Lyapunov Functionals and Stability of Stochastic

Functional Differential Equations

Leonid Shaikhet

Lyapunov Functionals
and Stability of
Stochastic Functional
Differential Equations
Leonid Shaikhet
Department of Higher Mathematics
Donetsk State University of Management
Donetsk, Ukraine

ISBN 978-3-319-00100-5 ISBN 978-3-319-00101-2 (eBook)

DOI 10.1007/978-3-319-00101-2
Springer Cham Heidelberg New York Dordrecht London

Library of Congress Control Number: 2013935832

© Springer International Publishing Switzerland 2013

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. Exempted from this legal reservation are brief excerpts in connection
with reviews or scholarly analysis or material supplied specifically for the purpose of being entered
and executed on a computer system, for exclusive use by the purchaser of the work. Duplication of
this publication or parts thereof is permitted only under the provisions of the Copyright Law of the
Publisher’s location, in its current version, and permission for use must always be obtained from Springer.
Permissions for use may be obtained through RightsLink at the Copyright Clearance Center. Violations
are liable to prosecution under the respective Copyright Law.
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.
While the advice and information in this book are believed to be true and accurate at the date of pub-
lication, neither the authors nor the editors nor the publisher can accept any legal responsibility for any
errors or omissions that may be made. The publisher makes no warranty, express or implied, with respect
to the material contained herein.

Printed on acid-free paper

Springer is part of Springer Science+Business Media (www.springer.com)

Preface

This book deals with stability of stochastic functional differential equations and
continues and complements the previous book of the author, “Lyapunov Functionals
and Stability of Stochastic Difference Equations” [278].
Functional differential equations (also called hereditary systems [9], or systems
with aftereffect [140], or equations with memory [11], or equations with deviating
arguments [66, 167], or equations with delays [10, 24–29, 54, 93, 97, 226], or equa-
tions with time lag [65, 72, 103], or retarded differential equations [52, 78], or differ-
ential difference equations [22]) are infinite-dimensional ones [88, 181], contrary to
ordinary differential equations, and describe the processes whose behavior depends
not only on their present state but also on their past history [104–106, 131–133].
Systems of such type are widely used to model processes in physics, mechanics,
automatic regulation, economy, finance, biology, ecology, sociology, medicine, etc.
(see, e.g., [15, 18, 20, 24–26, 32, 36–41, 57, 58, 71, 91, 92, 98, 160, 161, 165, 177,
186, 189, 192, 199, 212, 213, 218, 219, 224, 245, 253, 257, 286, 287, 293, 303]).
The first mathematical models with functional differential equations have been
studied during the 18th century (L. Euler, J. Bernoulli, M.J. Condorcet, J. Lagrange,
P. Laplace). In the beginning of 20th century the development of the delay systems
study started from the works by Vito Volterra [302], linked to viscoelasticity and
ecology. This pioneer work was continued by Tsypkin [298–300], Myshkis [220,
221], Kac and Krasovskii [120, 156–158], Elsgoltz [66], Razumikhin [239–242],
and many other (see, for example, [5, 9, 22, 43, 59, 62, 67, 100, 103–106, 111, 126,
132, 133, 140, 194, 206, 308, 313]).
An important direction in the study of hereditary systems is their stability [3, 12,
55, 56, 88, 96, 103, 125, 134, 135, 181, 182, 191, 195–197, 204–207, 211, 215–
217, 231, 236, 258–260, 301]. As it was proposed by Krasovskii in the 1950s [156–
159], a stability condition for differential equation with delays can be obtained using
an appropriate Lyapunov functional. The construction of different Lyapunov func-
tionals for one differential equation with delay allows one to get different stability
conditions for the solution of this equation. The method of Lyapunov–Krasovskii
functionals is very popular and developing until now [78, 79]. However, the con-
struction of each Lyapunov functional required a unique work from its author. In

v
vi Preface

1975, Shaikhet [261] introduced a parametric family of Lyapunov functionals, so

that an infinite number of Lyapunov functionals were used simultaneously. This ap-
proach allowed one to get different stability conditions for the considered equation
using only one Lyapunov functional.
During twenty last years, a general method of constructing Lyapunov func-
tionals was proposed and developed by Kolmanovskii and Shaikhet for stochastic
functional differential equations, for stochastic difference equations with discrete
time and continuous time, and for partial differential equations [48, 130, 131, 136–
139, 145–149, 265, 269–272, 278, 280, 281]. This method was successfully applied
to stability research of some mathematical models in mechanics, biology, ecology,
etc. (see, for instance, [27, 36–39, 267, 273, 276]). Nevertheless, it should be noted
that the stability theory for stochastic hereditary system dynamically develops and
has yet a number of unsolved problems [277, 279].
Usually the books devoted to the stability theory for functional differential equa-
tions do not concern numerical determination of stability domains or the behavior
of the solutions. In spite of that, this book, along with the modern theoretical results,
includes also many numerical investigations showing both the stability domains and
structure of the solutions. It offers a certain amount of analytical mathematics, prac-
tical numerical procedures, and actual implementations of these approaches.
In this book, consisting of twelve chapters, a general method of construction of
Lyapunov functionals for stochastic functional differential equations is expounded.
Introductory Chap. 1 presents general classification and some peculiarities of
functional differential equations, some properties of their solutions, the method of
steps and the characteristic equation for retarded differential equations, the depen-
dence of solution stability on small delay in equation, and the Routh–Hurwitz sta-
bility conditions for systems without delay. This section covers some theoretical
backgrounds of the differential equations used in the book with concentration on
mathematical rigor.
In Chap. 2 short introduction to stochastic functional differential equations is
presented, in particular, the definition of the Wiener process and its numerical sim-
ulation, the Itô integral, and the Itô formula. Different definitions of stability for
stochastic functional differential equations are also considered, basic Lyapunov-
type stability theorems, and description of the procedure of constructing Lyapunov
functionals for stability investigation. In this section some useful statements, some
useful inequalities, and some unsolved problems are also included.
In Chap. 3 the procedure of constructing Lyapunov functionals is used to obtain
conditions for stability of scalar stochastic linear delay differential equations with
constant and variable coefficients and with constant and variable delays. It is shown
that different ways of constructing Lyapunov functionals for a given equation al-
low us to get different conditions for asymptotic mean-square stability of the zero
solution of this equation.
In Chap. 4 the procedure of constructing Lyapunov functionals is demonstrated
for stability investigation of stochastic linear systems of two equations with constant
and distributed delays and with constant and variable coefficients.
In Chap. 5 the stability of the zero solution and positive equilibrium points for
nonlinear systems is studied. In particular, differential equations with nonlinearities
Preface vii

in deterministic and stochastic parts and with fractional nonlinearity are consid-
ered. It is shown that investigation of stability in probability for nonlinear systems
with the level of nonlinearity higher than one can be reduced to investigation of
the asymptotic mean-square stability of the linear part of the considered nonlinear
system.
In Chap. 6 the general method of construction of Lyapunov functionals is used
to get the asymptotic mean-square stability conditions for stochastic linear differen-
tial equations with constant delay, with distributed delay, and with variable bounded
and unbounded delays. Sufficient stability conditions are formulated in terms of the
existence of positive definite solutions of some matrix Riccati equations. Using the
procedure of constructing Lyapunov functionals, it is shown that for one stochastic
linear differential equation, several different matrix Riccati equations can be ob-
tained that allow one to get different stability conditions.
In Chap. 7 sufficient conditions for asymptotic mean-square stability of the so-
lutions of stochastic differential equations with delay and Markovian switching are
obtained. Taking into account that it is difficult enough in each case to get analytical
stability conditions, a numerical procedure for investigation of stability of stochastic
systems with Markovian switching is considered. This procedure can be used in the
cases where analytical conditions of stability are absent. Some examples of using
the proposed numerical procedure are considered. Results of the calculations are
presented by a lot of figures.
Chapter 8 is devoted to the classical problem of stabilization of the controlled
inverted pendulum. The problem of stabilization for the mathematical model of
the controlled inverted pendulum during many years is very popular among the re-
searchers. Unlike the classical way of stabilization in which the stabilized control is
a linear combination of the states and velocities of the pendulum, here another way
of stabilization is proposed. It is supposed that only the trajectory of the pendulum
can be observed and stabilized control depends on the whole trajectory of the pen-
dulum. Via the general method of construction of Lyapunov functionals, sufficient
conditions for stabilization by stochastic perturbations are obtained, and nonzero
steady-state solutions are investigated.
In Chap. 9 the well-known Nicholson blowflies equation with stochastic pertur-
bations is considered. Sufficient conditions for stability in probability of the trivial
and positive equilibrium points of this nonlinear differential equation with delay are
obtained.
In Chap. 10 the mathematical model of the type of predator–prey with afteref-
fect and stochastic perturbations is considered. Sufficient conditions for stability in
probability of the positive equilibrium point of the considered nonlinear system are
obtained.
Chapter 11 deals with a mathematical model of the spread of infectious diseases,
the so-called SIR epidemic model. Sufficient conditions for stability in probability
of two equilibrium points of the SIR epidemic model with distributed delays and
stochastic perturbations are obtained.
In Chap. 12 mathematical models are considered that describe human behaviors
related to some addictions: consumption of alcohol and obesity. The existence of
viii Preface

positive equilibrium points for these models are shown, and sufficient conditions
for stability in probability of these equilibrium points are obtained.
The bibliography at the end of the book does not pretend to be complete and
includes some of the author’s publications [261–279], his publications jointly with
coauthors [27, 36–39, 48, 64, 77, 136–148, 155, 197, 232, 246, 280, 281], and the
literature used by the author during preparation of this book.
The book is addressed both to experts in stability theory and to a wider audi-
ence of professionals and students in pure and computational mathematics, physics,
engineering, biology, and so on.
The book is mostly based on the results obtained by the author independently or
jointly with coauthors, in particular, with the friend and colleague V. Kolmanovskii,
with whom the author is glad and happy to collaborate for more than 30 years.
Taking into account that the possibilities for further improvement and develop-
ment are endless, the author will appreciate receiving useful remarks, comments,
and suggestions.
Donetsk, Ukraine Leonid Shaikhet
Contents

1 Short Introduction to Stability Theory of Deterministic Functional

Differential Equations . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1 Some Peculiarities of Functional Differential Equations . . . . . . 1
1.1.1 Description of Functional Differential Equations . . . . . . 1
1.1.2 Reducing to Ordinary Differential Equations . . . . . . . . 2
1.2 Method of Steps for Retarded Functional Differential Equations . . 3
1.3 Characteristic Equation for Differential Equation with Discrete
Delays . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.4 The Influence of Small Delays on Stability . . . . . . . . . . . . . 21
1.5 Routh–Hurwitz Conditions . . . . . . . . . . . . . . . . . . . . . 25
2 Stochastic Functional Differential Equations and Procedure
of Constructing Lyapunov Functionals . . . . . . . . . . . . . . . . . 29
2.1 Short Introduction to Stochastic Functional Differential Equations . 29
2.1.1 Wiener Process and Its Numerical Simulation . . . . . . . 29
2.1.2 Itô Integral, Itô Stochastic Differential Equation, and Itô
Formula . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.2 Stability of Stochastic Functional Differential Equations . . . . . . 34
2.2.1 Definitions of Stability and Basic Lyapunov-Type Theorems 34
2.2.2 Formal Procedure of Constructing Lyapunov Functionals . 39
2.2.3 Auxiliary Lyapunov-Type Theorem . . . . . . . . . . . . . 40
2.3 Some Useful Statements . . . . . . . . . . . . . . . . . . . . . . . 43
2.3.1 Linear Stochastic Differential Equation . . . . . . . . . . . 43
2.3.2 System of Two Linear Stochastic Differential Equations . . 47
2.3.3 Some Useful Inequalities . . . . . . . . . . . . . . . . . . 50
2.4 Some Unsolved Problems . . . . . . . . . . . . . . . . . . . . . . 51
2.4.1 Problem 1 . . . . . . . . . . . . . . . . . . . . . . . . . . 51
2.4.2 Problem 2 . . . . . . . . . . . . . . . . . . . . . . . . . . 52
3 Stability of Linear Scalar Equations . . . . . . . . . . . . . . . . . . 53
3.1 Linear Stochastic Differential Equation of Neutral Type . . . . . . 53

ix
x Contents

3.1.1 The First Way of Constructing a Lyapunov Functional . . . 53

3.1.2 The Second Way of Constructing a Lyapunov Functional . 56
3.1.3 Some Particular Cases . . . . . . . . . . . . . . . . . . . . 58
3.2 Linear Differential Equation with Two Delays in Deterministic Part 62
3.2.1 The First Way of Constructing a Lyapunov Functional . . . 62
3.2.2 The Second Way of Constructing a Lyapunov Functional . 63
3.2.3 The Third Way of Constructing a Lyapunov Functional . . 64
3.2.4 The Fourth Way of Constructing a Lyapunov Functional . . 65
3.2.5 One Generalization for Equation with n Delays . . . . . . 70
3.3 Linear Differential Equation of nth Order . . . . . . . . . . . . . . 72
3.3.1 Case n > 1 . . . . . . . . . . . . . . . . . . . . . . . . . 72
3.3.2 Some Particular Cases . . . . . . . . . . . . . . . . . . . . 77
3.4 Nonautonomous Systems . . . . . . . . . . . . . . . . . . . . . . 85
3.4.1 Equations with Variable Delays . . . . . . . . . . . . . . . 85
3.4.2 Equations with Variable Coefficients . . . . . . . . . . . . 91
4 Stability of Linear Systems of Two Equations . . . . . . . . . . . . . 97
4.1 Linear Systems of Two Equations with Constant Delays . . . . . . 97
4.2 Linear Systems of Two Equations with Distributed Delays . . . . . 102
4.3 Linear Systems of Two Equations with Variable Coefficients . . . . 107
5 Stability of Systems with Nonlinearities . . . . . . . . . . . . . . . . 113
5.1 Systems with Nonlinearities in Stochastic Part . . . . . . . . . . . 113
5.1.1 Scalar First-Order Differential Equation . . . . . . . . . . 113
5.1.2 Scalar Second-Order Differential Equation . . . . . . . . . 120
5.2 Systems with Nonlinearities in Both Deterministic and Stochastic
Parts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
5.3 Stability in Probability of Nonlinear Systems . . . . . . . . . . . . 126
5.4 Systems with Fractional Nonlinearity . . . . . . . . . . . . . . . . 131
5.4.1 Equilibrium Points . . . . . . . . . . . . . . . . . . . . . . 132
5.4.2 Stochastic Perturbations, Centering, and Linearization . . . 134
5.4.3 Stability of Equilibrium Points . . . . . . . . . . . . . . . 135
5.4.4 Numerical Analysis . . . . . . . . . . . . . . . . . . . . . 141
6 Matrix Riccati Equations in Stability of Linear Stochastic
Differential Equations with Delays . . . . . . . . . . . . . . . . . . . 153
6.1 Equations with Constant Delays . . . . . . . . . . . . . . . . . . . 153
6.1.1 One Delay in Deterministic Part and One Delay
in Stochastic Part of Equation . . . . . . . . . . . . . . . . 153
6.1.2 Several Delays in Deterministic Part of Equation . . . . . . 157
6.2 Distributed Delay . . . . . . . . . . . . . . . . . . . . . . . . . . 160
6.3 Combination of Discrete and Distributed Delays . . . . . . . . . . 164
6.4 Equations with Nonincreasing Delays . . . . . . . . . . . . . . . . 170
6.4.1 One Delay in Deterministic and One Delay in Stochastic
Parts of Equation . . . . . . . . . . . . . . . . . . . . . . . 170
6.4.2 Several Delays in Deterministic Part of Equation . . . . . . 174
Contents xi

6.5 Equations with Bounded Delays . . . . . . . . . . . . . . . . . . 179

6.5.1 The First Way of Constructing a Lyapunov Functional . . . 179
6.5.2 The Second Way of Constructing a Lyapunov Functional . 180
6.6 Equations with Unbounded Delays . . . . . . . . . . . . . . . . . 182
7 Stochastic Systems with Markovian Switching . . . . . . . . . . . . . 187
7.1 The Statement of the Problem . . . . . . . . . . . . . . . . . . . . 187
7.2 Stability Theorems . . . . . . . . . . . . . . . . . . . . . . . . . . 189
7.3 Application to Markov Chain with Two States . . . . . . . . . . . 193
7.3.1 The First Stability Condition . . . . . . . . . . . . . . . . 194
7.3.2 The Second Stability Condition . . . . . . . . . . . . . . . 194
7.4 Numerical Simulation of Systems with Markovian Switching . . . 198
7.4.1 System Without Stochastic Perturbations . . . . . . . . . . 199
7.4.2 System with Stochastic Perturbations . . . . . . . . . . . . 202
7.4.3 System with Random Delay . . . . . . . . . . . . . . . . . 203
7.4.4 Some Generalization of Algorithm of Markov Chain
Numerical Simulation . . . . . . . . . . . . . . . . . . . . 206
8 Stabilization of the Controlled Inverted Pendulum by a Control
with Delay . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 209
8.1 Linear Model of the Controlled Inverted Pendulum . . . . . . . . . 209
8.1.1 Stabilization by the Control Depending on Trajectory . . . 209
8.1.2 Some Examples . . . . . . . . . . . . . . . . . . . . . . . 214
8.1.3 About Stabilization by the Control Depending on Velocity
or on Acceleration . . . . . . . . . . . . . . . . . . . . . . 219
8.1.4 Stabilization by Stochastic Perturbations . . . . . . . . . . 220
8.2 Nonlinear Model of the Controlled Inverted Pendulum . . . . . . . 222
8.2.1 Stabilization of the Trivial Solution . . . . . . . . . . . . . 222
8.2.2 Nonzero Steady-State Solutions . . . . . . . . . . . . . . . 226
8.2.3 Stable, Unstable, and One-Sided Stable Points
of Equilibrium . . . . . . . . . . . . . . . . . . . . . . . . 228
8.3 Numerical Analysis of the Controlled Inverted Pendulum . . . . . 229
8.3.1 Stability of the Trivial Solution and Limit Cycles . . . . . . 229
8.3.2 Nonzero Steady-State Solutions of the Nonlinear Model . . 234
8.3.3 Stabilization of the Controlled Inverted Pendulum Under
Influence of Markovian Stochastic Perturbations . . . . . . 240
9 Stability of Equilibrium Points of Nicholson’s Blowflies Equation
with Stochastic Perturbations . . . . . . . . . . . . . . . . . . . . . . 251
9.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 251
9.2 Two Points of Equilibrium, Stochastic Perturbations, Centering,
and Linearization . . . . . . . . . . . . . . . . . . . . . . . . . . 252
9.3 Sufficient Conditions for Stability in Probability for Both
Equilibrium Points . . . . . . . . . . . . . . . . . . . . . . . . . . 253
9.4 Numerical Illustrations . . . . . . . . . . . . . . . . . . . . . . . . 255
xii Contents

10 Stability of Positive Equilibrium Point of Nonlinear System of Type

of Predator–Prey with Aftereffect and Stochastic Perturbations . . . 257
10.1 System Under Consideration . . . . . . . . . . . . . . . . . . . . 257
10.2 Equilibrium Points, Stochastic Perturbations, Centering, and
Linearization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 260
10.2.1 Equilibrium Points . . . . . . . . . . . . . . . . . . . . . . 260
10.2.2 Stochastic Perturbations and Centering . . . . . . . . . . . 261
10.2.3 Linearization . . . . . . . . . . . . . . . . . . . . . . . . . 263
10.3 Stability of Equilibrium Point . . . . . . . . . . . . . . . . . . . . 264
10.3.1 First Way of Constructing a Lyapunov Functional . . . . . 265
10.3.2 Second Way of Constructing a Lyapunov Functional . . . . 272
10.3.3 Stability of the Equilibrium Point of Ratio-Dependent
Predator–Prey Model . . . . . . . . . . . . . . . . . . . . 277
11 Stability of SIR Epidemic Model Equilibrium Points . . . . . . . . . 283
11.1 Problem Statement . . . . . . . . . . . . . . . . . . . . . . . . . . 283
11.2 Stability in Probability of the Equilibrium Point E0 = (bμ−11 , 0, 0) 284
11.3 Stability in Probability of the Equilibrium Point E∗ = (S ∗ , I ∗ , R ∗ ) 289
11.4 Numerical Simulation . . . . . . . . . . . . . . . . . . . . . . . . 295
12 Stability of Some Social Mathematical Models with Delay Under
Stochastic Perturbations . . . . . . . . . . . . . . . . . . . . . . . . . 297
12.1 Mathematical Model of Alcohol Consumption . . . . . . . . . . . 297
12.1.1 Description of the Model of Alcohol Consumption . . . . . 297
12.1.2 Normalization of the Initial Model . . . . . . . . . . . . . 299
12.1.3 Existence of an Equilibrium Point . . . . . . . . . . . . . . 300
12.1.4 Stochastic Perturbations, Centralization, and Linearization . 301
12.1.5 Stability of the Equilibrium Point . . . . . . . . . . . . . . 302
12.1.6 Numerical Simulation . . . . . . . . . . . . . . . . . . . . 309
12.2 Mathematical Model of Social Obesity Epidemic . . . . . . . . . . 310
12.2.1 Description of the Considered Model . . . . . . . . . . . . 311
12.2.2 Existence of an Equilibrium Point . . . . . . . . . . . . . . 312
12.2.3 Stochastic Perturbations, Centralization, and Linearization . 314
12.2.4 Stability of an Equilibrium Point . . . . . . . . . . . . . . 316
12.2.5 Numerical Simulation . . . . . . . . . . . . . . . . . . . . 321
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 325
Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 339
Chapter 1
Short Introduction to Stability Theory
of Deterministic Functional Differential
Equations

1.1 Some Peculiarities of Functional Differential Equations

In comparison with ordinary differential equations, functional differential equations

have some peculiarities. Below we consider some of them for deterministic func-
tional differential equations.

