if p = 1 or 00. Then
98 |(G/)(0|< [ \gait,z)V\m\dz
Jo
= fkaaT)|^ka (/,T)|^|/(T)UT
Jo
1/9 1/P
< {£
I
gait, T) |
</rj
{ £ |
r) |
• |/(t) 1?
by Holder’s inequality. Next, we have
99 I
(G/)(0 1*
< git, X) rf*} | |
git, T) |
• [/(T) |* dx
{ Jj { J“ j
fV.a,T)M/(T)|'rfT
Jo
100
f"
Jo
mm* * < r Jo \git, z)\-\m\
Jo
p dx dt
= p
c~ q
\gln)V\m\ p dt dx
= <£• I
git, t) dt~\ | \mv dx
J“ [j;
< dH*c x ll/ll?
Raising both sides of (100) to the power 1/p gives (95).
We shall summarize the results of this subsection up to now:

1. We have given a set of necessary and sufficient conditions for
an operator G of the form (68) to be L r stable and to be In-

stable.
2. We have shown that if G is L r stable and Zn-stable, then it is
Z^-stable for all p e [1, oo].
3. We
have proved the useful inequality (95), which gives an
upper bound for the Z^-norm of ((!/)(•)•
We shall conclude this subsection by showing that if G is a time-
invariant operator, then Theorems (70), (85), and (94) together reduce to
Theorem (37). Accordingly, suppose
101 g (t )
t
=h t
(a constant), V t >0
102 ga (t, T) = ha(t — t), V t, T > 0
so that G corresponds to a system with impulse response
103 h(t) — h 5{t

t
— £/) T" h a (t)
i =0
Then the condition (74) for L„,-s lability becomes
104 sup
t
(
Uei(t)
2 \K\ + Jofl h (t — xa ) |
dx)
)
< oo
However, as t —+ oo, I(t ) eventually includes all z>0. Thus (104)

is equivalent to requiring h(-) to belong to Q. Similarly, the condition
(89) for -stability becomes
105 sup ff; \h,\ + dx[ \h a (t - x)\dx] < oo

T V 0 l ~~ J
which also is equivalent to requiring /*(•) to belong toft. Finally, because
we have
106 cx = = P(-)IU
in this case, (95) reduces to (53).
Problem 6.6. Suppose /(•) e C£. Show that whenever s e C+ we have
\f(s)\<\\f(-)\\a
Problem 6.7. Show that iff(-) e then the function 5 e~ sTf(s) also
belongs to & for all positive T. [Hint: consider the inverse Laplace transform
of e~ M]
sT
Problem 6.8. Prove Theorem (60).

to a and/or to a e. If /(•) e G, find ||/(-)lla-
(i) m=£ i =0
S(t — iT), T> 0
(ii) f(t) = S(t — T) + e~ at

sin cot, T, a, co > 0.
(iii) fit) = £ e-**

to &.
= e~ sT s 2 + 5^ + 5
2
(0 f(s) s + s + 10
(ii) f(s) = 1/cosh V ^ (Hint: use partial fractions.)
Problem 6.11. Determine whether each of the operators below is In-

stable and Loo-stable, using Theorems (70) and (85).
(i) (Hu)(t) = u(t - S) + £ sin t e-^'-ûix) dr, 5 > 0
(ii) (Hu)(t) = I' e (_ +2,) «(t) dz

'
6.6
LINEAR TIME-INVARIANT FEEDBACK SYSTEMS
we study the conditions under which a feedback inter-
In this section,
connection of linear time-invariant subsystems results in a stable sys-
tem. Since the results for linear time-invariant systems are the ones
most often used in practice, and since many of the stability results for
nonlinear and/or time-varying systems use the results for linear time-
invariant systems as a point of departure, it is important to have a good
understanding of what makes a linear time-invariant feedback system
stable or unstable.
)
S ec. 6.6 L inear Time - In varian t Feedback S ystems 261
This section is divided into two subsections. In Sec. 6.6.1, we

deal exclusively with single-input single-output systems. The emphasis
here is on the properties of the sets d and <$, and how they can be ex-
ploited to obtain necessary and sufficient conditions for feedback stabil-
ity. In Sec. 6.6.2, we tackle multi-input/multi-output systems, where
some additional difficulties arise from dealing with matrix transfer
functions.
6.6.1 Single-Input/Single-Output Systems
In this subsection, we study systems of the form shown in Fig.

6.6. Initially, we assume thatgu g2 e i.e., that the system is “open-
loop stable.” We then consider the case where g 2 is a constant, and

contains a finite number of right half-plane poles, but is otherwise
stable. In both cases, we present necessary and sufficient conditions for
stability that involve the behavior of g (s) and g 2 ( s ) as s varies over
t
the closed right half-plane. Next, we present a graphical procedure,

whereby the stability conditions can be tested by examining only g^jco)
and g 2 (jco). This graphical procedure is a natural generalization of the
standard Nyquist criterion. As an application, we show how the results
presented here can be used to study the stability of networks containing
uniform or nonuniform RC transmission lines and operational ampli-
fiers.
The following result is the key tool used to derive all the stability
criteria of this section.
1 lemma ( Paley-Wiener Suppose p(-) e (J. Then the function

#(•) = 1 //?(•) belongs to & if and only if
2 inf |/?0)| > 0
sGC+
where C+ is the closed right half-plane, defined by
3 C+ = fa: Res > 0}

proof
(i) “ Only If” Suppose q = 1/p e &, and let q(-) denote the inverse
Laplace transform of q(-). Then by Problem 6.6, we have
4 |9(«)| ^ ||?(-)lla, V s e C+
Therefore
5 l^y^lkCOIIa, V s S C+
But this shows that
6 lfwls >0 V s e C.
iT«M -
which establishes (2).

l
(ii) “T/ ” The proof of the “if” part is far beyond the scope of this book
and can be found in [14] Hille and Phillips [p. 150]. M
Lemma (1) states the following: Suppose p(-) is the Laplace
transform of an element of ft. Then its reciprocal 1 /p(*) is also the
Laplace transform of an element of ( if and only if /?(•) is bounded
away from zero over the right half-plane. In particular, this means that
(1) p(*) has no zeroes in C + and (2) no sequence ,

in C+ with
| |
—> oo can be found such that p(s ) t
— > 0 as / —> oo.
Even though the “if” part of lemma

difficult to prove (1) is very
in the general case where p is an arbitrary element of (I, it is very easy
to prove in the special case where p(s) is a rational function of s.
fact Suppose p(s) is a rational function of ^ and belongs to Ct [i.e., p(s) is
proper and has no poles in C+ Then]. l/p(s) belongs to (I if and only if (1)
p has no zeroes in C+, and (2) p( oo) ^0 [i.e., if (2) holds].
The proof is
left as an exercise.
Notice that in proving the “only if” part of lemma (1) we never
made use of the fact that p e &. In view of this, we can state the fol-
lowing.
lemma Suppose p(*) is a distribution, and let p(*) denote its Laplace
transform. Under these conditions, if !/]?(•) belongs to then (2) holds.
Lemma (8) states that if p is any Laplace transform, then (2) is a

necessary condition for 1/p to belong to A (although it may not be suf-
ficient).
Consider now the feedback system shown in Fig. 6.6. By definition

[6.3(34)], this system is Z^-stable if yu y2 e Lp whenever u u u 2 e Lp .
Now, note that
Us) = ,
i+ g,(s)g
,
444r ,-M - +f^7vw
2 (s)
ii
î(«)f
) 1
i 2 (-y)
yi(s)
iMUs) Mi (s) - Ms) u 2 (s)
i + ii(s)g 2 (s) 1 + £i(^)£zO)
FIG. 6.6
2
Sec. 6.6 Linear Time-Invariant Feedback Systems 263
We can express these equations in matrix form as

.piOr fillis) finis) u t
(s)
-his). Jnis) finis). ACO.

where the definition of hij(s) is obvious. Now, from Theorem [6.5(60)],
we have the following result.
12 fact The system described by (9)-(10) is Loo-stable 11 if and only if
hij(-) E (3 for /, j = 1, 2.
,
Thus the objective of this section is to derive necessary and suffi-
cient conditions for /z

/y ( -
) to belong to &, for i,j= 1, 2. We shall tackle
first the so-called open-loop stable case, i.e., where g 1? g 2 belong to
because this is the easiest case.
13 lemma Suppose g l9 g 2 e <3. Then £,/•) E (3 for i, j = 1,2 if and

only if
1
14 K-)A 1+ #l(0#2(0
e <3
proof We shall do the proof in some detail, because the manipulations

carried out below are repeated, in essentially the same form, in several
later proofs.
(i) Suppose (14) is true. Clearly the function
15 A*) EE 1
belongs to Ofc, because it is the Laplace transform of the unit impulse.

