Stability Analysis of Nonlinear Systems 2016

Systems & Control: Foundations & Applications
Vangipuram Lakshmikantham
Srinivasa Leela
Anatoly A. Martynyuk
Stability Analysis
of Nonlinear Systems
Second Edition

Series Editor
Tamer Basar, University of Illinois at Urbana-Champaign, Urbana, IL,
USA
Editorial Board
Karl Johan strm, Lund University of Technology, Lund, Sweden
Han-Fu Chen, Academia Sinica, Beijing, China
Bill Helton, University of California, San Diego, CA, USA
Alberto Isidori, Sapienza University of Rome, Rome, Italy
Miroslav Krstic, University of California, San Diego, CA, USA
H. Vincent Poor, Princeton University, Princeton, NJ, USA
Mete Soner, ETH Zrich, Zrich, Switzerland;
Swiss Finance Institute, Zrich, Switzerland
Roberto Tempo, CNR-IEIIT, Politecnico di Torino, Italy
More information about this series at http://www.springer.com/series/4895
Vangipuram Lakshmikantham
Srinivasa Leela Anatoly A. Martynyuk
Stability Analysis
of Nonlinear Systems
Second Edition
Vangipuram Lakshmikantham (1924 2012)

Florida Institute of Technology
Department of Mathematical Sciences
Melbourne, FL, USA
Srinivasa Leela
State University of New York
Geneseo, NY, USA
Anatoly A. Martynyuk
National Academy of Sciences of Ukraine
Kiev, Ukraine
ISSN 2324-9749
ISSN 2324-9757 (electronic)
ISBN 978-3-319-27199-6
ISBN 978-3-319-27200-9 (eBook)
DOI 10.1007/978-3-319-27200-9
Library of Congress Control Number: 2015958665
Mathematics Subject Classification (2010): 34Dxx, 37C75, 93Dxx, 34Gxx, 34Kxx, 70K20
Springer Cham Heidelberg New York Dordrecht London
1st edition Taylor & Francis Group LLC (successor of Marcel Dekker, Inc.), Boca Raton, Florida, 1989
Springer International Publishing Switzerland 2015
This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part
of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations,
recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission
or information storage and retrieval, electronic adaptation, computer software, or by similar or
dissimilar methodology now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this
publication does not imply, even in the absence of a specific statement, that such names are exempt
from the relevant protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this
book are believed to be true and accurate at the date of publication. Neither the publisher nor the
authors or the editors give a warranty, express or implied, with respect to the material contained
herein or for any errors or omissions that may have been made.
Printed on acid-free paper
Springer International Publishing AG Switzerland is part of Springer Science+Business Media
(www.birkhauser-science.com)
PREFACE TO THE SECOND EDITION
The first edition of this book was published in 1989 by Marcel

Dekker, Inc. For the last two decades the methods of stability investigation of nonlinear systems presented in this book have acquired
further development and have been widely applied in the investigation of some types of contemporary dynamical systems. We note
here the following directions only: practical stability; fuzzy differential equations and inclusions; set differential equations in metric
spaces; fractional dynamic systems; causal differential equations; differential equations in cones; uncertain dynamical systems; dynamic
equations on time scales; weakly connected nonlinear systems.
Since then this book became a rare one, meanwhile the methods
of stability analysis set out in it possess a considerable potential for
further development of qualitative theory of equations.
The present edition constitutes some expansion of the First Edition of this book. Namely, each of the Chapters 14 is added with
new results displaying the development of stability theory. Chapter
5 is new. It contains applications of general methods for solution
of applied problems. In particular, new stability conditions are presented for motion of a robot interacting with a dynamic medium.
Conditions for stability are established for affine systems. Stability
conditions in a physical system are indicated for synchronization of
beams of two connected lasers. Stability conditions are found for the
Takagi-Sugeno system and their application for impulse control in the
predator-prey system is shown. These results have been obtained
for the recent years at the Stability of Processes Department of the
S.P. Timoshenko Institute of Mechanics of NAS of Ukraine.
The List of References at the end of the book is supplied with new
citations of the authors papers published after 1989. This new list of
references will enable all potential readers to approach the boundary
vi
Preface to the Second Edition
of the investigations beyond which there are open problems in the

elds.
The book is intended for graduate students and researchers in
the applied mathematics, the physical sciences and any of the engineering disciplines, who are interested in the qualitative theory of
equations. The authors strongly hope that any 2-places in the
book will be revealed by potential readers and each author would
appreciate receiving any comments from scientic community.
We undertook our work over the second edition of the book in
2009. To our deepest regret the work was darkened by the death of
Professor V.Lakshmikantham who passed away in 2012 and it is only
now when we are completing the second edition.
We appreciate the eorts and patience of many colleagues at the
Stability of Processes Department of the S.P. Timoshenko Institute of
Mechanics, National Academy of Sciences of Ukraine in the preparing
and correcting the manuscript.
Geneseo-Kiev, June 2015
S. Leela
A. A. Martynyuk
PREFACE
The problems of modern society are both complex and interdisciplinary. Despite the apparent diversity of problems, however, often
tools developed in one context are adaptable to an entirely dierent
situation. For example, consider Lyapynovs second method. This
interesting and fruitful technique has gained increasing signicance
and has given decisive impetus for modern development of stability
theory of dierential equations. A manifest advantage of this method
is that is does not require the knowledge of solutions and therefore
has great power in applications. There are several books available
expounding the main ideas of Lyapynovs second method, including
some extensions and generalizations.
It is now recognized that the concept of Lyapynov-like function
and the theory of dierential and integral inequalities can be utilized to study qualitative and quantitative properties of non-linear
dierential equations. Lyapunov-like function serves as a vehicle to
transform a given complicated dierential system and therefore it is
enough to investigate the properties of this simpler dierential system. It is also being realized that the same versatile tools are adaptable to discuss entirely dierent nonlinear systems, and other tools
such as the method of variation of parameters and monotone iterative technique provide equally eective methods to investigate problems of similar nature. Moreover, interesting new notions and ideas
have been introduced which seem to possess great potential. Due to
the increased interdependency and cooperation among mathematical
sciences across the traditional boundaries and the accomplishments
thus far achieved, there is every reason to believe that many breakthroughs are there waiting and oering an exiting prospect for this
versatile technique to advance further.
It is in this spirit that wee see the importance of our monograph.
vii
viii
Preface
Its aim is to present a systematic account of the recent developments,

describe the current state of the useful theory, show the essential
unity achieved by the wealth of applications and provide a unied
general structure applicable to a variety of nonlinear problems.
Some important features of the monograph are as follows: This is
the rst book that (i) presents a systematic study of stability theory
in terms of two dierent measures and exhibits the advantage of the
employing families of Lyapunov functions; (ii) treats the theory of a
variety of inequalities clearly bringing out the underlying theme; and
(iii) demonstrates the manifestations of general Lyapunov method by
showing how this eective technique can be adapted to study various
apparently diverse nonlinear problems.
This book also stresses the importance of utilizing dierent forms
of nonlinear variations of parametric formulae to discuss qualitative
behaviors of nonlinear problems, examines the constructive methods
generated by monotone iterative technique and the method of upper
and lower solutions, and illustrates the application of theoretical results to several dierent models chosen from real-world phenomena
for the benet of the practitioners.
In view of the existence of several excellent books on stability by
Lyapunovs second method, we have restricted ourselves to presenting new developments, illustrative examples and useful applications,
and consequently there is a minimum of overlap with the existing
books.
V. Lakshmikantham
S. Leela
A. A. Martynyuk
CONTENTS
Preface to the Second Edition
Preface
1 Inequalities
1.0 Introduction . . . . . . . . . . . . . .
1.1 GronwallType Inequalities . . . . .
1.2 WendorType Inequalities . . . . .
1.3 BihariType Inequalities . . . . . . .
1.4 Multivariate Inequalities . . . . . . .
1.5 Dierential Inequalities . . . . . . .
1.6 Integral Inequalities . . . . . . . . .
1.7 General Integral Inequalities . . . . .
1.8 IntegroDierential Inequalities . . .
1.9 Dierence Inequalities . . . . . . . .
1.10 Interval-Valued Integral Inequalities
1.11 Inequalities for Piecewise Continuous
1.12 Reaction-Diusion Inequalities . . .
1.13 Notes . . . . . . . . . . . . . . . . .
vii
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
Functions
. . . . . .
. . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
2 Variation of Parameters and Monotone Technique

2.0 Introduction . . . . . . . . . . . . . . . . . . . . . . .
2.1 Nonlinear Variation of Parameters . . . . . . . . . .
2.2 Estimates of Solutions . . . . . . . . . . . . . . . . .
2.3 Global Existence and Terminal Value Problems . . .
2.4 Stability Criteria . . . . . . . . . . . . . . . . . . . .
2.5 Method of Upper and Lower Solutions . . . . . . . .
2.6 Monotone Iterative Technique . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
1
1
2
12
16
24
26
31
35
38
46
52
56
61
65
.
.
.
.
.
.
.
67
67
68
73
83
89
93
96
ix
Contents
2.7
2.8
Method of Mixed Monotony .

Method of Lower and Upper
Analysis . . . . . . . . . . . .
2.9 Integro-Dierential Equations
2.10 Stability in Variation . . . . .
2.11 Dierence Equations . . . . .
2.12 Notes . . . . . . . . . . . . .
. . . . . . . . . . . . .
Solutions and Interval
. . . . . . . . . . . . .
. . . . . . . . . . . . .
. . . . . . . . . . . . .
. . . . . . . . . . . . .
. . . . . . . . . . . . .
. 100
.
.
.
.
.
103
106
116
123
133
3 Stability of Motion in Terms of Two Measures

135
3.0 Introduction . . . . . . . . . . . . . . . . . . . . . . . . 135
3.1 Basic Comparison Results . . . . . . . . . . . . . . . . 136
3.2 Stability Concepts in Terms of Two Measures . . . . . 140
3.3 Stability Criteria in Terms of Two Measures . . . . . . 143
3.4 A Converse Theorem in Terms of Two Measures . . . 150
3.5 Boundedness and Lagrange Stability in Terms of Two
Measures . . . . . . . . . . . . . . . . . . . . . . . . . 156
3.6 Stability Results for Autonomous or Periodic Systems 158
3.7 Perturbing Family of Lyapunov Functions . . . . . . . 162
3.8 M0 -Stability Criteria . . . . . . . . . . . . . . . . . . . 170
3.9 Several Lyapunov Functions . . . . . . . . . . . . . . . 177
3.9.1 Vector Lyapunov functions method . . . . . . . 177
3.9.2 Matrix-valued Lyapunov functions method . . 190
3.10 Cone Valued Lyapunov Functions . . . . . . . . . . . . 195
3.11 Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . 199
4 Stability of Perturbed Motion
4.0 Introduction . . . . . . . . . . . . . . . . . . . .
4.1 Stability of Perturbed Motion in Two Measures
4.2 Stability of Perturbed Motion (Continued) . . .
4.3 A Technique in Perturbation Theory . . . . . .
4.4 Stability of Delay Dierential Equations . . . .
4.5 Stability of Integro-Dierential Equations with
Finite Memory . . . . . . . . . . . . . . . . . .
4.6 Stability of Integro-Dierential Equations of
Volterra Type . . . . . . . . . . . . . . . . . . .
4.7 Integro-Dierential Equations (Continued) . . .
4.8 Stability of Dierence Equations . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
201
. 201
. 202
. 205
. 209
. 216
. . . . 227
. . . . 231
. . . . 234
. . . . 237
Contents
xi
4.9 Impulse Dierential Equations . . . . . . . . . . . . . 241

4.10 Reaction-Diusion Equations . . . . . . . . . . . . . . 249
4.11 Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . 252
5 Stability in the Models of Real World Phenomena 253
5.0 Introduction . . . . . . . . . . . . . . . . . . . . . . . . 253
5.1 Stability of a Robot Interacting with a Dynamic Medium254
5.2 Stabilization of Motions of Ane System . . . . . . . 270
5.3 Synchronization of Motions . . . . . . . . . . . . . . . 273
5.4 Stability of Regular Synchronous Generation of Optically Coupled Lasers . . . . . . . . . . . . . . . . . . . 276
5.5 Models of World Dynamics and Sustainable
Development . . . . . . . . . . . . . . . . . . . . . . . 285
5.6 Stability Analysis of Impulsive Takagi-Sugeno Systems 293
5.6.1 General results . . . . . . . . . . . . . . . . . . 293
5.6.2 Impulsive Fuzzy Control for Ecological
PreyPredator Community . . . . . . . . . . . 304
5.7 Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . 308
References
311
Index
327
1
INEQUALITIES
1.0
Introduction
This chapter introduces the theory of inequalities of various types

and therefore forms a basis for the remaining chapters.
In Section 1.1, we discuss Gronwall type integral inequalities
clearly bringing out the main ideas involved in several variants and
extensions of such inequalities. Section 1.2 is devoted to Wendorff
type integral inequalities where we consider some important extensions of Gronwall type inequalities to multivariate integral inequalities. Section 1.3 deals with nonlinear integral inequalities of separable
type which are known as Bihari type inequalities and discusses several variants of importance, while Section 1.4 considers some typical
Bihari integral inequalities with several independent variables.
Theory of differential inequalities which plays a prominent role in
the study of qualitative and quantitative behavior of nonlinear differential systems is developed in Section 1.5, where the extension of
this theory relative to abstract cones is also presented. The study of
nonlinear integral inequalities forms the content of Section 1.6. In
Section 1.7, we deal with nonlinear integral inequalities of general
type which have a broad applicability in the sense that they contain
several special cases found in the literature. Section 1.8 concerns itself with the theory of integro-differential inequalities where we also
discuss general comparison results that enable us to reduce the study
of integro-differential inequalities to that of differential inequalities.
Some interesting special cases are given here to illustrate the advantage of the approach. Theory of differential inequalities is discussed

V. Lakshmikantham et al., Stability Analysis of Nonlinear Systems,
Systems & Control: Foundations & Applications, DOI 10.1007/978-3-319-27200-9_1
Inequalities
in Section 1.9 where among others the discrete versions of Gronwall

and Bihari inequalities are dealt with. Section 1.10 is devoted to
the study of interval-valued integral inequalities. Here the intrinsic advantage of interval mappings, namely, the inclusion monotonic
property, is exploited. As a by-product, generalization of Gronwall
inequality to interval maps is obtained. Section 1.11 deals with differential and integral inequalities with impulsive eect. Finally in
Section 1.12, we discuss necessary theory for reaction-diusion inequalities.
1.1
GronwallType Inequalities
We begin with one of the simplest and most useful integral inequalities.
Theorem 1.1.1 Let m, v C[R+ , R+ ] where R+ denotes the
nonnegative real line. Suppose further that, for some c 0, we have
t
m(t) c +
v(s)m(s) ds,
t t0 0.
(1.1.1)
t t0 .
(1.1.2)
t0
Then,
t
m(t) c exp
v(s) ds ,
t0
Proof If c > 0, it follows from (1.1.1) that

m(t)v(t)
v(t) ,
t
c + v(s)m(s) ds
t0
which, upon integration yields

t
t
v(s)m(s) ds log c
log c +
t0
v(s) ds.
t0
Gronwalltype inequalities
This, together with (1.1.1) implies (1.1.2).

If c = 0, then (1.1.1) holds for every constant > 0 and therefore
the previous argument gives (1.1.2) with c = . Letting 0, we
get m(t) 0. This proves the theorem.
We have given the classical proof of Theorem 1.1.1 because of its
importance. We could also prove it by employing a linear dierential
inequality and the variation of constants formula. We shall use this
idea for a more general case.
Let us rst prove the basic result on linear dierential inequalities
which we use often.
Lemma 1.1.1 Let m C 1 [R+ , R+ ], v, h C[R+ , R+ ] and
m (t) v(t)m(t) + h(t),
m(t0 ) = c 0,
t t0 .
Then
t
m(t) c exp
t
t
v(s) ds +
t0
h(s) exp

v() d ds ,
t0

t
Proof Set q(t) = m(t) exp v(s) ds so that we have
t0

q (t) = [m (t) v(t)m(t)] exp

h(t) exp
v(s) ds
t0
t
t
v(s) ds ,
q(t0 ) = c.
t0
Hence it follows that

t
s
h(s) exp
q(t) c +
t0

v() d ds
t0
and this implies the stated conclusion proving the lemma.
t 0.
Inequalities
Theorem 1.1.2 Let m, v, h C[R+ , R+ ] and suppose that

t
m(t) h(t) +
t t0 .
v(s)m(s) ds,
(1.1.3)
t0
Then,
t
t
m(t) h(t) +

t t0 .
(1.1.4)
t
t

v(s) ds + h (s) exp
v() d ds,
t t0 .
[v(s)h(s)] exp
v() d ds,
s
t0
If h is dierentiable, then
t
m(t) h(t0 ) exp
t0
t0
(1.1.5)
Proof To prove (1.1.4), set p(t) =
t
v(s)m(s) ds, so that p(t0 ) = 0
t0
and
p (t) v(t)m(t),
t t0 .
Since m(t) h(t) + p(t), we get

p (t) v(t)p(t) + v(t)h(t),
Letting q(t) = p(t) exp(
t
t t0 .
v(s) ds), we see that q(t0 ) = 0 and
t0

q (t) = [p (t) v(t)p(t)] exp
v(s) ds
t0

h(t)v(t) exp
t
t
v(s) ds ,
t0
which implies

t
q(t)
h(s)v(s) exp
s
t0
t t0 .
v() d ds,
t0
Consequently, we obtain
t
t
p(t)
v(s)h(s) exp
t t0
v() d ds,
s
t0
and this gives (1.1.4) immediately.

To prove (1.1.5), set the right hand side of (1.1.3) as p(t) so that
p (t) = v(t)m(t) + h (t),
p(t0 ) = h(t0 )
which, in view of (1.1.3), yields the linear dierential inequality

p (t) v(t)p(t) + h (t),
p(t0 ) = h(t0 ).
Now, we get
t
p(t) h(t0 ) exp
t
t

v(s) ds + h (s) exp
v() d ds,
t0
t t0
t0
and the estimate (1.1.5) follows.

The estimates (1.1.4) and (1.1.5) apparently look dierent. In
fact, they are equivalent. For, integrating by parts,
t
t
t
v() d ds = h(t) h(t0 ) exp
h (s) exp
t0
t
t
h(s)v(s) exp
+
t0

v() d
t0
v() d ds,
s
which, in view of (1.1.5), yields (1.1.4). Thus, it is clear that assuming dierentiability of h does not oer anything new.
Inequalities
If, in Theorem 1.1.2, h is assumed to be nondecreasing and positive, then the estimate (1.1.4) reduces to
t

m(t) h(t) exp
t t0 .
v(s) ds ,
(1.1.6)
t0
For, setting w(t) =
m(t)
, we get from (1.1.3)
h(t)
t
w(t) 1 +
t t0 .
v(s)w(s) ds,
t0
Hence, by Theorem 1.1.1, we have w(t) exp
t

v(s) ds and (1.1.6)
t0
follows.
We can also get (1.1.6) from (1.1.4) because
t

t
v(s)h(s) exp
h(t) +
t0
t
t
h(t) 1 +

v() d ds
v(s) exp

t
(s)
v() d ds
s
t0
= h(t) 1
d(s)
t
= h(t) exp
t0

v() d ,
t t0 .
t0
t
Example 1.1.1 Let m C[R+ , R+ ] and m(t) (A + Bm(s)) ds,
t t0 , where A 0 and B > 0. Then,
m(t)
A
[exp(B(t t0 )) 1] ,
B
t0
t t0 .
Example 1.1.2 Let m C[R+ , R+ ] and

t
m(t) A +
(B + cm(s)) ds,
t0
t t0 ,
where A, B 0 and c > 0. Then,

m(t)
B
[exp(c(t t0 )) 1] + A exp(c(t t0 )),
c
t t0 .
Example 1.1.3 Let m C[R+ , R+ ] and for B > 0

t
m(t) e + B
t
m(s) ds,
t t0 .
t0
Then,
t t0 .
m(t) exp((B + 1)t Bt0 ) ,
We shall now consider integral inequalities with separable kernels

since these can also be reduced to linear dierential inequalities.
Theorem 1.1.3 Let m, h, q, v C[R+ , R+ ] and suppose that
t
m(t) h(t) +
t t0 .
q(t)v(s)m(s) ds,
(1.1.7)
t0
Then,
t
t
m(t) h(t) + q(t)
v(s)h(s) exp
We set p(t) =
t t0 . (1.1.8)
v() d ds,
s
t0
Proof
t
v(s)m(s) ds, so that p(t0 ) = 0 and
t0
p (t) = v(t)m(t). Since m(t) h(t) + q(t)p(t), we obtain

p (t) v(t)q(t)p(t) + v(t)h(t),
t t0 ,
and consequently
t
t
p(t)
v(s)h(s) exp
t0

v()q() d ds,
t t0 ,
Inequalities
which yields (1.1.8).

Example 1.1.4 Let m C[R+ , R+ ] and for t t0 ,
t
m(t) m(t0 ) exp(r(t t0 )) +
[exp(r(t s))](Am(s) + B) ds,

t0
where r, A, B 0 and r A > 0. Then, for t t0 ,

m(t) m(t0 ) exp[(r A)(t t0 )]
B
[1 exp((r A)(t t0 ))] .
+
rA
Corollary 1.1.1 Let m, h, gi , vi C[R+ , R+ ], i = 1, 2, . . . , n and
m(t) h(t) +
n

i=1
t
gi (t)
vi (s)m(s) ds,
t t0 .
t0
Then,
t
t
m(t) h(t) + G(t)
V (s)h(s) exp

V ()G() d ds,
t t0 ,
t0
where G(t) = sup gi (t) and V (t) =

i
vi (t).
i=1
Corollary 1.1.2 Let m, h, gi , 1 , 2 C[R+ , R+ ], i = 1, 2, . . . , n

and
t
m(t) h(t) + 1 (t)
g1 (s)m(s) ds + 2 (t)
t0
n

i=1
t
ci
gi (s)m(s) ds,
t0
for t [t0 , T ], where ci 0, ti [t0 , T ] and t0 t1 t2 . . .

tn T . Then,
m(t) K1 (t) + M K2 (t) ,
t [t0 , T ],
where
t
t
K1 (t) = h(t) + 1 (t)
g1 (s)h(s) exp
t
t
g1 (s)2 (s) exp
n

ci
i=1
1
n

i=1
ci
ti
1 ()g1 () d ds,
ti
gi (s)K1 (s) ds
t0
n

i=1
provided
t0
M=
1 ()g1 () d ds,
s
t0
K2 (t) = 2 (t) + 1 (t)
ti
ci
,
gi (s)K2 (s) ds
t0
gi (s)K2 (s) ds < 1.
t0
An integral inequality with a non-separable kernel that is dierentiable can be reduced to an integro-dierential inequality which
can further be reduced to a dierential inequality.
2 , R ],
Theorem 1.1.4 Suppose that h, m C[R+ , R+ ], K C[R+
+
Kt (t, s) exists, is continuous and nonnegative. Let
t
m(t) h(t) +
K(t, s)m(s) ds,
t t0 .
(1.1.9)
t0
Then,
t
t
m(t) h(t) +
(s, t0 ) exp
t0

A(, t0 ) d ds,
t t0 , (1.1.10)
where
t
A(t, t0 ) = K(t, t) +
Kt (t, s) ds,
t0
and
t
(t, t0 ) = K(t, t)h(t) +
Kt (t, s)h(s) ds.

t0
10
Inequalities
t
Proof Setting p(t) =
K(t, s)m(s) ds, we have p(t0 ) = 0 and
t0
t
p (t) = K(t, t)m(t) +
Kt (t, s)m(s) ds.

t0
Using the fact that p(t) is nondecreasing and m(t) h(t) + p(t), we
arrive at

t
t

p (t) K(t, t) + Kt (t, s) ds p(t) + K(t, t)h(t) + Kt (t, s)h(s) ds
t0
t0
A(t, t0 )p(t) + (t, t0 ),
t t0 ,
which yields the estimate (1.1.10).

Corollary 1.1.3 Under the assumption of Theorem 1.1.1, we
have

t
c
, t I = t t0 : v(s) ds < 1 .
m(t)
t
1 v(s) ds
t0
t0
2 , R ] and
Corollary 1.1.4 Let m, , v C[R+ , R+ ], w C[R+
+
for t t0 ,

t
s
m(t) c1 + (t) c2 + c3 (v(s)m(s) + w(s, )m( ) d ) ds ,
t0
t0
where c1 , c2 , c3 0. Then, for t t0 ,

t
m(t) c1 + (t) c2 exp c3
v2 (s) ds
t0
t
t
+
v1 (s) exp
t0
v2 () d
s

ds ,
11
where
t
v1 (t) = c3 c1 v(t) +
w(t, )d
t0
and
t
v2 (t) = c3 (t)v(t) +
w(t, )( ) d .
t0
2 , R ],
Corollary 1.1.5 Suppose that m C[R+ , R+ ], K C[R+
+
3
g C[R+ , R+ ] and Kt (t, s), gt (t, s, ) exist, are continuous and nonnegative. If
t
m(t) m(t0 )+
t s
K(t, s)m(s) ds+
t0
g(t, s, )m( ) d ds,
t t0 ,
t0 t0
then,
t
m(t) m(t0 ) exp

t t0 ,
A(s, t0 ) ds ,
t0
where
t
A(t, t0 ) = K(t, t)+
t
Kt (t, s) ds+
t0
t s
g(t, t, ) d +
t0
gt (t, s, ) d ds.
t0 t0
Finally, we shall prove a result on integral inequality that contains

Volterra and Fredholm integrals.
Theorem 1.1.5 Let m, v C[R+ , R+ ], T > t0 0 and
t
m(t) m(t0 ) +
T
v(s)m(s) ds +
t0
t0 t T . Then, if exp
v(s)m(s) ds,
(1.1.11)
t0
T

v(s) ds < 2, we have
t0

t

m(t0 )
v(s) ds ,
m(t)
T
exp
t0
2 exp
v(s) ds
t0
t0 t T.
12
Inequalities
Proof Set p(t) equal to the right hand side of (1.1.11). We obtain
T
p (t) v(t)p(t)
and p(t0 ) = m(t0 ) +
v(s)m(s) ds,
t0
which yields
t
p(t) p(t0 ) exp

v(s) ds ,
t0 t T.
t0
T

T
v(s) ds , and
But p(T ) = m(t0 ) + 2 v(s)m(s) ds p(t0 ) exp
t0
t0
T
v(s)m(s) ds = p(t0 ) m(t0 ).
t0
Consequently, p(t0 )
1.2
2exp
m(t0 )
T
and the result follows.
v(s) ds
t0
WendorType Inequalities
Most of Gronwalltype inequalities can be extended to multivariate

integral inequalities. We shall only prove some important extensions
to demonstrate the ideas involved. Let us begin by proving a simple
Wendor inequality.
2 , R ] and for some c 0,
Theorem 1.2.1 Let m, v C[R+
+
x y
m(x, y) c +
v(s, t)m(s, t) ds dt,
x x0 , y y 0 .
(1.2.1)
x 0 y0
Then,
x y
m(x, y) c exp

v(s, t) ds dt ,
x 0 y0
x x0 , y y 0 .
(1.2.2)
13
Wendorfftype inequalities
Proof Denote the right hand side of (1.2.1) by p(x, y) so that

p(x0 , y0 ) = 0 and
y
y
v(x, t)m(x, t) dt
px (x, y) =
y0

v(x, t) dt p(x, y),
y0
in view of the fact that p(x, y) is nondecreasing in x, y and

m(x, y) p(x, y). It then follows that
x y
m(x, y) p(x, y) c exp
v(s, t) ds dt ,
x x0 , y y 0 .
x 0 y0
The next result is an extension of Theorem 1.1.2.

2 , R ] and
Theorem 1.2.2 Let m, h, v C[R+
+
x y
m(x, y) h(x, y) +
v(s, t)m(s, t) ds dt,
x x0 , y y 0 .
x 0 y0
(1.2.3)
Then, for x x0 , y y0 ,
m(x, y) h(x, y)
x y
x y
v(s, t)h(s, t) exp
+
x 0 y0

v(, ) d d ds dt.
(1.2.4)
If, in addition, h is positive and nondecreasing in x and y, we have

x y
m(x, y) h(x, y) exp

v(s, t) ds dt ,
x x0 , y y 0 .
x 0 y0
(1.2.5)
14
Inequalities
Proof Let p(x, y) =
Rx Ry
v(s, t)m(s, t) ds dt so that p(x0 , y 0) = 0
x 0 y0
and
Zy
px (x, y)
v(x, t) dt p(x, y) +
y0
Zy
v(x, t)h(x, t) dt,
y0
using the fact that m(x, t) h(x, t) + p(x, t) and p(x, t) is nondecreasing in x and y. Hence, we get
p(x, y)
Zx Zy
v(s, t)h(s, t) exp
x 0 y0
Zx Zy
s
v(, ) d d ds dt,
from which follows the estimate (1.2.4).

If h is positive and nondecreasing in x, y, by setting K(x, y) =
m(x, y)
, we get from (1.2.4), the inequality
h(x, y)
K(x, y) 1 +
Zx Zy
v(s, t)K(s, t) ds dt.
x 0 y0
The estimate (1.2.5) is now a consequence of Theorem 1.2.1.

Corollary 1.2.1 If, in Theorem 1.2.2, h(x, y) = ax + by with
a (x), b (y) 0, then, for x x0 , y y0 ,
(a(x) + b(y0 ))(a(x0 ) + b(y))
m(x, y)
exp
(a(x0 ) + b(y0 ))
Zx Zy
v(s, t) ds dt .
x 0 y0
Finally, we give a typical multivariate integral inequality result in

which the inequalities between vectors are understood to be componentwise.
For x0 x, x0 , x Rn , let us use the notation
Zx
x0
v(s) d(s) =
Zx1 Zx2
x01 x02
...
Zxn
x0n
v(s1 , s2 , . . . , sn ) ds1 ds2 . . . dsn .
15
Wendorfftype inequalities
2 , R ] and for x x ,
Theorem 1.2.3 Let m, h, v C[R+
+
0
m(x) h(s) +
Zx
v(s)m(s) ds.
(1.2.6)
x0
Then
m(x) h(s) +
Zx
v(s)h(s) exp
x0
Zx
x x0 .
v() d ds,
(1.2.7)
If, in addition, h(x) is positive and nondecreasing, we get

Zx
m(x) h(s) exp

Rx
Proof Set p(x) =
x x0 .
v() d ds,
(1.2.8)
v(s)m(s) ds so that p(x10, x2 0, . . . , xn 0) = 0 and
x0
px1 (x) =
Zx2
...
x02
Zxn
v(x1 , s2 , . . . , sn ) ds2 . . . dsn .
x0n
Using the relation m(x) h(x) + p(x) and the nondecreasing character of p(x), we get
Px1
+
Zx2
x02
x
2
Z
...
...
x02
Zxn
v(x1 , s2 , . . . , sn ) ds2 . . . dsn p(x)
x0n
x
n
Z
v(x1 , s2 , . . . , sn )h(x1 , s2 , . . . , sn ) ds2 . . . dsn .
x0n
Consequently, we obtain
p(x)
Zx1 "
x01
exp
Zx
s
v() d
! Zx2
x02
...
Zxn
x0n
v(x1 , s2 , . . . , sn )
16
Inequalities
h(x1 , s2 , . . . , sn ) ds2 . . . dsn ds

which gives the estimate (1.2.7).
If h(x) is positive and nondecreasing, then, setting K(x) =
(1.2.6) yields
K(x) 1 +
Zx
v(s)K(s) ds,
x0
m(x)
h(x) ,
x x0
and in turn, this implies that

Zx
K(x) exp
v(s) ds ,
x0
proving the estimate (1.2.8).
1.3
BihariType Inequalities
The theory of Gronwalltype integral inequalities discussed in Section 1.1 can be extended to separable type nonlinear integral inequalities which are known as Biharitype inequalities. In this section, we
prove several such results that correspond to results in Section 1.1.
Theorem 1.3.1 Let m, v C[R+ , R+ ], g C[(0, ), (0, )] and
g(u) be nondecreasing in u. Suppose, for some c > 0,
m(x) c +
Then
Zt
v(s)g(m(s)) ds,
t t0 > 0.
"
t0
m(t) G1 G(c) +
Zt
v(s) ds ,
t0
t0 t < T
holds, where
G(u) G(u0 ) =
Zu
u0
ds
,
g(s)
G1 (u)
(1.3.1)
17
Biharitype inequalities

t
is the inverse of G(u) and T = sup t t0 : G(c) + v(s) ds
t0

domG1 .
Proof Denote the right hand side of (1.3.1) by p(t) so that
p(t0 ) = c and p (t) v(t)g(m(t)). Since g is nondecreasing and
m(t) p(t), we get
p (t) v(t)g(p(t)) ,
p(t0 ) = c.
Integrating from t0 to t, we obtain

p(t)
G(p(t)) G(c) =
t
dz
g(z)
v(s) ds
t0
and consequently

m(t) p(t) G1 G(c) +
t
v(s) ds ,
t0 t < T.
t0
A function g C[R+ , R+ ] is said to be subadditive if

g(u + v) g(u) + g(v)
and supadditive if the inequality is reversed.
Theorem 1.3.2 Let m, v, h C[R+ , R+ ], g C[(0, ), (0, )]
and g(u) be nondecreasing such that
t
m(t) h(t) +
v(s)g(m(s)) ds,
t t0 > 0.
t0
Then,
(i) if g(u) is subadditive,

1
m(t) h(t) + G
t
G(c) +

v(s) ds ,
t0
t0 t T0 < T, (1.3.2)
18
Inequalities
where G, G1 and T are as in Theorem 1.3.1 and

c=
ZT0
v(s)g(h(s)) ds ;
t0
(ii) if h is nondecreasing,
1
m(t) h(t0 ) + G
"
Proof Setting p(t) =
G(h(t0 )) +
Zt
v(s) ds ,
t0
Rt
t0 t < T. (1.3.3)
v(s)g(m(s)) ds, we get p(t0 ) = 0 and
t0
p (t) v(t)g(p(t)) + v(t)g(h(t)) ,

because of the properties of g. Note (t) =
Rt
v(s)g(h(s)) ds is non-
t0
decreasing and hence, setting c = (T0 ) for some T0 , t0 T0 < T ,

we obtain
p(t) c +
Zt
t0
v(s)g(p(s)) ds,
t0 t T0 < T.
Now, the estimate (1.3.2) follows in view of Theorem 1.3.1.

If h is nondecreasing, we have, from the definition of p(t),
g(m(t)) g(h(t0 ) + p(t)). Setting h(t0 ) + p(t) = w(t), we obtain
w (t) = p (t) = v(t)g(m(t)) v(t)g(m(t)) ,
w(t0 ) = h(t0 ).
Hence, by Theorem 1.3.1, we arrive at the estimate (1.3.3).

The estimate (1.3.2) can also be obtained if, in Theorem 1.3.2,
g(u) is assumed to be nonincreasing and superadditive in u.
Using the ideas of Theorems 1.1.2 and 1.1.4, one can prove the
following result.
19
Theorem 1.3.3 Let m, h C[R+ , R+ ], g C[(0, ), (0, )] and

3 , R ], K (t, s)
g(u) be nondecreasing in u. Suppose that K C[R+
+
t
exists, is continuous, nonnegative and for t t0 ,
t
m(t) h(t) +
K(t, s)g(m(s)) ds .
t0
Then
(i) if g is subadditive,

1
m(t) h(t) + v2 (t) + G
t
t0 t T0 < T,
v1 (s) ds ,
G(c) +
t0
where G, G1 , T are as in Theorem 1.3.1 and

T0
t
v1 (s)g(v2 (s)) ds,
c=
v1 (t) = K(t, t) +
t0
Kt (t, s) ds
t0
and
t
v2 (t) = K(t, t)h(t) +
Kt (t, s)g(h(s)) ds ;
t0
(ii) if h is nonincreasing,

m(t) h(t) h(t0 ) + G1 G(h(t0 )) +
t
v1 (s) ds ,
t0 t < T.
t0
2 , R ] such
Theorem 1.3.4 Let m, v C[R+ , R+ ] and w C[R+
+
that
t
m(t) c +
[v(s)m(s) + w(s, m(s))] ds,

t0
t t0 ,
(1.3.4)
20
Inequalities
where c > 0. Suppose further that

t

t
v(s) ds ) (t)g(z) exp
w(t, z exp

v(s) ds ,
t0
(1.3.5)
t0
where C[R+ , R+ ], g C[(0, ), (0, )] and g(u) is nondecreasing in u. Then,

t

t
1
G(c) + (s) ds exp
v(s) ds ,
(1.3.6)
m(t) G
t0
t0
t0 t < T , G, G1 and T being the same as in Theorem 1.3.1.

Proof
Set the right hand side of (1.3.4) equal

t
v(s) ds , so that we have, using (1.3.4) and (1.3.5),
p(t) exp
t0
[p (t) + v(t)p(t)] exp
t

v(s) ds
= v(t)m(t) + w(t, m(t))
t0
v(t)p(t) + (t)g m(t) exp
exp
v(s) ds
t0
t
t
v(s) ds .
t0
Since g is nondecreasing and m(t) p(t) exp
t

v(s) ds , we get
t0
p (t) (t)g(p(t)) ,
Hence, by Theorem 1.3.1,

1
p(t) G
t
G(c) +
p(t0 ) = c .
(s) ds ,
t0
t0 t < T ,
to
21
which yields the estimate (1.3.6).

Example 1.3.1 Let m, v, h C[R+ , R+ ] such that for c > 0,
0 p < 1,
t
m(t) c +
t
h(s)(m(s))p ds,
v(s)m(s) ds +
t0
t t0 .
t0
Then,
s
t
m(t) cq + q
h(s) exp
t0

v() d
1
t
ds
exp
t0

v(s) ds ,
t0
for t t0 , where q = 1 p.
If the kernel K(t, s) in Theorem 1.3.3 is such that Kt (t, s) 0,
then the arguments of that theorem oer only a crude estimate. However, a dierent approach gives a better bound which we discuss in
the next result.
Theorem 1.3.5 Suppose m C[R+ , R+ ], g C[(0, ), (0, )],
g(u) is nondecreasing in u and for some c > 0, > 0,
t
m(t) c +
e(ts) g(m(s)) ds,
t t0 .
(1.3.7)
t0
Then,
m(t) (1 + 0 )c ,
t t0 ,
where 0 > 0 satises

g((1 + 0 )c) c0 = 0
and
g((1 + )c) c > 0 ,
(1.3.8)
[0, 0 ).
Proof Set the right hand side of (1.3.7) equal to p(t) so that
p(t0 ) = c and
p (t) = g(m(t)) (p(t) c) g(p(t)) p(t) + c .
(1.3.9)
22
Inequalities
With the transformation p(t) = (1 + z(t))c and = t, it is easy to

reduce (1.3.9) to
dz
1
g((1 + z)c) z,
d
c
= 0.
We claim that z( ) 0 , [0 , ). If this is not true, there exists
a < such that z( ) = 0 and z( ) 0 , [0 , ]. In view

1
of (1.3.8), it is clear that c
g((1 + z)c) < 0 for all [0, 0 ] and
hence,
0
=
0
ds
0 s
z( )

0
1
c g((1
ds
0 ,
+ s)c) s
which is a contradiction. Therefore, z( ) 0 for [0 , ), which

in turn leads to
m(t) p(t) (1 + 0 )c ,
t t0 .
The following result is a nonlinear analogue of Theorem 1.1.5.

Theorem 1.3.6 Let m, v C[R+ , R+ ], g C[(0, ), (0, )],
ds
= . For c > 0 and T0 > t0 ,
g(u) is nondecreasing in u and g(s)
let
t
m(t) c +
T0
v(s)g(m(s)) ds +
t0
v(s)g(m(s)) ds,
(1.3.10)
t0 t T0 ,
(1.3.11)
t0
t [t0 , T0 ]. Then,

m(t) G1 G(c0 ) +
t

v(s) ds ,
t0
where G, G1 are the same as in Theorem 1.3.1 and c0 is a zero of

T0
the functional equation Z(s) = G(2s c) G(s) v(s) ds = 0.
t0
23
Proof Designating the right hand side of (1.3.10) by p(t), we get

T0
p(t0 ) = c +
p (t) v(t)g(p(t)) .
v(s)g(m(s)) ds,
t0
Note that the range of G is R+ , in view of the assumption

ds
= .
g(s)
Hence, by Theorem 1.3.1, we obtain

t
p(t) G1 G(p(t0 )) +
v(s) ds ,
t0 t T0 ,
(1.3.12)
t0
Now, observe that

T0
2p(t0 ) c = c + 2
v(s)g(m(s)) ds
t0
= p(T0 ) G1 G(p(t0 )) +
T0
v(s) ds .
t0
As a result, we get
T0
G(2p(t0 ) c) G(p(t0 ))
v(s) ds .
t0
Letting Z(s) = G(2s c) G(s)
T0
t0
v(s) ds, we see that Z( 2c ) < 0,
Z (s) > 0 for s > 2c and Z(Kc) > 0 for suciently large K > 0.
Thus, it is clear that there exists a c0 such that 2c < c0 < Kc which
is a solution of Z(s) = 0 and therefore, p(t0 ) c0 . Hence, by (1.3.12),
it follows that the estimate (1.3.11) holds.
24
Inequalities
Example 1.3.2 Taking g(u) = up , 0 p < 1 in Theorem 1.3.6,

the estimate (1.3.11) reduces to
1
t
m(t) cq0 + q
v(s) ds
t0 t T0 ,
q = 1 p.
t0
1.4
Multivariate Inequalities
We shall consider, in this section, some typical integral inequalities

with several independent variables . Our rst result deals with two
independent variables to bring out the ideas and to make the arguments simple.
2 , R ], g C[(0, ), (0, )] and
Theorem 1.4.1 Let m, v C[R+
+
g(u) is nondecreasing in u. Suppose that, for some c 0,
x y
m(x, y) c +
v(s, t)g(m(s, t)) ds dt ,
x x0 , y y0 , (1.4.1)
x 0 y0
holds. Then, we have for x0 x a, y0 y b,

1
m(x, y) G
x y
v(s, t) ds dt ,
G(c) +
(1.4.2)
x 0 y0
where G1 is the inverse of G, G(u) G(u0 ) =
u
u0
given by

ds
g(s)
x y
(a, b) = sup x x0 , y y0 : G(c) +
and a, b are
v(s, t) ds dt dom G
x 0 y0
Proof Denoting the right hand side of (1.4.1) by p(x, y), we obtain
2 p(x, y)
= v(x, y)g(m(x, y)) v(x, y)g(p(x, y))
xy
(1.4.3)
25
Multivariate inequalities
and
p(x0 , y) = p(x, y0 ) = p(x0 , y0 ) = c .
Since
(1.4.4)
2p
p p
2 G(p)
= G (p)
+ G (p)
,
xy
xy
x y
using (1.4.3) and noting that

G (p) =
1
,
g(p)
G (p) 0 and
p p
,
0,
x y
we get the dierential inequality

2 G(p)
v(x, y)
xy
integration of which yields, in view of (1.4.4), the relation
x y
G(p(x, y)) G(c) +
v(s, t) ds dt .
(1.4.5)
x 0 y0
The desired estimate (1.4.2) easily follows from (1.4.5) and the
proof is complete.
Example 1.4.1 Let v(x, y) = exp[(xx0 )+(yy0 )] and g(u) = u2 .
Then, for x0 x a, y0 y b,
m(x, y) c[1 c(exp(x x0 ) 1)(exp(y y0 ) 1)]1 ,
where a, b satisfy a0 = exp(a x0 ) 1 > 0 and b = c +
log(1 + ca0 ) log ca0 .
The next result generalizes the inequality (1.4.1) and deals with
several independent variables. We follow the notation of Section 1.2.
n , R ], h(x) > 0 and be deTheorem 1.4.2 Let m, h, v C[R+
+
creasing in x, g C[(0, ), (0, )], g(u) be nondecreasing in u and
1
u
v g(u) g( v ) for any v > 0. Suppose further that
x
m(x) h(x) +
v(s)g(m(s)) ds,
x0
x x0 .
(1.4.6)
26
Inequalities
Then, we have the estimate

x
m(x) h(x) + G1 G(1) + v(s) ds ,
x0 x a ,
(1.4.7)
x0
where G, G1 are the same functions as in Theorem 1.4.1 and a is

x
given by a = sup{x x0 : G(1) + v(s) ds dom G1 }.
x0
Proof Taking into account the assumptions on h and g, we get

from (1.4.6), the inequality

x
x
g(m(s))
m(s)
m(x)
1 + v(s)
ds 1 + v(s)g
ds,
h(x)
h(s)
h(s)
x0
x0
for x x0 . Applying Theorem 1.4.1, we now obtain the estimate

(1.4.7).
1.5
Dierential Inequalities
It is well known that the theory of dierential inequalities plays an

important role in the study of qualitative and quantitative behavior
of dierential systems of various kinds. This theory is also useful
in the investigation of integral and integrodierential inequalities
since, in many cases, their study cad be reduced to that of dierential
inequalities. We shall develop, in this section, the necessary theory
of dierential inequalities. We shall use vectorial inequalities freely
with the understanding that the same inequalities hold between their
respective components.
Consider the dierential system
u = g(t, u) ,
u(t0 ) = u0 ,
(1.5.1)
where g C[R+ Rn , Rn ]. The function g is said to be quasimonotone nondecreasing if x y and xi = yi for some i, 1 i n, implies
gi (t, x) gi (t, y) for each t. To appreciate this denition, note that
if g(t, u) = Au, where A is an n n matrix, g is quasimonotone
implies that aij 0, i, j = 1, 2, . . . , n, i = j. Let us now introduce
Differential inequalities
27
the notion of extremal solution of (1.5.1). Let r(t) be a solution of

(1.5.1) existing on some interval J = [t0 , t0 + a]. Then r(t) is said to
be the maximal solution of (1.5.1) existing on J, we have
u(t) r(t) ,
t J.
(1.5.2)
A minimal solution is dened similarly by reversing the inequality

(1.5.2).
We need the following known results for our discussion whose
proof may be found in Lakshmikantham and Leela [1] and Walter
[1].
Theorem 1.5.1 Let g C[E, Rn ], where E is an open (t, u) set
in Rn+1 and g(t, u) be quasimonotone nondecreasing in u for each t.
Then,
(i) if (t0 , u0 ) E, the system (1.5.1) has extremal solutions which
can be extended to the boundary of E;
(ii) if J is the largest interval of existence of the maximal solution
r(t) of (1.5.1), given any compact interval [t0 , T ] of J, there exists
an 0 > 0 such that for 0 < < 0 , solutions u(t, ) of
u = g(t, u) + ,
u(t0 ) = u0 +
(1.5.3)
exist on [t0 , T ] and lim u(t, ) = r(t) uniformly on [t0 , T ].

0
Lemma 1.5.1 Let v, w C[J, R] and for some xed Dini derivative Dv(t) w(t), t J S, where S is an almost countable subset of J. Then D v(t) w(t) on J, where D v(t) =
lim inf(1/h)[v(t + h) v(t)].
h0
The following scalar comparison result contains the central idea

of the theory of inequalities.
Theorem 1.5.2 Let g C[R+ R+ , R] and r(t) be the maximal
solution of (1.5.1) existing on [t0 , ). Suppose that m C[R+ , R+ ]
and Dm(t) g(t, m(t)), t t0 , where D is any xed Dini derivative.
Then m(t0 ) u0 implies m(t) r(t), t t0 .
Proof By Lemma 1.5.1, it follows that
D m(t) g(t, m(t)) ,
t t0 .
28
Inequalities
Let t0 < T < for all suciently small > 0 and lim u(t, ) = r(t)
0
uniformly on [t0 , T ]. Hence it is enough to show that

m(t) u(t, ) ,
t [t0 , T ].
(1.5.4)
If (1.5.4) is not true, then there exists a t1 (t0 , T ] such that

m(t1 ) = u(t1 , )
and m(t) u(t, ) ,
t [t0 , t1 ].
Consequently, we have
D m(t1 ) u (t1 , )
which, in turn, leads to the contradiction
g(t1 , m(t1 )) D m(t1 ) u (t1 , ) = g(t1 , u(t1 , )) + .
Hence (1.5.4) is true and the proof is complete.
To avoid repetition, we have not so far considered the lower estimate for m(t) which can be obtained by reversing the inequalities.
For later reference, we shall state the following result which yields a
lower bound for m(t).
Theorem 1.5.3 Let g C[R+ R+ , R] and (t) be the minimal
solution of (1.5.1) existing on [t0 , ). Suppose that m C[R+ , R+ ]
and Dm(t) g(t, m(t)), t t0 , where D is any xed Dini derivative.
Then m(t0 ) u0 implies m(t) (t), t t0 .
Proof The proof runs parallel to the proof of Theorem 1.5.2.
Instead of (1.5.3), we now have to consider solutions v(t, ) of
v = g(t, v) ,
v(t0 ) = u0 ,
for suciently small > 0 on [t0 , T ] and note that lim v(t, ) = (t)
0
uniformly on [t0 , T ]. Then it is enough to show that

m(t) > v(t, ) ,
to complete the proof.
t [t0 , T ]
Differential inequalities
29
An extension of Theorem 1.5.2 to systems requires g to be quasimonotone nondecreasing which is also a necessary condition for the
existence of extremal solutions of (1.5.1). Thus, we have the following
generalization of Theorem 1.5.2.
n , Rn ], g(t) be quasimonoTheorem 1.5.4 Let g C[R+ R+
tone nondecreasing in u for each t and r(t) be the maximal solution of (1.5.1) existing on [t0 , ). Suppose that Dm(t) g(t, m(t)),
t t0 holds for a xed Dini derivative. Then m(t0 ) u0 implies
m(t) r(t), t t0 .
As mentioned earlier, the inequalities in Theorem 1.5.4 are componentwise. Instead of considering componentwise inequalities between
vectors, we could utilize the notion of a cone to induce partial order
on Rn and prove Theorem 1.5.4 in that framefork. Naturally, this
approach is more general and is useful when we deal with conevalued
functions. We shall therefore develop theory of dierential inequalities in arbitrary cones and prove a result corresponding to Theorem
1.5.4 in that setting.
A proper subset K Rn is called a cone if the following properties
hold:

K K , 0 , K + K K , K = K ,
(1.5.5)
K {K} = {0} and
K 0 = ,
where K denotes the closure of K, K 0 is the interior of K. We shall
denote by K the boundary of K. The cone K induces the order
relations on Rn dened by
K
xy
yxK
and x < y
y x K 0.
(1.5.6)
The set K dened by K = { Rn : (x) 0 for all x K},

where (x) denotes the scalar product , x , is called the adjoint
cone and satises the properties (1.5.5). We note that K = (K ) ,
x K 0 if (x) > 0 for all K0 and x K if (x) = 0 for some
K0 where K0 = K 0.
We can now dene quasimonotone property of a function relative
to the cone K. A function F C[Rn , Rn ] is said to be quasimonotone
30
Inequalities
K
nondecreasing relative to K if x y and (xy) = 0 for some K0

imply (f (x) f (y)) 0. If f is linear, that is, f (x) = Ax where
A is an n by n matrix, the quasimonotone property of f means the
following: x 0 and (x) = 0 for some K0 imply (Ax) 0.
n , the quasimonotonicity of f reduces precisely to what we
If K = R+
dened before.
In the general setting of a cone, Theorem 1.5.1 is true. In fact, one
can prove existence of extremal solutions for dierential equations in
a Banach space, details of which may be found in Lakshmikantham
and Leela [2]. We shall now prove the comparison result relative to
the cone K.
Theorem 1.5.5 Let g C[R+ Rn , Rn ] and g(t, u) be quasimonotone nondecreasing in u relative to the cone K for each t R+ .
Let r(t) be the maximal solution of (1.5.1) relative to K existing on
[t0 , ) and for t t0 ,
K
D m(t) g(t, m(t)) ,
(1.5.7)
where m C[R+ , K]. Then, m(t0 ) u0 implies

K
m(t) r(t) ,
t t0 .
(1.5.8)
Proof We proceed as in the proof of Theorem 1.5.2 recalling that

the inequalities are now relative to the cone K. It is enough to show
that
K
(1.5.9)
m(t) < u(t, ) , t [t0 , T ].
If (1.5.9) is not true, there exists a t1 (t0 , T ] such that
u(t1 , ) m(t1 ) Kandu(t, ) m(t) K 0
for all t [t0 , t1 ). This means that there exists a K0 such that
(u(t1 , ) m(t1 )) = 0.
The quasimonotonicity of g now yields
{g(t1 , u(t1 , )) g(t1 , m(t1 ))} 0.
31
Integral inequalities
Set w(t) = (u(t, ) m(t)), t [t0 , t1 ) and w(t1 ) = 0. Hence

D w(t1 ) = lim
h0
1
[w(t1 + h) w(t1 )] 0
h
and as a result, we get

D w(t1 ) = (u (t1 , ) D m(t1 ))
> (g(t1 , u(t1 , )) g(t1 , m(t1 ))) 0,
which is a contradiction. Hence the proof is complete.
We observe that quasimonotonicity of g(t, u) relative to a cone P
need not imply the quasimonotonicity of g(t, u) relative to cone Q
whenever P Q. However, the order relations relative to P imply
the same order relations relative to Q if P Q. From this observation results the following corollary which is useful in applications.
Corollary 1.5.1 Let P, Q be two cones in Rn such that P Q.
Let the assumptions of Theorem 1.5.5 hold with K P . Then
P
m(t0 ) u0 implies m(t) r(t), t t0 .
1.6
Integral Inequalities
A result which includes Gronwall and Bihari type inequalities is the

following.
2 , R ], g(t, u) be nondecreasing in
Theorem 1.6.1 Let g C[R+
+
u for each t R+ and r(t) be the maximal solution of
u = g(t, u) ,
u(t0 ) = u0 ,
(1.6.1)
existing on [t0 , ). Suppose that m C[R+ , R+ ] and satises

t
m(t) m(t0 ) +
g(s, m(s)) ds,
t t0 .
t0
Then, m(t) r(t) for all t t0 , provided m(t0 ) u0 .
(1.6.2)
32
Inequalities
Proof Set m(t0 ) +
t
g(s, m(s)) ds = v(t) so that m(t) v(t),
t0
m(t0 ) = v(t0 ) and v g(t, v), in view of the fact that g is nondecreasing in u. An application of Theorem 1.5.2 yields v(t) r(t),
t t0 and this completes the proof.
Corollary 1.6.1 Let the assumptions of Theorem 1.6.1 hold
except that (1.6.2) is replaced by
t
m(t) n(t) +
g(s, m(s)) ds,
t t0 ,
t0
where n C[R+ , R+ ]. Then, we have

m(t) n(t) + r(t),
t t0
where r(t) is the maximal solution of

u = g(t, n(t) + u) ,
u(t0 ) = 0.
A more general result than Theorem 1.6.1 deals with Volterra integral inequality which, in general, cannot be reduced to a dierential
inequality. To prove such a result, we consider the integral equation
of Volterra type given by
t
u(t) = h(t) +
K(t, s, u(s)) ds,
(1.6.3)
t0
2 R, R]. One can prove the exwhere h C[R+ , R] and K C[R+
istence of extremal solutions of (1.6.3) employing arguments similar
to those in case of ordinary dierential equations. However, we do
need monotone nondecreasing property of K(t, s, u) with respect to
u. We shall merely state the result.
Theorem 1.6.2 Let m C[R+ , R+ ], h C[R+ , R+ ], K

3 , R ] and K(t, s, u) be nondecreasing in u for each (t, s), such
C[R+
+
that for t t0 ,
t
m(t) h(t) +
K(t, s, m(s)) ds
t0
(1.6.4)
33
Integral inequalities
holds. Assume that r(t) is the maximal solution of (1.6.3) existing

on [t0 , ). Then m(t) r(t), t t0 .
Proof Let t0 < T < and let u(t, ) be a solution of
t
u(t, ) = h(t) + +
K(t, s, u(s, )) ds,
t [t0 , T ],
t0
for suciently small > 0. Then, since lim u(t, ) = r(t) uniformly
0
on [t0 , T ], it is sucient to show that

t [t0 , T ].
m(t) < u(t, ),
If this is not true, since m(t0 ) < u(t0 , ), there exists a t1 > t0 such
that m(t1 ) = u(t1 , ) and m(t) u(t, ), t [t0 , t1 ]. Using the
monotone property of K, we are led to the contradiction
t1
m(t1 ) h(t1 ) +
K(t1 , s, m(s)) ds < h(t1 )

t0
t1
K(t1 , s, u(s, )) ds = u(t1 , )
+ +
t0
and the theorem is proved.

Corollary 1.6.2 If, in Theorem 1.6.2, K(t, s, u) = K(t, s)u, then
m(t) r(t), t t0 , where r(t) is the solution of (1.6.3) satisfying
t
r(t) = h(t)
R(t, s)h(s) ds,

t0
R(t, s) being the resolvent kernel given by

t
R(t, s) = K(t, s) +
R(t, )K(, s)d.

s
34
Inequalities
One can reduce the integral inequality (1.6.4) to an integrodierential inequality if h and K are smooth enough. Our next result
deals with this situation.
Theorem 1.6.3 In addition to the assumptions on m, h and K
in Theorem 1.6.2, suppose that h (t), Kt (t, s, u) exist, are continuous, nonnegative and Kt (t, s, u) is nondecreasing in u for each (t, s).
Suppose that r(t) is the maximal solution of the dierential equation
t
u (t) = h (t) + K(t, t, u(t)) +
K(t, s, u(t)) ds,
u(t0 ) = h(t0 )
t0
existing on [t0 , ). Then, m(t) r(t), t t0 .

Proof Set v(t) = h(t) +
t
K(t, s, m(s)) ds so that
t0
v (t) = h (t) + K(t, t, m(t)) +
t
Kt (t, s, m(s)) ds.
t0
In view of the assumption, we note that v(t) is nondecreasing and

hence
t
v (t) h (t) + K(t, t, v(t)) +
Kt (t, s, v(t)) ds,
v(t0 ) = h(t0 ).
t0
Now, by Theorem 1.5.2, we obtain the desired conclusion.

Theorem 1.6.3 shows that it is fruitful, in some situations, to consider integro-dierential inequalities rather than integral inequalities.
We shall therefore discuss the theory of integro-dierential inequalities in a later section.
Note that pseudo-linear representations of some nonlinear inequalities enable one to obtain new estimates for their solutions. The
following theorem is the illustration of the idea.
35
General integral inequalities
Theorem 1.6.4 Let m C(R+ R+ ), f, g C(R+ , R+ ) and

for all t [a, b]
t
m(t) c +
(f (s)m(s) + g(s)m (s)) ds,
(1.6.5)
where c 0 and > 1. If for all (t, s) [a, b]

t
g(s) exp ( 1)
1 ( 1)c1
s
f ( ) d ds > 0,
a
then

c exp
t
a
m(t)
1 ( 1)c1
t

f (s)ds

g(s) exp ( 1)
4
1
(1)
f ( )d ds
(1.6.6)
for all t [a, b].
Proof Inequalities (1.6.5) for all t [a, b] can be written in the
pseudo linear form
t
m(t) c +
(f (s) + g(s)m1 (s)) m(s) ds.

a
Applying GronwallBellman lemma (see Theorem 1.1.1) to this inequality and having performed some non-complex transformations,
we get estimate (1.6.6).
1.7
General Integral Inequalities
We shall continue, in this section, to consider integral inequalities of

a general kind. The results given below have a broad applicability
in the sense that they contain several special results found in the
literature.
36
Inequalities
Theorem 1.7.1 Assume that

(i) m, h, f C[R+ , R+ ], f (u) is strictly increasing in u and
Dom f 1 = R+ ;
2 , R ], g C[R3 , R ], H(t, u) is nondecreasing in
(ii) H C[R+
+
+
+
u for each t and g(t, s, u) is nondecreasing in (t, u) for each s;
(iii) for t t0
t

f (m(t)) h(t) + H t, g(t, s, m(s)) ds ;
(1.7.1)
t0
(iv) for any T > t0 , r(t, t0 ; T ) is the maximal solution of

u = g(T, t, f 1 (h(t) + H(t, u))),
u(t0 ) = 0
(1.7.2)
existing on [t0 , ).
Then, for t t0 ,
m(t) f 1 (h(t) + H(t, r (t))),
(1.7.3)
where r (t) = r(t, t0 ; t).

Proof We set v(t; t) =
t
g(t, s, m(s)) ds. Since g(t, s, u) is nonde-
t0
creasing in t, for T > t0 , we have

t
v(t; t)
g(T, s, m(s)) ds = v(t; T ),
t0 t T.
(1.7.4)
t0
In view of the assumptions (i) and (ii), the relations (1.7.1) and
(1.7.4) imply that
m(t) f 1 (h(t) + H(t, v(t, t)))
f 1 (h(t) + H(t, v(t, T ))),
t0 t T.
(1.7.5)
)
= g(T, t, m(t)), the monotone nature of g(t, s, u) in u
Since dv(t;T
dt
together with (1.7.5) yields the dierential inequality
dv(t; T )
g(T, t, f 1 (h(t) + H(t, v(t; T )))),
dt
v(t0 , T ) = 0, (1.7.6)
37
General integral inequalities
for t0 t T . Hence, by Theorem 1.5.2, we get

v(t; T ) r(t, t0 ; T ),
t0 t T,
(1.7.7)
where r(t, t0 ; T ) is the maximal solution of (1.7.2). Since r (t) =

r(t, t0 ; t), the relations (1.7.5) and (1.7.7) show that the desired estimate (1.7.3) holds, thus completing the proof.
The inequality (1.7.1) also yields a lower estimate for m(t) provided the monotone properties of the functions involved are suitably
changed. This is the content of the next result.
(i) m, h, f C[R+ , R+ ], f (u) is strictly increasing in u and
Dom f 1 = R+ ;
2 , R ], g C[R3 , R ], H(t, u) is nondecreasing in
(ii) H C[R+
+
+
+
u for each t and g(t, s, u) is nondecreasing in t for each (s, u) and
nondecreasing in u for each (t, s);
(iii) the inequality (1.7.1) holds;
(iv) for any T > t0 , (t, t0 ; T ) is the minimal solution of (1.7.2)
Then, if (t) = (t, t0 ; t), we have
m(t) f 1 (h(t) + H(t, (t))),
Proof Setting v(t; t) =
t
t t0
(1.7.8)
g(t, s, m(s)) ds as before, we now get
t0
v(t; t) v(t; T ),
(1.7.9)
since g is nondecreasing in t. Assumptions (i) and (ii) together with

inequalities (1.7.1) and (1.7.9) yield
m(t) f 1 (h(t)+H(t, v(t; t))) f 1 (h(t)+H(t, v(t; T ))), (1.7.10)
t0 t T , and for t0 t T ,
dv(t; T )
g(T, t, f 1 (h(t) + H(t, v(t; T )))),
dt
v(t0 ; T ) = 0. (1.7.11)
38
Inequalities
Hence, by Theorem 1.5.3, we get

v(t; T ) (t, t0 ; T ),
t0 t T,
(1.7.12)
where (t, t0 ; T ) is the minimal solution of (1.7.2). Since (t) =

(t, t0 ; t), the relations (1.7.10) and (1.7.12) imply the conclusion
(1.7.8) and the proof is complete.
It is easy to observe that, if instead of (1.7.1) we suppose that

t
f (m(t)) h(t) + H t, g(t, s, m(s)) ds ,
t t0 ,
(1.7.13)
t0
then one could state and prove theorems that are duals to Theorems
1.7.1 and 1.7.2 with appropriate monotone conditions on the functions involved. In order to avoid repetition, we omit the details, and
give an example to illustrate the scope of the results developed.
Example 1.7.1 Let f (u) = u, H(t, u) = b(t)u1/p , g(t, s, u) =
K(s)up , so that (1.7.1) reduces to
1/p
t
K(s)(m(s))p ds
,
m(t) h(t) + b(t)
t t0 .
t0
The comparison equation (1.7.2) then becomes

u = K(t)(h(t) + b(t)u1/p )p ,
u(t0 ) = 0,
and thus, we have the estimate

t
1/p
t
p
,
m(t) h(t)+b(t) cp K(s)(h(s)) exp(cp K()(b())p d) ds
t0
for t t0 , where cp = 1 if 0 p < 1 and cp = 2p1 if p 1.
1.8
IntegroDierential Inequalities
We shall begin with the following result concerning integro-dierential inequality.
Integrodifferential inequalities
39
2 , R ], H C[R3 , R ], H(t, s, u)
Theorem 1.8.1 Let f C[R+
+
+
+
be nondecreasing in u for each (t, s) and for t t0 ,
t
D m(t) f (t, m(t)) +
H(t, s, m(s)) ds,
(1.8.1)
t0
where m C[R+ , R+ ]. Suppose that r(t) is the maximal solution of

t
u (t) = f (t, u(t)) +
H(t, s, u(s)) ds,
u(t0 ) = u0 0,
(1.8.2)
t0
existing on [t0 , ). Then m(t) r(t), t t0 , provided m(t0 ) u0 .