1.1.1 Description of Functional Differential Equations

This section covers some theoretical backgrounds of the equations used in the book
with concentration on mathematical rigor. General classification of the equations
and some properties of their solutions will be discussed.
Let us consider equations with an unknown function depending on a continuous
argument t , which may be treated as time. The equations can be scalar or vector
equations and have the same dimension as the unknown function. It is assumed that
all variables under consideration are real.
A functional equation is an equation involving an unknown function for different
argument values. The equations
   
2x(3t) + 3x(2t) = 1, x x(t) = x t 2 + 1

are examples of functional equations. The differences between the argument values
of an unknown function and t in a functional equation are called argument devia-
tions. If all argument deviations are constant, as in the example

x(t) = t 2 x(t + 1) − x 2 (t − 2),

then the functional equation is called a difference equation [278].

Above we have given some examples of functional equations with discrete (or
concentrated) argument deviations. By increasing in the equation “the number of

L. Shaikhet, Lyapunov Functionals and Stability of Stochastic Functional 1

Differential Equations, DOI 10.1007/978-3-319-00101-2_1,
© Springer International Publishing Switzerland 2013
2 1 Short Introduction to Stability Theory of Deterministic Functional

summands” and simultaneously decreasing the differences between neighboring ar-

gument values, one naturally arrives at functional equations with continuous (or
distributed) argument deviations
 t
x(t) = x(0) + x(s) ds
0

and mixed (both continuous and discrete) argument deviations [35, 66, 167, 233,
234, 269–272]
 t
x(t) = x(t − 1) + x(s) ds.
t−1

These equations are called integral and integral functional equations (in particular,
integral difference equations).
Combining the notions of differential and functional equations, we obtain the
notion of functional differential equation [5, 44, 104–106, 122, 126, 131–133]
or, equivalently, differential equation with deviating argument [66, 167]. Thus,
this is an equation connecting the unknown function and some of its derivatives
for, in general, different argument values. Here also the argument values can be
discrete, continuous, or mixed. Correspondingly, one introduces the notions of
differential–difference equation, differential equation of neutral type, integral or
Volterra integro–differential equation [4, 13, 55, 57, 63, 64, 74–78, 89, 90, 104,
139, 150, 151, 171, 172, 185, 208, 222, 229, 230, 237, 238, 292], etc.
The order of a functional differential equation is the order of the highest deriva-
tive of the unknown function entering in the equation. So, a functional equation may
be regarded as a functional differential equation of order zero. Hence the notion of
functional differential equation generalizes all equations of mathematical analysis
for function of a continuous argument. A similar assertion holds for function de-
pending on several arguments.

1.1.2 Reducing to Ordinary Differential Equations

Sometimes Volterra-type integro–differential equations can be reduced to equiva-

lent ordinary differential equations. We explain this phenomenon by the following
equation with exponentially fading memory [281]:
  t 
ẋ(t) = f t, x(θ )e−k(t−θ) dθ , t ≥ t0 .
t0

Put
 t
y(t) = x(θ )e−k(t−θ) dθ.
t0
1.2 Method of Steps for Retarded Functional Differential Equations 3

Differentiating this equality, we obtain that the initial equation is equivalent to the
system of ordinary differential equations
 
ẋ(t) = f t, y(t) , ẏ(t) = x(t) − ky(t),

with the additional initial condition y(t0 ) = 0.

Consider another example of Volterra-type equation that can be reduced to ordi-
nary differential equations. Consider the equation
 t
ẋ(t) = −2x(t) + sin(t − s)x(s) ds.
0

Differentiating this equation two times,

 t
ẍ(t) = −2ẋ(t) + cos(t − s)x(s) ds,
0
 t
...
x (t) = −2ẍ(t) + x(t) − sin(t − s)x(s) ds,
0

and adding the last equation to the initial equation, we obtain

...
x (t) + 2ẍ(t) + ẋ(t) + x(t) = 0.

This equivalence looks like a surprise because integro–differential equations de-

scribe processes whose rate is determined by all previous states, while ordinary
differential equations describe processes determined by the current states only.

1.2 Method of Steps for Retarded Functional Differential

Equations
For discrete delays, very often to solve the Cauchy problem on a finite interval, one
can apply a step method. Let us explain this method using the scalar equation with
one discrete delay
 
ẋ(t) = f t, x(t), x(t − h) , t ≥ t0 , h = const > 0,
(1.1)
xt0 (s) = φ(s), s ∈ [−h, 0].

Let f : [t0 , ∞) × R2 → R be a continuous function satisfying the Lipschitz condi-

tion with respect to the second argument. The initial function φ for (1.1) must be
assigned on the interval [t0 − h, t0 ].
If t ∈ [t0 , t0 + h] (it is the first step), then t − h ∈ [t0 − h, t0 ]. Therefore, the delay
differential equation (1.1) converts into the ordinary differential equation
 
ẋ(t) = f t, x(t), φ(t − t0 − h) , t0 ≤ t ≤ t0 + h,
x(t0 ) = φ(0).
4 1 Short Introduction to Stability Theory of Deterministic Functional

Fig. 1.1 Solutions of (1.2) for a = −1, h = 1, and different initial values of x0

Solving this equation for the initial value x(t0 ) = φ(0), we get the solution on the
interval [t0 , t0 + h]. If now t0 + h ≤ t ≤ t0 + 2h (it is the second step), then t − h ∈
[t0 , t0 + h], and so x(t − h) is known from the first step. Hence, (1.1) for t0 + h ≤
t ≤ t0 + 2h once again converts into the ordinary differential equation, which with
the known initial value x(t0 + h) defines the solution x(t) on this interval. After
that we consider the interval [t0 + 2h, t0 + 3h], etc. In this way the solution can be
obtained for arbitrarily large t (theoretically, for the whole semiaxis [t0 , ∞)).
For the simplest case where the right-hand side of (1.1) does not contain x(t), at
each step the solution is reduced to the integration of a given function.

Example 1.1 Consider the scalar differential equation with discrete delay

ẋ(t) = ax(t − h), x(s) ≡ x0 = const, −h ≤ s ≤ 0. (1.2)

Using the method of steps, it is easy to get the solution of this equation in the form



k
− lh)l+1
l+1 (t t
x(t) = x0 1 + a , k= . (1.3)
(l + 1)! h
l=0

The solutions of (1.2) obtained via (1.3) are shown in Fig. 1.1 for a = −1, h = 1,
and different initial values x0 . One can see that the solutions for different x0 intersect
each other at the point t = 1 and many other points. It is known that this situation is
impossible for ordinary differential equations.
1.3 Characteristic Equation for Differential Equation with Discrete Delays 5

1.3 Characteristic Equation for Differential Equation with

Discrete Delays
Consider the nth-order differential equation with discrete delays

 
n−1 m(i)
x (n) (t) + aij x (i) (t − hij ) = 0, hij ≥ 0. (1.4)
i=0 j =1

With this equation, the following function is associated:

 
n−1 m(i)
(z) = zn + aij zi e−zhij , (1.5)
i=0 j =1

which is called the characteristic quasipolynomial of (1.4).

It is known [80] that the trivial solution of (1.4) is asymptotically stable if and
only if all zeros z = α + iβ of the characteristic quasipolynomial (1.5) (or all roots of
the characteristic equation (z) = 0) satisfy the condition α < 0. The bound of the
region of asymptotic stability can be defined by the equality (iβ) = 0, i 2 = −1,
β ∈ R.

Example 1.2 The characteristic quasipolynomial for (1.2) is (z) = z − ae−hz .

Thus, the characteristic equation (iβ) = 0 gives the system of two equations

a cos βh = 0, β + a sin βh = 0

with the solution βh = 12 π or h = − 2a 1

π . This means that the trivial solution of
1
(1.2) is asymptotically stable if and only if h < 2|a| π , a < 0.
Put a = −1. Then, for h = 1, the trivial solution of (1.2) is asymptotically stable,
and all solutions go to zero (Fig. 1.1). If h = 12 π , then the trivial solution of (1.2)
is stable but not asymptotically stable, and all solutions are bounded (Fig. 1.2). If
h > 12 π , then the trivial solution of (1.2) is unstable, and all solutions go to infinity
(Fig. 1.3, h = 2).
Note also that as h → 0, (1.2) goes to an ordinary differential equation with
asymptotically (for a < 0) stable trivial solution and different solutions without in-
tersections. In Fig. 1.4 the solutions of (1.2) are shown for h = 0.5.

Example 1.3 Consider the scalar differential equation

ẋ(t) + ax(t − h1 ) + bx(t − h2 ) = 0, 0 ≤ h1 < h2 . (1.6)

The characteristic quasipolynomial of this equation is

(z) = z + ae−h1 z + be−h2 z .

Suppose that a + b ≤ 0. Then (0) = a + b ≤ 0 and (∞) = ∞. Therefore,

there exists α ≥ 0 such that (α) = 0. This means that by the condition a + b ≤ 0
6 1 Short Introduction to Stability Theory of Deterministic Functional

Fig. 1.2 Solutions of (1.2) for a = −1, h = 12 π , and different initial values of x0

Fig. 1.3 Solutions of (1.2) for a = −1, h = 2, and different initial values of x0

the trivial solution of (1.6) cannot be asymptotically stable. So, we will suppose that
a + b > 0.
1.3 Characteristic Equation for Differential Equation with Discrete Delays 7

Fig. 1.4 Solutions of (1.2) for a = −1, h = 0.5, and different initial values of x0

The equality (iβ) = 0 gives the system of two equations

a cos βh1 + b cos βh2 = 0,

(1.7)
a sin βh1 + b sin βh2 = β.

From (1.7) it follows that the bounds of the stability region for (1.6) are formed in
the (a, b)-plane by the straight line a + b = 0 and the parametric curve
β cos βh2 β cos βh1
a=− , b= . (1.8)
sin β(h2 − h1 ) sin β(h2 − h1 )

Example 1.4 Consider the first-order scalar linear differential equation with delay

ẋ(t) + ax(t) + bx(t − h) = 0, (1.9)

which is the special case of (1.6) with h1 = 0, h2 = h. Thus, (1.8) takes the form
β cos βh β
a=− , b= , βh ∈ [0, π).
sin βh sin βh
The bound of the asymptotic stability region for the trivial solution of (1.9) is defined
by the conditions
a + b = 0, bh < 1,
   (1.10)
a + b cos h b2 − a 2 = 0, bh ≥ 1.
The bound of stability region given by conditions (1.10) is shown in Fig. 1.5 for
h = 1.3.
8 1 Short Introduction to Stability Theory of Deterministic Functional

Fig. 1.5 Stability region for (1.9) given by conditions (1.10) for h = 1.3

Fig. 1.6 Solutions of (1.9) for h = 1.3, x(s) = 0.5, −h ≤ s ≤ 0, a = −0.5, and different values
of b: (1) b = 0.4, (2) b = 0.5, (3) b = 0.7, (4) b = 0.909, (5) b = 1

In Fig. 1.6 the solutions of (1.9) are shown for x(s) = 0.5, −1.3 ≤ s ≤ 0, a =
−0.5 and different values of b: (1) b = 0.4 (x(t) → ∞); (2) b = 0.5 (x(t) = const);
(3) b = 0.7 (x(t) → 0); (4) b = 0.909 (x(t) is a periodical solution); (5) b = 1
(x(t) → ±∞).
In Fig. 1.7 the solutions of (1.9) are shown for x(s) = 0.5, −1.3 ≤ s ≤ 0, a = 0.5
and different values of b: (1) b = −0.6 (x(t) → ∞); (2) b = −0.5 (x(t) = const);
1.3 Characteristic Equation for Differential Equation with Discrete Delays 9

Fig. 1.7 Solutions of (1.9) for h = 1.3, x(s) = 0.5, −h ≤ s ≤ 0, a = 0.5, and different values of
b: (1) b = −0.6, (2) b = −0.5, (3) b = 0, (4) b = 1, (5) b = 1.537, (6) b = 1.7

(3) b = 0 (x(t) → 0); (4) b = 1 (x(t) → 0); (5) b = 1.537 (x(t) is a periodical
solution); (6) b = 1.7 (x(t) → ±∞).
The points that were used in Figs. 1.6 and 1.7 are shown also in Fig. 1.5. The
solutions of (1.9) were obtained via its difference analogue in the form

xi+1 = (1 − a)xi − bxi−m ,

where  is the step of discretization, and xi = x(i), m = h/,  = 0.01.

Example 1.5 Consider the first-order scalar linear differential equation of neutral
type

ẋ(t) + ax(t) + bx(t − h) + cẋ(t − h) = 0, |c| < 1, (1.11)

which is a generalization of (1.9).

The corresponding characteristic quasipolynomial is (z) = z + a + be−hz +
cze−hz . The equality (iβ) = 0 gives the system of two equations

a + b cos βh + cβ sin βh = 0,
β − b sin βh + cβ cos βh = 0,

with the solution

β(c + cos βh) β(1 + c cos βh)

a=− , b= , βh ∈ [0, π).
sin βh sin βh
10 1 Short Introduction to Stability Theory of Deterministic Functional

Fig. 1.8 Stability regions for (1.11) given by (1.12) are shown in space of parameters (a, b) for
h = 1 and different values of c: (1) c = −0.999, (2) c = −0.5, (3) c = 0, (4) c = 0.6, (5) c = 0.85,
(6) c = 0.95, (7) c = 0.995

The bound of the region of asymptotic stability of the trivial solution of (1.11) is
defined by the conditions

a + b = 0, bh < 1 + c,

  (1.12)
b2 − a 2
a + bc + (ac + b) cos h = 0, b > |a|.
1 − c2

In Fig. 1.8 the stability regions given by (1.12) for (1.11) are shown in the space
of the parameters (a, b) for h = 1 and different values of c: (1) c = −0.999, (2)
c = −0.5, (3) c = 0, (4) c = 0.6, (5) c = 0.85, (6) c = 0.95, (7) c = 0.995. In Fig. 1.8
one can see that if a > |b|, then the trivial solution of (1.11) is asymptotically stable
for all h > 0 and |c| < 1.
In Fig. 1.9 the stability regions for (1.11) are shown in the space of the parameters
(c, b) for a = 0.4 and different values of h: (1) h = 2, (2) h = 1, (3) h = 0.7. In
Figs. 1.10 and 1.11 the similar stability regions are shown for a = 0 and a = −0.4,
respectively.
In Fig. 1.12 the solutions of (1.11) are shown for c = 0.1, x(s) = 0.5, s ≤ 0,
a = −0.5, h = 1.3 and different values of b: (1) b = 0.4 (x(t) goes to infinity), (2)
b = 0.5 (x(t) = const), (3) b = 0.7 (x(t) goes to zero), (4) b = 1.004 (x(t) is a
bounded periodical solution), (5) b = 1.1 (|x(t)| goes to infinity). In Fig. 1.13 the
similar solutions are shown for the initial condition x(s) = cos(s), s ≤ 0, and the
same values of other parameters.
1.3 Characteristic Equation for Differential Equation with Discrete Delays 11

Fig. 1.9 Stability regions for

(1.11) given by (1.12) are
shown in space of parameters
(c, b) for a = 0.4 and
different values of h:
(1) h = 2, (2) h = 1,
(3) h = 0.7

Fig. 1.10 Stability regions

for (1.11) given by (1.12) are
shown in space of parameters
(c, b) for a = 0 and different
values of h: (1) h = 2,
(2) h = 1, (3) h = 0.7

Fig. 1.11 Stability regions

for (1.11) given by (1.12) are
shown in space of parameters
(c, b) for a = −0.4 and
different values of h:
(1) h = 2, (2) h = 1,
(3) h = 0.7
12 1 Short Introduction to Stability Theory of Deterministic Functional

Fig. 1.12 Solutions of (1.11) are shown for c = 0.1, h = 1.3, x(s) = 0.5, s ∈ [−h, 0], a = −0.5,
and different values of b: (1) b = 0.4, (2) b = 0.5, (3) b = 0.7, (4) b = 1.004, (5) b = 1.1

Fig. 1.13 Solutions of (1.11) are shown for c = 0.1, h = 1.3, x(s) = cos(s), s ∈ [−h, 0],
a = −0.5, and different values of b: (1) b = 0.4, (2) b = 0.5, (3) b = 0.7, (4) b = 1.004, (5) b = 1.1

In Fig. 1.14 the solutions of (1.11) are shown for c = −0.1, x(s) = 0.5, s ≤ 0,
a = −0.5, h = 1.3 and different values of b: (1) b = 0.4 (x(t) goes to infinity), (2)
b = 0.5 (x(t) = const), (3) b = 0.7 (x(t) goes to zero), (4) b = 0.804 (x(t) is a
bounded periodical solution), (5) b = 0.9 (|x(t)| goes to infinity).
The solutions of (1.11) were obtained via its difference analogue in the form

xi+1 = (1 − a)xi + (c − b)xi−m − cxi+1−m ,

where  is the step of discretization, and xi = x(i), m = h/,  = 0.01.

1.3 Characteristic Equation for Differential Equation with Discrete Delays 13

Fig. 1.14 Solutions of (1.11) are shown for c = −0.1, h = 1.3, x(s) = 0.5, s ∈ [−h, 0], a = −0.5,
and different values of b: (1) b = 0.4, (2) b = 0.5, (3) b = 0.7, (4) b = 0.804, (5) b = 0.9

Example 1.6 Let us consider the scalar second-order differential equation

ẍ(t) + a ẋ(t − h1 ) + bx(t − h2 ) = 0. (1.13)

The characteristic quasipolynomial of (1.13) is (z) = z2 + aze−h1 z + be−h2 z . The

equality (iβ) = 0 gives the following system of two equations:

βa cos βh1 − b sin βh2 = 0,

(1.14)
βa sin βh1 + b cos βh2 = β 2 ,

with the solution

β sin βh2 β 2 cos βh1

a= , b= .
cos β(h1 − h2 ) cos β(h1 − h2 )

In Fig. 1.15 the stability regions for (1.13) are shown for h1 = 1 and different
values of h2 : (1) h2 = 0.2, (2) h2 = 0.6, (3) h2 = 1, (4) h2 = 1.4, (5) h2 = 1.8. In
Fig. 1.16 the stability regions for (1.13) are shown for h2 = 1 and different values
of h1 : (1) h1 = 0.8, (2) h1 = 0.9, (3) h1 = 1, (4) h1 = 1.1, (5) h1 = 1.2.
Let us consider some particular cases of (1.13).
(a) h1 = h2 = h. In this case the solution of (1.14) has the form a = β sin βh, b =
β 2 cos βh, 0 ≤ βh ≤ π2 . In Fig. 1.17 the stability regions are shown for different
values of h: (1) h = 0.5, (2) h = 0.75, (3) h = 1, (4) h = 1.25, (5) h = 1.5.
2
(b) h1 = 0, h2 = h. In this case we obtain a = β tan βh, b = cosβ βh , 0 ≤ βh < π2 . In
Fig. 1.18 the stability regions are shown for different values of h: (1) h = 1, (2)
h = 2, (3) h = 3, (4) h = 4, (5) h = 5.
14 1 Short Introduction to Stability Theory of Deterministic Functional

Fig. 1.15 Stability regions for (1.13) for h1 = 1 and different values of h2 : (1) h2 = 0.2,
(2) h2 = 0.6, (3) h2 = 1, (4) h2 = 1.4, (5) h2 = 1.8

Fig. 1.16 Stability regions for (1.13) for h2 = 1 and different values of h1 : (1) h1 = 0.8,
(2) h1 = 0.9, (3) h1 = 1, (4) h1 = 1.1, (5) h1 = 1.2

(c) h1 = h, h2 = 0. In this case from the first equation of (1.14) it follows that
cos βh = 0. Therefore,
π
β= (2l + 1), l = 0, 1, . . . ,
2h
1.3 Characteristic Equation for Differential Equation with Discrete Delays 15

Fig. 1.17 Stability regions for (1.13) in the case h1 = h2 = h for different values of h: (1) h = 0.5,
(2) h = 0.75, (3) h = 1, (4) h = 1.25, (5) h = 1.5

Fig. 1.18 Stability regions for (1.13) in the case h1 = 0, h2 = h for different values of h:
(1) h = 1, (2) h = 2, (3) h = 3, (4) h = 4, (5) h = 5

and from system (1.14) it follows that the stability region consists of a sequence
of triangles, formed by parts of the a- and the b-axes and the line segments
 
l+1 2l +1 2l + 1 2
b = (−1) πa + π , l = 0, 1, . . . .
2h 2h
16 1 Short Introduction to Stability Theory of Deterministic Functional

Fig. 1.19 Stability regions for (1.13) in the case h1 = 1, h2 = 0

Fig. 1.20 Stability regions for (1.13) in the case h1 = 2, h2 = 0

The stability regions (the triangles OB0 A0 , A0 B1 A1 , A1 B2 A2 , A2 B3 A3 , . . .)

for this case are shown in Fig. 1.19 (h = 1) and Fig. 1.20 (h = 2). The points Al
1.3 Characteristic Equation for Differential Equation with Discrete Delays 17

Fig. 1.21 Stability regions for (1.13) in the case h1 = h, h2 = 2h for different values of h:
(1) h = 1, (2) h = 1.5, (3) h = 3

and Bl have the coordinates respectively

   
2l + 1 2
Al = 0, π , l = 0, 1, . . . ,
2h
   
π π (2l − 1)(2l + 1) 2
B0 = ,0 , Bl = (−1)l , π , l = 1, 2, . . . .
2h h 4h2

(d) h1 = h, h2 = 2h. In this case from the first equation of system (1.14) it follows
that
βa cos βh − b sin 2βh = (βa − 2b sin βh) cos βh = 0.
If cos βh = 0, i.e., βh = π2 , then from the second equation of system (1.14) it
follows that a part of the stability region bound is defined by the straight line
 
π π
b= a− . (1.15)
2h 2h

The stability regions in this case are shown in Fig. 1.21 for different values of h:
(1) h = 1, (2) h = 1.5, (3) h = 3. One can see that the part of the stability region
bound is defined by the straight line (1.15).