Therefore, if r E fi, then 1 — r e because is a linear space.
However,
16 1 - r(s) = ii(s)g 2 (s)

1 + &i(s)gi(s)
This shows that hi 2 h 21 e ,

(3. Next, we have
17 hn(s) = gi(s)r(s)
Because g 1 and r both belong to (3, so does their product. (See
lemma [6.5(19)].) Hence /z
nG <3. Finally, h 2 because
18 h 22 (s) = g 2 (s)r(s)
(ii) “Only If ” Suppose £,/•) e (3 for i, j = 1,2. Then, in particular,
h 21 (-) e <3. This means that 1 h 21 — e (&. However, 1 — h 2l =
r. This shows that r E (3. _
The utility of lemma (13) lies in that it leads to a simple necessary
and sufficient condition for the stability of the system (9) -(10).
11 Or, equivalently, (1) L\ -stable or (2) L^-stable for all p e [1, 00].
0
19 Theorem Suppose g u g 2 e A. Then the system (9)— (10) is Te-

stable 12 if and only if
20 inf
SEC+
1 1 + gi(s)g 2 (s) |
>0
proof By lemma (13), the system (9)-(10) is Too-stable if and only if
r e tt. Now, because g g2 e u, we have t , that 1 + g g2 e

x Hence, by
lemma (1), r = 1/(1 + gig2 ) e (I if and only if (20) holds. ^
In engineering terms, the quantity 1 + £i(s)g 2 (s) is usually re-
ferred to as the return difference.Thus Theorem (19) states that the sys-
tem described by (9)-(10) is Z^-stable if and only if the magnitude of the
return difference is bounded away from zero as s varies in the closed
right half-plane.
In Theorem (19), we assumed that both and g 2 e &. We next
consider the situation where g 2 is a nonzero constant, and contains
a finite number of right half-plane poles, but is otherwise stable.
21 lemma Suppose
22 g 2 (s) = g 20 =£ 0 V s
A A
Then e A for i,j = 1,2 if and only if red, where r is defined in (14).
Remarks: Note that there are no assumptions about g x

.
proof
(i) “If” Suppose fed. Then,
- gl(s)g2
23 1 r(s) =
[1 + #l(^20]
also belongs to &. This shows that /* 12 , ^ 2 i e S. Also, since g 2
is a nonzero constant, we have that
^n0) = [1 “ r(s)]/g 20
belongs to &. Finally,
24 ^ 22 ^) = gioffs)
belongs to d.
• (ii) “ Only If Exactly

’
as in the proof of lemma (13).
25 Theorem Suppose g 2 is of the form (22), and suppose

26 g^s) = g u (s) + n^/d^s)
where g lb e d, h 1
and d t are polynomials in s with no common
zeroes, the degree of h is less than or equal to that of d u and all
1
12 Or equivalently, (i) Ti-stable, (ii) L p -stable for all p e [1, oo].

zeroes of d x
are in C+. Under these conditions, the system (9)— (10)
12
is Loo-stable if and only if
inf |1
sec+
+g 20 liO)| > 0
Remarks:
(i) The condition (27) has exactly the same interpretation as
(20); namely, for the class of systems under study, a necessary
and sufficient condition for stability is that the magnitude of
the return difference is bounded away from zero as s varies
over the closed right half-plane.
(ii) Note that there is noassuming that all
loss of generality in
zeroes of d 1
are in C+
g has some poles in the open
,
since if t
left half-plane, the contribution of these poles can be removed
by partial fraction expansion, and lumped in with g lb .
proof By lemma (21), we only need to show that (27) is a necessary

and sufficient condition for r to belong to CL
(i)
“
Only If” Since r = 1/(1 +g 2 o£i), the necessity of (27) follows
from lemma (8).
“
(ii) If ” Let a denote the degree of the polynomial d u and define
28 p(s)
_
= «i0)
(s + 1)«
29 ? 0) = dM
(s + 1)«
Since the degree of n 1 is less than or equal to a, it is clear that

p,q e d. Also, it is easy to verify that
= + gf “
(J)
30 Us) - #,. W + 1
f«
,,, q(s)
31 ns ~
{?('’) + gio[p(s) + ?0)£i&0)]}
We now claim that, if (27) is true, then
32 inf \q(s)
sec+
+g 2 o[p(s) + q(x)gib(s)] > \
0
To show this, rewrite (32) equivalently as
33 inf
SEC+
|^»[1 + g 2 o#i(s)]| > 0
Now, suppose (33) is false. Then there is a sequence in C+ such that

^fc)[l + iôg'ite)] —>0 as i — > oo. There are two possibilities to con-
sider: (i) (s t )“ is bounded, and (ii) is unbounded. Each of these pos-
sibilities is studied separately.
bounded, it has a convergent subsequence. Renumber it as
If (s*)™ is
(-Vi)T, and suppose that it converges to s 0 e C+. Since q + gioip

+ qgib) e S, it is continuous over C + which means that ,
34 q(s o) + g 2 o[p(s 0 + 9(s 0 )i lb (s 0

) )]
= 0
>
However, (34) leads to a contradiction, as shown next. Either is a pole

of g u or it is not. If s 0 is a pole of g u then it is a zero of d u i.e., of q. Hence,
at s =s 0, we have
35 q(s 0 ) + g 2 o[p(.s 0 ) + ‘K.SoîiCs'o)] = g 2 oP(s 0 ) ^ 0
because n 1 and d x have no common zeros; thus (34) is contradicted. If s 0
is not a pole of then ^(.So) is a nonzero
gu finite number, and so is
1 + g 2 ogiCso) [because of (27)]. Hence

36 q(so) + g 2 o[pC?o) + #Cso)£i&Cso)] = <Ks 0 )[l + g 2 ogiOo)] ^ 0
Thus (32) is once again contradicted.
Now consider the situation where (j/)f is an unbounded sequence. Then
we can pick a subsequence, which is again renumbered as (j/)~, such that
| |
— » 00 as i — 00 . Now, since q is a rational function whose numerator
and denominator are of equal degree, it is easy to see that q(s t) approaches
a nonzero constant as |^| — * 00 . Also, by (27), [1 + g 2 ogi(*y,)] remains
bounded away from zero, which contradicts the assumption that q(s ) t
[1 + ^ 2 o£ife)] > 0 as i > 00 — — .
Hence we conclude that no such sequence fe)? exists, i.e., that (32) is
true. Next, by lemma (1) and the fact that q + g 2 o(p + qgn) e d, we
have that 1 /[q + g 20 (p + qgn)] e d, whence red, by (31). By lemma
(21), this shows that the system (9)-(10) is Loo-stable. M
Consider next the case where g x and g 2 are both of the form (26).
In this case, the necessary and sufficient conditions for stability are
slightly more complicated.
37 Theorem Suppose g i9 i = 1, 2 is of the form

38 g;(s) =g ib
(s) + n^/d^s)
where g ib e d, h and d are polynomials in s with no common zeros,
t t
the degree of hi is less than or equal to that of d and all zeros of d t, t
are in C + Under. these conditions, the system (9)-(10) is L^-stable 13

if and only if
39 (i) inf 1 1 +g t
(s) g 2 (s) |
> 0,
s GC+ ^
40 (ii) d (s) g (s)^0 whenever d 2 (s)
t x
= 0,
41 (iii) d2 (s) g 2 (s) ^ 0 whenever d (s) x
= 0.
Remarks: Theorem (37) states that three conditions together are
necessary and sufficient for stability:
(i) the return difference condition (39), which is the same as
before;
(ii) the quantity d g does not vanish at the poles of g 2
1 1
in C+ ;
and
(iii) the symmetrical condition about d2 g 2 and g l9
13 Or equivalently, (i) L\ -stable, (ii) L^-stable for all p e [1 ,

00].
v
Conditions (ii) and (iii) assure the complete shifting of the C+ poles
of ii and g 2 and can be rationalized as follows: suppose s 0 e C+ is
,
a pole of g,, and that d 2 (s 0 ) g 2 (s 0 ) =

0; then in effect the feedback
loop is open at the frequency s =
s 0 so that there is nothing to ,
stabilize the unstable subsystem £ As a result, the closed-loop trans-

fer function contains a pole at s 0 . Similar reasoning applies to the
C+ poles of g 2 .
The proof of Theorem (37) is omitted in the interests of brevity,

and can be found in [10] Desoer and Vidyasagar, together with some
generalizations of Theorem (37).
6.6.2 Graphical Stability Criterion
In applying Theorems (19) or (25), we have to verify a condition

of the form (20), which involves the behavior of the return difference
r(s) as s varies over C+. Similarly, in applying Theorem (37), conditions
(40) and can be checked easily, and only (39) is involved. The rest
(41)
of this subsection is devoted to the presentation of a graphical stability
criterion for verifying a condition of the form (20) [or (27) or (39)], by
examining only 1 +
gfjto) § 2 (jco) as co varies over R. The criterion here
is a natural generalization of the well known Nyquist criterion, and has
the same advantages associated with the Nyquist criterion, namely (i) :
the test can be applied directly to the experimentally measured data,

i.e., 1 + gi(ja>)g 2 (j(o), and (ii) even if the stability criterion is not satis-
fied, it is easy to visualize the form of “compensation” needed to satisfy

the criterion, i.e., to stabilize the system.
To present the graphical test for stability, we assume that the
return difference r(-) is of the form
42 r(s ) = + giCÔO = + S
1
=
rfi~
sU
+ rfs) 1
/ 0
"
+ SE
^Pi(s- Piy
where tt are nonnegative constants, p e C+ is
t ;
the multiplicity of
the pole pp, and n p is the total number of poles of f in C+ ;
in addition,
43 Sk/I<°°
/=o
44 r a { •) E Li
Basically, in (42), we have splitup r into a part that belongs to &
(namely the first three terms on the right side), plus the contribution of
the right half-plane poles of r expressed in partial fraction form.
,
Now,
if Theorem (19) or Theorem (25) is being applied, then the return dif-
ference r is of the form (42). However, if Theorem (37) is being applied,

then r may not be of the form (42), since it may contain terms of the
type
45 f(s)
— pf
(s
where /e d and p e C+. Thus the graphical stability criterion that we

are about to present does not apply to all the systems covered by
Theorem (37), but it does apply to all the systems covered by Theorems
(19)and (25).
Thus the problem at hand is as follows Given r of the form (42), :
determine conditions on r(jco), co e R, which insure that
46 inf
sec+
|
r(s ) |
>0
In general, this problem is difficult to solve. However, if the “delays”
t, in (42) are rationally related, i.e., if there exist a number T and inte-
gers such that
47 t, = n[T , V i
then a relatively simple criterion can be presented.