For a proof of existence of the maximal solution for (1.8.2) see
Hu, Zhuang and Khavanin [1] and Zhuang [1].
Proof Dene the functions
F (t, u) = f (t, p(t, u)),
K(t, s, u) = H(t, s, p(s, u)),
where p(t, u) = max[m(t), u] and let u(t) be any solution of

t
u (t) = F (t, u(t)) +
K(t, s, u(s)) ds,
u(t0 ) m(t0 ).
t0
We shall show that m(t) u(t) for t t0 . If this is not true, suppose
that for some t > t0 , m(t ) > u(t ). Then, there exists a t1 > t0
such that
u(t1 ) m(t1 ) and u (t1 ) < D + m(t1 ).
By denition of p(t, u), it is clear that m(t) p(t, u(t)) and
m(t1 ) = p(t1 , u(t1 )). Hence, using the monotone character of H,
we have
+
t1
D m(t1 ) f (t1 , m(t1 )) +
H(t1 , s, m(s)) ds
t0
t1
f (t1 , p(t1 , u(t1 ))) +
H(t1 , s, p(s, u(s))) ds

t0
40
Inequalities
= F (t1 , u(t1 )) +
Zt1
t0
K(t1 , s, u(s)) ds u (t1 ),
which is a contradiction. Thus, m(t) u(t) which implies that u(t)

is also a solution of (1.8.2). Since r(t) is the maximal solution of
(1.8.2), it follows that m(t) r(t), t t0 and the proof is complete.
Note that the proof of Theorem 1.8.1 is different from the proofs
of Theorems 1.5.2 and 1.6.3 which do not depend on the manner in
which the maximal solution is obtained. Of course, one could prove
Theorems 1.5.2 and 1.6.3 by the present method.
Corollary 1.8.1 Let (1.8.1) hold with f = a(t)u, H = H(t, s)u.
Then, m(t) R(t, t0 ), t t0 where R(t, s) is the solution of
R(t, s)
+ R(t, s)a(s) +
s
Zt
R(t, )H(, s)d = 0, R(t, t) = 1. (1.8.3)
on the interval t0 s t.
For the proof, it is enough to observe that r(t) = R(t, t0 )u(t0 ) is
the solution of (1.8.2).
Even in the case when f and H are linear, finding R(t, s) is difficult in general. Hence, a comparison result which enables us to
reduce integrodifferential inequalities to differential inequalities will
provide great advantage. This is what we shall consider next. First,
we need the following lemma.
2 , R ] satisfy
Lemma 1.8.1 Let g0 , g C[R+
+
g0 (t, u) g(t, u),
2
(t, u) R+
.
(1.8.4)
Then, the right maximal solution r(t, t0 , u0 ) of

u = g(t, u),
u(t0 ) = u0 0
(1.8.5)
and the left maximal solution (t, T, v0 ) of

u = g0 (t, u),
u(T ) = u0 0
(1.8.6)
41
satisfy the relation

r(t, t0 , u0 ) (t, T, v0 ),
t [t0 , T ],
(1.8.7)
whenever r(T, t0 , u0 ) v0 .
Proof It is known that lim u(t, ) = r(t, t0 , u0 ) and lim v(t, ) =
0
(t, T, v0 ) where u(t, ) is any solution of

u = g(t, u) + ,
u(t0 ) = u0 + ,
existing to the right of t0 and v(t, ) is any solution of

v = g0 (t, v) ,
v(T ) = v0 ,
existing to the left of T and > 0 is suciently small. Note that

(1.8.7) follows if we rst establish the inequality
t0 t < T.
u(t, ) < v(t, ),
Since g0 g and r(T, t0 , u0 ) v0 , it is easy to see that for a suciently small > 0, we have
u(t, ) < v(t, ),
T t < T,
and in particular, u(T , ) < v(T , ). We claim that

u(t, ) < v(t, ),
t0 t T .
If this is not true, there exists a t [t0 , T ) such that

u(t, ) < v(t, ),
t < t T
and u(t , ) < v(t , ).
This leads to the contradiction

g(t , u(t , )) + = u (t , ) v (t , ) = g0 (t , v(t , )) .
Hence u(t, ) < v(t, ),
is complete.
t0 t T and the proof of the lemma
We are now in a position to prove the following comparison result

which plays an important role in the study of integrodierential
inequalities.
42
Inequalities
2 , R], H
Let m C[R+ , R+ ], f C[R+
Theorem 1.8.2
3 , R] and
C[R+
t
D m(t) f (t, m(t)) +
H(t, s, m(t)) ds,
t I0 ,
(1.8.8)
t0
where I0 = {t t0 : m(s) (s, t, m(t)), t0 s t}, (t, T, v0 ) being

the left maximal solution of (1.8.6) existing on [t0 , T ]. Assume that
g0 (t, u) F (t, u; t0 )
where
(1.8.9)
t
F (t, u; t0 ) = f (t, u) +
H(t, s, u) ds
(1.8.10)
t0
and r(t) is the maximal solution of

u = F (t, u; t0 ),
u(t0 ) = u0 ,
(1.8.11)
existing on [t0 , ). Then m(t0 ) u0 implies

m(t) r(t)
t t0 .
(1.8.12)
Proof Since it is known that lim u(t, ) = r(t) where u(t, ) is a

0
solution of
u = F (t, u; t0 ) + ,
u(t0 ) = u0 + ,
for > 0 suciently small, on any compact set [t0 , T ] [t0 , ), it is

enough to prove that
m(t) < u(t, )
t0 t T.
If this is not true, then there exists a t (t0 , T ] such that

m(s) < u(s, )
t0 s < t
and m(t ) = u(t , ).
This implies that

D m(t ) u (t , ) = F (t , u(t , ); t0 ) + .
(1.8.13)
43
Consider now the left maximal solution (s, t , m(t )), t0 s t ,

of
u = g0 (t, u), u(t ) = m(t ).
By Lemma 1.8.1,
r(s, t0 , u0 ) (s, t , m(t )),
t0 s t .
Since r(t , t0 , u0 ) = lim u(t , ) = m(t ) = (t , t , m(t )) and

0
m(s) u(s, ), t0 s t , it follows that

m(s) r(s, t0 , u0 ) (s, t , m(t )),
t0 s t .
This inequality implies that t I0 and as a result (1.8.6) yields

D m(t ) F (t , m(t ); t0 )
which contradicts (1.8.13). Thus, m(t) r(t) for t t0 and the
proof is complete.
In Theorem 1.8.2, we could choose g0 = f if H 0.
We shall use comparison Theorem 1.8.2 to study integro-dierential inequality (1.8.1). The advantage of this approach will be
clear when we consider some special cases.
Theorem 1.8.3 Let (1.8.1) hold with m, f and H satisfying the
2 , R] be such
assumption of Theorem 1.8.1. Suppose that g0 C[R+
that g0 F , where
t
F (t, u; t0 ) = f (t, u) +
K(t, s, u) ds,
t0
and
K(t, s, u) = H(t, s, (s, t, u)),
where (s, T, v0 ) is the left maximal solution of (1.8.6) existing on
t0 t T . Then, m(t) r(t), t t0 , whenever m(t0 ) u0 , where
r(t) is the maximal solution of (1.8.11) existing on [t0 , ).
44
Inequalities
Proof We set I0 = {t > t0 : m(s) (s, t, m(t)), t0 s t}.

Since H is nondecreasing in u, (1.8.1) yields the inequality
t
D m(t) f (t, m(t)) +
H(t, s, (s, t, m(t))) ds,
t I0 .
t0
Hence, by Theorem 1.8.2, we have the stated result.

Some of the interesting special cases of Theorem 1.8.3 are given
below as corollaries.
Corollary 1.8.2 Consider the special case when f 0 and
H 0 in Theorem 1.8.3. Then m(t) r(t), t t0 , where r(t) is the
maximal solution of
t
u (t) = f (t, u(t)) +
H(t, s, u(t)) ds,
u(t0 ) = u0 ,
t0
existing on [t0 , ), provided m(t0 ) u0 . If, in particular,

f (t, u) = a(t)u, H(t, s, u) = H(t, s)u with a(t) 0, H(t, s) 0,
then
t

A(s, t0 ) ds , t t0 ,
m(t) m(t0 ) exp
t0
where A(t, t0 ) = a(t) +
t
H(t, s) ds.
t0
Corollary 1.8.3 Consider the special case with f = g0 and

H 0 in Theorem 1.8.3. Then m(t) r(t), t t0 , where r(t) is
the maximal solution of (1.8.11) existing on [t0 , ). If, in particular,
f = g0 = u, > 0 and H(t, s, u) = H(t, s)u with H(t, s) 0,
then
t

B(s, t0 ) ds , t t0 ,
m(t) m(t0 ) exp
t0
where B(t, t0 ) = +
t
t0
H(t, s) exp((t s)) ds.
45
Corollary 1.8.4 Consider the special case with f = g0 = c(u)

and H 0, where c(u) is continuous, nondecreasing in u and
c(0) = 0. Then m(t) r(t), t t0 , where r(t) is the maximal
solution of
t
u (t) = f (t, u(t)) +
H(t, s, J 1 (J(u(t)) (t s))) ds,
u(t0 ) = u0
t0
existing on [t0 , ), provided m(t0 ) u0 . Here J(u) =

J 1 is the inverse of J.
u
u0
ds
c(s)
and
Finally, we give a useful corollary that is a special case of Theorem

1.6.3.
Corollary 1.8.5 Consider the integral inequality
t
m(t) h(t) +
K(t, s)m(s) ds,
t t0 .
t0
2 , R ] and K (t, s)
where m C[R+ , R+ ], h C[R+ , R+ ], K C[R+
+
t
exists, is continuous and nonnegative. Then,
t

t
m(t) h(t) +
(s, t0 ) exp
d(, t0 ) d
t t0 ,
t0
where d(t, t0 ) = K(t, t) +
ds,
t
Kt (t, s) ds and
t0
t
(t, t0 ) = K(t, t)h(t) +
Kt (t, s)h(s) ds.

t0
Note that Corollary 1.8.5. is precisely Theorem 1.1.4, which was

proved by using the monotone character of
t
v(t) = h(t) +
K(t, s)m(s) ds.

t0
46
Inequalities
Here, we see a broader principle involved in the proof of which

the earlier trick is embedded as a special case.
Observe that if H in (1.8.1) is nondecreasing in u, then the estimate on m(t) becomes rather simple. For, using the fact m(t) 0,
we arrive at
t
D m(t) f (t, m(t)) +
H(t, s, 0) ds,
t t0 ,
t0
which yields m(t) r(t), t t0 , where r(t) is the maximal solution

of
t

u (t) = f (t, u) + H(t, s, 0) ds, u(t0 ) m(t0 ),
t0
1.9
Dierence Inequalities
One of the ecient methods of obtaining information on the behavior

of solutions of dierence equations, even when they cannot be solved
explicitly, is to estimate the function satisfying a dierence inequality
by the solution of the corresponding dierence equation. In this
section, we present the theory of dierence (or discrete) inequalities.
We begin with the following general result.
Theorem 1.9.1 Let n Nn+0 = {n0 , n0 + 1, . . . , n0 + k, . . . }, g be
dened on Nn+0 R+ and g(n, u) be nondecreasing in u for any xed
n. Suppose that, for n n0 , the inequalities
yn+1 g(n, yn )
un+1 g(n, un )
(1.9.1)
n n0 .
(1.9.2)
hold. Then, yn0 un0 implies

y n un
for all
47
Difference inequalities
Proof Suppose that (1.9.2) is not true. Then, because of the fact
yn0 un0 , there exists a k Nn+0 such that yk uk and yk+1 > uk+1 .
It then follows that, using (1.9.1) and the monotone character of g,
g(k, uk ) uk+1 < yk+1 g(k, yk ) g(k, uk )
which is a contradiction. Hence the theorem is provided.
Normally, when applying Theorem 1.9.1, one of the relations in
(1.9.2) is an equation and correspondingly yn (or un ) is the solution
of the difference equation. Let us now consider the simplest difference
equation given by
y(x) = g(x)
(1.9.3)
where y(x) = y(x + 1) y(x), y, g : Jx+ R and Jx+ =
{x, x + 1, . . . , x + k, . . . }. Then the solution of (1.9.3) is a function
of y such that (1.9.3) is satisfied and we denote it by
y(x) = 1 g(x),
where 1 is called the antidifference operator. This solution y(t)
is not unique since y(x) + w(x), w(x) being an arbitrary function of
period one, is also a solution of (1.9.3). It is also easy to compute
that
n
X
i=0
g(x + i) = y(x + n + 1) y(x) y(x + i)|i=n+1

.
i=0
(1.9.4)
In case Jx+ = N + = {0, 1, 2, . . . , k, . . . }, this formula becomes

n
X
i=0
yi = 1 yi |i=n+1
.
i=0
(1.9.5)
This information can be used to solve the linear difference equation

z(x + 1) = p(x)z(x) + q(x) z(x0 ) = z0 .
(1.9.6)
Setting P (x) =
x1
Q
t=x0
p(t), p(x0 ) = 1, we get

z(x + 1)
z(x)
q(x)
=
+
.
p(x + 1)
p(x) p(x + 1)
48
Inequalities
q(x)
If z(x)
p(x) = y(x), and p(x+1) = g(x), the equation (1.9.6) now assumes
the form (1.9.3). Consequently, the solution of (1.9.6) is
" x1
#
X q(s)
1
z(x) = P (x)[ g(x) + z0 ] = P (x)
+ z0
(1.9.7)
p(s + 1)
s=x
0
x1
X
q(x)
s=x0
x1
Y
p(t)
t=s+1
+ z0
x1
Y
p(t).
t=x0
We are now in a position to deduce from Theorem 1.9.1 the discrete version of Gronwall inequality.
Corollary 1.9.1 Let n Nn+0 , kn 0 and
yn+1 yn0 +
Then, for all n n0 ,
yn yn0 exp
n1
X
ks
s=n0
n
X
(ks ys + ps ).
s=n0
n1
X
n1
X
ps exp
s=n0
=s+1
k .
Proof We shall apply Theorem 1.9.1. The comparison equation

is
un = un 0 +
n1
X
(ks us + ps ),
un0 = yn0 .
s=n0
This is equivalent to
un = kn un + pn ,
the solution of which, based on (1.9.7), is
un = un0
n1
Y
s=n0
(1 + ks ) +
n1
X
s=n0
ps
n1
Y
(1 + k ) .
=s+1
The proof is complete by observing that 1 + ks exp(ks ).

Next, we shall consider the discrete version of Bihari inequality .
49
Theorem 1.9.2 Suppose that yn , hn are nonnegative functions

dened on Nn+0 and W is a positive, strictly increasing function dened on R+ . If for n n0 , yn Vn , where
Vn = y 0 + M
n1

hs W (ys ),
s=n0
then, for n N1 ,

1
yn G
G(y0 ) + M
n1

hs ,
s=n0
where G is the solution of

G(Vn ) =
and N1 = {n Nn+0 : M
n1

s=n0
Vn
W (Vn )
hs G() G(x0 )}.
Proof We have Vn = M hn W (yn ) M hn W (Vn ). It then follows

that
G(Vn+1 ) G(Vn ) + M hn ,
from which we get, in view of Theorem 1.9.1,
G(Vn ) G(y0 ) + M
n1

hs .
s=n0

n1

hs .
Hence, for n N1 , Vn G1 G(y0 ) + M
s=n0
Theorem 1.9.3 Let an , bn be two nonnegative functions dened

on Nn+0 and Pn be a positive, nondecreasing function dened on N + .
Let
n1
n1

n1

yn Pn +
as ys +
as
bk yk .
s=n0
s=n0
s=n0
50
Inequalities
Proof Since Ps is positive and nondecreasing, we obtain

n1
n1
X ys
X n1
X yk
yn
1+
as
+
as
bk
,
Pn
Ps s=n
Pk
s=n
0
un0 = 1,
k=n0
so that we have
un = an
hy
Pn
n1
X
bk
k=n0
yk i
.
Pk
This in turn yields

n1
h
i
X
un an un +
bk uk .
k=n0
By setting xn = un +
n1
P
bk uk , we obtain
k=n0
xn = un + bn un an xn + bn un (an + bn )un ,
form which one gets xn+1 (1 + an + bn )un , and therefore,
xn
n1
Y
(1 + ak + bk ).
k=n0
Consequently we arrive at
un an
n1
Y
(1 + ak + bk ),
k=n0
which implies
un 1 +
n1
X
s=n0
This in turn yields yn Pn un .
as
n1
X
k=n0
(1 + ak + bk ).
51
Theorem 1.9.4 Let yj (j = 0, 1, . . . ) be a positive sequence satisfying

n2
n1

yn+1 g yn ,
yj ,
yj ,
j=0
j=0
where g(y, z, w) is a nondecreasing function with respect to its arguments. If y0 u0 and

n1
n2

un+1 = g un ,
uj ,
uj ,
j=0
then, for all n 0
j=0
yn un .
Proof The proof is by induction. The claim is true for n = 0.

Suppose it is true for n = k. Then, we have
yk+1 g yk ,
k1

j=0
yj ,
k2

yj g uk ,
j=0
k1

uj ,
j=0
k2

uj = uk+1 .
j=0
If we take uk = tk+1 tk = tk , t0 = 0, the comparison equation

becomes
tk = g(tk , tk , tk1 ).
Theorem 1.9.5 Let yi (i = 0, 1, . . . ) be a positive sequence satisfying the inequality
yn+1 g(yn , yn1 , . . . , ynk )
where g is nondecreasing with respect to its arguments. Then,
y n un
where un is the solution of
un+1 = g(un , un1 , . . . , unk ),
yj uj ,
j = 0, 1, . . . , k.
52
Inequalities
Proof Suppose that the conclusion is not true. Then there exists
an index m k such that ym+1 > um+1 and yy uj , j m. It then
follows that
g(ym , ym1 , . . . , ymk ) ym+1 > um+1 = g(um , um1 , . . . , umk ),
which is a contradiction.
1.10
Interval-Valued Integral Inequalities
By an interval we mean a closed, bounded set of real numbers and

if Y is an interval, we shall denote its end points by y and y so that
Y = [y, y]. If Y = (Y1 , . . . , Yn ) is an interval vector, we denote the
vector norm by
kY k = max(|Y1 |, . . . , |Yn |),
where |Yi | = max(|y|, |y|) for each i = 1, 2, . . . , n. Accordingly, the
norm of an interval matrix is given by
X
kAk = max
|Aij |.
i
Furthermore, it is easy to see that [a, b] [c, d] if and only if

c a b d.
Let f be an arbitrary mapping. Define the interval function
f (Y ) = {f (y) : y Y }, where Y is an interval. Then, for any
interval Z such that Z Y we have f (Z) f (Y ). Note that even
though f is not assumed to be monotonic, the interval function is
inclusion monotonic. This intrinsic advantage of interval functions
is exploited in interval analysis.
Interval valued functions have interval integrals and interval integration preserves set inclusion. Thus,
Z(t) Y (t),
0ta
implies
Zt
0
Z(s) ds
Zt
0
Y (s) ds.
Interval-valued integral inequalities
53
We note also that if Y (t) = [y(t), y(t)], then the interval integral of
Y (t) is an interval between the lower Darboux integral of y(t) and
the upper Darboux integral of y(t).
A basic result which uses the inclusion monotone property of interval mappings is the following theorem. Before stating the theorem,
let us introduce some convenient notation.
Let BIn [t0 , t0 + a] denote the set of all bounded ndimensional
interval vector-valued functions on [t0 , t0 + a]. Suppose that a > 0
and U, H BIn [t0 , t0 + a] with
B < b1 U (t), H(t) b2 < B
for all
t [t0 , t0 + a],
(1.10.1)
where B = [B, B] is an ndimensional interval vector. Let be

a qdimensional interval vector of parameter values. Suppose that
F : [t0 , t0 +a]2 B Rn is a bounded real vector-valued function
and M is an ndimensional interval vector such that
F ([t0 , t0 + a], [t0 , t0 + a], B, ) M.
(1.10.2)
Now, define
ph, [u](t) = h(t) +
Zt
F (t, s, u(s), ) ds,
(1.10.3)
t0
where h(t) H(t), , u(t) B, t [t0 , t0 + a] and

P [Y ](t) = H(t) +
Zt
F (t, s, Y (s), ) ds,
(1.10.4)
t0
for Y (t) B, t [t0 , t0 + a]. Also, define P i [Y ] recursively by

P 0 [Y ] = Y and P i [Y ] = P [P i1 [Y ]] for i = 1, 2, . . .
Theorem 1.10.1 Let U, H BIn [t0 , t0 + a] and F Rn be
functions such that (1.10.1) and (1.10.2) are satisfied and (1.10.3),
(1.10.4) are defined. Suppose that
U (t) P [U ](t),
for all
t [t0 , t0 + a].
(1.10.5)
54
Inequalities
Then there exists > 0 such that

P [B](t) B,
for all
t [t0 , t0 + ].
(1.10.6)
P i [U ] P i [B]
for all
i, j = 0, 1, 2, . . .
(1.10.7)
Moreover,
and there exist U and U BIn [t0 , t0 + a] such that

U (t) = lim P i [B](t)
i
and
U (t) = lim P i [U ](t),

i
(1.10.8)
and
U U .
(1.10.9)
Finally, U contains the set of all fixed points such that u(t) B,
t [t0 , t0 + a] of (1.10.3) and if the set of all fixed points of (1.10.3)
contains U and is an element of BIn [t0 , t0 + a], then
P i [U ] U {all fixed points in
B} U P j [B],
(1.10.10)
for all i, j = 0, 1, 2, . . .
Proof The existence of a positive such that (1.10.6) holds follows
from the relation
P [B](t) H(t) + M (t t0 ).
(1.10.11)
In view of (1.10.1) and (1.10.11), we get (1.10.6) if

H(t) + M (t t0 ) B
t [t0 , t0 + ],
holds. Thus, for , we can choose the supremum of t such that

(b1 )k (B)k M k (t t0 ),
(B)k (b2 )k M k (t t0 )
are satisfied for all k = 1, 2, . . . , n.

The relations given in (1.10.7) follows from the set inclusion property, i.e. Y Z implies P [Y ] P [Z]. Using (1.10.5) and (1.10.6),
we have
P [B] B
implying P i [B] P j [B] for
ij
Interval-valued integral inequalities
55
and
U P [U ] implying P j [U ] P i [U ]
for i j.
Thus, since U B, we obtain

P j [U ] P i [U ] P i [B] P j [B]
and (1.10.7) is proved.
The pointwise limits implied by (1.10.8) exist, since these represent pairs of sequences of real numbers, when considered componentwise. Each of these is either a monotonic nondecreasing sequence of
real numbers bounded above or a monotonic nonincreasing sequence
of real numbers bounded below. In either case, there is a unique
limit for each sequence and we obtain the end points of the intervals
involved. The relation (1.10.9) follows from (1.10.7) and (1.10.8).
If u(t) is a fixed point of (1.10.3) for some h H and some
and if u(t) B for all t [t0 , t0 + a], then it follows from u = ph, [u]
that u P [B]. Furthermore,
u = pih, [u] P i [B] for all
i = 0, 1, 2, . . . ,
and thus u U .
If the set of all such fixed points is an element of BIn [t0 , t0 + a]
and if it contains U , then it also contains P i [U ] and U as well by
inclusion argument as before. We can summarize these results by
(1.10.10) and thus the proof is complete.
This theorem is an extension of Theorem 1.7.1 and 1.7.2 to interval mappings and contains their dual results which are obtained by
reversing the inequalities.
The following corollary of Theorem 1.10.1 is a generalization of
Gronwall inequality to interval maps.
Corollary 1.10.1 Let Z be an interval vector valued function
which is bounded, i.e. kZ(t)k b for t [t0 , t0 + a]. Suppose that A
is an interval vector and M is an interval matrix. If
Z(t)
Zt
t0
(M Z(s) + A) ds,
t [t0 , t0 + a]
56
Inequalities
then
Z(t)
M iA
i=0
1.11
(t t0 )i+1
,
(i + 1)!
t [t0 , t0 + a].
Inequalities for Piecewise Continuous Functions
Let us begin by proving a Gronwalltype inequality.

(i) for each t t0 , the function m(t) is nonnegative and piecewise continuous with discontinuities of the rst type at which m(t) is
continuous from the left;
(ii) for each t t0 , the function (t) is nonnegative and continuous;
(iii) the sequence {ti }, i = 1, 2, . . . , satises the condition
t0 < t1 < t2 < . . . , lim ti = ;
i
(iv) for t t0 , the following inequality

t
m(t) m(t0 ) +
(s)m(s) ds +
i m(ti )
(1.11.1)
t0 <ti <t
t0
holds, where i 0 are constants. Then,

m(t) m(t0 )
t
(1 + i ) exp
t0 <ti <t

(s) ds ,
t t0 .
t0
Proof Let t [t0 , t1 ]. Then, by Theorem 1.1.1, we have

t
m(t) m(t0 ) exp

(s) ds ,
t0
t [t0 , t1 ]
(1.11.2)
57
Inequalities for piecewise continuous functions
Now, let t (t1 , t2 ]. Then, we get from (1.11.1),

t
m(t) m(t0 ) +
(s)m(s) ds + 1 m(t1 )
t0
s
t1
m(t0 ) +
(s)m(t0 ) exp

()d
t0
t0
t1
t
(s)m(s) ds + 1 m(t0 ) exp
ds

(s) ds
t1
t0
t1
= m(t0 )(1 + 1 ) exp

(s) ds
t
+
t0
(s)m(s) ds.
t1
Hence it follows that

t1
m(t) m(t0 )(1 + 1 ) exp
t
(s) ds exp

(s) ds
t0
t1
t
= m(t0 )(1 + 1 ) exp

(s) ds .
t0
Using induction argument, it is easy to verify that inequality

(1.11.2) holds for t (ti , ti+1 ], i = 1, 2, . . . . The proof is therefore
complete.
We shall next consider a dierential inequality result for a right
continuous function m(t) with a countable set of discontinuities of
the rst type.
Theorem 1.11.2 Let m(t) 0 be right continuous on [t0 , )
with jump discontinuities at ti , i = 1, 2, . . . ,, ti > t0 such that
|m(ti ) m(t
i )| i ,
58
Inequalities
where

i=1
2 , R] and g(t, u) be nondei is convergent. Let g C[R+
creasing in u for each t and for any Dini derivative,

Dm(t) g(t, m(t)),
t [ti , ti+1 ),
i = 0, 1, 2, . . .
Then, m(t0 ) u0 implies that

m(t) r(t, t0 , u0 + ),
where =

i=1
t t0 ,
i and r(t, t0 , u0 ) is the maximal solution of

u = g(t, u),
u(t0 ) = u0 0,
(1.11.3)
Proof By Theorem 1.5.2, we have
m(t) r0 (t),
t [t0 , t1 ),
m(t) r1 (t),
t [t1 , t2 ),
(1.11.4)
and
where r0 (t) and r1 (t) are the maximal solutions of
u = g(t, u)
(1.11.5)
starting at (t0 , m(t0 )) and (t1 , m(t1 )) respectively. Since m(t1 )

m(t
1 ) + 1 and r1 (t) g(t, r1 (t)), by Theorem 1.5.2, it follows that
m(t) r11 (t),
t [t1 , t2 ),
where r11 (t) is the maximal solution of (1.11.5) through
(t1 , m(t
1 ) + 1 ). By (1.11.4), m(t1 ) r0 (t1 ) and therefore, again
applying Theorem 1.5.2, we get
m(t) r12 (t),
t [t1 , t2 ),
where r12 (t) is the maximal solution of (1.11.5) through

(t1 , r0 (t1 ) + 1 ). Dene a function 0 (t) as follows:

r0 (t) + 1 , t [t0 , t1 ]
0 (t) =
t [t1 , t2 ).
r12 (t),
Inequalities for piecewise continuous functions
59
0 (t) is well dened and now, by the monotonicity of g(t, u), we have
0 (t) = r0 (t) = g(t, r0 (t)) g(t, r0 (t)+1 ) = g(t, 0 (t)),
t [t0 , t1 ],
and

(t) = g(t, r12 (t)) = g(t, 0 (t)),
0 (t) = r12
t [t1 , t2 ).
Hence
0 (t) g(t, 0 (t)),
t [t0 , t2 ),
which yields, by Theorem 1.5.2,

0 (t) R0 (t),
t [t0 , t2 ),
where R0 (t) is the maximal solution of (1.11.5) through

(t0 , m(t0 ) + 1 ). Clearly, m(t) 0 (t) R0 (t), t [t0 , t2 ). Proceeding in the same way and repeating the arguments as given above, we
obtain
m(t) R1 (t), t [t0 , t3 ),
where R1 (r) is the maximal solution of (1.11.5) through
(t0 , m(t0 ) + 1 + 2 ), and it therefore follows, repeating those arguments successively,

m(t) r t, t0 , m(t0 ) +
i ,
t t0 ,
i=1
thus completing the proof of theorem.

Finally we shall prove a dierential inequality result for dierential
equations with impulsive eect.
Consider the dierential equation with impulsive eect given by

u (t) = g(t, u(t)), t = k ,
(1.11.6)
u(k+ ) = k (u(k )), k = 1, 2, . . .
u(t+
0 ) = u0 ,
2 , R], : R R and t < < < < . . . ,
where g C[R+
+
0
1
k
k
with k as k . Using Theorem 1.5.2, we can prove the
following result.
60
Inequalities
2 , R], : R R and
Theorem 1.11.3 Assume that g C[R+
+
k
k (u) is nondecreasing in u, m : R+ R+ is continuous for t = k ,
left continuous at k and satises
Dm(t) g(t, m(t)),
t = k ,
t > t0 ,
with
m(k+ ) k (m(k )),
k = 1, 2, . . . ,
where D is any xed Dini derivative. Let r(t) be the maximal solution of (1.11.6) existing on [t0 , ). Then m(t+
0 ) u0 implies that
m(t) r(t), t t0 .
Proof By Theorem 1.5.2, we have m(t) r(t), t (t0 , 1 ] and
hence m(1 ) r(1 ). Consequently,
m(1+ ) 1 (m(1 )) 1 (r(1 )) = r(1+ ).
Similar arguments show that
for t (k , k+1 ] [t0 , ),
m(t) r(t),
k = 1, 2, . . . and therefore the desired result follows.

Corollary 1.11.1 If in Theorem 1.11.3, g(t, u) = p(t)u + q(t),
where p and q are continuous on R+ and k (u) = dk u + bk , bk , dk 0
are constants, then for t t0 ,
m(t) m(t0 )
t
dk exp
t0 <k <t
t0
dk exp
s<k <t

t0 <k <t
k <i <t
p(s) ds
t0
t
t

p()d q(s) ds
t
di exp
p(s) ds
k

bk .
Reactiondiffusion inequalities
1.12
61
Reaction-Diusion Inequalities
Let be a bounded domain in Rn and let H = (t0 , ) , t0 0.

Suppose that the boundary H of H is split into two parts H0 , H1
such that H = H0 H1 , {t0 } H0 and H0 H1 is
empty. A vector is said to be an outer normal at (t, x) H1 if
(t, x h) H for small h > 0. The outer normal derivative is then
given by
u(t, x) u(t, x h)
u(t, x)
= lim
.
h0
h
We shall always assume that an outer normal exists on H1 and the
functions in question have outer normal derivatives on H1 .
If u C[H, R] and if the partial derivatives ut , ux , uxx exist and
are continuous in H, then we shall say that u C. The following
lemma is basic to our discussion.
Lemma 1.12.1 Assume that
(i) m C and m(t, x) < 0 on H0 ;
(ii)
m(t,x)
< 0 on H1 , where is the outer normal on H1 ;
(iii) m(t, x) is such that if (t1 , x1 ) H, m(t1 , x1 ) = 0, mx (t1 , x1 ) =

n

i j mxi xj (t1 , x1 ) 0, Rn ,
0 and the quadratic form
i,j=1
then mt (t1 , x1 ) < 0.

Under these conditions, we have m(t, x) < 0 on H.
Proof If the conclusion is not true, there would exist a t1 > t0 ,
x1 such that
m(t, x) < 0
on
[t0 , t1 )
and m(t1 , x1 ) = 0.
It is easy to see that m(t1 , x) has its maximum at x1 , which is equal

/ H0 because of (i). Also (t1 , x1 )
/ H1 .
to zero. Clearly (t1 , x1 )
Otherwise, we would have
m(t1 , x1 ) u(t1 , x1 h)
0.
h0
h
lim
62
Inequalities
which contradicts (ii). Hence (t1 , x1 ) H. Consequently, we get

mx (t1 , x1 ) = 0 and
i j mxi xj (t1 , x1 ) 0,
Rn .
i,j=1
Also, m(t1 h, x1 ) m(t1 , x1 ) < 0 for h > 0 suciently small. It

therefore follows that mt (t1 , x1 ) 0. But by (iii), we have mt (t1 , x1 )
is less than zero, which is a contradiction. The lemma is therefore
proved.
2
Let f C[H R Rn Rn , R]. Then, f is said to be elliptic at

a point (t1 , x1 ) if for any u, Pi , Qij , Rij , the quadratic form
n

(Qij Rij )i j 0,
Rn
i,j=1
implies
f (t1 , x1 , u, P, Q) f (t1 , x1 , u, P, R).
If this property holds for every (t, x) H, then f is said to be elliptic
in H.
We can now prove the following comparison results.
Theorem 1.12.1 Suppose that
2
(i) v, w C, f C[H R Rn Rn , R], f is elliptic in H and
vt f (t, x, v, vx , vxx ), wt f (t, x, w, wx , wxx ) in H;
v
(ii) v < w on H0 and
< w
on H1 .
Then, v < w on H, if one of the inequalities in (i) is strict.
Proof Assume that one of the inequalities in (i) is strict. It is
enough to show that m = v w satises the hypotheses of Lemma
1.12.1. It is easy to see that conditions (i) and (ii) of Lemma 1.12.1
hold. We only need to verify condition (iii). Let (t1 , x1 ) H be such
that
m(t1 , x1 ) = 0 = mx (t1 , x1 )
and
n

i,j=1
i j mxi xj (t1 , x1 ) 0,
Rn .
63
Reactiondiffusion inequalities
This implies that

v(t1 , x1 ) = w(t1 , x1 ),
n

vx (t1 , x1 ) = wx (t1 , x1 ) and
i j [vxi xj (t1 , x1 ) wxi xj (t1 , x1 )] 0,
Rn .
i,j=1
Thus, in view of (i), it follows that

mt (t1 , x1 ) = vt (t1 , x1 ) wt (t1 , x1 )
< f (t1 , x1 , v(t1 , x1 ), vx (t1 , x1 ), vxx (t1 , x1 ))
f (t1 , x1 , w(t1 , x1 ), wx (t1 , x1 ), wxx (t1 , x1 )) 0,
which proves the theorem.
We can dispense with the strict inequalities needed in Theorem
1.12.1 if f satises the following condition:
> > 0 on H1 and for suciently
(C0 ) z C, z > 0 on H, z(t,x)
small > 0,
either
(a) zt > f (t, x, v, vx , vxx ) f (t, x, v z, vx zx , vxx zxx )
or
(b) zt > f (t, x, w + z, wx + zx , wxx + zxx ) f (t, x, w, wx , wxx )
holds on H, where v, w C.
Theorem 1.12.2 Let the assumption (i) of Theorem 1.12.1 hold
and suppose that (C0 ) is satised. Then the relation
vw
on
H0
and
v
w
on
H,
imply v w on H.
= w + z. Then,
Proof Assume that (b) of (C0 ) holds and set w
we have
w
x , w
xx )
w
t = wt + zt f (t, x, w, wx , wxx ) + zt > f (t, x, w,
z
v
on H. Also, v < w
on H0 and on H1 , w = w
+ + .
Thus, the functions v and w
satisfy the assumptions of Theorem
64
Inequalities
1.12.1 and hence v < w

on H. Taking limit as 0 yields the
desired result. When (a) of (C0 ) holds, the proof is similar and
therefore, the proof is complete.
As an example, consider
f (t, x, u, ux , uxx ) = auxx + bux + F (t, x, u)
where auxx =
n

i,j=1
aij uxi xj , bux =
n

j=1
bj uxj and suppose that
F is locally Lipschitzian in u. Assume that the boundary H1

is regular, that is, there exists a function h C such that
h(x) 0, h
Let
1 on H1 and hx , hxx are bounded.
M > 1, H(x) = exp(LM h(x)) 1, z(t, x) = H(x) exp(N t), where
N = LM + A, L is the Lipschitz constant and aHxx + bHx A.
Then,
h
z
= LM z LM > 0 on H1
and
zt azxx bzx (N A)z = LM z > Lz.
Consequently, using Lipschitz condition of F , we arrive at
zt > [azxx + bzx ] + F (t, x, w + z) F (t, x, w)
which is exactly the condition (b) of (C0 ).
If H1 is empty so that H0 = H, then assumption (C0 ) can be
replaced by a one sided Lipschitz condition
(C1 )f (t, x, u, P, Q) f (t, x, v, P, Q) L(u v),
u v.
In this case, it is enough to set w

= w + exp(2Lt) in the proof of
Theorem 1.12.2 so that v < w
on H and
w
t f (t, x, w, wx , wxx ) + 2L exp(2Lt)
xx ) + L exp(2Lt)
f (t, x, w,
w
x , w
xx ),
> f (t, x, w,
w
x , w
on
H.
65
Notes
Even when H1 is not empty, condition (C1 ) is enough provided that

the assumption
v
w
+ (t, x, v)
+ (t, x, w),
on
H1
where C[H R, R] and is strictly increasing in u. To see this,

observe that w
> w and hence (t, x, w) < (t, x, w)
which gives the
desired strict inequality needed in the proof. Of course, if is not
strictly increasing or 0, then condition (C0 ) becomes essential.
1.13
Notes
Theorem 1.1.1 which is the well-known fundamental linear integral

inequality, known as GronwallBellman inequality, was made popular
by Bellman [1] in this form. The inequality contained in Example
1.1.1 is actually due to Gronwall [1]. See Lakshmikantham and Leela
[1] for Theorem 1.1.2. Observations concerning Theorem 1.1.2 are
new. Theorem 1.1.3 appears in Beesack [1]. For Corollaries 1.1.1
and 1.1.2 see Gamidov [1]. Theorem 1.1.4 and Corollaries 1.1.4 and
1.1.5 are taken from Martynyuk and Gutowski [1]. Theorem 1.1.5 is
new.
For Theorem 1.2.1 and 1.2.2 see Abramovich [1] and Corduneanu
[1]. Corollary 1.2.1 is from Beckenback and Bellman [1]. See Bainov
and Zahariev [1] for Theorem 1.2.3.
Theorem 1.3.1 is due to Bihari [1], while Theorem 1.3.2 is from
Muldowney and Wong [1]. Theorem 1.3.3 is given in Achmedov,
Yakubov and Veisov [1]. Theorems 1.3.4 and 1.3.5 are taken from
Martynyuk and Gutowski [1] and Gutowski and Radziszewski [1].
Theorem 1.3.6 is new.
For the results of Section 1.4 see Bainov and Hristova [1,2,3],
Bainov and Simeonov [1], Bondge and Pachpatte [1, 2], Bondge,
Pachpatte and Walter [1] and Young [1].
The results of Section 1.5 are adapted from Lakshmikantham and
Leela [1,4]. See also Walter [1].
The contents of Section 1.6 are taken from Lakshmikantham and
Leela [1]. See also Rama Mohan Rao [1], Viswanatham [1] and Louartassi et al. [1].
66
Inequalities
Section 1.7 contains the results that are due to Lakshmikantham

[2] and Beesack [2]. See Martynyuk and Gutowski [1] for a good
exposition of such results.
Most of the results of Section 1.8 are given in Lakshmikantham,
Leela and Rama Mohan Rao [1]. For Lemma 1.8.1 see Shendge [1].
The material concerning dierence inequalities in Section 1.9 is
adapted from Lakshmikantham and Trigiante [1].
See Moore [1,2] for the results of Section 1.10.
Theorem 1.11.2 is taken from Lakshmikantham and Leela [2] while
the rest of the results of Section 1.11 are adapted from Bainov, Lakshmikantham and Simeonov [1].
The results of Section 1.12 are taken from Lakshmikantham [3].
For several results and references on integral inequalities see Martynyuk, Lakshmikantham and Leela [1] and Turinici [2]. For abstract GronwallBellman and multivariate inequalities see Chandra
and Fleishman [1] and Turinici [1].
2
VARIATION OF PARAMETERS
AND MONOTONE TECHNIQUE
2.0
Introduction
In this chapter we stress the importance of nonlinear variation of

parameters formulae for a variety of nonlinear problems as well as
monotone iterative technique that oers constructive methods for the
existence of solution in a sector.
In Section 2.1, we present dierent forms of nonlinear variation
of parameters formula relative to a dierential system including the
well-known Alekseevs formula. Section 2.2 utilizes various nonlinear variation formulae to obtain estimates of solutions of perturbed
dierential system. While doing so, we trace the successive developments of using integral and dierential inequalities and consequent
advantages. Section 2.3 deals with results on global existence, terminal value problems and asymptotic equilibrium. Some simple stability criteria using norm as a candidate are given in Section 2.4.
Section 2.5 employs the method of upper and lower solutions to
prove existence results in a sector. Section 2.6 is devoted to the development of monotone iterative technique when the system involved
enjoys a certain mixed monotone properties, while Section 2.7 investigates the method of mixed monotony which shows that it is possible to construct monotone sequences that converge to solutions of
a given problem even when the system considered does not possess
any monotone properties. Using interval analytic methods together
with the method of upper and lower solutions, we obtain in Section
2.8 simultaneous bounds for solutions of dierential equations.

67
68
Variation of parameters and monotone technique
Section 2.9 considers nonlinear integro-dierential equations, investigates the problem of continuity and dierentiability of solutions with respect to initial values, obtains nonlinear variation of
parameters formula and discusses qualitative behavior of solutions
of perturbed integro-dierential systems as an application of this
formula. Stability in variation is studied for integro-dierential systems in Section 2.10 where, given a linear integro-dierential system, a method of nding an equivalent linear dierential system, is
discussed and then exploiting this approach stability properties of
nonlinear integro-dierential systems are investigated.
Finally Section 2.11 deals with stability results for dierence equations. For this purpose, we develop suitable variation of parameters
formulae and employ the corresponding theory of dierence inequalities.
2.1
Nonlinear Variation of Parameters
It is well known that the method of variation of parameters is a very

useful tool in the investigation of the properties of solutions of nonlinear systems. In this section, we shall discuss dierent forms of the
variation of parameters formula which oer exibility in applications.
Consider the dierential system
x = f (t, x),
x(t0 ) = x0 ,
t0 R+ ,
(2.1.1)
and its perturbed system

y = f (t, y) + R(t, y),
y(t0 ) = x0 ,
(2.1.2)
where f, R C[R+ Rn , Rn ]. Let us begin with the following result.

Theorem 2.1.1 Assume that the system (2.1.1) admits unique
solutions x(t, t0 , x0 ) for t t0 . Suppose further that
(t, t0 , x0 ) =
x(t, t0 , x0 )
x0
exists, is continuous and 1 (t, t0 , x0 ) exists for all t t0 . Let v(t)

be any solution of
v (t) = 1 (t, t0 , x0 )R(t, x(t, t0 , v(t))),
v(t0 ) = x0 ,
(2.1.3)
69
Nonlinear variation of parameters
existing on t t0 . Then, any solution y(t, t0 , x0 ) of (2.1.1) satises

the relations
y(t, t0 , x0 ) = x(t, t0 , x0 )
t
+ 1 (s, t0 , v(s))R(s, y(s, t0 , x0 )) ds,
(2.1.4)
t0
y(t, t0 , x0 ) = x(t, t0 , x0 )
t
+ (t, t0 , v(s))1 (s, t0 , v(s))R(s, y(s, t0 , x0 )) ds
t0
(2.1.5)
for t t0 .
Proof Let x(t, t0 , x0 ) be the unique solution of (2.1.1) existing for
t t0 . The method of variation of parameters requires determining
a function v(t) such that
y(t, t0 , x0 ) = x(t, t0 , v(t)),
v(t0 ) = x0 ,
(2.1.6)
is a solution of (2.1.2). Dierentiating with respect to t yields

y (t, t0 , x0 ) = x (t, t0 , v(t)) +
x(t, t0 , v(t))
v (t)
x0
and thus, we have

f (t, y(t, t0 , x0 )) + R(t, y(t, t0 , x0 ))
= f (t, x(t, t0 , v(t))) + (t, t0 , v(t))v (t),
which reduces to (2.1.3) because of (2.1.6) and the fact that
1 (t, t0 , x0 ) exists. Consequently, if v(t) is a solution of (2.1.3),
then y(t, t0 , x0 ) given by (2.1.6) is solution of (2.1.2). From (2.1.3),
v(t) must satisfy the integral equation
t
v(t) = x0 +
t0
1 (s, t0 , v(s))R(s, x(s, t0 , v(s))) ds
70
and hence (2.1.4) follows.

To prove (2.1.5), for t0 s t, we note that
d
x(t, t0 , v(s))
(x(t, t0 , v(t))) =
v (s) = (t, t0 , v(s))v (s)
ds
x0
and hence, integrating from t0 to t, we obtain
t
x(t, t0 , v(t)) = x(t, t0 , x0 ) +
(t, t0 , v(s))v (s) ds.
t0
If v(t) is any solution of (2.1.3) then the result follows immediately

in view of (2.1.3) and (2.1.6).
Corollary 2.1.1 If f (t, x) = A(t)x, where A(t) is an n by n
continuous matrix, then x(t, t0 , x0 ) = (t, t0 )x0 , (t, t0 ) being the
fundamental matrix solution of
x = A(t)x,
with (t, t0 ) = I, I being the identity matrix. Then, the relations
(2.1.4) and (2.1.5) both yield the well known form
y(t, t0 , x0 ) = (t, t0 )x0
t
+ (t, t0 )1 (s, t0 )R(s, y(s, t0 , x0 )) ds,
(2.1.7)
t0
since (t, t0 ) is nonsingular for t t0 .

To prove another type of variation of parameters formula known
as Alekseevs formula, we need the following result on dierentiability
of solutions with respect to initial values. We merely state such a
result since we shall prove later a more general result relative to
integro-dierential equations.
Theorem 2.1.2 Assume that f has continuous partial derivatives
f
on R+ Rn . Let x(t, t0 , x0 ) be the unique solution of (2.1.1)
x
f
(t, x(t, t0 , x0 )). Then,
existing for t t0 and let H(t, t0 , x0 ) =
x
Nonlinear variation of parameters
(i) (t, t0 , x0 ) =
71
x(t, t0 , x0 )
exists and is solution of
x0
y = H(t, t0 , x0 )y,
(2.1.8)
such that (t0 , t0 , x0 ) is the identity matrix;

(ii)
x(t, t0 , x0 )
exists, is the solution of (2.1.8) with
t0
x(t0 , t0 , x0 )
= f (t0 , x0 )
t0
and satises the relation
x(t, t0 , x0 )
+ (t, t0 , x0 )f (t0 , x0 ) = 0,
t0
t t0 .
(2.1.9)
Theorem 2.1.3 Under the assumptions of Theorem 2.1.2, any

solution y(t, t0 , x0 ) of (2.1.2) satises the integral equation
y(t, t0 , x0 ) = x(t, t0 , x0 )
t
+ (t, s, y(s, t0 , x0 ))R(s, y(s, t0 , x0 )) ds,
(2.1.10)
t0
x(t, t0 , x0 )
. Furthermore, the relax0
tions (2.1.5) and (2.1.10) are equivalent.
for t t0 where (t, t0 , x0 ) =
Proof Writing y(t) = y(t, t0 , x0 ), we see that

x(t, s, y(s)) x(t, s, y(s))
dx
(t, s, y(s)) =
+
y (s)
ds
t0
x0
= (t, s, y(s))[y (s) f (s, y(s))],
(2.1.11)
in view of (2.1.9). Noting that x(t, t, y(t)) = y(t) and R(s, y(s)) =
y (s) f (s, y(s)), the desired relation (2.1.10) follows by integrating
(2.1.11) from t0 to t. Now, to show the equivalence of (2.1.5) and
(2.1.10), rst note that for t0 s t,
x(t, s, y(s)) = x(t, t0 , v(s)),
(2.1.12)
72
where v(t) is a solution of (2.1.3). Then, dierentiating with respect

to s, we get
x(t, s, y(s)) x(t, s, y(s))
x(t, s, v(s))
+
y (s) =
v (s).
t0
x0
x0
This implies, substituting for y (s) and v (s) and using (2.1.9),
(t, s, y(s))R(s, y(s)) = (t, t0 , v(s))1 (s, t0 , v(s))R(s, y(s))
which proves the stated result.
All the foregoing variation of parameters formulae provide integral
representation for solutions of (2.1.2) in terms of solutions of (2.1.1)
with the same initial data (t0 , x0 ). In order to derive a formula when
the initial data are dierent , we need the following result.
Theorem 2.1.4 Under the assumptions of Theorem 2.1.2, if
x(t, t0 , x0 ) and x(t, t0 , x0 ) are any two solutions of (2.1.2) through
(t0 , y0 ) and t0 , x0 respectively, existing for t t0 , then
x(t, t0 , y0 ) = x(t, t0 , x0 )
1
(t, t0 , y0 s + (1 s)x0 ) ds (y0 x0 )
(2.1.13)
holds for t t0 .
Proof Set w(s) = x(t, t0 , y0 s + (1 s)x0 ) for 0 s 1 so that
dw(s)
= (t, t0 , y0 s + (1 s)x0 ) (y0 x0 ).
ds
Integrating this relation from 0 to 1 yields (2.1.13).
Now we shall prove the promised result.
Theorem 2.1.5 Under the assumptions of Theorem 2.1.1, if
y(t, t0 , y0 ) and x(t, t0 , x0 ) are solutions of (2.1.1) and (2.1.2) respec-
73
Estimates of solutions
tively, then we have for t t0 ,

1
(t, t0 , y0 s + (1 s)x0 ) ds
y(t, t0 , y0 ) = x(t, t0 , x0 ) +
0
(2.1.14)
t
(y0 x0 ) +
(t, s, y(s, t0 , y0 ))R(s, y(s, t0 , y0 )) ds.

t0
Proof
Using (2.1.10) and (2.1.13), we obtain

t
y(t, t0 , y0 ) = x(t, t0 , x0 ) +
(t, s, y(s, t0 , y0 ))R(s, y(s, t0 , y0 )) ds

t0
and
1
(t, t0 , y0 s + (1 s)x0 ds (y0 x0 ),
x(t, t0 , y0 ) = x(t, t0 , x0 ) +
0
which when added yield representation (2.1.14).
2.2
Estimates of Solutions
Consider the dierential system (2.1.1) which is equivalent to the

integral equation
t
x(t) = x0 +
f (s, x(s)) ds.
(2.2.1)
t0
Suppose that f satises the condition

f (t, x) g(t, x),
(t, x) R+ Rn ,
(2.2.2)
2 , R ], then we arrive at the integral inequality

where g C[R+
+
t
m(t) m(t0 ) +
g(s, m(s)) ds,

t0
t t0 ,
(2.2.3)
74
where m(t) = x(t, t0 , x0 ), x(t, t0 , x0 ) being any solution of (2.2.1)

existing on [t0 , ). If we assume in addition that g(t, u) is nondecreasing in u, then by Theorem 1.6.1, we obtain the estimate
x(t, t0 , x0 ) r(t, t0 , x0 ),
t t0 ,
(2.2.4)
where r(t, t0 , u0 ) is the maximal solution of

u = g(t, u),
u(t0 ) = u0 0,
(2.2.5)
Instead of using the integral equation (2.2.1), if we utilize the
dierential equation (2.1.1) directly, we get
m+ (t) = f (t, x(t, t0 , x0 )) g(t, m(t)),
t t0 ,
(2.2.6)
where m+ (t) is the right hand derivative of m(t). Then, Theorem
1.5.2 provides the same estimate (2.2.4) without the extra assumption of nondecreasing nature of g.
Note also that (2.2.2) demands that g be nonnegative which in
turn implies that the solutions r(t, t0 , u0 ) of (2.2.5) are nondecreasing. Thus the estimate (2.2.4) does not provide the best possible
information. To remove this drawback, we need to replace (2.2.2) by
[x, f (t, x)]+ lim
h0
1
[x + hf (t, x) x] g(t, x),
h
for (t, x) R+ Rn . If f (t, x) = Ax, where A is an nn matrix, then

it is easy to compute that g(t, u) = (A)u, (A) is the logarithmic
norm dened by (A) = lim (1/h)[I + hA 1], I being the identity
h0
matrix, the value of (A) depends, of course, on the particular norm

used for vectors and matrices. For example, if x represents the
Euclidian norm, then (A) is the largest eigenvalue of 1/2[A + A ],
A being the transpose of A , where as the corresponding matrix
norm A is the square root of the largest eigenvalue of A A. On
the other hand, if
x =
n

i=1
|xi | and A = sup

k
n

i=1
|aik |,
75
then

(A) = sup Re akk +
k

|aik | .
i=k
Also we note that x[x, y]+ = (x, y), the inner product. For more
details on (A), direction derivatives [x, y] and the generalized inner
products (x, y) in an arbitrary Banach space see Lakshmikantham
and Leela [1,2]. The relation (2.2.6) yields the dierential inequality
Dm(t) g(t, m(t)),
t t0 ,
for a suitable Dini derivative Dm(t), which implies (2.2.4), by Theorem 1.5.2 as before. We observe that g in (2.2.6) need not be nonnegative and consequently the bound obtained in (2.2.4) is much better.
These considerations prove the following result.
Theorem 2.2.1 Assume that either (2.2.2) or (2.2.6) holds.
Then if x(t, t0 , x0 ) is any solution of (2.1.1) existing for t t0 and
r(t, t0 , u0 ) is the maximal solution of (2.2.5) existing on [t0 , ), we
have the estimate (2.2.4).
We shall next utilize the various nonlinear variation of parameters
formulae to obtain estimates of solutions of the perturbed system
(2.1.2).
Theorem 2.2.2 In addition to the assumptions of Theorem 2.1.1,
suppose that
1 (t, t0 , x0 )R(t, x(t, t0 , x0 )) g(t, x0 )
(2.2.7)
2 , R ] and r(t, t , u ) is the maximal solution of

holds where g C[R+
+
0 0
(2.2.5) existing for t t0 . Assume further that
x(t, t0 , x0 ) a(t0 , x0 )(t t0 ),
t t0 ,
(2.2.8)
2 , R ], a(t, u) is nondecreasing in u for each t,

where a C[R+
+
C[R+ , R+ ] and (t) 0 as t . Then, any solution y(t, t0 , x0 )
of (2.1.2) existing on [t0 , ) satises the estimate
y(t, t0 , x0 ) a(t0 , r(t, t0 , x0 ))(t t0 ),
t t0 .
(2.2.9)
76
Proof By Theorem 2.1.1, any solution y(t, t0 , x0 ) of (2.1.2) satises

y(t, t0 , x0 ) = x(t, t0 , v(t))
(2.2.10)
where v(t) is a solution of (2.1.3). Setting m(t) = v(t) and using

(2.2.7), we obtain the inequality
D + m(t) g(t, m(t))
which, by Theorem 1.5.2, yields the relation
v(t) r(t, t0 , x0 ),
t t0 .
(2.2.11)
We then have, from (2.2.8), (2.2.10) and (2.2.11),

y(t, t0 , x0 ) a(t0 , v(t))(t t0 )
a(t0 , r(t, t0 , x0 ))(t t0 ),
t t0 ,
using the monotonic nature of a(t, u).

suppose that
(i) (t, t0 , x0 ) a(x0 ), 1 (t, t0 , x0 ) a(x0 ), t t0
where a C[R+ , R+ ] and a(u) is nondecreasing in u;
2 , R ] and g(t, u) is non(ii) R(t, y) g(t, y) where g C[R+
+
decreasing in u for each t;
(iii) r(t, t0 , x0 ), r0 (t, t0 , a(x0 )x0 ) are the maximal solutions
of
u = g(t, u),
u = g0 (t, u),
u(t0 ) = x0 ,
u(t0 ) = a(x0 )x0 ,
existing for t t0 , where g(t, u) = a(u)g(t, ua(u)), g0 (t, u) =

r (t, t0 , x0 ))g(t, u), a2 (u) being the composition a(a(u)).
a2 (
Then, any solution y(t, t0 , x0 ) of (2.1.2) satises the estimate
y(t, t0 , x0 ) r0 (t, t0 , a(x0 )x0 ),
t t0 .
(2.2.12)
77
Proof The solution x(t, t0 , x0 ) of (2.1.1) is related to (t, t0 , x0 )

by
1
x(t, t0 , x0 ) =

(t, t0 , x0 s) ds x0
as can be seen by integrating

dx
(t, t0 , sx0 ) = (t, t0 , sx0 )x0
ds
from s = 0 to s = 1. Hence, assumption (i) implies
x(t, t0 , x0 ) a(x0 )x0 ,
t t0 .
(2.2.13)
Furthermore, it follows from (2.1.3) and in view of (i) and (ii), that
t
v(t) x0 +
a(v(s))g(s, x(s, t0 , v(s))) ds

t0
t
x0 +
g(s, v(s)) ds,
t t0 .
t0
Since g(t, u) is also nondecreasing in u, Theorem 1.6.1 gives

v(t) r(t, t0 , x0 ),
t t0 .
(2.2.14)
Now, using the relations (2.1.5), (2.2.13) and (2.2.14), we get

t
y(t, t0 , x0 ) a(x0 )x0 +
a2 (v(s))g(s, y(s, t0 , x0 )) ds
t0
t
a(x0 )x0 +
g0 (s, y(s, t0 , x0 )) ds,

t0
which implies, by Theorem 1.6.1, the desired estimate (2.2.12).

Theorem 2.2.4 In addition to assumptions of Theorem 2.1.2,
suppose that
78
(i) (t, t0 , x0 ) a(x0 ), t t0 , where a C[R+ , R+ ] and a(u)

is nondecreasing in u;
(ii) R(t, y) g(t, y), where g C[R+ R+ , R+ ] and g(t, u) is
nondecreasing in u;
(iii) r(t, t0 , a(x0 )x0 ) is the maximal solution of
u = a(u)g(t, u),
u(t0 ) = a(x0 )x0 ,
existing for t t0 .
Then, for any solution y(t, t0 , x0 ) of (2.1.2) we have
y(t, t0 , x0 ) r(t, t0 , a(x0 )x0 ),
t t0 .
(2.2.15)
Proof Using the integral representation (2.1.10) and the estimate

(2.2.13), we easily obtain the inequality
t
y(t, t0 , x0 ) a(x0 )x0 +
a(y(s, t0 , x0 ))g(s, y(s, t0 , x0 )) ds

t0
which yields, by Theorem 1.6.1, the desired relation (2.2.15).

Theorem 2.2.5 Assume that the hypotheses of Theorem 2.1.2
and the conditions (i) and (ii) of Theorem 2.2.4 hold. Suppose that
r(t, t0 , u0 ) is the maximal solution of
u = a(u)g(t, u),
u(t0 ) = u0
existing for t t0 , such that

a(x0 )x0 + a(x0 + y0 )y0 x0 u0 .
Then, any solution y(t, t0 , y0 ) of (2.1.2) with y(t0 ) = y0 satises
y(t, t0 , y0 ) r(t, t0 , u0 ),
t t0 .
(2.2.16)
79
Proof The integral equation (2.1.14) together with (2.2.13) shows

that, for t t0 we have
y(t, t0 , y0 ) a(x0 )x0 + a(x0 + y0 )y0 x0
t
+ a(y(s, t0 , x0 ))g(s, y(s, t0 , y0 )) ds,
t0
which immediately gives, by Theorem 1.6.1, the relation (2.2.16).

Further we assume that the system (2.1.2) is such that
A1 . For all t 0 there exists a nonnegative integrable function b(t)
such that
f (t, y) b(t)y
for all (t, y) R+ Rn .
A2 . For all t 0 there exists a nonnegative integrable function c(t)
such that
R(t, y) c(t)y
for all (t, y) R+ Rn and > 1.
Theorem 2.2.6 For the system of equations (2.1.2) let conditions
of assumptions A1 and A2 be satised. Then for the norm of solutions
y(t) = y(t, t0 , y0 ) the estimate
y0 exp
t
t0
y(t)
1 ( 1)y0
1
t
b(s)ds

t
c(s) exp ( 1) b( )d ds
1
1
t0
(2.2.17)
holds true for all t t0 0 whenever
t
( 1)y0
t
c(s) exp ( 1)
1
t0

b( )d ds < 1.
(2.2.18)
80
Proof Let y(t) be the solution of system of equations (2.1.2) with

the initial conditions y(t0 ) = y0 , t 0. Under conditions A1 and A2
equation (2.1.2) yields the estimate of the norm of solutions y(t) in
the form
t
y(t) y0 +
t
c(s)y(s) ds.
b(s)y(s) ds +
t0
(2.2.19)
t0
We transform inequality (2.2.19) to the pseudo-linear form

t
y(t) y0 +

b(s) + c(s)y(s)1 y(s) ds,
(2.2.20)
t0
and applying the Gronwall-Bellman lemma arrive at the estimate

t
y(t) y0 exp

b(s) + c(s)y(s)1 ds

(2.2.21)
t0
for all t t0 0.
Further, estimating the expression

t
c(s)y(s)1 ds
exp
t0
we get estimate (2.2.17).