Example 1.7 Consider the scalar second-order differential equation

ẍ(t) = ax(t) + b1 x(t − h1 ) + b2 x(t − h2 ), h1 > h2 . (1.16)

18 1 Short Introduction to Stability Theory of Deterministic Functional

The characteristic quasipolynomial of (1.16) is

(z) = z2 − a − b1 e−h1 z − b2 e−h2 z , z = α + iβ.

Suppose that a + b1 + b2 ≥ 0. Then (0) ≤ 0 and (∞) = ∞. Therefore, there

exists α ≥ 0 such that (α) = 0. This means that by the condition a + b1 + b2 ≥
0 the trivial solution of (1.16) cannot be asymptotically stable. So, the inequality
a + b1 + b2 < 0 is the necessary condition for asymptotic stability of the trivial
solution of (1.16).
The equality (iβ) = 0 gives the following system of two equations:
 
b1 cos βh1 + b2 cos βh2 = − a + β 2 ,
(1.17)
b1 sin βh1 + b2 sin βh2 = 0,

with the solution

(a + β 2 ) sin βh2 (a + β 2 ) sin βh1
b1 = , b2 = − .
sin β(h1 − h2 ) sin β(h1 − h2 )

Example 1.8 Consider the scalar differential equation

ẍ(t) = −ax(t) + bx(t − h), (1.18)

where a, b, h are nonnegative constants. The characteristic quasipolynomial of

(1.18) is
(z) = z2 + a − be−zh , z = α + βi.
By the condition b ≥ a we have (0) ≤ 0, (∞) = ∞. So, there exists α ≥ 0
such that (α) = 0. Therefore, the inequality b < a is the necessary condition for
asymptotic stability of the trivial solution of (1.18).
The characteristic equation (iβ) = 0 can be represented in the form of the sys-
tem of two equations

b cos βh = a − β 2 , b sin βh = 0
2
with the following solutions: (1) b = a, (2) b = 0, a ≥ 0, (3) a + b = πh2 . So, the
necessary and sufficient condition for asymptotic stability of the trivial solution of
(1.18) takes the form
π2
b<a< − b. (1.19)
h2
In Fig. 1.22 the stability regions given by condition (1.19) are shown for different
values of h: (1) h = 2, (2) h = 1.5, (3) h = 1.2, (4) h = 1, (5) h = 0.8.
Put x(s) = es , s ∈ [−h, 0], a = 1, and consider the behavior of the solution of
(1.18) for different values of b and h.
In Fig. 1.23 the solutions of (1.18) are shown for h = 0 and the values of b:
(1) b = 0.3, (2) b = 0.9, (3) b = 1.01. For h = 0, condition (1.19) is the necessary
1.3 Characteristic Equation for Differential Equation with Discrete Delays 19

Fig. 1.22 Stability regions for (1.18) given by condition (1.19) for different values of h: (1) h = 2,
(2) h = 1.5, (3) h = 1.2, (4) h = 1, (5) h = 0.8

Fig. 1.23 Solutions of (1.18) for a = 1, h = 0 and different values of b: (1) b = 0.3, (2) b = 0.9,
(3) b = 1.01

stability condition only, so, the solution is periodical in the cases (1), (2) and goes
to infinity in the case (3).
In Fig. 1.24 the solutions of (1.18) are shown for h = 0.5 and the same values of
a and b. Condition (1.19) holds in the cases (1), (2) and does not hold in the case (3).
So, the solution goes to zero in the cases (1), (2) and goes to infinity in the case (3).
The similar picture one can see in Fig. 1.25 for h = 2.
In Fig. 1.26 the solutions of (1.18) are shown for h = 2.755359. In this case
2 2
condition (1.19) does not hold since a = πh2 − b = 1 for b = 0.3 and a = 1 > πh2 −
b = 0.4 for b = 0.9. So, the solution of (1.18) is periodical for b = 0.3 and goes
20 1 Short Introduction to Stability Theory of Deterministic Functional

Fig. 1.24 Solutions of (1.18) for a = 1, h = 0.5 and different values of b: (1) b = 0.3, (2) b = 0.9,
(3) b = 1.01

Fig. 1.25 Solutions of (1.18) for a = 1, h = 2 and different values of b: (1) b = 0.3, (2) b = 0.9,
(3) b = 1.01

to infinity for b = 0.9. In Fig. 1.27 the solutions of (1.18) are shown for h = 0.8,
2
b = 0.6 and different conditions on a: (1) a = b, (2) a = 1.3b, (3) a = πh2 − b.
In Figs. 1.22–1.27 one can see that the solution of (1.18) goes to constant on the
2
bound a = b of stability region, is a periodical on the bound a + b = πh2 , goes to
zero between these bounds, and goes to infinity out of these bounds. The solutions
of (1.18) were obtained via its difference analogue in the form
 
xi+1 = 2 − a2 xi − xi−1 + b2 xi−m ,

where  is the step of discretization, and xi = x(i), m = h/,  = 0.01.

1.4 The Influence of Small Delays on Stability 21

Fig. 1.26 Solutions of (1.18) for a = 1, h = 2.755359, and different values of b: (1) b = 0.3,
(2) b = 0.9

Fig. 1.27 Solutions of (1.18) for h = 0.8, b = 0.6, and different values of a: (1) a = b,
2
(2) a = 1.3b, (3) a = πh2 − b

1.4 The Influence of Small Delays on Stability

The effect of time delay is very essential in different real processes [46, 109, 225].
Many actual phenomena involve small delays, which are often neglected in the pro-
cess of mathematical modeling. Sometimes this leads to false conclusions. We shall
show that sometimes even small delay can change essentially the properties of the
solution.
22 1 Short Introduction to Stability Theory of Deterministic Functional

Fig. 1.28 Solutions of (1.20) for a = 1, h = 2, x0 (s) = 0.7, s ∈ [−h, 0], and different values of
b: (1) b = −1.1, (2) b = −1, (3) b = −0.7, (4) b = 0, (5) b = 0.7, (6) b = 1, (7) b = 1.1

Example 1.9 Consider the first-order differential equation of neutral time

 
ẋ(t) + ax(t) − b ẋ(t − h) + ax(t − h) = 0, a > 0. (1.20)

For h = 0, we have the equation (1 − b)[ẋ(t) + ax(t)] = 0 with the solution x(t) =
x(0) exp(−at), which is asymptotically stable for a > 0 and arbitrary b. If h > 0,
then the appropriate characteristic equation (z + a)(1 − be−hz ) = 0 has all roots

1 
zk = ln |b| + i2kπ , i 2 = −1, k = 0, ±1, ±2, . . . ,
h

with real parts Rezk = h1 ln |b|. So, if |b| > 1, then the trivial solution of (1.20) is
asymptotically stable for h = 0 and unstable for each h > 0. If |b| ≤ 1, we have
another picture.
In Fig. 1.28 the solutions of (1.20) are shown for the initial function x0 (s) = 0.7,
s ∈ [−h, 0], h = 2, a = 1, and for different values of the parameter b: (1) b = −1.1,
(2) b = −1, (3) b = −0.7, (4) b = 0, (5) b = 0.7, (6) b = 1, (7) b = 1.1. We can
see that in the cases (1) and (7) the solution of (1.20) goes to infinity, in the case (2)
it is a periodic solution, in the cases (3), (4), and (5) the solution goes to zero, and
in the case (6) it is a constant. In Fig. 1.29 a similar picture is shown for the initial
function x0 (s) = cos s, s ∈ [−h, 0], and the same values of other parameters. We
can see again that in the cases (1) and (7) the solution of (1.20) goes to infinity, in
the cases (2) and (6) it is a periodic solution, and in the cases (3), (4), and (5) it goes
to zero. In Fig. 1.30 the same situation with the solution of (1.20) is shown for small
1.4 The Influence of Small Delays on Stability 23

Fig. 1.29 Solutions of (1.20) for a = 1, h = 2, x0 (s) = cos(s), s ∈ [−h, 0], and different values
of b: (1) b = −1.1, (2) b = −1, (3) b = −0.7, (4) b = 0, (5) b = 0.7, (6) b = 1, (7) b = 1.1

Fig. 1.30 Solutions of (1.20) for a = 1, h = 0.3, x0 (s) = cos(s), s ∈ [−h, 0], and different values
of b: (1) b = −1.1, (2) b = −1, (3) b = 0.7, (4) b = 0.9, (5) b = 0.97, (6) b = 1, (7) b = 1.1
24 1 Short Introduction to Stability Theory of Deterministic Functional

Fig. 1.31 Solutions of (1.20) for a = 1, h = 0.3, x0 (s) = cos(s), s ∈ [−h, 0], and different val-
ues of b: (1) b = −1.1, (2) b = 0.7, (3) b = 1.1, and appropriate process y(t): (4) b = −1.1,
(5) b = 0.7, (6) b = 1.1

delay h = 0.3, the initial function x0 (s) = cos s, s ∈ [−h, 0], a = 1, and for different
values of the parameter b: (1) b = −1.1, (2) b = −1, (3) b = 0.7, (4) b = 0.9, (5)
b = 0.97, (6) b = 1, (7) b = 1.1.
Note that the process y(t) = x(t) − bx(t − h) goes to zero for all b. In Fig. 1.31
the solutions of (1.20) are shown for the initial function x0 (s) = cos s, s ∈ [−h, 0],
a = 1, and for different values of the parameter b: (1) b = −1.1, (2) b = 0.7, (3)
b = 1.1 and the appropriate process y(t): (4) b = −1.1, (5) b = 0.7, (6) b = 1.1.
One can see a similar situation for the second-order differential equation of neu-
tral type
 
ẍ(t) + a ẋ(t) + bx(t) = c ẍ(t − h) + a ẋ(t − h) + bx(t − h) ,
a > 0, b > 0, c > 1. (1.21)

If h = 0, then the trivial solution of (1.21) is asymptotically stable. But for any
h > 0, the trivial solution of (1.21) is unstable since for h > 0, the appropriate char-
acteristic equation (z2 + az + b)(1 − ce−hz ) = 0 has all roots with positive real
parts:
1
zk = (ln c + i2kπ), i 2 = −1, k = 0, ±1, ±2, . . . .
h
The solutions of (1.20) were obtained via its difference analogue in the form
1.5 Routh–Hurwitz Conditions 25
 
xi+1 = (1 − a)xi + b xi+1−m − (1 − a)xi−m ,

where  is the step of discretization, and xi = x(i), m = h/,  = 0.01.

1.5 Routh–Hurwitz Conditions

Now we consider some important statements [80] for the stability of determinis-
tic system of linear autonomous differential equations that will be essentially used
below.
Linear autonomous system of ordinary differential equations has the general
form
ẋ(t) = Ax(t), (1.22)
where x ∈ Rn , A is an n × n matrix with real elements aij , i, j = 1, . . . , n.

Theorem 1.1 (Lyapunov theorem) The zero solution of (1.22) is asymptotically sta-
ble if and only if all roots λ of the characteristic equation

det(λI − A) = 0 (1.23)

(I is the n × n identity matrix) have negative real parts, i.e., Re λ < 0.

Definition 1.1 Let us define the trace of the kth order of a matrix A as follows:
 
  a i i . . . ai i 
 11 1 k
Sk =  . . . . . . . . .  , k = 1, . . . , n.
 
1≤i1 <···ik ≤n aik i1 . . . aik ik 


Here, in particular, S1 = Tr(A), Sn = det(A), Sn−1 = ni=1 Aii , where Aii is the
algebraic complement of the diagonal element aii of the matrix A.
Using the traces of the kth order, we can represent the characteristic equation
(1.23) in the form

λn − S1 λn−1 + S2 λn−2 − · · · + (−1)n Sn = 0. (1.24)

Besides, via Sk we can define the Hurwitz matrix as follows:

⎛ ⎞
−S1 −S3 −S5 . . . 0
⎜ 1 S2 S4 . . . 0 ⎟
⎜ ⎟
⎜ 0 −S −S . . . 0 ⎟
⎜ 1 3 ⎟. (1.25)
⎜ 0 1 S2 . . . 0 ⎟
⎜ ⎟
⎝ ... ... ... ... ... ⎠
... ... . . . . . . (−1)n Sn
26 1 Short Introduction to Stability Theory of Deterministic Functional

Theorem 1.2 (Routh–Hurwitz criterion) All roots λ of the characteristic equation

(1.24) have negative real parts if and only if
 
−S −S3 
1 = −S1 > 0, 2 =  1  > 0,
1 S2 
 
−S1 −S3 −S5 
 
3 =  1 S2 S4  > 0, . . . ,
 0 −S1 −S3 
  (1.26)
−S1 −S3 −S5 . . . 0 

 1 S2 S4 . . . 0 

 0 −S −S . . . 0 
n =  1 3
> 0.
 0 1 S 2 . . . 0 
 ... ... ... ... . . . 

 ... ... . . . . . . (−1)n Sn 

Definition 1.2 A symmetric n × n matrix Q is called a positive (negative) defi-

nite matrix (Q > 0) (Q < 0) if there exists c > 0 such that x  Qx ≥ c|x|2 (x  Qx ≤
−c|x|2 ) for all x ∈ Rn .

Theorem 1.3 (Lyapunov theorem) All roots λ of the characteristic equation (1.24)
have negative real parts if and only if for an arbitrary positive definite matrix Q, the
matrix equation
A P + P A = −Q (1.27)
has a positive definite solution P .

Corollary 1.1 Let A be a 2 × 2 matrix. Then the zero solution of (1.22) is asymp-
totically stable if and only if

Tr(A) < 0 and det(A) > 0. (1.28)

Proof It is enough to note that from (1.25) and (1.26) it follows that 1 = −S1 =
−Tr(A) > 0 and 2 = −S1 S2 = −Tr(A) det(A) > 0. 

Corollary 1.2 Let A be a 3 × 3 matrix. Then the zero solution of (1.22) is asymp-
totically stable if and only if

S1 < 0, S1 S2 < S3 < 0.

Proof It is enough to note that from (1.25) and (1.26) it follows that 1 = −S1 > 0,
2 = S3 − S1 S2 > 0, 3 = S3 (S1 S2 − S3 ) > 0. 

Remark 1.1 If
     
a a12 q q12 p p12
A = 11 , Q = 11 , P = 11 ,
a21 a22 q12 q22 p12 p22
1.5 Routh–Hurwitz Conditions 27

then
 
p11 a11 + p12 a21 p11 a12 + p12 a22
PA = ,
p12 a11 + p22 a21 p12 a12 + p22 a22
and therefore the matrix equation (1.27) can be represented in the form of the system
of the equations
2(p11 a11 + p12 a21 ) = −q11 ,
2(p12 a12 + p22 a22 ) = −q22 ,
p11 a12 + p12 Tr(A) + p22 a21 = −q12 ,
which, for arbitrary positive definite matrix Q, has a positive definite solution P
with the elements
2 + det(A))q + a 2 q − 2a a q
(a22 11 21 22 22 21 12
p11 = ,
2|Tr(A)| det(A)
2 + det(A))q + a 2 q − 2a a q
(a11 22 12 11 11 12 12
p22 = ,
2|Tr(A)| det(A)
a12 a22 q11 + a21 a11 q22 − 2a11 a22 q12
p12 = .
2Tr(A) det(A)
If, in particular, q11 = q > 0, q22 = 1, and q12 = 0, then
2 + det(A))q + a 2
(a22
p11 = 21
,
2|Tr(A)| det(A)
2 + det(A) + a 2 q
a11
p22 = 12
, (1.29)
2|Tr(A)| det(A)
a12 a22 q + a21 a11
p12 = .
2Tr(A) det(A)

Remark 1.2 In the general case the elements pij of the solution P of the matrix
equation (1.27) are defined as follows [21]:

1  (r)
n−1
pij = γij 1,r+1 ,
2n
r=0

where n is the determinant (1.26) of the Hurwitz matrix (1.25), 1,r+1 is the alge-
braic adjunct of the element of the first line and (r + 1)th column of the determinant
(r)
n , γij are defined by the identity


n 
n−1
(r)
(−1)n−1 qkm Dik (λ)Dj m (−λ) ≡ γij λ2(n−r−1) ,
k,m=1 r=0
28 1 Short Introduction to Stability Theory of Deterministic Functional

qkm are the elements of the matrix Q, and Dik (λ) are the algebraic adjuncts of the
determinant
 
a11 − λ . . . a1n 

D(λ) =  . . . ... . . . .  .
 an1 . . . ann − λ

Example 1.10 Consider the second-order scalar differential equation

ẍ(t) + a ẋ(t) + bx(t) = 0. (1.30)

 0 1  x1 = x, x2 = ẋ, represent (1.30) in the form of (1.22) with

Using the new variables
the matrix A = −b −a
. Since Tr(A) = −a, det(A) = b, then via Corollary 1.1 the
inequalities a > 0, b > 0 are necessary and sufficient conditions for the asymptotic
stability of the zero solution of (1.30).
Chapter 2
Stochastic Functional Differential Equations
and Procedure of Constructing Lyapunov
Functionals

2.1 Short Introduction to Stochastic Functional Differential

Equations
Here the basic notation of the theory of stochastic differential equations [16, 17, 81,
84–87, 210] is considered.
Let {Ω, F, P} be a probability space, {Ft , t ≥ 0} be a nondecreasing family of
sub-σ -algebras of F , i.e., Ft1 ⊂ Ft2 for t1 < t2 , P{·} be the probability of an event
enclosed in the braces, and E be the mathematical expectation.

2.1.1 Wiener Process and Its Numerical Simulation

The Wiener process sometimes is also called the Brownian motion process. Origi-
nally, the Brownian motion process was posed by the English botanist Robert Brown
as a model for the motion of a small particle immersed in a liquid and thus subject to
molecular collisions. The Brownian motion assumes a central role in the theory of
stochastic processes and statistics. It is basic to descriptions of financial markets, the
construction of a large class of Markov processes called diffusions, approximations
to many queuing models, and the calculation of asymptotic distributions in large
sample statistical estimation problems.

Definition 2.1 A stochastic process w(t) is called the standard Wiener process (rel-
atively to the family {Ft , t ≥ 0}) if it is Ft -measurable and
– w(0) = 0 (P-a.s.);
– w(t) is a process with stationary and mutually independent increments;
– the increments w(t) − w(s) have the normal distribution with
   2
E w(t) − w(s) = 0, E w(t) − w(s) = |t − s|;

– for almost all ω ∈ Ω, the functions w(t) = w(t, ω) are continuous on t ≥ 0 [84].

L. Shaikhet, Lyapunov Functionals and Stability of Stochastic Functional 29

Differential Equations, DOI 10.1007/978-3-319-00101-2_2,
© Springer International Publishing Switzerland 2013
30 2 Stochastic Functional Differential Equations and Procedure

We will also consider an m-dimensional Wiener process

 
w(t) = w1 (t), . . . , wm (t) ∈ Rm

where the components wi (t), i = 1, . . . , m, are mutually independent scalar Wiener

processes, and
    
E w(t) − w(s) = 0, E w(t) − w(s) w(t) − w(s) = I |t − s|.

Here I is the m × m identity matrix, and the prime denotes the transposition.
The trajectories of the Wiener process are nondifferentiable functions, although
formally the derivative of the Wiener process ẇ(t) is called the white noise.
There are different ways to get numerical simulation of trajectories of a Wiener
process. One of them is the following [244].
Let Yi , i = 1, . . . , n, be independent
√ random variables
√ that are uniformly
distributed on [0, 1]. Then Xi = 12(Yi − 0.5) = 3(2Yi − 1), i = 1, . . . , n,
are independent identically distributed random variables such that EXi = 0 and
Var(Xi ) = 1. Define the random walk Sn , n ≥ 0, by S0 = 0 and Sn = X1 + · · · + Xn
for n > 0. By the central limit theorem, √1n Sn converges in distribution to N (0, 1),
i.e., √1 Sn
n
→ N (0, 1).
Define the continuous-time process Wn (t) = √1n S[nt] , t ≥ 0, where [t] is the in-
teger part of t , i.e., the greatest integer less than or equal to t . Therefore, for any
t > 0, we have

[nt] S[nt]
Wn (t) = √ → N (0, t).
n [nt]
Also, for t > s, we obtain
[nt]
S[nt] − S[ns] j =[ns]+1 Xj S[nt]−[ns]
Wn (t) − Wn (s) = √ = √ = √
n n n

[nt] − [ns] S[nt]−[ns]
= √ → N (0, t − s).
n [nt] − [ns]
Since the process {Wn (t), t ≥ 0} is not continuous, let us modify it in the following
way:
S[nt]   X[nt]+1
Wn(c) (t) = √ + nt − [nt] √ , t ≥ 0.
n n
It is easy to see that EWn(c) (t) = 0 and
 
 (c)  nt − [nt] (nt − [nt])2
lim Var Wn (t) = lim t − + = t.
n→∞ n→∞ n n
(c)
So, as n → ∞, Wn (t) converges in distribution to the Wiener process w(t). This
(c)
means that for large enough n, the process Wn (t) approximates the Wiener process
2.1 Short Introduction to Stochastic Functional Differential Equations 31

Fig. 2.1 50 trajectories of the Wiener process

well enough. 50 trajectories of the Wiener process obtained via this algorithm are
shown in Fig. 2.1.