Before presenting the testing procedure formally, we introduce
two preliminary notions: (i) the indented jco- axis, and (ii) the phase of
r(jco). Suppose we study r(jco), where r is of the form (42). The first
problem is that if Re = 0 for some i, then r(jco) becomes unbounded
/>,.
as jco —* p,. To circumvent this problem, we indent the yea-axis as shown

in Fig. 6.7, by going around the pole, p„ via a semicircle of radius f,
centered at p Let
t
. <5, denote the half-disk shaded in Fig. 6.7; i.e., let
48 <5 ( = {j:| 5— p,|<e,}
Then, since r(s) is unbounded at p i9 clearly we can choose e t sufficiently

small that 5 contains no poles of r other than p t ; and such that
t
49 inf
s G=5i
|
r(s) |
>0
If we repeat thisprocedure at each j co-axis pole of P, we get a finite num-
ber of half-disk regions where (49) holds. Let us now delete these half-
]
disks from C+ ,
and denote the resulting indented right half-plane as Ci+
(see Fig. 6.8). Now, because of (49), we have
50 inf
sECi+
|
r(s) |
> 0
if and only if (46) holds.

After we indent the jco- axis, it is clear that, corresponding to each
co e R, there exists exactly one point on the indented yon-axis whose
imaginary part is co, but whose real part may or may not be zero. By
a slight abuse of notation, let r(jco) denote the value of r at the point on
the indented j co-axis whose imaginary part is co (this too is illustrated in
Fig. 6.8).
Once the yen-axis is indented, f {jco) is continuous in co, and we can
define the argument or phase of r(ycn).
51 [ definition The argument of r(jco), denoted by 0(y on), is the com-

plex number such that
52 r(jco)= |
r(ja>) |
exp [j<t>(j(o)]
53 0(0) = 0 if r(0) > 0, 7i if f (0)<0

In applying (53), note that if r(0) = 0, then (50) is immediately
violated, and there is no need to test any further.
We now quote, without proof, the desired result.
FIG. 6.8
54 Theorem {Graphical Stability Test) Given r(-) of the form (42), we

have that (46) is true if and only if
55 (i) inf \r(jco)\

COER
> 0
56 (ii) lim cj)(j2nn/T ) — <j)(—j27in/T) = 2nv p where v p

,
is the num-
n~* °°
her of poles of f with positive real part.
Remarks: Theorem (54) is a natural generalization of the well-

known Nyquist criterion: (55) states that the polar (or Nyquist) plot
of r(jco) is bounded away from the origin, while (56) states that the
polar plot of f(jco) encircles the origin, in the counterclockwise sense,
exactly as many times as there are poles of r with positive real part.
The proof of Theorem (54), as well as of its generalization to the

case where the assumption (47) is dropped, can be found in [10] Desoer
and Vidyasagar.
One of the common applications of Theorem (54) is in finding the
range of values of the constant k for which the feedback system shown
in Fig. 6.9 is stable. Suppose g(-) is of the form (26); then by lemma (21),
we know that the closed-loop system is -stable if and only if
57 inf
sEC+
1 1 + & g(y) > |
0
By applying Theorem (54) to this system, we get a graphical stability

criterion that is well suited for applications.
58 Theorem Suppose g is of the form (26), and let k ^ 0. Then the

14
system of Fig. 6.9 is Testable if and only if
(i) The polar plot of g(jco), co e R is bounded away from the

point — 1 /k + jO
(ii) The polar plot of g(jco) encircles the point — 1 /k + jO exactly
v p times in the counterclockwise sense as co increases from
— oo to oo, where v is the number of poles of g with positive
p
real part.
FIG. 6.9
14 Or equivalently, L\ -stable,
(i) (ii) Z^-stable for all p e [1,°°].
Sec. 6.6 Linear Time-Invariant Feedback Systems
271
59 Example. Consider a system of the form shown in Fig. 6.9, with
= e - ^ + 4—* + 2
2
60 $,(*) 2
s 1
In this case, represents a system with a rational transfer function,

followed by a transport delay. It is easy to verify that and hence A
gj
that vp = 1. The Nyquist plot of gi(ym) is shown in Fig. 6.10. Note that,
as co increases, the Nyquist plot approaches the unit circle,
because
IjO'co) approaches a periodic function. Applying Theorem
(58), we see
that this system is Z^-stable whenever 0.5 < k < 0.586.
61 Example. To illustrate the power of Theorem (58), we shall

consider the stability of a nonuniform RC
transmission line with an
operational amplifier in the feedback, as shown in Fig. 6.11. This system
is of the type shown in Fig. 6.9, where
g is given by t
62 Si(s)
A{s)
where A(-) is one of the so-called chain parameters of the transmission

line. It is known that, for a general nonuniform line, A(s) is of the form
63 A(s) =n
i=
0 + Pi)
1
where is asymptotically of the form
64 pi ~ a/
2
with a being a physical constant depending on the parameters of the

line.[For an infinitely long uniform line, A{s) cosh */Xs, where 2 is =
a physical constant.] It is also known that
d, and, in fact, e Ll e .
Hence applying Theorem (58) to the present case, we see that the sys-
tem under study is Testable if and only if the Nyquistplot of g,(yco)
neither intersects nor encircles the critical point — \\k + /().
6.6.3 Multi-Input/Multi-Output Systems
We shall now turn to the stability of multi-input/multi-output

(m.i.m.o.) feedback: systems. It turns out that, in the case of open-loop
stable systems, the fact that the system is m.i.m.o. as opposed to s.i.s.o.
causes no real difficulties. However, in the case where some of the

subsystem transfer functions contain singularities in C+, the stability
conditions are substantially more complex than in the s.i.s.o. case.
Specifically, in the s.i.s.o. case,conditions (40) and (41) assure the
complete shifting of the C+ poles of the subsystem transfer functions
giand g 2 However, the analogous conditions in the m.i.m.o. case are
.
more complex and involve what are known as coprime factorizations.

We shall not develop the theory of coprime factorizations, because this
would take us too far afield. Rather, we shall state the basic results for
the open-loop stable case, and refer the reader to [10] Desoer and
Vidyasagar for further details.
We shall begin with a result that extends lemma (1) to the matrix
case.
_1
65 lemma Let F(*) e & nxn Then
. [F(-)] e & nXn if and only if
66 inf |
det F (s) > |
0
iGC +
Remarks: Lemma (65) is very valuable because it allows us to

determine whether or not the matrix-valued function [F(*)] _1 belongs
nXn
to & by examining the scalar function det F(-).
proof
(i) “/jf” We can write
1
67 [F (s)r l = Adj F (s)
det FCy)
where Adj F(V) denotes the adjoint matrix of F (s), i.e., the matrix
consisting of the cofactors of F(V). Because every cofactor of F(s )
consists of sums and products of the elements of F(s) and because
every element of F(-) belongs to <3, it follows that Adj F(-) e &
nXn .
Similarly, det F(-) e (3. Now, if (66) holds, then by lemma (1),
-1 nXn and
1/det F(*) e (3. Hence [F(-)] is the product of matrix in &
a scalar in (3 and therefore belongs to (3” x ".

68 (ii) “Only If” Suppose [F(*)] _1 e & nXn Then, by the same reasoning as .
above, det[F(*)] _1 e <3. However, det[F(-)] _1 = 1/det [F(-)L and

this belongs to <3. Hence, by lemma (8), we have that (66) holds.
—
Lemma (8) can be similarly extended to the multivariable case,
as follows.
lemma Suppose that F(-) is a matrix of distributions and that

_1 (3” x ”.
[F(*)] e Then (66) holds.
The proof is left as an exercise.

Throughout the rest of this subsection, we shall study feedback
systems described by
69 = u, - y 2
70 e2 =u +f 2 1
71 Yi —G e 1 1
72 y2 =Ge 2 2
y ,(s), y 2 (s) E R and G^s), G 2 (s) e R

where & (s) 9 e 2 (s), u n nxn
t 1 ( y),u 2 (^),
1
.
The system (69)-(72) is the same as that in Fig. 6.5, except that gj and
g 2 are replaced by the matrix transfer functions Gj and G 2 As in the .
s.i.s.o. case, (69)-(72) can be rewritten as

73 e =u - Hy
74 II <o a>
where
75 e = [«', ei]'
76 a = [a; KY
77 y = [y'i y'lY
"
78 G(s) = "O.W 0
0 G 2 (s)_
79 H.r°_ ‘i
-I 0
80 Theorem The system (73)— (74) is Testable 13 if and only if one of

the following equivalent conditions holds: (1)
81 [I + h6(-)] _1 e
or (2)
82 G(-)[I + HG(.)]
_1
e & xn
proof From (73)— (74), we can solve for e and y to get
83 e = (I + HG) _1 u
84 y = G(I + HG) _1 u
The conditions (81) and (82) now follow readily. M
Thus our objective is to find conditions on the subsystem transfer
functions G and G, which ensure that (81) or (82) holds. In the case
,
where both 6] and G 2 belong to fi" ” (i.e., the open-loop stable case),
x
this is very easily done.
85 Theorem Suppose G 2 e & nr ". Then the system (69)-(72) is
L„-stable 16 if and only if
86 inf
s EC+
|
det[I + h6(s)] > 0 |
proof Since G G 2 e & nXn G

1} ,
e (£
2nx2n
,
whence I + HG e & 2nx2n
Hence (81) holds if and only if (86) holds, by lemma (65). H
Remarks: Theorem (85) states that, in the open-loop stable case,
the feedback system (69)-(72) is Z^-stable if and only if the deter-
minant of the return difference matrix I + HG is bounded away from

zero over C + Note that .
r j g n
= det (I + G
^ ^
+ HG) =
2
87 det (I det a 2 6i)
G i
1 _
so that the stability criterion (86) can be expressed in terms of the

subsystem transfer matrices and 2 Also, once the quantity G .
det (I + H6) is computed, the condition (86) can be verified graph-

ically by using Theorem (54).
In the case where Gj and G 2 have singularities in C+ ,

the situation
is substantially more complicated in the m.i.m.o. case than it is in the
s.i.s.o. case, because of the complexity of poleshifting in the m.i.m.o.