Corollary 2.2.1 (cf. Demidovich [1]) Consider system (2.1.2)
for f (t, y) 0 for all t t0 0
dy
= R(t, y),
dt
y(t0 ) = y0 .
If condition A2 is fullled with the function c(t) such that

tk+1
c(s) ds > 0
tk
(2.2.22)
81
for any (tk , tk+1 ) R+ , tk < tk+1 , k = 0, 1, 2, . . . , then

t
y(t) y0 +
c(s)y(s) ds.
t0
Applying to this inequality the same procedure as in the proof of

Theorem 2.2.6 it is easy to show that if
t
1 ( 1)y0
c(s) ds > 0
t0
for all t t0 0, then

y0
y(t)
1 ( 1)y0
1
t

c(s)ds
1
1
(2.2.23)
t0
for all t t0 0.
Corollary 2.2.2 In system (2.1.2) let R(t, y) B(t, y)y, where
B : R+ Rn Rnn is an n n-matrix continuous with respect to
(t, x) R+ Rn .
Consider a system of non-autonomous linear equations with
pseudo-linear perturbation
dy
= f (t, y) + B(t, y)y,
dt
y(t0 ) = y0 .
(2.2.24)
Assume that condition A1 is satised and there exists a nonnegative

integrable function h(t) such that
B(t, y) h(t)y,
for all (t, y) R+ Rn .
Equation (2.2.24) implies that
t
y(t) y0 +
0
(b(s)y(s) + h(s)y(s)2 ) ds.
(2.2.25)
82
Applying to inequality (2.2.25) the same procedure as in the proof

of Theorem 2.2.6 we get the estimate
ky(t)k
ky0 k exp
1 ky0 k
Rt
Rt
b(s)ds
h(s) exp
Rs
b( )d ds
which holds true for the values of (t) [0, ) for which
1 ky0 k
Zt
h(s) exp
Zs
b( ) d ds > 0.
Corollary 2.2.3 In system (2.1.2) let R(t, y) = A2 (t)y 2 +

+ An (t)y n , where y i = col(y1i , y2 i, . . . , yni ) for all i = 2, 3, . . . , n.
Further we shall consider the system of nonlinear equations
n
dy X
=
Ak (t)y k ,
dt
y(t0 ) = y0 ,
(2.2.26)
k=1
where Ai C(R+ , Rnn ) are (n n)-matrices with the elements

continuous on any finite interval.
Assume that there exist nonnegative integrable on [0, ) functions bk (t), k = 1, 2, . . . , n, such that
kAk (t)k bk (t),
k = 1, 2, . . . , n.
(2.2.27)
In view of (2.2.27) we get from (2.2.26) the inequality
ky(t)k ky0 k +
Zt X
n
ky0 k +
0 k=1
Zt n
0 k=1
kAk (s)kky(s)kk ds
(2.2.28)
bk (s)ky(s)kk ds.
83
Global existence and terminal value problems
Applying to inequality (2.2.28) the same procedure as in the proof

of Theorem 2.2.6 it is easy to show that
ky(t)k ky0 k exp
(
1 (n 1)
Zt
b1 (s) ds
0
t
Z X
n
0 k=2
k1
bk (s)ky0 k
1 (n 1)
0 k=2
exp (k 1)
Zt
b1 ( )d ds
1
n1
(2.2.29)
for all t [0, ) if
Zt X
n
"
k1
bk (s)ky0 k
exp (k 1)
Zt
0
b1 ( )d ds > 0.
Estimate (2.2.17) allows boundedness and stability conditions for

solution of system (2.1.2) to be established in the following form.
Theorem 2.2.7 If conditions A1 and A2 of Theorem 2.2.6 are
satisfied for all (t, y) R+ Rn and there exists a constant > 0
such that ky(t)k < for all t t0 , where may depend on each
solution, then the solution y(t, t0 , y0 ) of system (2.1.2) is bounded.
Theorem 2.2.8 If conditions A1 and A2 of Theorem 2.2.6 are
satisfied for all (t, y) R+ Rn and f (t, y) = 0 for y = 0, and
for any > 0 and t0 0 there exists a (t0 , ) > 0 such that if
ky0 k < (t0 , ), then the estimate ky(t)k < is satisfied for all
t t0 , then the zero solution of system (2.1.2) is stable.
The proofs of Theorems 2.2.7 and 2.2.8 follow immediately from

the estimate of norm of solutions x(t) in the form of (2.2.17). The
notations ky(t)k < and ky(t)k < mean that the right hand part of
inequality (2.2.17) must satisfy these inequalities under appropriate
initial conditions.
2.3
Global Existence and Terminal Value Problems
The assumption (2.2.2) of Theorem 2.2.1 is strong enough to guarantee global existence of solutions of (2.1.1). In fact, we can prove
the following result.
84
Theorem 2.3.1 Assume that condition (2.2.2) holds and that

the maximal solution r(t, t0 , u0 ) of (2.2.5) exists for t t0 . Suppose
further that g(t, u) is nondecreasing in u. Then, the largest interval
of existence of any solution x(t, t0 , x0 ) of (2.1.1) with x0 u0
is [t0 , ). If, in addition, r(t, t0 , u0 ) is bounded on [t0 , ), then
lim x(t, t0 , x0 ) exists.
t
Proof Let x(t) = x(t, t0 , x0 ) be a solution of (2.1.1) such that

x0 u0 which exists on t0 t < for t0 < < and that the
value of cannot be increased. Set m(t) = x(t) for t0 t < and
obtain as in Theorem 2.1.1 the estimate
x(t) r(t),
t0 t < ,
(2.3.1)
where r(t) = r(t, t0 , u0 ). For any t1 , t2 such that t0 < t1 < t2 < ,
we see, by using (2.3.1) and the monotone nature of g, that
t2
x(t2 ) x(t1 )
g(s, x(s)) ds
t1
(2.3.2)
t2
g(s, r(s)) ds = r(t2 ) r(t1 ).

t1
Since lim r(t) exists and is nite, taking limit as t1 , t2 and

t
using Cauchy criterion for convergence, it follows from (2.3.2) that

lim x(t) exists. Now, dene x() = lim x(t) and consider the
initial value problem

x = f (t, x),
x() = lim x(t).

t
Because of local existence, it follows that x(t) can be continued

beyond , contradicting our assumption. Hence, every solution x(t)
of (2.1.1) such that x0 u0 exists on [t0 , ) and so the relations
(2.3.1) and (2.3.2) hold with = . Since r(t) is bounded and nondecreasing on [t0 , ), it follows that lim r(t) exists, which implies
t
in view of (2.3.1) and (2.3.2) that lim r(t) exists and the proof is
t
complete.
85
Global existence can be proved by relaxing condition (2.2.2) to

(2.2.6) and also not requiring g(t, u) to be nondecreasing. This motivates the next result.
Theorem 2.3.2 Assume that the condition (2.2.6) holds and that
the maximal solution r(t) = r(t, t0 , u0 ) of (2.2.5) exists for t t0 .
Then for every x0 u0 , the problem (2.1.1) has a solution x(t) on
[t0 , ) satisfying x(t) r(t), t t0 .
Proof Let S denote the set of all functions x dened on
Ix = [t0 , cx ) such that x(t) is a solution of (2.1.1) on Ix and
x(t) r(t) on Ix . We dene a partial order on S as follows: the
relation x y implies that Ix Iy and y(t) x(t) on Ix . We shall
rst show that S is nonempty. By the local existence, there is a solution x(t) of (2.1.1) dened on Ix . Setting m(t) = x(t) for t Ix
and using assumption (2.2.6), it is easy to obtain the dierential
inequality
D + m(t) g(t, m(t)), t Ix .
Now, by Theorem 1.5.2, it follows that
x(t) r(t),
t Ix ,
(2.3.3)
where r(t) is the maximal solution of (2.2.5). This shows that x S

and so S is nonempty.
If (x ) is a chain (S, ), then there is a uniquely dened map y
on Iy = [t0 , sup cx ) that coincides with x on Ix . Clearly y S and
hence y is an upper bound of (x ) in (S, ). Then, Zorns lemma

assures the existence of a maximal element z in (S, ). The proof of
the theorem is complete if we show that cz = . Suppose that it is
not true, so that cz < . Since r(t) is assumed to exist on [t0 , ),
r(t) is bounded on Iz . The relation x(t) r(t) on Iz implies that
z(t) is bounded on Iz which yields that there is an M > 0 such
that f (t, z(t)) M on Iz . We then have for all t1 , t2 Iz , t1 t2 ,
t2
z(t2 ) z(t1 )
f (s, z(s)) ds M (t2 t1 ),

t1
86
which shows that z is Lipschitzian on Iz and consequently, has a

continuous extension z0 on [t0 , cz ). By continuity, we get
cz
z0 (cz ) = x0 +
f (s, z0 (s)) ds.

t0
This implies that z0 (t) is a solution of (2.1.1) on [t0 , cz ] and clearly

z0 (t) r(t), t [t0 , cz ]. Consider the problem
x = f (t, x),
x(t0 ) = z0 (cz ).
By local existence, there exists a solution x0 (t) on [cz , cz + ), > 0.

Dene

z0 (t) for t0 t cz
z1 (t) =
x0 (t) for cz t < cz + .
Clearly z1 (t) is a solution of (2.1.1) on [t0 , cz + ) and by repeating
the arguments that were used to obtain (2.3.3), we get
z1 (t) r(t),
t [t0 , cz + ).
This contradicts the maximality of z and hence cz = . The proof

is complete.
Theorem 2.3.3 Under the assumptions of Theorem 2.3.1, the
terminal value problem
x = f (t, x),
x() = x ,
(2.3.4)
has a solution.
Proof We shall construct a solution x(t) of (2.1.1) such that
lim x(t) = x . We shall rst observe that for every (t0 , )

g(s, ) ds < .
(2.3.5)
t0
To see this, let r(t) = r(t, t0 , ) be the maximal solution of (2.2.5)

through (t0 , ). Since lim r(t) = r exists, we have
t
t
r r(t) = +
t
g(s, r(s)) ds +
t0
g(s, ) ds.
t0
87
Next, consider the maximal solution r(t) = r(t, t0 , x ) of (2.2.5).

Setting r = lim r(t), we see, because of (2.3.5), that we can choose
t
T suciently large so that

g(s, 2r ) ds < r .
(2.3.6)
For each n = 0, 1, 2, . . ., construct the maximal solution rn (t) =

r(t, T + n, x ) of (2.2.5) and a solution xn (t) = x(t, T + n, x )
of (2.1.1). By Theorem 2.3.1, xn (t) exists on [T + n, ), tends to a
nite limit as t and
xn (t) rn (t),
T + n t < .
(2.3.7)
We shall prove that xn (t) can be continued backwards up to T

and
(2.3.8)
xn (t) 2r , T t T + n.
Before we do this, let Rn (t) = Rn (t, T +n, x ) be maximal solution
of
(2.3.9)
u = g(t, u).
We claim that Rn (t) exists on [T, T + n]. To prove this, it is enough
to show that Rn (t) remains bounded on [T, T + n]. If not, there exist
t1 , t2 [T, T + n] such that Rn (t1 ) = 2r , Rn (t2 ) = r and on
[t1 , t2 ], r Rn (t) 2r . Then, from (2.3.9) we have
t2
Rn (t2 ) = Rn (t1 )
g(s, Rn (s)) ds,

t1
and consequently
t2
t2
g(s, Rn (s)) ds
r =
t1

g(s, 2r ) ds
t1
g(s, 2r ) ds.
T
This contradicts (2.3.6). Now, an argument similar to that of Theorem 2.1.1 shows that xn (t) exists on [T, T + n].
88
We shall next establish (2.3.8). If it were false, there exist

t1 , t2 [T, T + n] such that xn (t1 ) = 2r , xn (t2 ) = r and
on [t1 , t2 ], r xn (t) 2r . Then , we get
t2
r 2r
t2
f (s, xn (s)) ds 2r
t1
2r
g(s, 2r ) ds
t1
g(s, 2r ) ds > r
T
and this contradiction proves (2.3.8).

The solutions xn (t) are therefore dened on [T, ) and are uniformly bounded by 2r . Furthermore, since
xn (t) = f (t, xn (t)) g(t, 2r ),
the sequence {xn (t)} is equicontinuous on every bounded t interval.
Applying Ascoli-Arzela theorem, there is a subsequence, which we
shall denote by {xnk (t)} that converges uniformly on every bounded
t interval as n to a continuous function x(t). The function x(t)
is the desired solution of the terminal value problem (2.3.4). In fact,
we have
t
xnk (t) = xnk (T ) + f (s, xnk (s)) ds
T
and passing to the limit, we see that x(t) is a solution of (2.3.4). Since
lim x(t) = x() exists and lim xnk (t) = x(t), xnk (T + n) = x ,
nk
we can conclude that x() = x . The proof is complete.

We say that the system (2.1.1) has asymptotic equilibrium if every
solution of (2.1.1) tends to a nite limit as t and conversely,
for every , there exists a solution of the terminal value problem
(2.3.4). Thus, we see that under the assumptions of Theorem 2.3.1,
the system (2.1.1) possesses asymptotic equilibrium.
Stability criteria
2.4
89
Stability Criteria
In this section, we shall consider some simple stability results. We

list a few denitions concerning the stability of the trivial solution
of (2.1.2) which we assume to exist.
Denition 2.4.1 The trivial solution x = 0 of (2.1.1) is said to
be
(S1 ) equi-stable if, for each > 0 and t0 R+ , there exists a positive
function = (t0 , ) that is continuous in t0 for each such that
x0 < implies x(t, t0 , x0 ) < , t t0 ;
(S2 ) uniformly stable, if the in (S1 ) is independent of t0 ;
(S3 ) quasi-equi asymptotically stable, if for each > 0 and t0 R+ ,
there exists positive 0 = 0 (t0 ) and T = T (t0 , ) such that
x0 < 0 implies x(t, t0 , x0 ) < for t t0 + T ;
(S4 ) quasi-uniformly asymptotically stable, if 0 and T in (S3 ) are
independent of t0 ;
(S5 ) equi-asymptotically stable, if (S1 ) and (S3 ) hold simultaneously;
(S6 ) uniformly asymptotically stable, if (S2 ) and (S4 ) hold simultaneously.
Corresponding to the denitions (S1 ) to (S6 ), we can dene the
stability notions of the trivial solution u = 0 of (2.2.5). For example,
the trivial solution u = 0 of (2.2.5) is equi-stable if, for each > 0
and t0 R+ , there exists a function = (t0 , ) that is continuous
in t0 for each , such that u0 < implies u(t, t0 , u0 ) < , t t0 .
We are now in a position to prove some simple criteria for stability.
Theorem 2.4.1 Assume that (2.2.6) holds for (t, x) R+ S
where S = {x Rn : x < }. Suppose further that f (t, 0) 0
and g(t, 0) 0. Then, the stability properties of the trivial solution
of (2.2.5) imply the corresponding stability properties of the trivial
solution of (2.1.1).
Proof Let the trivial solution of (2.2.5) be equi-stable. Then, given
> 0 and t0 R+ , there exists a with the property that u0 <
90
implies u(t, t0 , u0 ) < for t t0 , where u(t, t0 , u0 ) is any solution

of (2.2.5). It is easy to show that with these and , the trivial
solution of (2.1.1) is equi-stable. If this were false, there would exist
a solution x(t) = x(t, t0 , x0 ) of (2.2.1) with x0 < and a t1 > t0
such that
x(t1 ) =
and x(t) ,
t [t0 , t1 ].
For t [t0 , t1 ], using condition (2.2.6), we obtain the estimate (2.2.4)

as in Theorem 2.2.1, namely,
x(t) r(t, t0 , x0 ),
t [t0 , t1 ],
where r(t, t0 , u0 ) is the maximal solution of (2.2.5). At t = t1 , we

arrive at the contradiction
= x(t1 ) r(t1 , t0 , x0 ) <
proving our claim. One can prove in a similar way other concepts of
stability and we omit the details.
Theorem 2.4.2 Under the assumptions of Theorem 2.2.2, if we
suppose that f (t, 0) 0, g(t, 0) 0, a(t, 0) 0, then the stability
properties of p(t, t0 , x0 ) a(t0 , r(t, t0 , x0 ))(t t0 ) imply the
corresponding stability properties of the trivial solution of (2.1.2).
The proof is immediate from the estimate (2.2.9) in view of our
assumptions.
suppose that
(i) (t, t0 , x0 ), 1 (t, t0 , x0 ) K whenever x0 < and
t t0 ;
(ii) R(t, y) g(t, y), where g C[R+ R+ , R+ ], g(t, u) is
non-decreasing in u, f (t, 0) 0, R(t, 0) 0 and g(t, 0) 0;
(iii) the trivial solution of
u = K 2 g(t, u)
is equi-stable.
u(t0 ) = u0 0.
(2.4.1)
91
Stability criteria
Then, the trivial solution of (2.1.2) is equi-stable.

1

R
Proof Since x(t, t0 , x0 ) =
(t, t0 , x0 s) ds x0 , we get
0
kx(t, t0 , x0 )k Kkx0 k,
t t0 ,
kx0 k < .
if
(2.4.2)
We wish to use relation (2.1.5). For this purpose, we consider v(t) of

(2.1.3) and suppose that for some t1 > t0 , we have kv(t1 )k = and
kv(t)k for t [t0 , t1 ]. Then, using (2.1.3), we get for t [t0 , t1 ],
kv(t)k kx0 k + K
kx0 k + K
Zt
g(s, kx(s, t0 , v(s))k) ds
Zt
g(s, Kkv(s)k) ds.
t0
t0
Thus
Kkv(t)k Kkx0 k + K
Zt1
g(s, Kkv(s)k) ds
t0
and by Theorem 1.6.1, we obtain

Kkv(t)k r(t, t0 , Kkx0 k),
t [t0 , t1 ],
where r(t, t0 , u0 ) is the maximal solution of (2.4.1), choosing kx0 k <

/K we see, in view of assumption (iii), that
= kv(t1 )k r(t1 , t0 , Kkx0 k) < ,
.
K
Now, the relations (2.1.5) and (2.4.2) yield, because of (i) and (ii),
which is a contradiction. Hence kv(t)k < for t t0 , if kx0 k <
ky(t, t0 , x0 )k Kkx0 k + K 2
Zt1
t0
g(s, ky(s, t0 , x0 )k) ds,
and hence, by Theorem 1.6.1, it follows that

ky(t, t0 , x0 )k r(t, t0 , KkX0k),
t t0 .
t t0 ,
92
Now, assumption (iii) implies the stated claim.

suppose that
(i) f (t, 0) 0, R(t, 0) 0;
(ii) k(t, t0 , x0 )k e(tt0 ) , if t t0 , kx0 k , with > 0;
(iii) kR(t, y)k (t)g(kyk) were C[R+ , R+ ], g(u) is nondecreasing, positive, submultiplicative on (0, ) and
!
Z
g(kx0 k)
1
M (t0 ) = G
G(1) +
(s) ds <
kx0 k
t0
where G(w) =
Ru
u0
(1/g(s))ds, u0 > 0 and G() = .
Then, the trivial solution of (2.1.2) is exponentially asymptotically

stable.
Proof Let kx0 k <

. Suppose that for some t1 > t0 , we have
M (t0 )
ky(t, t0 , x0 )k ,
t0 t t1
and ky(t1 , t0 , x0 )k = ,
where y(t, t0 , x0 ) is any solution of (2.1.2). Then, we get from (2.1.10)

and the assumptions of the theorem
ky(t, t0 , x0 )k kx0 ke(tt0 ) +
Zt
e(ts) (s)g(ky(s, t0 , x0 )k) ds
t0
for t0 t t1 . Setting m(t) = ky(t, t0 , x0 )ke(tt0 ) and using the

properties of g, we see that
!
Zt
m(t)
g(kx0 k)
m(s)
1 + (s)
g
ds, t0 t t1 .
kx0 k
kx0 k
kx0 k
t0
Hence, by Theorem 1.3.1,

1
m(t) kx0 kG
g(kx0 k)
G(1) +
kx0 k
Zt
t0
(s) ds ,
t0 t t1
93
Method of upper and lower solutions
which implies that

= ky(t1 , t0 , x0 )k kx0 ke(t1 t0 ) M (t0 ) < .
This contradiction shows that
ky(t, t0 , x0 )k M (t0 )kx0 ke(tt0 ) ,
t t0
which proves exponential asymptotic stability of the zero solution

of (2.1.2).
2.5
Method of Upper and Lower Solutions
One of the methods of proving existence of solutions of nonlinear

problems in a sector is the method of upper and lower solutions.
This method coupled with the monotone iterative technique offers a
flexible and effective mechanism for proving constructive existence
results. The upper and lower solutions that generate the sector serve
as bounds for solutions which can be improved by the monotone
iterative procedure. Moreover, the iterative schemes are also useful
to investigate qualitative properties of solutions. In this and the
following few sections, we shall systematically develop these ideas.
Consider the differential system
u = f (t, u),
u(0) = u0 ,
t J = [0, T ].
(2.5.1)
A function v C 1 [J, Rn ] is said to be a lower solution of (2.5.1) if

v f (t, v),
v(0) u0
and an upper solution of (2.5.1) if the inequalities above are reserved.

We shall now prove an existence result in a sector for (2.5.1).
Theorem 2.5.1 Let v, w C 1 [J, Rn ] be lower and upper solutions
of (2.5.1) such that v(t) w(t) on J and let f C[, Rn ], where
= {(t, u) J Rn : v(t) u w(t),
t J}.
If f is quasimonotone nondecreasing in u, then there exists a solution u(t) of (2.5.1) such that v(t) u(t) w(t) on J, provided
v(0) u(0) w(0).
94
In fact, the conclusion of Theorem 2.5.1 is true without demanding

f to satisfy the restrictive assumption of quasimonotonicity. However, we need to strengthen the notion of upper and lower solutions
of (2.5.1). We list below such lower and upper solutions as an assumption

v fi (t, ) for all such that
i
v(t) w(t) and vi (t) = i
(2.5.2)

w fi (t, ) for all such that
i
v(t) w(t) and wi (t) = i .
Theorem 2.5.2 Let v, w C 1 [J, Rn ] with v(t) w(t) on
J satisfying (2.5.2). Let f C[, Rn ]. Then there exists a solution u of (2.5.1) such that v(t) u(t) w(t) on J provided
v(0) u(0) w(0).
Since the assumptions of Theorem 2.5.1 imply the assumptions of
Theorem 2.5.2, it is enough to prove Theorem 2.5.2.
Proof of Theorem 2.5.2 Consider P : J Rn Rn dened by
Pi (t, u) = max{vi (t), min[ui , wi (t)]},
for each
i = 1, 2, . . . , n.
Then f (t, P (t, u)) denes a continuous extension of f to J Rn and

is bounded since f is bounded on . Therefore,
u = f (t, P (t, u))
has a solution u on J with u(0) = u0 . If we show that v(t) u(t)
w(t), we have u as a solution of (2.5.1), due to the denition of
P (t, u). For > 0 and e = (1, . . . , 1) consider
w (t) = w(t) + (1 + t)e
and
v (t) = v(t) (1 + t)e.
We have v (0) < u0 < w (0). Suppose that t1 J is such that
v (t) < u(t) < w (t)
on [0, t1 )
Method of upper and lower solutions
95
with uj (t1 ) = wj (t1 ). Then we have

v(t1 ) P (t1 , u(t1 )) w(t1 )
and
Pj (t1 , u(t1 )) = wj (t1 ).
Hence
wj (t1 ) fj (t1 , P (t1 , u(t1 ))) = uj (t1 ),
(t ), contradicting u (t) < w (t) for

which implies uj (t1 ) < wj
1
j
j
t1 < t. Therefore, v (t) < u(t) < w (t) on J. Now, letting 0
yields v(t) u(t) w(t) and the proof is complete.
Next we give an alternate proof of Theorem 2.5.2 using the idea

of dierential inequalities to show that any solution u(t) of (2.5.1)
lies in the sector provided that v(0) u0 w(0), assuming that f
satises a Lipschitz condition of the form
|fi (t, x) fi (t, y)| Li
|xi yi |.
i=1
Proof We shall rst assume that v, w in (2.5.2) satisfy strict inequalities and v(0) u(0) w(0). If the conclusion is false, there
exists a t1 > 0 and an i, 1 i n such that v(t1 ) u(t1 ) w(t1 )
and either vi (t1 ) = ui (t1 ) or ui (t1 ) wi (t1 ). Then,
fi (t1 , u(t1 )) = ui (t1 ) vi (t1 ) < fi (t1 , 1 , . . . , vi (t1 ), . . . , n )
= fi (t1 , u(t1 ))
or
fi (t1 , u(t1 )) = ui (t1 ) wi (t1 ) > fi (t1 , 1 , . . . , wi (t1 ), . . . , n )
= fi (t1 , u(t1 ))
which leads to a contradiction. In order to prove the conclusion in
case of nonstrict inequalities, consider
w
i (t) = wi (t) + e(n+1)Li t ,
vi (t) = vi (t) e(n+1)Li t ,
96
where > 0 is suciently small. Further let

Pi (t, u) = max{vi (t), min[ui , wi (t)]},
for each
i.
Then it is clear that if

is such that v(t)
w(t)
and
i = w
i (t),
it follows that = P (t,
) satises v(t) w(t) and i = wi (t).
Hence, using the Lipschitz condition it follows that
w
i (t) = wi (t) + (n + 1)Li e(n+1)Li t fi (t, ) + (n + 1)Li e(n+1)Li t
fi (t,
) + e(n+1)Li t > fi (t,
)
for all
such that v(t)
w(t)
and
i = w
i (t). Here we have
used the fact that
|
j Pj (t, )| e(n+1)Lj t ,
for each
j.
Similarly, we get vi < fi (t,

) for all
such that v(t)
w(t)
and
i = vi (t). Since v(0) < u0 < w(0),
we can conclude from the

previous argument that v(t) < u(t) < w(t),
t J. As is arbitrary,
the result follows letting 0.
2.6
Monotone Iterative Technique
In order to develop monotone iterative technique for the system

(2.5.1) so as to include several possibilities, we need to begin with
some new notions. For each xed i, 1 i n, let pi , qi be two
nonnegative integers such that pi + qi = n 1 so that we can express
the vector u as u = (ui , [u]pi , [u]qi ). Then, the system (2.5.1) can be
written as
(2.6.1)
ui = fi (t, ui , [u]pi , [u]qi ), u(0) = u0 .
Let v, w C 1 [J, Rn ]. Then v, w, are said to be coupled lower and
upper quasi solutions of (2.6.1) if
vi fi (t, vi , [v]pi , [w]qi ),
v(0) u0 .
(2.6.2)
wi fi (t, wi , [w]pi , [v]qi ),
w(0) u0 .
(2.6.3)
are satised. If, on the other hand, we have

vi = fi (t, vi , [v]pi , [w]qi ),
v(0) = u0
Monotone iterative technique
and
wi = fi (t, wi , [w]pi , [v]qi ),
97
w(0) = u0 ,
then v, w are said to be coupled quasi solutions of (2.6.1). One can

then dene coupled extremal quasi solutions of (2.6.1) accordingly.
A function f C[J Rn , Rn ] is said to possess a mixed quasimonotone property (mqmp for short) if for each i, fi (t, ui , [u]pi , [u]qi )
is monotone nondecreasing in [u]pi and monotone nonincreasing
in [u]qi .
Theorem 2.6.1 Let f C[J Rn , Rn ] possess mixed quasimonotone property and let v0 , w0 be coupled lower and upper quasi
solutions of the system (2.6.1) such that v0 w0 on J. Suppose
further that
fi (t, ui , [u]pi , [u]qi ) fi (t, ui , [u]pi , [u]qi ) Mi (ui ui )
(2.6.4)
whenever v0 w0 and v0i ui w0i and Mi 0. Then, there

exist monotone sequences {vn }, {wn } such that vn v, wn w
as n uniformly and monotonically to coupled minimal and
maximal quasi solutions of (2.6.1) on J provided v0 (0) u w0 (0).
Further, if u is any solution of (2.6.1) such that v0 u w0 , then
v u w on J.
Proof For any , C[J, Rn ] such that v0 , w0 on J, we
dene
Fi (t, ) = fi (t, i , []pi , []qi ) Mi (ui i )
and consider the uncoupled linear dierential system
ui = Fi (t, u),
u(0) = u0 .
(2.6.5)
Clearly, for a given , , the system (2.6.5) possesses a unique solution

u(t) dened on J.
For each , C[J, Rn ] such that v0 , w0 on J, we dene
the mapping A by
A[, ] = u
where u is the unique solution of (2.6.5). This mapping denes the
sequences {vn }, {wn }. First we prove that
98
(i) v0 A[v0 , w0 ] and w0 A[w0 , v0 ]; and

(ii) A possesses the mixed monotone property on the segment
[v0 , w0 ] where [v0 , w0 ] = {u C[J, Rn ] : v0 u w0 }.
To prove (i), set A[v0 , w0 ] = v1 where v1 is the unique solution of
(2.6.5) with = v0 , = w0 . Setting mi = v1i v0i , it easily follows
that

v0i
fi (t, v0i , [v0 ]pi , [w0 ]qi ) fi (t, v0i , [v0 ]pi , [w0 ]qi )
mi = v1i
mi (v1i v0i ) = Mi mi .
It thus follows that mi (t) mi (0)eMi t 0 on J and hence v0i v1i .
Similarly we can show w0i w1i which proves (i).
In order to prove (ii), let 1 , 2 , [v0 , w0 ] be such that
1 2 . Suppose A[1 , ] = u1 and A[2 , ] = u2 . Then, setting
mi = u1i u2i , we nd, using the mixed monotone property of f and
(2.6.4), that
mi = fi (t, 1i , [1 ]pi , []qi ) fi (t, 2i , [2 ]pi , []qi )
Mi (u1i 1i ) + Mi (u2i 2i ) Mi mi .
Also, since mi (0) = 0, we get u1i u2i . Similarly, if 1 , 2 ,
[v0 , w0 ] such that 1 2 , then, as before one can prove that
A[, 1 ] A[, 2 ]. It therefore follows that the mapping A satises
property (ii). Consequently this implies A[, ] A[, ] whenever
and , [v0 , w0 ]. In view of (i) and (ii) above, we can dene
the sequences
vn = A[vn1 , wn1 ],
wn = A[wn1 , vn1 ].
It is then easy to prove that the sequences {vn }, {wn } are monotonic
and converge uniformly and monotonically to coupled quasi-solutions
v, w of (2.6.1). Letting v = lim vn , w = lim wn , we nd
n
vi = fi (t, vi , [v]pi , [w]qi ),
v(0) = u0 ,
Monotone iterative technique
99
and
wi = fi (t, wi , [w]pi , [v]qi ),
w(0) = u0 .
We shall show that v, w, are coupled minimal and maximal quasisolutions respectively. Let u1 , u2 , be any coupled quasi-solutions of
(2.6.1) such that u1 , u2 [v0 , w0 ]. Let us assume that for some integer, k > 0, vk1 u1 , u2 wk1 on J. Then, setting mi = vki u1i
and employing the mqmp of f and (2.6.1), we arrive at
mi fi (t, vk1i , [u1 ]pi , [u2 ]qi ) fi (t, u1i , [u1 ]pi , [u2 ]qi )
Mi (vki vk1i ) Mi mi .
Since mi (0) 0, this implies that vk u1 . Using similar arguments
it is easy to conclude that vk u1 , u2 wk on J. It therefore follows
by induction that vn u1 , u2 wn on J for all n, since v0 u1 ,
u2 w0 on J. Hence, we have v u1 , u2 w on J showing that v, w
are coupled minimal and maximal quasi-solutions of (2.6.1). Since
any solution u of (2.6.1) such that u [v0 , w0 ] can be considered as
u, u coupled quasi-solutions of (2.6.1) we also have v u w on J.
This completes the proof.
We note rst that if qi = 0 so that pi = n 1 for each i, then
f is quasi-monotone nondecreasing and consequently Theorem 2.6.1
yields minimal and maximal solutions v, w of (2.6.1) respectively.
On the other hand, if pi = 0, qi = n 1 for each i, then f is quasimonotone nonincreasing and in this case, we have coupled extremal
quasi-solutions of extreme type. As it is, this theorem considers
a general class of quasi-solutions. Thus, it becomes an important
question to examine the relation between quasi-solutions and actual
solutions. It is not dicult to show that if f satises a Lipschitz
condition in the sector [v0 , w0 ], then a quasi solution is actually a
solution. In fact, we state this fact in the following corollary, whose
proof is immediate.
Corollary 2.6.1 In addition to the assumptions of Theorem
2.6.1, suppose that f (t, u) f (t, v) Lu v for u, v [v0 , w0 ],
then u = v = w is the unique solution of (2.6.1) such that v0 u
w0 on J.
100
2.7
Method of Mixed Monotony
The question whether it is possible to construct monotone sequences

that converge to the unique solution of (2.5.1) even when f does not
possess any monotone properties is natural and interesting. In this
section we develop a technique, which we call the method of mixed
monotony that investigates the foregoing question and is applicable
to a variety of nonlinear problems.
Suppose that there exist functions F C[J Rn Rn , Rn ],
, C 1 [J, Rn ] satisfying the following conditions:
(i) F (t, , ), F (t, , ) and (t) (t) on J;
(ii) F is mixed monotone, that is, F (t, u, v) is monotone nondecreasing in u for each (t, v) and monotone nonincreasing in v
for each (t, u);
(iii) F (t, u, u) = f (t, u) and
B(z z) F (t, y, z) F (t, y, z) B(y y)
whenever (t) y y (t), (t) z z (t), B being
an n by n matrix with nonnegative elements. In particular,
F (t, y, z) F (t, z, y) = B(y z)
holds.
Then we say that the dierential system (2.5.1) admits a process
of mixed monotonization. The procedure is called the method of
mixed monotony.
Let us begin with the following result.
Theorem 2.7.1 Assume that the system (2.5.1) admits a process of mixed monotonization. Then there exist monotone sequences
{n (t)}, {n (t)} which converge uniformly on J to the unique solution u(t) of (2.5.1) provided that (0) u0 (0). Furthermore,
we have
(t) 1 (t) n (t) u(t)
101
Method of mixed monotony
n (t) 1 (t) (t),
on
J.
Proof Let , C[J, Rn ], , [, ] = {u C[J, Rn ] : (t)

u(t) (t), t J} and Mi > 0 be any constants. Consider the
uncoupled linear initial value problem for each i,
ui = Fi (t, , ) + Mi (i ui ),
u(0) = u0 ,
(2.7.1)
with (0) u0 (0). Given any , , clearly there exists a unique

solution u of (2.7.1) on J. We dene a mapping A by A[, ] = u
and show that
(i) A[, ], A[, ];
(ii) A is a mixed monotone operator on [, ].
To prove (i), let P = 1 where 1 = A[, ]. Then
Pi Fi (t, , ) Fi (t, , ) Mi (i 1i ) = Mi Pi ,
and Pi (0) 0 imply that A[, ]. Similarly A[, ]. To
prove (ii), let 1 2 , 1 , 2 , C[J, Rn ] and 1 , 2 , [, ].
Setting u1 = A[1 , ], u2 = A[2 , ] and P = u1 u2 , we get
Pi = Fi (t, 1 , ) + Mi (1i u1i ) Fi (t, 2 , ) Mi (2i u2i )
Mi Pi ,
using the mixed monotony of F . Also, Pi (0) = 0. This implies that
A[1 , ] A[2 , ]. Similarly, one can prove A[, 1 ] A[, 2 ] if
1 2 and this proves (ii).
It is now easy to dene the sequences {n }, {n } with 0 = ,
0 = by
n+1 = A[n , n ], n+1 = A[n , n ],
and conclude that
1 n n 1 ,
on
J.
Using standard arguments, we can show that lim n = and

n
lim n = r exist uniformly on J and the functions , r satisfy
= F (t, , r),
r = F (t, r, ),
r(0) = (0) = u0 ,
t J.
102
Let P = r . Then, from (iii), we get

P = F (t, r, ) F (t, , r) = BP,
P (0) = 0.
Hence P (t) = P (0)eBt which implies r = = u on J. Since

F (t, u, u) = f (t, u), we see that u = = r is a solution of (2.5.1).
Furthermore, (iii) implies that
B(y y) f (t, y) f (t, y) B(y y),
whenever (t) y y (t). It is immediate that u is the unique
solution of (2.5.1). The proof is complete.
We shall next give sucient conditions for the system (2.5.1) to
admit the method of mixed monotony.
Theorem 2.7.2 Consider the system (2.5.1).
, C 1 [J, Rn ], (t) (t) on J such that
Suppose that
f (t, ) B( ),
f (t, ) + B( )
and
B(x y) f (t, x) f (t, y) B(x y),
whenever (t) y x (t), B being n by n matrix of nonnegative elements. Then the system (2.5.1) admits the method of mixed
monotony.
Proof We dene
1
F (t, y, z) = [f (t, y) + f (t, z) + B(y z)].
2
(2.7.2)
It is easy to see that F (t, y, z) is mixed monotone and note that

B(z z) F (t, y, z) F (t, y, z) B(y y),
(2.7.3)
whenever (t) y y (t) and (t) z z (t). In

particular, we have
F (t, y, z) F (t, z, y) = B(y z).
Lower and upper solutions and interval analysis
103
From (2.7.3), we see that

B( ) F (t, , ) f (t, )
which yields
f (t, ) B( ) F (t, , ).
Similarly, we can show that
F (t, , ).
Finally F (t, x, x) = f (t, x) follows from (2.7.2) by setting x = y = z
and the proof is complete.
2.8
Method of Lower and Upper Solutions and Interval

Analysis
Interval analytic methods, in which monotone inclusion property is

inherent, provide simultaneous iterative bounds for solutions of differential equations. In this section, we shall combine the method of
lower and upper solutions with the interval method to obtain existence of solutions (and sometimes solution sets) of an initial value
problem.
We begin by recalling the notation that is employed in Section
1.10. If U1 = [u1 , u1 ] and U2 = [u2 , u2 ] are intervals, then U1 U2 if
u2 u1 u1 u2 . Let p : C[J, R] C[J, R] be an integral operator
where J is an interval. The corresponding interval integral operator
P which is inclusive monotonic is dened by
P Y = {py : y Y }
where Y is an interval function. For our present discussion, we shall
use the following simple version of Theorem 1.10.1 which is a basic
result of interval analysis.
Theorem 2.8.1 Suppose that there exists an interval function
U0 such that P U0 U0 . Then the successive iterates given by
Un+1 = P Un form a decreasing sequence of interval valued functions
104
and lim Un (t) = U (t) exists as an interval function J. Moreover, if

n
u is any solution of pu = u, then u(t) U (t) on J.

Consider the initial value problem
u = f (t, u),
u(0) = u0 ,
(2.8.1)
where f C[J, R] and J = [0, T ]. Let us list the following assumptions for convenience:
(A0 ) , C 1 [J, R], (t) (t) on J and (t) f (t, ),
(t) f (t, ) with (0) u0 (0);
(A1 ) f (t, x) f (t, y) M (x y) whenever (t) y x (t)
and M 0;
(A2 ) , C 1 [J, R], (t) (t) on J with (0) u0 (0) and
f (t, ) M ( ), f (t, ) M ( ) for all such
that (t) (t), M > 0.
We note that whenever (A0 ) and (A1 ) hold, (A2 ) is satised.
Theorem 2.8.2 Assume either (A0 ) and (A1 ) or (A2 ). Then
there exists a decreasing sequence {Un (t)} of interval functions with
U0 (t) = [(t), (t)] such that lim Un (t) = U (t) exists and is an
n
interval function on J. If u(t) is any solution of (2.8.1) in the sector

[, ] = {u C[J, R] : u }, then u(t) U (t) on J. In case
(A0 ) and (A1 ) are satised, then U (t) = [(t), r(t)], where , r are
extremal solutions of (2.8.1) relative to [, ]. If further f satises
a Lipschitz condition relative to the sector, then U (t) is the unique
Proof Let us rst suppose that (A0 ) and (A1 ) hold. Consider the
integral operator p : C[J, R] C[J, R] dened by
M t
pu = u(0)e
t
+
eM (ts) F (s, u(s)) ds,
(2.8.2)
where F (t, u) = f (t, u) + M u. Because of (A1 ), F is increasing

in u whenever (t) u (t). For any interval valued function
Lower and upper solutions and interval analysis
105
Y [, ] we let P Y be the corresponding interval integral operator

dened by
P Y = {py : y Y }
which is inclusion monotonic. Since pu is increasing, we have
P Y = [py, py] if
Y = [y, y].
(2.8.3)
Because of (A0 ) and the fact that F is increasing, we get easily

p p and hence
P [, ] = [p, p] [, ].
Hence, by Theorem 2.8.1, we have, setting U0 = [, ], a
decreasing sequence {Un } of interval valued functions such that
lim Un (t) = U (t) exists and is an interval function on J. Fur-
thermore, if u(t) is any solution of (2.8.1) such that u [, ], then

u(t) U (t) on J.
Suppose that and r are minimal and maximal solutions of (2.8.1)
in the sector [, ]. Then, observing that any solution u of pu = u
is also a solution of (2.8.1) and vice versa, if U (t) = [u(t), u(t)], we
have [, u] [u, u]. Since P Y is continuous, we also have U = P U .
Hence by (2.8.3), it follows that pu = u and pu = u which imply
that u = f (t, u), u = f (t, u), u(0) = u(0) = u0 . Consequently,
[u, u] [, r] on J which implies that U = [, r].
Now suppose that (A2 ) holds. We consider the same operator p
dened in (2.8.2). Since F is no longer increasing, the relation (2.8.3)
need not hold. In order to apply Theorem 2.8.1, we need to show
that P [, ] [, ]. For this purpose, we shall show that for any
u [, ], pu [, ] which implies that P [, ] [, ].
Using (A2 ) and (2.8.2), we see that for all [, ],
t
pu
eM (ts) [F (s, ) F (s, u(s))] ds.
Choosing = u, it follows that pu. Similarly one can prove that

pu and hence P [, ] [, ].
106
If f (t, x) satises the Lipschitz condition

|f (t, x) f (t, y)| L|x y|,
(t) x,
y (t),
t J,
we get immediately u(t) = u(t) = u(t) as the unique solution of

(2.8.1) such that (t) u(t) (t) on J. The proof of theorem is
therefore complete.
We shall give an example to show that condition (A2 ) is realizable.
Example 2.8.1 Let f (t, u) = 1 + u u2 and consider (2.8.1). We
choose (t) 0, (t) = 2et 1 and M = 4eT 3. Then it is easy to
check that assumption (A2 ) is veried.
2.9
Integro-Dierential Equations
In this and the next section, we shall discuss some problems for
integro-dierential equations. We rst develop nonlinear variation
of parameters formula for integro-dierential equations. For this
purpose, we need to investigate the problem of continuity and differentiability of solutions of nonlinear integro-dierential equations
and obtain the relation between the derivatives. We then prove the
nonlinear variation of parameters formula for solutions of perturbed
integro-dierential equations.
We need the following result before we proceed further.
Theorem 2.9.1 Consider the initial value problem for linear
integro-dierential equations given by
t
x (t) = A(t)x(t) +
B(t, s)x(s) ds + F (t),
x(t0 ) = x0 ,
(2.9.1)
t0
2
where A(t), B(t, s) are continuous n by n matrices on R+ and R+
respectively and F C[R+ , Rn ]. Then, the unique solution of (2.9.1)
is given by
t
x(t) = R(t, t0 )x0 +
R(t, s)F (s) ds,

t0
t t0 ,
(2.9.2)
107
Integro-differential equations
where R(t, s) is the solution of the initial value problem

t
R(t, s)
+ R(t, s)A(s) +
s
R(t, )B(, s) d = 0,
(2.9.3)
R(t, t) = I for 0 s t < , I being the identity matrix.

Proof Let R(t, s) be the solution of (2.9.3) and set v(s) =
R(t, s)x(s), where x(t) = x(t, t0 , x0 ) is the solution of (2.9.1) on
[t0 , ). Then, we have

s
R(t, s)
x(s) + R(t, s) A(s)x(s) + B(s, u)x(u) du + F (s) .
v (s) =
s
t0
Integrating from t0 to t, this yields

t
R(t, t)x(t) R(t, t0 )x0 =

R(t, s)
+ R(t, s)A(s) x(s) ds
s
t0
s
t
R(t, s)
+
t0
t
B(s, u)x(u) du ds +
t0
R(t, s)F (s) ds.

t0
By Fubinis theorem, we get

t
s
t
t

R(t, u)B(u, s) du x(s) ds.
R(t, s) B(s, u)x(u) du ds =
t0
t0
t0
Hence we obtain
t
R(t, t)x(t) R(t, t0 )x0 =
R(t, s)
+ R(t, s)A(s)
s
t0
t
+
t
R(t, u)B(u, s) du x(s) ds +

s
R(t, s)F (s) ds.

t0
108
Since R(t, s) satises (2.9.3), the desired relation (2.9.2) follows immediately and the proof is complete.
Now, we shall discuss the problem of continuity and dierentiability of solutions x(t, t0 , x0 ) of the initial value problem
t
x (t) = f (t, x(t)) +
g(t, s, x(s)) ds,
x(t0 ) = x0
(2.9.4)
t0
with respect to (t0 , x0 ).

2
Theorem 2.9.2 Assume that f C[R+ Rn , Rn ], g C[R+
n
2
fx and gx exist and are continuous on R+ R , R+ Rn
respectively. Let x(t, t0 , x0 ) be the unique solution of (2.9.2) existing
on some interval t0 t < a < and set J = [t0 , t0 + T ], t0 + T < a.
Then
Rn , Rn ],
(i) (t, t0 , x0 ) =
x(t, t0 , x0 )
exists and is the solution of
x0
t
y (t) = H(t, t0 , x0 )y(t) +
G(t, s; t0 , x0 )y(s) ds,
(2.9.5)
t0
such that (t0 , t0 , x0 ) = I, where

H(t, t0 , x0 ) = fx (t, x(t, t0 , x0 )),
G(t, s; t0 , x0 ) = gx (t, s, x(s, t0 , x0 ));
(ii) (t, t0 , x0 ) =
x(t, t0 , x0 )
t0
z (t) = H(t, t0 , x0 )z(t)

t
+ G(t, s; t0 , x0 )z(s) ds g(t, t0 , x0 ),
t0
such that (t0 , t0 , x0 ) = f (t0 , x0 );
(2.9.6)
109
(iii) the functions (t, t0 , x0 ), (t, t0 , x0 ) satisfy the relation

(t, t0 , x0 ) + (t, t0 , x0 )f (t0 , x0 )
t
+ R(t, ; t0 , x0 )g(, t0 , x0 ) d = 0,
(2.9.7)
t0
where R(t, s; t0 , x0 ) is the solution of the initial value problem

R(t, s; t0 , x0 )
+ R(t, s; t0 , x0 )H(s, t0 , x0 )
s
t
+ R(t, ; t0 , x0 )G(, s; t0 , x0 ) d = 0,
(2.9.8)
R(t, t; t0 , x0 ) = I on the interval t0 s t and R(t, t0 ; t0 , x0 )

= (t, t0 , x0 ).
Proof Under the assumptions on f, g it is clear that solution
x(t, t0 , x0 ) exists, is unique and continuous in (t, t0 , x0 ) on some interval. Consequently the functions H and G are continuous in (t, t0 , x0 )
and therefore, the solutions of linear initial value problems (2.9.5) and
(2.9.6) exist and are unique on the same interval for which x(t, t0 , x0 )
is dened.
Let ek = (e1k , . . . , enk ) be the vector such that ejk = 0 if k = j
and ekk = 1. Then, for small h, x(t, h) = x(t, t0 , x0 + ek h) is dened on J and lim x(t, h) = x(t, t0 , x0 ) uniformly on J. Setting
h0
x(t) = x(t, t0 , x0 ), it follows by the mean value theorem that

d
[x(t, h) x(t)] =
dt
1
fx (t, sx(t, h) + (1 s)x(t)) ds
0
t 1
(x(t, h) x(t)) +
gx (t, s, x(s, h)
t0 0
+ (1 )x(s)) d(x(s, h) x(s)) ds.
110
If xh (t) =
x(t, h) x(t)
, h = 0, we see that the existence of
h
x(t, t0 , x0 )
is equivalent to the existence of lim xh (t). Since
h0
x0
x(t0 , h) = x0 + ek h, xh (t0 ) = ek , it is clear that xh (t) is the solution
of the initial value problem
y (t) = H(t, t0 , x0 , h)y(t)
t
+ G(t, s, t0 , x0 , h)y(s) ds,
y(t0 ) = ek ,
(2.9.9)
t0
where
1
fx (t, sx(t, h) + (1 s)x(t)) ds,
H(t, t0 , x0 , h) =
0
and
1
gx (t, s, x(s, h) + (1 )x(s)) d.
G(t, s, t0 , x0 , h) =
0
Since lim x(t, h) = x(t) uniformly on J, continuity of fx , gx imh0
plies that
lim H(t, t0 , x0 , h) = H(t, t0 , x0 )
h0
and
lim G(t, s, t0 , x0 , h) = G(t, s; t0 , x0 )
h0
uniformly on J and J J respectively. Consider (2.9.9) as a family of

initial value problems depending on a parameter, where H and G are
continuous, h being small. Since the solutions of (2.9.9) are unique,
the general solution of (2.9.9) is a continuous function of h for xed
(t, t0 , x0 ). In particular, lim xh (t) exists and is the solution of (2.9.5)
on J. This implies that
such
that
h0
x(t,t0 ,x0 )
x0
exists and is the solution of (2.9.5)
0 ,t0 ,x0 )
= I. Furthermore, it is easy to see from (2.9.5)
that x(tx
0
x(t,t0 ,x0 )
is also continuous with respect to its arguments.
x0
111
To prove (ii), dening x(t, h) = x(t, t0 + h, x0 ), we have, by the

mean value theorem,
d
(x(t, h) x(t)) =
dt
1
fx (t, sx(t, h) + (1 s)x(t)) ds(x(t, h) x(t))
0
t 1
gx (t, s, x(s, h) + (1 )x(s)) d(x(s, h) x(s)) ds
+
t0 +h 0
t
0 +h
g(t, s, x(s)) ds.

t0
x(t, h) x(t)
, h = 0, it is obvious that
h
xh (t) is the solution of the initial value problem
Setting as before, xh (t) =
y (t) = H(t, t0 , x0 , h)y(t)

t
G(t, s, t0 , x0 , h)y(s) ds L(t, t0 , x0 , h),
+
t0 +h
such that y(t0 + h) = a(h), where

1
L(t, t0 , x0 , h) =
h
t
0 +h
g(t, s, x(s)) ds
t0
and
1
a(h) =
h
t
0 +h
1
f (s, x(s)) ds
h
t0
t
0 +hs
g(s, , x()) d ds.

t0
t0
Noting that lim L(t, t0 , x0 , h) = g(t, t0 , x0 ), lim a(h) = f (t0 , x0 )

h0
h0
and using an argument similar to the proof of (i), we see that

x(t, t0 , x0 )
exists, is continuous in its argument and is the solution
t0
of (2.9.6).
112
Finally, to prove (iii), we observe that (t, t0 , x0 ) and (t, t0 , x0 )

are the solutions of (2.9.5) and (2.9.6) respectively and hence, by
Theorem 2.9.1, we get the desired relation (2.9.7). The proof of
theorem is therefore complete.
We are now in position to prove the nonlinear variation of parameters formula for solutions y(t, t0 , x0 ) of the perturbed integro
dierential equation
t
y (t) = f (t, y(t)) +
g(t, s, y(s)) ds
t0
+ F (t, y(t), (Sy)(t)),
(2.9.10)
y(t0 ) = x0 .
Theorem 2.9.3 Suppose that the hypotheses of Theorem 2.9.2

hold. Assume further that F C[R+ Rn Rn , Rn ] and S(y)(t) =
t
2 Rn , Rn ]. If x(t, t , x ) is the
K(t, s, y(s)) ds with K C[R+
0 0
t0
solution of (2.9.4) existing on J, any solution y(t, t0 , x0 ) of (2.9.10)

existing on J satises the integral equation
y(t, t0 , x0 ) = x(t, t0 , x0 )
t
+ (t, s, y(s))F (s, y(s), (Sy)(s)) ds
(2.9.11)
t0
t
t
[(t, , y()) R(t, ; s, y(s))]g(, s, y(s)) d ds
+
t0 s
for t t0 , where y(t) = y(t, t0 , x0 ) and R(t, s; t0 , x0 ) is the solution

of the initial value problem (2.9.8).
Proof Setting p(s) = x(t, s, y(s)) where y(s) = y(s, t0 , x0 ), we
have
p (s) = (t, s, y(s)) + (t, s, y(s))y (s).
Substituting for y (s), integrating from t0 to t and using Fubinis
113
theorem, we get
t
p(t) p(t0 )
t
(t, s, y(s))F (s, y(s), (Sy)(s)) ds

t0
(t, s, y(s)) + (t, s, y(s))f (s, y(s))
=
t0
t
(t, , y())g(, s, y(s)) d ds.
+
s
Now, the relation (2.9.7) together with the fact x(t, t, y(t)) =
y(t, t0 , x0 ) yields (2.9.11), completing the proof.
In the special case when f (t, x) = A(t)x and g(t, s, x) = B(t, s)x
in (2.9.4) where A, B are continuous n by n matrices, we have
0 ,x0 )
= R(t, t0 ), where
x(t, t0 , x0 ) = R(t, t0 )x0 and (t, t0 , x0 ) = x(t,t
x0
R(t, s) is the solution of (2.9.8) such that R(t, t) = I. Consequently it
is easy to see that the relation (2.9.11) reduces to the linear variation
of parameters formula given by
t
y(t, t0 , x0 ) = R(t, t0 )x0 +
R(t, s)F (s, y(s), (Sy)(s)) ds,
t t0 ,
t0
since (t, , y()) = R(t, , s, y()) being independent of y and s.

If g(t, s, x) 0, the relation (2.9.11) reduces to the usual Alekseevs formula (2.1.10). However, when g = 0 or is not linear in x,
(t, t0 , x0 ) = R(t, s, t0 , x0 ) and hence, the relation (2.9.11) cannot
have a simpler form that corresponds to Alekseevs formula as one
would like to expect.
Finally, we shall give an application of nonlinear variation of parameters formula to discuss the qualitative behavior of the perturbed
equation (2.9.10). For this purpose, we require the following result
which is interesting in itself.
Theorem 2.9.4 Let the assumptions of Theorem 2.9.2 hold. Suppose that x(t, t0 , x0 ) and x(t, t0 , x0 ) are any two solutions of (2.9.4)
114
existing on J. Then, for t J, we have

x(t, t0 , x0 ) x(t, t0 , y0 )
1
(t, t0 , sx0 + (1 s)y0 ) ds(x0 y0 ).
(2.9.12)
Proof We set p(s) = x(t, t0 , x0 s + (1 s)y0 ), for 0 s 1, so

that
x
p (s) =
(t, t0 , x0 s + (1 s)y0 )(x0 y0 ).
x0
Integrating this expression from 0 to 1 and noting that p(1) =
x(t, t0 , x0 ) and p(0) = x(t, t0 , y0 ), we obtain the stated result (2.9.11).
(i) the assumptions of Theorem 2.9.2 hold and f (t, 0) = g(t, s, 0)
0;
(ii) there exist positive numbers , M and such that if x0 ,
we have
R(t, s; t0 , x0 ) M e(ts) ,
(t, t0 , x0 ) M e(tt0 ) ,
for t0 s t; and
(iii) g(t, s, x) Kxe2t whenever x and
t
F (t, x, (Sy)(t)) (t)x +
(t, s)y(s) ds,

t0
2 , R ] and x, y(s) .

where C[R+ , R+ ], C[R+
+
Then, any solution y(t, t0 , x0 ) of (2.9.10) with x0 small enough

satises the estimate
(tt0 )+
y(t, t0 , x0 ) M x0 e
t
t0
(s,t0 ) ds
t t0 ,
115
where
2M Ke2t
+M
(t, t0 ) = M (t) +
t
(t, s)e(ts) ds,
t0
and
L1 [t
0 , ).
Proof Let x(t) = x(t, t0 , x0 ) and y(t) = y(t, t0 , x0 ) be solutions of

(2.9.4) and (2.9.10) respectively , existing on [t0 , ). Then, in view
of (ii), it follows from (2.9.12) that
x(t, t0 , x0 ) M x0 e(tt0 ) ,
t t0 .
For those values of t t0 for which y(t) , assumptions (ii), (iii)

and relation (2.9.11) yield
(tt0 )
y(t) M x0 e
t
+M
(s)e(ts) y(s) ds
t0
t
+M
e(ts)
t0
t
s
s
(s, )y() d ds
t0
2M Ke(t) e2 y(s) d ds.
t0 t0
Using Fubinis theorem and dening V (t) = y(t)et , we get

t
V (t) M x0 e
t0
t
0 (s)V (s) ds
t0
t
(2.9.13)
M e(s) (, s) d V (s) ds,
+
s
t0
2M K 2t
e
. Setting the right hand side of
(2.9.13) equal to p(t), we get, noting p(t) is nondecreasing,

t

(ts)
p (t) 0 (t) + M e
(t, s) ds p(t), p(t0 ) = M x0 et ,
where 0 (t) = M (t) +
t0
116
which yields
(tt0 )+
y(t) M x0 e
provided x0
is complete.
2.10
t
t0
(s,t0 ) ds
for
t t0 ,
p N (t0 )
e
, where (s, t0 ) ds N (t0 ). The proof
M
t0
Stability in Variation
One of the important techniques for investigating the qualitative behavior of nonlinear systems is to utilize the corresponding linear variational systems. To bring out the inherent rich behavior of linear
integro-dierential systems, we shall rst discuss a method of nding an equivalent linear dierential system and then exploiting this
method, we investigate the stability properties of nonlinear integrodierential system by means of the corresponding variational system.
Let us continue to consider the system (2.9.4) and suppose that
fx , gx exist and are continuous and f (t, 0) = g(t, s, 0) 0. Let x(t) =
x(t, t0 , x0 ) be the unique solution of (2.9.4) existing on [t0 , ) and
let us consider the variational systems
t
y (t) = fx (t, x(t))y(t) +
gx (t, s, x(s))y(s) ds,

t0
(2.10.1)
y(t0 ) = x0 ,
and
t
z (t) = fx (t, 0)z(t) +
gx (t, s, 0)z(s) ds,
z(t0 ) = x0 .
(2.10.2)
t0
Theorem 2.10.1 Assume that there exists an n by n continuous

2 such that L (t, s) exists, is continuous
matrix function L(t, s) on R+
s
and satises
t
K(t, s) + Ls (t, s) + L(t, s)A(s) +
L(t, u)K(u, s) du = 0 (2.10.3)

s
117
Stability in variation
2
where A(t), K(t, s) are continuous n by n matrices on R+ and R+
respectively. Then, the initial value problem for linear integrodierential system
t
u (t) = A(t)u(t) +
K(t, s)u(s) ds + F (t),
u(t0 ) = x0 , (2.10.4)
t0
where F C[R+ , Rn ], is equivalent to the initial value problem for

linear dierential system
v (t) = B(t)v(t) + L(t, t0 )x0 + H(t),
v(t0 ) = x0 ,
where B(t) = A(t) L(t, t) and H(t) = F (t) +
t
(2.10.5)
L(t, s)F (s) ds.
t0
Proof Let u(t) be any solution of (2.10.4) existing on [t0 , ). Set

p(s) = L(t, s)u(s) so that p (s) = Ls (t, s)u(s) + L(t, s)u (s). Substituting for u (s) in (2.10.4) and integrating, we get
t
L(t, t)u(t) L(t, t0 )x0 =
[Ls (t, s) + L(t, s)A(s)]u(s) ds

t0
t
s
L(t, s)
+
t0

K(s, )u() d ds +
t0
Since, by Fubinis theorem,

s

t
L(t, s)
t0
t
t0
t t
K(s, )u() d ds =
t0
L(t, s)F (s) ds.
L(t, u)K(u, s) du u(s) ds,

t0
it follows, using (2.10.3) and (2.10.4) that

t
Ls (t, s) + L(t, s)A(s)
L(t, t)u(t) L(t, t0 )x0 =
t0
t
L(t, u)K(u, s) du u(s) ds + H(t) F (t)
+
s
118
t
=
K(t, s)u(s) ds + H(t) F (t) = u (t) + A(t)u(t) + H(t).
t0
Hence, u(t) satises (2.10.5).

Let v(t) be any solution of (2.10.5) existing on [t0 , ). Then,
dening
t
z(t) = v (t) A(t)v(t)
K(t, s)v(s) ds F (t),

t0
we shall show that z(t) 0 which proves that v(t) satises (2.10.4).
Now, substituting for v (t) from (2.10.5) and using (2.10.3) together
with Fubinis theorem, we get

z(t) = L(t, t)v(t) L(t, t0 )x0
t
+
t
Ls (t, s)v(s) ds
t0
s
L(t, s) A(s)v(s) + F (s) +

t0
K(s, u)v(u) du ds.

t0
d
Since (L(t, s)v(s)) = Ls (t, s)v(s) + L(t, s)v (s), we have, by intedt
gration,
t
L(t, t)v(t) L(t, t0 )x0 =
[Ls (t, s)v(s) + L(t, s)v (s)] ds.
t0
It therefore follows, using (2.10.4), that

t
z(t) =
t0
s

L(t, s) v (s) + A(s)v(s)
t
K(s, u)v(u) du + F (s) ds =
+
t0
L(t, s)z(s) ds,

t0
119
which implies z(t) 0 because of uniqueness of solutions of Volterra

linear integral equations. The proof is therefore complete.
Let us now dene the concept of stability in variation for the
equation (2.9.4) and prove a simple result.
Denition 2.10.1 The zero solution of (2.9.4) is said to possess stability properties in variation if the zero solution of (2.10.2)
possesses the corresponding stability properties.
Theorem 2.10.2 Suppose that the assumption of Theorem 2.10.1
holds. Then the stability properties of the function
p(t, t0 , x0 ) = (t, t0 )x0
t
+ (t, s)L(s, t0 )x0 ds,
(2.10.6)
t0
where (t, t0 ) is the fundamental matrix solution of

v = B(t)v
(2.10.7)
imply the corresponding stability properties of (2.9.4) in variation.

Proof Setting A(t) = fx (t, 0), K(t, s) = gx (t, s, 0) and F (t) 0,
we see, in view of Theorem 2.10.1, that it is enough to consider the
initial value problem (2.10.5) with H(t) 0 in order to deduce information concerning the variational system (2.10.2). By the variation
of parameters formula applied to the initial value problem
v (t) = B(t)v + L(t, t0 )x0 ,
we have
v(t0 ) = x0 ,
t
v(t) = (t, t0 )x0 +
(t, s)L(s, t0 )x0 ds.

t0
It then follows easily that

v(t) p(t, t0 , x0 ),
t t0 ,
where p(t, t0 , x0 ) is given by (2.10.6). The conclusion of theorem

is therefore immediate.
120
In order to discuss the stability properties of zero solution of

(2.9.4), it is enough to consider the stability properties of the zero solution of the variational system (2.10.1) in view of Theorem 2.9.4. To
see this, note that any
solution
v(t) of (2.10.1) such that v(t0 ) = x0
x(t,t0 ,x0 )
is given by v(t) =
x0 . Furthermore, if x(t, t0 , x0 ) is any
x0
solution of (2.9.4), then, because of Theorem 2.9.4, we have
1
x(t, t0 , x0 ) =
0
x(t, t0 , x0 s)
ds x0 .
x0
(2.10.8)
Hence, if v(t) , t t0 , whenever x0 (), then we get

x(t, t0 , x0 ) .
We shall therefore concentrate on system (2.10.1), which can be
written as
t
y (t) = A(t)y(t) +
K(t, s)y(s) ds + F (t, y),

t0
(2.10.9)
y(t0 ) = x0 ,
where
F (t, y) = [fx (t, x(t)) fx (t, 0)]y(t)
t
+ [gx (t, s, x(s)) gx (t, s, 0)]y(s) ds.
t0
Then, because of Theorem 2.10.1, it is sucient to investigate the

stability properties of the zero solution of
v (t) = B(t)v(t) + L(t, t0 )x0 + H(t, v),
v(t0 ) = x0 ,
(2.10.10)
where
t
H(t, v) = F (t, v) +
L(t, s)F (s, v(s)) ds.

t0
(2.10.11)
121
We shall assume that

fx (t, x) fx (t, 0) (t),
C[R+ , R+ ]
(2.10.12)
and
2
C[R+
, R+ ]
gx (t, s, x) gx (t, s, 0) (t, s),
(2.10.13)
hold whenever x for some > 0. For convenience, let us dene
the following functions
t
L(t, )(, s) ds
(t, s) = (t, s) + (s)L(t, s) +

s
and
t
(t, )(, s) d .
d(t, s) = (t, s) (s) +

s
Now, consider the linear integral equation

t
r(t) = p(t, t0 , x0 ) +
d(t, s)r(s) ds,
(2.10.14)
t0
whose solution r(t) is given by

t
r(t) = p(t, t0 , x0 )
R(t, s)p(s, t0 , x0 ) ds,

t0
where R(t, s) is the resolvent kernel satisfying

t
R(t, s) = d(t, s) +
R(t, ) d(, s) d.
s
Now, we can prove the following result.