2.1.2 Itô Integral, Itô Stochastic Differential Equation, and Itô

Formula

Let H2 [0, T ] be the space of random functions f (t) that are defined and Ft -
measurable for each t ∈ [0, T ] and for which
 t
Ef 2 (s) ds < ∞.
0

Then for all functions from H2 [0, T ], the Itô integral with respect to the Wiener
process w(t)
 t
f (s) dw(s)
0
32 2 Stochastic Functional Differential Equations and Procedure

is defined and has the following properties:

 t  t 2  t
E f (s) dw(s) = 0, E f (s) dw(s) = Ef 2 (s) ds.
0 0 0

Let Hp , p > 0, be the space of F0 -adapted stochastic processes ϕ(θ ), θ ≤ 0, with

continuous trajectories that are independent on the σ -algebra B[0,∞) (dw), where
B[t1 ,t2 ] (dw) is the minimal σ -algebra generated by the random variables w(s) −
w(t) for arbitrary s, t : t1 ≤ t ≤ s ≤ t2 . In the space Hp two norms are defined:
p

ϕ
0 = sups≤0 |ϕ(s)| and
ϕ
1 = sups≤0 E|ϕ(s)|p .
We will consider the Itô stochastic functional differential equation [84]

dx(t) = a1 (t, xt ) dt + a2 (t, xt ) dw(t),

t ≥ 0, x ∈ Rn , (2.1)

with the initial condition

x0 = φ ∈ H p . (2.2)
Here x ∈ Rn , xt = x(t + s), s ≤ 0, w : [0, ∞) → Rm is the standard Wiener
process, the continuous functionals a1 (t, ϕ), a2 (t, ϕ) are defined on [0, ∞) × Hp ,
a1 ∈ Rn , a2 is an n × m-dimensional matrix. It is assumed also that the functionals
ai , i = 1, 2, satisfy the following conditions: ai (t, 0) ≡ 0, and for arbitrary functions
ϕ1 (θ ), ϕ2 (θ ) from Hp ,

∞
ai (t, ϕ1 ) − ai (t, ϕ2 ) 2 ≤ ϕ1 (−θ ) − ϕ2 (−θ ) 2 dRi (θ ),
0
 ∞
(2.3)
dRi (θ ) < ∞, i = 1, 2,
0

where Ri (θ ) are nondecreasing bounded functions.

A solution of problem (2.1)–(2.2) is a process x(t) such that x(θ ) = φ(θ) for
θ ≤ 0 and with probability 1
 t  t
x(t) = x(0) + a1 (s, xs ) ds + a2 (s, xs ) dw(s), t ≥ 0,
0 0
x(0) = φ(0).

The last integral is understood in the Itô sense.

Sometimes, instead of x(t), we will write x(t, φ) for the solution of (2.1) with
the initial function (2.2). Existence and uniqueness theorems for problem (2.1)–(2.2)
are considered in [84–87, 132, 133].
To calculate the stochastic differential of the process η(t) = u(t, x(t)), where
x(t) is a solution of problem (2.1)–(2.2), and the function u : [0, ∞) × Rn → R has
2.1 Short Introduction to Stochastic Functional Differential Equations 33

continuous partial derivatives

 
∂u(t, x) ∂u(t, x) ∂u(t, x)
ut = , ∇u = ,..., ,
∂t ∂x1 ∂xn
 2 
∂ u(t, x)
∇ u=
2
, i, j = 1, . . . , n,
∂xi ∂xj

the following Itô formula [84] is used:

   
dη(t) = Lu t, x(t) dt + ∇u t, x(t) a2 (t, xt ) dw(t). (2.4)

The operator L is called the generator of (2.1) and is defined in the following way:
     
Lu t, x(t) = ut t, x(t) + ∇u t, x(t) a1 (t, xt )
1
   
+ Tr a2 (t, xt )∇ 2 u t, x(t) a2 (t, xt ) , (2.5)
2
where Tr denotes the trace of a matrix.
The generator L can be applied also for some functionals V (t, ϕ) : [0, ∞) ×
Hp → R+ . Suppose that a functional V (t, ϕ) can be represented in the form
V (t, ϕ) = V (t, ϕ(0), ϕ(θ )), θ < 0, and for ϕ = xt , put
 
Vϕ (t, x) = V (t, ϕ) = V (t, xt ) = V t, x, x(t + θ ) ,
x = ϕ(0) = x(t), θ < 0. (2.6)

Denote by D the set of the functionals for which the function Vϕ (t, x) defined
by (2.6) has a continuous derivative with respect to t and two continuous deriva-
tives with respect to x. For functionals from D, the generator L of (2.1) has the
form
∂Vϕ (t, x(t))  
LV (t, xt ) = + ∇Vϕ t, x(t) a1 (t, xt )
∂t
1
   
+ Tr a2 (t, xt )∇ 2 Vϕ t, x(t) a2 (t, xt ) . (2.7)
2
From the Itô formula it follows that for functionals from D,
 t


E V (t, xt ) − V (s, xs ) = ELV (τ, xτ ) dτ, t ≥ s. (2.8)
s

Together with (2.1), we will also consider the stochastic differential equation of
neutral type [134]
 
d x(t) − G(t, xt ) = a1 (t, xt ) dt + a2 (t, xt ) dw(t), t ≥ 0,
x0 = φ ∈ H p , (2.9)
34 2 Stochastic Functional Differential Equations and Procedure

with the additional conditions on the functional G(t, ϕ): G(t, 0) = 0,

 ∞  ∞

G(t, ϕ) ≤ ϕ(−s) dK(s), dK(s) < 1. (2.10)
0 0

2.2 Stability of Stochastic Functional Differential Equations

2.2.1 Definitions of Stability and Basic Lyapunov-Type Theorems

Definition 2.2 The solution x(t) of (2.1) with the initial function (2.2) for some
p > 0 is called:
– Uniformly p-bounded if supt≥0 E|x(t)|p < ∞.
– Asymptoticallyp-trivial if limt→∞ E|x(t)|p = 0.
∞
– p-integrable if 0 E|x(t)|p dt < ∞.

Definition 2.3 The trivial solution of (2.1) for some p > 0 is called:
– p-stable if for each ε > 0, there exists δ > 0 such that E|x(t, φ)|p < ε, t ≥ 0,
p
provided that
φ
1 < δ.
– Asymptotically p-stable if it is p-stable and for each initial function φ, the solu-
tion x(t) of (2.1) is asymptotically p-trivial.
– Exponentially p-stable if it is p-stable and there exists λ > 0 such that for
each initial function φ, there exists C > 0 (which may depend on φ) such that
E|x(t, φ)|p ≤ Ce−λt for t > 0.
– Stable in probability if for any ε1 > 0 and ε2 > 0, there exists δ > 0 such that the
solution x(t, φ) of (2.1) satisfies the condition P{supt≥0 |x(t, φ)| > ε1 /F0 } < ε2
for any initial function φ such that P{
φ
0 < δ} = 1.
In particular, if p = 2, then the solution of (2.1) is called respectively mean-
square bounded, mean-square stable, asymptotically mean-square stable, and so on.

Definition 2.4 A nonnegative functional V (t, ϕ), defined on [0, ∞) × Hp , such

that V (t, 0) ≡ 0 and limt→0 EV (t, xt ) = 0 if limt→0 E|x(t)|p = 0, p > 0, is called
an Fp -functional.

Certain stability conditions for stochastic functional differential equations can

be stated in terms of Lyapunov functionals. In the sequel, ci are different positive
numbers.

Theorem 2.1 Let V : [t0 , ∞) × Hp → R+ be a continuous functional such that for

any solution x(t) of problem (2.1)–(2.2) and p ≥ 2, the following inequalities hold:
p
EV (t, xt ) ≥ c1 E x(t) , t ≥ 0, (2.11)
2.2 Stability of Stochastic Functional Differential Equations 35

p
EV (0, φ) ≤ c2
φ
1 , (2.12)
 t

 p
E V (t, xt ) − V (0, φ) ≤ −c3 E x(s) ds, t ≥ 0. (2.13)
0

Then the trivial solution of (2.1) is asymptotically p-stable.

Proof From (2.11)–(2.13) we have

p
c1 E x(t) ≤ EV (t, xt ) ≤ EV (0, φ) ≤ c2
φ
1 .
p
(2.14)

This proves the p-stability.

To prove the asymptotic p-stability, let us show that the solution of (2.1) is
asymptotically p-trivial for any initial function φ. Note that from (2.14) we obtain
p c 2
sup E x(t) ≤
φ
1 .
p
(2.15)
t≥0 c1

From (2.13) and (2.12) it follows that

 ∞
p 1 c2
E x(s) ds ≤ EV (0, φ) ≤
φ
1 < ∞,
p
(2.16)
0 c3 c3
i.e., the solution of (2.1) is p-integrable. Applying the generator L to the function
|x(t)|p via (2.5), we obtain
p p
p−2 
EL x(t) = 2E x(t) x (t)a1 (t, xt )
2
p−2
 
+ E x(t) Tr a2 (t, xt )a2 (t, xt )
p−4  
+ (p − 2)E x(t) x (t)a2 (t, xt ) 2 .

By the Hölder inequality, (2.3), and (2.15), there is a constant c4 such that
p−2 
2E x(t) x (t)a1 (t, xt )
p−2

≤ E x(t) x(t) 2 + a1 (t, xt ) 2
 ∞
p p−2
≤ E x(t) + E x(t) x(t − θ ) 2 dR1 (θ )
0

p ∞ p  p−2  p  2
≤ E x(t) + E x(t) p E x(t − θ ) p dR1 (θ )
0

≤ c4 .

Analogously,
p−2
 
E x(t) Tr a (t, xt )a2 (t, xt ) ≤ c4 ,
2
36 2 Stochastic Functional Differential Equations and Procedure
p−4 
E x(t) x (t)a2 (t, xt ) 2 ≤ c4 .

Hence, there exists a constant c5 such that |EL|x(t)|p | ≤ c5 , and using (2.8) for
t2 ≥ t1 ≥ 0, we obtain

E x(t2 ) p − E x(t1 ) p ≤ c5 (t2 − t1 ),

i.e., the function E|x(t)|p satisfies the Lipschitz condition. From this, (2.15), and
(2.16) it follows that limt→∞ E|x(t)|p = 0. The proof is completed. 

Remark 2.1 From (2.8) it follows that for the functional V ∈ D, condition (2.13) in
Theorem 2.1 follows from the inequality
p
ELV (t, xt ) ≤ −c3 E x(t) , t ≥ 0. (2.17)

Theorem 2.2 Let there exist a functional V (t, ϕ) ∈ D such that for any solution
x(t) of problem (2.1)–(2.2) and p ≥ 2, the following inequalities hold:
p
V (t, xt ) ≥ c1 x(t) , (2.18)
p
V (0, φ) ≤ c2
φ
0 , (2.19)
LV (t, xt ) ≤ 0, t ≥ 0, (2.20)

ci > 0, for any initial function φ such that P{

φ
0 ≤ δ} = 1, where δ > 0 is small
enough. Then the zero solution of (2.1) is stable in probability.

Proof Let us suppose that P{

φ
0 ≤ δ} = 1. From (2.20) it follows that the process
V (t, xt ) is a supermartingale. By (2.18), (2.19), and the inequality for supermartin-
gales [84–87] we have
  V (t , φ) c δ p
P sup x(t, φ) > ε1 /F0 ≤ P sup V (t, xt ) > c1 ε1 /F0 ≤
p 0 2
p ≤ p < ε2
t≥t0 t≥t0 c 1 ε1 c 1 ε1

for δ < ε1 (c1 ε2 /c2 )1/p . The theorem is proven. 

Theorem 2.3 Let there exist a functional W : [0, ∞) × H2 → R+ satisfying the

condition EW (t, ϕ) ≤ c1
ϕ
21 and such that for the functional
2
V (t, ϕ) = W (t, ϕ) + ϕ(0) − G(t, ϕ) , (2.21)

where G(t, ϕ) satisfies condition (2.10), the following estimates are valid:

EV (0, φ) ≤ c2
φ
21 ,
 t
2 (2.22)
EV (t, xt ) − EV (0, φ) ≤ −c3 E x(s) ds, t ≥ 0,
0

where ci , i = 1, 2, 3, are some positive constants. Then the zero solution of (2.9) is
asymptotically mean-square stable.
2.2 Stability of Stochastic Functional Differential Equations 37

The proof of Theorem 2.3 is similar to Theorem 2.1 and can be found
in [132–135].

Theorem 2.4 Let there exist a functional V (t, ϕ) ∈ D such that for some p > 0 and
λ > 0, the following conditions hold:
p
EV (t, xt ) ≥ c1 eλt E x(t) , t ≥ 0, (2.23)
p
EV (0, φ) ≤ c2
φ
1 , (2.24)
ELV (t, xt ) ≤ 0, t ≥ 0. (2.25)

Then the trivial solution of (2.1) is exponentially p-stable.

Proof Integrating (2.25) via (2.8), we obtain EV (t, xt ) ≤ EV (0, φ). From this and
from (2.23)–(2.24) it follows that
p
c1 E x(t) ≤ e−λt EV (0, φ) ≤ c2
φ
1 .
p

p
The inequality c1 E|x(t)|2 ≤ c2
φ
1 means that the trivial solution of (2.1) is p-
stable. Besides, from the inequality c1 E|x(t)|2 ≤ e−λt EV (0, φ) it follows that the
trivial solution of (2.1) is exponentially p-stable. The proof is completed. 

Corollary 2.1 Let there exist a functional V0 (t, ϕ) ∈ D such that for some p > 0,
the following conditions hold:
p
c1 E x(t) ≤ EV0 (t, xt )
 ∞ t
p m
p
≤ c2 E x(t) + (s − t + θ )i E x(s) ds dKi (θ ), (2.26)
i=0 0 t−θ
p
ELV0 (t, xt ) ≤ −c3 E x(t) , (2.27)

where m ≥ 0, Ki (θ ), i = 0, 1, . . . , m, are nondecreasing functions such that, for

some small enough λ > 0,


m  ∞
1
ηi (λ) < ∞, ηi (λ) = eλθ θ i+1 dKi (θ ). (2.28)
i +1 0
i=0

Then the trivial solution of (2.1) is exponentially p-stable.

Proof It is enough to show that by the conditions (2.26)–(2.28) there exists a func-
tional V (t, ϕ) that satisfies the conditions of Theorem 2.4. Indeed, put V1 (t, ϕ) =
eλt V0 (t, ϕ). By (2.26) and (2.27) we have
38 2 Stochastic Functional Differential Equations and Procedure

ELV1 (t, xt ) = λeλt EV0 (t, xt ) + eλt ELV0 (t, xt )

  m  ∞ t

p  p

≤ e λ c2 E x(t) +
λt i
(s − t + θ ) E x(s) ds dKi (θ )
i=0 0 t−θ

p
− c3 E x(t)

p
= eλt (λc2 − c3 )E x(t)

m 

 ∞ t p
+λ
(s − t + θ ) E x(s) ds dKi (θ ) .
i

i=0 0 t−θ

Now put

  ∞ t
m
1 p
V2 (t, xt ) = λ eλ(s+θ) (s − t + θ )i+1 x(s) ds dKi (θ ).
i +1 0 t−θ
i=0

Then
m 
 p
LV2 (t, xt ) = λ eλt ηi (λ) x(t)
i=0
 ∞ t 
p

i
− e λ(s+θ)
(s − t + θ ) x(s) ds dKi (θ )
0 t−θ
m 
 p
= λeλt ηi (λ) x(t)
i=0
 ∞ t 
p
− eλ(s−t+θ) (s − t + θ )i x(s) ds dKi (θ ) ,
0 t−θ

and via s ≥ t − θ , for the functional V = V1 + V2 , it follows that

   

m
p
ELV (t, xt ) ≤ eλt λ c2 + ηi (λ) − c3 E x(t)
i=0
m  ∞ t

  λ(s−t+θ)  p
−λ e − 1 E x(s) ds dKi θ
i=0 0 t−θ
   

m
p
≤ eλt λ c1 + ηi (λ) − c3 E x(t) .
i=0
2.2 Stability of Stochastic Functional Differential Equations 39

From this and from (2.28), for small enough λ > 0, we obtain (2.25). It is easy to
check that conditions (2.23)–(2.24) hold too. The proof is completed. 

Example 2.1 Consider the linear stochastic differential equation

ẋ(t) = ax(t) + bx(t − h) + σ x(t − τ )ẇ(t). (2.29)

For the functional

 t  t
V (t, xt ) = x (t) + |b|
2
x (s) ds + σ
2 2
x 2 (s) ds, (2.30)
t−h t−τ

we have
 
LV (t, xt ) = 2x(t) ax(t) + bx(t − h) + σ 2 x 2 (t − τ )
   
+ |b| x 2 (t) − x 2 (t − h) + σ 2 x 2 (t) − x 2 (t − τ )
   
≤ 2 a + |b| + σ 2 x 2 (t). (2.31)

So, by the condition

1
a + |b| + σ 2 < 0 (2.32)
2
the functional (2.30) satisfies the conditions of Theorem 2.1 with p = 2, and there-
fore the trivial solution of (2.29) is asymptotically mean-square stable.
On the other hand, by (2.31)–(2.32) the functional (2.30) satisfies the conditions
of Corollary 2.1 with p = 2, m = 0, and dK0 (s) = (|b|δ(s − h) + σ 2 δ(s − τ )) ds.
Therefore, the trivial solution of (2.29) is exponentially mean-square stable. By The-
orems 2.1–2.4 the construction of stability conditions is reduced to the construc-
tion of some Lyapunov functionals satisfying the assumptions of these theorems.
Below we will use the general method of construction of Lyapunov functionals
[136–139, 146–148, 150, 265, 269–272, 278] allowing one to obtain stability con-
ditions immediately in terms of parameters of systems under consideration. This
method was successfully used for functional differential equations, for difference
equations with discrete and continuous time [278], and for partial differential equa-
tions [48].

2.2.2 Formal Procedure of Constructing Lyapunov Functionals

The formal procedure of constructing Lyapunov functionals consists of four steps.

Step 1 Let us represent (2.9) in the form

       
dz(t, xt ) = b1 t, x(t) + c1 (t, xt ) dt + b2 t, x(t) + c2 (t, xt ) dw(t), (2.33)
40 2 Stochastic Functional Differential Equations and Procedure

where z(t, xt ) is some functional of xt , z(t, 0) = 0, the functionals bi (t, x(t)), i =

1, 2, depend on t and x(t) only and do not depend on the previous values x(t + s),
s < 0, of the solution, and bi (t, 0) = 0.

Step 2 Consider the auxiliary differential equation without memory

   
dy(t) = b1 t, y(t) dt + b2 t, y(t) dw(t). (2.34)

Let us assume that the zero solution of (2.34) is asymptotically mean-square stable
and therefore there exists a Lyapunov function v(t, y) such that c1 |y|2 ≤ v(t, y) ≤
c2 |y|2 and L0 v(t, y) ≤ −c3 |y|2 . Here L0 is the generator of (2.34), ci > 0, i =
1, 2, 3.

Step 3 Replacing the second argument y of the function v(t, y) by the functional
z(t, xt ) from left-hand part of (2.33), we obtain the main component V1 (t, xt ) =
v(t, z(t, xt )) of the functional V (t, xt ). Then it is necessary to calculate LV1 , where
L is the generator of (2.33), and in a reasonable way to estimate it.

Step 4 Usually, the functional V1 almost satisfies the requirements of stability

theorems. In order to satisfy these conditions completely, an auxiliary component
V2 can be easily chosen by some standard way. As a result, the desired functional
V (t, xt ) takes the form V = V1 + V2 .

Let us make remarks on some peculiarities of this procedure.

(1) It is clear that the representation (2.33) in the first step of the procedure is not
unique. Hence, for different representations, one can construct different Lya-
punov functionals, and, as a result, one can get different stability conditions.
(2) In the second step, for the auxiliary equation (2.34), one can choose different
Lyapunov functions v(t, y). So, for the initial equation (2.9), different Lyapunov
functionals can be constructed, and, as a result, different stability conditions can
be obtained.
(3) It is necessary to emphasize also that by choosing different ways of estimation
of LV1 one can construct different Lyapunov functionals and, as a result, one
can get different stability conditions.
(4) At last, some standard way of the construction of the additional functional V2
sometimes allows one to simplify the fourth step and do not use the functional
V2 at all. Below the corresponding auxiliary Lyapunov-type theorems are con-
sidered.

2.2.3 Auxiliary Lyapunov-Type Theorem

The following theorem in some cases allows one to use the procedure of construct-
ing Lyapunov functionals without an additional functional V2 .
2.2 Stability of Stochastic Functional Differential Equations 41

Theorem 2.5 Let there exist a functional V1 (t, xt ) ∈ D of type (2.21) such that

l 
 ∞

ELV1 (t, xt ) ≤ Ex (t)D(t)x(t) + Ex  (t − s)Si (t − s)x(t − s) dνi (s)
i=1 0
n 
 ∞
+ Exi2 (t − s) dKi (s)
i=1 0


k
     
+ Ex  t − τi (t) Qi t − τi (t) x t − τi (t)
i=1
m 
 ∞  t
+ dμj (s) (θ − t + s)j Ex  (θ )Rj (θ )x(θ ) dθ
j =0 0 t−s
 ∞  t
+ dμ(s) Ex  (τ )R(τ + s, t)x(τ ) dτ, (2.35)
0 t−s

where L is the generator of (2.9), D(t) is a negative definite matrix, Si (t), i =

1, . . . , l, Qi (t), i = 1, . . . , k, Rj (t), j = 0, . . . , m, R(s, t), s ≥ t ≥ 0, are non-
negative definite matrices, τi (t), i = 1, . . . , k, t ≥ 0, are differentiable nonnega-
tive functions with τ̇i (t) ≤ τ̂i < 1, Ki (s), i = 1, . . . , n, νi (s), i = 0, . . . , l, μj (s),
j = 0, . . . , m, and μ(s), s ≥ 0, are nondecreasing functions of bounded variation
such that
 ∞  ∞
ki = dKi (s) < ∞, qi = dνi (s) < ∞,
0 0
 ∞
(2.36)
s j +1
rj = dμj (s) < ∞,
0 j +1
and the matrix


l 
k
1
G(t) = D(t) + K + qi Si (t) + Qi (t)
1 − τ̂i
i=1 i=1


m  ∞  t+s
+ rj Rj (t) + dμ(s) R(t + s, θ ) dθ, (2.37)
j =0 0 t

where K is the diagonal matrix with elements ki , i = 1, . . . , n, is uniformly negative

definite matrix with respect to t ≥ 0 , i.e.,

x  G(t)x ≤ −c|x|2 , c > 0, x ∈ Rn . (2.38)

Then the zero solution of (2.9) is asymptotically mean-square stable.