case. To illustrate this point, consider the system of Fig. 6.6, and let
15 Or, equivalently,
(1) Li-stable, (2) Z^-stable for p e [1, oo].
16 Or, equivalently, (1) Li-stable, (2) L^-stable for all p e [1, oo].
S ec. 6. 6 Linear Time-In variant Feedback Systems 275
u2 — Si —
ku a nonzero constant. Then the closed-loop transfer
function from u x to is
88 h ll() M- 1
liw
+ kg^s)
Now, suppose s 0 e C+ is a pole of |i Because we can write
89 * ll(j)
k + 1/^(5)
it is clear that h lx is analytic at s 0 and that R u (s 0 ) = 1/k. Thus, if we
place a nonzero feedback around all poles of £, in C+ are shifted
from their original locations. Now, if we can assure that the new pole
locations are not in C+, the system is stable (loosely speaking). This is
the purpose of condition (39). Conditions (40) and (41) ensure that the
effective feedback at the pole locations of g, and
g2 is nonzero, so that
pole shifting does take place.
In contrast, in the multivariable case, it is quite possible for both
the open-loop and closed-loop transfer functions to have poles at the
same point in C. For example, let
"
1 /(s - 0
'
'2 2
"
Gifc) =
1)
; ^ 2 (s) =
0 Vis + 1)J 2 0
and suppose u 2 =
0. Then G 2 represents a constant nonsingular feed-
back matrix, and the transfer function matrix from u, to y is ,
s + 1 -2
90 Hn (5) = 6,(1 + 626,)- (s - Vis + 3) - Vis + 3)
-2 s + 1
L(j-lX» + 3) (s-1Xj + 3)J

Thus both G t
and Hn have poles at s = 1, even though the feedback
matrix is nonsingular. This phenomenon cannot occur in s.i.s.o. sys-
tems and the source of the complications in the m.i.m.o. case. This
is
difficulty can be overcome by using the concept of coprime factoriza-

tions, and the details can be found in [10] Desoer and Vidyasagar.
Problem 6.12. Consider the feedback system shown in Fig. 6.9. Deter-
mine the range of values of the constant k for which the feedback system is
Coo-stable, with
(i) g(s) = e -I (s 2 + 1)
s2 + 2s
= 1 — e~ 2ss
(ii) g(s)
s + 1
As (s 2
= e~ -f 4s + 4)
(iii) g(s)
s2 — 2s — 3
(iv) #(s) = 1 - e~ + e~ + s 3s
s + 5
Chap. 6 Input-Output Stability
276
Problem 6.13. Consider the feedback system shown in Figure 6.6.
Suppose
" S
e s
1
= s +2
Gi (j)
Lj
—+TT 1
1 - e
'3*
Is the feedback system loo-stable with
- n 11
=
1
1 —-1
+
1
7“*
< i)G ^ )=
Lo 2J
(ii) Ga (j) s 3
_2 0 -
01
(iii) G 2 O) = P
L_0 e 'J
.7
~
TIME-VARYING AND/OR NONLINEAR SYSTEMS
In the previous section, we studied the feedback stability of linear time-
invariant systems. In this section, we study nonlinear and/or time-
varying systems. Our attention is centered exclusively on single-input/
single-output systems. The stability of multi-input/multi-output sys-
tems, which requires more elaborate techniques, is studied in detail in
[10] Desoer and Vidyasagar.
In this section, we prove two well known results, known as the
circle criterion and the (generalized) Popov criterion, respectively. The
circle criterion provides a sufficient condition for the L 2 - stability of a
feedback system containing a linear time-invariant element in the for-
ward path and a memoryless time-varying and/or nonlinear element in
the feedback path. We show that the circle criterion is also a necessary
condition for stability, in a sense to be made precise later on. The gen-
eralization of the Popov criterion that is presented here provides a

sufficient condition for the Instability of a feedback system containing
a linear time-invariant system in the forward path and a time-invariant
memoryless nonlinearity in the feedback path. Both criteria are graph-
ical in nature, i.e., they are frequency domain criteria, and this feature
makes them easy to apply.
6.7.1 The Circle Criterion
To develop the circle criterion, we shall first introduce a defini-

tion.
] 1 (
Sec. 6. 7 Time- Varying AND/OR Nonlinear Systems 211
1 [definition A continuous function (f>:

R+ X R — > R is said to
belong to the sector [«, fi], a < fi,
if
2 o« 2 < x<f>(t, x) < fix 2

, \J t>0, \/ x e R
or, equivalently, if (1) <f>(t, 0) = 0, V > 0, and (2) t
3 a < <)?, V t > 0, V x ^ 0

Remarks : A continuous function <>
j : R + x R—+R belongs to
the sector [a, /?] if, for each fixed t, the graph of <f>(t , x) vs. x lies
between the straight lines passing through the origin with slopes a
and p, respectively. This is illustrated in Fig. 6.12.
We shall next prove a result very similar to Theorem [6.5(54)].
4 lemma Suppose that h{-) e d and that u e L le Then h* u e L 2e

.
,
and, moreover,
5 II
* u) T 1 2 h * u is causal. Hence by the definition of

causality,
6 (h * u) T = (h * uT)T
Thus
7 ||(^ * u)t\\i — II
(h* u T) T || 2 < \
\h * uT \\ 2
< sup 03
|
h(j q)\*\\ut \\ 2
where the last step follows from Theorem [6.5(54)]. —

We shall now prove a result that is commonly known as the circle
criterion.
278 Chap. 6 nput-Output Stability
8 Theorem ( Circle Criterion ) Consider a feedback system of the

type shown in Fig. 6.13. The element G is linear, time-invariant, and
described by
9 yi (t) = (Geît) = P g(t - T)e 1

(t)dT
Jo
FIG. 6.13 F
where the transfer function g is of the form
10 g(s) = g„(s) + “(•S')
where gb e h and d are polynomials with no common zeroes, all

(i,
zeroes of d are in C+ and the degree of h is less than or equal to that

,
of d. Further, gb is of the form
11 gb (s) = g (s) + S= gte-

a
0
isT
, g a (-)€L x
i
(i.e., the “delays” are all equally spaced). The element F is memory-
less and is described by
12 y 2 (t) = (Fe 2 )(t) = (/)[t, e 2 (t)]
where 0(., •) is continuous and belongs to the sector [a, /?]. Under
these conditions, the system under study is L 2 -stable if one of the
following sets of conditions, as appropriate, holds : If a, ft ^ 0, let
D be the disk in the complex plane centered on the real axis and
passing through the points 1/a j 0 and 1 //? j 0. — + — +
(i) IfO <a< p: (a) When the 70-axis is indented as in Sec.
6.6.2, the graph of g(ja>) is bounded away from the disk D ;
i.e.,
13 inf g(y'o)
|
—z >0 |
(OdzR
ZED
(b) The graph of g(jcci) encircles the disk D exactly v p times,

where vp is the number of poles of g with positive real part;
i.e.,
>
Sec. 6.7 Time-Varying AND/OR Nonlinear Systems 279
n
_ Arg [g = 2 nvp
14
^ D
lim Arg [gSBl
^ f -
)
,
\/ z
(ii) I/O = a < fi: (a) v p = 0, i.e., g has no poles with positive
real parts, and (b)
15 inf Re g(jco) +~>0

coER p
(iii) If a <0 < fi: (a) g has no poles in C+ ,
and (b) the graph of
g(jco) is entirely contained in the disk D and is bounded away
from the boundary of D.
(iv) If a < fi < 0: Replace g by — g, a by — /?, f by —a, and
apply (i), or (ii) above, as appropriate.
Remarks:
1. Suppose
16 g(s) = c'(sl - A) _1 b + d
where the matrix A is Hurwitz (i.e., all eigenvalues of A have
negative real parts), the pair (A, b) is controllable, and the
pair (A, c) is observable. Then Theorem (8) gives a sufficient
condition for the L 2 -stability of the system of Fig. 6.5. By
Problem 6.5, the same conditions also guarantee the global
asymptotic stability of the equilibrium point 0 of the system
[6.4 (27)]— [6 4(29)] .
;
Thus Theorem (8)
see also Problem 6.14.
contains, as a special case, the circle criterion for Liapunov
stability (Theorem [5.5(60)]). The details are left as an exercise.
On the other hand, Theorem (8) is more general than Theorem
[5.5(60)], because the former is applicable even to distributed
systems, i.e., to systems where the transfer function g is irra-

tional.
2. If /? — a— 0 and /?, a approach k ^ 0, then Theorem (8)
reduces to the sufficiency part of Theorem [6.6(58)].
proof Define
17 c =^+ «
2
II fc*
1 &
18 >
2
so that
19 =c—r
OC
20 P=c+r
Rearrange the system of Fig. 6.13 as in Fig. 6.14. In this way, the system
under study is transformed into another equivalent system whose forward
element G t has the transfer function
21 gt(s) =
1 + cg(s)
and whose feedback element F t is described by
22 (F y)(t)
t = 01 \t, y(t)] - cy(t) A y(t)]
FIG. 6.14
Clearly, the new function 0,0, •) belongs to the sector [— r, r], Now,
regardless of which set of conditions (i)-(iv) is applicable, one can verify,
using Theorem [6.6(58)], that g e & and
t that
23 sup [#,0'«)l
CO
= < v '
The details of the verification are left as a problem (see Problem 6.15).
Now, with regard to the rearranged system of Fig. 6.14, we have that
g g
t Qj and that 0,0, •) belongs to the sector [— r, r]. The latter condition
means that y lt e L le whenever e 2t e L le and that
24 1 1 (y 2 f)r 1 12 ^r \\ig2 t)T\\i, V f> 0