Theorem 2.10.3 Suppose that the assumption of Theorem 2.10.1
holds together with (2.10.12) and (2.10.13). Then, any solution v(t)
of (2.10.10) satises the estimate
v(t) r(t),
t t0 ,
(2.10.15)
122
provided x(t) for t t0 , where r(t) = r(t, t0 , x0 ) is the

Proof Let v(t) be any solution of (2.10.10). Then, we have
v(t) = (t, t0 )x0
t
+ (t, s)[L(s, t0 )x0 + H(s, v(s))] ds.
(2.10.16)
t0
Since
F (t, v) fx (t, x(t)) fx (t, 0)v(t)
t
+ gx (t, s, x(s)) gx (t, x, 0)v(s) ds
t0
we get, using Fubinis theorem, (2.10.11), (2.10.12) and (2.10.13),

t
H(t, v) (t)v(t) +
(t, s)v(s) ds.

t0
Employing Fubinis theorem again, it follows that

t
t
(t, s)H(s, v(s)) ds
t0
(t, s) (s)

t0
t
(t, )(, s) d v(s) ds.
+
s
Consequently, we obtain from (2.10.11), (2.10.6) and the denition

of d(t, s),
t
v(t) p(t, t0 , x0 ) +
d(t, s)v(s) ds,

t0
t t0 .
Difference equations
123
It therefore follows, from Corollary 1.6.2, that

v(t) r(t, t0 , x0 ),
t t0 ,
r(t, t0 , x0 ) being the linear integral equation (2.10.14). The proof
is therefore complete.
Since (2.10.10) and (2.10.9) are equivalent by Theorem 2.10.1 and
(2.10.9) is nothing but a restatement of (2.10.1), it is clear from the
conclusion of Theorem 2.10.3 that the stability properties of (2.10.14)
imply the corresponding stability properties of (2.10.1) and therefore,
imply the stability properties of (2.9.4).
2.11
Dierence Equations
In this section, we shall rst discuss the variation of parameters formula for dierence equations and then, utilizing the theory of dierence inequalities, investigate the stability properties of solutions of
dierence equations.
Consider the dierence equation
yn+1 = A(n)yn + f (n, yn ),
y n0 = y 0 ,
(2.11.1)
Where An is an N by N nonsingular matrix and f : Nn+0 RN RN

and Nn+0 = {n0 , n0 + 1, . . . , n0 = k, . . .}.
Theorem 2.11.1 The solution y(n, n0 , y0 ) of (2.11.1) satises
the equation
yn = (n, n0 )y0 +
n1

(n, j + 1)f (j, yj )
(2.11.2)
j=n0
where (n, n0 ) is the fundamental matrix of the equation

xn+1 = A(n)xn .
(2.11.3)
Proof Let
y(n, n0 , y0 ) = (n, n0 )xn ,
xn 0 = y 0 .
(2.11.4)
124
Then, substituting in (2.11.1), we get

(n + 1, n0 )xn+1 = A(n)(n, n0 )xn + f (n, yn )
from which we see that
xn = 1 (n + 1, n0 )f (n, yn )
and
xn =
n1

(n0 , j + 1)f (j, yj ) + xn0 .
j=n0
From (2.11.4), it then follows that

y(n, n0 , y0 ) = (n, n0 )y0 +
n1

(n, j + 1)f (j, yj ).
(2.11.5)
j=n0
Consider now the dierence equation

xn+1 = f (n, xn ).
(2.11.6)
Lemma 2.11.1 Assume that f : Nn+0 Rs Rs and f possesses

partial derivatives on Nn+0 Rs . Let the solution x(n) x(n, n0 , x0 )
of (2.11.6) exist for n n0 and let
H(n, n0 , x0 ) =
f (n, x(n, n0 , x0 ))
.
x
(2.11.7)
x(n, n0 , x0 )
x0
(2.11.8)
Then,
(n, n0 , x0 ) =
(n + 1, n0 , x0 ) = H(n, n0 , x0 )(n, n0 , x0 ),
(n0 , n0 , x0 ) = 1.
(2.11.9)
(2.11.10)
125
Proof By dierentiating (2.11.6) with respect to x0 , we have

xn+1
f xn
=
x0
x x0
which coincides with (2.11.8). Then (2.11.9) follows from the denition of .
We are now able to generalize Theorem 2.11.1 to the equation
yn+1 = f (n, yn ) + F (n, yn ).
(2.11.11)
f
exist
x
Rs . If x(n, n0 , x0 ) is the
Theorem 2.11.2 Let f, F : Nn+0 Rs Rs and let
and be continuous and invertible on Nn+0

solution of
xn+1 = f (n, xn ), xn0 = x0 ,
(2.11.12)
then any solution of (2.11.11) satises the equation

y(n, n0 , x0 ) = x(n, n0 , x0
+
n1

1 (j + 1, n0 , vj , vj+1 )F (j, yj ))
(2.11.13)
j=n0
where
1
(n, n0 , svj + (1 s)vj1 ) ds
(n, n0 , vj , vj+1 ) =
0
and vj satises the implicit equation

(n + 1, n0 , vn , vn+1 )(vn+1 vn ) = F (n, yn ).
(2.11.14)
Proof Let us put y(n, n0 , x0 ) = x(n, n0 , vn ) and v0 = x0 . Then

y(n + 1, n0 , x0 ) = x(n + 1, x0 , vn+1 ) x(n + 1, n0 , vn )
+ x(n + 1, n0 , vn )
= f (n, x(n, n0 , vn )) + F (n, x(n, n0 , vn ))
126
from which we get

x(n + 1, n0 , vn+1 ) x(n + 1, n0 , vn ) = F (n, yn ).
Applying the mean value theorem, we have
1
0
x(n + 1, n0 , svn+1 + (1 s)vn )

ds(vn+1 vn ) = F (n, yn )
x0
and hence by (2.11.7),

t
(n + 1, n0 , svn+1 + (1 s)vn ) ds(vn+1 vn ) = F (n, yn )
0
which is equivalent to (2.11.14). It now follows that

vn+1 vn = 1 (n + 1, n0 , vn , vn+1 )F (n, yn )
and
vn = v0 +
n1

1 (j + 1, n0 , vj , vj+1 )F (j, yj )
(2.11.15)
j=n0
from which the conclusion results.

Corollary 2.11.1 Under the hypothesis of Theorem 2.11.2, the
solution y(n, n0 , x0 ) can be written in the following form
y(n, n0 , x0 ) = x(n, n0 , x0 )
+ (n, n0 , vn , x0 )
n1

1 (j + 1, n0 , vj , vj+1 )F (j, yj )
(2.11.16)
j=n0
Proof Applying the mean value theorem to (2.11.13) once more,

the result follows.
Corollary 2.11.2 If f (n, x) = A(n)x, then (2.11.16) reduces to
(2.11.2).
127
Proof In this case, xn = (n, n0 )x0 and (n, n0 , x0 ) (n, n0 )

(n, n0 , vn , vn+1 ) and therefore, we have
vn+1 vn = 1 (n + 1, n0 , vn , vn+1 )F (n, yn )
and
vn = x 0 +
n1

1 (j + 1, n0 , vj , vj+1 )F (j, yj )
j=n0
from which follows the claim.

Having developed the necessary variation of parameters formula,
we shall next consider qualitative behavior of solutions of dierence
equations. As in the case of dierential equations, comparison results
play an important role in this investigation. Let us begin with the
following simple result.
Theorem 2.11.3 Let g(n, u) be a nonnegative function, nondecreasing in u. Suppose that
(i) f, F : Nn+0 S S , > 0 where S = {x RN : x < };
(ii) f (n, 0) 0 and g(n, 0) 0;
(iii) f (n, u) g(n, u).
Then, the stability of the trivial solution of the equation
un+1 = g(n, un )
(2.11.17)
implies the stability of the trivial solution of (2.11.6).

Proof From (2.11.6) we have
yn+1 f (n, yn ) g(n, yn ),
and hence the comparison equation is (2.11.17). Theorem 1.9.1 can
be applied, provided that y0 u0 , to get yn un for n n0 .
Suppose now that the zero solution of (2.11.17) is stable. Then, for
> 0, there exists a (, n0 ) such that for |u0 | < we have |u0 | < .
This means the stability of the trivial solution of (2.11.6).
128
Since g 0, the trivial solution of the comparison equation

(2.11.17) may not be, for example, asymptotically stable and therefore we cannot conclude from Theorem 2.11.3 that the trivial solution
of (2.11.6) is also asymptotically stable. This is due to the strong
assumption (iii) which can be replaced by the following condition
(iv) f (n, y) y + w(n, y)
where g(n, u) = u + w(n, u) is nondecreasing in u. Noting that
w in (iv) need not be positive and hence the trivial solution of
un+1 = g(n, un ) can have dierent stability properties, we can conclude from Theorem 2.11.3, with (iii) replaced by (iv), that the stability properties of the trivial solution of (2.11.17) imply the corresponding stability properties of the trivial solution of (2.11.6). This
version of Theorem 2.11.3 is more useful and we shall denote it as
Theorem 2.11.3*, for which the proof needs minor modications.
From Theorems 2.11.3 and 2.11.3*, we can easily obtain several
important variants. Before we do that, we need the following result
which characterizes the stability properties of solutions of (2.11.3) in
terms of fundamental matrix solution.
Lemma 2.11.2 Let (n, n0 ) be the fundamental matrix solution
of (2.11.3). Then the trivial solution of (2.11.3) is
(i) uniformly stable, if there exists an M > 0 such that
(n, n0 ) M for n n0 ;
(ii) uniformly asymptotically stable, if there exist constants > 0,
0 < < 1 such that (n, n0 ) nn0 for n n0 .
Proof To prove (i), since xn = (n, n0 )x0 , we get
xn (n, n0 )x0 M x0
. This proves suciency.

and hence x0 < for n n0 , if x0 < M
To prove necessity, uniform stability implies (n, n0 ) < 1, n n0
if x0 < . Taking y0 = xx00 , we see that sup (n, n0 )y0 is
y0 =1
bounded and since this is just the denition of the norm of , we are
done.
129
The proof of suciency for (ii) is similar as in (i). The proof of

necessity follows by considering that if there is uniform asymptotic
stability, then xing > 0, there exists a > 0, N () > 0 such that
x0 < implies
xn (n, n0 )x0 < ,
n n0 + N ().
As before, it is easy to see that (n, n0 )x0 < , n n0 + N (),

where can be chosen suciently small. Moreover, uniform stability
implies that (n, n0 ) for n n0 for some . We then have
for n [n0 + mN (), n0 + (m + 1)N ()],
(n, n0 ) (n, n0 + mN ())
(n0 + mN (), n0 + (m 1)N ()) . . .
(n0 + N (), n0 ) < n
and this proves the lemma.
This result shows that for linear dierence equation, asymptotic
stability implies exponential asymptotic stability.
Theorem 2.11.4 Let (n, n0 ) be the fundamental matrix of the
linear equation (2.11.3). Let F : Nn+0 RN RN and
1 (n + 1, n0 )F (n, (n, n0 )yn ) g(n, yn )
(2.11.18)
where the function g(n, u) is nondecreasing in u. Assume that the

solutions un of
(2.11.19)
un+1 = un + g(n, un )
are bounded for n n0 . Then the stability properties of the linear
equation (2.11.3) imply the corresponding stability properties of the
null solution of
(2.11.20)
xn+1 = A(n)xn + F (n, xn ).
Proof The linear transformation xn = (n, n0 )yn reduces

(2.11.20) to
yn+1 = yn + 1 (n + 1, n0 )F (n, (n, n0 )yn ).
130
We then have
yn+1 yn + g(n, yn ).
If y0 u0 , we obtain yn un where un is the solution of (2.11.9).
It then follows that
xn (n, n0 )yn (n, n0 )un .
If the solution of the linear system is, for example, uniformly
asymptotically stable, then from Lemma 2.11.2, we see that
(n, n0 ) L nn0 ,
for some suitable L > 0 and 0 < < 1. Then
xn L nn0 un
and this shows that the solution x = 0 is uniformly asymptotically
stable because un is bounded. The proof of other cases is similar.
We shall merely state another important variant of Theorem
2.11.3* which is widely used in numerical analysis.
Theorem 2.11.5 Given the dierence equation
yn+1 = yn + hA(n)yn + f (n, yn )
(2.11.21)
where h is a positive constant, suppose that

(i) f (n, 0) 0 for n n0 and
(ii) g(n, u) is a nondecreasing function in u, g(n, 0) 0 and
f (n, yn ) g(n, y).
Then, the stability properties of the null solution of
un+1 = I + hA(n)un + g(n, un )
(2.11.22)
imply the corresponding stability properties of the null solution of

(2.11.21).
131
In this form, the foregoing theorem is used in the estimation of

the growth of errors in numerical methods for dierential equations.
Usually, instead of (2.11.22), one uses the comparison equation
un+1 = (I + hA)un + g(n, un )
which is less useful because (I + hA) > 1. The form (2.11.22) is
more interesting because when the eigenvalues of A are all negative,
I + hA can be less than 1. This will happen, for example, if the
logarithmic norm
I + hA(n) 1
h0
h
(A(n)) = lim
is less than zero. From the denition, it follows that

I + hA(n) = 1 + h(A(n)) + o(h2 A(n)).
Letting
(A) = h[(A(n)) + o(h2 A(n))],
the comparison equation becomes

)un + g(n, un ).
un+1 = (1 + u
We shall now use the variation of parameters formula (2.11.2) and
consider stability by rst approximation.
f (n, yn ) gn (yn )
where gn are positive and
(2.11.23)
gn < . Then, if the zero solution of
n=n0
(2.11.3) is uniformly stable (or uniformly asymptotically stable),then

the zero solution of (2.11.1) is uniformly stable (or uniformly asymptotically stable).
Proof By (2.11.2) we get
yn = (n, n0 )yn +
n1

j=n0
(n, j + 1)f (j, yj )
132
By Lemma 2.11.2, we then have, using (2.11.23),

yn M y0 + M
n1

gj (yj ).
j=n0
Now Corollary 1.9.1 yields the estimate

n1

yn M y0 exp M

gj
j=n0
from which follows the proof, provided that y0 is small enough such
that
n1

gj < .
M y0 exp M
j=n0
In the case of uniform asymptotic stability, it follows that for

n > N , (n, n0 )x0 < , the previous inequality can be written as

yn exp M
gj
n0
from which we conclude that lim yn = 0.

Corollary 2.11.3 If the matrix A is constant such that the
solutions of (2.11.3) are bounded, then the solutions of the equation
yn+1 = (A + B(n))yn
are bounded provided that

n=n0
B(n) < .

f (n, yn ) Lyn
(2.11.24)
where L > 0 is suciently small and the solution xn = 0 of (2.11.3)

is uniformly asymptotically stable. Then the solution yn = 0 of
(2.11.1) is exponentially asymptotically stable.
Proof By Lemma 2.11.2, we have
(n, n0 ) H nn0 ,
H > 0,
0 < < 1,
133
Notes
and because of (2.11.24), we get

yn H
nn0
y0 + LH
n1

n1
1 yj .
j=n0
By introducing the new variable pn = 1 yn , we see that

pn H n0 y0 + LH 1
n1

pj .
j=n0
Using Corollary 1.9.1 again, we arrive at

pn H
n0
= H
n0
y0
n1

(1 + LH 1 )
j=n0
y0 (1 + LH 1 )nn0 ,
which implies
yn Hy0 ( + LH)nn0 .
1
, the conclusion follows.
H
Corollary 2.11.4 Consider the equation
If + LH < 1, that is, L <
yn+1 = Ayn + f (n, yn )
(2.11.25)
where A has all the eigenvalues inside the unit disk and moreover,
f (n, y)
= 0,
y0
y
lim
uniformly with respect to n. Then the zero solution of (2.11.25) is

exponentially asymptotically stable.
2.12
Notes
Theorem 2.1.1 is from Lord and Mitchell [1] and Theorem 2.1.3 is
due to Alekseev [1]. Theorems 2.1.2 and 2.1.4 are standard, see
Lakshmikantham and Leela [1]. Theorem 2.1.5 is new.
134
For the development leading to Theorem 2.2.1 in Section 2.2 see

the survey paper of Martynyuk, Leela and Mitropolsky [1]. The remaining results of Section 2.2 are adapted from Lord and Mitchell
[1], Lakshmikantham and Leela [1] and Martynyuk [15]. The contents of Section 2.3 may be found in Lakshmikantham and Leela [2]
where more general results in a Banach space are given. The results of Section 2.4 follow the works of Lord and Mitchell [1] and
Lakshmikantham and Leela [1]. The material contained in Sections
2.5 and 2.6 is taken from Ladde, Lakshmikantham and Vatsala [1]
while the results of Section 2.7 are due to Khavanin and Lakshmikantham [1]. Section 2.8 consists of the work of Lakshmikantham and
Sivasundarum [1]. The contents of Section 2.9 are taken from Hu,
Lakshmikantham and Rama Mohan Rao [1]. Theorem 2.10.1 is due
to Rama Mohan Rao and Srinivas [1] while the remaining results of
Section 2.10 are based on the work of Lakshmikantham and Rama
Mohan Rao [1]. For the results of Section 2.11 see Lakshmikantham
and Trigiante [1].
3
STABILITY OF MOTION IN TERMS OF TWO
MEASURES
3.0
Introduction
This chapter is devoted essentially to the investigation of stability

theory in terms of two measures. It also provides a converse theorem for uniform asymptotic stability in this set up and stresses the
importance of employing families of Lyapunov-like functions and the
theory of inequalities.
In Section 3.1 we formulate the basic comparison results in terms
of Lyapunov-like functions and the theory of dierential inequalities
that are necessary for our later discussion. Section 3.2 denes stability concepts in terms of two dierent measures and demonstrates how
such concepts enable us to unify a variety of stability notions found
in the literature. Section 3.3 oers sucient conditions for various
types of stability in terms of two measures. This section also introduces families of Lyapunov functions to consider uniform stability
properties where one can relax the usual assumptions. In this general setup, generalizations of Marachkovs theorem are considered.
In Section 3.4, we prove a converse theorem for uniform asymptotic
stability in terms of two measures which includes the well-known
Masseras converse theorem. Due to the interplay between the two
measures, proving such a result poses diculties and to construct a
smooth Lyapunov function one has to pay a reasonable price. Interestingly, the converse theorem considered yields, in particular, the
converse theorem for partial uniform asymptotic stability and thus
provides a result that is exible enough to warrant its use in several

135
136
Stability of Motion in Terms of Two Measures
directions. We indicate in Section 3.5 the criteria for boundedness

and Lagrange stability in terms of two measures.
In order to avoid the complications that arise in employing two
measures and to keep the exposition simpler, we restrict ourselves
in the rest of the chapter to a simple measure. We devote Section
3.6 to stability results for autonomous or periodic systems and consider the modifications needed when the derivative of the Lyapunov
function satisfies weaker assumptions. Section 3.7 deals with perturbing families of Lyapunov functions where it is shown that the
Lyapunov function found in a given problem does not satisfy the desired requirements, it is fruitful to perturb it rather than discard it.
Furthermore, this section demonstrates the usefulness of employing
families of Lyapunov functions. In fact, we actually construct a family of Lyapunov functions to prove asymptotic stability when the set
on which the derivative of the Lyapunov function is zero is a larger
set containing the origin.
Section 3.8 introduces the notion of M0 -stability which describes
a very general type of invariant set and its stability behavior. This
notion includes eventual stability and the stability of asymptotically
invariant sets. The advantage of using several Lyapunov functions
rather than a single one is demonstrated in Section 3.9, where we
discuss the method of vector and matrix-valued Lyapunov functions
indicating the fruitfulness of employing several Lyapunov functions.
Finally in Section 3.10, we develop the method of cone valued Lyapunov functions. It is shown that choosing an appropriate cone other
n is more useful and provides a flexible
than the standard cone R+
mechanism suitable for applications such as large scale dynamical
systems.
3.1
Basic Comparison Results
Consider the differential system

x = f (t, x),
x(t0 ) = x0 ,
t0 R+ ,
(3.1.1)
where f C[R+ Rn , Rn ]. For any Lyapunov-like function

V C[R+ Rn , Rn ], we define the function
D + V (t, x) = lim sup

h0+
1
[V (t + h, x + hf (t, x)) V (t, x)]
h
137
(3.1.2)
for (t, x) R+ Rn . Occasionally, we shall denote (3.1.2) by

D + V (t, x)(3.1.1) to emphasize the definition of D + V with respect to
the system (3.1.1). One could also use other generalized derivative,
for example
D V (t, x) = lim inf
h0
1
[V (t + h, x + hf (t, x)) V (t, x)].
h
(3.1.3)
We note that if V C 1 [R+ S(, R+ )], then D + V = V (t, x) where

V (t, x) = Vy (t, x) + Vx (t, x)f (t, x).
We can now formulate the basic comparison result in terms of
Lyapunov function V .
Theorem 3.1.1 Let V C[R+ Rn , R+ ] and be locally Lipschitzian in x. Assume that the function D + V (t, x) satisfies
D + V (t, x) g(t, V (t, x)),
(t, x) R+ Rn ,
(3.1.4)
2 , R]. Let r(t) = r(t, t , u ) be the maximal solution

where g C[R+
0 0
of the scalar differential equation
u = g(t, u),
u(t0 ) = u0 0,
(3.1.5)
existing for t t0 . Then, V (t0 , x0 ) u0 implies

V (t, x(t) r(t)),
t t0 ,
(3.1.6)
where x(t) = x(t, t0 , x0 ) is any solution of (3.1.1) existing for t t0 .

Proof. Let x(t) be any solution of (3.1.1) existing for t t0 such
that V (t0 , x0 ) u0 . Define m(t) = V (t, x(t)). For sufficiently small
h > 0, we have
m(t + h) m(t) = V (t + h, x(t + h))
V (t + h, x(t) + hf (t, x(t)))
+ V (t + h, x(t) + hf (t, x(t))) V (t, x(t)).
138
Since V (t, x) is locally Lipschitzian in x, we get, using (3.1.4), the

dierential inequality
D + m(t) g(t, m(t)),
m(t0 ) u0 ,
and Theorem 1.5.2 now gives the desired result (3.1.6).

Corollary 3.1.1 If, in Theorem 3.1.1, we suppose that d(t, u)
0, then V (t, x(t)) is nonincreasing in t and V (t, x(t)) V (t0 , x0 ),
t t0 .
The next result plays an important role whenever we use several
Lyapunov functions.
Theorem 3.1.2 Let V C[R+ Rn , RN ] and be locally Lipschitzian in x. Assume that
D + V (t, x) g(t, V (t, x)),
(t, x) R+ Rn ,
where g C[R+ Rn , RN ] and g(t, u) is quasimonotone nondecreasing in u. Let r(t) = r(t, t0 , u0 ) be the maximal solution of
u = g(t, u),
u(t0 ) = u0 0,
(3.1.7)
existing for t t0 and x(t) = x(t, t0 , x0 ) be any solution of (3.1.1)

existing for t t0 . Then V (t0 , x0 ) u0 implies
V (t, x(t)) r(t),
t t0 .
We recall that inequalities between vectors are componentwise inequalities and quasimonotonicity of g(t, u) means that u v, ui vi ,
1 i N implies gi (t, u) gi (t, v). Theorem 3.1.2 is a special case
of the following result which deals with cone-valued Lyapunov functions.
Theorem 3.1.3 Assume that V C[r+ Rn , K] and V (t, x)
is locally Lipschitzian in x relative to the cone K RN and for
(t, x) R+ Rn ,
K
D + V (t, x) g(t, V (t, x)).
139
Let g C[R+ K, RN ], g(t, u) be quasimonotone nondecreasing in

u with respect to K and r(t) = r(t, t0 , u) be the maximal solution of
(3.1.7) existing for t t0 . Then, any solution x(t) = x(t, t0 , x0 ) of
(3.1.1) existing for t t0 satises the estimate
K
V (t, x(t)) r(t),
t t0 ,
provided V (t0 , x0 ) u0 .
Proof Proceeding as in Theorem 3.1.1 with suitable modications
we arrive at the dierential inequality
K
D + m(t) g(t, m(t)),
m(t0 ) u0 ,
t t0 .
Then, Theorem 1.5.5 yields the stated result.

The next theorem which is a variant of Theorem 3.1.3 is more
exible in applications. We merely state the result since its proof
follows from Corollary 1.5.1.
Theorem 3.1.4 Let P and Q be two cones in RN such that
R Q. Suppose that V C[R+ Rn , Q], V (t, x) satises a local
Lipschitz condition relative to P and
P
D + V (t, x) g(t, V (t, x)),
(t, x) R+ Rn .
Assume that g C[R+ Q, RN ] and g(t, u) is quasimonotone nondecreasing in u relative to P and x(t) = x(t, t0 , x0 ) is any solution of
P
(3.1.1) existing for t t0 such that V (t0 , x0 ) u0 . Then,

Q
V (t, x(t)) r(t),
t t0 ,
(3.1.8)
where r(t) = r(t, t0 , u0 ) is the maximal solution of (3.1.7) relative to

N , then (3.1.8)implies the componenwise
P . In particular, if Q = R+
estimate V (t, x(t) r(t)), t t0 .
In all the comparison results discussed so far, we have estimated
(t, x) by a function of t and V (t, x) only. However, in some
situations, it is more natural to estimate D + V (t, x) by a function of
D+ V
140
t, x and V (t, x). The following result is in that direction whose proof
is similar to the proof of Theorem 3.1.1.
Theorem 3.1.5 Let V C[R+ Rn , R+ ] and V (t, x) be locally
Lipschitzian in x. Assume that g C[R+ Rn R+ , R] and for
(t, x) R+ Rn ,
D + V (t, x) g(t, x, V (t, x)).
Let x(t) = x(t, t0 , x0 ) be any solution of (3.1.1) existing on [t0 , )
and t(t, t0 , x0 , u0 ) be the maximal solution of
u = g(t, x(t), u),
u(t0 ) = u0 0,
existing for t t0 . Then V (t0 , x0 ) u0 implies

V (t, x(t)) r(t, t0 , x0 , u0 ),
3.2
t t0 .
Stability Concepts in Terms of Two Measures
Let us begin by dening the following classes of functions for future

use
K = {a C[R+ , R+ ] : a(u) is strictly increasing in u and
a(0) = 0},
L = { C[R+ , R+ ] : (u) is strictly decreasing in u and
lim (u) = 0},
2
, R+ ] : a(t, s) K for each s and a(t, s) L for
KL = {a C[R+
each t},
2
, R+ ] : a(t, s) K for each t},
CK = {a C[R+
= {h C[R+ Rn , R+ ] : inf h(t, x) = 0},

0 = {h : inf h(t, x) = 0 for each t R+ }.
x
We shall now dene the stability concepts for the system (3.1.1)
in terms of two measures h0 , h .
Denition 3.2.1 The dierential system (3.1.1) is (3.1.1) is said
to be
Stability Concepts in Terms of Two Measures
141
(S1 ) (h0 , h)-equistable, if for each > 0 and t0 R+ , there exists a

function = (t0 , ) > 0 which is continuous in t0 for each
such that
h0 (t0 , x0 ) < implies h(t, x(t)) < ,
t t0 ,
where x(t) = x(t, t0 , x0 ) is any solution of (3.1.1);

(S2 ) (h0 , h)-uniformly stable, if (S1 ) holds with being independent
of t0 ;
(S3 ) (h0 , h)-quasi-equi-asymptotically stable, if for each > 0 and
t0 R+ , there exist positive numbers 0 = 0 (t0 ) and
T = T (t0 , ) such that
h0 (t0 , x0 ) < 0 implies h(t, x(t)) < ,
t t0 + T ;
(S4 ) (h0 , h)-quasi-uniform asymptotically stable, if (S3 ) holds with

0 and T being independent of t0 ;
(S3 ) (h0 , h)-quasi-equi-asymptotically stable, if for any > 0 and
> 0, t0 R+ , there exists a positive number T = T (t0 , , )
such that
h0 (t0 , x0 ) < implies h(t, x(t)) < ,
t t0 + T ;
(S4 ) (h0 , h)-quasi-uniform asymptotically stable, if (S3 ) holds with

T being independent of t0 ;
(S5 ) (h0 , h)-asymptotically stable, if (S1 ) holds and given t0 R+ ,
there exists a 0 = 0 (t0 ) > 0 such that
h0 (t0 , x0 ) < 0 implies
lim h(t, x(t) = 0);
(S6 ) (h0 , h)-equi asymptotically stable, if (S1 ) and (S3 ) hold together;
(S7 ) (h0 , h)-uniformly asymptotically stable, if (S2 ) and (S4 ) hold
simultaneously;
(S8 ) (h0 , h)-unstable, if (S1 ) fails to hold.
142
Sometimes the notion of quasi asymptotic stability may be relaxed

somewhat as in (S3 ) and (S4 ). We shall use these notions to define
Lagrange stability in Section 3.5.
A few choices of the two measures (h0 , h) given below will demonstrate the generality of the Definition 3.2.1. Furthermore, (h0 , h)stability concept enable us to unify a variety of stability notions
found in the literature. It is easy to see that Definition 3.2.1 reduces
to
(1) the well known stability of the trivial solution x(t) 0 of
(3.1.1) or equivalently, of the invariant set {0}, if h(t, x) =
h0 (t, x) = kxk;
(2) the stability of the prescribed motion x0 (t) of (3.1.1) if
h(t, x) = h0 (t, x) = kx x 0 (t)k;
(3) the partial stability of the trivial solution of (3.1.1) if h(t, x) =
kxks , 1 s n and h0 (t, x) = kxk;
(4) the stability of asymptotically invariant set {0}, if h(t, x) =
h0 (t, x) = kxk + (t), where L;
(5) the stability of the invariant set A Rn if h(t, x) = h0 (t, x) =
d(x, A), where d(x, A) is the distance of x from the set A;
(6) the stability of conditionally invariant set B with respect to A,
where A B Rn , if h(t, x) = d(x, B) and h0 (t, x) = d(x, A).
We recall that the set {0} is said to be asymptotically invariant
if given > 0, there exists a () > 0 such that x0 = 0 implies
kx(t, t0 , x0 )k < for t t0 (). Several other combinations
of choices are possible for h0 , h in addition to those given in (1)
through (6).
Definition 3.2.2 Let h0 , h . Then, we say that
(i) h0 is finer than h if there exists a > 0 and a function CK
such that h0 (t, x) < implies h(t, x) (t, h0 (t, x));
(ii) h0 is uniformly finer than h if in (i) is independent of t;
Stability Criteria in Terms of Two Measures
143
(iii) h0 is asymptotically ner than h if there exists a > 0

and a function KL such that h0 (t, x) < implies
h(t, x) (h0 (t, x), t).
Denition 3.2.3 Let V C[R+ Rn , R+ ]. Then V is said to
be
(i) h-positive denite if there exists a > 0 and a function b K
such that b(h(t, x)) V (t, x) whenever h(t, x) < ;
(ii) h-decrescent if there exists a > 0 and a function a K such
that V (t, x) a(h(t, x)) whenever h(t, x) < ;
(iii) h-weakly decrescent if there exists a > 0 and a function
a CK such that V (t, x) a(t, h(t, x)) whenever h(t, x) < ;
(iv) h-asymptotically decrescent if there exists a > 0 and a
function a KL such that V (t, x) a(h(t, x), t) whenever
h(t, x) < .
Corresponding to Denition 3.2.1, we need the stability denition
for the trivial solution of the comparison equation
u = g(t, u),
u(t0 ) = u0 0,
(3.2.1)
2 , R] and g(t, 0) 0. We merely state one of the

where g C[R+
concepts.
Denition 3.2.4 The trivial solution u(t) 0 of (3.2.1) is

said to be equistable if for any > 0 and t0 R+ , there exists a
= (t0 , ) > 0 which is continuous in t0 such that u0 < implies
u(t, t0 , u0 ) < , t t0 , u(t, t0 , x0 ) being any solution of (3.2.1).
3.3
Let us now establish some sucient conditions for the (h0 , h)-stability properties of the dierential system (3.1.1).
(A0 ) h, h0 and h0 is uniformly ner than h;
144
(A1 ) V C[R+ Rn , R+ ], V (t, x) is locally Lipschitzian in x, V is

h-positive denite and h0 -decrescent;
2 , R] and g(t, 0) 0;
(A2 ) g C[R+
(A3 ) D + V (t, x) g(t, V (t, x)) for (t, x) S(h, ) for some > 0,
where S(f, ) = {(t, x) R+ Rn : h(t, x) < }.
Then, the stability properties of the trivial solution of (3.1.5) imply
the corresponding (h0 , h)-properties of (3.1.1).
Proof We shall only prove (h0 , h)-equiasymptotic stability of
(3.1.1). For this purpose, let us rst prove (h0 , h)-equistability.
Since V is h-positive denite, there exists a (0, ] and b K
such that
b(h(t, x)) V (t, x), (t, x) S(h, ).
(3.3.1)
Let 0 < < and t0 R+ be given and suppose that the trivial
solution of (3.1.5) is equistable. Then, given b() > 0 and t0 R+ ,
there exists a function 1 = 1 (t0 , ) that is continuous in t0 such
that
u0 < 1 implies u(t, t0 , u0 ) < b(),
t t0 ,
(3.3.2)
where u(t, t0 , u0 ) is any solution of (3.1.5). We choose u0 = V (t0 , x0 ).

Since V is h0 -decrescent and h0 is uniformly ner than h, there exists
a 0 > 0 and a function a K such that for (t0 , x0 ) S(h0 , 0 ),
h(t0 , x0 ) < and V (t0 , x0 ) a(h0 (t0 , x0 )).
(3.3.3)
It then follows from (3.3.1) that

b(h(t0 , x0 )) V (t0 , x0 ) a(h0 (t0 , x0 )),
(t0 , x0 ) S(h0 , 0 ).
(3.3.4)
Choose = (t0 , ) such that (0, 0 ], a() < 1 and let
h0 (t0 , x0 ) < . Then (3.3.4) shows that h(t0 , x0 ) < since 1 < b().
We claim that h(t, x(t)) < , t t0 whenever h0 (t0 , x0 ) < , where
x(t) = x(t, t0 , x0 ) is any solution of (3.1.1) with h0 (t0 , x0 ) < . If
this is not true, then there exists a t1 > t0 and a solution x(t) of
(3.1.1) such that
h(t1 , x(t1 )) = and
h(t, x(t)) < ,
t0 t t1 ,
(3.3.5)
145
in view of the fact that h(t0 , x0 ) < whenever h0 (t0 , x0 ) < . This
means that x(t) S(h, ) for [t0 , t1 ] and hence by Theorem 3.1.1, we
have
V (t, x(t)) r(t, t0 , u0 ), t0 t t1 ,
(3.3.6)
where r(t, t0 , x0 ) is the maximal solution of (3.1.5). Now the relations
(3.3.1), (3.3.2), (3.3.5) and (3.3.6) yield
b() V (t1 , x(t1 )) r(t1 , t0 , x0 ) < b(),
a contradiction proving (h0 , h)-equistability of (3.1.1).
Suppose next that the trivial solution of (3.1.5) is quasi-equi
asymptotically stable. From the (h0 , h)-equistability, we set = so
that 0 = 0 (t0 , ). Now let 0 < < . Then, by quasi-equi asymptotic stability of (3.1.5), we have that, given b() > 0 and t0 R+ ,
there exist positive numbers 1 = 1 (t0 ) and T = T (t0 , ) > 0 such
that
u0 < 1
implies
u(t, t0 , u0 ) < b(),
t t0 + T.
(3.3.7)
Choosing u0 = V (t0 , x0 ) as before, we nd a 0 = 0 (t0 ) > 0 such that

0 (0, 0 ] and a(0 ) < 1 . Let 0 = min(0 , 0 ) and h0 (t0 , x0 ) < 0 .
This implies that h(t, x(t)) < , t t0 and hence the estimate (3.3.6)
is valid for all t t0 . Suppose now that there exists a sequence {tk },
tk t0 + T , tk as k such that h(tk , x(tk )) where x(t)
is any solution of (3.1.1) such that h0 (t0 , x0 ) < 0 . This leads to a
contradiction
b() V (tk , x(tk )) r(tk , t0 , u0 ) < b()
because of (3.3.6) and (3.3.7). Hence the system (3.1.1) is (h0 , h)equi asymptotically stable and the proof is complete.
We have assumed in Theorem 3.3.1 stronger requirements on V ,
h, h0 only to unify all the stability criteria in one theorem. This
obviously puts burden on the comparison equation (3.1.5). However,
to obtain only non-uniform stability criteria, we could weaken certain
assumptions of Theorem 3.1.1 as in the next result. The details of
proof are omitted.
Theorem 3.3.2 Assume that conditions (A0 ) (A3 ) hold with
the following changes:
146
(i) h0 , h 0 and h0 is ner than h and

(ii) V is h0 -weakly decrescent.
Then, the equi or uniform stability properties of the trivial solution
of (3.1.5) imply the corresponding equi (h0 , h)-stability properties of
(3.1.1).
We shall next consider a result on (h0 , h)-asymptotic stability
which generalizes classical results.
(i) h0 , h 0 and h0 is ner than h;
(ii) V C[R+ Rn , R+ ], V (t, x) is locally Lipschitzian in x, V is
h-positive denite and h0 -weakly decrescent;
(iii) W C[R+ Rn , R+ ], W (t, x) is locally Lipschitzian in x,
W is h-positive denite, D + W (t, x) is bounded from above or
from below on S(h, ) and for (t, x) S(h, ), D V (t, x)
C(W (t, x)), C K.
Then, the system (3.1.1) is (h0 , h)-asymptotically stable.
Proof By Theorem 3.3.2 with g 0, it follows that the system
(3.1.1) is (h0 , h)-equistable. Hence it is enough to prove that given
t0 R+ , there exists a 0 = 0 (t0 ) > 0 such that h0 (t0 , x0 ) < 0
implies h(t, x(t)) as t .
For = , let 0 = 0 (t0 , ) be associated with (h0 , h)-equistability. We suppose that h0 (t0 , x0 ) < 0 . Since W (t, x) is hpositive denite, it is enough to prove that lim W (t, x(t)) = 0 for
t
any solution x(t) of (3.1.1) with h0 (t0 , x0 ) < 0 . We rst note

that lim inf W (t, x(t)) = 0. For otherwise, in view of (iii), we get
t
V (t, x(t)) as t .
Suppose that lim W (t, x(t)) = 0. Then, for any > 0, there exist
t
divergent sequences {tn }, {tn } such that ti < ti < ti+1 , i = 1, 2, . . . ,

and
W (ti , x(ti )) = , W (ti , x(ti )) = , and

2
(3.3.8)
< W (t, x(t)) < , t (tn , tn ).

2
147
Of course, we could also have, instead of (3.3.8),

W (ti , x(ti )) = ,
W (ti , x(ti ))
= , W (t, x(t))
2
, . (3.3.9)
2
Suppose that D + W (t, x) M . Then, it is easy to obtain, using
(3.3.8), the relation ti ti >

. In view of (iii), we have for
2M
large n,
0 V (tn , x(tn )) V (t0 , x0 ) +
ti

D + V (s, x(s)) ds
1in t

< 0,
V (t0 , x0 ) nC
2 2M
which is a contradiction. Thus, W (t, x(t)) 0 as t and hence
h(t, x(t)) 0 as t . The argument is similar when D + W
is bounded from below and we use (3.3.9). The proof is therefore
complete.
Corollary 3.3.1 (Marachkovs theorem) Suppose that f is
bounded on R+ S(). Then the trivial solution of (3.1.1) is asymptotically stable if there exist C K and V C[R+ S(), R+ ] such
that
(i) V is positive denite, V (t, 0) 0 and V (t, x) is locally Lipschitzian in x;
(ii) D + V (t, x) C(x), (t, x) R+ S(), C K.
Proof Take h0 = h = x, W = x and note that |D + W (t, x)|
f (t, x). Then, all the hypotheses of Theorem 3.3.3 are satised and
the proof is complete.
Corollary 3.3.2 The positive deniteness of V in Corollary
3.3.1 can be weakened to positive semi deniteness of V , that is,
V (t, 0) 0 and V (t, x) 0 on R+ S(). Then the conclusion of
Corollary 3.3.1 holds.
Proof It is enough to prove that x = 0 is stable. Let 0 < <
and t0 R+ be given. Let f (t, x) M on R+ S(). Choose
148
= (t0 , ) so that (0, /2) and V (t0 , x0 ) < C(/2) 2M

for
x0 < . Let x0 < and suppose that t1 , t2 > t0 such that
x(t2 ) = ,
x(t1 ) =
and x(t) < ,

2
t [t0 , t1 ).
Then, by (ii), we have

0 V (t2 , x(t2 )) V (t0 , x0 ) + V (t2 , x(t2 )) V (t1 , x(t1 ))

(t2 t1 )
V (t0 , x0 ) C
2

<0
V (t0 , x0 ) C
2 2M
which is a contradiction. Hence the proof is complete.
Corollary 3.3.3 Suppose that there exist a, C K and
V C[R+ S(), R+ ] such that
(i) V (t, x) is locally Lipschitzian in x, V (t, x) 0 and V is positive
denite;
(ii) D + V (t, x) C(V (t, x)), (t, x) R+ S().
Then x = 0 is asymptotically stable.
Proof Take W = V and h = h0 = x in Theorem 3.3.3.
If we dene only uniform stability properties, we can relax the
assumptions of Theorems 3.3.1 and 3.3.2 by employing a family of
Lyapunov functions each of which satises less restrictive conditions.
Theorem 3.3.4 Assume (A0 ) and (A2 ) of Theorem 3.3.1. Suppose further (A1 ) for each (0, ), > 0, there exists a oneparameter family of functions V C[S(h, ) S c (h0 , ), R+ ] such
that V is locally Lipschitzian in x, V is h-positive denite and h0 decrescent, where S c (h0 , ) is the complement of S(h0 , );
(A3 ) D V (t, x) g(t, V (t, x)) for (t, x) S(h, ) S c (h0 , ).
149
Then, the uniform stability of the trivial solution of (3.1.5) implies

(h0 , h)-uniform stability of (3.1.1).
Proof Suppose that the trivial solution of (3.1.5) is uniformly
stable. Because of (A0 ) and (A1 ), there exists a (0, ] such that
the relations (3.3.1) and (3.3.4) hold for (t, x) S(h, ) S c (h0 , )
and 0 is independent of t0 . Also, (3.3.2) holds with 1 independent
of t0 . We choose = () > 0 such that (0, 0 ] and a() < 1 .
We let h0 (t0 , x0 ) < and note that h(t0 , x0 ) < as before. If (h0 , h)uniform stability of (3.1.1) does not hold, then there will exist a
solution x(t) of (3.1.1) and t1 , t2 > t0 such that

h0 (t1 , x(t1 )) = , h(t2 , x(t2 )) = and
(3.3.10)
for t [t1 , t2 ].
x(t) S(h, ) S c (h0 , )
Hence, choosing = and using Theorem 3.1.1, we have
V (t, x(t)) r(t, t1 , u0 ),
t [t1 , t2 ],
(3.3.11)
where r(t, t1 , u0 ) is the maximal solution of (3.1.5) through (t1 , u0 ).

The relations (3.3.1), (3.3.2), (3.3.10) and (3.3.11) lead us to the
contradiction
b() V (t2 , x(t2 )) r(t2 , t1 , u0 ) < b(),
proving (h0 , h)-uniform stability of (3.3.1).
Theorem 3.3.5 Assume that A0 holds and suppose further that
for each 0 < < , there exist > 0 and V C[s(h, )
S c (h0 , ), R+ ], V (t, x) is locally Lipschitzian in x, V is bounded
and
D + V (t, x)
for
(t, x) S(h, ) S c (h0 , ).
Then, if the system (3.3.1) is (h0 , h)-uniformly stable, it is (h0 , h)uniformly asymptotically stable
Proof Assume that (3.3.1) is (h0 , h)-uniformly stable. Then taking = , we set 0 = (). Let t0 R+ and h0 (t0 , x0 ) < 0 . Clearly
h(t, x(t)) < , t t0 , where x(t) is any solution of (3.3.1). Let
150
0 < < and = () corresponding to (h0 , h)-uniform stability.

Let us prove that there exists a t0 > t0 such that h0 (t0 , x(t0 )) < ().
If there is no such t0 , we would have h0 (t, x(t)), t t0 . Then,
using the assumptions of Theorem, it follows that
V (t, x(t)) V (t0 , x0 ) (t t0 ),
t t0 ,
and hence V (t, x(t)) as t . This contradicts

the boundedness of V and therefore there exists a t0 satisfying
h0 (t0 , x(t0 )) < (). We therefore have h(t, x(t)) < , t t0 . Now
let |V | M for (t, x) S(h, ) S c (h0 , ) and we choose t0 so that
2M
. We can thus conclude that the system (3.3.1) is
t0 = t0 =
(h0 , h)-quasi uniform asymptotic stable completing the proof.
3.4
A Converse Theorem in Terms of Two Measures
The importance of uniform asymptotic stability in the investigation

of stability properties of perturbed dierential equations need no emphasis. The converse theorem of Massera which results from uniform
asymptotic stability of the origin has been widely utilized in perturbation theory. A converse theorem of Massera type for uniform
asymptotic stability in terms of two measures poses diculties due
to the interplay between two measures and hence it is not possible
to construct a smooth Lyapunov function unless we pay a price. In
this section, we shall prove such a converse theorem and we see that
the price we have to pay to overcome the diculties is reasonable.
Let us recall the following modication of a result of Massera
which is useful in our discussion.
Lemma 3.4.1 Let L and > 0. Then, there exists a
function a K such that
a((s)) exp(s),
s 0.
We shall assume that h0 , h and h0 is uniformly ner than

h, (h0 , h) being the two measures discussed in Section 3.3. Recall
151
that this means that there exists a > 0 and a K such that
h0 (t, x) < implies h(t, x) (h0 (t, x)).
We are now in position to prove our main result of this section.
(i) |h(t, x)h(t, y)| Lxy and f (t, x)f (t, y) M (xy)
for (t, x), (t, y) R+ Rm , where L, M > 0 are constants;
(ii) the system (3.1.1) is (h0 , h)-uniformly stable and (h, h)-quasi
uniformly asymptotically stable.
Then, for a constant > 0 there exist two functions U, W
C[S(h, ), R+ ] which are Lipschitzian in x and such that
(a) U (t, x) b(h(t, x)) for (t, x) S(h, ) and U (t, x) a(h0 (t, x))
for (t, x) S(h0 , 0 ), where a, b K and 0 (0, ) is a constant with (0 ) < ;
(b) D + U (t, x) 0 for (t, x) S(h, );
(c) W (t, x) N for (t, x) S(h, ) and W (t, x) b1 (h0 (t, x)) for
(t, x) S(h0 , 0 ), where b1 K and N > 0 is a constant;
(d) D + W (t, x) C(U (t, x)) for (t, x) S(h, ), where c K.
Proof We shall use some standard arguments. Choose a constant
> 0 and for any > 0, a T = T () > 0, both associated with
(h, h)-quasiuniform asymptotic stability assumed in (ii). Obviously,
the function T can be assumed to be decreasing. For (t, x) S(h, )
and j = 1, 2, . . . , dene
Uj (t, x) = sup{Gj (h(t + , x(t + , t, x))) : 0} exp[M T (j 1 )],
(3.4.1)
where Gj (u) = u j 1 for u j 1 and Gj (u) = 0 for 0 u j 1 .
Clearly, for every u, v 0,
|Gj (u) Gj (v)| |u v|.
Because of (h, h)-quasiuniform asymptotic stability and the continuity of Gj and h, Uj is well dened by (3.4.1) as a mapping from
152
S(h, ) into R+ . We have,

Uj (t, x) = sup{Gj (h(t + , x(t + , t, x))) :
0 T (j 1 )} exp[M T (j 1 )],
(3.4.2)
from which it easily follows that Uj is continuous in t. Moreover, by

(3.4.2), (i) and Gronwall inequality, we get
|Uj (t, x) Uj (t, y)| sup{|h(t + , x(t + , t, x))
h(t + , x(t + , t, y))| : 0 T (j 1 )} exp[M T (j 1 )]
L sup{x(t + , t, x) x(t + , t, y) : 0 T (j 1 )}
exp[M T (j 1 )]
Lx y.
(3.4.3)
Thus, Uj C[S(h, ), R+ ]. Now, we set, for (t, x) S(h, ),
U (t, x) =
2j Uj (t, x).
(3.4.4)
j=1
Taking into account the decreasing character of T (), it can be

easily seen that
Uj (t, x) sup{h(t + , x(t + , t, x)) : 0 T (j 1 )}
1 + sup{h(t + , x(t + , t, x)) : 0 T (1)},
j = 1, 2, . . . ,
which implies the uniform convergence of the series of (3.4.4) in any
compact subset of S(h, ). Then U C[S(h, ), R+ ]. By (3.4.3) we
have
|U (t, x) U (t, y)| Lx y.
(3.4.5)
As a consequence of (3.4.1) and (3.4.4) we see that U is decreasing
along the solution of (3.1.1). From this and (3.4.5), it is easy to
obtain the inequality
D + U (t, x) 0,
for (t, x) S(h, ),
following the standard arguments.
153
Now, given (0, ), we choose j 1 such that j 1 < . We

have, for (t, x) S(h, ) \ S(h, ),
U (t, x) 2j Uj (t, x) 2j [h(t, x) j 1 ] exp[M T (j 1 )] > 0,
where = 2j ( j 1 ) exp[M T (j 1 )]. From this, it follows that
there exists a function a K such that U (t, x) a(h(t, x)) for
(t, x) S(h, ).
Let (0, ). Since j = 1, 2, . . . and (t, x) S(h, ),
Uj (t, x) sup{h(t + , x(t + , t, x)) : 0},
(3.4.6)
then the (h0 , h)-uniform stability of (3.1.1) implies the existence of

a () > 0 such that h0 (t, x) < () implies U (t, x) , which is
equivalent to the statement that there exist a constant 0 (0, )
with (0 ) < and a function a K such that U (t, x) a(h0 (t, x))
for (t, x) S(h0 , 0 ). Thus U satises condition (a).
Next we consider the function W : S(h, ) R+ dened by

C(U (, x(, t, x))) d,
W (t, x) =
(3.4.7)
where C K is to be chosen later. By (3.4.4), (3.4.6) and assumption

(ii), the system (3.1.1) is obviously (h0 , U )-uniformly asymptotically
stable and (H, U )-quasi uniformly asymptotically stable. Then, there
exists a constant (0, 0 ] such that
U (, x(, t, x)) p(h0 (t, x))q( t),
(t, x) S(h0 , 0 ),
(3.4.8)
where p K, q L. Also, there exists a function L such that

U (, x(, t, x)) ( t) for t and (t, x) S(h, ). (3.4.9)
We can assume 0 = 0 . We now choose c K such that the integrals

C(()) d
0
and

[C(p(0 )q())]1/2 d
0
(3.4.10)
154
converge and c exists and belongs to class K with

c (()) exp(),
(3.4.11)
where M + 1. Such a choice is possible by Masseras Lemma

3.4.1. As a consequence of (3.4.9) and (3.4.10), W is well dened
and bounded in S(h, ).
Now for (t, x), (t, y) S(h, ), we have

|W (t, x) W (t, y)|
|C(U (, x(, t, x))) C(U (, x(, t, y)))| d

t

= [C ()|U (, x(, t, x)) U (, x(, t, y))|] d,
t
where U (, x(, t, x)) U (, x(, t, y)) or U (, x(, t, y))

U (, x(, t, x)). Using the fact that c K and U is Lipschitzian
in x, we obtain

|W (t, x) W (t, y)| L
C (( t)) x(, t, x) x(, t, y) d

exp[( t)(M )] d
Lx y
t

exp[(M )] d.
Lx y
0
Thus,
|W (t, x) W (t, y)| Lx y.
(3.4.12)
Here we have employed Gronwall inequality, the relations (3.4.9),

(3.4.11) and the choice M + 1. Clearly W is continuous in t and
this fact, together with (3.4.12), proves that W C[S(h, ), R+ ].
Because of (3.4.8) and (3.4.10), we get, for (t, x) S(h0 , 0 ),

W (t, x)
C(p(h0 (t, x))q( t)) d

t

1/2
[C(p(h0 (t, x))q(0))]
155

[C(p(0 )q())]1/2 d
0
b1 (h0 (t, x)),
b1 K.
Now, it is easy to show, using (3.4.12) that D + W (t, x) = c(U (t, x))
for (t, x) S(h, ).
The two functions U and W have all the desired properties. The
proof is therefore complete.
Conversely, it can be easily proved that the system (3.1.1) is
(h0 , h)-uniformly stable and (h, h)-quasi uniform asymptotically stable if the following conditions are satised:
(i) h0 is uniformly ner than h;
(ii) for every > 0, every solution x(t, t0 , x0 ) with h(t0 x0 ) <
exists for all t t0 ; and
(iii) there exist two functions U, W C[S(h, ), R+ ], which are locally Lipschitzian in x and satisfy the conditions (a), (b), (c)
and (d) of Theorem 3.4.1.
We have the following corollary of Theorem 3.4.1.
Corollary 3.4.1 Suppose that the assumptions (i), (ii) of Theorem 3.4.1 hold. Then, for a constant > 0, there exists a function
V C[S(h, ), R+ ], which is Lipscitzian in x for a constant and such
that
(a) V (t, x) b(h(t, x)) for (t, x) S(h, ) and V (t, x) a(h0 (t, x))
for (t, x) S(h0 , 0 ), where a, b K, 0 (0, ) is a constant
with (0 ) < ;
(b) D + V (t, x) (h(t, x)) for (t, x) S(h, ), where K.
Indeed, if U and W are the functions obtained in Theorem 3.4.1,
the function V = U + W has the desired properties with a = b + b1 ,
= coa.
If h(t, x) = h0 (t, x), Theorem 3.4.1 and its corollary become
two equivalent propositions. Thus, when h(t, x) = h0 (t, x) = x,
156
Theorem 3.4.1 reduces to the well-known Masseras converse theorem on uniform asymptotic stability (actually, in Masseras theorem further assertions about smoothness of V are made). If
h(t, x) = xs , and
h0 (t, x) = x, where is the Euclidean
norm and xs = x21 + + x2s , s < n, then Theorem 3.4.1 yields
a converse theorem for partial uniform asymptotic stability. It is
clear that various choices of h and h0 are possible and thus Theorem
3.4.1 oers a unied result that is exible enough to warrant its use
in several applications.
If we carefully examine the proof of Masseras theorem, we notice
that it is the domain of attraction which plays the prominent role
in obtaining a smooth Lyapunov function. Consequently, when two
measures are employed, this same feature shows that the price we
had to pay to prove Theorem 3.4.1 is reasonable and natural.
3.5
Boundedness and Lagrange Stability in Terms of

Two Measures
Corresponding to the dierent types of stability notions, there are

dierent types of boundedness concepts which we shall dene below.
Denition 3.5.2 Let h0 , h . Then, the dierential system
(3.1.1) is said to be
(B1 ) (h0 , h)-equibounded, if for each > 0, t0 R+ , there exists a
positive function = (t0 , ) which is continuous in t0 for each
such that
h0 (t0 , x0 ) implies
H(t, x(t, t0 , x0 )) < (t0 , ),
t t0 ;
(B2 ) (h0 , h)-uniformly bounded if in (B1 ) is independent of t0 ;

(B3 ) (h0 , h)-quasi-equi-ultimately bounded, if for each 0 and
t0 R+ , there exist positive numbers N and T = T (t0 , ) such
that
h0 (t0 , x0 ) implies
h(t, x(t, t0 , x0 )) < N,
t t0 + T ;
157
Boundedness and Lagrange Stability
(B4 ) (h0 , h)-quasi uniform ultimately bounded, if T in (B3 ) is independent of t0

(B5 ) (h0 , h)-equi ultimately bounded if (B1 ) and (B3 ) hold;
(B6 ) (h0 , h)-uniform ultimately bounded if (B2 ) and (B4 ) hold;
(B7 ) (h0 , h)-equi Lagrange stable if (B1 ) and (S3 ) hold;
(B8 ) (h0 , h)-uniform Lagrange stable if (B2 ) and (S4 ) hold.
Observe that if in (B1 ) is such that (t0 , ) K, then (h0 , h)equiboundedness implies (h0 , h)-equistability, since given > 0, there
exists a (t0 , ) such that (t0 , ) whenever (t0 , ). Moreover, if LK, then (h0 , h)-equibondedness implies (h0 , h)-eventual
stability, because given > 0, there exist () > 0 and () > 0 such
that (t0 , ) whenever () and t0 ().
(i) h0 , h and h0 is uniformly ner than h with occurring in
the Denition 3.2.2 being suciently large;
h-positive denite and h0 -decrescent with the function b K occurring in the Denition 3.2.3 satisfying b(u) as u ;
2 , R] and for (t, x) R Rn , D + V (t, x)
(iii) g C[R+
+
g(t, V (t, x)).
Then, boundedness properties of (3.1.5) imply the corresponding

(h0 , h)-boundedness properties of (3.1.1).
Proof We shall only indicate the proof of (h0 , h)-equiboundedness
of (3.1.1) since the proofs of other concepts are similar. The relations
(3.3.1) and (3.3.3) hold with 0 = 1 (), in view of assumption (ii).
Let 0 < 0 and t0 R+ . Set 1 = a() and suppose that (3.1.5)
is equibounded. Then, given 1 > 0 and t0 R+ , there exists a
1 = 1 (t0 , ) that is continuous in t0 for each such that
u0 < 1
implies
u(t, t0 , u0 ) < 1 ,
t t0 ,
(3.5.1)
158
where u(t, t0 , u0 ) is any solution of (3.1.5). Choose a = (t0 , )

such that b() 1 and let h0 (t0 , x0 ) < . Then (3.3.3) implies that
h(t0 , x0 ) < . If possible, let there exist a solution x(t) = x(t, t0 , x0 )
of (3.1.1) and t1 > t0 such that
h(t1 , x(t1 )) =
and h(t, x(t)) ,
t [t0 , t1 ].
(3.5.2)
Then, by Theorem 3.1.1, we get, setting u0 = V (t0 , x0 ),

V (t, x(t)) r(t, t0 , u0 ),
t [t0 , t1 ],
(3.5.3)
where r(t, t0 , u0 ) is the maximal solution of (3.1.5). Hence, the relations (3.3.1), (3.5.1),(3.5.2) and (3.5.3) lead to the contradiction
b() V (t1 , x(t1 )) r(t1 , t0 , u0 ) < b(),
proving the theorem.
3.6
Stability Results for Autonomous or Periodic Systems
In order to avoid the complications in using two measures and to

keep the discussions simple, we shall, hereafter, consider the stability properties of the trivial solution. This means that we choose
h(t, x) = h0 (t, x) = x. Let us consider the system
x = f (t, x),
x(t0 ) = x0 ,
(3.6.1)
where f C[R+ S(), Rn ] and f (t, 0) 0, S() being {x Rn :

x < }.
In this and the following sections, we shall be concerned with
the problem of nding the appropriate modications that are needed
to prove stability results when D + V satises a weaker assumption,
say, positive semideniteness. Clearly such changes require compensating conditions and it is reasonable to involve the set where
D + V (t, x) = 0. This problem is very important in applications. For
example, the dissipative mechanical systems lead to such a situation.
In this section, we shall deal with periodic systems and in the next
section, we shall deal with general systems.
159
The following simple result shows that if f is periodic in t or autonomous and is smooth enough to guarantee uniqueness of solutions,
then stability of the trivial solution of (3.6.1) is always uniform.
Theorem 3.6.1 Let f C[R S(), Rn ], f (t, x) be periodic in t
with a period and the system (3.6.1) admit unique solution. Then,
the stability of the trivial solution of (3.6.1) is necessarily uniform.
Proof By the periodicity of f (t, x) in t, it follows that, if x(t, t0 , x0 )
is a solution of (3.6.1), then x(t + , t0 , x0 ) is also a solution. Furthermore, the uniqueness of solution shows that, for any integer k,
x(t k, t0 kw, x0 ) = x(t, t0 , x0 ).
For each xed t0 , t0 (, ), let (t0 ) = sup (t0 , ). Since
0<<
(t0 , ) is continuous in t0 for each , if we let 0 =

is clear that 0 > 0. For < < 0 , we dene
inf [(t0 )], it
0t0
() = sup x(t0 + , t0 , x0 ).

0
Notice that () is a monotone increasing function of and hence, if

() is the inverse function of (), we have
x(t, t0 , x0 ) < ,
t t0 ,
for every t0 [0, ], provided x0 (). Let t0 be arbitrary. Then

we can choose an integer k such that
k t0 (k + 1)
or
(k + 1) t0 k.
Hence, either 0 t0 k or 0 t0 +(k+1) . Consequently,

if x0 ()| either
x(t, t0 , x0 ) = x(t k, t0 k, x0 ) <
or
x(t, t0 , x0 ) = x(t + (k + 1), t0 + (k + 1), x0 ) < ,
for t t0 and the uniformity of stability is evident.
160

(i) f is periodic in t with period and solutions of (3.6.1) are
unique;
(ii) V C[R+ S(), R+ ], V (t, x) is locally Lipschitzian in x and
periodic in t with period ;
(iii) for (t, x) R+ S(), V (t, x) b(x), V (t, 0) 0, b K
and D + V (t, x) 0; and
(iv) the set = {(t, x) R+ (S() \ {0}) : D + V (t, x) = 0} does
not contain any noncontinuable positive trajectory.
Then, x 0 is uniformly asymptotically stable.
Proof It follows by Theorems 3.3.1 and 3.6.1 that the origin is
uniformly stable. Let () > 0 be the number associated with this
property for > 0. Let 0 = () and x t0 R+ and x0 S(0 ).
Then, we have, for any solution of (3.6.1),
x(t, t0 , x0 ) < ,
t t0 .
(3.6.2)
We shall show that given any (0, ), there exists a t0 > t0

such that x(t0 , t0 , x0 ) < (). Hence x(t, t0 , x0 ) < for t t0
and we will be done. Suppose it happens that
x(t, t0 , x0 ) ()
for all t t0 ,
(3.6.3)
and set v(t) V (t, x(t, t0 , x0 )). By (iii), one has

lim v(t) = v,
v R+ .
(3.6.4)
Consider the sequence xk x(t+k, t0 , x0 ), k = 1, 2, . . . . Because of

(3.6.2) and (3.6.3), there exists a subsequence {xki } which converges
to a point x = 0 and x (). We have
v(t0 + ki ) = V (t0 + ki , xki ) = V (t0 , xki ).
Then, the continuity of V (t0 , ) and (3.6.4) imply V (t0 , x ) = v.
161
Now, consider the solution x(t, t0 , x ). Because of (iv) and the

fact that x(t, t0 , x ) , t t0 , there exists a t > t0 such that
lim V (t , x(t , t0 , xki )) = V (t , x(t , t0 , x )).
Therefore,
lim V (t , x(t , t0 , xki )) < v.
(3.6.5)
The periodicity of f implies that

x(t , t0 , xki ) = x(t + ki , t0 + ki , x(t0 + ki , t0 , x0 ))
= x(t + ki , t0 , x0 ).
Then,
V (t , x(t , t0 , xki )) = V (t + ki , x(t + ki , t0 , x0 )) = v(t + ki ),
and hence, by (3.6.4), it follows that
lim V (t , x(t , t0 , xki )) = v
in contradiction to (3.6.5). This completes the proof.

Analogously, we can prove a result for the instability of the trivial
solution.
Theorem 3.6.3 Assume that the hypotheses (i), (ii) and (iv) of
Theorem 3.6.2 hold. Suppose further that V (t, 0) 0 and
(iii ) for all (t, x) R+ S(), D + V (t, x) 0 and for any (0, ),
there is x S() with V (t0 , x) > 0.
Then, the solution x 0 of (3.6.1) is unstable.
Proof Suppose that x 0 is stable. Then, it is uniformly stable.
Hence, by the same arguments used in the proof of Theorem 3.6.2, we
obtain that x 0 is asymptotically stable. Let (0, ) such that
x0 S() implies x(t, t0 , x0 ) 0 as t . Choose x0 S() with
the condition V (t0 , x0 ) = V0 > 0. Then, we get the contradiction
0 < V0 lim V (t, x(t, t0 , x0 )) = 0
t
162
3.7
Perturbing Family of Lyapunov Functions
We shall consider, in this section, the general system (3.6.1) relative

to the problem posed in Section 3.6. We shall rst discuss a result on
nonuniform stability of the trivial solution of (3.6.1) under weaker
assumptions, which indicates that in those situations when the Lyapunov function found does not satisfy all the desired conditions, it is
fruitful to perturb that Lyapunov function rather than to discard it.
Moreover, this result demonstrates the usefulness and exibility of
employing a family of Lyapumov functions instead of a single function. We then prove asymptotic stability results involving the set
where D + V (t, x) = 0, where we actually construct a one parameter
family of Lyapunov functions as a result of perturbing the original
Lyapunov function.
(i) V1 C[R+ S(), R+ ], V1 (t, x) is locally Lipschitzian in x,
V1 (t, 0) 0 and
D + V1 (t, x) g1 (t, V1 (t, x)),
(t, x) R+ S(),
2 , R] and g (t, 0) 0;
where g1 C[R+
1
(ii) for every > 0, there exists a V2, C[R+ S() S c (), R+ ],
V2, is locally Lipschitzian in x and for (t, x) R+
S() S c (),
b(x) V2, (t, x) a(x),
a, b K
and
D + V1 (t, x) + D + V2, (t, x) g2 (t, V1 (t, x) + V2, (t, x)),
2 , R], g (t, 0) 0;
where g2 C[R+
2
(iii) the trivial solution of

u = g1 (t, u),
u(t0 ) = u0 0
(3.7.1)
is equistable and the trivial solution of

v = g2 (t, v),
is uniformly stable.
v(t0 ) = v0 0
(3.7.2)
163
Then, the trivial solution of the system (3.6.1) is equistable.