42 2 Stochastic Functional Differential Equations and Procedure

Proof Put

l 
 ∞  t
V2 (t, xt ) = dνi (s) x  (θ )Si (θ )x(θ ) dθ
i=1 0 t−s

n 
 ∞  t 
k  t
1
+ dKi (s) xi2 (θ ) dθ + x  (s)Qi (s)x(s) ds
1 − τ̂i
i=1 0 t−s i=1 t−τi (t)

m  ∞
  t (θ − t + s)j +1 
+ dμj (s) x (θ )Rj (θ )x(θ ) dθ
j +1
j =0 0 t−s
 ∞  t  τ +s
+ dμ(s) x  (τ )R(τ + s, θ )x(τ ) dθ dτ.
0 t−s t

Then


l  ∞
ELV2 (t, xt ) = qi Ex  (t)Si (t)x(t) − Ex  (t − s)S(t − s)x(t − s) dν(s)
i=1 0


n n 
 ∞
+ ki Exi2 (t) − Exl2 (t − s) dKl (s)
i=1 l=1 0


k
1
+ Ex  (t)Qi (t)x(t)
1 − τ̂i
i=1


k
1 − τ̇i (t)      
− Ex  t − τi (t) Qi t − τi (t) x t − τi (t)
1 − τ̂i
i=1


m
+ rj Ex  (t)Rj (t)x(t)
j =0
m 
 ∞  t
− dμj (s) (θ − t + s)j Ex  (θ )Rj (θ )x(θ ) dθ
j =0 0 t−s
 ∞  t+s
+ dμ(s) Ex  (t)R(t + s, θ )x(t) dθ
0 t
 ∞  t
− dμ(s) Ex  (τ )R(τ + s, t)x(τ ) dτ. (2.39)
0 t−s

From (2.35), (2.37), and (2.39) for the functional V (t, xt ) = V1 (t, xt ) + V2 (t, xt ) it
follows that
ELV (t, xt ) ≤ Ex  (t)G(t)x(t). (2.40)
2.3 Some Useful Statements 43

By (2.38) and Remark 2.1 this means that the functional V (t, xt ) satisfies conditions
(2.22) of Theorem 2.3, and therefore the zero solution of (2.9) is asymptotically
mean-square stable. The proof is completed. 

Corollary 2.2 Let in Theorem 2.5 inequality (2.35) be the exact equality, G(t) =
G = const, and the functional V = V1 + V2 be F2 -functional. Then in the scalar
case the condition G < 0 is a necessary and sufficient condition for asymptotic
mean-square stability of the zero solution of (2.9).

Proof In the considered case, for the functional V = V1 +V2 , we have ELV (t, xt ) =
GEx 2 (t). If G ≥ 0, then from this and from (2.8) it follows that
 t
EV (t, xt ) = EV (0, φ) + G Ex 2 (τ ) dτ ≥ EV (0, φ) > 0.
0

This means that limt→∞ EV (t, xt ) = 0 and therefore limt→∞ E|x(t)|2 = 0. The
proof is completed. 

Remark 2.2 Since the functional V (t, xt ) constructed

 t in Theorem 2.5 satisfies the
conditions (2.22) and V (t, xt ) ≥ 0, we have that c3 0 E|x(s)|2 ds ≤ EV (0, φ) < ∞.
This means that by conditions (2.35) and (2.38) the solution of (2.9) is also mean-
square integrable.

Remark 2.3 In the scalar case, from Remark 2.2 it follows that if by condition (2.35)
∞
the solution of (2.9) is mean-square nonintegrable, i.e., 0 Ex 2 (t) dt = ∞, then
supt≥0 G(t) ≥ 0.

Remark 2.4 Theorem 2.5 is a useful development and improvement of the general
method of construction of Lyapunov functionals. It allows one not to use Step 4 of
the procedure and get good stability conditions using much more simple Lyapunov
functional than via Theorem 2.3. It can be used in different applications.

2.3 Some Useful Statements

2.3.1 Linear Stochastic Differential Equation

Consider now conditions for asymptotic mean-square stability of the trivial solution
of the linear Itô stochastic differential equation

ẋ(t) = Ax(t) + Bx(t − τ ) + σ x(t − h)ẇ(t), (2.41)

where A, B, σ , τ ≥ 0, h ≥ 0 are known constants.

44 2 Stochastic Functional Differential Equations and Procedure

Lemma 2.1 A necessary and sufficient condition for asymptotic mean-square sta-
bility of the trivial solution of (2.41) is
1
A + B < 0, G−1 > σ 2 , (2.42)
2
where
⎧ −1 √
⎪
⎪
Bq sin(qτ )−1
, B + |A| < 0, q = B 2 − A2 ,
⎪
⎨ A+B cos(qτ )
G = 1+|A|τ
2|A| , B = A < 0, (2.43)
⎪
⎪
⎪
⎩ Bq −1 sinh(qτ )−1 √
A+B cosh(qτ ) , A + |B| < 0, q = A2 − B 2 .

Remark 2.5 If A = −a and B = 0, then the necessary and sufficient stability condi-
tion (2.42)–(2.43) takes the form a > 12 σ 2 .

Note that the proof of Lemma 2.1 is based on two old enough papers [243, 290]
as it was shown briefly in the author recent book [278]. Following to advices and
requests of some readers of the book [278], the author took the decision to write
here the proof of this lemma in more detail.

Proof of the Lemma 2.1 A necessary and sufficient stability condition (2.42) with
 ∞
G=2 x 2 (s) ds, (2.44)
0

where x(t) is a solution of (2.41) in the deterministic case, i.e., with σ = 0, was
obtained in [243]. By the Plancherel theorem the integral (2.44) coincides [243, 290]
with

2 ∞ dt
G= . (2.45)
π 0 (A + B cos τ t)2 + (t + B sin τ t)2
Let us obtain for this integral the representation (2.43) in elementary functions.
Following [243], consider the functional
 t
1 2
V (xt ) = Gx (t) + β(s − t)x(s)x(t) ds
2 t−τ
 t  t
+ δ(s − t, θ − t)x(θ)x(s) dθ ds, (2.46)
t−τ s

where G is a constant, and β(s) and δ(s, θ ) are continuously differentiable func-
tions. By (2.46) and (2.41) with σ = 0 we obtain
  t 
dV (xt )  
= Gx(t) + β(s − t)x(s) ds Ax(t) + Bx(t − τ )
dt t−τ
 t
dβ(s − t)
+ β(0)x (t) − β(−τ )x(t − τ )x(t) −
2
x(s)x(t) ds
t−τ ds
2.3 Some Useful Statements 45
 t  t
+ δ(s − t, 0)x(t)x(s) ds − δ(−τ, θ − t)x(θ )x(t − τ ) dθ
t−τ t−τ
  t 
t ∂δ(s − t, θ − t) ∂δ(s − t, θ − t)
− + x(θ )x(s) dθ ds
t−τ s ∂s ∂θ
   
= GA + β(0) x 2 (t) + GB − β(−τ ) x(t)x(t − τ )
 t  
dβ(s − t)
+ Aβ(s − t) − + δ(0, s − t) x(s)x(t) ds
t−τ ds
 t
 
+ Bβ(s − t) − δ(−τ, s − t) x(s)x(t − τ ) ds
t−τ
  t 
t ∂δ(s − t, θ − t) ∂δ(s − t, θ − t)
− + x(θ )x(s) dθ ds.
t−τ s ∂s ∂θ
(2.47)

Let us suppose that the functions β(s) and δ(s, θ ) satisfy the conditions

GA + β(0) = −1, GB − β(−τ ) = 0,

dβ(s)
Aβ(s) − + δ(s, 0) = 0, Bβ(s) − δ(−τ, s) = 0, (2.48)
ds
∂δ(s, θ ) ∂δ(s, θ )
+ = 0.
∂s ∂θ

Then from (2.47) and (2.48) it follows that dVdt(xt ) = −x 2 (t), and therefore (if the
condition for asymptotic stability of the trivial solution of the considered equation
holds, i.e., limt→∞ V (xt ) = 0),
 ∞  ∞
dV (xt )
x (t) dt = −
2
dt = V (x0 ). (2.49)
0 0 dt
Using the initial function

0 if − τ ≤ s ≤ −ε,
xε (s) =
1+ s
ε if − ε ≤ s ≤ 0,

and the limit ε → 0, from (2.46) we obtain V (x0 ) = 12 G. From this and from (2.49)
it follows that G in the functional (2.46) indeed coincides with (2.44) and has the
representation (2.45).
To get for G the representation (2.43), let us solve system (2.48). From the last
equation of (2.48) it follows that δ(s, θ ) = ϕ(s − θ ) and, by the forth equation of
(2.48), ϕ(s) = Bβ(−τ − s). Substituting this ϕ(s) into the third equation of (2.48),
we obtain that the function β(t) is defined by the differential equation

β̇(t) = Aβ(t) + Bβ(−t − τ ) (2.50)

46 2 Stochastic Functional Differential Equations and Procedure

with the conditions

GA + β(0) + 1 = 0, GB − β(−τ ) = 0. (2.51)

Suppose that q 2 = B 2 − A2 > 0. Then, by (2.50),

β̈(t) = Aβ̇(t) − B β̇(−t − τ )

   
= A Aβ(t) + Bβ(−t − τ ) − B Aβ(−t − τ ) + Bβ(t + τ − τ )
= −q 2 β(t)

or
β̈(t) + q 2 β(t) = 0. (2.52)
Substituting the general solution β(t) = C1 cos qt + C2 sin qt of (2.52) into (2.50)
and (2.51), we obtain two equations for G, C1 , and C2

GA + C1 = −1, GB = C1 cos qτ − C2 sin qτ, (2.53)

and two homogeneous linear dependent equations for C1 and C2

C1 (A + B cos qτ ) − C2 (q + B sin qτ ) = 0,
(2.54)
C1 (q − B sin qτ ) + C2 (A − B cos qτ ) = 0.

By (2.54) we have
A + B cos qτ q − B sin qτ
C2 = C1 = −C1 . (2.55)
q + B sin qτ A − B cos qτ
Substituting the first equality (2.55) into (2.53) and excluding C1 , we obtain
A sin qτ − q cos qτ
G= . (2.56)
q(A cos qτ + q sin qτ + B)
Multiplying the numerator and the denominator of the obtained fraction by
B sin qτ − q, one can convert (2.56) to the form of the first line in (2.43). Note
that the same result can be obtained using the second equality (2.55).
Suppose now that q 2 = A2 − B 2 > 0. Then, similarly to (2.52), we obtain the
equation β̈(t) − q 2 β(t) = 0 with the general solution β(t) = C1 eqt + C2 e−qt . Sub-
stituting this solution into (2.50) and (2.51), similarly to (2.53) and (2.54), we have

GA + C1 + C2 = −1, GB = C1 e−qτ + C2 eqτ ,

(2.57)
C1 (q − A) − C2 Beqτ = 0, C1 B + C2 (q + A)eqτ = 0.

By the two last equations of (2.57),

q − A −qτ B
C2 = C1 e = −C1 e−qτ . (2.58)
B q +A
2.3 Some Useful Statements 47

From the first equality of (2.58) and the two first equations of (2.57) we obtain
q − A + Be−qτ
G= . (2.59)
q(q − A − Be−qτ )

Put now sinh x = 12 (ex − e−x ) and cosh x = 12 (ex + e−x ) (respectively, hyperbolic
sine and hyperbolic cosine). Multiplying the numerator and the denominator of
(2.59) by B sinh qτ − q in the denominator, we have
 
q − A − Be−qτ (B sinh qτ − q)
  
= q − A − Be−qτ B cosh qτ − Be−qτ − q
= qB cosh qτ − AB cosh qτ − B 2 e−qτ cosh qτ
− Bqe−qτ + ABe−qτ + B 2 e−2qτ − A2 + B 2 + Aq + Bqe−qτ
  
= (q − A)(A + B cosh qτ ) + Be−qτ A + B eqτ + e−qτ − cosh qτ
 
= q − A + Be−qτ (A + B cosh qτ ).
As a result, we obtain (2.59) in the form of the third line in (2.43). Note also that the
same result can be obtained using the second equality of (2.58).
The second line of (2.43) can be obtained from the first (or the third) line in the
limit as q → 0. The proof is completed. 

2.3.2 System of Two Linear Stochastic Differential Equations

Consider the system of two stochastic differential equations without delays

ẋ1 (t) = a11 x1 (t) + a12 x2 (t) + σ1 x1 (t)ẇ1 (t),
(2.60)
ẋ2 (t) = a21 x1 (t) + a22 x2 (t) + σ2 x2 (t)ẇ2 (t),
where aij , σi , i, j = 1, 2, are constants, and w1 (t) and w2 (t) are mutually indepen-
dent standard Wiener processes.
Put A =
aij
, i, j = 1, 2, and
1
δi = σi2 , i = 1, 2. (2.61)
2
Remark 2.6 If σ1 = σ2 = 0, then (by Corollary 1.1) the trivial solution of (2.60) is
asymptotically stable if and only if
Tr(A) = a11 + a22 < 0, det(A) = a11 a22 − a12 a21 > 0. (2.62)
If a12 = a21 = 0, then (by Remark 2.5) the necessary and sufficient conditions for
asymptotic mean-square stability of the trivial solution of (2.60) are
a11 + δ1 < 0, a22 + δ2 < 0. (2.63)
48 2 Stochastic Functional Differential Equations and Procedure

Lemma 2.2 Let for some positive definite matrix P =

pij
, i, j = 1, 2, the param-
eters of system (2.60) satisfy the conditions

p12 a21 + p11 (a11 + δ1 ) < 0,

p12 a12 + p22 (a22 + δ2 ) < 0,
   (2.64)
4 p12 a21 + p11 (a11 + δ1 ) p12 a12 + p22 (a22 + δ2 )
 2
> p11 a12 + p22 a21 + p12 Tr(A) .

Then the trivial solution of system (2.60) is asymptotically mean-square stable.

Proof Let L0 be the generator of system (2.60). Using the Lyapunov function

v(t) = p11 x12 (t) + 2p12 x1 (t)x2 (t) + p22 x22 (t) (2.65)

and (2.61)–(2.62) for system (2.60), we have

  
L0 v(t) = 2 p11 x1 (t) + p12 x2 (t) a11 x1 (t) + a12 x2 (t) + p11 σ12 x12 (t)
  
+ 2 p12 x1 (t) + p22 x2 (t) a21 x1 (t) + a22 x2 (t) + p22 σ22 x22 (t)
   
= 2 p12 a21 + p11 (a11 + δ1 ) x12 (t) + 2 p12 a12 + p22 (a22 + δ2 ) x22 (t)
 
+ 2 p11 a12 + p22 a21 + p12 Tr(A) x1 (t)x2 (t).

By (2.64) L0 v(t) is a negative definite square form, i.e., the function v(t) satisfies
(2.17) with p = 2. So, the trivial solution of system (2.60) is asymptotically mean-
square stable. The proof is completed. 

Corollary 2.3 Suppose that conditions (2.62) hold, a12 = 0, and

| Tr(A)| det(A) | Tr(A)| det(A) − A2 δ1
δ1 < , δ2 < , (2.66)
A2 A1 − | Tr(A)|δ1
where
A1 = det(A) + a11
2
, A2 = det(A) + a22
2
. (2.67)
Then the trivial solution of system (2.60) is asymptotically mean-square stable.

Proof By Remark 1.1, from (1.29) and (2.64) it follows that if, for some q > 0,

−q + 2p11 δ1 < 0, −1 + 2p22 δ2 < 0, (2.68)

then the trivial solution of (2.60) is asymptotically mean-square stable.

Using (1.29), we can represent (2.68) in the form

(A2 q + a21
2 )δ
1 (A1 + a122 q)δ
2
< q, < 1.
| Tr(A)| det(A) | Tr(A)| det(A)
2.3 Some Useful Statements 49

From this we have

2 δ
a21 1 | Tr(A)| det(A) − A1 δ2
<q < . (2.69)
| Tr(A)| det(A) − A2 δ1 2 δ
a12 2

So, if
2 δ
a21 1 | Tr(A)| det(A) − A1 δ2
< , (2.70)
| Tr(A)| det(A) − A2 δ1 2 δ
a12 2

then there exists q > 0 such that (2.69), and therefore (2.68) holds.
Let us show that (2.70) holds. Indeed, by the first condition (2.66) we can rewrite
(2.70) in the form
  
2 2
a12 a21 δ1 δ2 < Tr(A) det(A) − A2 δ1 Tr(A) det(A) − A1 δ2
 2
= Tr(A) det(A) − Tr(A) det(A)(A1 δ2 + A2 δ1 ) + A1 A2 δ1 δ2 .
(2.71)

By (2.67) we have
  
A1 A2 = det(A) + a11
2
det(A) + a22
2

 
= det(A) + a11
2
+ a22
2
det(A) + a11
2 2
a22
 
= Tr(A) − (a11 a22 + a12 a21 ) det(A) + a11
2 2 2
a22
2
= Tr(A) det(A) + a12
2 2
a21
2
≥ Tr(A) det(A). (2.72)

So, by (2.71) and (2.72) it is enough to show that

 
0 < Tr(A) det(A) − A2 δ1 − A1 − Tr(A) δ1 δ2 . (2.73)

Note that from (2.66) and (2.72) it follows that

| Tr(A)| det(A) A1
δ1 < ≤ .
A2 | Tr(A)|

So, (2.73) is equivalent to (2.66). The proof is completed. 

Remark 2.7 If a12 = 0, then conditions (2.66) coincide with (2.63). Indeed, by
(2.62) from (2.66) we obtain

|a11 + a22 |a11 a22 |a11 + a22 |a22 (a11 + δ1 )

δ1 < = −a11 , δ2 < = −a22 .
(a11 + a22 )a22 (a11 + a22 )(a11 + δ1 )
50 2 Stochastic Functional Differential Equations and Procedure

Remark 2.8 From the conditions (2.63) and a12 a21 ≤ 0 it follows that

| Tr(A)| det(A) | Tr(A)| det(A)

|a11 | ≤ , |a22 | ≤ .
A2 A1

So, from the conditions (2.63) and a12 a21 ≤ 0 it follows that

| Tr(A)| det(A) | Tr(A)| det(A)

δ1 < , δ2 < .
A2 A1

Corollary 2.4 Suppose that the parameters of system (2.60) satisfy the condi-
tions (2.62),

a21 > 0, A2 > Tr(A) δ2 , (2.74)

and the intervals

 
a12 (a22 + δ2 ) a11 + δ1
I1 = − ,− (2.75)
A2 − | Tr(A)|δ2 a21
and
   
| Tr(A)| −
(a11 − a22 )2 + 4 det(A) | Tr(A)| + (a11 − a22 )2 + 4 det(A)
I2 = ,
2a21 2a21
(2.76)
have common points. Then the trivial solution of system (2.60) is asymptotically
mean-square stable.

Proof Consider the function v(t) given by (2.65) with p11 = 1, p12 = μ, p22 = γ ,
−1
where γ = a21 (μ| Tr(A)| − a12 ). From (2.64) it follows that μ ∈ I1 . On the other
hand, the function v(t) is positive definite if and only if γ > μ2 , which is equivalent
to μ ∈ I2 . So, the appropriate μ exists if and only if the intervals I1 and I2 have
common points. The proof is completed. 

2.3.3 Some Useful Inequalities

Lemma 2.3 For arbitrary vectors a ∈ Rn , b ∈ Rn and an n × n matrix R > 0, it

follows that

a  b + b a ≤ a  Ra + b R −1 b.

Proof The proof follows from the simple equality

   
0 ≤ a − R −1 b R a − R −1 b = a  Ra + b R −1 b − a  b − b a. 
2.4 Some Unsolved Problems 51

Lemma 2.4 For positive P2 , x and nonnegative P1 , Q such that P2 > Qx, the
following inequality holds:
 2 
P1 + Qx −1 Q + P1 P2 + Q 2
≥ .
P2 − Qx P2

Proof It is enough to check that the function

P1 + Qx −1
f (x) =
P2 − Qx
reaches its minimum at the point
P2
x0 = 
Q2 + P1 P2 + Q

and this minimum equals x0−2 . The proof is completed. 

2.4 Some Unsolved Problems

In spite of the fact that the theory of stability for stochastic hereditary systems is
very popular in researches, there are simply and clearly formulated problems with
unknown decisions. In order to attract attention to such problems, one of them for
stochastic difference equation with continuous time is represented in [277], and two
unsolved stability problems for stochastic differential equations with delay are de-
scribed below.