Also, by lemma (4), we have G e u e L le whenever
t e lt e L le and
25 ||(^îf)rll2 ^ yo|IOif)rll2j 0
We can now complete the proof. Suppose that u u u 2 e L 2 and that

e lu e 2t e L le satisfy the system equations
26 eu = u - F e 2t x t
27 e 2t = u 2 + G e u t
— [
Taking the truncations to [0, T ] of (26)-(27) and noting that F G

t, t
are
causal, we get
28 (.
— U \T ~~
e lt)r \Fti.e 2t)AT
29 ( e 2t)r
— u2 t ~ [Gt(e u)r]T
Taking norms, we get
30 Il(û)r|l2 < llîrlb + ||[^r(^2f)T]r|l2
< 1 +
1 1 12 1 1
-^*(^ 2 f)r II 2
<ll«irll 2 + r\\(e ) T 2 2t \\
[by (24)]
31 1 1 2 r)r 1 1 2 ^ Il
w 2rll2 + [^(^lf)r]rll2
II
< Il
w 2 r ||2 + \\G (e U T t ) \\2
< II
«2 r|| 2 + yo||(eif)rll 2 [by (25)]
Eliminating ||(e 2 f)rl 2 I

from (30) and (31) gives
32 Il(îf)rll2 < llîrlta + Kl I Marita + ôl (îr)r

I I 2]
33 (1 - ryo)||(û)rll 2 <ll«irll 2 + r \\u 2T \\ 2 < ll«i II 2 +r||w 2 || 2
Because ry Q < 1 [by (23)], we can divide both sides of (33) by 1 — ry 0 to

,
obtain
34
Because the right-hand side of (34) is finite and independent of T, we see

that e u SL 2 . Similarly, if we eliminate II {.e u)T 2 from (30) and (31), we
\\
get
35 ll (e2()r || 2
<lMk^MM
ry 0
which shows that e 2t e L 2 Thus . the system is Testable, in the sense of

definition [6.3(34)]. M
36 Example. Consider a system of the form shown in Fig. 6.13,
with g(-) as in Example [6.6(59)], i.e.,
s(s )
s.
= e_ 5 s
2
~+2
+ 4s
s2
Suppose it is desired to find the largest sector [a, fi] such that the circle
criterion is satisfied. Note that g has one pole with a positive real part,
vp = 1 ;
therefore if a < 0, the conditions for cases (ii)-(iv) of Theorem
17
(8) can never be satisfied. Thus, for the circle criterion to hold, we
must have a > 0.
The plot of g(jco) isshown in Fig. 6.10. For the system to be L 2 -
stable, the plot of g(j co) must encircle the critical disk D exactly once, in
17 The circle criterion is a sufficient condition for Instability. Hence if it is not

satisfied, we cannot conclude that the system is unstable.
a counterclockwise sense. From Fig. 6.10, we see that this is the case
provided 0.5 <a< < 0.55.
Hence the system is Z, 2 -stable for all
ft
nonlinearities in the sector [a, /?] with a > 0.5 and < 0.55.
We now show that the circle criterion of Theorem (8) is also a
necessary condition for the stability of a certain class of systems. First
of all, by applying Theorem (8), case (iii), we get the following result:
37 corollary Consider a feedback system of the form shown in Fig. 6.13.

Let G, F be as in Theorem (8), and suppose 0(-, •) belongs to the sector
[— r, r] for some r > 0. Then the system under study is Z, 2 -stable if g has
no poles in C+ and ,
38 sup \g(jCO)\ < r- 1
CO
Now, consider the linear time-invariant feedback system shown in

Fig. 6.15, where g is of the form
39 g(s) = g + g (s) +
0 a ga (-) G L x
FIG. 6.15
Thus g is of the form (10)— (1 1), except that there are no delay terms.
Suppose we ask: What conditions must g satisfy in order for the system
of Fig. 6.15 to be Z^-stable for all gains k e [— r, r] and for all delays
t>0? To put it another way, under what conditions does g(s)/
[1 + k e~ g(s)\ e (I for all k e [— r, r] and for all x
sx
0? The answer >
is readily obtained from an application of the graphical stability crite-
rion Theorem [6.6(58)], and is given next.
40 Theorem Suppose g is of form (39). Then g(.s)/[l + k e~ STg(s)] e (I

V k e [— r, r] and V t >
0, if and only if g e (I and (38) is satisfied.
proof
(i) “ If ” If g e & and (38) is satisfied, then the polar plot of g(jco) is
contained within a circle of radius less than r _1 Since \e~ Jcor .

\
= 1
V co and V T > 0, we see that the polar plot of e~ jcox

g(jco) is also
contained within a circle of radius less than r" 1 Thus the polar .
plot of e~ jcor g(jco) is bounded away from, and does not encircle, the
point —
1 /(k +/0) whenever k e [— r, r]. Also e~ ST g(s) e <5,
——
S ec. 6 7
. Time- Varying AND/OR Nonlinear Systems 283
V T > 0.
Hence by Theorem [6.6(58)], we see that g(y)/[l
+ ke~ ST
fi, V k G [— r, r], V T
g(s)] E 0. >
(ii) “ Only If” Suppose g(s)/[l + ke~ STg(s )] e & V k e [— r, r],
>
V T 0. Then, by setting k = 0, t = 0, we see that g e &. Next,
by the Riemann-Lebesgue lemma, ga (jco) >0 as |o| * oo, —
which means that g(jol) >g 0 as |co —
» oo. The hypothesis |
—
implies that g 0 < r~\ because if |g 0 > r” 1 then the function
| 1 | ,
g(s)/[l + kg(s)\ $ & with k = 1/g 0 Finally, suppose (38) is false; .
then there exist a complex number z with z r and a sequence | |

>
(co f)“
such that g(jcOt) z as i — —
> oo. Clearly the sequence (co,)~
is bounded, because g{j(o) g 0 as \co —

> oo and we have just
\
—
shown that |^ 0 < r~ 1 (whereas \z1
r
-1
). So there exists a finite
\
>
frequency co 0 such that g(jco 0 ) = z. Now let z \z\ exp (y‘0), and =
let r = + 2n)lco Then we have
(</> 0.
[e~ JcoT g(jCO)]co=coo = e~ g(jCQo) = \z\>r~ JcooZ 1

41
Hence the polar plot of e~ Jcorg(jco) intersects the point |z| + j0,
which shows that £(s)/[l + ke~ srg(s)] ^ ft with k = \z\~ 1 , because
the stability criterion [6.6(58)] is violated. This contradicts the
hypothesis that g(s)/[l ke~ sz g(s)\ g S + V k e [— r, r ] and
V t > 0. Hence we conclude that no such number z exists, i.e.,
that (38) holds. —
Thus Theorem (40) shows that the circle criterion is not at all
conservative when applied to systems of the form shown in Fig. 6.15.
Actually, we have restricted k to belong to the symmetrical interval
[— r, r] only to keep things simple, and we can prove similar results for
the case where k in Fig. 6.15 varies over an arbitrary interval [a, ft].
Combining Theorems (40) and (8), we get the following important
result.
42 Theorem Suppose g is of the form (39) with g 0 — 0 and that the

system of Fig. 6.15 is L^-stable for all k e [— r, r] and for all
t > 0. Then the system of Fig. 6.13 is L 2 -stable for all nonlineari-
ties 0(«, •) in the sector [—r,r].
Finally, combining Theorem (42) and Problem 6.5, we can prove

the following result, which is reminiscent of Aizerman’s conjecture
except that the result below is not a conjecture but a fact.
43 Theorem Suppose
44 gC?) = c'(sl — A) -1 b + d
where the matrix A is Hurwitz, the pair (A, b) is controllable, and the
pair (A, c) is observable. Suppose the system of Fig. 6.15 is L
stable V k e [— r, r] and V t > 0. Then the equilibrium point 0

of the system
45 ±(t) = Ax(/) — b u(t)

46 y(t = c'x(0 + du(t)
)
47 u(t) = j(?)]
is globally asymptotically stable whenever $(•, •) belongs to the

sector [— r, r].
6.7.2 The Popov Criterion
In this subsection we shall state and prove a generalization of

what is known as the Popov criterion. This criterion pertains to the sys-
tem of Fig. 6.16. Note that this system is the same as that in Fig. 6.13,
except that the nonlinear element 0( ) is replaced by the time-invariant
• , •
nonlinearity </>(•). For the purpose of proving the Popov criterion, we

shall rearrange the system as shown in Fig. 6.17.
The following two lemmas pave the way to Popov’s criterion.
48 lemma Suppose <j>: R —* R belongs to the sector [0, k]\ i.e., suppose
49 0 < (70(<7) < ka 2

, V cr e R
With the symbols as defined in Fig. 6.17, whenever q > 0, we have
50 yi{t)v(t)dt>^ yl{t)dt. V r> 0, V V e L 2e
J ^
a
proof Define <D : R —* R by
51 <3>(<7) = 0(<7) do
J*
Then <D(cr) >0, V o e R, because of the sector condition (49). From
Fig. 6.17, we have
52 y 2 (t) = 2 (t)]
</>[e
53 v(t) = e 2 (t) + qe 2 (t)

so that
54 y 2 (t)v(t) dt = e 2 (t)<f>[e 2 (t)\ dt +q e 2 (t)<f>[e 2 (t)] dt
> 4- f {<l>[e 2 (t)]}

2
dt +q f e 2 (t)<f>[e 2 (t)] dt
Jo Jo
= ^(0 dt + q
T Jo\ (*
Jo
e 2 (t)<f>[e 2 {t)\ dt
where we use (49). Hence (50) is established if we show that
55 q e 2 (t)<f>[e 2 (t)] dt> 0

If q = 0, (55) follows immediately, so suppose q > 0. Then
r
56 <7 f
Jo
e 2 (t)<f>[e 2 (t)] dt = q P^’ 0(<r) c/<x
J«2(0)
= ri®[« 2 (T)]-®[« 2 (0)B