Proof Let 0 < < and t0 R+ be given. Since the trivial
solution of (3.7.2) is uniformly stable, given b() > 0 and t0 R+ ,
there exists a 0 = 0 () > 0 such that
v(t, t0 , v0 ) < b(),
t t0 ,
(3.7.3)
provided v0 < 0 , where v(t, t0 , x0 ) is any solution of (3.7.2). In view

of a K, there is a 2 = 2 () > 0 such that
a(2 ) <
0
.
2
(3.7.4)
By the equistability of u 0 relative to (3.7.1), given

t0 R+ , there exists a = (t0 , ) such that
u(t, t0 , u0 ) <
0
,
2
t t0 ,
0
> 0 and
2
(3.7.5)
whenever u0 < , u(t, t0 , x0 ) being any solution of (3.7.1).

Choose u0 = V1 (t0 , x0 ). Since V1 (t, x) is continuous and
V1 (y, 0) 0, there exists a 1 > 0 such that
x0 < 1
and V1 (t0 , x0 ) <
(3.7.6)
hold simultaneously. Set = min(1 , 2 ). Then we claim that

x0 < implies x(t, t0 , x0 ) < for t t0 . If this were false,
there would exist a solution x(t, t0 , x0 ) of (3.6.1) with x0 < and
t1 , t2 > t0 such that
x(t1 , t0 , x0 ) = 2 ,
x(t2 , t0 , x0 ) =
(3.7.7)
and x(t, t0 , x0 ) S() S(2 ) on [t1 , t2 ]. Let 2 = so that the

existence of a V2, satisfying hypothesis (ii) is assured. Hence, setting
m(t) = V1 (t, x(t, t0 , x0 )) + V2, (t, x(t, t0 , x0 )),
we obtain the dierential inequality
D + m(t) g2 (t, m(t)),
t [t1 , t2 ],
t [t1 , t2 ],
164
which yields
m(t2 ) r2 (t2 , t1 , m(t1 )),
r2 (t, t1 , v0 ) being the maximal solution of (3.7.2) such that r2 (t1 ,

t1 , v0 ) = v0 . We also have
V1 (t1 , x(t1 , t0 , x0 )) r1 (t1 , t0 , V1 (t0 , x0 )),
where r1 (t, t0 , u0 ) is the maximal solution of (3.7.1). By (3.7.5) and
(3.7.6), we get
0
V1 (t1 , x(t1 , t0 , x0 )) <
.
(3.7.8)
2
Also, by (3.7.4), (3.7.7) and the assumptions on V2, , we have
V2, (t1 , x(t1 , t0 , x0 )) a(2 ) <
0
.
2
(3.7.9)
The inequalities (3.7.8), (3.7.9) together with (3.7.3), (3.7.7),

V1 0 and V2, b(kxk) lead to the contradiction b() < b().
Hence the proof of the theorem is complete.
When the hypothesis on D + V (t, x) is weakened, we need to impose a compensating assumption in the region close to the set where
D + V (t, x) = 0. We need the following notations before we can proceed to discuss our results.
Let : R+ Rn R. For any t R+ , we denote by At ( = 0)
the set of all x Rn for which (t, x) = 0. If is independent of t,
the set At ( = 0) coincides with a unique set which will be denoted
by A( = 0). Let (, ) = {(t, x) R+ Rn : kxk < }, for
0 < < and C be the class of f C[R+ Rn , Rn ] such that for
any compact set K Rn , there exists a L = L(K) > 0 satisfying
the Lipschitz condition
kf (t, x) f (t, y)k Lkx y k, t R+ ,
x, y K.
Definition 3.7.1 Let > 0. A function W : R+ Rn R is

said to be -definitely nonvanishing on the sets At (V = 0) or on the
set A(V = 0) when V is independent of t, if for every (0, ),
there exist , > 0 such that
(t, x) R+ (, )
and V (t, x) <
implies
|W (t, x)| > ,
165
W
W
+
f (t, x).
t
x
Theorem 3.7.2 Suppose that V C[R+ Rn , R+ ], V (t, x) is
locally Lipschitzian in x and W C 1 [R+ Rn , R] such that
where W (t, x) =
(i) a(x) V (t, x) b(x) for all (t, x) R+ S(), a, b K;

(ii) there exist C[R+ Rn , R+ ], C and c K such that for
all (t, x) R+ S(),
D + V (t, x) c((t, x));
(iii) W and W f are bounded in R+ S(); and
(iv) W is -denitely nonvanishing on the sets At ( = 0).
Then, x 0 is uniformly asymptotically stable.
Proof Let (0, ). By virtue of (iv), there exist , > 0 such
that
(t, x) (, ),
(t, x) <
implies
|W (t, x)| > . (3.7.10)
Consider the two sets

G(1)
= {(t, x) (, ) : (t, x) < , W (t, x) < },

G(2)
= {(t, x) (, ) : (t, x) < , W (t, x) > }
(i)
and the functions : R+ (, ) R+ , i = 1, 2, dened by

(i)
(t, x) for (t, c) G ,
(i)
=
(i)
for (t, c) (, ) \ G .
Obviously these two functions are of class C. Let C 1 [R+ , [0, 1]]
be such that

, ( ) = 0 for [ , ),
( ) = 1 for 0,
1
(i)
and consider the two functions : (, ) [0, 1], i = 1, 2, such

that
(i)
(i)
for (t, x) (, ).
(t, x) = ( (t, x)),
166

(i)
It is easy to show that C, i = 1, 2. Now, we set for every

(t, x) (, ),
(2)
k (t, x) = [(1)
(t, x) (t, x)]W (t, x).
The function k is bounded and along the solution of (3.6.1),

(2)

D + k (t, x) [(1)
(t, x) (t, x)]W (t, x)
(2)
+ W (t, x)D[(1)
(t, x) (t, x)],
(3.7.11)
where D = D + or D+ according as W (t, x) 0 or W (t, x) 0 re(i)

spectively. Since C, i = 1, 2, we see that for all (t, x) (, ),

(2)
2
2
|W (t, x)D[(1)
(t, x) (t, x)]| n W (t, x)[1 + f (t, x) ],
(3.7.12)
where n > 0 is a constant. Then (iii) implies that the left hand
side of (3.7.12) is bounded. On the other hand, from the denition
(i)
of and from (3.7.10) it follows
(2)

[(1)
(t, x) (t, x)]W (t, x) 0,
for all
(t, x) (, ).
Thus D + k is bounded from above. Furthermore the following property:

(t, x) (, )
and (t, x)
imply D + k (t, x) <
(3.7.13)
holds. Now, we shall show that we can determine a constant > 0
such that, setting for all (t, x) (, ),
V (t, x) = V (t, x) + k (t, x),
the function V satises the hypotheses of Theorem 3.3.5 and V is
bounded. Moreover, for > 0, we have
D + V (t, x) D + V (t, x) + D + k (t, x).
We now prove that we can choose > 0 such that D + V (t, x) <
, then from
. Indeed, if (t, x) (, ) and (t, x)
2
167
(3.7.13) and (ii) it follows that D + V (t, x) < for every choice
, then using again the condition (ii) we

of > 0. If (t, x) >
2
have

+
D V (t, x) < e
+ A ,
2
where A > 0 is an upper bound for D + k (t, x). Hence, for all
(t, x) (, ), we shall have D + V (t, x) < , if we choose
so that
0 < < (A + 1 )c( /2).
On the other hand, by virtue of (i), (ii), the solution x 0 of (3.6.1)
is uniformly stable. Thus, all the hypotheses of Theorem 3.3.5 are
satised and the proof is complete.
The assumption g1 (t, u) = g2 (t, u) 0 is admissible in Theorem
3.7.1, which implies that
+
D + V1 (t, x) 0 on
+
D V1 (t, x) + D V2, (t, x) 0 on
R+ S(),
and
R+ S() S c ().
If, on the other hand, we demand that D + V1 (t, x) satises a strengthened condition, then we can conclude from Theorem 3.7.2 uniform
asymptotic stability of the trivial solution of (3.6.1). This is stated
in the following corollary.
Corollary 3.7.1 Let, in Theorem 3.7.1, g1 (t, u) = g2 (t, u) 0.
Suppose that
D + V1 (t, x) c((t, x)),
(t, x) R+ S(),
where c, are the same functions dened in (ii) of Theorem 3.7.2.

Assume also that hypotheses (iii) and (iv) of Theorem 3.7.2 hold.
Then x 0 of (3.6.1) is uniformly asymptotically stable.
Proof We now have D + V1 (t, x) 0 and hence D + V2, (t, x) 0
which implies uniform stability of x 0 of (3.6.1). Then, Theorem
3.7.2 shows that we can reduce to Theorem 3.3.5 which yields the
stated conclusion.
When f is bounded, conclusion (i) of Theorem 3.7.2 can be weakened and replaced by the conditions that V is positive semi-denite
168
and V (t, 0) 0. But the uniformity of asymptotic stability is lost.

We have the following theorem in this direction.
Theorem 3.7.3 Assume that V C[R+ Rn , R+ ], V (t, x) is
locally Lipschitzian in x and W C 1 [R+ Rn , R] such that
(i) V (t, 0) 0, V (t, x) 0 for (t, x) R+ S();
(ii) D + V (t, x) c((t, x)) 0 where C and C[R+
S(), R+ ];
(iii) f and W are bounded on R+ S(); and
(iv) W (t, x) is -denitely nonvanishing on the sets At ( = 0).
Then, the solution x 0 of (3.6.1) is asymptotically stable.
Proof We shall proceed with the proof in four stages for convenience.
(I) Let > 0 and x1 , x2 S() be such that d(x1 , x2 ) > . Let
x(t) be a solution of (3.6.1) with x(t1 ) = x1 , x(t2 ) = x2 , t2 > t1 .
The boundedness of f implies the existence of a number > 0,
depending on only, such that t2 t1 > , namely, = M
, M
being the bound for f (t, x) on R+ S().
(II) By virtue of (iii) and (iv), we have that for each (0, ),
there exist two numbers > 0, > 0 such that (t, x) (, )
and (t, x) < imply |W (t, x)| < r , |W (t, x)| > , where
r [0, 1/2) and is dened as in the above paragraph. Consider
the functions k , V dened as in the proof of Theorem 3.7.2, namely,
(2)
k (t, x) = [(1)
(t, x) (t, x)]W (t, x)
and
V (t, x) = V (t, x) + k (t, x),
(3.7.14)
where > 0 is chosen such that for all (t, x) (, ),

D + V (t, x) .
(3.7.15)
Obviously now we have that V is bounded from below only and

|k (t, x)| < 2r .
169
(III) Given t0 R+ and > 0, we choose (0, /2) and dene

> 0 with the following property
x <
implies
V (t0 , x) < ,
where = (1 2r) > 0. Consider a solution x() of (3.6.1)

passing through (t0 , x0 ), with x0 < . Suppose that there exist
t1 , t2 R+ , t0 < t1 < t2 with
, x(t2 ) = and
x(t) for t [t1 , t2 ].
2
2
From (3.7.15) and the consideration in (I), it follows that
x(t1 ) =
V (t2 , x(t2 )) V (t1 , x(t1 )) <

and consequently, because of (3.7.14) and (3.7.15),
V (t2 , x(t2 )) V (t1 , x(t1 )) < .
Therefore
V (t2 (x, t2 )) < V (t1 , x(t1 )) V (t0 , x(t0 )) < 0
which is a contradiction. This proves that x 0 is stable.
(IV) For t0 R+ , dene 0 = 0 (t0 ) (0, ) such that x0 < 0
implies x(t) < for all t t0 . Let x0 S(0 ). We shall now
prove that for each (0, ) and t0 , there exists a t such
that x(t) < . Indeed, if x(t) for all t t0 , we should have
V (t, x(y)) as t which is a contradiction.
In order to prove the asymptotic stability of x 0, we have only
to show that for every 2 (0, ), there exists T > 0 (which will
depend on t0 , x0 , ) so that x(t) < 2 for all t t0 + T . Suppose
this is not true. Then we can nd two divergent sequences {ti },{ti }
such that for each xed i, i = 1, 2, . . . , x(ti ) = , x(ti ) = 2 and
< x(t) < 2 for t [ti , ti ]. With the aid of the function V (t, x)
dened by (3.7.14) relative to an (0, ) proceeding as in (III) we
can see that in each interval [ti , ti ] one gets
V (ti , x(ti )) V (ti , x(ti )) < .
Because > 0 is independent of i and D + V (t, x) 0 for all
(t, x) R+ S(), we get the contradiction V (t, x(t)) as
t . The proof is complete.
170
3.8
M0 -Stability Criteria
Stability in the sense of Lyapunov investigates the stability properties

of invariant sets. Since in many concrete problems such as adaptive
control systems, one needs to consider the stability of sets which are
not invariant, the notion of eventual stability was introduced to deal
with such situations. It is subsequently recognized that although the
set which is eventually stable is not invariant in the usual sense, it is
so in the asymptotic sense. This observation leads to a new concept
of asymptotically invariant sets, which form a special subclass of invariant sets, and their stability properties. In this section, a concept
of stability called M0 -stability, which describes a very general type
of invariant set and its stability behavior, is introduced. This notion
naturally leads us to consider the initial values on surfaces that critically depend on initial time and to use dierent topologies in the
denition of stability of M0 -invariant sets.
Consider the generalized initial value problem
x = f (t, x),
x(t0 ) = (t0 , x ),
t0 0,
(3.8.1)
where f, C[R+ Rn , Rn ] and f is smooth enough to ensure

existence of solutions of (3.8.1). For convenience, let us introduce
the following notation:
M = M (R+ , Rn ) is the space of all measurable mappings from
R+ to Rn such that x M if and only if x(t) is locally integrable on R+ and
t+1
x(s) ds < ;
sup
t>0
M0 = M0 (R+ , Rn ) is the subspace of M = M (R+ , Rn ) consisting of all x(t) such that

t+1
x(s) ds
t
as
t ;
171
the set S(M0 , ) is the subset of M = M (R+ , Rn ) dened by

t+1
S(M0 , ) = x M : lim inf
x(s) ds .
t
By x S(M0 , ), we mean that for each > 0, there exists a () > 0

with the property that () as 0 such that
t+1
x(s) ds < ,
t > ().
Let us now give the denitions for M0 -invariant set and the various
types of M0 -stability. As usual, let x(t, s, (s, x )), t s, represent
a solution of (3.8.1) starting at (s, (s, x )) .
Denition 3.8.1 Let A Rn . A is M0 -invariant with respect to
the system (3.8.1) if whenever x A and (s, x ) M0 , we have
x(, s, (s, x )) M0 .
Denition 3.8.2 With respect to the system (3.8.1), the set A
is said to be
(M1 ) M0 -equistable if for each > 0, there exists 1 (), 1 ()
as and 1 (t0 , ), 2 (t0 , ) such that
t
0 +1
x(t, s, (s, x )) ds < ,
t0
provided x S(A, 1 ) and
t0 +1
t0
t t0 + 1,
(s, x ) ds < 2 , t0 1 ();
(M2 ) M0 -uniformly stable if 1 and 2 in (M1 ) are independent of t0 ;

(M3 ) M0 -quasi equiasymptotically stable if for every > 0, there exist
positive numbers 10 (t0 ), 20 (t0 ), 0 and T (t0 , ) such that
t
0 +1
x(t, s, (s, x )) ds < ,
t t0 + 1 + T (t0 , ),
t0
provided x S(A, 10 ) and (s, x ) S(M0 , 20 );
t0 0 ,
172
(M4 ) M0 -quasi uniformly asymptotically stable if 10 , 20 and T in

(M3 ) are independent of t0 ;
(M5 ) M0 -equi asymptotically stable if (M1 ) and (M3 ) hold;
(M6 ) M0 -uniformly asymptotically stable if (M2 ) and (M4 ) hold.
Consider the example
x = e1 ,
where (s, x ) = x +
x(t0 ) = (t0 , x ),
t0 0,
1
. The solution is given by
s
x(t, s, (s, x )) = x +
1
+ es et ,
s
and it is clear that the set x = 0 is M0 -uniformly stable. For this

example, the set x = 0 is also eventually stable. If, on the other
hand, we choose (s, x ) = x + (s) where : [0, ) R is a C 1
function coinciding with et except at some peaks where it reaches
the value 1. There is one peak for each integer value of t and the
width of the peak corresponding to abscissa n is smaller than ( 12 )n .
Then x = 0 is not eventually stable, but it is M0 -uniformly stable.
This example shows that stability behavior depends also on the initial
values.
Now, consider another initial value problem
x = (t),
where is dened by
n,
2n4 (t n) + n,
(t) =
2n4 (t n) + n,
x(t0 ) = x
(3.8.2)
when t = n, an integer,
1
when n 3 < t < n,
2n
1
when n < t < n + 3 ,
2n
for all other t 0.
Then (t) exists except on a set of measure zero. Considering only

positive solutions of (3.8.2), we obtain
x(t, s, (s, x )) = x + (t0 ) (t) x + (t0 ).
173
t +1
The set x = 0 is M0 -uniformly stable since t00 (s) ds is at most
1
for n [t0 , t0 + 1]. But x = 0 is not eventually uniformly stable
2n2
since (t0 ) does not approach zero as t0 .
We need a preliminary result and some convenient notation before
we can proceed to prove M0 -stability criteria.
Lemma 3.8.1 (Jensen inequality) Let be a convex function
and f integrable. Then

f (t) dt (f (t)) dt.
Consider the comparison equation

u = g(t, u),
u(t0 ) = (t0 , u ),
t0 0,
(3.8.3)
2 , R], C[R2 , R ]. The set u = 0 is M -invariant

where g C[R+
+
0
+
if u(, s, (s, 0)) M0 whenever (s, 0) M0 . Concepts analogous
to (M1 )(M6 ) can be dened. For example, the set u = 0 is
(M1 ) M0 -equistable, with respect to (3.8.3), if for each > 0, there

exist 1 (), 1 (t0 , ) and 2 (t0 , ) such that
t
0 +1
u(t, s, (s, u )) ds < ,
t t0 + 1,
t0
provided u < 1 and
t0 +1
t0
(s, u ) ds < 2 , t0 1 ().
The remaining notions (M2 )(M6 ) corresponding to (M2 )(M6 )

can be easily formulated.
We shall present results concerning M0 -uniform stability and M0 uniform asymptotic stability. Based on these, one can construct
proofs for other cases.
Theorem 3.8.1 Assume that there exist functions V (t, x) and
g(t, u) satisfying the following conditions:
2 , R];
(i) g C[R+
174
(ii) V C[R+ Rn , R+ ] and V (t, x) is Locally Lipschitzian in x;

(iii) b(x) V (t, x) a(t, x) where a P and b KC, with
KC = { C[R+ , R+ ] : K and is convex},
2
, R+ ] : given > 0, there exists () > 0 such that
P = { C[R+
(s, (s, x )) S(M0 , ) provided (s, x ) S(M0 , )};
(iv) for (t, x) R+ Rn , D + V (t, x) g(t, V (t, x)).

Then the M0 -uniform stability of the set u = 0 implies the M0 uniform stability of the set A, that is (M2 ) implies (M2 ).
Proof Let > 0. By (M2 ), there exist 1 (), 2 () and 1 () such
that
t
0 +1
u(t, s, (s, u )) ds < b(),
t t0 + 1,
t0 1 ()
t0
t +1
provided u < 1 and t00 (s, u ) ds < 2 . By (iii) and denition of A, we can nd 1 (), 2 () and 2 () such that the following
inequalities will hold simultaneously:
t
0 +1
a(s, (s, x )) ds < 2 ,
t0 2 ()
t0
and x S(A, 1 ),
t
0 +1
(s, x ) ds < 2 .
t0
Let 1 () = min(1 (), 1 ()) and () = max(1 (), 2 ()). If we

t +1
choose x such that x S(A, 1 ) and t00 (s, x ) ds < 2 , then
t
0 +1
x(t, s, (s, x )) ds < ,
t0
t t0 + 1,
t0 ().
175
If this is not true, then there exists t1 > t0 + 1, t0 (), such

that
t
0 +1
x(t1 , s, (s, x )) ds = ,
t0
t
0 +1
x(t, s, (s, x )) ds < ,
t0 + 1 t < t1 ,
t0 ().
t0
Let r(t, s, (s, u )) be the maximal solution of (3.8.3). Then, by

Theorem 3.1.1, we have
V (t, x(t, s, (s, x ))) r(t, s, (s, u )),
(3.8.4)
since V (s, (s, x )) a(s, (s, x )) (s, d(x , A)) = (s, u )
letting u = d(x , A). Using (3.8.4) and assumption (iii), we obtain
the following contradiction:
t0 +1

b() b
x(t1 , s, (s, x )) ds
t0
t
0 +1
V (t1 , x(t1 , s, (s, x ))) ds
t0
t
0 +1
r(t1 , s, (s, u )) ds < b(),
t0
since u < 1 and
t0 +1
t0
(s, u ) ds < 2 . This completes the proof.
Corollary 3.8.1 The function g(t, u) = (t) where

M0 C 1 [R+ , R+ ], is admissible in Theorem 3.8.1.
Corollary 3.8.2 The function g(t, u) 0 is also admissible in
Theorem 3.8.1.
These corollaries correspond to the rst theorem of Lyapunov in
classical stability theory. Note that stability of an invariant set and
176
the stability of an asymptotically invariant set imply M0 -stability of

the invariant set but the converse is not true.
Theorem 3.8.2 Assume the conditions of Theorem 3.8.1 hold.
Then, if the set u = 0 is M0 -uniformly asymptotically stable with
respect to (3.8.3), the set A is M0 -uniformly asymptotically stable
with respect to the system (3.8.1).
Proof By Theorem 3.8.1, the set A is M0 -uniformly stable and
we need to prove that (M4 ) holds. It follows from (M4 ) that there
, , and T () such that
exist positive numbers 10
20
0
t
0 +1
u(t, s, (s, u )) ds < b(),
t t0 + 1 + T (),
t0 0 ,
t0

and t0 +1 (s, u ) ds < .
provided u < 10
20
t0
As in the proof of Theorem 3.8.1, we can nd positive numbers
10 , 20 , 0 which satisfy the inequalities
t
0 +1
a(s, (s, x )) ds < 20

,
t 0 ,
t0
t +1
and x S(A, 10 ), t00 (s, x ) ds < 20 . Then 0 does not
is independent of t . Let
depend on t0 since 20
0
10 = min(10 , 10 )
and 0 = max(0 , 0 ). Then, we claim that

t
0 +1
x(t, s, (s, x )) ds < ,
t t0 + 1 + T (),
t0 0 ,
t0
when x is chosen so that x S(A, 10 ) and

t
0 +1
(s, x ) ds < 20 .
t0
If this is not true, there exists a sequence {tk }, tk as k ,

such that
t
0 +1
x(tk , s, (s, x )) ds > .
t0
Several Lyapunov Functions
177
This leads to a contradiction

t
0 +1
b()
t
0 +1
r(tk , s, (s, u )) ds < b()
V (tk , x(tk , s, (s, x ))) ds

t0
t0
3.9
As we have seen in earlier sections, it is possible to study a variety of

problems in a unied way by just using a single Lyapunov function.
However, it is natural to ask whether it might be more advantageous,
in some situations, to use several Lyapunov functions. The answer
is positive and this approach oers a more exible mechanism and
each function can satisfy less rigid requirements. In this section, we
shall consider vector and matrix-valued Lyapunov functions, indicating the fruitfulness of employing several Lyapunov functions and
applying it to study large scale dynamical systems.
3.9.1
Vector Lyapunov functions method
Let us rst consider the method of vector Lyapunov functions. Naturally, Theorem 3.1.2 plays an important role whenever we employ
vector Lyapunov functions. As a typical result, we shall merely state
a theorem that gives sucient conditions in terms of vector Lyapunov functions for the stability properties of the trivial solution of
(3.6.1).
(i) g C[R+ RN , RN ], g(t, 0) 0 and g(t, u) is quasimonotone
nondecreasing in u for each t R+ ;
N ], V (t, x) is locally Lipschitzian in x and
(ii) V C[R+ S(), R+
the function
N

Vi (t, x)
(3.9.1)
V0 (t, x) =
i=1
is positive denite and decrescent;
178
(iii) f C[R+ S(), Rn ], f (t, 0) 0 and D + V (t, x)

g(t, V (t, x)), (t, x) R+ S().
Then, the stability properties of the trivial solution of
u = g(t, u),
u(t0 ) = u0 0,
(3.9.2)
imply the corresponding stability properties of the trivial solution of

(3.6.1).
Note that in condition (ii) of Theorem 3.9.1, we have used the
measure V0 (t, x) dened by (3.9.1). We could use other convenient
measures such as
V0 (t, x) = max Vi (t, x),
1iN
V0 (t, x) =
di Vi (t, x)
i=1
for a positive vector d > 0.

To exhibit the advantage in using vector Lyapunov functions, consider the following example.
Example 3.9.1 Let us consider the two systems

x = et x + y sin t (x3 + xy 2 ) sin2 t,
y = x sin t + et y (x2 y + y 3 ) sin2 t.
(3.9.3)
Suppose we choose a single Lyapunov function V given by

V (t, x) = x2 + y 2 .
Then, it is evident that
D + V (t, x) 2(et + | sin t|)V (t, x)
using the inequality 2|ab| a2 + b2 and observing that [x2 + y 2 ]2
sin2 t 0. Clearly, the trivial solution of the scalar dierential equation
u = 2(e1 + | sin t|)u, u(t0 ) = u0 0
is not stable and so, we cannot deduce any information about the stability of the trivial solution of (3.9.3) from Theorem 3.3.1, although
179
it is easy to check that it is stable. On the other hand, let us seek

a Lyapunov function as a quadratic form with constant coecients
given by
1
V (t, x) = [x2 + 2Bxy + y 2 ].
(3.9.4)
2
Then, the function D + V (t, x) with respect to (3.9.3) is equal to the
sum of two functions w1 (t, x), w2 (t, x) where
w1 (t, x) = x2 [et + B sin t] + xy[2Bet + (A + 1) sin t]
+ y 2 [Aet + B sin t],
w2 (t, x) = sin2 t[x2 + y 2 ](x2 + 2Bxy + y 2 ).
For arbitrary A and B, the function V (t, x) dened in (3.9.4)
does not satisfy Lyapunovs theorem on the stability of motion.
Let us try to satisfy the conditions of Theorem 3.3.1 by assuming
w1 (t, x) = (t)V (t, x). This equality can occur in two cases:
(i) A1 = 1, B1 = 1, 1 (t) = 2[et + sin t] when V1 (t, x) =
1
2
2 (x + y) ;
(ii) A2 = 1, B2 = 1, 2 (t) = 2[et sin t] when V2 (t, x) =
1
2
2 (x y) .
The functions V1 , V2 are not positive denite and hence, do not
satisfy Theorem 3.3.1. However, they do fulll the conditions of
Theorem 3.9.1. In fact,
(a) the functions V1 (t, x) 0, V2 (t, x) 0 and
and therefore V0 (t, x) =
crescent;
Vi (t, x) = x2 +y 2
i=1
Vi (t, x) is positive denite and de-
i=1
(b) the vectorial inequality D + V (t, x) g(t, V (t, x)) is satised

with the functions
g1 (t, u1 , u2 ) = 2(et + sin t)u1 ,
g2 (t, u1 , u2 ) = 2(et sin t)u2 .
180
It is clear that g(t, u) is quasi-monotone nondecreasing in u, and

the null solution of u = g(t, u) is stable. Consequently, the trivial
solution of (3.9.3) is stable by Theorem 3.9.1.
We shall indicate an application of Theorem 3.9.1 to large scale
dynamical systems where several Lyapunov functions result in a natural way in the study of such systems by the decomposition and
aggregation method. Let us consider the overall large scale system
of the form
x = F (t, x), x(t0 ) = x0 ,
(3.9.5)
which is assumed to consist of isolated sub-systems
zi = gi (t, zi ) + Ri (t, x),
where x Rn , zi Rni such that
zi (t0 )xi0 ,
(3.9.6)
ni = n and the Ri are the
i=1
interaction terms, usually of the form

Ri (t, x) =
(Rij (t, zj )).
j=1
If there exist functions Vi , di and wi such that

D + Vi (t, zi ) di (t, zi ) + wi (t, zi , x),
(3.9.7)
and if the right hand side of (3.9.7) can be majorized such that (3.9.7)
takes the general form
D + V (t, z) H(t, V (t, z))
(3.9.8)
where H(t, u) is quasi-monotone nondecreasing in u, then the method

of vector Lyapunov functions can be employed eectively, with suitable conditions on V and H to obtain various stability results of
(3.9.5) by means of Theorem 3.9.1. We note that the vector Lyapunov function V (t, z) can be constructed from the good behavior
(uniform stability or asymptotic stability) of subsystems which can
then be utilized to determine the behavior of the large scale system
(3.9.5) as indicated above. In many situations, H(t, V ) = AV where
A is an n by n matrix of nonnegative o-diagonal elements, which
181
implies that H(t, V ) is quasi-monotone nondecreasing in u. Since

this needed property of A is not a necessary condition for the matrix
A to be a stable matrix, the limitation of this method is clear.
Consider a nonlinear system
dy
= g(y),
dt
(3.9.9)
where y = (y1 , . . . , ym )T and g(y) is a vector function dened and

continuous in the domain G = {y Rm : y < H}, 0 < H < +.
Let system (3.9.9) have the following properties:
C1 . A unique solution passes through each point of the domain G.
C2 . g(0) = 0 and for all y G such that y = 0, we have g(y) = 0.
C3 . System (3.9.9) is a Wazewski system, i.e., the components
of the vector function g(y) are quasimonotonically increasing
functions.
Recall that for the dierential function g(y) to be quasimonotonically increasing, it is necessary and sucient that
gj
0 for
yi
i = j,
i, j = 1, 2, . . . , m.
Remark 3.8.1 If the condition C2 is satised, system (3.9.9)

has an equilibrium position y = 0 that is isolated (there are no other
equilibrium positions in the domain G).
By virtue of the assumptions, the nonnegative cone R+
m is an
invariant set for system (3.9.9).
Let us establish the asymptotic stability conditions for the zero
solution in R+
m.
K 0 = {y Rm : yi > 0, i = 1, . . . , m}, .. K 0 K + .
Denition 3.9.1 (see Aleksandrov and Platonov [1]) The
MartynyukObolenskii (MO) condition is satised for system (3.9.9)
if there exists a positive vector G such that g() < 0.
Theorem 3.9.2 For the zero solution of system (3.9.9) to be
asymptotically stable in a nonnegative cone, it is necessary and sufcient that this system satisfy the MO-condition.
182
The proof of this theorem is based on the following auxiliary statements.

Theorem 3.9.3 (see Spenier [1]) Let a continuius tangent vector
eld w(y) be given on a set Qr = {y R+
m : y = r, r = const > 0}
T
and be such that the inequality z w(y) 0, where Qr = Qr \ K 0 ,
holds for any y and z such that y Qr , z Qr , y T z = 0. Then
there exists a point y Qr such that w(
y ) = 0.
We choose a number r, 0 < r < H, and denote by Ar a set of
points y belonging to Qr and being such that for each y Ar , there
exists a number satisfying the condition g(y) = y. It is obvious
that if for y Qr , there exists (y), then
(y) =
y T g(y)
.
y2
(3.9.10)
Lemma 3.9.1 For any r (0, H), the set Ar is nonempty and is
a compact set, while function (3.9.10) is continuous on Ar and equal
to zero at no point of the set.
Proof. Let us rst show that Ar = . Consider the vector eld
w(y) = g(y)
y T g(y)
y.
y2
(3.9.11)
It is easy to verify that if y Qr , then y T w(y) = 0, i.e., the vector

eld (3.9.11) is tangent to Qr .
Let y Qr , z Qr , y T z = 0. Then there exists an ordered
set (i1 , . . . , ip ), where 1 i1 < i2 < . . . < ip m, 1 p < m, such
/ {i1 , . . . , ip }. Since
that yi > 0 for i {i1 , . . . , ip } and yi = 0 for i
the right-hand sides of system (3.9.9) are quasimonotonic, we have
/ {i1 , . . . , ip }, and since the vectors y and z are
gi (y) 0 for all i
orthogonal to each other, zi = 0 for i {i1 , . . . , ip }. Then

zi gi (y) 0.
z T w(y) = z T g(y) =
i{i
/ 1 ,...,ip }
According to Theorem 3.9.3, there exists a vector y Qr such

that w(
y ) = 0. The last equality can be written as g(
y ) = (
y )
y,
where (
y ) = = yT /
y 2 .
183
Thus, we have proved that the set Ar is non empty.

The vector function g(y) is continuous in the domain G. Therefore, Ar is a compact set, and the function (y) is continuous on Ar .
Moreover, since g(y) = (y)y and the equilibrium position y = 0 is
isolated, we have (y) = 0 on the set Ar .
Lemma 3.9.2 Let r (0, H). If
= max
yAr
y T g(y)
< 0,
y2
then there exists a point Qr such that > 0, g() < 0.

Proof. If B = K 0 Ar = , then for any point B we have
> 0, () < 0 and g() = (). Hence, g() < 0.
Let B = . Then Ar Qr . For each point y Ar , we construct
a neighborhood Uy such that the following inequalities hold for all
z Uy :
r 2
r
.
z y < , z T g(z) <
2
2
The sets Uy form an open covering of the compact set Ar . According
to Mazko [1], a nite subcovering U1 , . . . , Ul can be isolated in this
open covering. Let us set up continuous functions 1 (y), . . . , l (y)
such that s (y) > 0 for y Us and s (y) = 0 for y
/ Us , s = 1, . . . , l.
T
m
Let e = (1, . . . , 1) R . Consider the vector eld

l

y T g(y)
s (y) .
w(y) = g(y)
y+e

l

T
s=1
y
s (y)
y+e
(3.9.12)
s=1
The vector eld w(y) has the following properties:

(a) w(y) is tangent to Qr , i.e., y T w(y) = 0 for all y Qr ;
(b) the function w(y) is continuous on Qr ;

l
!
Us
(c) if y
Qr , z Qr , and z T y = 0, then
s=1

(y T g(y))(z T e)
s (y) > 0.

l

T
s=1
y
s (y)
y+e
l
z T w(y) = z T g(y)
s=1
184

l
Since z T e > 0 for y

Us , we have y T g(y) < 0,
s=1

l
l

s (y) > 0, y T y + e
s (y) > 0, and, as was shown
s=1
s=1
in the proof of Lemma 3.9.1, the inequality z T g(y) 0 holds

for all y, z Qr such that y T z = 0;
l

(d) if y Qr \
Us , z Qr , and z T y = 0, then
s=1
z T w(y) = z T g(y) 0.
Thus, the conditions of Theorem 3.9.3 are satised for the vector
eld (3.9.12). Then, there exists a point Qr such that

l

T g()
s () .
g() =

+e
l

T
s=1
+e
s ()
(3.9.13)
s=1
Us
The point does not belong to the set

T g() < 0 for
Qr because
s=1
Us , and since (3.9.9) is a Wazewski syss=1
tem, at least one coordinate of the vector g() must be nonnegative for Q
r . However, the point cannot belong to the set
Qr \
Us
either. Indeed, all the functions s (y), s = 1, . . . , l,
s=1
are equal to zero on this set, and relation (3.9.13) becomes

g() =
T g()
.
2
On one hand, Ar , and on the other hand, the sets U1 , . . . , Ul

form a covering Ar and, hence,
/ Ar . We arrive at a contradiction.
T
Thus,
/ Qr , ans g() < 0. Hence, > 0 and g() < 0.
Lemma 3.9.3 The sets D + = {y G R+
m : g(y) 0},
+
D = {y G Rm : g(y) 0} are positively invariant sets for
system (3.9.9).
Proof. Let y D + , and y(t, y) be the solution of system (3.9.9)
satisfying the initial condition y(0, y) = y. Denote by [0, ) the right
185
maximum interval of existence of this solution. Let us rst prove

y ), where R+
y ) = {y Rm : y y}.
that y(t, y) R+
m (
m (
Consider the function h(y) = g(y) g(
y ). We have h(y) g(y),
h(
y ) = 0. Hence, the system
dv
= h(v)
dt
has a solution v(t) y. Also
d
v (t)
g(
v (t)),
dt
dy(t, y)
g(y(t, y)),
dt
y(0, y) = v(0) = y.
But then, according to Rouche, Habets, and Laloy [1], the estimate
y ).
y(t, y) v(t) = y holds for all t [0, ), i.e., y(t, y) R+
m (
Let us now choose two times t1 and t2 such that 0 t1 < t2 < .
We get y(t2 t1 , y) y. Then y(t2 , y) y(t1 , y). Hence, all the
components of the vector y(t, y) are nondecreasing functions on the
interval [0, ). Therefore, g(y(t, y)) 0 for t [0, ), i.e., D + is a
positively invariant set.
It can similarly be shown that y D , then the solution y(t, y)
remains in the set K (
y ) = {y Rm : 0 y y} with time.
Then this solution can be extended to the interval [0, +) and all
its components on this interval are nonincreasing functions.
y < H/2. Then for the solution
Lemma 3.9.4 Let y D + ,
y(t, y) of system (3.9.9) emerging from the point y at t = 0, there
exists T > 0 such that y(T, y) = H/2.
Proof. Suppose that the solution y(t, y) remains in the domain
y < H/2 with time. Then it is dened on the interval [0, +).
In the proof of Lemma 3.9.3, we showed that all the components
of the vector y(t, y) are nondecreasing functions. Then, there exists
lim y(t, y) = z, where 0 < z H/2. Thus, the set of -limiting
t+
points of the trajectory of the solution consists of one point z. But

then the point z should be an equilibrium position for system (3.9.9)
(see Mazko [1]), which contradicts the property C2 of the system.
Lemma 3.9.5 Let there exists a vector such that > 0,
< H, g() < 0. Then the set K () is a positively invariant
186
set for system (3.9.9) and is in the attraction domain of the equilibrium position y = 0.
Proof. Consider the solution y(t, ) emerging from the point at
t = 0. According to Lemma 3.9.3, this solution is dened on the
interval [0, +), satises the condition y(t, ) K () for t 0,
and all its components on this interval are nonincreasing functions.
Then y(t, ) 0 as t +.
With Lemma 3.9.3, if y K (), then y(t, y) y(t, ) for
all t 0. Then y(t, y) K () for t [0, +) and y(t, y) 0 as
t +.
Proof of Theorem 3.9.2. Necessity. Let the zero solution of system
(3.9.9) be asymptotically stable in a nonnegative cone. If there exists
a number r, 0 < r < H, such that
= max
yAr
y T g(y)
< 0,
y2
then, according to Lemma 3.9.2, there exists a vector such that

= r, > 0, and g() < 0.
Assume now that the inequality > 0 holds for any r (0, H).
Then there is a point y = 0 in an indenitely small neighborhood of
the origin of coordinates such that y 0, g(
y ) = (
y )
y 0. Consider
the solution y(t, y) emerging from this point at t = 0. Without loss
of generality, we assume that
y < H/2. According to Lemma 3.9.4,
the solution falls onto the sphere y = H/2 at some time t = T .
Then, the equilibrium position y = 0 of system (3.9.9) is unstable.
Suciency. Let system (3.9.9) satisfy the MO-condition. We
choose a positive vector such that < H, g() < 0, and consider
the solution y(t, ) emerging from the point at t = 0.
According to Lemma 3.9.5, the set K () is in the attraction
domain of the equilibrium position y = 0.
Specify a positive number . For the solution y(t, ), we nd T > 0
such that y(t, ) < for t > T . The solutions of system (3.9.9)
have the property of integral continuity (see Mazko [1]). Therefore,
for the numbers > 0 and T > 0 it is possible to select > 0 such
that for 0 y (0) and y (0) < the following inequality holds
on the interval [0, T ]:
y(t, y (0) ) < .
(3.9.14)
187
But then estimate (3.9.14) will also be valid for t > T because
y(t, y (0) ) y(t, ) < for all t > T . Then the zero solution of
system (3.9.9) is asymptotically stable in a nonnegative cone. Theorem 3.9.2 is proved
Corollary 3.9.1 If the zero solution of system (3.9.9) is stable
on the cone R+
m , then it is asymptotically stable in this cone.
Indeed, when proving the necessity of Theorem 3.9.2, we showed
that if the zero solution is stable in the cone R+
m , it satises the MOcondition. From the proof of the suciency of the theorem, it follows
that the zero solution is asymptotically stable.
With Lemma 3.9.4, it is easy to prove the following theorem.
Theorem 3.9.4 Let the following conditions be satised:
(1) system (3.9.9) has properties C1 and C3 ;
(m) 0
(2) there exists a sequence of points y (m) R+
m such that y
(m)
(m)
for m , y
= 0, g(y ) 0;
(m) ) G.
(3) g(0) = 0, g(y) = 0 for y R+
m (y
Then the zero solution of system (3.9.9) is unstable in R+

m.
Remark 3.9.1 Theorem 3.9.2 allows us to reduce the analysis
of the zero solution of Wazewskis nonlinear autonomous system for
stability in R+
m to testing the solvability of an auxiliary system of
inequalities in a positive cone.
Example 3.9.2 Consider a proportional-plus-oating control system
dxi
= i xi + , i = 1, 2, . . . , n,
dt
n
(3.9.15)
d
=
ai xi p f (),
dt
i=1
where i > 0, p > 0, f () > 0 for = 0, f (0) = 0. Let us

estimate the ranges of parameters in which the state x = 0, = 0 of
system (3.9.15) is asymptotically stable. Replacing variable yi = 12 x2i ,
188
i [1, n], z = 12 2 , we reduce system (3.9.15) to the form

dyi
1
1 2
= i x2i + xi i x2i +
, i [1, n],
dt
2
2 i
n
dz
=
ai xi p 2 f ()
dt
i=1

n
n

x2i
|ai | 2
|ai |i 2p
f ().
2
i
2
i=1
i=1
According to Theorem 3.9.2, the asymptotic stability of the state

x = 0, = 0 follows from the asymptotic stability of the state y = 0,
z = 0 of the system
1
dyi
= i yi + z,
dt
i

n
n

|ai |
dz
=
|ai |i yi 2p
z g(z),
dt
i
i=1
i=1
where g(z) = min{f (), f ()}.

The state y = 0, z = 0 is asymptotically stable if and only if the
following inequalities are simultaneous:
z
, yi0 > 0, i = 1, 2, . . . , n,
2i

n
n

|ai |
|ai |i yi0 , z 0 > 0.
z 0 + g(z 0 ) >
2p
i
yi0 >
i=1
i=1
Performing some transformations of this system, we nd the estimate

n

|ai |
i=1
p,
(3.9.16)
which denes the range of parameters in which the state y = 0, z = 0

is asymptotically stable.
Remark 3.9.2 In Piontkovskii and Rutkovskaya [1], the range
of parameters for the system of equations (3.9.15) with n = 4 is
189
estimated as follows:

2
min(i )
i
4
X
i=1
|ai |2 <
p 2
4
(3.9.17)
It is obvious that when n = 4 estimate (3.9.16) defines a wider range

of parameters than estimate (3.9.17) does.
Example 3.9.3 Consider the system
li
k
dyi X
mj
=
pij yj
,
dt
i = 1, 2, . . . , k,
(3.9.18)
j=1
where mj and li are odd natural numbers, pij are constant coefficients, pij 0 for i 6= j, i, j = 1, 2, . . . , k.
Let P = (pij )ki,j=1 . It is easy to verify that if det P 6= 0, then system (3.9.18) has properties C1 C3 , where H is any positive number.
Then, the zero solution of this system is asymptotically stable in a
nonnegative cone if and only if there exist positive numbers 1 , . . . , k
satisfying the inequalities
k
X
mj
pij j
< 0,
i = 1, . . . , k,
j=1
whence it follows that for the zero solution of system (3.9.18) to be

asymptotically stable in a nonnegative cone, it is necessary and sufficient that the matrix P satisfy the SevastyanovKotelyanskii conditions.
Example 3.9.4 Consider the system
(
y1 = p11 y1l + p12 y2l ,
y2 = p21 y1m + p22 y2m ,
(3.9.19)
where l and m are odd natural numbers, pij are constant coefficients,
p12 0, p21 0.
190
With Theorem 3.9.2, it is easy to show that conditions p11 < 0,

l
m l
pm
11 p22 > p12 p21 are the necessary and sucient ones for the asymptotic stability of the zero solution of system (3.9.19) in a nonnegative
cone.
Remark 3.9.3 To apply Theorem 3.9.2, it is necessary that
no equilibrium positions, except y = 0 (property C2 ) exist in some
neighborhood of the origin of coordinates.
It is clear that the isolation of the equilibrium position is a necessary condition for its asymptotic stability. Note that if the system
has properties C1 and C3 and satises the MO-condition, and a positive vector such that g() < 0 exists in any indenitely small
neighborhood of the point y = 0, then the equilibrium position may
not be isolated. Hence, these conditions are not sucient for the
zero solution to be asymptotically stable in a nonnegative cone.
3.9.2
Matrix-valued Lyapunov functions method
We consider a problem on stability in a product space for a system

of dierential equations of the perturbed motion (cf. Lefschetz [1])
dy
= g(t, y) + G(t, y, z),
dt
dz
= h(t, z) + H(t, y, z).
dt
(3.9.20)
Here y Rp , z Rq , g : R+ Rp Rp , G : R+ Rp Rq Rp ,
h : R+ Rq Rq , H : R+ Rp Rq Rq . In addition, functions
g, G; h, H are continuous on R+ Rp , R+ Rq , R+ Rp Rq and
they vanish for y = z = 0.
The problem itself is to point out the connection between the
stability properties of equilibrium state y = z = 0 with respect to
system (3.9.20) on Rp Rq and its nonlinear approximation
dy
= g(t, y),
dt
dz
= h(t, z).
dt
(3.9.21)
191
Assumption 3.9.1 Let there exist the time-invariant neighborhood Ny Rp and Nz Rq of the equilibrium state y = 0 and
z = 0, respectively and let there exist a matrix-valued function

v11 (t, y) v12 (t, y, z)
U (t, y, z) =
(3.9.22)
v21 (t, y, z) v22 (t, z)
the element vij of which satisfy the estimations (cf. Krasovskii [1],
Djordjevic [1])
c11 y2 v11 (t, y) c11 y2
2
c22 z v22 (t, z) c22 z
for all
(t, y = 0) R+ Ny ;
for all
(t, z = 0) R+ Nz ;
c12 yz v12 (t, y, z) c12 yz

for all (t, y = 0, z = 0) R+ Ny Nz ;
v12 (t, y, z) = v21 (t, y, z)
cii > 0,
for all
cii > 0,
(t, y = 0, z = 0) R+ Ny Nz ;
cij , cij = const R,
i = j.
(3.9.23)
Assumption 3.9.2 Let there exist functions ij (t), i = 1, 2;

j = 1, 2, . . . , 8 which are bounded on any nite interval on R+ such
that

v11 T
v11
+
g 11 y2 ;
t
y

v11 T
v22
+
G 12 y2 + 13 yz;
t
y

v22 T
h 21 z2 ;
z

v22 T
(3.9.24)
H 22 z2 + 23 yz;
z

v12 T
v12
+
g 14 y2 + 15 yz;
t
y

v12 T
G 16 y2 + 17 yz + 18 z2 ;
y

v12 T
h 24 z2 + 25 yz;
z
192
v12
z
T
H 26 y2 + 27 yz + 28 z2 .
For system (3.9.20) we have the following result.

(1) all conditions of Assumptions 3.9.1, 3.9.2 are fullled;
(2) the matrices

c11 c12
C=
,
c21 c22

c12 = c21 ,
C=

c11 c12
,
c21 c22
c12 = c21 ,
are positive denite;

1 T
(S (t) + S(t)) where
2

11 12
,
12 = 21
S(t) =
21 22
(3) the matrix M (t) =
and
11 = 12 (11 + 12 ) + 21 2 (14 + 16 + 26 );
22 = 22 (21 + 22 ) + 21 2 (18 + 24 + 28 );

1 2
1 13 + 23 22 + 1 2 (15 + 25 + 17 + 27 ),
12 =
2
1 , 2 being positive numbers, have the characteristic roots
1 (t), 2 (t) and let Re i (t) , for all t t0 .
Then the state of equilibrium y = z = 0 of the system (3.9.20)
is uniformly stable if the number is equal to zero and exponentially
stable if < 0.
If conditions of Assumptions 3.9.1, 3.9.2 are fullled for Ny =
Rp , Ny = Rq and conditions (2), (3) of the theorem hold, then the
equilibrium state y = z = 0 of the system (3.9.20) is uniformly stable
in the whole if = 0 and exponentially stable in the whole if < 0.
193
Proof The proof uses the direct method. On the basis of estimations (3.9.23), it is not dicult to show that the function
v(t, y, z) = T U (t, y, z) satises the estimates
uT T Cu v(t, y, z) uT T Cu,
(3.9.25)
uT
where
= (y, z), = diag [1 , 2 ].
Also, in view of Assumption 3.9.1 and the estimates (3.9.24), the
derivative Dv(t, y, z) dened by Dv(t, y, z) = T DU (t, y, z) satises
(3.9.26)
Dv(t, y, z) uT S(t)u = uT M (t)u uT u.
By virtue of (2) and (3) and the inequalities (3.9.25), (3.9.26), we
see that all conditions of Theorems 1.2.1, 1.2.5 from the book by
Martynyuk [4] are veried for the function v(t, y, z) and its derivative.
Hence the proof is complete.
If in estimate (3.9.24) we change the sign of inequality for the
opposite one, then by means of the method similar to the given one
we can obtain the estimate

Dv(t, y, z) uT S(t)u
which allows us to formulate instability conditions for the equilibrium
state y = z = 0 of system (3.9.20) on the basis of Theorem 1.2.7
from the book by Martynyuk [4].
The statement of Theorem 3.9.5 shows that uniform stability or
exponential stability of the equilibrium state y = z = 0 of system
(3.9.20) can hold even if the equilibrium state y = z = 0 of system
(3.9.21) has no properties of asymptotic quasi-stability (cf. Lefschetz
[1]).
Example 3.9.5 Consider the system describing the motion of two
nonautonomously connected oscillators
dx1
dt
dx2
dt
dy1
dt
dy2
dt
= 1 x2 + v cos ty1 v sin ty2 ,

= 1 x1 + v sin ty1 + v cos ty2 ,
= 2 y2 + v cos tx1 + v sin tx2 ,
= 2 y1 + v cos tx2 v sin tx1 ,
(3.9.27)
194
where 1 , 2 , v, , + 1 2 = 0, are some constants.

For the independent subsystems
dx1
= 1 x2 ,
dt
dy1
= 2 y2 ,
dt
dx2
= 1 x1 ,
dt
dy2
= 2 y1 ,
dt
(3.9.28)
the auxiliary function vii , i = 1, 2, is taken in the form

v11 (x) = xT x, x = (x1 , x2 )T ,
v22 (y) = y T y,
y = (y1 , y2 )T .
(3.9.29)
We use the equation (3.9.28) to determine the nondiagonal element v12 (x, y) of the matrix-valued function U (t, x, y) = [vij ()],
i, j = 1, 2. To this end set = (1.1)T and v12 (x, y) = xT P12 y,
where P12 C 1 (T , R22 ). For the equation

dP12
0 1
0
2
cos t sin t
+
P12 +P12
+2v
= 0,
1
0
2 0
sin t cos t
dt
(3.9.30)
the matrix

2v
sin t cos t
P12 =
+ 1 2 cos t sin t
is a partial solution bounded for all t T .
Thus, for the function v(t, x, y) = T U (t, x, y) it is easy to establish the estimate of (3.9.25) type with matrices C and C in the
form

c11 c12
c11 c12
C=
, C=
,
c12 c22
c12 c22
.
where c11 = c11 = 1; c22 = c22 = 1; c12 = c12 = |+|2v|
1 2 |
T
T
Besides, the vector u1 = (x, y) = u2 since the system (3.9.28)
is linear.
For system (3.9.27) the estimate (3.9.26) becomes
&
Dv(t, x, y)&(3.9.28) = 0
for all (x, y) R2 R2 because M = 0.
Cone Valued Lyapunov Functions
195
Due to (3.9.30) the motion stability conditions for system (3.9.27)

are established based on the analysis of matrices C and C property
of having xed sign.
It is easy to verify that the matrices C and C are positive denite,
if
4v 2
> 0.
1
( + 1 2 )2
Consequently, the motion of nonautonomously connected oscillators
is uniformly stable in the whole, if
|v| <
3.10
1
| + 1 2 |.
2
As we have noted in the last section, an unpleasant fact in the approach of several Lyapunov functions is the requirement of quasimonotone property of the comparison system. Since comparison system with a desired stability property can exist without satisfying
the quasi-monotone property, the limitation of this otherwise eective technique is obvious. It is observed that this diculty is due
to the choice of the cone relative to the comparison system, namely,
N , the cone of nonnegative elements of RN and a possible answer
R+
N to work in a given
lies in choosing a suitable cone other than R+
situation.
Using the comparison results 3.1.3 and 3.1.4, it is now easy to
discuss the method of cone valued Lyapunov functions. We shall
merely state two typical results.
(i) V C[R+ S(), K], V (t, x) is locally Lipschitzian in x relative to the cone K RN and for (t, x) R+ S(),
K
D + V (t, x) g(t, V (t, x));

(ii) g C[R+ K, RN ], g(t, 0) 0 and g(t, u) is quasi-monotone
nondecreasing in u relative to cone K;
196
(iii) f (t, 0) 0 and for some 0 K0 , 0 (V (t, x)) is positive denite and decrescent for (t, x) R+ S(), where K0 = K \ {0}
and K0 is the adjoint of K0 .
Then, the stability properties of the trivial solution u = 0 of
u = g(t, u),
u(t0 ) = u0 ,
(3.10.1)
imply the corresponding stability properties of the trivial solution

x = 0 of (3.6.1).
The following version of Theorem 3.10.1 is in a more exible setting so as to be useful in applications.
(i) P and Q are two cones in RN such that P Q;
(ii) V C[R+ S(), Q], V (t, x) is locally Lipschitzian in x relative to P and
P
D + V (t, x) g(t, V (t, x)),
(t, x) R+ S();
(iii) g C[R+ Q, RN ], g(t, 0) 0 and g(t, u) is quasi-monotone

nondecreasing in u relative to P ; and
(iv) f (t, 0) 0 and for some 0 Q0 , 0 (V (t, x)) is positive denite and decrescent for (t, x) R+ S().
Then the stability properties of the trivial solution u = 0 of (3.10.1)
imply the corresponding stability properties of (3.6.1).
N , = (1, 1, . . . , 1), we obtain Theorem 3.9.1 from
If K = R+
0
N

Vi (t, x) = V0 (t, x). One could
Theorem 3.10.1 since 0 (V (t, x)) =
i=1
also use other measures in place of 0 (V (t, x)). For example, let
C[K, R+ ], (u) is nondecreasing in U relative to K. Then it
is enough to suppose (V (t, x)) be positive denite and decrescent
N in Theorem 3.10.2,
in Theorem 3.10.1. Moreover, if P Q = R+
the unpleasant fact concerning the quasimonotonicity of g(t, u) mentioned earlier can be removed. This, of course, means that we have
197
to choose an appropriate cone P which necessarily depends on the

nature of g(t, u). Let us demonstrate this by means of a simple
example.
Example 3.10.1 Consider the comparison system
u1 = a11 u1 + a12 u2 , u1 (t0 ) = u10 ,
u2 = a21 u1 + a22 u2 , u2 (t0 ) = u20 .
(3.10.2)
2 . Suppose that we do not demand a , a

Let Q = R+
12
21 to be nonnegative. Then, the function g(t, u) violates the quasi-monotone
nondecreasing condition in u = (u1 , u2 ) relative to Q. Hence, the
dierential inequalities
D + V1 (t, x) g1 (t, V1 (t, x), V2 (t, x)),

D + V2 (t, x) g2 (t, V1 (t, x), V2 (t, x)),
(3.10.3)
do not yield the componentwise estimates of V (t, x(t)) in terms of

the maximal solution of (3.10.2).
Suppose now that there exist two numbers , such that
0 < < and
2 a21 + a22 a11 + a12 ,
(3.10.4)
2 a21 + a22 a11 + a12 .
(3.10.5)
These conditions can hold with no restriction of nonnegativity of a21

2 dened by
and a12 . We shall now choose the cone P Q = R+
2
: u2 u1 u2 }.
P = {u R+
This cone has two boundaries u2 = u1 and u2 = u1 . On the

boundary u2 = u1 , we take = ( 1 , 1) so that

(
'
u1
1
=0
, 1 , u1 ,
and
'

(
u1
u1
1
0 for all u = 0.
, a , a11 u1 + a12 , a21 u1 + a22
198
This reduces to the condition (3.10.4). Similarly, we can obtain

(3.10.5). Thus, if the inequalities (3.10.3) are relative to P , we obtain
the componentwise estimates on V as
Vi (t, x(t)) ri (t, t0 , V (t0 , x0 )),
(3.10.6)
by Theorem 3.1.4. We note that the estimate (3.10.6) is precisely the

one we would have obtained if a12 , a21 0, by the standard method
of vector Lyapunov functions. Since a12 , a21 need not be nonnegative
in our example, the usefulness of cone-valued Lyapunov function is
clear.
Example 3.10.2 Let in an isodromic control system
di
= i i + ai , i = 1, 2, . . . , n,
dt
n

d
= a1 1
ai i n+1 f (),
dt
(3.10.7)
i=2
the conditions 0 < 1 < min(k ), k [2, n + 1], f () > 0, for each
k
= 0, f (0)=0, ik > 0 be fullled.

Remark 3.10.1 In contrast to the classical statement of the
problem on stability of isodromic control system here we take into
account the signs before coecients ai and this allows us to separate
from the general class of isodromic control systems those generating
quasimonotone semi-groups. Also it should be noted that the signs
of coecients ai are really dierent when the real automatic control
systems are under consideration.
Systems of equations (3.10.1) generate quasimonotone
semi)
n+1
:
groups with respect to a circular cone K = (i , ) R
*
n

2i + 2 .
1 0, 21
i=2
The zero solution of (3.10.1) is uniformly asymptotically stable if

and only if the system of inequalities
1 1 + a1 < 0,
199
Notes
n

(1 1 + a1 )
(i i + ai )2
2
i=2

+
2
n

(ai i n+1 g()) ,
i=2
1 0,
n

2i + 2
21
i=2
has solution. This is valid if and only if

n+1 +
n

a2
i
i=2
>
a21 f ()
(3.10.8)
for = 0.
Remark 3.10.2 In Martynyuk and Obolensky [1] sucient conditions for uniform asymptotic stability were estableshed as follows
n+1
n

a2
i
i=2
(3.10.9)
where the signs before coecients ai were not taken into account. It
is easily seen that condition on parameter (3.10.8) extends (3.10.9).
3.11
Notes
For various comparison results given in Section 3.1 see Lakshmikantham and Leela [1, 4] and Martynyuk and Obolensky [2]. Stability
in terms of two measures was introduced by Movchan [1] and successfully developed by Salvadori, see Bernfeld and Salvadori [1].
The contents of Sections 3.2 and 3.3 are modelled on the basis
of the works of Bernfeld and Salvadori [1] and Lakshmikantham and
Leela [1]. Corollary 3.3.1 is due to Marachkov [1].
The converse theorem given in Section 3.4 is due to Larshmikanthan and Salvadori [1].
Section 3.5 contains new results and the results of Section 3.6 are
taken from Krasovski [1].
200
Theorem 3.7.1 may be found in Lakshmikantham and Leela [3]

while Theorems 3.7.2 and 3.7.3 are from Salvadori [1, 2, 3] which are
generalizations of a result of Matrosov [1].
The contents of Section 3.8 are based on the work of Jane Moore
[1].
The method of vector Lyapunov functions described in Section
3.9 is due to Matrosov [2, 3] and is taken from Lakshmikantham
and Leela [1], Martynyuk and Obolensky [1, 2], and Aleksandrov
and Platonov [1]. See also Bellman [2], Lakshmikantham [1], and
C. Corduneanu [1]. For the method of matrix Lyapunov functions
see Martynyuk [16, 13].
Section 3.10 contains the work of Lakshmikantham and Leela [4].
See Martynyuk [8, 11, 12] for related results in matrix Lyapunov
functions. For large scale dynamic systems see Siljak

[1]. For various results on stability of motion by Lyapunovs direct method see
the books of Bhatia and Szeg
o [1], Cesari [1], Hahn [1], Malkin [1],
Yoshizawa [1], Zubov [1], Krasovski [1], LaSalle and Lefschetz [1],
Lakshmikantham and Leela [1], and Rouche, Habets and Laloy [1].
For recent development in the method of vector Lyapunov functions
see Martynyuk [16] and Aleksandrov and Platonov [2].
4
STABILITY OF PERTURBED
MOTION
4.0
Introduction
In this chapter, stability considerations are extended to a variety of nonlinear systems utilizing the same versatile tools, namely,
Lyapunov-like functions, theory of appropriate inequalities and different measures, that were developed in the previous chapters. In order to avoid monotony, we have restricted ourselves to present only
typical extensions which demonstrate the essential unity achieved
and pave the way for further work.
Sections 4.1 and 4.2 deal with stability criteria for perturbed differential equations. The main features are the use of converse theorem for uniform asymptotic stability in terms of two measures and
the coupled comparison equation which depends on the solutions of
the given system. In Section 4.3, a technique in perturbation theory
is presented which combines Lyapunov method and the method of
variation of parameters to help preserve the inherent rich behavior
of perturbations.
Section 4.4 is devoted to the extension of Lyapunov method to
dierential equations with innite delay using both Lyapunov functionals and functions. A unied approach is presented which is parallel to the corresponding theory for dierential equations without
delay. In Section 4.5 we describe a technique which oers better
qualitative information compared to the method of Section 4.4. The
idea is to use upper and lower estimates simultaneously together
with certain auxiliary functions so as to obtain growth of solutions

201
202
Stability of perturbed motion
in a best possible way. Section 4.6 investigates stability criteria for

integro-dierential equations of Volterra type following the pattern of
Section 4.4, while Section 4.7 employs variational systems and their
nice properties. Here given integro-dierential system, a method of
nding an equivalent linear dierential system is developed which is
then utilized to discuss the stability properties of nonlinear integrodierential systems.
Section 4.8 concentrates on the stability theory of dierence equations, while Section 4.9 investigates impulsive dierential equations
extending Lyapunov method to such systems. It is shown that impulsive eects can help or destroy the qualitative behavior. Finally
in Section 4.10, we devote our attention to weakly coupled reactiondiusion systems and study the stability properties via Lyapunov
methods.
4.1
Stability of Perturbed Motion in Two Measures
In order to unify the investigation of stability properties of perturbed

motion, it is useful to utilize coupled comparison functions as in Theorem 3.1.5. Of course, the use of coupled functions is also benecial
in the study of stability properties of unperturbed motion, since estimating D + V (t, x) by a function of t, x and V is more advantageous
than by a function of t and V only.
Let us consider the perturbed dierential system
x = f (t, x) + R(t, x),
x(t0 ) = x0 ,
(4.1.1)
where f, R C[R+ Rn , Rn ] together with the unperturbed system

(3.1.1). We can dene stability concepts of the trivial solution u = 0
of
(4.1.2)
u = g(t, x(t), u), u(t0 ) = u0 ,
where g C[R+ Rn R+ , R] and x(t) = x(t, t0 , x0 ) is any solution
of (4.1.1) existing on [t0 , ), as follows
Denition 4.1.1 The trivial solution of (4.1.2) is said to be h0 conditionally equistable, if given any > 0 and t0 R+ , there exist
Stability of perturbed motion in two measures
positive functions 1 = 1 (t0 , ),

in t0 such that
203
2 = 2 (t0 , ) that are continuous
t t0 ,
(4.1.3)
where u(t, t0 , x0 , u0 ) is any solution of (4.1.2), h0 , h are functions as
in Section 3.2.
h0 (t0 , x0 ) < 2
and u0 < 1
imply u(t, t0 , x0 , u0 < ),
The other stability concepts can be dened in a similar fashion.