2.4.1 Problem 1

Consider the linear stochastic differential equation with delays


m
ẋ(t) = Ax(t) + Bi x(t − τi ) + σ x(t − h)ẇ(t), (2.77)
i=1

where A, Bi , σ , τi > 0, h ≥ 0 are known constants, and w(t) is the standard Wiener
process.
It is known [290] that a necessary and sufficient condition for asymptotic mean-
square stability of the zero solution of (2.77) can be represented in the form
 ∞
σ2 2 dt
G−1 > , G= m  .
2 π 0 (A + i=1 Bi cos τi t)2+ (t + m
i=1 Bi sin τi t)
2
(2.78)
52 2 Stochastic Functional Differential Equations and Procedure

By Lemma 2.1, in the particular case m = 1, B1 = B, τ1 = τ the integral (2.78)

can be calculated in elementary functions, and the stability conditions take the form
(2.42)–(2.43).
The problem is: to calculate the integral (2.78) in elementary functions for m ≥ 2,
in particular, for m = 2.

2.4.2 Problem 2

From (2.42) and (2.43) it follows that the zero solution of the differential equation
with a constant delay ẋ(t) = −bx(t − h) is asymptotically stable if and only if
π
0 < bh < . (2.79)
2
It is also known [221, 318] that the zero solution of the differential equation with a
varying delay ẋ(t) = −bx(t − τ (t)) is asymptotically stable for an arbitrary delay
τ (t) such that τ (t) ∈ [0, h] if and only if
3
0 < bh < . (2.80)
2
Consider the stochastic differential equation with a constant delay

ẋ(t) = −bx(t − h) + σ x(t)ẇ(t). (2.81)

From (2.42) and (2.43) it follows that the zero solution of (2.81) is asymptotically
mean-square stable if and only if

b2 − p 2 σ2
0 < bh < arcsin , p= . (2.82)
b2 + p 2 2
In the deterministic case (σ = 0) condition (2.82) coincides with (2.79).
Consider the stochastic differential equation
 
ẋ(t) = −bx t − τ (t) + σ x(t)ẇ(t) (2.83)

with a varying delay τ (t) such that τ (t) ∈ [0, h].

The problem is: to generalize condition (2.80) for (2.83).
Chapter 3
Stability of Linear Scalar Equations

3.1 Linear Stochastic Differential Equation of Neutral Type

Using the proposed procedure of Lyapunov functionals construction, let us inves-
tigate asymptotic mean-square stability of the linear scalar stochastic differential
equation of neutral type with constant coefficients

ẋ(t) + ax(t) + bx(t − h) + cẋ(t − h) + σ x(t − τ )ẇ(t) = 0, |c| < 1,

  (3.1)
x(s) = φ(s), s ∈ − max(h, τ ), 0 .

3.1.1 The First Way of Constructing a Lyapunov Functional

Following Step 1 of the procedure, we rewrite (3.1) in the form

ż(xt ) = −ax(t) − bx(t − h) − σ x(t − τ )ẇ(t),

z(xt ) = x(t) + cx(t − h).

Suppose that a > 0. Then the function v = y 2 is a Lyapunov function for the auxil-
iary differential equation ẏ(t) = −ay(t) since v̇ = −2ay 2 . Thus, the trivial solution
of the auxiliary differential equation is asymptotically stable. Put V1 = z2 (xt ). Then
 
LV1 = 2z(xt ) −ax(t) − bx(t − h) + σ 2 x 2 (t − τ )
= −2ax 2 (t) − 2bcx 2 (t − h) − 2(ac + b)x(t)x(t − h) + σ 2 x 2 (t − τ )
 
≤ −2a + |ac + b| x 2 (t) + ρx 2 (t − h) + σ 2 x 2 (t − τ ),

where

0 if |ac + b| ≤ 2bc,
ρ= (3.2)
|ac + b| − 2bc if |ac + b| > 2bc.

L. Shaikhet, Lyapunov Functionals and Stability of Stochastic Functional 53

Differential Equations, DOI 10.1007/978-3-319-00101-2_3,
© Springer International Publishing Switzerland 2013
54 3 Stability of Linear Scalar Equations

By Theorem 2.5 with S = Qi = Rj = R(s, t) = 0 and

 
D = −2a + |ac + b|, dK1 (s) = ρδ(s − h) + σ 2 δ(s − τ ) ds

(here and below δ(s) is the Dirac function) we obtain that if σ 2 + ρ + |ac + b| < 2a,
then the trivial solution of (3.1) is asymptotically mean-square stable.
Note that, by (3.2), ρ ≥ 0. So, from the obtained condition it follows that a > 0.
Using two representations for ρ in (3.2), we obtain the following two stability con-
ditions:
|ac + b| ≤ 2bc, σ 2 + |ac + b| < 2a (3.3)
and
σ2
|ac + b| > 2bc, p + |ac + b| − bc < a, p= . (3.4)
2
By (3.3) we have bc = |bc| and a > 0. So, |ac + b| = a|c| + |b|, and inequalities
(3.3) take the forms 2|bc| ≥ a|c| + |b| and σ 2 + a|c| + |b| < 2a. The former is wrong
if 2|c| < 1. Suppose that 2|c| ≥ 1. Then
 
σ 2 + |b| 1
<a≤ 2− |b|,
2 − |c| |c|

which is impossible since from this the contradiction σ 2 |c| + 2|b|(1 − |c|)2 < 0
follows. Thus, inequalities (3.3) are incompatible.
Consider now conditions (3.4). Suppose first that bc ≥ 0. From this and from
a > 0 we have bc = |bc| and |ac + b| = a|c| + |b|, and inequalities (3.4) take the
forms
p
2|bc| < a|c| + |b|, a > |b| + .
1 − |c|
If 2|c| < 1, then the first inequality holds for a > 0 and arbitrary b. If 2|c| ≥ 1, then
the second inequality implies the first one. So, if bc ≥ 0, then from (3.4) it follows
that
p
bc ≥ 0, a > |b| + . (3.5)
1 − |c|
Let now bc < 0. Then the first inequality in (3.4) holds, and (3.4) takes the form

bc < 0, p + |ac + b| − bc < a. (3.6)

Since bc < 0, |ac + b| = |a|c| − |b||. So, if a|c| ≥ |b|, then from (3.6) we have
p
− a < |b| ≤ a|c|. (3.7)
1 − |c|
If a|c| < |b|, then
p
a|c| < |b| < a − . (3.8)
1 + |c|
3.1 Linear Stochastic Differential Equation of Neutral Type 55

Fig. 3.1 Stability regions for

(3.1) given by conditions
(3.11) for c = −0.5, h = 1
and different values of p:
(1) p = 0, (2) p = 0.5,
(3) p = 1, (4) p = 1.5

Combining (3.7) and (3.8), we obtain

p p
bc < 0, − a < |b| < a − . (3.9)
1 − |c| 1 + |c|

Note that since bc < 0, the system

p p
|b| = − a, |b| = a −
1 − |c| 1 + |c|

has the solution

p pc
a= , b=− . (3.10)
1 − c2 1 − c2

Combining (3.5), (3.9), and (3.10), we obtain the stability condition in the form
p pc
1−c + b, b > − 1−c 2,
a> p pc (3.11)
1+c − b, b ≤ − 1−c 2.

Thus, if condition (3.11) holds, then the trivial solution of (3.1) is asymptotically
mean-square stable.
The stability regions for (3.1), given by the stability conditions (3.11), are shown
in Fig. 3.1 for c = −0.5, h = 1, and different values of p: (1) p = 0; (2) p = 0.5;
(3) p = 1; (4) p = 1.5. In Fig. 3.2 the stability regions are shown for c = 0.5 and
the same values of the other parameters.
56 3 Stability of Linear Scalar Equations

Fig. 3.2 Stability regions for

(3.1) given by conditions
(3.11) for c = 0.5, h = 1 and
different values of p:
(1) p = 0, (2) p = 0.5,
(3) p = 1, (4) p = 1.5

3.1.2 The Second Way of Constructing a Lyapunov Functional

To get another stability condition, rewrite (3.1) in the form

ż(xt ) = − (a + b)x(t) − σ x(t − τ )ẇ(t),

t
z(xt ) =x(t) + cx(t − h) − b x(s) ds.
t−h

By condition (2.10) it is necessary to suppose that

|c| + |b|h < 1. (3.12)

Suppose also that a + b > 0. Then the function v = y 2 is a Lyapunov function for the
auxiliary differential equation ẏ(t) = −(a + b)y(t) since v̇ = −2(a + b)y 2 . Thus,
the trivial solution of the auxiliary differential equation is asymptotically stable. Put
V1 = z2 (xt ). Then

LV1 = −2(a + b)x(t)z(xt ) + σ 2 x 2 (t − τ )

= −2(a + b)x 2 (t) − 2(a + b)cx(t)x(t − h)

t
+ 2(a + b)b x(t)x(s) ds + σ 2 x 2 (t − τ )
t−h
 
≤ (a + b) −2 + |c| + |b|h x 2 (t) + σ 2 x 2 (t − τ )

t 
+ (a + b) |c|x (t − h) + |b|
2 2
x (s) ds .
t−h
3.1 Linear Stochastic Differential Equation of Neutral Type 57

Fig. 3.3 Stability regions for

(3.1) given by conditions
(3.13) for |c| = 0.5, h = 0.2,
and different values of p:
(1) p = 0.2, (2) p = 0.6,
(3) p = 1, (4) p = 1.4

Thus, using (3.12) and Theorem 2.4 with m = S = Qi = R(s, t) = 0, n = 1 and

 
D = (a + b) −2 + |c| + |b|h , R0 (θ ) = (a + b)|b|, dμ0 (s) = δ(s − h) ds,
 
dK1 (s) = (a + b)|c|δ(s − h) + σ 2 δ(s − τ ) ds,

we obtain the stability condition in the form

p 1 − |c|
a> − b, |b| < . (3.13)
1 − |c| − |b|h h

The stability regions for (3.1), given by the stability condition (3.13), are shown
in Fig. 3.3 for |c| = 0.5, h = 0.2, and different values of p: (1) p = 0.2, (2) p = 0.6,
(3) p = 1, (4) p = 1.4 and in Fig. 3.4 for |c| = 0.5, p = 0.4, and different values of
h: (1) h = 0.1, (2) h = 0.15, (3) h = 0.2, (4) h = 0.25.
It is easy to see that for b ≤ 0, condition (3.11) is better than (3.13). So, condition
(3.13) is better to use for b > 0 only in the form

p 1 − |c|
a> − b, 0<b< . (3.14)
1 − |c| − bh h

The stability regions for (3.1), given by both stability conditions (3.11) and
(3.14), are shown in Fig. 3.5 for c = −0.6, p = 0.4, and different values of h:
(1) h = 0.05; (2) h = 0.1; (3) h = 0.15, (4) h = 0.2. In Fig. 3.6 the stability re-
gions are shown for c = 0.6 and the same values of the other parameters.
58 3 Stability of Linear Scalar Equations

Fig. 3.4 Stability regions for

(3.1) given by conditions
(3.13) for |c| = 0.5, p = 0.4,
and different values of h:
(1) h = 0.1, (2) h = 0.15,
(3) h = 0.2, (4) h = 0.25

Fig. 3.5 Stability regions for

(3.1) given by conditions
(3.11) and (3.14) together for
c = −0.6, p = 0.4, and
different values of h:
(1) h = 0.05, (2) h = 0.1,
(3) h = 0.15, (4) h = 0.2

3.1.3 Some Particular Cases

(1) Note that, as h → 0, condition (3.14) takes the form

p
a> − b, b > 0. (3.15)
1 − |c|

On the other hand, in the case h = 0, for the functional


t 
V = (1 + c) x +
2 2
x (s) ds ,
t−τ
3.1 Linear Stochastic Differential Equation of Neutral Type 59

Fig. 3.6 Stability regions for

(3.1) given by conditions
(3.11) and (3.14) together for
c = 0.6, p = 0.4, and
different values of h:
(1) h = 0.05, (2) h = 0.1,
(3) h = 0.15, (4) h = 0.2

we have LV = [−2(a + b)(1 + c) + σ 2 ]x 2 (t). So, by Corollary 2.2 for h = 0 a nec-

essary and sufficient condition for asymptotic mean-square stability of the trivial
solution of (3.1) is
p
a> − b. (3.16)
1+c
For b > 0 and c > 0, condition (3.15) is essentially worse than (3.16), but for b > 0
and c ≤ 0, condition (3.15) coincides with (3.16). The second condition in (3.11)
coincides with (3.16) as well.
(2) For c = 0, the necessary and sufficient condition for asymptotic mean-square
stability of the trivial solution of (3.1) has the form (Lemma 2.1)

a + b > 0, G−1 > p, (3.17)

where
⎧ 1+bq −1 sin(qh) √
⎪
⎪ a+b cos(qh) , b > |a|, q = b2 − a 2 ,
⎪
⎨
σ2
p= , G = 1+ah
2a , b = a > 0, (3.18)
2 ⎪
⎪
⎪ 1+bq −1 sinh(qh)
⎩ √
a+b cosh(qh) , a > |b|, q = a 2 − b2 .

If, in addition, a = 0, then the necessary and sufficient stability condition (3.17)–
(3.18) takes the form
1 b2 − p 2
h< arcsin 2 .
b b + p2
In Fig. 3.7 the stability regions for (3.1) given by the sufficient conditions (3.11),
(3.14) and the necessary and sufficient conditions (3.17)–(3.18) are shown for c = 0,
h = 1, and different values of p: (1) p = 0; (2) p = 0.5; (3) p = 1; (2) p = 1.5;
60 3 Stability of Linear Scalar Equations

Fig. 3.7 Stability regions for

(3.1) given by sufficient
conditions (3.11), (3.14) and
necessary and sufficient
conditions (3.17), (3.18) are
shown for c = 0, h = 1 and
different values of p:
(1) p = 0, (2) p = 0.5,
(3) p = 1, (4) p = 1.5,
(5) p = 2

Fig. 3.8 Stability regions for

(3.1) given by sufficient
conditions (3.11), (3.14) and
necessary and sufficient
conditions (3.17), (3.18) are
shown for c = 0, h = 1, and
different values of p:
(1) p = 0, (2) p = 0.5,
(3) p = 1, (4) p = 1.5,
(5) p = 2

(5) p = 2. In Fig. 3.8 the same stability regions are shown in another scale. We
can see that the sufficient conditions (3.11), (3.14) give the stability region that is
sufficiently close to the exact one.
(3) For p = 0, from (3.11), (3.14) it follows that

b, b ≥ 1−|c|
h ,
a> 1−|c|
(3.19)
−b, b < h .

In Fig. 3.9 the exact stability regions for (3.1) are shown for p = 0, h = 1, and
(1) c = 0.5; (2) c = −0.5 and the stability region given by the sufficient condition
(3.19) for (3) c = |0.5|. In Fig. 3.10 the similar regions are shown for (1) c = 0.85;
(2) c = −0.85; (3) c = |0.85|.
3.1 Linear Stochastic Differential Equation of Neutral Type 61

Fig. 3.9 Exact stability regions are shown for p = 0, h = 1, and (1) c = 0.5, (2) c = −0.5, and
stability region, given by sufficient condition (3.19) for (3) c = |0.5|

Fig. 3.10 Exact stability regions are shown for p = 0, h = 1, and (1) c = 0.85, (2) c = −0.85, and
stability region, given by sufficient condition (3.19) for (3) c = |0.85|
62 3 Stability of Linear Scalar Equations

3.2 Linear Differential Equation with Two Delays in

Deterministic Part

Consider the scalar differential equation

ẋ(t) + ax(t − h1 ) + bx(t − h2 ) + σ x(t − τ )ẇ(t) = 0, 0 ≤ h1 < h2 ,

  (3.20)
x(s) = φ(s), s ∈ − max(h2 , τ ), 0 .

Using different representations of this equation, we will obtain different conditions

for asymptotic mean-square stability of the zero solution of (3.20).

3.2.1 The First Way of Constructing a Lyapunov Functional

Rewrite (3.20) in the form

ż(xt ) = −ax(t) − bx(t − h2 ) − σ x(t − τ )ẇ(t),

t
z(xt ) = x(t) − a x(s) ds.
t−h1

Putting V1 = z2 (xt ), we obtain

 
LV1 = 2z(xt ) −ax(t) − bx(t − h2 ) + σ 2 x 2 (t − τ )
= −2ax 2 (t) − 2bx(t)x(t − h2 )

t
t
+ 2a 2 x(t)x(s) ds + 2ab x(t − h2 )x(s) ds + σ 2 x 2 (t − τ )
t−h1 t−h1


 2  t
≤ −2ax (t) + |b| x (t) + x (t − h2 ) + a x 2 (t)h1 +
2 2 2 2
x (s) ds
t−h1

t 
+ |ab| x 2 (t − h2 )h1 + x (s) ds + σ 2 x 2 (t − τ )
2
t−h1

    t
= −2a + a 2 h1 + |b| x 2 (t) + a 2 + |ab| x 2 (s) ds
t−h1
 
+ |b| + |ab|h1 x 2 (t − h2 ) + σ 2 x 2 (t − τ ).

From this and from (2.10) and Theorem 2.5 with m = S = Qi = R(s, t) = 0,
n = 1 and

D = −2a + a 2 h1 + |b|, R0 (θ ) = a 2 + |ab|, dμ0 (s) = δ(s − h1 ) ds,

  
dK1 (s) = |b| + |ab|h1 δ(s − h2 ) + σ 2 δ(s − τ ) ds
3.2 Linear Differential Equation with Two Delays in Deterministic Part 63

it follows that if
a(1 − ah1 ) − p 1
|b| < , 0<a< , (3.21)
1 + ah1 h1
then the trivial solution of (3.20) is asymptotically mean-square stable.

Remark 3.1 Rewriting (3.20) in the form

ż(xt ) = −bx(t) − ax(t − h1 ) − σ x(t − τ )ẇ(t),

t
z(t) = x(t) − b x(s) ds
t−h2

and using symmetry, we obtain another sufficient condition for asymptotic mean-
square stability of the trivial solution of (3.20):

b(1 − bh2 ) − p 1
|a| < , 0<b< . (3.22)
1 + bh2 h2

3.2.2 The Second Way of Constructing a Lyapunov Functional

Rewrite (3.20) in the form

ż(xt ) = −(a + b)x(t) + bx(t − h1 ) − bx(t − h2 ) − σ x(t − τ )ẇ(t),

t
z(xt ) = x(t) − (a + b) x(s) ds.
t−h1

Putting V1 = z2 (xt ), we obtain

 
LV1 = 2z(xt ) −(a + b)x(t) + bx(t − h1 ) − bx(t − h2 ) + σ 2 x 2 (t − τ )
= −2(a + b)x 2 (t) + 2bx(t)x(t − h1 ) − 2bx(t)x(t − h2 )

t
t
+ 2(a + b)2
x(t)x(s) ds − 2(a + b)b x(t − h1 )x(s) ds
t−h1 t−h1

t
+ 2(a + b)b x(t − h2 )x(s) ds + σ 2 x 2 (t − τ )
t−h1
   
≤ −2(a + b)x 2 (t) + |b| x 2 (t) + x 2 (t − h1 ) + |b| x 2 (t) + x 2 (t − h2 )

t 
+ (a + b) x (t)h1 +
2 2
x (s) ds + σ 2 x 2 (t − τ )
2
t−h1


  2 t
 
+ (a + b)b x (t − h1 )h1 + x (t − h2 )h1 + 2
2 2
x (s) ds
t−h1
64 3 Stability of Linear Scalar Equations
 
= −2(a + b) + (a + b)2 h1 + 2|b| x 2 (t) + σ 2 x 2 (t − τ )
  
+ |b| 1 + |a + b|h1 x 2 (t − h1 ) + x 2 (t − h2 )

   t
+ (a + b)2 + 2b(a + b) x 2 (s) ds.
t−h1

Thus, we obtain the representation (2.35) with m = S = Qi = R(s, t) = 0, n = 1

and

D = −2(a + b) + (a + b)2 h1 + 2|b|,

  
R0 (θ ) = (a + b)2 + 2b(a + b) , dμ0 (s) = δ(s − h1 ) ds,
    
dK1 (s) = |b| 1 + |a + b|h1 δ(s − h1 ) + δ(s − h2 ) + σ 2 δ(s − τ ) ds.

From this and from Theorem 2.5 it follows that if

(a + b)(1 − (a + b)h1 ) − p 1
2|b| < , 0<a+b< , (3.23)
1 + (a + b)h1 h1
then the trivial solution of (3.20) is asymptotically mean-square stable.

Remark 3.2 Similarly to Remark 3.1, using symmetry, we obtain another sufficient
condition for asymptotic mean-square stability of the trivial solution of (3.20):

(a + b)(1 − (a + b)h2 ) − p 1
2|a| < , 0<a+b< . (3.24)
1 + (a + b)h2 h2

3.2.3 The Third Way of Constructing a Lyapunov Functional

Rewrite (3.20) in the form

ż(xt ) = −(a + b)x(t) − σ x(t − τ )ẇ(t),

t
t (3.25)
z(xt ) = x(t) − a x(s) ds − b x(s) ds.
t−h1 t−h2

Putting V1 = z2 (xt ), we obtain

 
LV1 = 2z(xt ) −(a + b) x(t) + σ 2 x 2 (t − τ )

t
= −2(a + b)x (t) + 2a(a + b)
2
x(t)x(s) ds
t−h1

t
+ 2b(a + b) x(t)x(s) ds + σ 2 x 2 (t − τ ). (3.26)
t−h2
3.2 Linear Differential Equation with Two Delays in Deterministic Part 65

Therefore,


  t
LV1 ≤ −2(a + b)x 2 (t) + a(a + b) x 2 (t)h1 + x 2 (s) ds
t−h1


  t
+ b(a + b) x 2 (t)h2 + x 2 (s) ds + σ 2 x 2 (t − τ )
t−h2
  
= −2(a + b) + |a + b| |a|h1 + |b|h2 x 2 (t) + σ 2 x 2 (t − τ )

t
t 
+ |a + b| |a| x 2 (s) ds + |b| x 2 (s) ds .
t−h1 t−h2

Thus, we obtain the representation (2.35) with m = S = Qi = 0, n = 1 and

 
D = −2(a + b) + |a + b| |a|h1 + |b|h2 , dK1 (s) = σ 2 δ(s − τ ) ds,
 
R0 (θ ) = a(a + b), dμ0 (s) = δ(s − h1 ) ds,
 
R(s, t) = b(a + b), dμ(s) = δ(s − h2 ) ds.