However, because e 2 (-) is the convolution of v(-) and the function (1/#)
exp (— tjq ), we see that e 2 (0) = 0, whence d>[> 2 (0)] = 0. Hence
r
57 q e 2 (t)<f> [e 2 (t j\ dt = q<& [e 2 (T)\ > 0
JQ
This establishes (50). m
58 lemma Let led, and define
59 <5 = inf Re h(jco)

coER
Then, for all / e L 2e we have ,
T
60 P r(t)f(t) dt> 5 Jo f\t)
Jo
[ dt, V T> 0
where r = h*f
proof By causality, we have rT = (h* fT T ) . Now,
61 r(t)f(t) dt = rT (t)fT (t) dt

Jo
= \~(h*fT)(t)fT(t)d t
,
By ParsevaPs equality, we now have
62 r (t)f(t)dt = ~Re (h *fT)(jco)f*Uco) dco

f J
Re h(jco)\Mjco)\* dco
s
2n
'
i
2n \fT (jco)\*dco
I
= <5|IMII
where fT denotes the Fourier transform of /r (*). Clearly (62) is the same
as (60). ^
Now we come to the following.
63 Theorem Generalized Popov Criterion ) Consider the system

(< shown
in Fig. 6.17. Suppose the transfer function g satisfies
64 g(-),g(-)ea
Suppose the nonlinearity 0(«) satisfies (49). Under these conditions,
thissystem with inputs u and z is L 2 -stable if there exist constantsx
q > 0, 5 > 0 such that
65 Re [(1 + j coq)g(jco)] + -i- > > 0, <5 V E i?
Remarks : Theorem (63) guarantees Instability, in the sense that

both the “errors” e and v belong to L 2 whenever the two “inputs”
x
u and z belong to L 2 However, we are interested in concluding that

x
.
e1 and e2 belong to L2 ,
because these are the quantities of interest in
the original system (shown in Fig. 6.16). This is not very difficult. If
q > 0, then the function s 1/(1 + qs) e Ofc; hence e 2 E L2 when-

ever v e L 2 The restriction on the
. inputs cannot be removed; how-
ever, L 2 whenever u l9 z e L 2 Now, if
we can conclude that e u e 2 e .
q > 0 and u 2
an arbitrary element
is of L 2 the “input” z may or may ,
not belong to L 2 Hence, if q > 0, we need to make extra assumptions

.
about the nature of u 2 For example, we can assume that u 2 e L 2

. .
These comments are summarized in the following.
66 fact Under the conditions of Theorem (63), the system of Fig. 6. 1 6 has the
property that e u e 2 E I 2 whenever u u u2 e L2 ,
with the added assump-
tion that u 2 E L2 if q > 0.
proof of Theorem (63) Suppose (65) holds, and suppose u u z E L2 ,

eu
v E L 2e Note that,
. because g, g £ &, the function s i—> g(s)(l + qs) e a.
Next, given two functions fu f2 e L le ,

define their truncated inner product
as
67 <\fufi)T = f /i OO/iW dt
Jo
Note that </i,/2 >r is the inner product (in the L 2 -sense) of the L 2 - func-
tions and f2T , so that by Schwarz’s inequality, we have
68 K/ij/iX 1
! < ll/irlta’IIArlk
For the system of Fig. 6.17, we have
69 et = u — y2t
70 v =z+w
Hence
71 <<?i, w> r + <>, y 2 >r = <e u v — z) T + <v, «! — ei> r
= — (e z) T + (e u v) T ~ (v, e y T +
l9 t (v 9 u t y T
= - <e u z)T + <v, Mj> r

< Ikirll z + 2 *|l rll 2 ||^rll 2 *ll M ir II 2
< Ikirlk’IMta + 1 1 lb’ll ui II 2
On the other hand, by lemma (48), we have
72 (v, y 2 yT > -g || J^rlll
Now, if we define
73 V = inf Re [(1 + J COq)g(j CO)]

coER
74 m = sup I
(1 + jcoq)g(jco) |
coER
we have
75 ||wr l| 2 <m\\e 1T \\ 2
by lemma (4). Next, by lemma (58), we have
76 (e l9 wy T >V 1 1 tfirlli — (v
2"^ llîrlll
+ "j" Ikirlli 1
>0-4)11^+2^11^
= (y-l)\\e 1T \\l + £-2 \\(v-z)T \\l
Combining (72) and (76), we have
77 <(e l9 wy T + (v, y2 yT > || J^rlll + (v

2)
II Kirill
Because y 2 = u — eu
1 (77) becomes
<e i’ W>T + - e^rWl + (v - -|-)||e 12.|||
+ ^\\{v-z)T \\l
Combining (71) and (78) gives
” e drWi + — y) IkirIB + 2^2 \\(v “ z)t\\i

X (v
< IkirlMMb + Ikrlta’lki II2
Expanding ||(«! - e )T and ||(v - z)T gives

x \\l \\\
<c i Ikirlb + c2 \\vT \\ 2 + c3
where c l9 c 2 c 3 are
, finite constants. Now, by (65),
and of course 3/2m 2 > 0. Thus the left-hand side of (80) is a positive
definite form in e 1T and vT This implies that \\e 1T 2 \\v T 2 are bounded
.
\\ , \\
independently of T. Hence e u v e L 2 whenever u u z e L 2 .

—
Remarks:
1. The inequality (65) can be given a graphical interpretation,
which makes it quite useful in practice. Suppose we plot
g(jco) vs. co Im g(jco), with co as a parameter ranging from 0
to 00 . [Note that it is only necessary to do the plot for co e
[0, 00 ), because both Re g(jco) and co Im g(jco) are even func-
tions of co.] This graph is sometimes referred to as the Popov
The inequality (65) means that, for some 8 > 0, one
plot of g.
can draw a straight line with a nonnegative slope (namely l/q)
through the point (— l/k +
8 0) such that the plot of Re g(jco)
,
vs. co Im g(jco) lies entirely to the right of such a line. Such a

line is known as a Popov line. This situation is depicted in
Fig. 6.18.
2. If we set q = 0 in (65), (65) becomes the same as (15); i.e., the
Popov criterion becomes the same as the circle criterion
[Theorem (8), case (ii)] with a =
0. Because of the flexibility of
choosing the constant q >

0 in (54), the inequality (54) is less
restrictive than (15). This can be explained by noting that the
circle criterion [case (ii)] guarantees Instability for all feed-
back nonlinearities in the sector [0, /?], whereas the Popov
criterion guarantees Instability only for all time-invariant non-
linearities in the sector [0, k].
3. With the proof given above, the Popov criterion is a special

case of what is known as the passivity theorem.
81 Example. Let
^= s2 + 4s + 4
The Nyquist and Popov plot of g are shown in Figs. 6.19 and 6.20,
plot
respectively. From Fig. 6.20, we see that no matter how small 1 fk is
(i.e., no matter how large k is) we can always construct a suitable Popov
line, as indicated. Hence the system of Fig. 6.16 isZ 2 -stable for all time-
invariant nonlinearities 0(-) in the sector [0, k\ 9 for all finite k. On the
other hand, if we apply the circle criterion, case (ii), we see that
inf
oER
c
Re g(ja>) = OA
)
Hence (15) is satisfied whenever (1//?) > (1/32), i.e.,

fi < 32. Hence the
system of Fig. 6.13 isL 2 -stable for all possibly time-varying nonlinearities
0(», •) in the sector [0, fi] 9 for all /? < 32. This example shows that in
general the circle criterion and the Popov criterion yield different sta-
bility bounds, which should be interpreted differently.
An Application : Aizerman's Conjecture
In Sec. 5.5, we stated Aizerman’s conjecture, which we shall

restate here in a slightly different form for the sake of convenience.
Conjecture (Aizerman Suppose that g, g e Q and that
g/(l + hg) e Ct for all h e [0, k\. Then the system of Fig. 6.16 is
T 2 -stable for all nonlinearities 0(*) in the sector [0, k\.
Thus Aizerman’s conjecture states that if the system of Fig. 6.16

is stable whenever 0(*) is a constant gain of value h e [0, k] 9 then it is
stable for all nonlinearities </>(•) in the sector [0, k]. In general, Aizer-
man’s conjecture is false. However, Popov’s criterion provides a means
of identifying a large class of transfer functions g(-) for which Aizer-
man’s conjecture holds.
Suppose g, g e 6t. If g contains any impulses, g would contain
higher-order distributions and hence would not belong to a. Thus
g, g E Ct implies that g does not contain any impulses, i.e., that g e L 1 .
Now, because g e L u g/(l +

hg) e G, V h e [0, k], implies, by the
graphical stability criterion of Theorem [6.6(58)], that the Nyquist plot
of g neither intersects nor encircles the half-line segment (— oo, —1/A:].
Because the only difference between the Nyquist plot and the Popov
plot is in the vertical axis, the same is true of the Popov plot as well.
Now, suppose the Popov plot of g has the shape shown in Fig. 6.21.
Reg(/a;)
Then the stability of the nonlinear feedback system is assured for all
0(«) in the sector [0, k] 9 because of Thus § satisfies