Note that in case of h0 = x, the above denition reduces to partial stability of the trivial solution x = 0, u = 0 of the combined
systems (4.1.1) and (4.1.2) with respect to the component u.
Let us prove the following result.
(i) h0 , h and h0 is uniformly ner than h;
h-positive denite and h0 -decrescent;
(iii) g C[R+ Rn R+ , R], g(t, 0, 0) 0 and for (t, x) S(h, p),
for some > 0,
D + V (t, x) g(t, x, V (t, x)).
Then any one of the h0 -conditional stability properties of the
trivial solution u = 0 of (4.1.2) implies the corresponding (h0 , h)stability property of the system (4.1.1).
Proof The proof is very much similar to the proof of Theorem 3.3.1
except that we now employ Theorem 3.1.5 instead of Theorem 3.1.1.
The relations (3.3.1), (3.3.3) and (3.3.4) remain the same. Assuming
h0 -conditional equistability of the trivial solution of (4.1.2), we have
1 , 2 satisfying (4.1.3) when t0 R+ and b() is given. We set
= min(, 2 ) where is the same one dened in the proof of
Theorem 3.3.1. With this , one can show, as in Theorem 3.3.1,
that (h0 , h)-equistability holds for the system (4.1.1). Based on these
modications, it is not dicult to construct proofs of other stability
properties and hence, the theorem is proved.
204
Let us next apply the converse Theorem 3.4.1 to discuss total

stability of the system (3.1.1) in the general context of two measures.
Denition 4.1.2 The system (3.1.1) is said to be (h0 , h)-totally
stable, if given > 0 and t0 R+ , there exist two numbers 1 , 2 > 0
such that
h0 (t0 , x0 ) < 1 and R(t, x) < 2 for (t, x) S(h, ),
(4.1.4)
imply h(t, y(t, t0 , x0 )) < , t t0 where y(t, t0 , x0 ) is any solution of

the perturbed system (4.1.1).
Theorem 4.1.2 Suppose that the assumptions (i) and (ii) of
Theorem 3.4.1 hold. Then, the system (3.1.1) is (h0 , h)-totally stable.
Proof Let U, W C[S(h, p), R+ ] be two Lyapunov functions
which satisfy the conditions of Theorem 3.4.1. By (b), (d) and the
boundedness of W , we see that the unperturbed system (3.1.1) is
(h, U ) quasi-uniformly asymptotically stable and that is a constant
associated with this property. Then, given > 0, there exists a
T () > 0 such that
(t, x) S(h, p) implies U (, x(, t, x)) < for t + T ().
(4.1.5)
Let (0, ) and t0 R+ be given. Choose 1 , 2 > 0 so that
1 < 0
and a(1 ) < b(),
(4.1.6)
and
k2 eM < a(1 )/2,
a(1 ) + k2 < b(),
(4.1.7)
where k > 0 is a Lipschitz constant for U and = T (b(1 )/2). Let

x0 and R be given in such a way that (4.1.4) is satised. Then, let
us suppose that for a solution y(t, t0 , x0 ) of (4.1.1) and a t t0 , we
have h(t, y(t, t0 , x0 )) . Since
U (t0 , x0 ) < a(1 ) < b()
and U (t, y(t, t0 , x0 )) b(),
it is clear that there exist t1 , t2 > t0 , t2 > t1 , such that

U (t1 , y(t1 , t0 , x0 )) = a(1 ),
U (t2 , y(t2 , t0 , x0 )) = b()
Stability of perturbed motion (Continued)
205
and
a(1 ) U (t, y(t, t0 , x0 )) b()
on
[t1 , t2 ].
Setting x1 = y(t1 , t0 , x0 ), if t2 t1 , by Gronwall-inequality we

have
|U (t1 + , y(t1 + , t1 , x1 )) U (t1 + , x(t1 + , t1 , x1 ))| k2 eM .
Thus, by means of (4.1.5) and (4.1.7), we obtain
U (t1 + , y(t1 + , t1 , x1 )) < a(1 )
which is a contradiction. Therefore, t2 t1 < . Because of the
condition (b) of Theorem 3.4.1, we get for t [t1 , t2 ],
D + U (t, y(t, t0 , x0 )) D + U (t, y(t, t0 , x0 ))
+ kR(t, y(t, t0 , x0 )) k2 .
By (4.1.7) we then have
b() U (t2 , y(t2 , t0 , x0 )) U (t1 , y(t1 , t0 , x0 )) + k2 (t2 t1 )
a(1 ) + k2 < b()
which is a contradiction. The proof is complete.
4.2
Stability of Perturbed Motion (Continued)
We shall consider in the rest of this chapter only the case h =

h0 = x to avoid technical complications. Relative to the coupled system (4.1.2), we need another related stability concept which
we shall dene below.
Denition 4.2.1 The trivial solution of (4.1.2) is said to be
(i) uniformly stable, if given 0 < < and t0 R+ , there exists a = () > 0 such that x0 < and u0 < imply
u(t, t0 x0 , u0 ) < on any interval t0 t t1 , where x(t) ;
(ii) quasi uniformly asymptotically stable, if given 0 < < and
t0 R+ , there exist positive numbers 0 > 0 and T = T () > 0
such that x0 < 0 and u0 < 0 imply u(t, t0 x0 , u0 ) < for
t t0 + T whenever x(t) < , t t0 .
206
In the Denition 4.2.1, x(t) = x(t, t0 , x0 ) is any solution of (4.1.1)

and u(t, t0 , x0 , u0 ) is any solution of (4.1.2).
One can prove a result similar to Theorem 4.1.1 which we merely
state.
(i) V C[R+ S(), R+ ], V (t, x) is locally Lipschitzian in x, V
is positive denite and decrescent;
(ii) g C[R+ S() R+ , R], g(t, 0, 0) 0 and for (t, x)
R+ S(),
D + V (t, x) g(t, x, V (t, x)).
Then any one of stability properties of the trivial solution of coupled system (4.1.2) implies the corresponding stability property of
the trivial solution of (4.1.1).
Let us now discuss an important special case of (4.1.2). Suppose
that
g(t, x, u) = c(u) + w(t, x)
(4.2.1)
where c K and |w(t, x)| (t) whenever x and
t+1
(s) ds 0
as
t .
We claim that u = 0 if (4.1.2) is uniformly asymptotically stable.

For this purpose, let us rst prove uniform stability. Using the assumption on (t), we note that
t
t
() d =
t0

() ds d
t0

s+1
t
t
() d ds =
G(s) ds,
t0 1
t0 1
(4.2.2)
Stability of perturbed motion (Continued)
where G(t) =
t+1

207
(s)ds. Let
Q(t) = sup[G(s) : t 1 s < ]

so that Q L. Let 0 < < and t0 R+ be given. Choose a
= () > 0 such that 2 < and a = () 1 so that
2Q( ) < min(c(), ).
Let x0 < , t0 and x(t) for t [t0 , t1 ]. Also let u0 <
and if possible, suppose that u(t1 ) = u(t1 , t0 , x0 , u0 ) . Then there
would exist a t2 t0 such that u(t2 ) = and u(t) for
t [t2 , t1 ]. We then have
t1
u(t1 ) u(t2 )
t1
|w(s, x(s))| ds
c(u(s)) ds +
t2
t2
t1
c()(t2 t1 ) +
G(s) ds
t2
(t1 t2 )[c() + Q( )] + Q( )

+ Q( ) < + = ,
2 2
which is a contradiction. Hence u = 0 is uniformly stable.
Next, we shall prove quasi uniform asymptotic stability. Taking
= , set 0 = () and 0 = (). Because of uniform stability
of x = 0, it follows that x0 < 0 implies x(t) < , t t0 . Let
0 < < and t0 R+ be given. Choose = () and = () as
before. Choose
T = [c() () + 2Q(1) + 2]/c() > ()
and note that T = T () only. Let us suppose that u0 < 0 but
u(t) for t [t0 + , t0 + T ]. Then we get
0 < u(t0 + T ) u(t0 + ) + [c() + Q(t0 + )](T )
1
+ Q(t0 + ) c()(T ) + Q(1) = 0
2
208
which is a contradiction. Thus, there exists a t1 such that

u(t1 , t0 + , x0 , u0 ) <
which implies
u(t, t0 + , x0 , u0 ) < for
t t0 + T.
Hence u = 0 is quasi uniformly asymptotically stable, proving the

claim.
The prof is complete.
We note that whenever either (t) 0 as t or L1 [0, ]
t+1

(s)ds 0 as t and hence these cases are inwe have
t
cluded in foregoing discussion. The special case discussed above

results when x = 0 of (3.1.1) is uniformly asymptotically stable
and w(t, x) = KR(t, x), K being the Lipschitz constant for the
Lyapunov function obtained by Masseras converse theorem. Consequently, one can conclude the quasi uniform asymptotic stability of
u = 0 of (4.1.2) by Theorem 4.2.1.
Sometimes, it is more advantageous to discuss directly the inequality
(4.2.3)
D + V (t, x) g(t, x, V (t, x))
rather than the corresponding comparison equation (4.1.2). In this
approach, we can also weaken the assumption on the perturbation
term.
Let us assume that in (4.2.1), we have |w(t, x)| (t) whenever x < for any 0 < < , satisfying the condition
t+1

(s) d(s) 0 as t . Then we claim that x = 0 of (4.1.1)
t
is quasi uniform asymptotically stable. Proceeding as before, let us

choose and as follows:
2a() < b()
and 2Q( ) < min(c(), b()),
where a, b K are the functions resulting from properties of V . If

x0 < and there exist t1 , t2 such that, for t0 ,
x(t2 ) = ,
x(t1 ) =
and x(t) < ,
t [t2 , t1 ],
A technique in perturbation theory
209
then, using (4.2.3), we are led to the contradiction

b() V (t1 , x(t1 ))
a(x(t2 )) + [Q( ) c()](t1 t2 ) + Q( ) + a()
Q( ) + a() < b().
Hence x = 0 is uniformly stable. Proceeding again, as before, we
choose T () satisfying
T = [c() () + 2Q(1) + 2a()]/c() > ()
and suppose, if possible, that x(t) for t [t0 + , t0 + T ]
whenever x0 < 0 . Then, setting y0 = x(t0 + , t0 , x0 ) and using
(4.2.3), we are again led to the contradiction
0 < b() V (t0 + T, x(t0 + T, t0 + , y0 ))
1
a() + Q(1) + c()(T ) = 0.
2
Hence there exists a t1 [t0 + , t0 + T ] such that x(t1 ) < and
consequently, we can conclude that x = 0 of (4.1.1) is quasi uniform
asymptotically stable, completing our discussion.
4.3
A Technique in Perturbation Theory
A study of the eect of perturbations of dierential equations depends on the method employed and on the nature of perturbations.
One of the most used techniques is that of Lyapunov method and
the other is the nonlinear variations of parameters. These methods
dictate that we measure the perturbations by means of a norm and
thus destroy the ideal nature, if any of the perturbing terms.
In this section, we develop a new comparison theorem that connects the solutions of perturbed and unperturbed dierential systems
in a manner useful in the theory of perturbations. This comparison
result bounds, in a sense, the two approaches mentioned earlier and
consequently provides a exible mechanism to preserve the nature
of perturbations. The results that are given in this section show
that the usual comparison theorem (Theorem 3.1.1) in terms of a
210
Lyapunov function is included as a special case and that the perturbation theory could be studied in a more fruitful way.
Consider the two dierential systems
and
y = f (t, y),
y(t0 ) = x0
(4.3.1)
x = F (t, y),
x(t0 ) = x0
(4.3.2)
where f, F C[R+
Relative to the system (4.3.1), let
us assume that the following assumption (H) holds:
S(), Rn ].
(H) the solutions y(t, t0 , x0 ) of (4.3.1) exist for all t t0 ,

are unique and continuous with respect to the initial data and
y(t, t0 , x0 ) is locally Lipschitzian in x0 .
Let x0 < and suppose that y(t, t0 , x0 ) < for t0 t < T .
For any V C[R+ S()], R+ and any xed t (t0 , T ), we dene
D V (s, y(t, s, x))
1
lim inf [V (s + h, y(t, s + h, x + hF (s, x)))
h
h0
V (s, y(t, s, x))],
(4.3.3)
for t0 < s t and x S().

The following comparison result which relates the solutions of
(4.3.2) to the solutions of (4.3.1) is an important tool in the subsequent discussion.
Theorem 4.3.1 Assume that the assumption (H) holds. Suppose
that
(i) V C[R+ S(), R+ ], V (s, x) is locally Lipschitzian in x and
for t0 < s t, x S(),
D V (s, y(t, s, x)) g(s, V (s, y(t, s, x)));
(4.3.4)
2 , R] and the maximal solution r(t, t , u ) of

(ii) g C[R+
0 0
u = g(t, u),
exists for t0 t < T .
u(t0 ) = u0 0
(4.3.5)
211
Then, if x(t) = x(t, t0 , x0 ) is any solution of (4.3.2), we have

V (t, x(t, t0 , x0 )) r(t, t0 , u0 ),
t0 t < T,
(4.3.6)
provided V (t0 , y(t, t0 , x0 )) u0 .

Proof Let x(t) = x(t, t0 , x0 ) be any solution of (4.3.2) such that
x0 < . Set
m(s) = V (s, y(t, s, x(s))),
t0 s t,
so that m(t0 ) = V (t0 , y(t, t0 , x0 )). Then using the assumptions (H)
and (i), it is easy to obtain
D + m(s) g(s, m(s)),
t0 < s t,
which yields by Theorem 1.5.2 the estimate

m(s) r(s, t0 , u0 ),
t0 s t,
(4.3.7)
provided m(t0 ) u0 . Since

m(t) = V (t, y(t, t, x(t))) = V (t, x(t, t0 , x0 )),
the desired result (4.3.6) follows from (4.3.7) by setting s = t.
Taking u0 = V (t0 , y(t, t0 , x0 )), the inequality (4.3.6) becomes
V (t, x(t, t0 , x0 )) r(t, t0 , V (t0 , y(t, t0 , x0 ))),
t0 t < T,
(4.3.8)
which shows the connection between the solutions of systems (4.3.1)

and (4.3.2) in terms of the maximal solution of (4.3.5).
A number of remarks can now be made:
(1) The trivial function f (t, y) 0 is admissible in Theorem
4.3.1 to yield the estimate (4.3.6) provided V (t0 , x0 ) u0 . In this
case, y(t, t0 , x0 ) = x0 and hypothesis (H) is trivially veried. Since
y(t, s, x) = x, the denition (4.3.3) reduces to
1
D + V (s, x) lim sup [V (s + h, x + hF (s, x)) V (s, x)]
h0+ h
(4.3.9)
212
which is the usual denition of generalized derivative of the Lyapunov

function relative to the system (4.3.2). Consequently, Theorem 4.3.1
reduces, in this special case, to Theorem 3.1.1.
(2) Suppose that f (t, y) = A(t)y where A(t) is an n by n continuous matrix. The solutions y(t, t0 , x0 ) of (4.3.1) then satisfy
y(t, t0 , x0 ) = (t, t0 )x0 ,
where (t, t0 ) is the fundamental matrix solution of y = A(t)y,
with (t0 , t0 ) = I (identity matrix). The assumption (H) is clearly
veried. Suppose also that g(t, u) 0. Then (4.3.6) yields
V (t, x(t, t0 , x0 )) V (t0 , (t, t0 )x0 ),
t t0 .
(4.3.10)
If V (t, x) = x, (4.3.10) leads to

x(t, t0 , x0 ) (t, t0 )x0 ),
t t0 .
(4.3.11)
If, on the other hand, g(t, u) = u, > 0, we get a sharper estimate

V (t, x(t, t0 , x0 )) V (t0 , (t, t0 )x0 ) exp((t t0 )),
t t0 .
(4.3.12)
which, in the special case V (t, x) = x, reduces to

x(t, t0 , x0 ) (t, t0 )x0 ) exp((t t0 )),
t t0 .
(4.3.13)
Clearly the relation (4.3.13) helps in improving the behavior of

solutions of (4.3.2) relative to the behavior of solutions of (4.3.1).
This is a great asset in perturbation theory and it can be seen by
setting
F (t, x) = f (t, x) + R(t, x)
where R(t, x) is perturbation term.
(3) Suppose that f (t, y) is nonlinear, fy (t, y) exists and is continuous for (t, y) R+ S(). Then, the solutions y(t, t0 , x0 ) are
dierentiable with respect to (t0 , x0 ) and by Theorem 2.1.2, we have
y
(t, t0 , x0 ) = (t, t0 , x0 )f (t0 , x0 ),
t0
y
(t, t0 , x0 ) = (t, t0 , x0 ),
0
t t0
(4.3.14)
213
where (t, t0 , x0 ) is the matrix solution of the variational equation

z = fy (t, y(t, t0 , x0 ))z.
If V (s, x) is also assumed to be dierentiable, then by (4.3.14),
we have, for a xed t,
D V (s, y(t, s, x)) Vs (s, y(t, s, x))
+ Vx (s, y(t, s, x))(t, s, x)[F (s, x) f (s, x)].
(4.3.15)
The relation (4.3.15) gives an intuitive feeling of the denition (4.3.3).

If, in addition, V (t, x) = x2 and f (t, x) = f (t, x) + R(t, x), (4.3.15)
yields
D V (s, y(t, s, x)) 2y(t, s, x)(t, s, x)R(s, x)
which exhibits how the perturbation term is involved in the computation.
(4) When the solutions of (4.3.1) are known, a possible Lyapunov
function for (4.3.2) is
W (s, x) = V (s, y(t, s, x)).
(4.3.16)
where V (s, x) and y(t, s, x) are as before. One could use V (s, x) =
x so that W (s, x) = y(t, s, x) is a candidate for Lyapunov function for (4.3.2). If y(t, s, x) x, condition (4.3.14) reduces to
lim
h0
1
[x + hF (t, x) x] g(t, x)
h
which is an often used assumption in comparison results.

As an application of Theorem 4.3.1, we shall consider some results
on stability and asymptotic behavior of solutions of (4.3.2).
Theorem 4.3.2 Assume that (H) holds and (i) of Theorem 4.3.1
2 , R], g(t, 0) 0, f (t, 0) 0,
is veried. Suppose that g C[R+
F (t, 0) 0 and for (t, x) R+ S(),
b(x) V (t, x) a(x), b K.
(4.3.17)
Furthermore, suppose that the trivial solution of (4.3.1) is uniformly

stable and u = 0 of (4.3.5) is uniformly asymptotically stable. Then,
the trivial solution of (4.3.2) is uniformly asymptotically stable.
214
Proof Let 0 < < , t0 R+ be given. The uniform stability of

u = 0 of (4.3.5) implies that given b() > 0, t0 R+ , there exists a
1 = 1 () > 0 such that if u0 1 , then
t t0 .
u(t, t0 , u0 ) < b(),
(4.3.18)
Let 2 = a1 (1 ). Since x = 0 of (4.3.1) is uniformly stable, given

2 > 0, t0 R+ , there exists a = () > 0 such that
y(t, t0 , x0 ) < 2 ,
t t0 ,
if
x0 < .
(4.3.19)
We claim that x0 < also implies that x(t, t0 , x0 ) < , t t0 ,

where x(t, t0 , x0 ) is any solution of (4.3.2). If this is not true, there
would exist a solution x(t, t0 , x0 ) of (4.3.2) with x0 < and a
t1 > t0 such that x(t1 , t0 , x0 ) , t0 t t1 . Then, by Theorem
4.3.1, we have
V (t, x(t, t0 , x0 )) r(t, t0 , V (t0 , y(t, t0 , x0 ))),
t0 t t1 .
Consequently, by (4.3.17), (4.3.18) and (4.3.19), we get

b() V (t1 , x(t1 , t0 , x0 )) r(t1 , t0 , a(y(t1 , t0 , x0 )))
r(t1 , t0 , a(2 )) r(t1 , t0 , 1 ) < b().
This contradiction proves that x = 0 of (4.3.2) is uniformly stable.
To show uniform asymptotic stability, we set = and () = 0 .
Then, from the foregoing argument, we have
b(x(t, t0 , x0 )) V (t, x(t, t0 , x0 )) r(t, t0 , V (t0 , y(t, t0 , x0 )))
for all t t0 , if x0 < 0 . From this it follows that
b(x(t, t0 , x0 )) r(t, t0 , 1 ()),
t t0
which implies the uniform asymptotic stability of x = 0 because of

the assumption that u = 0 of (4.3.5) is uniformly asymptotically
stable. Hence the proof is complete.
Setting F (t, x) = f (t, x) + R(t, x) in Theorem 4.3.2, we see that
although the unperturbed system (4.3.1) is only uniformly stable,
215
the perturbed system (4.3.2) is uniformly asymptotically stable, an

improvement caused by the perturbing term.
Let us present a simple but illustrative example. Consider
y = et y 2 ,
y(t0 ) = x0 ,
(4.3.20)
x0
. The
1 + x0 (et et0 )
fundamental matrix solution of the corresponding variational equation is
1
.
(t, t0 , x0 ) =
1 + x0 (et et0 )2
whose solutions are given by y(t, t0 , x0 ) =
Consequently, choosing V (t, x) = x2 , we see that

D V = 2y(t, s, x)(t, s, x)R(s, x)
2
where R(t, x) is the perturbation. Let R(t, x) = x

2 so that the
perturbed equation is
x = et x2
x2
,
2
x(t0 ) = x0 ,
(4.3.21)
Accordingly, it is easily seen that g(t, y) = y 3/2 and hence the

solutions of
u = u3/2 , u(t0 ) = u0 0
are u(t, t0 , u0 ) =
[2 +
the relation
|x(t, t0 , x0 )|2
4u0
1/2
u0 (t
t0 )2 ]
. Thus, by Theorem 4.3.1, we get
|x0 |2
[1 + x0 (et et0 +
t t0 2 ,
)]
2
t t0
which shows that all solutions x(t, t0 , x0 ) 0 as t , although

only some solutions y(t, t0 , x0 ) are bounded. For example, setting
t0 = 0 and x0 = 1 shows that et is the corresponding solution
of (4.3.20) where as for the same initial conditions the solution of
2
(4.3.21) is
.
2 + t + 2et
216
4.4
Stability of Delay Dierential Equations
The extension of Lyapunovs method to dierential equation with

nite or innite delay can be carried out using both Lyapunov functions and functionals. When Lyapunov functions are used, it becomes
necessary to estimate their derivatives relative to a suitable minimal
class of functions, so that one can reduce the study of a given problem with delay to the study of a corresponding ordinary dierential
equation. This approach has recently been recognized as an important tool and a natural framework in the discussion of qualitative
behavior of solutions of delay dierential equations.
In this section, we shall study in a unied way the stability theory
of dierential equations with innite delay.
Let B be a linear space of Rn -valued functions dened on (, 0]
with a semiform B and B B be the space of functions (s) on
(, 0] such that (s) is continuous on [, 0] and B, 0,
where denotes the function on (, 0] dened by t (s) = (t+s).
The space B is said to be admissible if, for any 0 and any B ,
we have
(i) t B for all t [, 0];
(ii) t is continuous in t with respect to B for all t [, 0];
and
(iii) (0) B K( ) sup (s) + M ( ) B , where
s0
> 0 is a constant and K(s), M (s) are continuous functions.

Clearly B and Rn are also admissible spaces with B =
sup B and |nR = (0) respectively, being the Eu-
clidian norm. The space B is said to have a fading memory, if, in

condition (iii), K(s) K is a constant and M (s) 0 as s .
In the set of the admissible spaces, we shall introduce the order
relation X < Y to mean X Y and Y N X for a constant
N . Therefore, condition (iii) implies that B < B < Rn . Typical
and more important admissible spaces are
Ch = { C[[h, 0], Rn ] : es (s) a limit as s ,
Stability of delay differential equations
217
when h = , }
0
Mh
= { : measurable on (h, 0] and
es (s)ds < },
h
with the seminorms given by

C =
h
sup es (s),
hs0
0
M = (0) +
es (s) ds,
, h being constant such that 0 < , 0 .

Let us consider the functional dierential equation
x (t) = f (t, xt ),
xt0 = 0 Bt0 ,
t0 > 0,
(4.4.1)
where f Rn , f (t, 0) 0 and f (t, ) is completely continuous on

R+ B for an admissible space B. Let x(t) = x(t0 , 0 )(t) denote
the solution of (4.4.1). We shall assume existence and fundamental properties of solution of (4.4.1). Let us list for convenience the
following hypotheses:
3 , nonde(H1 ) There exists a continuous function L(t, s, r) on R+
creasing in r with L(t, s, 0) 0 such that any solution x(t) of (4.4.1)
satises
x(t0 , 0 )(t) L(t, t0 , 0 Bt0 ), t t0 .
2 , nondecreas(H2 ) There exists a function p(t, r) continuous on R+
ing in r such that p(t, r) t and p(t, r) as t .
(H1 ) The function L in (H1 ) is of the form L(t, s, r) = L(ts, 0, r).

(H2 ) In H2 , q(r) = t p(t, r) is positive and independent of t.
It is easy to see that when (H2 ) holds, we have
(t, r) = sup{s : p(s, r) t}
218
2 , nonincreasing in r and such

as a function that is continuous on R+
that (t, r) t. Moreover, p(t, r) whenever t (, r) and
r 0.
Before presenting the comparison result, let us assume that there
exist a Lyapunov function V and two comparison functions g0 , g satisfying the following hypotheses:
(H3 ) V C[R+ S(), R+ ], V (t, x) is locally Lipchitzian in x

and for (t, x) R+ S(), V (t, x) b(x), where b K.
2 , R], g (t, u) g(t, u), r(t, t, , u ) is the right
(H4 ) g0 , g C[R+
0
0
0
maximal solution of
u = g(t, u),
u(t0 ) = u0 0,
(4.4.2)
existing for t t0 and (t, T, v0 ) is the left maximal solution of

u = g0 (t, u),
u(T0 ) = v0 0,
(4.4.3)
existing for t0 t T .
(H5 ) The derivative of V with respect to (4.4.1) dened by
D V (t, (0), ) = lim inf
h0
1
[V (t + h, (0) + hf (t, )) V (t, (0))]
h
satises the dierential inequality

D V (t, (0), ) g(t, V (t, (0)))
whenever p(t, V (t, (0)) t0 and , where, for t t0 ,
= { Btt0 : V (s, (s)) (s, t, V (t, (0)))},
s [p(t, V (t, (0))), t).
We are now in a position to prove a general comparison principle
in terms of Lyapunov function which provides a exible mechanism
to choose dierent minimal classes of functions that are required in
studying stability results for the equation (4.4.1).
Theorem 4.4.1 Assume that hypotheses (H1 )(H5 ) hold. Let
x(t0 , 0 )(t) be any solution of (4.4.1) through (t0 , 0 ) existing on
[t0 , ) such that

(4.4.4)
b sup L(s, t0 , 0 B(t0 ) ) u0 .
t0 st1
219
Then,
V (t, x(t0 , 0 (t))) r(t, t0 , u0 ),
where

r(t, t0 , u0 ) =
u0 ,
r(t, t1 , u0 ),
t t0 ,
t [t0 , t1 ];
t t1 and t1 = (t0 , 0).
Proof We set m(t) = V (t, x(t0 , 0 )(t)), t t0 , where x(t0 , 0 (t))

is any solution of (4.4.1). Since (t0 , r) t0 , r 0, we dene
t1 = (t0 , 0) t0 . In view of (4.4.4), (H1 ) and (H3 ), we have

m(s) b sup L(s, t0 , 0 B(t0 ) ) u0 .

t0 st1
Because of the fact that r(t, t0 , u0 ) = lim u(t, ) and the denition of
0
r(t, t0 , u0 ), it is enough to prove that

m(t) < u(t, ),
t t1
(4.4.5)
where u(t, ) is the solution of

u = g(t, u) + ,
u(t1 ) = u0 + ,
and > 0 is suciently small.

Note that m(t1 ) u0 < u0 + . If (4.4.5) is not true, then there
exists a t > t1 such that
m(t ) = u(t , )
and
m(t) < u(t, ),
t [t1 , t ).
This implies that

D (m(t )) u (t , ) = g(t , m(t )) + .
(4.4.6)
Now, consider the left maximal solution (s, t , m(t )), t0

s t , of the problem
u = g0 (t, u),
u(t ) = m(t ).
By Lemma 1.8.1, we obtain

r(s, t1 , u0 ) (s, t , m(t )),
s [t1 , t ].
220
Since
r(t , t1 , u0 ) = lim u(t , ) = m(t ) = (t , t , m(t ))
0
and
m(t) u(t, ),
t [t1 , t ],
it follows that
m(s) r(s, t1 , u0 ) (s, t , m(t )),
t [t1 , t ].
Since t > t1 = (t0 , 0), we have t p(t , m(t )) p(t , 0) t0 .

Consequently, by (H5 ), using a standard argument, we get
D m(t ) g(t , m(t ))
which contradicts (4.4.6). Thus,
m(t) r(t, t1 , u0 ), t t1 ,
and hence the conclusion of the theorem follows.
An observation of the proof of the Theorem 4.4.1 allows us to focus
on the idea and to formulate the following comparison theorem.
Theorem 4.4.2 Suppose that (H2 ) and (H4 ) hold. Let m
C[[t0 , ), R+ ] satisfy the dierential inequality
D m(t) g(t, m(t)),
t I0 ,
where I0 = {t > t0 : p(t, m(t)) t0 , m(s) (s, t, m(t)),

s [p(t, m(t)), t]} and (t, T, v0 ) is the left maximal solution of
(4.4.3) on t0 , T . Then,
m(t) r(t, t0 , u0 ),
whenever
t t0 ,
sup m(s) u0 , t1 being equal to (t0 , 0).
t0 st1
Note that when p(t, r) = t , > 0, so that (t, r) = t + , the

set I0 reduces to
I0 = {t > t0 + : m(s) (s, t, m(t)), t s t]}.
221
As a result, the corresponding special case of Theorem 4.4.2 can be

used in studying stability results for an equation with nite delay.
Let X, Y be two admissible spaces such that X < Y . The concepts
of stability in (X, Y ) for the zero solution of (4.4.1) may be dened
as follows. Let us consider the denition of the equistability.
Denition 4.4.1 The zero solution of (4.4.1) is said to be equistable in (X, Y ) if for any > 0 and any t0 t0 , there exists a
= (t0 , ) such that
0 X <
implies
xt Y <
for all
t t0 .
The other concepts of uniform stability, equi- and uniform asymptotic stability can be dened in a similar fashion. We shall now give
sucient conditions for the stability of the null solution of (4.4.1)
in (Bt0 , Rn ), since many practical phenomena associated with functional dierential equations with innite delay suggest the advantage
of stability in Rn rather than a function space. Thus, Denition 3.4.1
with X = Bt0 and x(t) in place of xt Y will suce for our discussion.
Theorem 4.4.3 Let the assumptions of Theorem 4.4.1 hold. Suppose that
a(x) V (t, x), (t, x) R+ S(),
where a K. Then, the equi-stability properties of the null solution
of (4.4.2) imply the corresponding equi-stability properties of the null
solution of (4.4.1). If, instead of (H1 ), (H2 ) in Theorem 4.4.1, hypotheses (H1 ), (H2 ) are satised, then the uniform stability properties
of the zero solution of (4.4.1) follow from the uniform stability properties of the zero solution of (4.4.2).
Proof. We shall indicate the proof for equi- and uniform stability
only. The other assertions can be proved similarly.
Let 0 < < and t0 0 be given. Then a() > 0 is dened.
Assume that the zero solution of (4.4.2) is equistable. Then, given
a() > 0 and t0 R+ , there exists a 1 = 1 (t0 , ) > 0 such that
0 u0 < 1
implies
u(t, t0 , u0 ) > a(),
where u(t, t0 , u0 ) is any solution of (4.4.2).
t t0 ,
222
0 , r) =
Let L(t
sup L(s, t0 , r), t1 = (t0 , 0), and note that
t0 st1
b(t0 , r) b(L(t
0 , r)) is continuous, b(t0 , 0) = 0 and b(t0 , r) is increasing in r. Consequently, there exists a = (t0 , ) > 0 such that
b(t0 , r) < 1 if r < .
Now choose 0 Bt0 < . Then, we claim that
x(t0 , 0 )(t) < ,
t t0 .
If this is not true, there exists a solution x(t0 , 0 ) and a t2 t1 t0

such that
x(t0 , 0 )(t2 ) = and
x(t0 , 0 )(t) < ,
t0 t t2 .
Dene m(t) = V (t, x(t0 , 0 )(t)), t t0 . Since (H1 )(H5 ) are satised, we have, by Theorem 4.4.1,
m(t) r(t, t0 , u0 ).
Hence, it follows that
a() = a(x(t0 , 0 )(t2 )) V (t2 , x(t0 , 0 )(t2 )) r(t2 , t0 , u0 ) < a().
This contradiction proves that the null solution of (4.4.1) is equistable.
If we assume uniform stability of the null solution of (4.4.2), then
1 (in the above paragraph) is independent of t0 . Note that, by
(H2 ), we get (t, r) = t + q(r), where q(r) is independent of t. Now
choose > 0 such that b(r) < 1 if r < , where b(r) = b(L(r))
and L(r) = sup L(, r). It is easy to see that = () depends

0q(1 )
only on . Hence, the uniform stability of the null solution of (4.4.1)

follows by an argument similar to one given above. We omit the
details in order to avoid repetition.
Let us now consider certain important special cases of Theorem
4.4.3.
Case (a). Suppose that g0 (t, u) = 0, so that g(t, u) 0. Then,
(s, t, v0 ) v0 . Consequently, the set in (H5 ) reduces to
= { Btt0 : V (s, (s)) V (t, (0)),
s [p(t, V (t, (0))), t]}
223
where t t0 .
Case (b). Suppose that g0 (t, u) = [A (t)/A(t)]u, where A(t) > 0
is continuous and dierentiable on [t0 , ) and A(t) as t .
2 ,R ]
Let g(t, u) = g0 (t, u) + [1/A(t)]g1 (t, A(t)u), where g1 C[R+
+
and r(t, t0 , u0 ) be the maximal solution of (4.4.2). Evidently, the
solution (s, t, v0 ) = v0 A(t)/A(s), s t. Hence
= { Btt0 : V (s, (s))A(s) V (t, (0))A(t),
s [p(t, V (t, (0))), t], t t0 }.
Case (c). Suppose that g0 (t, u) = c(u), where c K and
g(t, u) g0 (t, u). Computing (s, t, v0 ), we see that
(s, T, v0 ) = J 1 [J(v0 ) (s T )],
where J(u) J(u0 ) =
0 s T,
u ds
and J 1 is the inverse function of J.
u0 c(s)
Since (s, T, v0 ) is increasing in s to the left of T , choosing a xed

s0 < T and dening L(u) = (s0 , T, u), it is clear that L(u) > u
for u > 0, L(u) is continuous and increasing in u. Hence the set
reduces to
= { Btt0 : V (s, (s)) L(V (t, (0)),
s [p(t, V (t, (0))), t],
t t0 }.
The foregoing special cases clearly demonstrated that hypothesis

(H5 ) is instrumental in unifying the selection of minimal classes that
are needed for various stability results. Thus, the method discussed
in this section oers a unied approach to stability theory for equations with innite delay, analogous to stability theory in ordinary
dierential equations. In fact, Theorem 3.4.3 contains, as special
cases, several known results in stability theory for equations with
nite and innite delay.
Next, we shall consider the use of Lyapunov functionals in this
unifying approach. As appropriate modications of (H1 ), (H3 ) and
(H5 ), let us list the following hypotheses:
224
3 , nonde(H1 ) There exist continuous functions L(t, s, r) on R+

2
creasing in r with L(t, s, 0) = 0 and 0 (t, s) on R+ with 0 > 0, such
that for any solution x(t) of (4.4.1), we have the estimate
xt B L(t, s, xs B )

provided xs B < 0 (t, s), t s.
(H3 ) There exists a collection of continuous functionals {V (t, ;
t0 ), t0 0} dened on {(t, ) : Btt0 , t t0 }, satisfying
a(B ) V (t, ; t0 ) b(t, t0 , Btt0 ),
a K,
3 , b(t, s, r) is nondecreasing in r and

where b is continuous on R+
b(t, s, 0) = 0.
(H5 ) The derivative of V along the solutions dened by

D V (t, ; t0 ) = sup{lim inf
h0
1
[V (t + h, xt+h (t, ); t0 ) V (t, ; t0 )]}
h
for a solution xt of (4.4.1) where supremum is taken over all such

solutions, satises
D V (t, ; t0 ) g(t, V (t, ; t0 ))
whenever p(t, V (t, ; t0 )) t0 and , where
= { Btt0 : V (s, st ; t0 ) (s, t, V (t, ; t0 ),
s [p(t, V (t, ; t0 )), t]}.
With these changes in hypotheses, Theorem 4.4.1 takes the following form.
Theorem 4.4.4 Assume that (H1 ), (H2 ), (H3 ), (H4 ) and (H5 )
are satised. Then, for any solution x(t) of (4.4.1) we have
V (t, xt ; t0 ) r(t, t0 , u0 ),
whenever
sup
t t0 ,
(b(s, t0 , L(s, t0 , 0 Bt0 ))) u0 .
t0 s(t0 ,0)
225
Proof We set m(t) = V (t, xt ; t0 ) for t t0 , so that we have

m(s)
sup
(b(s, t0 , L(s, t0 , 0 Bt0 ))) u0
t0 s(t0 ,0)
and
0 Bt0
inf
t0 s(t0 ,0)
(0 (s, t0 )),
because of (H1 ) and (H3 ). Furthermore, let us note that by (H5 ), we

get
D m(t ) g(t , m(t )).
With this modications, the proof is a straightforward repetition of
the proof of Theorem 4.4.1.
Having this comparison theorem at our disposal, it is now easy
to give sucient conditions, in a unied way, for various stability
properties analogous to Theorem 4.4.3. We shall only indicate the
necessary changes in the proof of Theorem 4.4.3 in the current framework.
(b(s, t0 , L(s, t0 , ))) 1
We choose a > 0 such that
sup
and
t0 s(t0 ,0)
inf
(0 (s, t0 )).
Then, it is clear that whenever
t0 s(t0 ,0)
0 Bt0 < , it follows that m(t) 1 , t [t0 , t1 ], t1 = (t0 , 0).

The rest of the proof is similar to the proof of Theorem 4.4.3.
We shall now oer an example to illustrate the results discussed
in this section. Consider the equation
0
x (t) = ax(t) + bx(t h) +
g(t, s, x(t + s)) ds,
(4.4.7)
where a > 0, h > 0, b are the constants. Assume that g(t, s, x) is

3 and |g(t, s, x)| (s)|x|, where
continuous on R+
0
0
(s)e ds <
s
and
(s) ds < aeh |b|,
(4.4.8)
226
for some > 0. Let p(t, r) = t + q(r), where q(r) is continuous on

R+ , nondecreasing in r and q(r) h for r 0. In view of (4.4.8),
we can choose q(r) so that
0
a |b|e
(s) ds = > 0
(4.4.9)
and
q(m)
(s)e s ds m1/2 .
(4.4.10)
Let m(t) = V (t, (0)) = |(0)|2 and g0 (t, u) = 2u. Then, we have
0
m (t) 2a[(0)] + 2|b| |(0)| |(h)| + 2|(0)|
(s)|(s)| ds,
and the solution (t, s, v0 ) of

u = 2u,
u(t0 ) = v0
is given by (t, s, v0 ) = v0 exp(2(s t)), s < t. Now, estimating

m (t) along the set , we get
0
m (t) 2m(t)[a + |b|e ] + 2|(0)|

h
(s)|(s)| ds.
Furthermore, using (4.4.10), we obtain

q(m(t))
q(m(t))
(s)e s ds (0),
(s)|(s)| ds 2c
(4.4.11)
Integro-differential equations with finite memory
227
provided kkc 1, and using the fact that , we have

2
Z0
(s)|(s)| ds 2|(0)|e
q(m(t))
Z0
(s) ds
q(m(t))
2|(0)|eh
Z0
(4.4.12)
(s) ds.
Now, in virtue of (4.4.9), (4.4.11) and (4.4.12), we get
m (t) 2[a |b|e
Z0
(s)ds]m + m = m g(t, m)
and it is easy to see that all the assumptions of Theorem 4.4.3 are
satisfied. Since the zero solution of
u = u,
u(t0 ) = u0
is uniformly asymptotically stable, it follows that the zero solution

of (4.4.7) is uniformly asymptotically stable in (c , Rn).
4.5
Stability of Integro-Differential Equations with

Finite Memory
When we apply the results developed in Section 4.4. to example with

finite delay given by
x (t) = b[x(t) + x (t)] a
Zt
[x(s) + x3 (s)] ds,
a, b, > 0 (4.5.1)
with V (t, x) =
|x|2 ,
g(t, u) = g0 (t, u) 0 and
= { C[[, 0], R] : |(s)|2 |(0)|2 },

it is easy to see that the trivial solution of (4.5.1) is stable if a b.
However, for the equation
x (t) = a
Zt
[x(s) + x3 (s)] ds,
a, > 0
(4.5.2)
228
this method gives no information at all. In this section, we develop

a comparison result which oers a better mechanism for the study
of delay equations. The idea is to use upper and lower comparison
estimates simultaneously together with auxiliary functions that are
constructed so as to contain the growth of the function to be compared in a best possible way. When we apply this technique, we
get a large domain of stability for (4.5.1) and also some stability
information for (4.5.2).
Instead of considering delay equation of the form (4.4.1), we shall
consider integro-dierential equation with a nite memory. This will
also serve as a prelude to the next two sections, where we shall deal
with integro-dierential equations of Volterra type.
Consider the integro-dierential equation with nite memory
x (t) = f (t, x, T x),
where f C[R+ R2 , R] [Tx ](t) =
xt0 = 0 ,
t
t0 0
(4.5.3)
x(s) ds, > 0 , xt = xt (s) =
x(t + s), s 0 and 0 C[[, 0], R].

Let us list the following assumptions for convenience:
(A1 ) For i = 1, 2, i , ri C[[t0 , ) [t0 , ), R] such that
(i) r2 (s, t0 ) i (s, t) r1 (s, t0 ), ti s t, where ti = ti (t) is
chosen such that t0 ti t;
(ii) ri (t, t0 ) = i (t, t), t t0 ;
(iii) r1 (s, t0 ) = 2 (s, t), t s t2 and r2 (s, t0 ) = 1 (s, t),
t s t1 .
(A2 ).
(i) D + r1 (t, t0 ) > f (t, r1 (t, t0 ), [T u](t)) for all u C[R+ , R] such
that 1 (s, t) u(s) r1 (s, t0 ), s [t0 , t], D + r2 (t, t0 ) <
f (t, r2 (t, t0 ), [T u](t)) for all u C[R+ , R] such that r2 (s, t0 )
u(s) 2 (s, t), s [t0 , t];
(ii) D + 1 (, t) > f (, 1 (, t), [T u]()), t1 t and D + 2 (, t) <
f (, 2 (, t), [T u]()), t2 t for all u C[R+ , R] such that
2 (s, t0 ) u(s) r1 (s, t0 ), t0 s .
229
Integro-differential equations with finite memory
We shall now prove the comparison result.

Theorem 4.5.1 Assume that (A1 ) and (A2 ) hold. Suppose that
x(t) = x(t, t0 , 0 ) is any solution of (4.5.3) such that
r2 (t0 + s) < 0 (s) < r1 (t0 + s, t0 )
for
s 0.
Then, r2 (t, t0 ) < x(t, t0 , 0 ) < r1 (t, t0 ), t t0 .

Proof Suppose that the assertion of Theorem 4.5.1 is false. Then,
there exists a t > t0 such that
either
r1 (t , t0 ) = x(t ) or
and
r2 (s, t0 ) x(s) r1 (s, t0 ),
r2 (t , t0 ) = x(t ),
t0 s t .
(t , t
x(t )
(t , t ).
= 1
Let us consider the case r1
0) =
case can be proved similarly. We need to prove that
(4.5.4)
The other
1 (s, t ) x(s) for t1 s t ,

so that we have 1 (s, t ) x(s) for t s t , because of
(A1 ) (iii). For this purpose, let
v(s) = x(s) 1 (s, t ),
t1 s t
and note that v(t ) = 0. Hence

D + v(t ) < f (t , x(t ), [T x](t )) f (t , 1 (t , t ), [T u](t )).
Choosing u(s) = x(s), we obtain that D + v(t ) < 0. It then follows that v(t) is increasing for t < t in a suciently small interval
t t t , > 0, which implies that
x(t ) > 1 (t , t ).
We wish to show that
1 (s, t ) < x(s),
for
t1 s t .
If this is not true, there exists an s (t1 , t ) such that

x(s ) = 1 (s , t )
and x(s) > 1 (s, t ),
s (s , t ).
230
We therefore have x (s ) D + 1 (s , t ). However, using (A2 ), we

get
D + 1 (s , t ) > f (s , x(s ), [T u](s )).
Choosing u(s) = x(s), we then obtain
D + 1 (s , t ) > f (s , x(s ), [T x](s )) = x (s ),
which is a contradiction. Thus, 1 (s, t ) x(s) for t s t is
valid as claimed and therefore by (A2 ), it follows that
D + r1 (t , t0 ) > f (t , x(t ), [T u](t )) = f (t , x(t ), [T x](t )) = x (t ),
which implies that
D r1 (t , t0 ) > x (t ).
On the other hand, we obtain, because of (4.5.4), the relation
D r1 (t , t0 ) x (t ),
which is a contradiction. The proof of the theorem is therefore complete.
Let us now apply Theorem 4.5.1 to the example (4.5.1). We choose
functions i as follows:
r1 (t, t0 ) = r2 (t, t0 ) = ,
t < ,
2 (s, t) = 1 (s, t), t s t and t > 0,
,
t s t1 ,
1 (s, t) =
A
(t s) , t1 s t,
2
where t1 = t2 = max 0, t 4/A and A = 2(b + a )(1 + 2 ). It is easily
checked that (ii) of (A2 ) is veried for these 1 , t1 and A. From (i)
of (A2 ) we get
3
t
f b( + ) + a( + )[t1 (t )] a
t1
(1 + 31 ) ds.
231
Integro-differential equations of Volterra type
Since the value of the integral is zero, it suces to suppose that

4
3
< r1 (t, t0 ) = 0.
( + ) b + a
A
This in turn shows that it is enough to assume that

4
0
b + a
A
which implies that
a
1 + [1 + b2 2 (1 + 2 )2 ]1/2
.
2 (1 + 2 )
If b = 0, the above condition reduces to

a 2
2
.
1 + 2
If we restrict to 0 < 1, we see that it is enough to assume that

a 2 1 for the stability of the zero solution of (4.5.2).
4.6
Stability of Integro-Dierential Equations of

Volterra Type
This section deals with the extension of Lyapunovs method to

integro-dierential equations of Volterra type. The results are parallel to the results of Section 4.4 and hence we shall merely indicate
needed modications.
Let us consider the integro-dierential system
x = f (t, x, T x),
x(t0 ) = x0 ,
where f [R+ S() Rn , Rn ], T x =
t
t0
t0 0,
(4.6.1)
K(t, s, x(s)) ds and the
2 S(), Rn ].
kernel k C[R+
Assume that f (t, 0, 0) 0 and
K(t, s, 0) 0 so that (4.6.1) admits the trivial solution.
Before we proceed to state the theorem, let us list the following
hypotheses.
232
(H0 ) V C[R+ S(), R+ ], V (t, x) is locally Lipschitzian in x

and
b(x) V (t, x) a(x), (t, x) R+ S(),
where a, b K.
(H1 ) g0 , g C[R+ R+ , R], g0 (t, u) g(t, u), r(t, t0 , u0 ) is the
right maximal solution of
u = g(t, u),
u(t0 ) = u0 0,
(4.6.2)
existing on [t0 , ) and (t, t0 , v0 ) is the left maximal solution of

v = g0 (t, v),
v(t0 ) = v0 0,
t0 > t0 ,
(4.6.3)
existing on t0 t t0 .
1
[V (t+h, x+hf (t, x, T x))V (t, x)]
h
g(t, V (t, x)), (t, x) , where = {x C[R+ , Rn ] : V (s, x(s))
(s, t, V (t, x(t))), t0 s t}.
(H2 ) D V (t, x, T x) lim inf
h0
We are now in a position to prove the following general comparison

theorem which permits us to consider a unied theory of stability for
the integro-dierential system (4.6.1).
Theorem 4.6.1 Assume that (H0 ), (H1 ) and (H2 ) hold. Let
x(t, t0 , x0 ) be any solution of (4.6.1) such that V (t0 , x0 ) u0 . Then,
V (t, x(t, t0 , x0 )) r(t, t0 u0 ),
t t0 .
Proof We set m(t) = V (t, x(t, t0 , x0 )), t t0 , so that m(t0 ) u0 .

Because r(t, t0 , u0 ) = lim u(t, ) where u(t, ) is any solution of
0+
u = g(t, u) + ,
u(t0 ) = u0 + ,
for > 0, suciently small, it is enough to prove that m(t) < u(t, ),
for t t0 . If this is not true, there exists a t1 > t0 such that
m(t1 ) = and
m(t) < u(t, ),
t0 t < t1 .
233
Integro-differential equations of Volterra type
This implies that

D m(t1 ) u (t1 , ) = g(t1 , m(t1 )) + .
(4.6.4)
Consider now the left maximal solution (s, t1 , m(t1 )) of (4.6.3)

with v(t1 ) = m(t1 ) on the interval t0 s t1 . Since
r(t1 , t0 , u0 ) = lim u(t1 , ) = m(t1 ) = (t1 , t1 , m(t1 ))
0+
and m(s) u(s, ), for t0 s t1 , it follows that

m(s) r(s, t0 , u0 ) (s, t1 , m(t1 )),
t0 s t1 .
This inequality implies that (H2 ) holds for x(s, t0 , x0 ) on t0 s t1

and as a result, standard computation yields
D m(t1 ) g(t1 , m(t1 )),
which contradicts the relation (4.6.4). Thus, m(t) r(t, t0 , u0 ),
t t0 and the proof is complete.
Having this comparison theorem at our disposal, we merely state
the following theorem that oers various stability criteria in a single
setup.
Theorem 4.6.2 Let (H0 ), (H1 ) and (H2 ) hold. Then the stability
properties of the trivial solution of (4.6.2) imply the corresponding
stability properties of the trivial solution of (4.6.1).
The following special cases of Theorem 4.6.2 are important in
themselves.
Case I. Suppose that g0 (t, u) 0 so that g(t, u) 0. Then,
(s, t0 , v0 ) = v0 . Consequently, the set in (H2 ) reduces to
= {x C[R+ , Rn ] : V (s, x(s)) V (t, x(t)),
t0 s t}.

A (t)
u where A(t) > 0
Case II. Suppose that g0 (t, u) =
A(t)
is continuously dierentiable on [t0 , ) and A(t) as t .

234

1
2 , R ].
Let g(t, u) = g0 (t, u) +
g1 (t, A(t)u) with g1 C[R+
+
A(t)
0
0
Evidently, (s, t , v0 ) = v0 A(t)/A(s) for t0 s t . Hence
= {x C[R+ , Rn ] : V (s, x(s))A(s) V (t, x(t))A(t), t0 s t}.
Case III. Suppose that g0 (t, u) = (u) where C[R+ , R+ ],
K and g(t, u) g0 (t, u). Computing (s, t0 , v0 ) we see that
(s, t0 , v0 ) = J 1 [J(v0 ) (s t0 )],
0 s t0 ,
u ds
and J 1 is the inverse function of J.
u0 (s)
Since (s, t0 , v0 ) is increasing in s to the left of t0 , choosing a xed
s0 < t0 and dening L(u) = (s, t0 , v0 ), it is clear that L(u) > u
for u > 0, L(u) is continuous and increasing in u. Hence the set
reduces to
where J(u) J(u0 ) =
= {x C[R+ , Rn ] : V (s, x(s)) L(V (t, x(t))), t0 s t}.
4.7
Integro-Dierential Equations (Continued)
The approach developed in Sections 4.4 and 4.6 for delay as well
as integro-dierential equations oers useful results in a unied way
only when the ordinary dierential equations part of the given system have nice stability properties, since the delay term or integral
term is essentially treated as perturbation. The comparison method
presented in Section 4.5 is one of the ways to overcome the diculty
posed by lack of nice stability information for the unperturbed system. In this section, we shall discuss another method that makes use
of Theorem 2.10.1 and the corresponding variational system. However, a result analogous to Theorem 2.10.1 is not yet available for
delay equations.
Consider the integro-dierential system
t
x (t) = f (t, x(t)) +
g(t, s, x(s)) ds,

t0
x(t0 ) = x0 ,
(4.7.1)
Integro-Differential Equations (Continued)
235
2 S(), Rn ]. Assume that

where f C[R+ S(), Rn ] and g C[R+
fx , gx exist and are continuous and f (t, 0) 0 g(t, s, 0). Setting
fx (t, 0) = A(t),
gx (t, s, 0) = K(t, s),
(4.7.2)
and using the mean value theorem, equation (4.7.1) can be written
as
Zt
Zt
x (t) = A(t)x(t) + K(t, s)x(s) ds + F (t, x(t)) + G(t, s, x(s)) ds,

t0
t0
(4.7.3)
with x(t0 ) = x0 and
F (t, x) =
Z1
[fx (t, x fx (t, 0))] dx ,
G(t, s, x) =
Z1
[gx (t, s, x gx (t, s, 0))] dx .
We shall now prove the following result.

(i) the trivial solution of the linear integro-differential equation
x (t) = A(t)x(t) +
Zt
K(t, s)x(s) ds,
x(t0 ) = x0
(4.7.4)
t0
is exponentially asymptotically stable;

(ii) for (t, x) R+ s(), kF (t, x)k w1 (t, kxk), kG(t, s, x)k
2 , R ], w C[R3 , R ] and
w2 (t, s, kxk), where w1 C[R+
+
2
+
+
w1 (t, u), w2 (t, s, u) are nondecreasing in u.
Then, the stability properties of the trivial solution of the differential system
"
#
Zt
u (t) = u(t) + M w1 (t, u(t)) + w(t, s, u(t)) ds , u(t0 ) = u0 ,
t0
(4.7.5)
236
where > 0, M > 0 and

w(t, s, u(s)) = w2 (t, s, u(s)) + L(t, s)w1 (s, u(s))
s
+ L(s, )w2 (s, , u()) d
(4.7.6)
t0
L(t, s) being any solution of (2.10.3), imply the corresponding stability properties of the system (4.7.1).
Proof By Theorem 2.10.1, system (4.7.3) is equivalent to
x (t) = B(t)x(t) + L(t, t0 )x0 + F (t, x(t))
t
t
+ L(t, s)F (s, x(s)) ds + G(t, s, x(s)) ds
t0
(4.7.7)
t0
t s
L(t, )G(s, , x()) d ds,
+
t0 t0
with x(t0 ) = x0 , where B(t) = A(t)L(t, t) and L(t, s) is any solution

of (2.10.3). Also, it is clear that (4.7.4) is equivalent to
x = B(t)x(t) + L(t, t0 )x0 ,
x(t0 ) = x0
(4.7.8)
and hence assumption (i) shows that the trivial solution of (4.7.8)
is exponentially asymptotically stable. Consequently, there exists a
Lyapunov function V such that
(a) V C[R+ S(), R+ ], V (t, x) is Lipscitzian in x for a constant
M > 0 and x V (t, x) Kx, (t, x) R+ S();
(b) D + V (t, x) V (t, x), (t, x) R+ S().
Let x(t) = x(t, t0 , x0 ) be any solution of (4.7.3) existing on
some interval t0 t t1 . Then, using (a) and (b) and setting
m(t) = V (t, x(t)) we get the integro-dierential inequality
D + m(t) m(t) + M w1 (t, m(t))
Stability of difference equations
237
t
w(t, s, m(s))ds,
+M
t0 t t1 ,
t0
where w(t, s, u) is given by (4.7.6). Setting v(t) = m(t)e(tt0 ) , we

obtain, for t0 t t1 ,

t
D + v(t) M e(tt0 ) w1 (t, v(t)e(tt0 ) ) + w(t, s, v(s)e(st0 ) ) ds
t0
which implies by Corollary 1.8.2 that

v(t) r(t, t0 , m(t0 )),
t0 t t1 ,
where r(t, t0 , u0 ) is the maximal solution of the dierential equation

u (t) = M e(tt0 ) w1 (t, u(t)e(tt0 ) ) +
t

w(t, s, u(t)e(st0 ) ) ds ,
t0
u(t0 ) = u0 . As a result, we have

V (t, x(t)) r(t, t0 , V (t0 , x0 ))e(tt0 ) ,
t0 t t1 ,
and it is easy to show that R(t, t0 , u0 ) = r(t, t0 , u0 )e(tt0 ) is the

maximal solution of the dierential system (4.7.5). Hence the desired
stability properties of the trivial solution of (4.7.1) follow from the
corresponding stability properties of the trivial solution of (4.7.5).
The proof is complete.
4.8
Stability of Dierence Equations
In this section, we shall extend Lyapunovs method to dierence

equations. Let us now consider the dierence equation
yn+1 = f (n, yn ),
(4.8.1)
with f (n, 0) = 0, f : Nn+ S() Rs , f (n, x) is continuous in

x. Let y(n, n0 , y0 ) be the solution of (4.8.1) having (n0 , y0 ) as initial
condition which is dened for n Nn+0 .
238
Let V : Nn+0 S() R+ and consider the variation of V along

the solutions of (4.8.1)
V (n, yn ) = V (n + 1, yn+1 ) V (n, yn ).
(4.8.2)
If there is a function : Nn+0 R+ R+ such that

V (n, yn ) (n, V (n, yn )),
then we shall consider the inequality
V (n + 1, yn+1 ) V (n, yn ) + (n, V (n, yn )) g(n, V (n, yn )) (4.8.3)
to which we shall associate the comparison equation
un+1 = g(n, un ) un + (n, un ).
(4.8.4)
The auxiliary functions V (n, x) are called Lyapunov functions. In

the following we shall always assume that such functions are continuous with respect to the second argument.
Theorem 4.8.1 Suppose there exist two functions V (n, x) and
g(n, u) satisfying the following conditions
(1) g : Nn+0 R+ R, g(n, 0) = 0, g(n, u) is nondecreasing in u;
(2) V : Nn+0 S() R+ , V (n, 0) = 0, and V (n, u) is positive
denite and continuous with respect to the second argument;
(3) f satises (4.8.3).
Then
(a) the stability of u = 0 for (4.8.4) implies the stability of yn = 0.
(b) the asymptotic stability of u = 0 implies the asymptotic stability
of yn = 0.
Proof By Theorem 1.9.1, we know that V (n, yn ) un , n Nn+0 ,
provided that V (n0 , y0 ) u0 . From the hypothesis of positive deniteness we obtain for K
(yn ) V (n, yn ) un .
Stability of difference equations
239
If the zero solution of the comparison equation is stable, we get that

un < () provided that u0 < (, n0 ) which implies
(yn ) V (n, yn ) ()
from which we get
yn < .
(4.8.5)
From V (n0 , y0 ) u0 (, n0 ) using the hypothesis of continuity

of V with respect to the second argument, it is possible to nd a
(, n0 ) such that y0 < (, n0 ) will imply V (n, y0 ) u0 . We need
to show now that (4.8.5) holds for all n n0 . Suppose it is not true.
Then there exists an n1 such that yn1 > and yn < for n n1 .
We then have V (n, yn1 ) () and
() V (n, yn1 ) un1 < ()
which is a contradiction.
In the case of asymptotic stability, from
(yn ) V (n, yn ) un
we get lim (yn ) = 0 and consequently lim yn = 0.
n0
n0
Corollary 4.8.1 If there exists a positive denite function

V (n, x) such that on Nn+0 S(), Vn 0, then the zero-solution of
(4.8.1) is stable.
Proof. In this case (n, u) = 0 and the comparison equation
reduces to
un+1 = un
which has stable zero solution.
Theorem 4.8.2 Assume that there exist two functions V (n, x)
and (n, u) satisfying conditions (1), (2), (3) of Theorem 4.8.1 and
moreover suppose that V is decrescent. Then
(a) uniform stability of u = 0 implies uniform stability of yn = 0.
(b) uniform asymptotic stability of u = 0 implies uniform asymptotic stability of yn = 0.
240
Proof The proof proceeds as in the previous case except we need

to show that (, n0 ) can be chosen independent of n0 . This can be
done by using the hypothesis that V (n, x) is decrescent, because in
this case there exists a K such that V (n, yn ) (yn ). In fact,
as before, we have
(yn ) V (n, yn ) un
provided that V (n0 , yn0 ) u0 (). If we take (y0 ) < (),
that is y0 < () 1 (()), then yn < for all n n0 .
The uniform asymptotic stability will follow similarly.
Corollary 4.8.2 If there exists a positive denite and decrescent
function V such that V (n, yn ) 0, then yn = 0 is uniformly stable.
Corollary 4.8.3 If there exists a function V such that
(yn ) V (n, yn ) (yn )
and
V (n, yn ) (yn )
where , , K, then yn = 0 is uniformly asymptotically stable.
Proof Clearly yn 1 (V (n, yn )) and by substitution we have
V (u, yn ) (1 (V (n, yn ))) (V (n, yn )).
The comparison equation is now given by
un+1 = un (un )
(4.8.6)
which shows that the origin is asymptotically stable for (4.8.6) and
because the equation is autonomous, it follows that the stability is
also uniform.
If in Theorem 4.8.2 the condition that V is decrescent is removed,
the asymptotic stability will result.
Theorem 4.8.3 Assume that there exists a function V such that
(1) V : Nn+0 S() R+ , V (n, 0) = 0, positive denite;
241
Impulse Differential Equations
(2) V (u, yn ) (yn ).

Then the origin is asymptotically stable.
Proof By Theorem 4.8.1 we know that the origin is stable. Suppose it is not asymptotically stable. Then there exists a solution
y(n, n0 , y0 ) and a set Jn0 Nn+0 such that y(n, n0 , y0 ) > > 0.
We then have
V (n + 1, y(n + 1, n0 , y0 )) V (n, yn ) (),
n Jn0 .
Since V is decreasing, V (n + 1, yn+1 ) V (n, yn ). Summing we have

V (n + 1, y(n + 1, n0 , y0 )) V (n0 , y0 ) k()
where k is the number of elements less than n in Jn0 . Taking the
limit we get
lim V (n, yn ) =
n
which contradicts the hypothesis that V is positive.
4.9
Impulse Dierential Equations
In this section, we shall consider dierential systems with impulsive

eect and extend Lyapunov stability to such systems.
Consider the impulsive system

x = f (t, x), t = k , x(t+
0 ) = x0 ,
+
x(k ) = x(k ) + Ik (x(k )),
t0 0,
where
(i) 0 < 1 < 2 < . . . < k < . . . and k as k ;
(ii) f C[R+ S(), Rn ], f (t, 0) 0; and
(iii) Ik C[S(), Rn ], Ik (0) = 0, k = 1, 2, 3, . . .
(4.9.1)
242
We shall assume that the solutions x(t) = x(t, t0 , x0 ) of (4.9.1)

exist on [t0 , ). We note that the solutions x(t) are continuously
dierentiable for t = k , left continuous at t = k and satisfy (4.9.1).
We shall also assume that the solutions x(t) meet each of the hyperplanes t = k at most once.
Let us rst prove a comparison result in terms of Lyapunov functions using Theorem 1.11.3.
(A0 ) V C[R+ S(), R+ ], V (t, x) is locally Lipschitzian in x,
D + V (t, x) g(t, V (t, x)), t = k , (t, x) R+ S(),
and
V (t, x + Ik (x)) k (V (t, x)) for t = k , x S(), where
2 , R] and K.
g C[R+
k
Let r(t) = r(t, t0 , u0 ) be the maximal solution of the scalar dierential equation with impulsive eect given by
u = g(t, u),
u(k+ )
t = k ,
u(t+
0 ) = u0 0,
t0 0,
= k (u(k ))
(4.9.2)
existing on [t0 , ). Then, if x(t) is any solution of (4.9.1) existing

on [t0 , ) such that V (t0 , x0 ) u0 , we have
V (t, x(t)) r(t),
t t0 .
Proof Setting m(t) = V (t, x(t)), where x(t) is any solution of

(4.9.1) existing on [t0 , ), we see that m(t) is continuous for t = k ,
left continuous at t = k and satises the inequalities
D + m(t) g(t, m(t)),
and
t = k ,
m(k+ ) k (m(k )),
m(t+
0 ) u0
k = 1, 2, 3, . . .
Consequently, by Theorem 1.11.3, we obtain the conclusion of Theorem 4.9.1.
243
Theorem 4.9.2 Assume (A0 ) of Theorem 4.9.1 holds. Suppose

further
b(x) V (t, x) a(x),
(t, x) R+ S(),
where a, b K. Then the stability properties of the trivial solution

of (4.9.2) imply the corresponding stability properties of the trivial
In view of Theorem 4.9.1, the proof of this theorem is straight
forward with standard arguments and hence, we omit the details.
Denition 4.9.1 The function V : R+ Rn R belongs to
class V0 if the following conditions hold:
(1) V (t, x) is continuous in Y =
k=1 Yk , Yk = {(t, x) R+
Rn : k1 < t < k } and local Lipschitz on x in all Yk ;
(2) for any k = 1, 2, . . . and any point (t0 , x0 ) Yk , Yk = { (t, x)
R+ Rn : t = k } there exist nite limits
V (k , x) =
V (k+ , x) =
lim
V (t, y),
lim
V (t, y)
(t,y)(k ,x)0
(t,y)(k ,x)+0
and the equality V (k , x) = V (k , x) is fullled.