From this and from Theorem 2.5 it follows that if

p
a+b> , |a|h1 + |b|h2 < 1, (3.27)
1 − |a|h1 − |b|h2

then the trivial solution of (3.20) is asymptotically mean-square stable.

Remark 3.3 Note that condition (3.27) follows from condition (3.22), condi-
tion (3.22) follows from condition (3.24), and condition (3.21) follows from con-
dition (3.23).
Let us show, for instance, that condition (3.27) follows from (3.22). Indeed, by
(3.22) we have (p + |a|)b−1 + (|a| + b)h2 < 1. So, using h1 < h2 , we obtain |a|h1 +
bh2 < (|a| + b)h2 < 1. To get the first condition (3.27), rewrite (3.22) and (3.27) as
follows: p < b(1 − bh2 ) − |a|(1 + bh2 ), p < (a + b)(1 − |a|h1 − |b|h2 ). So, it is
enough to show that b(1 − bh2 ) − |a|(1 + bh2 ) ≤ (a + b)(1 − |a|h1 − |b|h2 ) or
|a|(a + b)h1 + abh2 ≤ a + |a| + |a|bh2 . For a = 0, this inequality holds. If a > 0,
then it is equivalent to (a + b)h1 ≤ 2. Using h1 < h2 and ah1 + bh2 < 1, we have
(a + b)h1 < ah1 + bh2 < 1 < 2. Let now a < 0. Then it is necessary to show that
(a + b)h1 ≤ 2bh2 . But this follows from (a + b)h1 < bh1 < bh2 < 2bh2 .

3.2.4 The Fourth Way of Constructing a Lyapunov Functional

Let us show that using the same representations of the initial equation but different
ways of LV1 estimation, we can get different stability conditions.
66 3 Stability of Linear Scalar Equations

Rewrite (3.20) in the form (3.25) and put V1 = z2 (xt ) again. Using (3.26) and the
condition a + b > 0, let us estimate LV1 in the following way:

t
LV1 = −2(a + b)x (t) + 2(a + b)
2 2
x(t)x(s) ds
t−h1

t−h1
+ 2(a + b)b x(t)x(s) ds + σ 2 x 2 (t − τ )
t−h2

t 
≤ −2(a + b)x (t) + (a + b) h1 x (t) +
2 2 2 2
x (s) ds
t−h1

t−h1 
+ (a + b)|b| (h2 − h1 )x 2 (t) + x 2 (s) ds + σ 2 x 2 (t − τ )
t−h2
 
= −2(a + b) + (a + b)2 h1 + (a + b)|b|(h2 − h1 ) x 2 (t) + σ 2 x 2 (t − τ )

t
t−h1
+ (a + b)2 x 2 (s) ds + (a + b)|b| x 2 (s) ds.
t−h1 t−h2

From representation (2.35) with m = S = Qi = 0, n = 1 and

D = −2(a + b) + (a + b)2 h1 + (a + b)|b|(h2 − h1 ), dK1 (s) = σ 2 δ(s − τ ) ds,

R0 (θ ) = (a + b)2 , dμ0 (s) = δ(s − h1 ) ds, dμ(s) = δ(s − h2 ) ds,

(a + b)|b|, τ ∈ [t − h2 , t − h1 ],
R(τ + h2 , t) =
0, τ ∈ (t − h1 , t].

and from Theorem 2.5 it follows that if

p
a+b> , (a + b)h1 + |b|(h2 − h1 ) < 1, (3.28)
1 − (a + b)h1 − |b|(h2 − h1 )
then the trivial solution of (3.20) is asymptotically mean-square stable.
It is easy to see that condition (3.28) coincides with (3.27) for a ≥ 0, b ≥ 0, but
in the case ab < 0 condition (3.28) is weaker than (3.27). Using Remark 3.3, we
can conclude that the stability conditions (3.21) and (3.28) together are better than
all other.
Put p = 0 and h2 = 1. In Fig. 3.11 the stability regions for (3.20), given by
conditions (3.21)–(3.24) and (3.27) (with numbers 1–5, respectively), are shown for
h1 = 0.1. In Fig. 3.12 the similar picture is shown for h1 = 0.2 with addition of
the bound of the stability region, given by the necessary and sufficient condition
of asymptotic stability obtained by the characteristic equation (Example 1.3). In
Fig. 3.13 we can see how the picture in Fig. 3.12 is changed for h1 = 0.25.
Consider now the case p > 0, h2 = 1. In Fig. 3.14 the stability regions, given
by conditions (3.21)–(3.24), (3.27), and (3.28) (with numbers 1–6, respectively),
are shown for p = 0.1, h1 = 0.1. In Fig. 3.15 these stability regions are shown for
p = 0.1, h1 = 0.5, and in Fig. 3.16 for p = 0.25, h1 = 0.5.
3.2 Linear Differential Equation with Two Delays in Deterministic Part 67

Fig. 3.11 Stability regions for (3.20), given by conditions (3.21)–(3.24), (3.27) (with numbers
1–5, respectively), are shown for p = 0, h1 = 0.1, h2 = 1

Fig. 3.12 Stability regions for (3.20), given by conditions (3.21)–(3.24), (3.27) that are similar to
Fig. 3.11, are shown for p = 0, h1 = 0.2, h2 = 1 with addition of the bound of stability region,
given by the necessary and sufficient condition
68 3 Stability of Linear Scalar Equations

Fig. 3.13 Here it is shown how Fig. 3.12 is changed for h1 = 0.25

Fig. 3.14 Stability regions for (3.20), given by conditions (3.21)–(3.24), (3.27), (3.28) (with num-
bers 1–6, respectively), are shown for p = 0.1, h1 = 0.1, h2 = 1
3.2 Linear Differential Equation with Two Delays in Deterministic Part 69

Fig. 3.15 Stability regions for (3.20), given by conditions (3.21)–(3.24), (3.27), (3.28) (with num-
bers 1–6, respectively), are shown for p = 0.1, h1 = 0.5, h2 = 1

Fig. 3.16 Stability regions for (3.20), given by conditions (3.21)–(3.24), (3.27), (3.28) (with num-
bers 1–6, respectively), are shown for p = 0.25, h1 = 0.5, h2 = 1
70 3 Stability of Linear Scalar Equations

3.2.5 One Generalization for Equation with n Delays

Consider the stochastic differential equation with n delays in the deterministic part


n
ẋ(t) + a0 x(t) + ai x(t − hi ) + σ x(t − τ )ẇ(t) = 0,
i=1 (3.29)
 
x(s) = φ(s), s ∈ − max(hn , τ ), 0 .

Here it is supposed that

0 < h1 < h2 < · · · < hn . (3.30)
For V1 = x 2 (t), we have
 

n
LV1 = −2x(t) a0 x(t) + ai x(t − hi ) + σ 2 x 2 (t − τ )
i=1
 

n 
n
≤ −2a0 + |ai | x 2 (t) + |ai |x 2 (t − hi ) + σ 2 x 2 (t − τ ).
i=1 i=1

By the representation (2.35) and Theorem 2.5 we obtain the sufficient condition
for asymptotic mean-square stability of the trivial solution of (3.29) in the form


n
σ2
a0 > |ai | + p, p= . (3.31)
2
i=1

Rewrite (3.29) as follows:

ż(xt ) = −S0 x(t) − σ x(t − τ )ẇ(t),

 n
t
z(xt ) = x(t) − ai x(s) ds,
i=1 t−hi (3.32)

n
Sj = ai , j = 0, 1, . . . , n.
i=j

Suppose that S0 > 0 and put V1 = z2 (xt ). Then


n
t
LV1 = −2S0 x 2 (t) + 2S0 x(t) ai x(s) ds + σ 2 x 2 (t − τ )
i=1 t−hi


n 
t 
≤ −2S0 x (t) + S0
2
|ai | x (t)hi +
2
x (s) ds + σ 2 x 2 (t − τ )
2

i=1 t−hi
3.2 Linear Differential Equation with Two Delays in Deterministic Part 71
 


n 
n t
= −2 + |ai |hi S0 x (t) + S0
2
|ai | x 2 (s) ds + σ 2 x 2 (t − τ ).
i=1 i=1 t−hi

Using Theorem 2.5 with m = S = Qi = Rj = 0 and

 

n
D = −2 + |ai |hi S0 , dK1 (s) = σ 2 δ(s − τ ) ds,
i=1


n
dμ(s) = δ(s − hi ) ds, R(s, t) = S0 |ai |,
i=1

we obtain the sufficient condition for asymptotic mean-square stability of the trivial
solution of (3.29) in the form
 −1

n 
n
σ2
S0 > p 1 − |ai |hi , |ai |hi < 1, p = . (3.33)
2
i=1 i=1

This condition is a generalization of condition (3.27).

Let us obtain a generalization of condition (3.28). Note that by (3.30) and h0 = 0
the process z(t) from (3.32) can be represented as follows:


n i−1

 t−hj 
n−1
t−hj
z(xt ) = x(t) − ai x(s) ds = x(t) − Sj +1 x(s) ds. (3.34)
i=1 j =0 t−hj +1 j =0 t−hj +1

Using the representation (3.34), the functional V1 = z2 (xt ), and S0 > 0, we get


n−1
t−hj
LV1 = −2S0 x 2 (t) + 2S0 x(t) Sj +1 x(s) ds + σ 2 x 2 (t − τ )
j =0 t−hj +1


n−1 
t−hj 
≤ −2S0 x 2 (t) + S0 |Sj +1 | (hj +1 − hj )x 2 (t) + x 2 (s) ds
j =0 t−hj +1

+ σ 2 x 2 (t − τ )
 

n−1
= −2 + |Sj +1 |(hj +1 − hj ) S0 x 2 (t)
j =0


n−1
t−hj
+ S0 |Sj +1 | x 2 (s) ds + σ 2 x 2 (t − τ ).
j =0 t−hj +1
72 3 Stability of Linear Scalar Equations

By Theorem 2.5 with m = S = Qi = 0 and

 

n−1
D = −2 + |Sj +1 |(hj +1 − hj ) S0 , dK1 (s) = σ 2 δ(s − τ ) ds,
j =0


n−1
dμ(s) = δ(s − hj +1 ) ds,
j =0

S0 |Sj +1 |, τ ∈ [t − hj +1 , t − hj ],
R(τ + hj +1 , t) =
0, τ ∈ (t − hj , t].

from this we obtain a generalization of condition (3.28) in the form

 −1

n−1 
n−1
S0 > p 1 − |Sj +1 |(hj +1 − hj ) , |Sj +1 |(hj +1 − hj ) < 1. (3.35)
j =0 j =0

To prove that condition (3.35) is weaker than (3.33), it is enough to show that


n−1 
n
|Sj +1 |(hj +1 − hj ) ≤ |ai |hi .
j =0 i=1

Using h0 = 0, rewrite this inequality in the form


n−1 
n 
n−1 
n
|Sj +1 |hj +1 = |Sj |hj ≤ |Sj +1 |hj + |ai |hi .
j =0 j =1 j =1 i=1

Now it is enough to note that |Sj | ≤ |Sj +1 | + |aj |, j = 1, . . . , n − 1, and Sn = an .

3.3 Linear Differential Equation of nth Order

3.3.1 Case n > 1

Consider the scalar stochastic differential equation of the nth order

n

 ∞
x (n) (t) = x (j −1) (t − s) dKj (s) + σ x(t − τ )ẇ(t), t ≥ 0, (3.36)
j =1 0

d j x(t)
where x (j ) (t) = dt j
, j = 1, . . . , n. The initial conditions for (3.36) have the form

x (j ) (θ ) = φ (j ) (θ ), θ ≤ 0, (3.37)
3.3 Linear Differential Equation of nth Order 73

where φ(θ ) is a given n − 1 times continuously differentiable function. The kernels

Kj (s) are functions of bounded variation on [0, ∞) such that

∞  
αij = s i dKj (s) < ∞, 0 ≤ i ≤ n, 1 ≤ j ≤ n. (3.38)
0

Put xi (t) = x (i−1) (t) and transform (3.36) to the system of stochastic differential
equations
ẋi (t) = xi+1 (t), i = 1, . . . , n − 1,
n
∞
(3.39)
ẋn (t) = xj (t − s) dKj (s) + σ x1 (t − τ )ẇ(t).
j =1 0

Put also

∞
βij = s i dKj (s) (3.40)
0
and note that for m = 1, . . . , n, we have

∞
xn (t − s) dKm (s)
0

∞
t 
d
= xn (t)β0m − dKm (s) xn (θ ) dθ . (3.41)
dt 0 t−s

Similarly, using (3.40), it is easy to check that for i = 1, . . . , n − 1,

∞
xn−i (t − s) dKm (s)
0


i+1
βj −1,m
= (−1)j −1 xn−i+j −1 (t)
(j − 1)!
j =1

∞

d t (θ − t + s)i
+ (−1)i+1 dKm (s) xn (θ ) dθ . (3.42)
dt 0 t−s i!

Putting


n−1
∞
t (θ − t + s)l
z(xt ) = (−1) l+1
dKn−l (s) xn (θ ) dθ,
0 t−s l!
l=0
(3.43)

n−1
i−l βi−l,n−i
al = (−1) , l = 0, 1, . . . , n − 1,
(i − l)!
i=l

by (3.42) and (3.43) we obtain

74 3 Stability of Linear Scalar Equations

 ∞ n−1

 ∞
xj (t − s) dKj (s) = xn−i (t − s) dKn−i (s)
j =1 0 i=0 0


n−1 
i+1
βj −1,n−i
= (−1)j −1 xn−i+j −1 (t) + ż(xt )
(j − 1)!
i=0 j =1


n−1 
i
βi−l,n−i
= (−1)i−l xn−l (t) + ż(xt )
(i − l)!
i=0 l=0


n−1
= al xn−l (t) + ż(xt ). (3.44)
l=0

Following the procedure of constructing Lyapunov functionals and using (3.44),

rewrite (3.39) as follows:
ẋi (t) = xi+1 (t), i = 1, . . . , n − 1,

d  n−1 (3.45)
xn (t) − z(xt ) = al xn−l (t) + σ x1 (t − τ )ẇ(t).
dt
l=0

For (3.45), consider the auxiliary system of the form


n−1
ẏi (t) = yi+1 (t), i = 1, . . . , n − 1, ẏn (t) = al yn−l (t). (3.46)
l=0

Let y = (y1 , . . . , yn ) , and A be the n × n matrix

⎛ ⎞
0 1 0 ... 0
⎜ 0 0 1 ... 0 ⎟
⎜ ⎟
A=⎜ ⎜ ... ... ... . . . . . .⎟
⎟. (3.47)
⎝ 0 0 0 ... 1 ⎠
an−1 an−2 an−3 . . . a0

By (3.46) and (3.47) we have ẏ = Ay.

Assume that the trivial solution of (3.46) is asymptotically stable. From Theo-
rem 1.3 it follows that for the matrix (3.47) and an arbitrary positive definite matrix
Q, the matrix equation (1.27) has a unique positive definite solution P .
Consider the Lyapunov function for (3.46) of the form v(y) = y  P y. Because of
(1.25), we have v̇(y) = −y  Qy. According to the procedure of constructing Lya-
punov functionals, we consider the functional
   
V1 (t, xt ) = x1 (t), . . . , xn−1 (t), xn (t) − z(xt ) P x1 (t), . . . , xn−1 (t), xn (t) − z(xt ) .
(3.48)
Let Q be a diagonal matrix with positive entries ql , l = 1, . . . , n. From (3.49) it
follows that LV1 (t, xt ) with respect to (3.45) equals
3.3 Linear Differential Equation of nth Order 75


n 
n
LV1 (t, xt ) = − ql xl2 (t) − 2 z(xt )(P A)nl xl (t) + pnn σ 2 x12 (t − τ )
l=1 l=1


n 
n
 
≤− ql xl2 (t) + 2 βl z(xt )xl (t) + pnn σ 2 x12 (t − τ ), (3.49)
l=1 l=1

where (P A)nl is (nl)th element of the matrix P A, βl = |(P A)nl |, and pnn = (P )nn .
Put also

 n−1

 ∞

n−1
αj +1,n−j  t (θ − t + s)j
α= , W (t, xt ) = dKn−j (s) xn2 (θ ) dθ,
(j + 1)! j!
j =0 j =0 0 t−s
(3.50)
and suppose that α > 0. Then using (3.43) and some positive numbers γl , l =
1, . . . , n, we obtain

   n−1 ∞  t   (θ − t + s)j
2z(xt )xl (t) ≤ dKn−j (s) γl xl2 (t) + γl−1 xn2 (θ ) dθ
j!
j =0 0 t−s

= αγl xl2 (t) + γl−1 W (t, xt ). (3.51)

By (3.51) from (3.49) it follows that


n 
n
LV1 (t, xt ) ≤ (αβl γl − ql )xl2 (t) + pnn σ 2 x12 (t − τ ) + βl γl−1 W (xt ). (3.52)
l=1 l=1

From this and from (3.50) the representation of the type of (2.35) follows, where
S(t) = Qi (t) = R(s, t) = 0, D is a diagonal matrix with the elements dll = αβl γl −
ql , dK1 (s) = pnn σ 2 δ(s − τ ) ds, dKl (s) = 0, l = 2, . . . , n, m = n − 1, dμj (s) =
|dKn−j (s)|, and Rj (s) is the matrix with all zero elements except for (Rj )nn =
1 n −1
j! l=1 βl γl .
So, by (2.37) the matrix G is the diagonal matrix with

G11 = αβ1 γ1 + pnn σ 2 − q1 , Gll = αβl γl − ql , l = 2, . . . , n − 1,

   n−1

1 βl
Gnn = α βn γn + + − qn .
γn γl
l=1

It is easy to see that Gnn reaches its minimum with respect to γn at γn = 1. Besides,
from the matrix equation (1.27) with the matrix A given by (3.47) it follows that
β1 = |an−1 |pnn , 2βn = qn . So, we can conclude that if there exist positive numbers
γ1 , γ2 , . . . , γn−1 such that
76 3 Stability of Linear Scalar Equations
 
1 q1 σ2 ql
γ1 < − , γl < , l = 2, . . . , n − 1,
α β1 |an−1 | αβl

n−1   (3.53)
βl 1
< − 1 qn , α < 1,
γl α
l=1
then the matrix G is a negative definite one, and therefore the zero solution of (3.36)
is asymptotically mean-square stable.
Let us rewrite inequalities (3.53) in the form
 −1
q1 σ2 1 αβl 1
0<α − < , < , l = 2, . . . , n − 1,
β1 |an−1 | γ1 ql γl
  (3.54)
β1  βl
n−1
1
+ < − 1 qn , α < 1.
γ1 γl α
l=2
From the system of inequalities (3.54) it follows that
 −1   
β 1  βl
n−1 2 n−1
q1 σ2 βl 1
αβ1 − +α < + < − 1 qn . (3.55)
β1 |an−1 | ql γ1 γl α
l=2 l=2
So, if the condition
 −1 
n−1 2  
q1 σ2 β 1
αβ1 − +α l
< − 1 qn
β1 |an−1 | ql α
l=2
or
  
n−1 2
q1 β1 β 1 1
σ < |an−1 |
2
−  , l
< Θqn , Θ= − ,
β1 Θqn − n−1 2 −1 ql α 2 α
l=2 βl ql l=2
(3.56)
holds, then there exist positive numbers γ1 , γ2 , . . . , γn−1 such that (3.55) also holds
and the zero solution of (3.36) is asymptotically mean-square stable.
So, we have proven

Theorem 3.1 Let there exist a diagonal matrix Q with positive entries q1 , . . . , qn
such that the matrix equation (1.27) has a positive definite solution P that satisfies
inequalities (3.56). Then the zero solution of (3.36) is asymptotically mean-square
stable.

Remark 3.4 Without the loss of generality, we may assume that qn = 1. Otherwise,
the matrix equation (1.27) can be divided by qn . Thus, all elements of the matrices
Q and P will be divided by qn .