Popov’s criterion.
Aizerman’s conjecture [see, for instance, g of Example (81)]. On the
other hand, suppose the Popov plot of g has the appearance shown in
d
Fig. 6.22. In this case, Popov’s criterion is not satisfied. However,

because Popov’s criterion is only a sufficient condition for stability, it
still does not follow that Aizerman’s conjecture is false for such a g.
Thus, in summary, Popov’s criterion provides a readily verifiable suf-
ficient condition for determining whether Aizerman’s conjecture is valid
for a particular transfer function g(«).
RegO'co)
Problem 6.14. Suppose g(-) is of the form (44), and that the criteria of
Theorem (8) are satisfied. Show that in this case 1 +k ^0 V k e [ -a, JJ]
Then, using the results of Problem 6.5, show that Theorem (8) contains
Theorem [5.5(60)] as a special case.
Problem 6.15. Show that if the criteria of Theorem (8) are satisfied, then
(23) holds.
Problem 6.16. Using the circle criterion, determine the largest range
[<%, /?] such that the system of Fig. 6.13 is Z, 2 -stable for all 0(«, •) in the
sector [a, /?] with
0) &(s)
s s )
g(s) =e 3s s
2
+ 3s + 2
s2 - 4
(ii)
mew - -^, 2
Problem 6.17. Using the Popov criterion, determine the largest con-
stant k such that the system of Fig. 6.16 is £ 2 -stable for all 0(-) in the
sector [0, k], with
(i) £ 0) =e j +4
i2 + +6
s 2
= e~ 2s i
(ii) g(s)
T+W
(
s 2 + 9s + 20
(iii) g(s) =e 3s
s3 + 6s 2 + 11s + 6
Do any of these transfer functions satisfy Aizerman’s conjecture?
Appendices
BELLMAN-GRONWALL INEQUALITY
1 Theorem Suppose c > 0, r(-) and k(-) are nonnegative valued
continuous functions, and suppose
2 r{t) <c+ f
•/o
k(r)r(r) dr, V t e [0, T]
Then
3 r(0 < c exp jj* &(t) V t e [0, T]
proof Let 5(0 denote the right-hand side of (1). Then (1) states that
4 r(t) < 5(0, V te [0,71

Further
5 = &(0 r(0 < k(t) 5(0, V t e [0, T]
6 5(0 - &(0 ^(0 < o, vre[o,r]

7 [i(0 — A:(0 *(0] exp —k(r) drj <,0, V t e [0, T]
d
8 5(0 exp -k(l) dx <0 V G T]
dt
X t [0,
9 s(t) exp jj"* —k(x) dxj < s(0) = c
292
)
Sec. II Summary of Matrix Exponentials and State Transition Matrices 293
s{t ) < c exp [J>H

Now, (10) and (4) together imply (3).
The utility of this inequality lies in that, given the implicit bound
(2) for r(t) [notice that the right-hand side of (2) also involves /*(•)],
we are able to obtain an explicit upper bound for r{t). In particular,

note that if r(.) is nonnegative valued and satisfies (2) with c = 0, then
r(t) = 0\/tG[0 9 T].
II
SUMMARY OF MATRIX EXPONENTIALS
AND STATE TRANSITION MATRICES
Given an n X n matrix A, the polynomial c(X) defined by
c(A) = det (21 — A)
is known as the characteristic polynomial of A. The zeros of the

characteristic polynomial of A are known as the eigenvalues of A. The
well known Cayley-Hamilton theorem states that
c(A) =0
A polynomial m(X of the lowest possible degree such that m(A) 0 =
is minimal polynomial of A. A minimal polynomial of A is a
called a
divisor of the characteristic polynomial of A, and every eigenvalue of
A is also a zero of a minimal polynomial of A.
Given an n X n matrix A, the n x n matrix-valued function
e
Af = 2 AT/z*!
= i 0
is known shown that the

as the matrix exponential of A. It can be
infinite series in (3) converges uniformly and absolutely over every
finite interval. Suppose m(X) is a minimal polynomial of A, with
zeroes A u X k of multiplicities m l9 ... ,m k respectively. Then it

. . . ,
can be shown that e At is of the form

k r m— 1
e Al =S
=
2j PijiWe**
i 1 j=0
where the p t - s are the so-called interpolating polynomials. The following

fact is obvious from the expansion (4).
fact e At
is bounded as a function of t if and only if Re A 0 V /, and t <
m =
t whenever Re At = 0 (i.e., if all eigenvalues of A have nonpositive
1
real parts, and all eigenvalues of A having zero real parts are simple zeros
of a minimal polynomial of A) e At » 0 as t ;
—
* oo if and only if Re At < 0 —
V / (i.e., all eigenvalues of A have negative real parts).
:
294 Appendix
Note that the infinite series
/
S= A
0
i+1
t‘/i\
also converges absolutely and uniformly over every finite interval to

At
Ae At So we
. see that e is a solution of the matrix differential equation
M(0 - AM(0 V t > Q, M(0) = I

Further, in view of the existence and uniqueness theory developed in
Sec. 3.4, we see that M(7) = e
At
is in fact the unique solution of (7).
The solution of the vector differential equation
x(0 = Ax(/) V > 0, x(0) = x t 0
is given by
x(t) = eA 'x(0) = e Atx 0
In the case of the time-varying linear vector differential equation
x(t) = A0)x(0 V > t t 09 x(t 0 ) =x 0
where A(-) is piecewise-continuous, the solution for x(t) is given by
x(0 = 0>(t, t 0 )x(t 0 )

= <P(t, t 0 )x 0
where 0(-, •) is the unique solution of the linear matrix differential

equation
t 0) = A(r)®0, t0) V t > t0 , ®(? 0 , t 0) =I

The matrix <!>(•, -) is known as the state transition matrix, and satisfies
the following properties
m, T)r = (T,o, Vf,t

®(t, s) = Q>(t, s), V t,T,s
For a detailed treatment of these topics, see [7] Chen.

,
References
[1] Arnold, V. I. Ordinary Differential Equations MIT Press, Cambridge, Mass.,

1973.
[2] Athans, M., and Falb, P. L. Optimal Control McGraw-Hill, New York, 1966.
,
[3] Barman, J. Well-Posedness of Feedback Systems and Singular Perturbations,
Ph.D. Thesis, Univ. of California, Berkeley, Ca. 1973.

[4] Bellman, R. E. Introduction to Matrix Analysis, 2nd ed., McGraw-Hill, New
York, 1970.
[5] Bergen, A. R. and Franks, R. L. “Justification of the Describing Function
Method,” SIAM J. Control, vol. 9, pp. 568-589, 1971.
[6] Blaquiere, A. Nonlinear System Analysis, Academic Press, New York, 1966.
[7] Chen, C. T. Introduction to Linear System Theory, Holt, Rinehart, & Winston,
New York, 1971.
[8] Chua, L., and Lin, P. M. Computer-Aided Analysis of Electronic Circuits:
Algorithms and Computational Techniques, Prentice-Hall, Englewood Cliffs,
N.J., 1975.
[9] Desoer, C. A., and Shensa, M. J. “Networks with Very Small and Very Large
Parasitics: Natural Frequencies and Stability,” Proc. IEEE, vol. 58, pp.
1933-1938, 1970.
[10] Desoer, C. A. and Vidyasagar, M. Feedback Systems: Input-Output Properties,
Academic Press, New York, 1975.
[11] Eggleston, H. G. Convexity, Cambridge University Press, Cambridge, 1966.
[12] Gelb, A. and Vander Velde, W. E. Multiple-Input Describing Functions and
Nonlinear System Design, McGraw-Hill, New York, 1968.
[13] Hahn, W. Stability of Motion, Springer-Verlag, Berlin, 1967.
295
296 References
[14] Hille, E., and Phillips, R. S. Functional Analysis and Semigroups, Amer.
Math. Soc., Providence, R.I., 1957.
[15] Jury, E. I. Inners and Stability of Dynamic Systems, John Wiley & Sons,
New York, 1975.
[16] Lefschetz, S. Stability of Nonlinear Control Systems, Academic Press,
New York, 1962.
[17] Narendra, K. S. and Taylor, J. H. Frequency Domain Criteria for Absolute
Stability, Academic Press, New York, 1973.
[18] Nemytskii, V. V. and Stepanov, V. V. Qualitative Theory of Differential
Equations, Princeton Univ. Press, Princeton, 1960.
[19] Ralston, A. A First Course in Numerical Analysis, McGraw-Hill, New York,
1965.
[20] Roy den, New York, 1963.
H. L. Real Analysis, Macmillan,
[21] Sandberg, W. “On the L 2 -Boundedness of Solutions of Nonlinear Func-
I.
tional Equations,” Bell Sys. Tech J., vol. 43, pp. 1581-1599, 1964.
.
[22] Sandberg, I. W. “A Frequency-Domain Condition for the Stability of

Feedback Systems Containing a Single Time-Varying Nonlinear Element,”
1601-1608, 1964.
Bell Sys. Tech. J., vol. 43, pp.
[23] Sandberg, W. “Some Results on the Theory of Physical Systems Gov-
I.
erned by Nonlinear Functional Equations,” Bell Sys. Tech. J ., vol. 44, pp.
871-898, 1965.
[24] Taussky, O. “A Generalization of A Theorem of Liapunov,” Siam J. Appl.
Math., vol, 9, pp. 640-643, 1961.
[25] Wilkinson, J. H. The Algebraic Eigenvalue Problem, Clarendon Press, Ox-
ford, 1965.
[26] Willems, J. C. The Analysis of Feedback Systems, MIT Press, Cambridge,
Mass., 1970.
[27] Zames, G. “On the Input-Output Stability of Time-Varying Nonlinear
Feedback Systems, Part I: Conditions Derived Using Concepts of Loop
Gain, Conicity and Positivity,” IEEE Trans. Auto. Control, vol. AC-11,
pp. 228-238, 1966.
[28] Zames, G. “On the Input-Output Stability of Time-Varying Nonlinear
Feedback Systems, Part II: Conditions Involving Circles in the Frequency
Plane and Sector Nonlinearities,” IEEE Trans. Auto. Control, vol. AC-11,
pp. 465-476, 1966.
Index
Index
A
a 246
, C*, 52
a ,249 Cn x », 64
tte, 250 C[a, b], 59
Absolute stability, 199 C"[a,
, b], 59
Adams-Bashforth algorithms, 121 Cauchy sequence, 55
Adams-Moulton algorithms, 121 Causality, 229
Aizerman’s conjecture Center, 19, 20
in input-output setting, 290 Cetaev’s theorem, 163
(see also Circle criterion, input- Circle criterion, input-output, 278
output, necessity of) necessity of, 282
Asymptotic stability, 135 Circle criterion, Liapunov, 207
theorems, 153, 158 Complete inner product space (see
Asymptotic stability in the large (see Hilbert space)
Global asymptotic stability) Complete normed space (see Banach
Autonomous system, 5 space)
Continuity (in a normed space), 55
uniform, 55
B Continuous dependence on initial con-
ditions
BCn [ta ,
oc
), 133 example of lack of, 4
Banach algebra, 247 theorem, 86
Banach space, 56 Contraction mapping theorem
Bellman-Gronwall inequality, 292 global, 73
Bendixson’s theorem, 31 local, 76, 78
299
300 Index
Convergence (in a normed space), 54 Function (cont.):