Let us discuss some special cases of (4.9.1) to illustrate the results.
Suppose that
g(t, u) = p(t)g(u)
(4.9.3)
where p C[R+ , R+ ] and g K. Assume that there exists a c0 > 0
such that for each 0 < c c0 ,
k (c)
ds
g(s)
is satised and
k
p(s) ds k ,
k 0,
k = 1, 2, . . . ,
(4.9.4)
k1

i=1
i = . Then the trivial solution of (4.9.2) is
uniformly asymptotically stable. To prove this, let 0 < c0
244
and t0 (j1 , j ] for some j 1. Since k K, there exists a

0 < < such that k (s) < whenever 0 s < . We choose
0 u0 < . Let us rst show that u(t) < , t t0 , where
u(t) = u(t, t0 , u0 ) is any solution of (4.9.2). Clearly u(t) is nondecreasing in each interval (k1 , k ]. In particular, u(j ) u0 <
and hence u(j+ ) = j (u(j )) < .
We shall show that the sequence {u(j+ )} is nonincreasing for i j
from which it follows that u(t) < for t > j .
Let i j + 1 and for each k = j + 1, j + 2, . . . , i 1, suppose that
+
). Then, we have
u(k+ ) u(k1
+
) . . . u(j+ ) < .
u(i ) u(i1
In view of (4.9.2), the assumption (4.9.3) yields

+
u(
i )
u(
i (u(i ))
u(i
ds/g(s)
u(i+ 1)
ds/g(s) i , (4.9.5)
p(s) ds +
u(i1 )
u(i
+
). Thus, by induction, we
which in turn shows that u(i+ ) u(i1
+
+
get u(k ) u(k1 ) for each k j + 1 and consequently, u(t) <
for t t0 .

i = , we can show that lim u(i+ ) = 0. Assume the
If
i
i=1
contrary so that there exists a > 0 such that u(i+ ) for i j.

+
Then g() g(u(i+ )) g(u(i1
)). Consequently, we get, using
(4.9.5),
+
u(i1
)
i
u(i+ )
+
u(i1
) u(i+ )
,
ds/g(s)
g()
+
) i g(), which yields
and therefore u(i+ ) u(i1
+
u(j+k
)
u(j+ )
g()
j+k

i .
i=j+1
As k , we are led to a contradiction and this proves that u 0

of (4.9.2) is uniformly asymptotically stable.
245
Next, we shall consider the case

g(t, u) = p(t)g(u)
(4.9.6)
where p(t) > 0 is continuous on R+ and g K. We assume that

(4.9.4) holds with k = 0 and show that the trivial solution of (4.9.2)
is uniformly stable. To prove this, we choose as before, t0 (j , j+1 ]
for some j 1 and > 0 such that k (s) < for s [0, ). Let
0 u0 < . If we suppose that for some t (t0 , j+1 ] we have
u(t ) , then

ds/g(s) <
j ()
u(t
)

ds/g(s)
ds/g(s) <
u0
u0
j+1
p(s) ds
ds/g(s) =
p(s) ds,
j
t0
which implies that

j ()
j+1
p(s) ds +
j
ds/g(s) > 0
contradicting (4.9.4). Hence u(t) < for t [t0 , j+1 ].

Let i j + 2 and suppose that u(t) < for t (j+1 , j ]. Then,
for t (j , j+1 ] we obtain
u(t)
t
ds/g(s)
u(i+ )
Since
u(i+ )
= i (u(i )),
i+1
p(s) ds
i+
+
u(
i )
i+
ds/g(s) =
u(i )
p(s) ds.
i (u(
i ))
ds/g(s) and hence,
u(i )
we get from (4.9.4) with k = 0,

u(t)
i+1
ds/g(s)
u(i )
i (u(
i ))
ds/g(s) 0.
p(s) ds +
i
u(i )
246
This proves that u(t) u(i ) < for t (j , i+1 ] which by induction
shows that u(t) < for t t0 proving the statement.
Further we consider the impulsive system (4.9.1) under the condition 0 < 1 k+1 k 2 < , k = 1, 2, . . . . For system (4.9.1)
the following Theorem is valid.
Theorem 4.9.3 Let there exists the function V (t, x) V0 for
system (4.9.1) such that the following inequalities are satised:
(1) 0 V (t, x) c(x) for (t, x) R+ D, D Rn ;
(2) V (t+ , x(t+ )) V (t, x) 0 for t = k , k = 1, 2, . . . ;
&
(3) D + V (t, x)&(4.9.1) b(x) for t (k , k+1 ], k = 1, 2, . . . ;
(4) V (k+ , x(k+ )) a(x(k+ )), where a, b, c K.
Then the zero solution of system (4.9.1) is asymptotically stable.
Proof Choose the Lyapunov function
x) V0 .
+ candidate V ,(t,
Next, consider the sequence of numbers V (k+ , x(k+ )) k=0 . Obviously this is nonincreasing sequence. Let t0 = 0 then using conditions
(1), (4) of Theorem 4.9.3 we get
a(x(k+ )) V (0+ , x(0+ )) c(x0 ),
where x(0+ ) = x0 .
At rst we show that for any > 0 there exists () > 0 such
that x0 < () implies x(k+ ) < eL2 , where L > 0 is Lipschitz
constant for the function f (t, x) k = 1, 2, . . . . Suppose the contrary.
+
Then there exists N > 0
such that x(N
) eL2 . Next, for any
> 0 choose () = c1
a(eL2 )
2
. Then we get
+
)) c(x0 )
a(eL2 ) a(x(N

L2 )
a(eL2 )
1 a(e
=
.
<c c
2
2
This contradiction proves that for any > 0 there exists () > 0
such that x0 < () implies x(k+ ) < eL2 < , k = 1, 2, . . . .
247
Next, show that for any > 0 there exists () > 0 such that
x0 < () implies x(t) < for t (k , k+1 ].
dx
Consider the system
= f (t, x) for t = k . The systems soludt
t
tion x(t) is expressed as x(t) = x(k+ ) + f (s, x(s))ds, t (k , k+1 ].
k
Then we have the estimates

x(t)
x(k+ )
t
f (s, x(s)) ds
x(k+ )
t
+
Lx(s) ds.
k
Using Theorem 1.1.1 and inequality x(k+ ) < eL2 < we obtain
x(t) x(k+ )eL(tk ) eL2 x(k+ ) < eL2 eL2 =
i.e. x(t) < as
x0 < () = c1

a(eL2 )
.
2
So, combining the obtained results we conclude that system (4.9.1)

is stable.
Next, we have to prove that there exists 0 > 0 and x0 < 0
implies x(k+ ) 0 as k+ and x(t)
, 0 as t .
Clearly the sequence V (k+ , x(k+ )) k=0 is bounded from below
by zero i.e. limk V (k+ , x(k+ )) = 0. Then we have
+
+
+
+
, x(k+1
)) = V (k+1
, x(k+1
)) V (k+1 , x(k+1 ))
V (k+1
+ V (k+1 , x(k+1 )) V (k+1 , x(k+1 )) = V (k+ , x(k+ ))

k+1
k+1
D + V ds V (k+ , x(k+ ))
b(x(s)) ds.
+
k
,
+
Next, suppose that x(k+ )) k=0 0 as k . Then we
+
,
can select subsequence x(n+k )) n , where n, k = 1, 2, . . . and
k
limnk x(nk ) > 0.
248
Let nk s < nk+1 then using the conditions of Theorem 4.9.3

we obtain the estimates V (n+k , x(n+k )) V (s, x(s)) c(x(s)) and
c(x(s)) a(x(n+k ))). Since functions b, c K
b(x(s)) b(c1 (a(x(n+k )))) > 0.
For this subsequence we get
nk+1
(n+k+1 , x(n+k+1 ))
(n+k , x(n+k ))
(n+k , x(n+k ))
b(x(s)) ds
nk
(nk+1 nk ).
It is also clear that

p

V (n+k+1 , x(n+k+1 ))
k=1
p

V (n+k , x(n+k )) (nk+1 nk ) ,
k=1
(n+p+1 , x(n+p+1 ))
V (n+1 , x(n+1 )) 1 np .
So, V (n+p , x(n+p )) as np .

+
,
This contradiction proves that x(k+ )) k=0 0 as k .
Obviously if t then k , therefore we derive 0 x(t)
eL2 x(k+ )
+
, 0. Thus we have proved that x0 < 0 implies
x(k+ )) k=0 0 as k and x(t) 0 as t ,
t (k , k+1 ], k = 1, 2, . . . . Hence zero solution of system (4.9.1)
is asymptotically stable. This completes the proof of Theorem 4.9.3.
To show that impulsive eect can help in the stability behavior of
(4.9.1), we give the following example.
Example 4.9.1 Consider
x
x = , t 1, t = i
t
and x(i+ ) = x(i) + pi x(i),
where i 2 is an integer and pi are constants such that |1 + pi |

i
for every i. Taking V (t, x) = x2 , we see that
i+1
D + V (t, x) =
2
V (t, x),
t
t = i,
249
Reaction-diffusion equations
V (i, x + Ii (x)) (1 + pi )2 V (i, x),

so that p(t) = 1t , g(s) = 2s, i (s) = (1 + pi )2 s and
Zi+1
p(s)ds+
Zi (c)
c
ds
=
g(s)
Zi+1
ds
+
s
2
(1+p
Z i) c

ds
i+1
= log
|1 + pi | 0.
2s
i
Hence u 0 of (4.9.2) is uniformly stable, which implies that x = 0

is uniformly stable.
4.10
Reaction-Diffusion Equations
This section is devoted to the study of stability properties of weakly

coupled systems of reaction-diffusion equations.
Let RN be a bounded, open and connected region, equipped
with a smooth boundary = , say, of class C 2+ , 0 < < 1. For
Rn , let us denote by kvk0 = sup kv(x)k
a continuous function v :
kv(x) v(y)k
and [] = sup
, x 6= y, 0 < < 1. The space C k+
kx
yk
x,y
and the norm in that space are defined in the usual way. For example,
kvk = kvk0 + [v]
and
kvkk+ = kvk0 +
where Di =
form by
N
X
kDi vk0 +
N
X
Di (aij (x)Dj v)
i=1
N
X
i,j=1
kDi Dj vk ,
. The elliptic operator L is given in the self-adjoint

xi
Lv =
i,j=1
We assume that
where aij are functions defined in .
and
(A1 ) aij C 1+ ()
0 < < 1, > 0;
N
P
i,j=1
aij (x)i j kk2 for RN ,
250
and = 0 on , where is the outer normal.

(A2 ) C 2+ ()
For any interval J = (0, T ], let c (J) consist of all u for which
We
ut , Di u, Di Dj u, 1 i, j N exist and are continuous on J .
consider the reaction-diusion system
ut = Lu + f (u) in J ,
(4.10.1)
u
u(0, x) = (x) in ,
= 0 on J ,
where f C[Rn , Rn ] and L is the same operator for all components

of u in (4.10.1), that is, the equation is such that all components
obey the same diusion law. Let us state the following existence
result due to Amann [1] before the proceed further.
Theorem 4.10.1 Let (A1 ) and (A2 ) hold and f be locally Lipschitzian. Then, the problem (4.10.1) has a local solution u in c (J).
Furthermore, the derivatives Di Dj u are H
older continuous in x and
2+
u(t, ) maps J continuously into C

(). If an a
` priori estimate
u(t, x) K, where K is independent of T , can be established then
the solution u exists for 0 t < and it is bounded in C 2+ (),

sup u(t, )2+ = K1 < .
t0
In order to extend Lyapunov stability theory to the system

(4.10.1), we need the following comparison result.
(A3 ) V C 2 [S(, R+ )], Vu (u)Lu LV and Vu (u)f (u) g(V )
where g is locally Lipschitz continuous;
(A4 ) r(t) is the solution of
r = g(r),
r(0) = r0 0
(4.10.2)
existing on [0, ).
implies
Then V ((x)) r0 in
V (u(t, x)) r(t),
t0
(4.10.3)
251
Reaction-diffusion equations
where u(t, x) is the solution of (4.10.1).

Proof Setting m(t, x) = V (u(t, x)), we obtain
mt Lm + g(m)
m
=0
on J ,
in J
m(0, x) r0
in .
Also, r = g(r), r(0) = r0 0. Consequently, an application of

Theorem 1.12.2 with m = v, = r yields the estimate (4.10.3) and
the proof is complete.
Utilizing Theorem 4.10.2, it is easy to reduce the study of stability
properties of (4.10.1) to that of (4.10.2). We state such a result,
leaving its proof to the reader.
Theorem 4.10.3 Assume (A1 ), (A2 ) and (A3 ). Suppose further
that
b(u) V (u) a(u),
u S(), a, b K,
(4.10.4)
and f (0) = g(0) = 0. Then the stability properties of the trivial

solution of (4.10.2) imply the corresponding stability properties of
the trivial solution of (4.10.1).
As an example, let us consider a special case of (4.10.1) where
Lui = dui , d > 0, f is such that u f (u) 0, u S(). Then,
taking V (u) = u2 , we see that (A3 ) and (4.10.4) are satised
so that Vu (u)Lu LV and g(u) 0. Consequently we obtain
whenever V ((x)) r0 , x ,
V (u(t, x)) V ((x)), t 0, x ,
which yields the uniform stability of the zero solution of (4.10.1).
We note that if V (u) is convex, then Vu L(u) LV holds. Also,
if the system (4.10.1) does not possess the same diusion law for all
components, the foregoing method does not work. Since the stability
properties depend solely on the reaction term, one concludes the same
stability property of (4.10.2) even when = 0 or > 0 in (A1 ). To
avoid this defect, we need to compare the system (4.10.1) with the
scalar reaction-diusion equation
rt = Lr + g(r) in J ,
(4.10.5)
r = 0 on J , r(0, x) 0 on .
252
Applying Theorem 1.12.2 with m = v and = r where r(t, x) is the

solution of (4.10.5), we get, instead of (4.10.3), the estimate
V (u(t, x)) r(t, x),
t 0,
x ,
(4.10.6)
On the basis of (4.10.6), it is now

provided V ((x)) r(0, x), x .
easy to formulate Theorem 4.10.3 under the same assumptions. This
implies that we have the problem of knowing the stability properties
of the trivial solution of (4.10.5). In those cases where it is simple
to determine the behavior of solutions of (4.10.5), this technique is
useful.
4.11
Notes
Theorem 4.1.1 is new while Theorem 4.1.2 is due to Lakshmikantham and Salvadori [1]. The contents of Section 4.2 are essentially
new and the special cases considered are based on Strauss and Yorke
[1]. Section 4.3 consists of the work of Lakshmikantham and Leela
[1], while Section 4.4 contains the work of Lakshmikantham, Ladde
and Leela [6]. See Hale and Kato [1], Kato [1,2] concerning allied results for equations with innite delay. For a survey on equations with
unbounded delay see Corduneanu and Lakshmikantham [1]. Section
4.5 introduces a new technique which is adapted from Lakshmikantham and Leela [5]. All the results of Section 4.6 are taken from
Lakshmikantham [4]. The contents of Section 4.7 are new while
the results given in Section 4.8 are adapted from Lakshmikantham
and Trigiante [1]. Section 4.9 consists of the results from the work
of Bainov, Lakshmikantham and Simenov [1] and Martynyuk and
Slynko [1]. The material of Section 4.10 is adapted from Lakshmikantham and Leela [1]. See also Hale [1] and Redheer and Walter
[1] for the use of Lyapunov functions. For the use of vector Lyapunov functions in reaction-diusion systems see Lakshmikantham
and Leela [7]. The concept of matrix valued function for perturbed
motion is well-established in stability theory (see Martynyuk [16]
and Khoroshun and Martynyuk [1]).
5
STABILITY IN THE MODELS OF REAL WORLD
PHENOMENA
5.0
Introduction
In this chapter we oer several examples of real world models to

illustrate the general methods of stability analysis developed in the
previous chapters.
Section 5.1 deals with the stability problem of robot motion whose
mathematical model takes into account the dynamics of the environment interacting with the robot. We apply here some integral
inequalities from Chapter 1 of the book.
In Sections 5.2 and 5.3 the estimates of norms of solutions to dierential equations are obtained in terms of nonlinear and pseudo-linear
integral inequalities. This approach facilitates establishing the estimates of norms of solutions for some classes of systems of equations
of perturbed motion found in various applied problems.
In Section 5.4, we consider a problem of stability of regular synchronous generator of optical connected lasers.
Section 5.5 presents models from economics and using the method
of vector Lyapunov functions proves that a market tends to some
given evolution independent of initial conditions.
Finally in Section 5.6, we analyze a model of impulsive TakagiSugeno systems with application to the mathematical model in population growth under the impulsive control.

253
254
5.1
Stability in the Models of Real World Phenomena
Stability of a Robot Interacting with a Dynamic

Medium
A dynamic robot model is described by the dierential equation

H(q)
q + h(q, q)
= + J T (q)F,
(5.1.1)
where q, q,
q Rn are the vectors of the generalized coordinates,
velocities and accelerations of the robot; H(q) is the positive denite
matrix of inertia moments of manipulator mechanisms; h(q, q)
is the
n-dimensional nonlinear vector function which takes into consideration centrifugal, Coriolis and gravitational moments; = (t) is the
n-dimensional vector on input (control); J T (q) is the n m Jacobi
matrix associated with the motion velocity of control robot devices
and its generalized coordinates; F (t) is the n-dimensional vector of
generalized forces or generalized forces and moments acting on the
executive robot device due to the dynamic environment.
Under the condition when the environment does not admit eigen
motions independent of the motion of the executive robot organs,
the mathematical model of environment is described by the nonlinear
vector equation
M (s)
s + L(s, s)
= F,
(5.1.2)
s = (q),
(5.1.3)
where s is the vector of the environment motions; (q) is the vector function connecting the coordinates s and q. Note that in the
case of traditional hybrid control, the environment plays the role of
kinematic limitation and the relationship (5.1.3) becomes
(q) = 0.
(5.1.4)
Under certain assumptions the equation (5.1.2) may be represented

as
(5.1.5)
M (q)
q + L(q, q)
= S T (q)F,
where M (q) is the n m non-singular matrix; L(q, q)
is the nonlinear
T
n-dimensional vector function; S (q) is the n m matrix of the n
rank.
Robot Interacting with a Dynamic Medium
255
Thus the equation set (5.1.1), (5.1.5) represents the closed mathematical model of the robot interacting with the environment.
Let qp (t) be the n-dimensional vector of the program value of the
generalized coordinates, qp (t) be the n-dimensional vector of the program value of the generalized velocities, Fp (t) be the m-dimensional
vector of forces corresponding to the program values of the generalized coordinates and velocities. The program values of force
Fp (t) and those of functions qp (t), qp (t), qp (t) cannot be arbitrary
and should satisfy the relationship Fp (qp (t), qp (t), qp (t)) where
C(Rn Rn Rn , Rm ). The connected system of equations
(5.1.1), (5.1.5) can easily be reduced to the form
.
L(qp , qp ) + S T (q) S T (qp ) Fp
M (q)
q M (qp ) + L(q, q)
(5.1.6)
= S T (q)(F Fp ).
The n-dimensional vector of deviations of the program trajectory
from real one is designated by y. Then the equation (5.1.6) becomes
y + K(t, y, y)
= M 1 (y + qp )S T (y + qp )(F Fp ),
where
)
K(t, y, y)
= M 1 (y + qp ) L(y + qp , y + qp )
.
L(q, qp ) + M (y + qp ) M (qp ) qp +
. *
+ S T (y + qp ) S T (qp ) Fp .
The equation set (5.1.7) is transformed to the following form

dx
= A(t)x + (t, x) + (t, x)(t),
dt
where
x = (x1 , x2 )T , x1 = y, x2 = y,
&On
&In
&
K &&
A(t) = K &
,
&
&
y (y,y)=(0,0)
y (y,y)=(0,0)
(5.1.7)
256
On and In are the respective zero and unit matrices of dimension n,

0
t) = o(y2 + y
2 )1/2 ,
(t, x) =
,
0 (y, y,
0 (t, x1 , x2 )

0
(t, x) =
,
M 1 (x1 + qp )S T (x1 + qp )
(t) = F (t) Fp (t).
Within the general statement, the problem of choosing the program forces Fp (t) is associated with studying the solutions of the
dierential equation
d
= Q(),
dt
where Q C(Rm , Rm ), (t) = F (t) Fp (t), (t0 ) = 0 and
Q(0) = 0.
Thus the problem of stability of the robot motion interacting with
a dynamic environment results in the need to analyze the solutions
of the systems of equations
dx
= A(t)x + (t, x) + (t, x)(t),
dt
d
= Q(), (t0 ) = 0
dt
x(t0 ) = x0 ,
(5.1.8)
(5.1.9)
under certain assumptions of functions specifying the action of dynamic environment on the robot.
Consider the independent equation (5.1.9) which species the inuence of dynamic environment on the executive organ of the robot.
From (5.1.9) it follows that
t
(t) = 0 +
Q((s)) ds,
t t0 .
(5.1.10)
t0
The term in the equation (5.1.8) which species the action of

environment on the robot is designated by u(t, x) = (t, x)(t) for
(t, x) R+ D, D = {x : x < H}, H suciently small, the
function u(t, x) satises the inequality
u(t, x) p(t)
(5.1.11)
257
where p(t) is the function integrable over any nite time interval.
With
(5.1.12)
(t) = F (t) Fp (t)
and (5.1.10) representing the deviation of program value of the force
Fp (t) from the force F (t) acting due to the dynamic environment,
the action of environment on the robot may be estimated by the
function p(t). We introduce the designations
p0 = sup p(t),
t0
t+1
p1 = sup
p(s) ds,
t0
t+1
p2 (s) ds
p2 = sup
t0
1/2
.
Further the following denition will be used.

Denition 5.1.1 Let for any > 0 the values > 0 and > 0
be those for which the inequality x(t) < occurs for solving the
equation (5.1.8) with t 0 if x(0) < and one of the following
conditions is satised
(1) p0 ;
(2) p1 ;
(3) p2 .
Here we consider that robot motion is:
(a) stable with limited action of environment on the robot (case 1);
(b) stable with limited, on the average, action of environment on
the robot (case 2);
(c) stable with limited, on the quadratic average, action of environment on the robot (case 3).
It is of interest to consider the action of environment on the robot
when the limiting relationship u(t, x) 0, t is uniformly
satised over x with suciently low values x. This corresponds to
the choice of control in (5.1.1) when F (t) Fp (t), t .
258
In the case when H in the estimate of the domain D is not small

(H < ) and consequently, the large neighborhood of the equilibrium state of the robot-mechanical system is considered, the estimate
u(t, x) (t)x
(5.1.13)
should be taken instead of (5.1.11), where (t) is the integrable function such that

(s) ds < +.
(5.1.14)
0
Let us make the following assumptions on the equations (5.1.8),

(5.1.9):
I. The fundamental matrix X(t) of solutions of the rst approximations of the system (5.1.8) satises the inequality
X(t)X 1 (s) N e(ts)
(5.1.15)
where N and are positive constants independent of t0 . Note that

the condition (5.1.15) guarantees the exponential stability of the zero
solution of
dx
= A(t)x.
(5.1.16)
dt
II. The vector function (t, x) in (5.1.8) satises the following
condition: for each L > 0 the values D = D(L) and T = T (L) exist,
such that
(t, x) Lx
(5.1.17)
with x D and t T .
III. The inuence of the vector function of dynamic environment
on robot satises the condition u(t, x) 0 with t , uniformly
over x with suciently small values x.
Theorem 5.1.1 The equations (5.1.8) and (5.1.9) of the robot
movement interacting with the environment are assumed to be those
where the conditions IIII are satised. Then there exists t0 R+
259
such that any movement x(t; t0 , x0 ) of the robot simulated by the system (5.1.8) will approach to zero with t and suciently small
values x(t0 ).
Proof When the condition I is satised, the value L in (5.1.17) is
dened by the formula L = (2N )1 :
(t, x)
x,
2N
t T.
(5.1.18)
From the condition III it follows that > 0 exists such that
u(t, x) <
NL
,
2N
t T.
(5.1.19)
For a certain t0 R+ with t t0 we have

t
x(t) = W (t, t0 )x0 +
.
W (t, ) (, x( )) + u(, x( )) ds. (5.1.20)
t0
With the estimates (5.1.15), (5.1.17)(5.1.19) we nd from (5.1.20)

(tt0 )
x(t) N e
t
x0 + N L
e(ts) x(s) ds
t0
t
+N
(5.1.21)
e(ts) u(s, x(s)) ds.
t0
Let us designate M (t) = max x(s) and represent (5.1.21) as

t0 st
M (t) N x0 +
NL
N
N
N
M (t) +
x0 +
.
NL
NL
Since 2N ( N L)1 < , then M (t) < with all t t0 as soon as

x0 <
( N L)
<
.
4N
4N
Set = lim sup x(t). It is evident that 0 < + and the

t
sequence {tj }, j = 1, 2, . . . exists such that the limiting relationship

x(tj ) is valid with tj +, j +.
260
From the inequality (5.1.21) we obtain

tj /2

x(tj ) N x0 e(tj t0 ) + N L
e(tj s) x(s) ds
t0
tj
+ NL
tj /2

e(tj s) x(s) ds + N
e(tj s) u(s, x(s)) ds
t0
tj /2
tj
+N
e(tj s) u(s, x(s)) ds.
tj /2
For a given > 0 there exists J such that x(tj ) > for all
j J and x(t) < + with t tj /2. Consequently, with j J
we nd
N L 1 tj N L( + ) N 1 tj
e 2
+
e 2
N x0 e(tj t0 ) +
+
NL
max u(t, x(s)).
+
12 tj stj
N L( + )
is obtained as j +. Since
N L 1 < 1/2, we have < 3. It follows from arbitrariness of

that = 0. With the denition of , we may conclude that the
motion x(t) at vanishing interactions of robot with the environment
tends to the equilibrium state corresponding to the zero solution of
(5.1.8).
Thus
Further we study the motion of the robot interacting with dynamic environment under the conditions (5.1.13) and (5.1.14). For
providing sucient stability conditions the following Lemma is
needed.
Lemma 5.1.1 Let be the positive constant and the function
(t) C(R+ , R+ ) be such that

(s) ds < +
0
or
lim (t) = 0.
t+
261
Then
t
t
es (s) ds = 0.
lim e
t+
Proof Let us rst prove the case when (t) is integrable. For the
given > 0 we choose t to be large enough so that
(s) ds < ,
2
t/2
(s) ds < .
2
0
t/2
Then
et
t/2
t/2
t
es (s) ds e 2
(s) ds < ,
2
0
0
t
e (s) ds
t/2

(s) ds
t/2
Consequently,
t
(s) ds <
.
2
t/2
t
es (s) ds <
with a suciently large t. Therefore

t
t
es (s) ds = 0.
lim e
Consider the case (t) 0 with t +. If
es (s) ds < +
the statement of Lemma 5.1.1 is evident. On the other hand, using

the LHospital rule we obtain
1
lim t
t+ e
t
es (s) ds = lim
(t)
= 0.
262
Lemma 5.1.2 Let the function u(t) be continuous and nonnegative and satisfy the inequality
t
u(t) c +
.
ku(s) + (s) ds,
t 0,
where c and k are non-negative constants. Then

t
u(t) ckt +
ek(ts) (s) ds,
t 0.
Proof of this lemma follows by the standard method developed

in the theory of integral inequalities.
Theorem 5.1.2 The equations (5.1.8), (5.1.9) of robot movement
interacting with the environment are supposed to be such that
(1) the condition I is satised;
(2) for any > 0 there exists L = L() > 0 such that (t, x)
L()x with x , t 0;
(3) the vector function u(t, x) = (t, x)(t) satises the estimate
u(t, x) <
NL
,
2N
t 0.
Then any robot motion beginning in the domain {x R2n :

x(0) < /(2N )} will remain in the domain {x R2n : x < /2}
for all t 0.
Proof From the inequality (5.1.21) we obtain
t
u(t) N x0 +
0
.
N L()u(s) + N es u(s, x(s)) ds,
(5.1.22)
263
where u(t) = et x(t). Applying Lemma 5.1.2 to (5.1.22) we nd

t
N Lt
s
N Ls
N x0 + N e u(s, x(s))e
u(t) e
ds
0
and consequently,
x(t) N x0 e(N L)t + N e(N L)t
t
u(s, x(s))e(N L)s ds.
(5.1.23)
From the inequality x0 /(2N ) it follows that the rst summand in (5.1.23) will be smaller than /2 for all t 0. From condition 3 of Theorem 5.1.2 it follows that
t
( N L) (N L)t
e
(5.1.24)
e(N L)s ds .
2
2
0
Consequently, from (5.1.24) we obtain

1 e(N L)t
x(t) N x0 e(N L)t +
2

(N L)t
+
1 e(N L)t =
e
2
2
2
for all t 0.
The proof is complete.
Remark 5.1.1 From Theorem 5.1.2 it follows that if u(t, x)

0 or u(s, x(s)) ds < , the robot motion tends to the equilibrium
0
state as t +.
Case A. Consider the interactions of the robot with dynamic
environment when functions (t, x)(t) satisfy the estimate
(t, x)(t) (t)
(5.1.25)
for x r, r > 0, t 0 and

t+1
(s) ds 0
G(t) =
t
(5.1.26)
264
as t .
It is evident that the condition (5.1.26) will be satised if (t) 0

with t + or
(s) ds < +. It is shown (see Strauss and
0
Yorke [1]) that the function (t) may be determined as follows:
1
(t) = 0
with
with
with
t = 3n,
3n + n1 t 3(n + 1)
0 t 2.
1
n+1 ,
The robot motion under the conditions (5.1.25) and (5.1.26) is

described by the following statement.
Theorem 5.1.3 Let us assume that the equations (5.1.8), (5.1.9)
of the perturbed motion of the robot interacting with environment are
such that
(1) for the equations of the rst approximation (5.1.16) the condition I is satised;
(2) for the vector function (t, x) nonlinearities with any L > 0,
= (L) > 0 and = (L) > 0 exist such that
(t, x) Lx with x and t ;
(3) for the arbitrary solution (t) of the relationships (5.1.10) and
(5.1.12) which determine the quality of unsteady response to the
robot interaction with environment, the conditions (5.1.25) and
(5.1.26) are satised.
Then the time 0 and the domain S = {x R2n :
x < , > 0} will be found, such that the robot motion starting in the domain S at any time moment t0 , will approach the
equilibrium state, i.e. x(t) 0 at t +.
Proof When the condition (1) is satised the Cauchy matrix
W (t, s) of the linear approximation (5.1.16) of the system (5.1.8)
satises the condition W (t, t0 ) N e(tt0 ) at all t t0 . Let
0 < L < min {(/N ), r}. By the condition (2), (L) and (L) can be
265
chosen such that (L) 1 and (L) L. Besides, let (L) be

such that with t the estimate
Zt
e(N L)(ts) (s) ds <
(L)
= 1
2N
(5.1.27)
is valid.
It is easy to show that for all t t0 1 the inequality
Zt
t0
(s) ds
Zt
G(s) ds
t0 1
is satisfied. Thus for any k > 0 the estimate

"Z
#
s+1
Zt
Zt
ks
t0
e (s) ds
k(s+1)
(u) du ds =
t0 1
Zt
ek(s+1) G(s) ds
t0 1
(5.1.28)
is valid
With (5.1.28) we obtain
kt
Zt
t0
ks
kt
e (s) ds e
Zt
ek(s+1) G(s) ds.
(5.1.29)
t0 1
Applying the LHospital rule to the expression in the right-hand

side of the inequality (5.1.29) we can find
kt
lim e
Zt
ek(s+1) G(s) ds = 0
t0 1
whence it follows that the inequality (5.1.27) is justified. Then let

(L)
t0 and kx(t0 )k < 1 =
. From the equality (5.1.20) we
2N
obtain
(tt0 )
kx(t)k N 1 e
+N
Zt
t0

e(ts) Lkx(s)k + (s) ds,
266
thus
t
x(t)e
N 1 e
t0
.
N Lx(s)es + N es (s) ds.
(5.1.30)
t0
Let us designate x(t)et = w(t) and use Lemma 5.1.2 for the
inequality (5.1.30). It is easy to see that
t0 N L(tt0 )
w(t) N 1 e
t
eN L(ts) N es (s) ds,
+
t0
or
(N L)(tt0 )
x(t) N 1 e
t
+N
e(N L)(ts) (s) ds.
t0
Then by the condition (5.1.27) we nd

t
x(t) N 1 + N
e(N L)(ts) (s) ds <
+ N 1 = .
2
t0
Thus, x(t) < for all t t0 and the limiting relationship

x(t) 0 is satised for t +. The statement is proved.
Case B. Three conditions for the vector function (t, x)(t),
(t) = F (t) Fp (t) will be taken into consideration which specify
the robot interacting with the dynamic environment. The following
estimate of the function of transient process in (5.1.8) is needed.
Lemma 5.1.3 Let the following conditions be satised for the
equations of perturbed motion (5.1.8):
(1) the Cauchy matrix W (t, t0 ) of the equations of the rst approximation (5.1.16) satises the condition I;
(2) for the vector function (t, x) with each L > 0, a certain value
H = H(L) > 0 exists such that (t, x) Lx for x H
and t 0;
267
(3) for any function (t), satisfying the relationships (5.1.10) and
(5.1.12) the estimation (5.1.11) holds for all x < H and
t 0.
Then for suciently small initial perturbations x0 = x(0) and
(0) = F (0) Fp (0) the transient process in (5.1.8) satises the
estimate
x(t) N (1 (t) + 2 (t)),
(5.1.31)
where
1 (t) = et x0 , x0 = x(0),
t
t
es p(s) ds, = N L.
2 (t) = e
0
The estimate (5.1.31) follows from Lemma 6.1 of Barbashin [1,

p. 185], where the function 2 (t) is shown to satisfy one of the following inequalities for all t 0
2 (t) p0 1 ,
2 (t) p1 e (1 e )1 ,

2
1 1/2
1 e
.
2 (t) p2 (1 e )
2
The sucient conditions which provide the stability of motion of

the robot interacting with the environment are given in the following
statement.
Theorem 5.1.4 The equations of perturbed motion of the robot
interacting with the environment are supposed to be such that
(1) for the equations of the rst approximation (5.1.16) the condition I is satised;
(2) for the vector function of nonlinearities (t, x) with any L > 0,
= (L) > 0 exists such that (t, x) Lx with x H
and t 0;
268
(3) for any function (t) which satisfies the relationships (5.1.10)
and (5.1.12) for all kxk H and t 0 the estimate (5.1.11)
and one of the inequalities are satisfied
;
2N

p1 <
e
1 e ;
2N

1/2
p2 <
1
e
.
2N e2 1
p0 <
(5.1.32)
(5.1.33)
(5.1.34)
Then under any initial condition

x0 = x(0),
(0) = F (0) Fp (0)
(5.1.35)
for which kx0 k < (2N )1 , the transient process of the system
(5.1.8) satisfies the estimate
kx(t)k N (1 (t) + 2 (t))
(5.1.36)
for all t 0 and kx(t)k < .

The Proof of Theorem 5.1.4 is based on the estimate of the transient process (5.1.31). Under the initial conditions (5.1.35) when
kx0 k < (2N )1 , the estimate 1 (t) < (2N )1 is valid for the
function 1 (t) for all t 0. When the conditions (5.1.32)(5.1.34)
are satisfied the estimate 2 (t) < (2N )1 is valid for the function
2 (t). Therefore it follows from (5.1.36) that kx(t)k < for all
t 0. The proof of Theorem is complete.
Next we will show that the motion of the robot interacting with
the environment may be dissipative under appropriate limitation on
the initial state x0 and the function (t) = F (t) Fp (t).
Theorem 5.1.5 Let us suppose that for the equation (5.1.8) of
perturbed motion of robot interacting with environment
(1) the conditions (1)(2) of Theorem 5.1.4 hold;
269
(2) the inequalities
;
N

p1 < e 1 e ;
N

1/2

2
1 e ,
p2 <
2
N e 1
p0 <
(5.1.37)
(5.1.38)
(5.1.39)
are satised in -neighborhood of the state x = 0, i.e. with all

(x = 0) {x : x < } where 0 < < 1, 0 < < (2N )1 .
Then the positive number R+ exists such that for t >
and x0 < the transient process in (5.1.8) satises the estimate
x(t) < for all t .
Proof Consider the estimate (5.1.36). Then choose >
ln N (1 p)1 and the estimate for the functions 1 (t) and 2 (t)
can be obtained. For t > we have 1 (t) = et x0 < (1 p)N 1
for all t . When at least one of the conditions (5.1.37)(5.1.39)
is satised, 2 (t) < N 1 is obtained for all t . It follows from
the estimate (5.1.36) that the transient process in the system will be
damping, i.e. x(t) < for all t > .
1
Further the equations of the perturbed motion (5.1.8) will be considered under the following assumptions:
I . The matrix A(t), the vector function of nonlinearity (t, x) and
the vector function (t, x)(t) where (t) = F (t) Fp (t) are
continuous and periodic with respect to t. The period of these
functions is supposed to be common, for example, unity.
II . As above, the assumption I is preserved for the case of periodic
matrix A(t), i.e.
W (t, s) N e(ts)
(5.1.40)
where W (t, s) = X(t)X 1 (s).

III . The vector function of nonlinearities (t, x) in the domain
x < H, t 0 satises the Lipschitz condition
(t, x) (t, y) Kx y.
(5.1.41)
270
IV . The constants N , , K in the inequalities (5.1.40), (5.1.41)

satisfy the inequality = N K > 0.
Theorem 5.1.6 Suppose that for the equations of perturbed motion (5.1.8) for the robot interacting with the environment, the conditions I IV are satised. Besides, one of the conditions (5.1.37)
(5.1.39) is satised. Then in the domain x < H(2N )1 the periodic
robot motion z(t) is possible, and for any other motion x(t) of the
robot, which is started in the domain x(0) H(2N )1 , the limiting relationship x(t) z(t) 0 with t + is valid, i.e. the
periodic robot motion is asymptotically stable.
The Proof of this theorem is based on the principle of contracting
mappings and Theorem 6.4 in Barbashin [1].
5.2
Stabilization of Motions of Ane System
Consider an ane system with many controlling bodies

dx
= A(t)x(t) +
Gi (t, x(t))ui (t) + Bu0 (t),
dt
l
(5.2.1)
i=1
y(t) = Cx(t),
(5.2.2)
x(t0 ) = x0 ,
(5.2.3)
where x Rn , A(t) is an n n-matrix with continuous elements on

any nite interval, Gi (t, x) is an n m-matrix, the control vectors
ui (t) Rm for all i = 1, 2, . . . , l, B is an nm-matrix and the control
u0 (t) Rm , C is a constant n n-matrix, x0 is a vector of the initial
states of system (5.2.1). With regard to system (5.2.1) the following
assumptions are made:
A1 . Functions Gi (t, 0) = 0, i = 1, 2, . . . , l, for all t 0;
A2 . There exists a constant n m-matrix K0 such that for the
system
dy
= (A(t) BK0 C) y
dt
Stabilization of Motions of Affine System
271
the fundamental matrix (t) satises the estimate

(t)1 (s) M e(ts) ,
for t s t0 , where M and are some positive constants.
A3 . There exist constants i > 0 and q > 1 such that
Gi (t, x) i xq ,
for all i = 1, 2, . . . , l.
The following assertion takes place.
Theorem 5.2.1 Let conditions of assumptions A1 A3 be satised and, moreover,
1 qM
q+1
t
(Ki C)x0 qI
i=1
where =
eqs ds > 0,
i .
i=1
Then the controls

ui (t) = Ki y(t),
i = 1, 2, . . . , l,
u0 (t) = K0 y(t)
stabilize the motion of system (5.2.1) to the exponentially stable one.

Proof. Let the controls ui (t) = Ki y(t) and u0 (t) = K0 y(t) be
used to stabilize the motion of system (5.2.1). Besides, we have

dx
= (A(t) BK0 C)x(t)
Gi (t, x(t))(Ki Cx(t)).
dt
l
i=1
and
x(t) = (t)
t
(t0 )x0
t0
(t)1 (s)
Gi (s, x(s))(Ki Cx(s)) ds.
i=1
(5.2.4)
272
In view of conditions of Theorem 5.2.1 we get from (5.2.4) the

estimate of norm of solution of system (5.2.1) in the form
t
x(t) x0 M e
t
+
M e(ts)
Ki Cx(s)q+1 ds.
i=1
(5.2.5)
We transform inequality (5.2.5) to the form

t
x(t)e x0 M +
t
M eqs
Ki Cx(s)es q+1 ds.
i=1
(5.2.6)
Applying Corollary 2.2.3 to inequality (5.2.6) we get

M x0
x(t)et
1
l
t

q
1 qM q+1 (Ki C)x0 qI eqs ds
0
i=1
M x0
1+
M q+1
l
(Ki C)x0 qI
i=1
1
q
(eqt 1)
M x0
M q+1
l
(Ki C)x0 q 1
i=1
for all t 0.
If condition
1 qM
q+1
t
(Ki C)x0 qI
i=1
eqs ds > 0,
of Theorem 5.2.1 is satised, then

M q+1
1
l

i=1
(Ki C)x0 q
>0
273
Synchronization of Motions
and for the norm of solution x(t) we have the estimate

x(t) M0 x0 et
for all t 0, where
M0 =
M

1
M q+1
l
(Ki C)x0 q q
i=1
This completes the proof of Theorem 5.2.1.
5.3
The theory of motion synchronization studies the systems of dierential equations of the form (see Rozo [1] and bibliography therein)
dx
= f (t, x, ),
dt
x(t0 ) = x0 ,
(5.3.1)
where f (t, x, ) : R+ Rn [0, 1] Rn , f is a function continuous

with respect to t, x, and periodic with respect to t with the period
T , and is a small parameter. Alongside system (5.3.1) we shall
consider an adjoint system of equations
d
x
= g(
x),
dt
where
1
g(x) =
T
x
(t0 ) = x0 ,
(5.3.2)
T
f (s, x, 0) ds.
0
Assume that in the neighborhood of point x0 for suciently small

value of for any t [0, T ] the vector-function f and its partial
derivatives are continuous. Designate
1 1 1
1

1 f 1 1 fi 1
1, 1
1
f (t, x, ), 1
M=
max
1 1 1 vj 1 .
t[0,T ],xx0 d,
It is clear that the solutions of equations (5.3.1) and (5.3.2) remain
in the neighborhood x x0 d for |t| < dM 1 .
274
With allowance for

t
x(t, ) = x0 +
f (s, x(s, ), ) ds
0
and
t
x
(t, ) = x0 +
g(
x(s, )) ds,
0
we compile the correlation

t
[f (s, x(s, ), ) f (s, x(s, ), 0)] ds
x(x, ) x
(t, ) =
0
t
[f (s, x(s, ), 0) f (s, x
(s, ), 0)] ds
(5.3.3)
t
[f (s, x
(s, ), 0) g(
x(s, ))] ds.
+
0
As it is shown in monograph Rozo [1] for the rst and third summands in correlation (5.3.3) the following estimates hold true
1
1 t
1
1
1
1
1 [f (s, x(s, ), ) f (s, x(s, ), 0)] ds1 M t,
1
1
(5.3.4)
1
1 t
1
1
1
1
(s, ), 0) g(
x(s, ))] ds1 2M T + 4M 2 T t.
1 [f (s, x
1
1
(5.3.5)
To estimate the second summand we assume that there exists an

integrable function N (t) : R+ R+ such that for any t1 , t2 [0, T ]
(0 t1 < t2 )
t2
N (s) ds > 0
t1
275
and > 1 such that

f (t, x, 0) f (t, x
, 0) N (t)x x

(5.3.6)
in the domain of values t [0, T ] and x, x

D.
In view of estimates (5.3.4)(5.3.6) we nd from (5.3.3)
x(s, ) x
(s, ) (2M T + (4M 2 T + M )t)
s
(, ) d
+ N ( )x(, ) x
(5.3.7)
for all s t.
Let there exist [0, 1] such that
.1
1 ( 1) (2M T + (4M 2 T + M )t)
T
N (s) ds > 0
0
for all < . Then the norm of divergence of solutions x(t, ) and
x
(t, ) under the same initial conditions is estimated as follows
x(t, ) x
(t, ) [2M T + (4M 2 T + M )t]

1
T
1
.1
2
1 ( 1) (2M T + (4M T + M )t)
N (s) ds
0
(5.3.8)
for all t [0, T ] and for < .
Estimate (5.3.8) is obtained from inequality (5.3.7) by the application of Corollary 2.2.1.
If in estimate (5.3.6) = 1 and N (t) = M , then the application of
the Gronwall-Bellman lemma to inequality (5.3.7) yields the estimate
of divergence between solutions in the form (see Rozo [1])
x(t, ) x
(t, ) [2M T + (4M 2 T + M )t] exp(M T )
for all t [0, T ].
276
5.4
Stability of Regular Synchronous Generation of

Optically Coupled Lasers
This section deals with the stability with respect to linear approximation of some periodic solutions to a system of nonlinear dierential
equations. This system describes some experimental realization of a
chaotic CO2 -laser with a 100 per cent depth-modulated periodic
pumping by alternate current.
Variation of the factor of strengthening g and amplitude E of a
synchronized eld of two optically coupled lasers is described by the
simplest model
g = g0 (t) g(1 + E 2 ),
(5.4.1)
E = (g
gth )E/2,
where is an ecient time of relaxation of the active medium
( 1), g0 (t) = A(1
+ sin t) is a (2/)-periodic pumping,
gth = gth + 2M (1 1 (/M )2 ) is a threshold coecient of
strengthening. Here gth means threshold strengthening, M is a real
positive coupling factor, is a value of resonance eigenfrequency detuning (further on detuning). For the problems considered below
the dierence of real medium kinetics of CO2 -from the model one is
not of essential importance (see Likhanski et al. [1]).
The mode of phase synchronization, for which the eld amplitudes
of both lasers are equal at any moment and the phase is constant and
depends on detuning, is realized under the condition || < M . Moreover, the dynamics of two coupled lasers coincides with the dynamics
of one equivalent laser whose threshold grows with the growth of detuning. In the mode of synchronous generation (for a xed M ) the
growth of detuning corresponds to lessening of the parameter A/
gth
for the equivalent laser. Due to the complex bifurcation diagram of
the laser with periodic pumping this results in generation of both
chaotic and regular signals.
Designate by (gT (t), ET (t))T , t [t0 , ) = T0 , t0 0, T -periodic
solution of system of equations (5.4.1) with the initial condition
g(t0 ) = g0 ,
E(t0 ) = E0
(5.4.2)
and dene variables y1 and y2 of the perturbed motion of system
277
Stability of Coupled Lasers
(5.4.1) as
y1 = g gT (t),
y2 = E ET (t).
Then the perturbed equations of motion(5.4.1) are

y 1 = (1 + ET2 (t))y1 2gT (t)ET (t)y2
2ET (t)y1 y2 gT (t)y22 y1 y22 ,
(5.4.3)
2y 2 = ET (t)y1 + (gT (t) gth )y2 + y1 y2 .

For the linear approximation of system (5.4.3) (designated as
(5.4.3 )) we construct an auxiliary matrix-valued function

p20 y12
p11 (t)y1 y2
U (t, y1 , y2 ) =
p11 (t)y1 y2
p02 y22
where p20 and p02 are nite positive constants, p11 (t) C 1 (R, R),
and a scalar Lyapunov function
v(t, y, ) = T U (t, y1 , y2 ),
(5.4.4)
where y = (y1 , y2 )T and = (1 , 2 )T > 0.

Total time derivative of function (5.4.4) found by virtue of linear
approximation of system (5.4.3) is
&
dv &&
= (212 p20 (1 + ET2 (t))/ + 1 2 p11T (t)ET (t))
dt &(5.4.3 )
y12 + (22 p02 (gT (t) gth ) 41 2 p11T (t)
gT (t)ET (t)/ )y22 = s20 (t)y12 + s02 (t)y22 ,
if p11T (t) is assumed to be a T -periodic solution of the linear dierential equation
gth )p11
p 11 = ((1 + ET2 (t))/ gT (t) +
+ (21 p20 gT (t)/( 2 ) 2 p02 /(21 ))ET (t).
(5.4.5)
Conditions of uniform asymptotic stability of T -periodic solution of

system of equations (5.4.3) (noncritical case) are established in the
form of a system of inequalities
p20 p02 p211T (t) > 0,
s20 (t) < 0,
s02 (t) < 0 for all
(5.4.6)
t [t , t + T ],
t T0 .
278
Thus, the problem on asymptotic stability of some signals of the

equivalent CO2 -laser is reduced to the problem of nding T -periodic
solutions to nonlinear non-stationary initial problem (5.4.1)(5.4.2)
and linear inhomogeneous equation (5.4.5) with periodic coecients
and the initial condition
p11 (t0 ) = p110 .
(5.4.7)
This, in its turn, involves preliminary study of the problem on

the domain where equations (5.4.1) and (5.4.5) form T -system (see
Samoilenko and Ronto [1]) and establishing existence conditions for
the corresponding T -periodic solutions passing through the point
(g0 , E0 , p110 ) at the initial instant t0 .
We set
T = k(2/),
(5.4.8)
where k is a positive integer, and dene the domain D R3 which
singles out T -system, by the inequalities
D:
|g| gmax ,
|E| Emax ,
|p11 | p11 max .
(5.4.9)
We introduce the vector M = (M1 , M2 )T and the scalar M3 which

bounds for all t T0 and (g, E, p11 ) D the absolute values of
the corresponding right-hand sides of equations (5.4.1) and (5.4.5)
(further on f1 , f2 and f3 ):
2
))/,
M1 = (2A + gmax (1 + Emax
M2 = (gmax + gth )Emax /2,

2
)/ + gmax + gth )p11 max
M3 = ((1 + Emax
(5.4.10)
+ (21 p20 gmax /( 2 ) + 2 p02 /(21 ))Emax .

Continuous vector function f = (f1 , f2 )T periodic in t with the period T satises in T0 [gmax , gmax ] [Emax , Emax ] the Lipschitz
condition with the matrix

2 )/
2gmax Emax /
(1 + Emax
,
K=
(gmax + gth )/2
Emax /2
and the scalar continuous periodic function f3 in T0
[p11 max , p11 max ] with the constant
2
K3 = (1 + Emax
)/ + gmax +
gth .
279
Following the denition of T -system and relating with vectorfunction (f T , f3 )T and domain D the nonempty set Df of points
R3 contained in D together with its T2 (M T , M3 )T -neighborhood the
conditions dening T -system are obtained in the form of a system of
inequalities
2gmax T M1 > 0, 2Emax T M2 > 0,
2p11 max T M3 > 0,

2
T
T K11 + K22 + (K11 K22 ) + 4K12 K21
< 1,
K3 < 1.
Moreover, it is also assumed that the initial value (g0 , E0 , p110 ) belongs to Df .
The immediate construction of the desired T -periodic solutions is
achieved, for example, by the method of trigonometric collocations
by a numerical-analytical scheme. To this end, we assume that the
values of functions fj (t, g, E, p11 ), j = 1, 2, 3, calculated basing on
the m-th approximation to the desired periodic solution coincide in
T
N = 2r + 1 collocation points ti = i N
, i = 0, 1, . . . , 2r, with the
values of the trigonometric polynomials
fjm = m
j0 +
r

m
(m
jl cos lt + jl sin lt),
(5.4.11)
l=1
where = 2/T . Then the vectors of the coecients

m
m
m
m T
fjm = (m
j0 , j1 , j1 , . . . , jr , jr )
(5.4.12)
of trigonometric polynomials (5.4.11) are expressed via the respective

vectors of values of these polynomials
2r
fjmM = (fj (ti , gm (ti ), E m (ti ), pm
11 (ti )))i=0
with the help of the matrix

= [pq ]N
p,q=1 ,
where
pq
p = 1,

N
2
p = 2, 4, . . . , 2r,
= N cos p(q 1) N ,
2 sin (p 1)(q 1) , p = 3, 5, . . . , N,
N
N
280
and
fjm = fjmM .
By introducing into consideration the N N -two-diagonal matrix
0 0
0
0
0
... 0
0
0 0 1 0
0
... 0
0
0 1
0
0
0
.
.
.
0
0
0
0 0
0
0 2 . . . 0
1
=
1
0
... 0
0
0
0 0
2
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1
0 0
0
0
0
. . . 0 r
1
0 0
0
0
0
. . . r
0
and N -dimensional vectors

T
r

m
m
m

m

(jl cos lt0 + jl sin lt0 ), 0, . . . , 0 ,
zj = j0 +
l=1
where

m
m
m
m T
1 m
(m
j0 , j1 , j1 , . . . , jr , jr ) = fj ,
we obtain the vectors of the coecients of (m+1)-th trigonometric

approximation to the desired T -periodic solution in the form
g m+1, = g0 + 1 f1m z1m ,
E m+1, = E 0 + 1 f2m z2m ,
1 m
pm+1,
= p0
z3m ,
11 + f3
11
where g0 , E 0 and p0
11 are the vectors of the coecients of appropriate zero approximations.
The form of the zero approximation (g0 (t), E 0 (t))T and the vector
of the initial values at the collocation points and the initial vector
of the coecients of the right-hand sides f1 , f2 of equations (5.4.1)
respectively are taken based on solution of system (5.4.1) linearized
by the equation for g
t
g0 (t) = Cg e +
A
(sin t cos t) + A,
1 + 2 2
g0 (t0 ) = g0 ,
281

t
1
Cg e + (A gth )t
E (t) = CE exp
2

A
1
cos t + sin t
,
1 + 2 2
0
E 0 (t0 ) = E0 ,
where constants Cg and CE are dened univalently. We take solution

of the corresponding homogeneous initial problem (5.4.5), (5.4.7) as
p011 (t), assuming T -periodic functions to be known
gT (t) g (t) =
m
gjm eijt ,
ET (t) E (t) =
m
j=r
Ejm eijt ,
j=r
m
m
m
m
m
m
where gjm = (m
gj igj )/2, gj = gj , Ej = (Ej iEj )/2,
m = E m , and m , m and m , m stand for coecients (5.4.12)
Ej
gj gj
Ej Ej
j
of the corresponding trigonometric series (5.4.11). Then
p011 (t)
= Cp11 exp{((1 +
m
Ejm Ej
)t
Ejm Esm
gth g0m )t
ei(j+s)t )/ + (
i(j + s)
s=j
gjm ijt

j=0
ij
},
p011 (t0 ) = p110 .
The control of convergence of the described iteration process

of nding T -periodic solution is performed by comparing with a
pre-given accuracy 1 the dierence between the vectors of coecients of the m-th and (m + 1)-th trigonometric approximations for gT (t), ET (t), p11T (t) with zero-vector, and by comparing with a pre-given accuracy 2 the mean values of functions
fj (t, gm (t), E m (t), pm
11 (t)), taken over a period, with zero. The latter
condition is necessary and sucient (see Samoilenko and Ronto [1])
for the existence of periodic solutions of the period T passing through
the point (g0 , E0 , p110 ) Df for t = t0 and is an indicator of a good
choice of the values k (see (5.4.8)), gmax , Emax , p11 max (see (5.4.9)),
t0 , g0 , E0 , p110 (see (5.4.2), (5.4.7)), p20 , p02 , 1 , 2 (see (5.4.10)) and
the parameter values of the system under consideration.
For the values = 40.96241, A = 0.39856, gth = 0.4, M = 0.1,
= 400, = 0.001, t0 = 0, gmax = 4.77321, g0 = 0.40001,
Emax = 9.27539, E0 = 0.03817 with the use of N = 31 collocation
282
Figure 5.2. Graphs of (4/)-periodic functions gT (t) and ET (t).
283
Figure 5.3. Phase trajectory corresponding to (4/)-periodic solution

of (gT (t), ET (t))T .
Figure 5.4. Graph of (4/)-periodic function p11T (t).
284
Figure 5.5. Phase trajectory corresponding to (6/)-periodic solution

of (gT (t), ET (t))T .
points during 5 iterations (1 = 1025 , 2 = 105 ) a periodic solution

was constructed for the initial problem (5.4.1), (5.4.2) with duplication of the period (k = 2). The corresponding graphs are shown
in Figures 5.2 and 5.3. Uniform asymptotic stability of the corresponding zero solution of system (5.4.3) is established during 4 iterations by constructing with the same accuracy the periodic function
p11T (t) (Figure 5.4) satisfying conditions (5.4.6). Here 1 = 8.01158,
2 = 4.38394, p20 = 2.97746, p02 = 0.14038, p11 max = 4.65370,
p110 = 0.42974 and 1 = 651, 2 = 10358490, 3 = 5168882277,
4 = 3194900 are the scale multiplies.
Uniform asymptotically stable signal with triple period with respect to the pumping period (Figure 5.5) is investigated in the same
way for the parameters changed as compared with the previous example = 52.116990, A = 0.399742, gmax = 0.904412, g0 = 0.399723,
Emax = 8.306199, E0 = 0.002538, 2 = 106 , 1 = 8.776919,
2 = 0.385523, p20 = 2.070760, p02 = 8.392944, p11 max = 8.833734,
p110 = 0.083020, 2 = 13105471, 3 = 8472130609.
Sustainable Development
285
The considered examples demonstrate the possibility of parallel

solution of some problems on the spectrum and structure of collective
mode as well as their stability and competition between the mode
of composed resonator. The method of constructing an auxiliary
function pointed out in the context of matrix-valued Lyapunov functions allows to calculate stability domains of some periodic signals of
coupled lasers with periodic pumping in the regime of synchronous
generation.
5.5
Models of World Dynamics and Sustainable

Development
The Forrester model of world dynamics (see Forrester [1]) is constructed in terms of the approach developed in the investigation of
complex systems with nonlinear feedbacks. In the modeling of world
dynamics the following global processes are taken into account:
(i) quick growth of the world population;
(ii) industrialization and the related production growth;
(iii) restricted food resources;
(vi) growth of industrial wastes;
(v) shortage of natural resources.
The main variables in the Forrester model are:
(1) population P (further on the designation X1 is used);
(2) capital stocks K (X2 );
(3) stock ratio in agricultural industry X (X3 );
(4) level of environmental pollution Z (X4 );
(5) quantity of nonrenewable natural resources R (X5 ).
Factors through which the variables X1 , . . . , X5 , eect one another
are:
286
- relative number of population Pp (population normed to its

number in 1970 .);
- specic stocks Kp ;
- level of living standard C;
- relative level of meals F ;
- normed value of specic stocks in agricultural industry Xp ;
- relative pollution Zs ;
- ratio of the resources left RR .
In addition to the enumerated factors Forrester also considers the
notion of quality of living Q. This factor depends on the variables
Pp , C, F and Zs : Q = QC QF QP QZ .
For the variables P , K, X, Z, R interpreted as system equations,
the equations of the type
dy
= y+ y,
dt
(5.5.1)
are written, where y + is a positive rate of velocity growth of the variable; y is a negative rate of velocity diminishing of the variable y.
In a simplied form the world dynamics equations are
dP
= P (B D),
dt
dK
1
= K+ TK
K,
dt
dX
= X+ TX1 X,
dt
dZ
= Z+ TZ1 Z,
dt
dR
= R ,
dt
(5.5.2)
where B is a birth rate, D is a death rate, K+ is a velocity of capital

stocks production, X+ is an increment of the ratio of agricultural
industry stocks, Z+ is a velocity of pollution generation, TZ is a
characteristic time of natural decay of pollutants, and R is a velocity
of resource consumption.
Mathematical analysis of model (5.5.2) reveals the existence of
stationary and quasi-stationary solutions which are interpreted as a
global equilibrium and a stable society.
287
Let a nation N (a totality of international organizations) form

the public opinion about global processes occurring on a certain level
of the system. The measure of the change of the public opinion (t)
will be modeled on each system by the equation (see Martynyuk [9])
d2
+ m2 = 0,
dt2
(t0 ) = 0 ,
(t0 ) = 0 .
(5.5.3)
Here the value m is a function of variables (1)(5) at times t = t0 .

Moreover, for the system levels the equations of (5.5.1) type are
written
dy
= y + y + b(t),
(5.5.4)
dt
where the discontent function b(t) is as follows
b(t) = ge|(t)| ,
= const > 0.
(5.5.5)
Here g is a factor of discontent reecting the change of the level

of living standard of the countries involved into world dynamics.
Correlation (5.5.5) models the increase (decrease) of discontent of
the current global processes depending on changes of the measure of
the public opinion.
Thus, the Forrester model (5.5.1)(5.5.2) is generalized by the
equations
dX1
= X1 (B D) + g1 e|(t)| ,
dt
dX2
1
= K+ TK
X2 + g2 e|(t)| ,
dt
dX3
= X+ TX1 X3 + g3 e|(t)| ,
dt
(5.5.6)
dX4
= Z+ TZ1 X4 + g4 e|(t)| ,
dt
dX5
= R + g5 e|(t)| ,
dt
d2
+ m2 = 0,
dt2
where g1 , . . . , g5 are the discontent factors on the corresponding level
of the system.
288
It is proposed to describe general nonlinear model of world dynamics by a system of dierential equations of the type
dXi
= Wi (X) + gi e|(t)| ,
dt
d2
+ m2 = 0, i = 1, 2, . . . , N.
dt2
(5.5.7)
(5.5.8)
Here X = (X1 , . . . , X5 , . . . , XN ) S(H), where X1 , . . . , X5 are

the Forrester variables and X5+1 , . . . , Xn are some other variables
N
involved into the world dynamics equations, Wi : S(H) R+
is a vector-function with the components describing the variation
of parameters on the appropriate system level. It is assumed
that the solution (X T (t), (t))T of system of coupled equations
(5.5.7)(5.5.8) exists for all t t0 with the initial conditions
N , R R).
(X0T , 0 , 0 )T int(R+
Assume that the system of nonlinear equations
W1 (X) + g1 e|(t)| = 0,

WN (X) + gN e|(t)| = 0
possesses a quasistationary solution Xn (t) = (X1n (t), . . . , XnN (t))T
for any bounded function (t) being a solution of equation (5.5.8).
Moreover, the Lyapunov substitution
Y (t) = X(t) Xn (t)
brings system of equations (5.5.7) to the form
dY
= Y (t, Y ),
dt
(5.5.9)
where Y (t, Y ) = W (Y + Xn (t)) + Ge|(t)| (W (Xn (t)) + Ge|(t)| ).

It is clear that Y (t, 0) = 0 for all t 0. System (5.5.9) is a system
of perturbed equations of world dynamics.
The problem of sustainable development is associated with the
analysis of solution Y = 0 of equation (5.5.9). The stability analysis
289
of solutions will be carried out with respect to two measures H0 and

H taking the values from the sets
= {H C(R+ RN , R+ ) : inf H(t, Y ) = 0};
(t,Y )
0 = {H { : inf H(t, Y ) = 0 for every t R+ }.

Y
We need the following denition.