Converse theorems (Liapunov), 165 locally positive definite, 141
Convolution, 231, 246 conditions for, 142, 143
measurable, 225
of class K, 142
D positive definite, 141
conditions for, 142, 143
Derivative along a trajectory, 39, 145
Describing function, 103
use in predicting limit cycles, 107
bounds on, 106, 1 12
G
Generating solution, 46
E Global asymptotic stability, 136
theorems, 154, 158
Epidemic equation, 28
Equilibrium point, 5
isolated, 7
Equivalent gain, 103
H
Equivalent linearization (see Optimal
Harmonic balance, 43, 102
quasilinearization)
Hilbert space, 62
Euclidean norm, 8, 57, 62
Holder’s inequality, 227
Euler method
graphical, 23
numerical, 114
Existence and uniqueness theorem I
for differential equations

global, 83 Impulse response, 97, 231
linear, 87 Indented j w -axis, 269
local, 79, 81 Index, 36
for feedback systems, 237 theorems, 37
Existence of solutions Induced matrix norm, 65
example of lack of, 3 formulae, 67
theorems, 79, 81, 83, 87, 237 submultiplicativity of, 66
Inner product space, 60
Instability (Liapunov), 132
F theorems, 160, 162, 163
Invariant set, 156
Feedback systems
Isocline method, 26
existence and uniqueness of
solutions, 237
stability conditions, 264, 266, 274,
278, 283, 286 K

Finite settling time, 94
Focus, 19, 20 Kalman- Yacubovitch lemma, 201
Function Kalman’s conjecture, 200
decrescent, 143 Krylov-Boguliubov method, 42
1
Index 301
L Lipschitz constant, 80
Loop transformations, 204, 216
Lur’e problem, 199
Lp , 226 Lur’e-Postnikov criterion, 211
extended, 228
1$ 229
Instability, 233 M
Liapunov function, 149
common, 202 Mathieu equation, 151, 159, 208
existence of quadratic, 175, 181 Measure of a matrix, 67
Lur’e-Postnikov type, 210 formulae, 71
Popov type, 213 properties, 67
Liapunov function candidate, 149 Memoryless nonlinearity, 97
Liapunov matrix equation, 172 odd, 101
closed-form solution, 174 Minkowski’s inequality, 227
stability conditions based on, 175
Multistep algorithms, 118
uniqueness of solutions, 173 exactness constraints, 119
Liapunov’s direct method, 147
Liapunov’s indirect method, 186
Limit cycle (see also Periodic solution), 31 N
analytical method for, 38
existence, 34 Node, 15, 20

nonexistence, 31 Nonautonomous system, 5
Limit point, 34 Norm, 54

Limit set, 34 Normed linear space, 54
positive, 156 Numerical solution techniques, 113
Linear systems Nyquist criterion, 270
asymptotic stability, 168
canonical forms for second order, 15
exponential stability, 170 O
global asymptotic stability, 168
Li -stability, 250, 254, 257 Optimal quasilinearization
L 2 -stability, 252, 254 with bias reference input, 100
Instability, 250, 254, 258 with sinusoidal reference input, 101
Loc-stability, 250, 254, 255
periodic, 183
singularly perturbed, 124 P
slowly varying, 184, 223
stability, 167 Paley-Wiener lemma, 261
stability of autonomous, 175 Pendulum equation, 6, 29, 40, 48, 149
uniform asymptotic stability, 169 Periodic solution, 13
uniform stability, 169 analytical method for, 38
Linear vector space, 51 existence, 34, 107

Linearization method, 21 nonexistence, 31
Lipschitz condition, 80 Phase-locked loop, 156
verification of, 88 Phase plane, 1
1 1 1 1
302 Index
Phase-plane plot, 1 Solution estimates for linear differential

Phase-plane trajectory, 1 equations
Picard’s iterations, 85 forced, 93
Poincare'-Bendixson theorem, 34 unforced, 89
Popov criterion (input-output), 286 Stability (feedback), 238
Popov criterion (Liapunov), 213 Stability (input-output), 233
Popov plot, 215, 288 relation to Liapunov stability, 240,
Power series method, 45 243, 244
Predator-prey equation (see Volterra Stability (Liapunov), 132
equation) theorem, 148
Predictor-corrector algorithm, 122 State plane, 1
Principle of harmonic balance, 43, 102 State-plane plot, 1
State-plane trajectory, 11
State variable, 2
R State vector, 2
Stationary point (see Equilibrium point)
R n 52
,
Subspace, 53
R nxn , 64
RC transmission line, 271
Rayleigh’s equation, 29, 48
Region of attraction, 135 T
Return difference, 264
Return difference matrix, 274 Taylor series method, 113
Runge-Kutta method Topological equivalence, 58
fourth order, 116 Truncation, 228
second order, 115
S U
Saddle point, 17, 20 Uniforpi asymptotic stability, 135

Schwarz’s inequality, 61 theorem, 153
Sector, 106, 202 Uniform continuity, 55
Secular solutions, 46 Uniform stability, 132
Set theorem, 148
connected, 31 Uniqueness of solutions
invariant, 156 example of lack of, 3
limit, 34 theorems, 79, 81, 83, 87, 237
of measure zero, 225
positive limit, 156
simply connected, 31 y
Singular perturbations
linear systems, 124 Vector field, 12
nonlinear systems, 196 direction, 12
Singular point (see Equilibrium point) method, 27
Slowly-varying oscillations, 42 30
radial,
Slowly-varying systems sum, 30
linear, 184, 223 Vector space (see Linear vector space)
nonlinear, 218 Volterra equation, 9, 29, 37, 39, 48, 134

Nonlinear Systems - Vydiasagar

Diunggah oleh

Informasi Dokumen

Hak Cipta

Format Tersedia

Bagikan dokumen Ini

Bagikan atau Tanam Dokumen

Opsi Berbagi

Apakah menurut Anda dokumen ini bermanfaat?

Apakah konten ini tidak pantas?

Hak Cipta:

Format Tersedia

Nonlinear Systems - Vydiasagar

Diunggah oleh

Hak Cipta:

Format Tersedia

Nonlinear Systems Analysis

Professor of Electrical Engineering

Prentice-Hall, Inc. Englewood Cliffs , New Jersey 07632

(Prentice-Hall electrical engineering series)

© 1978 by Prentice-Hall, Inc., Englewood Cliffs, N.J. 07632

All rights reserved. No part of this book

Printed in the United States of America

Prentice-Hall International, Inc., London

1 .2 Autonomy, equilibrium points, 5

Nonlinear differential equations, 50

3.1 Mathematical preliminaries, 51

Approximate analysis methods, 96

4.1 Describing functions, 97

Stability in the sense of Liapunov, 131

5.1 Basic definitions, 131

Input-output stability, 224

6.1 Introduction to L^-spaces, 225

This book is intended as a text for a one-term or one-quarter course in

nonlinear systems, at either the first-year graduate or senior-graduate

only prerequisite for using the book is a course in ordinary differential

classroom hours, while most of it can be covered in forty hours.

The book contains five chapters besides the introduction. Chapter

1. All items within each section of each chapter (equations, theorems,

reference is made an item in another section, the full number

given, e.g. example [5.1(13)] means example (13) in Sec. 5.1.

This means that 0 is the unique number in [0, 2 n) such that

sin ^ = (^rftir’ co ^ = wtxw ri

Thus tan -1 is a function of both variables x and x 2 and not just of x ,

the ratio x 2 /x Note that tan -1 is well-defined everywhere in R 2

except at (0, 0).

output function (or response function ). If we define the auxiliary

then the single nth-order equation (1) can be equivalently expressed as

Finally, if we define ^-vector-valued functions x(«)* R + —* Rn and

then the n first-order equations (5)-(8) can be combined into a first-

order vector differential equation, namely,

11 x(t) = f [t, x(t), i/(0], t >0

Similarly, suppose a system with p inputs and k outputs is

described by a set of k ordinary differential equations of the form

14 x(0 = [x t (f) . . • xn (t)]'

where we take m = Q 0, and define

17 ±{t) =f [t, x(t), u(0], t >0

depends continuously on the initial condition x(0).

18 Example. Consider the scalar differential equation

19 x(t) = —sign x(t), t> 0; x(0) =0

21 Example. Consider the scalar differential equation

22 *(0 = 2 x(t)’ t —° ; *(°) =0

tems containing time delays.

25 Example. Consider the scalar differential equation

but there is no continuously differentiable function x(-) defined over

ness of solutions to (17), and their continuous dependence on the initial

Problem Determine whether or not each of the following differ-

(a) x(t) = [*(/)] 1/3 ; x(0) =0

Sec. 1 2 Autonomy, Equilibrium Points 5

x(0 =f [t, x(/)]

problems, one is generally concerned with the behavior of a system of

Then [1.1(17)] can be rewritten as

x(0 = fjt, x(t)\

[ definition The system described by be autonomous

[definition] A vector x 0 e R" is said to be an equilibrium

f(t, x0) = 0, V t > t0

x 0 is an equilibrium point of (1) at time t 0 then it is clear that

x0 is an equilibrium point of (1) at all times t x

if (1) is autonomous, then x 0 e R is an equilibrium point of (1) at

some time if and only if it is an equilibrium point of (1) at all times.

22 (0 = 2 x(t)’ t —° ; (°) =0

0 = [io * 2 o]' is an equilibrium point of (12) 3 if and only if

i(0 =/i[i(0, * 2 (0]