Denition 5.5.1 The world dynamics (5.5.7)(5.5.8) has sustainable development with respect to two measures if for every > 0
and t0 R+ there exists a positive function (t0 , ) > 0 continuous
in t0 for every such that the condition H0 (t0 , Y0 ) < implies the
estimate H(t, Y (t)) < for all t t0 for any bounded solution (t)
of equation (5.5.8).
Note that if system (5.5.7) having no zero solution (W (0, (t)) = 0
for X = 0) has the nominal solution Xn (t) then the measures H0 and
H can be taken as follows: H(t, X) = H0 (t, X) = X Xn (t), where
is an Euclidean norm of the vector X. If it is of interest to study
stability of the development in the Forrester variables, the measures
H0 and H are taken as: H(t, X) = X Xn (t)s , 1 s 5, and
H0 (t, X) = X Xn (t). This corresponds to stability analysis of
system (5.5.7) in two measures with respect to a part of variables.
For system (5.5.9) assume that the elements uij (t, Y ) of the
matrix-valued function
U (t, Y ) = [uij (t, Y )],
i, j = 1, 2, . . . , m,
m < N,
are constructed, where uii C(R+ RN , R+ ) and uij C(R+

RN , R) for (i = j) [1, m]. The function
V (t, Y, w) = wT U (t, Y )w,
w Rm ,
(5.5.10)
is considered together with the function

D + V (t, Y, w) = wT D + U (t, Y )w,
(5.5.11)
where D + U (t, Y ) is the upper right Dini derivative calculated

element-wise for the matrix-valued function U (t, Y ).
290
Conditions of the sustainable development in two measures

(H0 , H) are established in the following result.
Theorem 5.5.1 Let the functions in equations of global dynamics
(5.5.7)(5.5.8) be dened and continuous in the domain of values
(t, Y, ) R+ S D. If, moreover,
(1) measures H0 and H are of class ;
(2) function (5.5.10) satises the condition V (t, Y, w) C(R+
S Rm , R+ ) and is locally Lipchitz in Y ;
(3) function V (t, Y, w) satises the estimates
(a) a(H(t, Y )) V (t, Y, w) b(t, H0 (t, Y )) for all (t, Y, w)
S(h, H) Rm or
(b) a(H(t, Y )) V (t, Y, w) c(H0 (t, Y ))
where a, c Kclass and b CKclass of comparison functions;
(4) there exists a matrix-valued function (Y, w), C(RN Rm ,
Rmm ) and (0, w) = 0 for all (w = 0) Rm such that
8
w)e
D + V (t, Y, w) eT (Y,
for all (t, Y, w) S Rm , where e = (1, 1, . . . , 1)T Rm ,
8
S (RN R+ ), (Y,
w) = 12 ((Y, w) + T (Y, w)) for any
bounded solution (t) of equation (5.5.8).
Then
(a) world dynamics (5.5.7)(5.5.8) has sustainable development
8
with respect to two measures if the matrix (Y,
w) is negative
semi-denite, the measure H is continuous with respect to the
measure H0 and condition (3)(a) is satised;
(b) world dynamics (5.5.7)(5.5.8) has uniformly sustainable de8
velopment with respect to two measures if the matrix (Y,
w)
is negative semi-denite, the measure H is uniformly continuous with respect to the measure H0 and condition (3)(b) is
satised.
291
Proof We note that function V (t, Y, w) determined by formula

(5.5.10) is a scalar pseudo-quadratic form with respect to w Rm .
Therefore, the property of denite sign of function (5.5.10) with respect to the measure H does not require the H-sign-deniteness of
the elements uij (t, x) of matrix U (t, Y ). First we shall prove assertion (a) of Theorem 5.5.1. Conditions (1), (2), and (3a) imply that
the function V (t, Y, w) is weakly H0 -decreasing. Thus, for t0 R,
(t0 R+ ) there exists a constant 0 = 0 (t0 ) > 0 such that for
H0 (t0 , x0 ) < 0 the inequality
V (t0 , Y0 , w) b(t0 , H0 (t0 , Y0 ))
(5.5.12)
holds true.
Also, condition (3a) implies that there exists a 1 (0, H) such
that
a(H(t, x)) V (t, x, w)
for
H(t, x) 1 .
(5.5.13)
The fact that the measure H is continuous with respect to the

measure H0 implies that there exist a function CK and a constant 2 = 2 (t0 ) > 0 such that
H(t0 , Y0 ) (t0 , H0 (t0 , Y0 ))
for H0 (t0 , Y0 ) < 2 ,
(5.5.14)
where 2 is taken so that

(t0 , 2 ) < 1 .
(5.5.15)
Let (0, 0 ) and t0 R (t0 T ) be given. Since the

functions a K and b CK, given and t0 , one can choose
3 = 3 (t0 , ) > 0 so that
b(t0 , 3 ) < a().
(5.5.16)
We take (t0 ) = min (1 , 2 , 3 ). Conditions (5.5.12)(5.5.16)

imply that for H0 (t0 , Y0 ) < the inequalities
a(H(t0 , Y0 )) V (t0 , Y0 , w) b(t0 , H0 (t0 , Y0 )) < a()
are fullled. From this we get
H(t0 , Y0 ) < .
292
Let Y (t; t0 , Y0 ) = Y (t) be a solution of system (5.5.9) with the

initial conditions for which H0 (t0 , Y0 ) < . We shall make sure that
under conditions of Theorem 5.5.1 the estimate
H(t, Y (t)) < for all
t t0
holds true. Assume that there exists a t1 t0 such that

H(t1 , Y (t)) =
and H(t, Y (t)) < ,
t [t0 , t1 ),
for solution Y (t; t0 , Y0 ) with the initial conditions H0 (t0 , Y0 ) < .

8
Condition (4) and the fact that the matrix (Y,
w) is negative
semi-denite in the domain S imply that the roots i = i (Y, w) of
the equation
8
det [(Y,
w) E] = 0
satisfy the condition i (Y, w) 0, i = 1, 2, . . . , m, in the domain S.
Therefore,
8
D + V (t, Y, w) eT (Y,
w)e 0
and for all t [t0 , t1 ] the sequence of inequalities
a() = a(H(t1 , Y (t1 ))) V (t, Y, w) V (t0 , Y0 , w)
b(t0 , H0 (t0 , Y0 )) < a()
is satised.
The contradiction obtained disproves the assumption that t1
[t0 , +). Thus, system (5.5.7)(5.5.8) is (H0 , H)-stable.
Assertion (b) of Theorem 5.5.1 is proved in the same way. Besides,
it is taken into account that condition (3)(b) is satised and the
measure H is uniformly continuous with respect to the measure H0 ,
the value can be taken independent of t0 R (t0 R+ ). Hence
the uniform (H0 , H)-stability of system (5.5.7)(5.5.8)) follows.
Note that the construction of a suitable function (5.5.10) in terms
of the matrix function U (t, Y ) is essentially simplied because the
elements uij (t, Y ) can be associated with the world dynamics equations on a certain system level.
Takagi-Sugeno Impulsive Systems
5.6
5.6.1
293
Stability Analysis of Impulsive Takagi-Sugeno Systems

General results
Consider the impulsive fuzzy dynamic model of Takagi-Sugeno.

Given the properly dened input variables and membership functions, the T-S fuzzy rules for a multivariable system considered herein
are of the form:
Ri , i = 1, r : if z1 (t) is Mi1 and . . . and zn (t) is Min then
dx(t)
dt = Ai x(t), t = k ,
(5.6.1)
+ ) = B x(t), t = , k = 1, 2, . . . (k N),
x(t
i
k
+
x(t0 ) = x0 ,
where x(t) = (x1 , . . . , xn )T Rn is the state vector, z =
(z1 , . . . , zn )T Rn is the premise variable vector associated with
the systems states and inputs, x(t+ ) is the right value of x(t),
Ai Rnn , Bi Rnn are the system matrices, Mij () are the
membership functions of the fuzzy sets Mij and r is the number
of fuzzy rules. We suppose that Bi are non-singular matrices and
0 < 1 k+1 k 2 < .
We also suppose that at the moments of impulsive eects {k } the
solution x(t) is left continuous, i.e., x(k ) = x(k ).
The state equation can be dened as follows
r
dx(t)
=
i (z(t))Ai x(t),
dt
i=1
r

i (z(t))Bi x(t),
x(t+ ) =
i=1
+
x(t0 ) = x0 ,
t = k ,
t = k , k N,
where
i (z)
i (z) = r
i=1 i (z)
with i (z) =
n

j=1
Mij (zj ).
(5.6.2)
294
Clearly
i (z) = 1 and i (z) 0, i = 1, r. Next, without loss of
i=1
generality we take z = x.
The stability analysis in the sense of Lyapunov of zero solution
x = 0 of system (5.6.2) is the aim of this section. Before the main
results are obtained, the following assumption is made regarding the
T-S fuzzy system (5.6.2).
Assumption 5.6.1 There exist > 0 and > 0 such
that the functions i (x) for system (5.6.2) satisfy the inequality
Dx+ i (x) x1+ , i = 1, r.
In this assumption Dx+ i (x) denotes the upper Dini derivative of
i (x) i.e.
Dx+ i (x) = lim sup{(i (x(t + )) i (x(t)))/ : 0 }.
Remark 5.6.1 It should be noted that Assumption 5.6.1 admits
unique existence of solutions for system (5.6.2).
Let E denote the space of symmetric n n -matrices with
scalar
product (X, Y ) = tr(XY ) and corresponding norm X = (X, X),
where tr() denotes the trace of corresponding matrix. Let K E
be a cone of positive semi-denite symmetric matrices. Next we
will dene the following linear operators Fi X = ATi X + XAi ,
Bij X = BiT XBj , for all X E, i, j = 1, r.
Several theorems are rst proved to demonstrate that if certain
hypotheses are satised, the stability of the above nonlinear system
can be obtained using the direct Lyapunov method. It is shown
that stability conditions can be formulated in terms of linear matrix
inequalities.
Theorem 5.6.1 Under Assumption 5.6.1 the equilibrium state
x = 0 of fuzzy system (5.6.2) is asymptotically stable if for all
[1 , 2 ] there exists a common symmetric positive denite matrix X such that

p1

(1)k+1 (Fi )k k
1
(Bji + Bij ) I +
X < 0,
2
k!
i, j = 1, r,
k=1
(5.6.3)
295
(1)p (Fi )p X 0.
(5.6.4)
Before we prove of Theorem 5.6.1 we have the following remark.

Remark 5.6.2 It should be noted that
(1) (Fi )p X = Fi1 Fi2 . . . Fip X, where i1 = i, i2 = j, i1 , . . . , ip = 1, r;
(2) for i1 , . . . , ip = 1, r

r
p
i (x)Fi
X=
r

ip =1
i=1
ip (x) . . . i1 (x)Fi1 Fi2 . . . Fip X.
i1 =1
Proof Choose the Lyapunov function namely from class V0 ,

V (t, x) = xT P (t, x)x, where

r
i (x)Fi (tk )
X
e i=1
r

t i (x)Fi (ts)
P (t, x) =
e i=1
dsQ, for t (k , k+1 ],
+
X,
for t = k+1
.
Q and X are symmetric positive denite n n-matrices. Later we
K
shall show that P (t, x) > 0 in some neighborhood of the origin. First
let us consider the derivative of V (t, x) with respect to time. If t = k ,
then we have
Dt+ V
r

&
T
&
(t, x) (5.6.2) = x
i (x)(ATi P (t, x) + P (t, x)Ai )x
i=1
+ xT Dt+ P (t, x)x
r

i (x)Fi P (t, x)x
= xT
i=1
+ xT Dt+ P (t, x)x,
where
&
Dt+ P (t, x)&(5.6.2) =
r

i=1
i (x)Fi (tk )
r

i=1
Dx+ i (x)
dx
Fi (t k )
dt
296
t
i (x)Fi X
i=1
i (x)Fi (ts)
dx
Dx+ i (x) Fi (t
dt
i (x)Fi e
r

i=1
i (x)Fi (tk )
t
X
r
i=1
i (x)Fi (tk )
r
i (x)Fi (ts)
i=1
Dx+ i (x)
Dx+ i (x)
i=1
i (x)Fi P (t) e
r

i=1
i (x)Fi
r

i=1
i (x)Fi (ts)

ds Q
dx
Fi X(t k )
dt
dx
Fi (t s)ds Q Q
dt
r
i (x)Fi (tk )
i=1
i=1
t
t

s) dsQ Q
i=1
i=1
i=1
i=1
r
Dx+ i (x)
i=1
r
i=1
i (x)Fi (ts)
Dx+ i (x)
i=1
dx
Fi X(t k )
dt
dx
Fi (t s)ds Q Q.
dt
&
Hence, for the derivative Dt+ V (t, x)&(5.6.2) , we have the estimates:
Dt+ V
r
r

&
T
T
&
(t, x) (5.6.2) = x
i (x)Fi P (t, x)x x
i (x)Fi P (t, x)x

x Qx x
T
i=1
i=1
i (x)Fi (tk )
dx
Dx+ i (x) Fi X(t
i=1
t
+ xT
i=1
r
r

i=1
i (x)Fi (ts)
r

i=1
dt

k ) x

dx
Dx+ i (x) Fi (t s)ds Q x
dt
min (Q)x2
r

i=1
+ 2 e
i (x)Fi 2
r

i=1
1 dx 1
1 1
Dx+ i (x) Fi X 1 1 x2
dt
297
r

(x)Fi 2
2 i=1 i
2 e
r

i=1
1 dx 1
1 1
Dx+ i (x) Fi Q 1 1 x2 ,
dt
where min () > 0 is the minimal eigenvalue of corresponding matrix. Denote by a = max Ai . Then since Fi X ATi X +
i=1,r
XAi 2Ai X, we get Fi 2Ai 2a, i = 1, r. It is also

clear that
r
1 dx 1
1 1
i (x)Ai x ax.
1 1
dt
i=1
Hence the following inequality is fullled

&
Dt+ V (t, x)&(5.6.2) min (Q)x2
2
2a2
+ 2a 2 e
Dx+ i (x) X x3
i=1
+ 2a2 22 e2a2
Dx+ i (x) Q x3
i=1

min (Q) + 2a2 r2 e2a2 X + 2 Q x x2 .
&
Therefore Dt+ V (t, x)&(5.6.2) < 0 for all x from the ball x < R, where
R=
min (Q)

2
2a
2a r2 e 2 X + 2 Q
1/
Consider the dierence V = V (t+ , x(t+ )) V (t, x):

&
V &(5.6.2) = xT (t+ )P (t+ )x(t+ ) xT (t)P (t)x(t) = xT (t+ )Xx(t+ )

r
r

k
xT e
i=1
i (x(k ))Fi (k k1 )
i=1
i (x(k ))Fi (k s)
ds Q x
k1
r
r

j (x)i (x)BjT XBi x
j=1 i=1
2
+x
e
0
r

i=1
i (x)Fi y
dy Qx,
x e
T
r

i=1
i (x(k ))Fi (k k1 )
Xx
298
where y = k s.
Next we shall prove the following inequality
r

i=1
i (x)Fi (k k1 )

I
p1 (1)k+1

i=1 i (x)Fi
(k k1 )k
k!
k=1
(5.6.5)
X.
Let us choose an arbitrary element K = K and consider an

expansion in a Maclaurin series of the scalar function
r

i (x)Fi (k k1 )h
i=1
X X
(h) = tr e
p1 (1)k+1

i=1 i (x)Fi
(k k1 )k hk
k!
k=1

X
h 0, restricting p-order terms

(0)h + +
(h) = (0) +
(p1)
(p)
(0)hp1 ()hp
+
,
(p 1)!
p!
(0, h).
(p1)
(0) = =
(0) = 0, we get
Let h = 1, then since (0) =
(p)
()
, where
(1) =
p!

p

r
(p)
i (x)Fi (k k1 )
() = tr (1)p
i=1
r

i=1
i (x)Fi (k k1 )

X
r

i=1
i (x)Fi (k k1 )
give
Inequality (5.6.4) and positivity of operator e
(p)
estimate () 0. Thus (1) 0 for all K and therefore

inequality (5.6.5) is satised.
299
Consider the function

2
fx (2 ) = x
r

i=1
i (x)Fi y
dy Qx.
By Lagrange theorem we have
fx (2 ) = fx ()2 = xT 2 e
r

i=1
i (x)Fi
Qx,
where (0, 2 ) and therefore

r
fx (2 ) x ei=1
i (x)Fi 2
Q2 2 e2a2 Qx2 .
(5.6.6)
Inequalities (5.6.3), (5.6.5)(5.6.6) yield

&
V &
(5.6.2)
r

i1 =1
i1 (x) . . . ip1 (x)Qi1 i2 ...ip1 x
ip1 =1
+ 2 e2a2 Qx2
r
r

ip1 (x) . . . i1 (x)min (Qi1 i2 ...ip1 )x2

ip1 =1
i1 =1

+ 2 e2a2 Q x2 + 2 e2a2 Q x2 ,
where Qi1 i2 ...ip1 are positive denite matrices,
= min min (Qi1 i2 ...ip1 ),
i1 , . . . , ip1 = 1, r.
&
2a2
It is clear that V &(5.6.2) 0 if Q
e
(we can choose, for
2
e2a2 I).
example, Q =
2 n2
K
Next we shall show that P (t, x) > 0 for all t R i.e., V (t, x) is
a positive denite function. Since V (t, x) is decreasing function, we
have for x < R and t [k , k+1 ), k N
xT P (t, x)x xT (k+1 )P (k+1 , x(k+1 ))x(k+1 )
300

+
+
+
+
xT (k+1
)P (k+1
, x(k+1
))x(k+ ) min (X)x(k+1
)2 > 0.
&
As a result, we have V (t, x) > 0, Dt+ V (t, x)&(5.6.2) < 0 and
&
V &(5.6.2) 0 for all x < R.
Hence, all conditions of Theorem 4.9.3 hold. Therefore the zero
solution of impulsive Takagi-Sugeno fuzzy system (5.6.2) is asymptotically stable. This completes the proof of Theorem 5.6.1.
Let p be xed then we shall name the LMIs (5.6.3)(5.6.4) by
p-order stability conditions of system (5.6.2).
Next we shall formulate 2-nd order stability conditions of system
(5.6.2).
Corollary 5.6.1 Under Assumption 5.6.1 the equilibrium state
[1 , 2 ] there exists a common symmetric positive denite matrix
X such that
1 T
(B XBi + BiT XBj ) X + (ATj X + XAj ) < 0,
2 j
ATi ATj X + XAj Ai + ATj XAi + ATi XAj 0,
i, j = 1, r,
i, j = 1, r.
Suppose that fuzzy system (5.6.2) is such that A1 = A2 = =

An = A. Then we have the following 4-th order stability conditions.
Corollary 5.6.2 Under Assumption 5.6.1 the equilibrium state
[1 , 2 ] there exists a common symmetric positive denite matrix
X such that
1 T
(B XBi + BiT XBj ) X + (AT X + XA)
2 j
1
1
( (AT )2 X + 2AT XA + XA2 ) 2 + 3 (AT )3 X
2
6

+ 3((AT )2 XA + AT XA2 ) + XA3 < 0, i, j = 1, r,
(5.6.7)
(AT )4 X + 4((AT )3 XA + AT XA3 ) + 6(AT )2 XA2 + XA4 0.

(5.6.8)
301
Example 5.6.1 Let us consider the impulsive system (5.6.2) with

the following system matrices

2 0.5
A1 = A2 = A =
,
0.4 0.1

1.1 0.1
1.2 0.15
, B2 =
.
B1 =
0.2 0.2
0.1 0.3
Let the period of control action 1 = 2 = = 0.12 and suppose
that Assumption
5.6.1holds. Then it is easy to check that matrix

0.0756 0.0102
X=
satises LMIs (5.6.7)(5.6.8). Therefore by
0.0102 0.3261
Corollary 5.6.2 the zero solution x = 0 of the considered fuzzy system
is asymptotically stable.
Remark 5.6.3 It is easy to verify that 2-nd order stability conditions are not available to discuss stability analysis of the above fuzzy
system.
Remark 5.6.4 It should be noted that it is impossible to take
stability analysis of fuzzy system from Example 5.6.1 via paper by
Zhang, Li and Dan [1] because the discrete components (matrices
B1 and B2 ) are unstable and stability conditions from the paper
are neglected. Note that matrix A is also unstable. So, our stability
conditions are available to investigate the impulsive T-S fuzzy system
in which continuous and discrete components may be all unstable.
Let p 2 and Gi1 i2p1 be positive denite matrices. Consider
the following matrix equations for i1 , . . . , ip1 = 1, r

p1

1
(1)k+1 (Fi )k k
(Bji + Bji ) I +
X = Gi1 i2 ...ip1 . (5.6.9)
2
k!
k=1
Similarly to Theorem 5.6.1 we have the following result.

Theorem 5.6.2 Under Assumption 5.6.1 the equilibrium state
[1 , 2 ] there exists a common symmetric positive denite solution X of (5.6.9) such that the following inequality is fullled
e2a
(2a)p
<
,
p!
X
302
where a = max Ai , = min min (Gi1 i2 ...ip1 ) for i1 , . . . , ip1 =

i=1,r
1, r.
Next, we state the following assumption.
Assumption 5.6.2. There exist R0 > 0, 1 > 0, 2 > 0 and > 0
such that the functions i (x), i = 1, r, satisfy the inequality

1 x1+ , for x R0 ,
Dx+ i (x)
2 x1 , for x R0 .
Taking into account Assumption 5.6.2 we can establish the following.
Theorem 5.6.3 Let in Assumption 5.6.2 constants 1 , 2 , R0
be such that
1 2 <
2min (Q)
4a4 r 2 22 e4a2 ( X + 2 Q )2
and
1/
min (Q)

< R0 ,
2a2 r2 2 e2a2 X + 2 Q

1/
min (Q)

,
R0 <
2a2 r2 1 e2a2 X + 2 Q
where a = max Ai , Q is a symmetric positive denite nn- matrix

i=1,r
and X is a common symmetric positive denite matrix such that conditions (5.6.3)(5.6.4) of Theorem 5.6.1 hold. Then the zero solution
of impulsive fuzzy system (5.6.2) is globally asymptotically stable.
Proof Choose for a candidate the Lyapunov function from class
V0 , V (t, x) = xT P (t, x)x, where

r
i (x)Fi (tk )
X
e i=1
r

t i (x)Fi (ts)
P (t, x) =
e i=1
ds Q, for t (k , k+1 ],
+
,
X,
for t = k+1
303
where Q and X are symmetric positive denite n n-matrices. Let

us consider the derivative of V (t, x) with respect to time (notice that
V (t, x) is radially unbounded function). If t = k then we have two
cases:
(1) if x R0 then similar to the proof of Theorem 5.6.1 we get
&
&
Dt+ V (t, x)&
(5.6.2)
&
&
Clearly Dt+ V (t, x)&
(5.6.2)

min (Q) + 2a2 r2 e2a2 1 X

+2 Q x x2 .
< 0 by conditions of Theorem 5.6.3;
(2) if x R0 then by analogy we get

&
&
Dt+ V (t, x)&
(5.6.2)

min (Q) + 2a2 r2 e2a2 2 X

+2 Q x
x2 .
&
Clearly Dt+ V (t, x)&(5.6.2) < 0 by conditions of Theorem 5.6.3. Thus
&
< 0 for all x Rn .
we have showed that D + V (t, x)&
t
(5.6.2)
Similar to the proof of Theorem 5.6.1 we can show &(taking

into account the conditions of Theorem 5.6.3) that V &(5.6.2) =
V (t+ , x(t+ )) V (t, x) 0 and V (t, x) > 0. Therefore by Theorem 4.9.3 the zero solution of impulsive Takagi-Sugeno fuzzy system
(5.6.2) is globally asymptotically stable.
Remark 5.6.5 In spite of advantages of LMI method, the existence of solution that satises the sucient conditions is not guaranteed. This happens when the number of fuzzy rules is increased or
too many systems matrices are imposed.
Remark 5.6.6 The result of this section can be utilized for
chaotic, inverted pendulum, biological, electrical dynamical systems
etc. Moreover in practice it is enough to verify (using, for example
Matlab LMI toolbox) 2-nd order or 4-th order stability conditions.
304
5.6.2
Impulsive Fuzzy Control for Ecological

PreyPredator Community
It is well-known that control problem is an important task for mathematical theory of articial ecosystems. Impulsive control of such
systems is more favorable due to seasonal functioning of this type of
systems. Some problem of impulsive control for homotypical model
has been considered in the paper by Liu [1]. But for practice it is
suitable to consider models with fuzzy impulsive control because it
is almost impossible to accurately measure the biomass of one or
another biological species but possible to roughly estimate those.
Consider a Lotka-Volterra type preypredator model (with interspecic competition among preys) whose evolution is described by
the following equations
dN1
= N1 N1 N2 N12 ,
dt
dN2
= mN2 + sN1 N2 ,
dt
(5.6.10)
where N1 (t) is the biomass of preys, N2 (t) is the biomass of predators, is the growth rate of the preys, m is the death rate of the
predators, is the rate of the interspecic competition among preys,
is the per-head attack rate of the predators, and s is the eciency
of converting preys to predators.
Suppose that the ecosystem is controlled via regulation of the
number of species at certain xed moments of time (impulsive control) , 2, . . . , k, . . . and the regulation is reduced either to eliminate or fulminate the representatives of species. Taking into account
these assumptions we have to add the regulator equations to the
system of the evolution as
N1 = u1 (N1 , N2 ),
N2 = u2 (N1 , N2 ),
t = k,
k N,
where u1 , u2 are feedback functions, is a period of control action.

Under these assumptions the equations of closed controlled ecosys-
305
tem become
dN1
= N1 N1 N2 N12 ,
dt
dN2
= mN2 sN1 N2 , t = k,
dt
N1 = u1 (N1 , N2 ),
N2 = u2 (N1 , N2 ),
t = k,
(5.6.11)
k N.
Besides the trivial equilibrium state, equation (5.6.10) has also the
positive asymptotically stable states
N1 =
m
,
s
N2 =
s m
.
s 2
It is clear that if the number of preys is much greater than the

equilibrium ones then some amount of preys is eliminated and vice
versa. Analogous situation occures with the predators. Thus, the
impulsive fuzzy controls are designed regarding the rules:
if Ni Ni , then ui (N1 , N2 ) = i (Ni Ni ), i > 0,
i = 1, 2;
if Ni Ni , then ui (N1 , N2 ) = i (Ni Ni ), i (0, 1),
i = 1, 2,
where i are the fulmination rates, i are the elimination rates.
The fuzzy relation x y (x is much larger than y) can be
formalized using the following membership function
, if x > y,
(x, y) = 1 + 1/(x y)2
0,
if x y.
Next, we dene the variables of disturbance of motion x1 (t) =
N1 (t) N1 , x2 (t) = N2 (t) N2 . Then the equations for system
(5.6.11) become (using linearization):
m
m
dx1
=
x1 ,
dt
s
s
dx2
s m
=
x1 , t = k,
(5.6.12)
dt
x1 = u1 ,
x2 = u2 , t = k.
306
The Takagi-Sugeno fuzzy model (5.6.1) of system (5.6.12) is specied by the following four rules:
R1 : if N1 N1 and N2 N2 , then
dx(t)
dt = Ax(t), t = k,
x(t+ ) = B1 x,
x(t+
0 ) = x0 .
t = k,
dx(t)
dt = Ax(t), t = k,
+
x(t ) = B2 x,
x(t+
0 ) = x0 .
t = k,
dx(t)
dt = Ax(t), t = k,
x(t+ ) = B3 x,
x(t+
0 ) = x0 .
t = k,
dx(t)
dt = Ax(t), t = k,
+
x(t ) = B4 x,
x(t+
0 ) = x0 .
t = k,
It is obvious that Assumption 5.6.1 holds for membership function (x, y). Using Corollary 5.6.1 the stability analysis of nontrivial
equilibrium position for ecosystem is reduced to checking the existence of symmetric positive denite matrix X such that the following
LMIs hold true:
1 T
(B XBj + BjT XBi ) X + (AT X + XA) < 0, i, j = 1, 4,
2 i
(AT )2 X + 2AT XA + XA2 0.
(5.6.13)
307
Matrices A, B1 , B2 , B3 and B4 are as follows
m
m
s
s
A = s m
,
0

1 1
1 1
0
0
B1 =
, B2 =
,
0
1 2
0
1 2

1 1
1 1
0
0
, B4 =
.
B3 =
0
1 2
0
1 2
Next we consider the stability analysis of the obtained Takagi
Sugeno fuzzy model for ecosystems evolution with the following parameters: = 4, = 0.3, = 0.5, m = 1.2, s = 0.4, = 0.5 and
the parameters of impulsive control: 1 = 0.9, 2 = 0.5, 1 = 0.99,
2 = 0.6.

1.7427 1.8779
It is easy to check that matrix X =
satises
1.8779 8.2018
inequalities (5.6.13). Therefore by Corollary 5.6.1 the equilibrium
state of ecological system is asymptotically stable (see Figure 5.6).
5
x1
x2
Time
Figure 5.6. Evolution of x1 (t) and x2 (t) (stable result).
Let us change the parameters of impulsive control: 1 = 6,

2 = 4, 1 = 0.6, 2 = 0.2. In this case the solution of LMIs (5.6.13)
308
is infeasible and computer simulation gives an unstable result (see

Figure 5.7).
30
x1
x
25
20
15
10
5
0
5
10
Time
Figure 5.7. Evolution of x1 (t) and x2 (t) (unstable result).
Based on the well-known Lyapunov direct method, sucient conditions have been derived to guarantee the asymptotic stability and
globally asymptotic stability of the equilibrium point of impulsive
T-S fuzzy systems. It is shown that these sucient conditions are
expressed easily as a set of LMIs. It is also concluded that the
obtained stability conditions allow to investigate the impulsive TS fuzzy system in which continuous and discrete components may be
all unstable.
5.7
Notes
Most of the results of Chapter 5 are taken from Martynyuk [15, 18].
The model of robot interacting with dynamic environment is due to
DeLuca and Manes [1]. It should be noted that the importance of
studying the problem of stability of motion of a robot interacting with
a dynamic environment was discussed in contemporary literature.
The contents of Sections 5.15.3 are adapted from the paper by
Notes
309
Martynyuk and Chernienko [1] and Martynyuk [15]. See also Louartassi et al. [1] and NDoye et al. [1].
Section 5.4 is based on the results by Lila and Martynyuk [1] and
Martynyuk [17].
In Section 5.5 the model of (5.5.2) is taken from the monograph
by Forrester [1]. The model (5.5.6) is new. Theorem 5.5.1 is taken
from Martynuyuk [14]. Some other models of the world dynamics
are in Egorov et al. [1], Levashov et al. [1], etc.
Section 5.6 is adapted from Denysenko et al. [1].
For recent development and applications of the main results of
the book see, for example: Martynyuk [1]; Lakshmikanthan, Leela
and Martynyuk [3]; Martynyuk and Sun Zhen qi [1]; Lakshmikanthan
and Mohapatra [1]; Lakshmikanthan, Bhaskar, and Devi [1]; Lakshmikanthan, Leela, and Devi [1]; Lakshmikanthan, Leela, Drici, and
McRae [1]; Lakshmikanthan, Leela, and Vatsala [1]; Martynyuk and
Yu. A. Martynyuk-Chernienko [1]; Martynyuk, Chernetskaya, and
V. Martynyuk [1].
REFERENCES
Abramovich, J.
[1] On Gronwall and Wendor type inequalities. Proc. Amer. Math.
Soc. 87 (1983) 481486.
Achmedov, K. T., Yakubov, M.A. and Veisov, I.A.
[1] Some integral inequalities. Izvestiya Acad. Nauk Uz. SSR.
(1972) 1622.
Aleksandrov, A. Yu. and Platonov, A. V.
[1] Construction of Lyapunovs functions for a class of nonlinear
systems. Nonlinear Dynamics and Systems Theory 6(1) (2006)
1729.
[2] Comparison Method and Motion Stability of Nonlinear Systems.
Izdat. Sankt.-Petersb. Univer., SPb., 2012.
Alekseev, V. M.
[1] An estimate for perturbations of the solution of ordinary dierential equations. Vestnik Mosk. Univ. Ser.1, Math., Mekh. (2)
(1961) 2836.
Amann, H.
[1] Invariant sets and existence theorems for semilinear parabolic
and elliptic systems. J. Math. Anal. Appl. 65 (1978) 432467.
Amundson, N. R.
[1] Nonlinear problems in chemical reactor theory.
Proc. Amer. Math. Soc. Vol. III (1974) 5984.
SIAM-AMS
Aris, R.
[1] The Mathematical Theory of Diusion and Reaction in Permeable Catalysts. Clarendon Press, Oxford, 1975.

Systems & Control: Foundations & Applications, DOI 10.1007/978-3-319-27200-9
311
312
References
Bainov, D. D. and Hristova, S. G.

[1] On some integral inequalities for a scalar function of a vector
argument. Univ. Annual. Appl. Math. 17 (1981) 6770.
[2] Some generalizations of integral inequalities of Bellman-Bihari
type for functions of several variables. Rendiconti del Circulo
Math. di Palermo 30 (1981) 378384.
[3] Integral inequalities for functions with a number of variables.
Bull. Math. de la Soc. Sci. Math., Roumanie 26 (1982) 1520.
Bainov, D. D. and Simeonov, P. S.
[1] On some generalizations of Gronwall-Bellman integral inequality
for scalar functions of many variables. Rendiconti del Circulo
Matem. di Palermo 32 (1983) 2738.
Bainov, D. D. and Zahariev, A. I.
[1] A note on Bellman-Gronwall inequality. J. Math. Anal. Appl.
80 (1981) 147149.
Bainov, D. D., Lakshmikantham, V. and Simeonov, P. S.
[1] Theory of Impulsive Dierential Equations. World Scientic,
Singapore, 1989.
Barbashin, E. A.
[1] Introduction to the Theory of Stability. Nauka, Moscow, 1967.
Bellman, R.
[1] Stability Theory of Dierential Equations. McGraw-Hill, London, 1953.
[2] Vector Lyapunov functions. SIAM J. Control 1 (1963) 3234.
Beckenbach, E. F. and Bellman, R.
[1] Inequalities. Springer-Verlag, Berlin, 1961.
Beesack, P. R.
[1] Gronwall Inequalities. Cavlenton Math. Lecture Notes, 1975.
[2] On Lakshmikanthams comparison method for Gronwall inequalities. Ann. Pol. Math. 35 (1977) 187222.
Bernfeld, S. R. and Salvadori, L.
[1] Stability and Bifurcation. Trento, May 2530, 1980.
References
313
Bhatia, N. P. and Szeg

o, G. P.
[1] Stability Theory of Dynamic Systems. Springer, Berlin, 1970.
Bihary, I.
[1] A generalization of a lemma of Bellman and its applications to
uniqueness problems of dierential equations. Acta Math. Acad.
Sci. Hungar. 7 (1956) 7194.
Bondge , B. K. and Pachpatte, B. G.
[1] On Wendor type integral inequalities in n independent variables. Chinese J. Math. 7 (1979) 3746.
[2] On nonlinear integral inequalities of Wendor type. J. Math.
Anal. Appl. 70 (1979) 161169.
Bondge , B. K., Pachpatte, B. G. and Walter, W.
[1] On generalized Wendor type inequalities and their applications.
Nonlinear Analysis 4 (1980) 491495.
Bloom, F.
[1] Well-posed problems for integro-dierential equations in mechanics and electromagnetis theory. SIAM, Philadelphia, 1981.
Burton, T. A.
[1] Volterra Integral and Dierential Equations. Academic Press,
New York. 1983.
Casten, R. G. and Holland, C.J.
[1] Stability properties of solutions to systems of reaction-diusion
equations. SIAM J. Appl. Math. 33 (1977) 353364.
Cesari, L.
[1] Asymptotic Behavior and Stability Problems in Ordinary Differential Equations. Ergebn. & Math., 2nd Edition, Springer,
Berlin, 1963.
Chandra, J. and Fleishman, B. A.
[1] On a generalization of Gronwall-Bellman lemma in partially ordered Banach spaces. J. Math. Anal. Appl. 31 (1970) 668681.
314
References
Corduneanu, A.
[1] A note on the Gronwall inequality in two independent variables.
J. Integral Equations 4 (1982) 261276.
Corduneanu, C.
[1] The contribution of R. Conti to the comparison method in differential equations. Libertas Math. XXIX (2009) 113115.
Corduneanu, C. and Lakshmikantham, V.
[1] Equations with unbounded delay: A survey. Nonlinear Analysis
4 (1980) 831877.
De Luca, F. and Manes, C.
[1] Hybrid force/position control for robots on contact with dynamic
environments. Proc. Robot Control, SYROCO91, 1988, 377
382.
Demidovich, B. P.
[1] Lectures on Mathematical Theory of Stability. Nauka, Moscow,
1967.
Denysenko, V. S., Martynyuk, A. A. and Slynko, V. I.
[1] Stability analysis of impulsive Takagi-Sugeno systems. Int.
J. Innovative Computing, Information and Control 5 (10(A))
(2009) 31413155.
NDoye I., Zasadzinski M., Darouach M., Radhy, N-E., Bouaziz, A.
[1] Exponential stabilization of a class of nonlinear systems: A generalized Gronwall-Bellman lemma approach. Nonlinear Analysis. 74 (18) (2011) 73337341.
Egorov, V.A., Kallistov, Yu.N., Mitrofanov, V.B., Piontkovski, A.A.
[1] Mathematical Models of Sustainable Development. Gidrometeoizdat, Leningrad, 1980.
Forrester, J. W.
[1] World Dynamics. Nauka, Moscow, 1978.
Gamidov, Sh. G.
[1] Some integral inequalities for boundary value problems of dierential equations. Di. Eqns. 5 (1969) 463472.
References
315
Gronwall, T. H.
[1] Note on the derivatives with respect to a parameter of the solutions of a system of dierential equations. Ann. of Math. 20
(1919) 292296.
Gutowski, R. and Radziszewski, B.
[1] Asymptotic behavior and properties of solutions of a system of
nonlinear second order ordinary dierential equations describing
motion of mechanical systems. Arch. Mech. Stosow. 6 (22)
(1970) 675694.
Hahn, W.
[1] Stability of Motion. Springer-Verlag, Berlin, 1967.
Hale, J. K.
[1] Large diusivity and asymptotic behaviour in parabolic systems.
J. Math. Anal. Appl. 118 (1986) 455466.
Hale, J. K. and Kato, J.
[1] Phase space for retarded equations with innite delay. Funkc.
Ekvac. 21 (1978) 1141.
Hatvani, L.
[1] On the asymptotic stability by nondecrescent Lyapunov function. Nonlinear Analysis 8 (1984) 6777.
[2] On partial asymptotic stability and instability. Acta Sci. Math.
49 (1985) 157167.
Howes, F. A. and Whitaker, S.
[1] Asymptotic stability in the presence of convection. Nonlinear
Analysis 12 (1988) 14511459.
Hu, S., Lakshmikantham, V. and Rama Mohan Rao, M.
[1] Nonlinear variation of parameter formula for integro-dierential
equations of Volterra type. J. Math. Anal. Appl. 129 (1988)
223230.
Hu, S., Zhuang, W. and Khavinin, M.
[1] On the existence and uniqueness of nonlinear integro-dierential
equations. J. Math. Phys. Sci. 21 (1987) 93103.
316
References
Kato, J.
[1] Stability problem in functional dierential equations with innite
delay. Funkc. Ekvac. 21 (1978) 6380.
[2] Stability in functional dierential equations. In: Lecture Notes
in Math. 799, Springer-Verlag, New York, 1980, 252262.
Khavanin, M. and Lakshmikantham, V.
[1] The method of mixed monotony and rst order dierential systems. Nonlinear Analysis 10 (1986) 873877.
Khoroshun, A. S. and Martynyuk, A.A.
[1] Novel approach to absolute parametric stability of the uncertain
singularly perturbed systems. Communications in Appl. Anal.
17, no. 3 & 4 (2013) 439450.
Krasovski, N. N.
[1] Problems of the Theory of Stability of Motion. Stanford University Press, Stanford, Calif. 1963.
Ladde, G. S., Lakshmikantham, V. and Leela, S.
[1] A new technique in perturbation theory. Rocky Mountain J. 6
(1977) 133140.
Ladde, G. S., Lakshmikantham, V. and Vatsala, A. S.
[1] Monotone Iterative Techniques for Nonlinear Dierential Equations. Pitman, Boston, 1985.
Lakshmikantham, V.
[1] Several Lyapunov functions. Proc. Int. Sym. Nonlinear Oscillations. Kiev, Ukr. SSR, 1969.
[2] A variation of constants formula and Bellman-Gronwall-Reid
inequalities. J. Math. Anal. Appl. 41 (1973) 199204.
[3] Comparison results for reaction-diusion equations in Banach
spaces. Proc. of SAFA Conference. Bari, Italy, 1979, 121156.
[4] Some problems in integro-dierential equations of Volterra type.
J. Integral Eqns. 10 (1985) 137146.
Lakshmikantham, V., Bhaskar T. G., and Devi, J. V.
[1] Theory of Set Dierential Equations in Metric Spaces. Cambridge Scientic Publishers, Cambridge, 2006.
References
317
Lakshmikantham, V. and Leela, S.

[1] Dierential and Integral Inequalities. Vol. I and II, Academic
Press, New York, 1969.
[2] An Introduction to Nonlinear Dierential Equations in Abstract
Spaces. Pergamon Press, Oxford, 1981.
[3]
On perturbing Lyapunov functions. Math. Sys. Theory 10

(1976) 8590.
[4] Cone valued Lyapunov functions. Nonlinear Analysis 1 (1977)

215222.
[5] A technique in stability theory for delay dierential equations.
Nonlinear Analysis 3 (1979) 317323.
[6] A unied approach to stability theory for dierential equations
with innite delay. J. Integral. Eqns. 10 (1985) 147156.
[7] Reaction-diusion systems and vector Lyapunov functions. Di.
and Integral Eqns. 1 (1988) 4147.
Lakshmikantham, V., Leela, S., and Devi, J. V.
[1] Theory of Fractional Dynamic Systems. Cambridge Scientic
Publishers, Cambridge, 2009.
Lakshmikantham, V., Leela, S., Drici, Z., and McRae F. A.
[1] Theory of Causal Dierential Equations. Atlantic Press/World
Scientic, Amsterdam Paris, 2009.
Lakshmikantham, V., Leela, S. and Martynyuk A.A.
[1] Stability of Motion: Method of Comparison. Naukova Dumka,
Kiev, 1991.
[2] Stability Analisys of Nonlinear Systems. Marcel Dekker, Inc.,
New York, 1989.
[3] Practical Stability of Nonlinear Systems. World Scientic, Singapore, 1990.
Lakshmikantham, V., Leela, S. and Rama Mohan Rao, M.
[1] Integral and integro-dierential inequalities. Appl. Anal. 24
(1987) 157164.
Lakshmikantham, V., Leela, S. and Vatsala, A. S.
318
References
[1] Theory of Dierential Equations in Cone. Cambridge Scientic

Publishers, Cambridge, 2011.
Lakshmikantham, V. and Liu, Xinzhi
[1] Perturbing families of Lyapunov functions and stability in terms
of two measure. J. Math. Anal. Appl. 140 (1989) 107114.
Lakshmikantham, V. and Mohapatra, R. N.
[1] Theory of Fuzzy Dierential Equations and Inclusions. Taylor
and Francis, London, 2003.
Lakshmikantham, V. and Rama Mohan Rao, M.
[1] Integro-dierential equations and extension of Lyapunov method. J. Math. Anal. Appl. 30 (1970) 435447.
[2] Stability in variation for nonlinear integro-dierential equations.
Appl. Anal. 24 (1987) 165173.
Lakshmikantham, V. and Salvadori, L.
[1] On Massera type converse theorem in terms of two dierent
measures. Bull. U.M.I. 13 (1976) 293301.
Lakshmikanthm, V. and Sivasundaram, S.
[1] The method of upper and lower solutions and interval method
for rst order dierential equations. Appl. Math. and Comp.
23 (1987) 15.
Lakshmikantham, V. and Trigiante, D.
[1] Theory of Dierence Equations and Numerical Analysis. Academic Press, New York, 1987.
LaSalle, J. P.
[1] Stability theory for ordinary dierential equations.
Eqns. 4 (1968) 5765.
J. Di.
LaSalle, J. P. and Lefschetz, S.

[1] Stability by Lyapunovs Direct Method with Applications. Academic Press, New York, 1961.
Levashov, V. K.
[1] Sustainable Development of Society. Academia, Moscow, 2001.
References
319
Likhanski, V. V. and Napartovich, A. P.

[1] Radiation emitted by optically coupled lasers. Usp. Phys. Nauk
160(3) (1990) 101143.
Likhanski, V. V., Napartovich, A. P. and Sykharev, A. G.
[1] Phase locking of optically coupled and periodically pumped
lasers. Kvantovaya Electronika 22(1) (1995) 4748.
Lila, D. M. and Martynyuk, A. A.
[1] On stability of some solutions of synchronous generation of optically connected lasers with periodical pump. Neliniyni Kolyvannya 12 (4) (2009) 451460.
Liu, X.
[1] Progress in stability of impulsive systems with applications to
populations growth models. In: Advances in Stability Theory on
the End of 20th Century (Ed.: A. A. Martynyuk). Stability and
Control: Theory, Methods and Applications, Vol. 13, Taylor
and Francis, London, 2003, 321338.
Lord, M. E. and Mitchell, A.R.
[1] A new approach to the method of nonlinear variation of parameters. J. Appl. Math. Comp. 4 (1978) 95105.
Louartassi Y., El Houssine El Mazoudi, and Elalami, N.
[1] A new generalization of lemma of GronwallBellman. Appl.
Math. Sci. 6(13) (2012) 621628.
Malkin, I. G.
[1] Theory of Motion Stability. Nauka, Moscow, 1966.
Marachkov, M.
[1] On a theorem on stability. Bull. Soc. Phys. Math. Kazan (12)
(1940) 171174.
Martynyuk, A. A.
[1] Stability Analysis: Nonlinear Mechanics Equations. Gordon and
Breach Science Publishers, Amsterdam, 1995.
[2] Stability by Liapunovs Matrix Functions Method with Applications. Marcel Dekker, New York, 1998.
[3] Qualitative Methods in Nonlinear Dynamics: Novel Approaches
320
References
to Liapunovs Matrix Function.

2002.
Marcel Dekker, New York,
[4] Stability of Motion: The Role of Multicomponent Liapunov Functions. Cambridge Scientic Publishers, Cambridge, 2007.
[5] Hierarchical matrix Liapunov function. Dierential and Integral
Eqns. 2(4) (1989) 411417.
[6] Analysis of stability problems via matrix Liapunov functions. J.
Appl. Math. Stoch. Anal. 3(4) (1990) 209226.
[7] A theorem on polystability. Dokl. Akad. Nauk SSSR 318(4)
(1991) 808811.
[8] On the matrix comparison method in the theory of motion stability. Prikl. Mekh. 29(10) (1993) 116122.
[9]
On a generalization of Richardsons model of the arms race.

Dokl. Akad. Nauk 339(1) (1994) 1517.
[10] On the boundedness with respect to two measures of solutions

of replicator equations. Dokl. Akad. Nauk 353(2) (1997) 155
157.
[11]
On stability of motion of discontinuous dynamical systems.

Dokl. Akad. Nauk 397(3) (2004) 308312.
[12] Stability analysis by comparison technique. Nonlin. Anal. 62

(2005) 629641.
[13] Stability of dynamical systems in metric space. Nonlinear Dynamics and Systems Theory 5(2) (2005) 157167.
[14] The models of the world dynamics and sustainable development. Dokl. Nats. Akad. Nauk of Ukraine (7) (2010) 1621.
[15] Novel bounds for solution of nonlinear dierential equations.
Applied Math. 6 (2015) 182194.
[16] Criterion of asymptotic stability of nonlinear monotone systems
and its applications (survey). Int. Appl. Mech. 47 (5) (2011)
367.
[17]
Novel trends in the theory of direct Liapunov method. In:

Advances in Nonlinear Analysis: Theory Methods and Applications. (Eds.: S. Sivasundaram et al.) Cambridge Scientic
References
321
Publishers, Cambridge, 2009, 223234.

[18] Stability in the models of real world phenomena. Nonlinear
Dynamics and Systems Theory 11 (4) (2011) 7 52.
Martynyuk, A. A, Chernetskaya, L. N., and Martynyuk, V. A.
[1] Weakly Connected Nonlinear Systems: Boundedness and Stability of Motion. CRC Press/Taylor & Francis Group, LLC, Boca
Raton, 2013.
Martynyuk, A. A. and Chernienko A. N.
[1] On the theory of motion stability of a robot interacting with a
dynamic environment. Engineering Simulations. 17 (5) (2000)
605620.
Martynyuk, A. A. and Gutowski, R.
[1] Integral Inequalities and Stability of Motion. Naukova Dumka,
Kiev, 1979.
Martynyuk, A. A., Lakshmikantham, V. and Leela, S
[1] Stability of Motion: Method of Integral Inequalities. Naukova
Dumka, Kiev, 1989.
Martynyuk, A. A., Leela, S. and Mitropolsky, Yu. A.
[1] A survey of collected works of V. Lakshmikantham. Dierential
Equations 22 (1986) 552572.
Martynyuk, A. A. and Martynyuk-Chernienko Yu. A.
[1] Uncertain Dynamical Systems. Stability and Motion Control.
CRC Press/Taylor & Francis Group, LLC, Boca Raton, 2012.
Martynyuk, A. A. and Obolensky A. Yu.
[1] On stability of autonomous Wazewskiis systems. Dierential
Equations 8 (1980) 13921407.
[2] On the theory of one-side models in spaces with arbitrary cones.
J. of Appl. Math. Stoch. Anal. bf3 (2) (1990) 8597.
Martynyuk, A. A. and Slynko V. I.
[1] Solutions of the problem of constructing Liapunov matrix func-
322
References
tion for a class of large-scale systems. Nonlinear Dynamics and

Systems Theory 1 (2) (2001) 193203.
[2] Stability of a nonlinear impulsive system. Prikl. Mech. 40 (2)
(2004) 134144.
Martynyuk, A. A. and Sun Zhen qi
[1] Qualitative Analysis of Nonlinear Systems with Small Parameter. Chinese Academy of Sciences Publishing Company, Beijing,
2006.
Matrosov, V. M.
[1] On the stability of motion. Prikl. Mat. Mekh. 26 (1962) 885
895.
[2] On the theory of stability of motion. Prikl. Mat. Mekh. 26
(1962) 9921002.
[3] Vector Lyapunov Functions Method: Nonlinear Analysis of Dynamical Properties. Fizmatlit, Moscow, 2000.
Mazko, A. G.
[1] Stability and comparison of states of dynamical systems relative
to variable cone. Ukr. Matem. Zhurn. 57(2) (2005) 198213.
Meadows, D. L. and Meadows, D. H.
[1] Toward Global Equilibrium. Wright-Allen Press, Cambridge,
1973.
Moore, Jane
[1] A new concept of stability, M0 -stability. J. Math. Anal. Appl.
112 (1985) 113.
Moore, R. E.
[1] Methods and Applications of Interval Analysis. SIAM Studies in
Appl. Math., Philadelphia, 1979.
[2] Set valued extensions and integral inequalities. J. Integral Eqns.
5 (1983) 187198.
Movchan, A. A.
[1] Stability of processes with respect to two measures. Prikl. Mat.
Mekh. 24 (1960) 9881001.
References
323
Muldowney, J. S. and Wong, J. S. W.

[1] Bounds for solutions of nonlinear integro-dierential equations.
J. Math. Anal. Appl. 23 (1968) 487499.
Oziraner, A. S.
[1] On the stability of motion in linear approximation. Prikl. Mat.
Mekh. 41(3) (1977) 413421.
Pachpatte, D. B.
[1] Explicit estimation on integral inequalities with time scale. J. of
Inequalities in Pure and Applied Mathematics 7(4) (2006) Art.
143, 8 p.
Piontkovskii, A. A., and Rutkovskaya, L. D.
[1] Investigations of certain problem of the stability theory via the
vector Lyapunov function method. Automatika i Telemekhanika
28(10) (1967) 2331.
Rama Mohan Rao, M.
[1] A note on an integral inequality. Ind. Math. Soc. 27 (1963)
6571.
Rama Mohan Rao, M. and Srinivas, P.
[1] Asymptotic behavior of solutions of Volterra integro-dierential
equations. Proc. Amer. Math. Soc. 94 (1985) 5560.
Redheer, R. and Walter, W.
[1] Solution of the stability problem for a class of generalized
Volterra prey-predator systems. J. Di. Eqns. 52 (1984) 245
263.
Rouche, N., Habets, P. and Laloy, M.
[1] Stability Theory by Lyapunovs Direct Method. Springer Verlag,
New York, 1977.
Rozo, M.
[1] Nonlinear Oscillations and Stability Theory. Nauka, Moscow,
1971.
324
References
Salvadori, L.
[1] Sulla stabilita del movimento. Matematiche 24 (1969) 218239.
[2] Famiglie ad un parametro di funzioni di Lyapunov nello studio
della stabilita. Symp. Math. 6 (1971) 309330.
[3] Sul problema della stabilita asintotica. Rondiconti dell. Accad.
Naz. Lineci 53 (1972) 3538.
[4] Sulleslensione ai sistemi dissipativi del criterio di stabilita del
Routh. Ricerche. Mat. 151 (1966) 162167.
Samoilenko, A. M. and Ronto, V. I.
[1] Numerical-Analitical Methods of Investigation of Periodic Solutions. Vussh. Shkola, Kiev, 1976.
Shendge, G. R.
[1] A new approach to the stability theory of functional dierential
systems. J. Math. Anal. Appl. 95 (1983) 319334.
Siljak,
D. D.
[1] Large scale dynamic systems. North-Holland, New York, 1978.
[2] Competitive economic systems: Stability, decomposition and aggregation. IEEE Trans. AE-21 (1976) 149160.
Spanier E. H.
[1] Algebraic Topology. Springer, New York, 1966.
Strauss, A. and Yorke, J.
[1] Perturbation theorems for ordinary dierential equations. J.
Di. Eqns. 3 (1967) 1530.
Turinici, M.
[1] Abstract Gronwall-Bellman inequalities in ordered metrizable
uniform spaces. J. Integral Eqns. 6 (1984) 105117.
[2]
Abstract comparison principles and multivariate GronwallBellman inequalities. J. Math. Anal. Appl. 117 (1986) 100
127.
Viswanatham, B.
[1] A generalization of Bellmans lemma. Proc. Amer. Math. Soc.
14 (1963) 1518.
325
References
Voronov, A. A. and Matrosov, V. M. (Eds.)

[1] Method of Vector Lyapunov Functions in the Theory of Stability.
Nauka, Moscow, 1987.
Vujicic, V. A. and Martynyuk, A. A.
[1] Some Problems of Mechanics of Nonautonomous Systems. Mathematical Institue of SANU, BeogradKiev, 1992.
Walter, W.
[1] Dierential and Integral Inequalities.
York, 1970.
Springer-Verlag, New
Young, E. C.
[1] Gronwalls inequality in n independent variables. Proc. Amer.
Math. Soc. 41 (1973) 241244.
Yoshizawa, T.
[1] Stability Theory by Liapunovs Second Method. The Mathematical Society of Japan, Tokyo, 1966.
Zhang, X., Li, D. and Dan, Y.
[1] Impulsive control of TakagiSugeno fuzzy systems. Fourth Int.
Conf. on Fuzzy Systems and Knowledge Discovery. Vol. 1, 2007,
321325.
Zhuang, Wan
[1] Existence and uniqueness of solutions of nonlinear integrodierential equations of Volterra type in a Banach space. Appl.
Anal. 22 (1986) 157166.
Zubov, V. I.
[1] The Methods of Liapunov and Their Applications.
Leningr. Univer., Leningrad, 1957.
Izdat.
INDEX
h0 is
asymptotically ner than h, 143
ner than h, 142
uniformly ner than h, 142
A technique in perturbation theory,
209
Ane system, stabilization, 270
Alekseevs formula, 70, 113
Antidierence operator, 47
Asymptotic equilibrium, 88
(h0 , h)-equistable, 141

stable
(h0 , h)-asymptotically, 141
(h0 , h)-equi asymptotically,
141
(h0 , h)-equi Lagrange, 157
(h0 , h)-quasi-equi-asymptotically, 141
(h0 , h)-quasi-uniform asymptotically, 141
(h0 , h)-uniformly, 141
(h0 , h)-uniformly asymptotically, 141
(h0 , h)-uniform
Lagrange,
157
(h0 , h)-unstable, 141
Cauchy criterion for convergence, 84

Comparison result in terms of Lyapunov function V , 137
Cone, 29
Control system, 187, 198
Coupled quasi solutions, 97
Elliptic operator L, 249
extremal, 97
Equation
dierence, 47
Darboux integral
dierential
lower, 53
with impulsive eect, 59
upper, 53
with innite delay, 216
Dierential system, 140
functional
dierential, 217
bounded
integro-dierential,
106
(h0 , h)-equi ultimately, 157
perturbed,
106
(h0 , h)-quasi uniform ultimawith nite memory, 228
tely, 157
Estimates
of solutions, 73
(h0 , h)-quasi-equi-ultimately,
156
(h0 , h)-uniform ultimately,
157
(h0 , h)-uniformly, 156
(h0 , h)-equibounded, 156
Forrester model of world dynamics,

285
Function
interval, 52

Systems & Control: Foundations & Applications, DOI 10.1007/978-3-319-27200-9
327
328
Index
quasimonotone nondecreasing, Lyapunovs method to integro26

differential equations of
in u with respect to K, 139
Volterra type, 231
relative to K, 30
Method
Fundamental matrix solution, 70
of mixed monotony, 100
General nonlinear model of world
of upper and lower solutions, 93
dynamics, 288
MOcondition, 181
Generalized inner products, 75
Monotone iterative technique, 96
Global existence of solutions, 83
Norm
Identity matrix, 70
Euclidian, 74
Inequality
logarithmic, 74, 131
Bihari-type, 16
One parameter family of Lyapunov
Bihari, discrete version, 48
functions, 162
dierential, 26
Optically coupled lasers, 276
in arbitrary cones, 29
for piecewise continuous funcProblem of continuity and dierentions, 56
tiability of solutions, 108
Gronwall-type, 2
Process of mixed monotonization,
integral, 31
100
integral
general, 35
Quasimonotonicity, 138
interval-valued, 52
Volterra, 32
Reaction-diusion equations, 249
integro-dierential, 38
Robot motion stable
pseudo-linear, 34, 35
with limited action, 257
reaction-diusion, 61
with limited, on the average, action, 257
Wendor-type, 12
with several independent variwith limited, on the quadratic
ables, 24
average, action, 257
Inner product, 75
Set
Interval analysis, 103
M0 -equistable, 173
Interval function, 103
{0} asymptotically invariant,
142
Lyapunov function
h-asymptotically
decrescent, Set with respect to the system
M0 -equi asymptotically stable,
143
172
h-decrescent, 143
M0 -equistable, 171
h-positive denite, 143
M0 -invariant, 171
h-weakly decrescent, 143
M0 -quasi
equiasymptotically
matrix-valued, 191
stable, 171
vector, 177
329
Index
M0 -quasi uniform asymptotically stable, 172
M0 -uniformly asymptotically
stable, 172
M0 -uniformly stable, 171
Several Lyapunov functions, 138
Solution
asymptotically stable in a nonnegative cone, 181
extremal, 27
left maximal, 218
maximal, 27
minimal, 27
right maximal, 218
stable
equi-asymptotically, 89
exponentially asymptotically,
92
quasi-equi asymptotically, 89
quasi-uniformly asymptotically, 89, 205
uniformly asymptotically, 89
uniformly, 89, 205
trivial, 89, 202, 205
h0 -conditionally equistable,
202
equi-stable, 89, 143
Space B, 216
admissible, 216
Stability
in variation, 116
of asymptotically invariant set,
142
of conditionally invariant set,
142
properties in variation, 119
of the invariant set, 142
of the prescribed motion, 142
of the trivial solution, 142
partial of the trivial solution,
142
Sustainable development with re-
spect to two measures, 289

Synchronization of motions, 273
System
(h0 , h)-totally stable, 204
impulsive Takagi-Sugeno, 293
integro-dierential, 234
linear, 117
perturbed dierential, 202
Theorem Ascoli-Arzela, 88
Theory of dierence inequalities, 46
Uncoupled linear initial value problem, 101
Uniform asymptotic stability of T periodic solution, 277
Variation of parameters
formula for dierence equations, 123
method, 69
nonlinear, 68
when the initial data are dierent, 72
World dynamics equations, 286
Zero solution, 119, 221
equi-stable in (X, Y ), 221

Stability Analysis of Nonlinear Systems 2016

Transféré par

Informations du document

Copyright

Formats disponibles

Partager ce document

Partager ou intégrer le document

Options de partage

Avez-vous trouvé ce document utile ?

Ce contenu est-il inapproprié ?

Droits d'auteur :

Formats disponibles

Stability Analysis of Nonlinear Systems 2016

Transféré par

Droits d'auteur :

Formats disponibles

Systems & Control: Foundations & Applications

Systems & Control: Foundations & Applications

More information about this series at http://www.springer.com/series/4895

Vangipuram Lakshmikantham (1924 2012)

PREFACE TO THE SECOND EDITION

The first edition of this book was published in 1989 by Marcel

Preface to the Second Edition

of the investigations beyond which there are open problems in the

Its aim is to present a systematic account of the recent developments,

Preface to the Second Edition

2 Variation of Parameters and Monotone Technique

Method of Mixed Monotony .

3 Stability of Motion in Terms of Two Measures

4.9 Impulse Dierential Equations . . . . . . . . . . . . . 241

This chapter introduces the theory of inequalities of various types

Springer International Publishing Switzerland 2015

in Section 1.9 where among others the discrete versions of Gronwall

Proof If c > 0, it follows from (1.1.1) that

which, upon integration yields

This, together with (1.1.1) implies (1.1.2).

q (t) = [m (t) v(t)m(t)] exp

Hence it follows that

and this implies the stated conclusion proving the lemma.

Theorem 1.1.2 Let m, v, h C[R+ , R+ ] and suppose that

Proof To prove (1.1.4), set p(t) =

v(s)m(s) ds, so that p(t0 ) = 0

Since m(t) h(t) + p(t), we get

v(s) ds), we see that q(t0 ) = 0 and

and this gives (1.1.4) immediately.

which, in view of (1.1.3), yields the linear dierential inequality

and the estimate (1.1.5) follows.

v() d ds = h(t) h(t0 ) exp

For, setting w(t) =

Hence, by Theorem 1.1.1, we have w(t) exp

Example 1.1.2 Let m C[R+ , R+ ] and

where A, B 0 and c > 0. Then,

Example 1.1.3 Let m C[R+ , R+ ] and for B > 0

m(t) exp((B + 1)t Bt0 ) ,

We shall now consider integral inequalities with separable kernels

v(s)m(s) ds, so that p(t0 ) = 0 and

p (t) = v(t)m(t). Since m(t) h(t) + q(t)p(t), we obtain

which yields (1.1.8).

[exp(r(t s))](Am(s) + B) ds,

where r, A, B 0 and r A > 0. Then, for t t0 ,

where G(t) = sup gi (t) and V (t) =

Corollary 1.1.2 Let m, h, gi , 1 , 2 C[R+ , R+ ], i = 1, 2, . . . , n

for t [t0 , T ], where ci 0, ti [t0 , T ] and t0 t1 t2 . . .

K2 (t) = 2 (t) + 1 (t)

gi (s)K2 (s) ds < 1.

K(t, s)m(s) ds,

Kt (t, s)h(s) ds.

Proof Setting p(t) =

K(t, s)m(s) ds, we have p(t0 ) = 0 and

p (t) = K(t, t)m(t) +

Kt (t, s)m(s) ds.

A(t, t0 )p(t) + (t, t0 ),

which yields the estimate (1.1.10).

where c1 , c2 , c3 0. Then, for t t0 ,

m(t) c1 + (t) c2 exp c3

g(t, s, )m( ) d ds,

Finally, we shall prove a result on integral inequality that contains

and p(t0 ) = m(t0 ) +

and the result follows.

Most of Gronwalltype inequalities can be extended to multivariate