Remark 3.5 Note that condition (3.56) is also correct without the assumption α > 0.
Indeed, if α = 0 (this means also that z(xt ) ≡ 0), then from (3.56) we have Θ =
∞ and σ 2 < |an−1 |q1 /β1 , which also follows immediately from (3.49) and β1 =
|an−1 |pnn .
3.3 Linear Differential Equation of nth Order 77

Remark 3.6 The stability condition obtained in Theorem 3.1 uses the representation
(3.42) where the integrals in the right-hand side depend only on xn for all i. Follow-
ing the same procedure, one can try to obtain other stability conditions using the
representations where the right-hand side depends on xm for m ≤ n. For example,
for n = 2, we have

∞
∞
d
x1 (t − s) dK1 (s) = β01 x1 (t) − β11 x2 (t) + dK1 (s)
0 dt 0

t
× (τ − t + s)x2 (τ ) dτ,
t−s

∞
∞
t (3.57)
d
xi (t − s) dKi (s) = β0i xi (t) − dKi (s) xi (τ ) dτ, i = 1, 2,
0 dt 0 t−s

∞
∞
d
x2 (t − s) dK2 (s) = x1 (t − s) dK2 (s).
0 dt 0

3.3.2 Some Particular Cases

It is easy to see that the stability condition (3.56) is the best one for those q1 , . . . , qn
for which the right-hand part of inequality (3.56) reaches its maximum. Let us con-
sider some particular cases of condition (3.56) in which it can be formulated imme-
diately in the terms of the parameters of the considered equation (3.36).
Let be n = 1. In this case, (3.36) has the form

∞
ẋ(t) = x(t − s) dK(s) + σ x(t − τ )ẇ(t), t ≥ 0. (3.58)
0

Condition (3.56) cannot be used immediately since it was obtained for n > 1. So,
note that for the functional V1 (t, xt ) = x 2 (t), similarly to (3.52) (for γ1 = q1 = 1),
we have

∞

  t 2
LV1 (t, xt ) ≤ (−1 + αβ1 )x 2 (t) + p11 σ 2 x 2 (t − τ ) + β1 dK(s) x (θ ) dθ,
0 t−s

where

∞   ∞
α = α11 = s dK(s), a0 = β01 = dK(s) < 0,
0 0
1 1 1
p11 = − = > 0, β1 = |p11 a0 | = .
2a0 2|β01 | 2
The stability condition for (3.58) takes the form σ 2 < 2|a0 |(1 − α).
If, in particular, dK(s) = (−aδ(s) − bδ(s − h)) ds, then α = |b|h, a0 = −a − b <
0, and the stability condition takes the form σ 2 < 2(a + b)(1 − |b|h). Note that the
last condition also follows immediately from (3.13) for c = 0.
78 3 Stability of Linear Scalar Equations

Let be n = 2. In this case, (3.36) has the form

∞
∞
ẍ(t) = x(t − s) dK1 (s) + ẋ(t − s) dK2 (s) + σ x(t − τ )ẇ(t), t ≥ 0.
0 0
(3.59)
Following Remark 3.4, we will consider the matrix equation (1.25) with
     
0 1 p11 p12 q 0
A= , P= , Q= , q > 0.
a 1 a0 p12 p22 0 1

By (3.43), a0 = β02 − β11 , a1 = β01 , where βij are defined by (3.40). By Corol-
lary 1.1 the matrix equation (1.27) has a positive definite solution if and only if
a0 < 0, a1 < 0. From (1.29) it follows that the last element of the matrix P is
p22 = (q + |a1 |)(2a0 a1 )−1 . Since β1 = |a1 |p22 = (q + |a1 |)(2|a0 |)−1 , the stabil-
ity condition (3.56) takes the form
 
q|a0 | (q + |a1 |)α 2
σ < 2|a1 |
2
− , (3.60)
q + |a1 | 4|a0 |(1 − α)

where α = α12 + 12 α21 < 1, αij are defined by (3.38).

√
The right-hand part of (3.60) reaches its maximum at q = 2|a0 |α −1 (1 − α)|a1 |
− |a1 |. So, as a result, we obtain a sufficient condition for asymptotic mean-square
stability of the zero solution of (3.58) of the form
  
|a1 |
σ 2 < 2|a1 | |a0 | − α , α < 1. (3.61)
1−α
Using the representations (3.57), we can get for (3.59) other stability conditions.
Put, for instance,

z1 (t) = x1 (t) = x(t), x2 (t) = ẋ(t),

∞
t
∞
t (3.62)
z2 (t) = x2 (t) + dK1 (s) x1 (θ ) dθ + dK2 (s) x2 (θ ) dθ.
0 t−s 0 t−s

By (3.62), (3.57), and (3.40), equation (3.59) can be represented in the form of the
system
ż1 (t) = x2 (t),
(3.63)
ż2 (t) = β01 x1 (t) + β02 x2 (t) + σ x1 (t − τ )ẇ(t),
By Corollary 1.1 the trivial solution of the auxiliary differential equation
   
0 1 y
ẏ = Ay, A = , y= 1 ,
β01 β02 y2

is asymptotically stable if and only if

β01 < 0, β02 < 0. (3.64)

3.3 Linear Differential Equation of nth Order 79

Besides, by (1.29) the matrix equation (1.27) by conditions (3.64) has the positive
definite solution P with the elements
2 + |β |)q + β 2
(β01 02 02 |β01 | + q q
p11 = , p22 = , p12 = − . (3.65)
2β01 β02 2β01 β02 2β01

Suppose that (3.64) holds and consider the Lyapunov function V1 = z P z where
z = (z1 , z2 ) . Calculating LV1 for (3.63), by (3.65) and (3.62) we obtain
 
LV1 = 2 p11 z1 (t) + p12 z2 (t) x2 (t)
  
+ 2 p12 z1 (t) + p22 z2 (t) β01 x1 (t) + β02 x2 (t) + p22 σ 2 x12 (t − τ )

= −qx12 (t) − x22 (t) + p22 σ 2 x12 (t − τ )

∞
t
+ 2p22 β01 dK1 (s) x1 (t)x1 (θ ) dθ
0 t−s

∞
t
− dK1 (s) x2 (t)x1 (θ ) dθ
0 t−s

∞
t
+ 2p22 β01 dK2 (s) x1 (t)x2 (θ ) dθ
0 t−s

∞
t
− dK2 (s) x2 (t)x2 (θ ) dθ.
0 t−s

Using some γ1 > 0, γ2 > 0 and (3.38), we have

LV1 ≤ −qx12 (t) − x22 (t) + p22 σ 2 x12 (t − τ )

∞

  t  2 
+ p22 |β01 | dK1 (s) x1 (t) + x12 (θ ) dθ
0 t−s

∞
 
1  t 1 2
+ dK1 (s) γ1 x22 (t) + x (θ ) dθ
2 0 t−s γ1 1

∞
 
 t 1 2
+ p22 |β01 | dK2 (s) γ2 x12 (t) + x (θ ) dθ
0 t−s γ2 2

∞

1  t  
+ dK2 (s) x22 (t) + x22 (θ ) dθ.
2 0 t−s

As a result, we obtain
 
LV1 ≤ − q − p22 |β01 |(α11 + α12 γ2 ) x12 (t)
 
1
− 1 − (α11 γ1 + α12 ) x22 (t) + p22 σ 2 x12 (t − τ )
2
80 3 Stability of Linear Scalar Equations
 
∞

1   t 2
+ p22 |β01 | + dK1 (s) x1 (θ ) dθ
2γ1 0 t−s
 
∞

p22 |β01 |1 1   t 2
+ + dK2 (s) x2 (θ ) dθ.
γ2 2 0 t−s

From Theorem 2.5 by (2.35) and (3.38) it follows that if

  α11
p22 |β01 |(2α11 + α12 γ2 ) + σ 2 + < q,
2γ1
(3.66)
α11 γ1 p22 α12 |β01 |
α12 + + < 1,
2 γ2

then the trivial solution of (3.59) is asymptotically mean-square stable.

To get stability conditions immediately in the terms of (3.59), transform (3.66)
by the following way. Substituting p22 from (3.65) into (3.66) and solving both
inequalities (3.66) with respect to q, we obtain that inequalities (3.66) are equivalent
to
  −1
μ α11 μ
0< + 2− <q
|β02 | γ1 |β01 β02 |
 
γ2 |β02 | α12 |β01 |
< 2(1 − α12 ) − α11 γ1 − , (3.67)
α12 γ2 |β02 |

where
μ = |β01 |(2α11 + α12 γ2 ) + σ 2 . (3.68)
So, if the inequality
  −1
μ α11 μ
0< + 2−
|β02 | γ1 |β01 β02 |
 
γ2 |β02 | α12 |β01 |
< 2(1 − α12 ) − α11 γ1 −
α12 γ2 |β02 |

holds, then there exists q > 0 such that (3.67) holds too.
From the last condition we have
  
μ α11 γ2 |β02 | α12 |β01 | μ
+ < 2(1 − α12 ) − α11 γ1 − 2−
|β02 | γ1 α12 γ2 |β02 | |β01 β02 |
 
γ2 |β02 |   μ μ
= 2(1 − α12 ) − α11 γ1 2 − − 2|β01 | +
α12 |β01 β02 | |β02 |
or
 
α11 γ2 |β02 |   μ
< 2(1 − α12 ) − α11 γ1 2 − − 2|β01 |.
γ1 α12 |β01 β02 |
3.3 Linear Differential Equation of nth Order 81

Using (3.68), rewrite the obtained inequality in the form

   
2|β01 | + α11 γ1−1 |β02 | α11 σ2
< γ2 2 − − − γ2 . (3.69)
2(1 − α12 ) − α11 γ1 α12 α12 2α12 |β01 |
Now the numbers γ1 and γ2 are separated, and we have minimize the left part of
this inequality with respect to γ1 > 0 and maximize the right part of this inequality
with respect to γ2 > 0.
By Lemma 2.4 the minimum of the left part of (3.69) is reached at
2(1 − α12 )
γ1 = 
2 + 4|β |(1 − α ) + α
α11 01 12 11

and equals γ1−2 . Besides, it is easy to see that the maximum of the right part of
(3.69) is reached at
|β02 | α11 σ2
γ2 = − −
α12 α12 2α12 |β01 |

and equals γ22 . So, condition (3.69) takes the form γ1−1 < γ2 or

α112 + 4|β |(1 − α ) + α
01 12 11 |β02 | α11 σ2
< − − .
2(1 − α12 ) α12 α12 2α12 |β01 |
As a result, from this and from (3.62) we obtain

 2 + 4|β |(1 − α ) + α 
α11 01 12 11
σ < 2|β01 | |β02 | − α11 − α12
2
,
2(1 − α12 )
α11 + α12 < 1. (3.70)
So, if (3.70) holds, then the trivial solution of (3.59) is asymptotically mean-square
stable.

Example 3.1 Consider the second-order differential equation

ẍ(t) + a ẋ(t − h1 ) + bx(t − h2 ) + σ x(t − τ )ẇ(t) = 0 (3.71)
with a > 0, b > 0, h1 ≥ 0, h2 ≥ 0, τ ≥ 0. Equation (3.71) is obtained from (3.59) if
dK1 (s) = −bδ(s − h2 ) ds and dK2 (s) = −aδ(s − h1 ) ds. In this case, α11 = bh2 ,
α12 = ah1 , α21 = bh22 , β02 = −a < 0, β11 = −bh2 , a0 = β02 −β11 = −a +bh2 < 0,
a1 = β01 = −b < 0. The conditions for asymptotic mean-square stability (3.61) and
(3.70) give respectively
   
1 2 b
σ < 2b a − bh2 − ah1 + bh2
2
,
2 1 − ah1 − 12 bh22
1
ah1 + bh22 < 1, (3.72)
2
82 3 Stability of Linear Scalar Equations

Fig. 3.17 Stability regions for (3.67), given by the conditions (3.68), (3.69), are shown for h1 = 1,
h2 = 0, and different values of σ : σ = 0.2 (red), σ = 0.1 (yellow), and σ = 0 (triangle A1 OB1 ).
The triangle A0 OB0 gives the region of stability obtained via the characteristic equation

Fig. 3.18 Stability regions for (3.67), given by conditions (3.68) (red and yellow) and (3.69) (green
and yellow) are shown for h1 = 0.2, h2 = 0.5, σ = 0.25

 ! 
(bh2 )2 + 4b(1 − ah1 ) + bh2
σ < 2b a − bh2 − ah1
2
,
2(1 − ah1 )
ah1 + bh2 < 1. (3.73)
Note that for h2 = 0, conditions (3.72)–(3.73) are equivalent, but for h2 > 0, both
conditions can give regions of stability that are different and complement each other.
3.3 Linear Differential Equation of nth Order 83

In Fig. 3.17 the stability regions given by conditions (3.72)–(3.73) in the space of
the parameters (a, b) are shown for h1 = 1, h2 = 0, and different values of σ : σ =
0.2 (red), σ = 0.1 (yellow), and σ = 0 (the triangle A1 OB1 ). For comparison, the
triangle A0 OB0 gives a part of the region of asymptotic stability that was obtained
via the characteristic equation in Example 1.6 (the case c, Fig. 1.19).
In Fig. 3.18 the stability regions given by conditions (3.72) (red and yellow) and
(3.73) (green and yellow) are shown for h1 = 0.2, h2 = 0.5, σ = 0.25. One can see
that the stability regions have a common part (yellow) but have also the different
parts: red for (3.72) and green for (3.73).

Example 3.2 Consider the second-order differential equation

ẍ(t) = ax(t) + bx(t − h1 ) + bx(t − h2 ) + σ x(t − τ )ẇ(t) = 0 (3.74)

with a > 0, h1 ≥ 0, h2 ≥ 0, τ ≥ 0. Equation (3.74) is obtained from (3.59) if

dK1 (s) = (aδ(s) + b1 δ(s − h1 ) + b2 δ(s − h2 )) ds and dK2 (s) = 0. In this case
we have a0 = −β11 = −b1 h1 − b2 h2 < 0, a1 = β01 = a + b1 + b2 < 0, α = 12 α21 =
2 (|b1 |h1 + |b2 |h2 ). The stability condition (3.61) takes the form
1 2 2

 
|a1 |
σ < 2|a1 | |a0 | − α21
2
, α21 < 2. (3.75)
2(2 − α21 )

Note that the stability condition (3.70) cannot be used here since β02 = 0.
Let n = 3. In this case, (3.36) has the form

∞
∞
...
x (t) = x(t − s) dK1 (s) + ẋ(t − s) dK2 (s)
0 0

∞
+ ẍ(t − s) dK3 (s) + σ x(t − τ )ẇ(t). (3.76)
0

By Remark 3.4 we will consider the corresponding matrix equation (1.27) with

⎛ ⎞ ⎛ ⎞ ⎛ ⎞
0 1 0 q1 0 0 p11 p12 p13
A=⎝0 0 1 ⎠, Q=⎝0 q2 0⎠ , P = ⎝p12 p22 p23 ⎠ .
a2 a1 a0 0 0 1 p13 p23 p33
(3.77)
By Corollary 1.2 the inequalities

a0 < 0, a2 < 0, A0 = a0 a1 + a2 > 0 (3.78)

84 3 Stability of Linear Scalar Equations

are necessary and sufficient conditions for the matrix equation (1.27), (3.77) to have
the positive definite solution P with the elements
 
1 a1 a02 a0 a2 q2 + a22 a 2 a1 |a2 |q2 + a1 a2
p11 = + q1 + , p12 = 0 q1 + ,
2 a2 A0 2A0 2a2 A0 2A0
q1 a03 + a2 (a 2 + |a1 |)q2 + a12 + a0 a2
p13 = , p22 = q1 + 0 ,
2|a2 | 2a2 A0 2A0
a02 |a0 |q2 + |a2 | a0 q2 + |a1 |
p23 = q1 + , p33 = q1 + .
2|a2 |A0 2A0 2a2 A0 2A0
(3.79)
Note that from (3.78) it also follows that a1 < 0. Calculating β1 = |a2 |p33 , β2 =
|p13 + a1 p33 |, we obtain the representation

βl = ρl1 q1 + ρl2 q2 + ρl3 , l = 1, 2, (3.80)

where
|a0 | |a2 | a1 a 2
ρ11 = , ρ12 = , ρ13 = ,
2A0 2A0 2A0
(3.81)
1 |a1 | a2
ρ21 = , ρ22 = , ρ23 = 1 .
2A0 2A0 2A0
So, the stability condition (3.56) can be written in the form
q1 β1
σ 2 < |a2 | sup f (q1 , q2 ), f (q1 , q2 ) = − . (3.82)
q1 >0,q2 >β22 Θ −1 β1 Θ − β22 q2−1

For the fixed ai , i = 0, 1, 2, using (3.78)–(3.82), the supremum of the function

f (q1 , q2 ) can be obtained numerically.

Example 3.3 Consider (3.76) with dKj (s) = kj δ(s − hj ) ds, αij = |kj |hij , βij =
kj hij , j = 1, 2, 3, i = 0, 1, 2. Then

1 1
a0 = β03 − β12 + β21 = k3 − k2 h2 + k1 h21 , a1 = β02 − β11 = k2 − k1 h1 ,
2 2
 
1
a2 = β01 = k1 , A0 = k3 − k2 h2 + k1 h21 (k2 − k1 h1 ) + k1 ,
2
1 1 1 1 1 1
α = α13 + α22 + α31 = |k3 |h3 + |k2 |h22 + |k1 |h31 , Θ= 2
− .
2 6 2 6 α α
Put, for example, h1 = h2 = h3 = 0.1, k1 = −1, k2 = −2, k3 = −3. Then a0 =
−2.805 < 0, a1 = −1.9 < 0, a2 = −1 < 0, A0 = 4.3295 > 0, α ≈ 0.310 < 1, Θ ≈
7.171, ρ11 ≈ 0.324, ρ12 ≈ 0.115, ρ13 ≈ 0.219, ρ21 ≈ 0.115, ρ22 ≈ 0.219, ρ23 ≈
0.417. Conditions (3.74) hold. The function f (d1 , d2 ) reaches its supremum for
q1 ≈ 4.49, q2 ≈ 0.54. The stability condition (3.82) takes the form σ 2 < 2.246.
3.4 Nonautonomous Systems 85

For h3 = 0.2 and the same values of all other parameters, the function f (q1 , q2 )
reaches its supremum for q1 ≈ 0.75, q2 ≈ 0.96, and the stability condition (3.82)
takes the form σ 2 < 0.1969.

3.4 Nonautonomous Systems

3.4.1 Equations with Variable Delays

Consider the scalar stochastic differential equation with variable delays

m
  m t  
ẋ(t) = ax(t) + bi x t − hi (t) + ci x(s) ds + σ x t − τ (t) ẇ(t).
i=1 i=1 t−hi (t)
(3.83)
Suppose that in (3.83)

hi (t) ≤ h0i , ḣi (t) ≤ ĥi < 1, τ̇ (t) ≤ τ̂ < 1 (3.84)

and put

m
|bi | 
m
B(h) =  , C0 (h) = |ci |h0i . (3.85)
i=1 1 − ĥi i=1

Let us consider (3.83) as the representation of type (2.33) with z(t, xt ) = x(t) and
the auxiliary differential equation ẏ(t) = ay(t). The zero solution of this equation is
asymptotically stable if and only if a < 0. Using the appropriate Lyapunov function
v(y) = y 2 , we obtain the functional V1 (t, xt ) in the form V1 (t, xt ) = x 2 (t).
Using (3.84), (3.85), and some positive numbers γi , i = 1, . . . , m, we have

t 
m
   m
LV1 = 2x(t) ax(t) + bi x t − hi (t) + ci x(s) ds
i=1 i=1 t−hi (t)
 
+ σ 2 x t − τ (t)
2
 
 m 
m
 
≤ 2a + C0 (h) + γi |bi | x 2 (t) + γi−1 |bi |x 2 t − hi (t)
i=1 i=1


m
t  
+ |ci | x 2 (s) ds + σ 2 x 2 t − τ (t) .
i=1 t−h0i

So, we obtain the representation of type (2.35), where


m
D = 2a + C0 (h) + γi |bi |, k = m + 1, τk (t) = τ (t), Qk = σ 2 ,
i=1
86 3 Stability of Linear Scalar Equations

R0 = 1, Qi = γi−1 |bi |, τi (t) = hi (t), i = 1, . . . , m,


m
 
dμ0 (s) = |ci |δ s − h0i ds
i=1

and all other parameters are zeros. By (2.36)–(2.37) we have

p  
σ2 γ −1
G = 2a + 2C0 (h) + + γi + i |bi |.
1 − τ̂ 1 − ĥi
i=1

To minimize G, put γi = √ 1 . By Theorem 2.5 we obtain the following asser-

1−ĥi
tion: if
σ2
+ B(h) + C0 (h) < |a|, a < 0, (3.86)
2(1 − τ̂ )
then the zero solution of (3.83) is asymptotically mean-square stable.
In addition to (3.84), assume that
 
ḣi (t) ≤ ĥ0 (3.87)
i

and put

p p
|bi |ĥ0i
B0 (h) = |bi |h0i , B1 (h) =  . (3.88)
i=1 i=1 1 − ĥi

Consider the representation (2.33) of (3.83) in the form of the differential equation
of neutral type
p 


  t
ż(t, xt ) = S0 x(t) + bi ḣi (t)x t − hi (t) + ci x(s) ds
i=1 t−hi (t)
 
+ σ x t − τ (t) ẇ(t), (3.89)

where


p
t 
p
z(t, xt ) = x(t) + bj x(s) ds, S0 = a + bi . (3.90)
j =1 t−hj (t) i=1

Condition (2.10) for (3.89) has the form B0 (h) < 1.

The auxiliary equation for (3.89) is ẏ(t) = S0 y(t), and the zero solution of
this equation is asymptotically stable if and only if S0 < 0. Using the appropri-
ate Lyapunov function v(y) = y 2 , we obtain the functional V1 (t, xt ) in the form
V1 (t, xt ) = z2 (t, xt ).
3.4 Nonautonomous Systems 87

By (3.84), (3.85), (3.87)–(3.90), for some positive numbers γ1i , γ2ij , we obtain


p
  
LV1 (t, xt ) ≤ 2S0 x (t) +
2
|bi |ĥ0i γ1i x 2 (t) + γ1i−1 x 2 t − hi (t)
i=1
 
+ σ x t − τ (t)
2 2


m 
m
t  −1 2  
+ |bj bi |ĥ0i γ2ij x 2 (s) + γ2ij x t − hi (t) ds
i=1 j =1 t−h0j


m 
m

t t  
+ |bj ci | x 2 (θ ) + x 2 (s) ds dθ
i=1 j =1 t−h0i t−h0j


p
t  
+ |S0 bj + cj | x 2 (t) + x 2 (s) ds.
j =1 t−h0j

As a result, we have the representation of type (2.35)

  m
  m t
LV1 (t, xt ) ≤ Dx 2 (t)+σ 2 x 2 t −τ (t) + Qi x 2 t −hi (t) + qj x 2 (s) ds,
i=1 j =1 t−h0j

where


m 
m 
m
 
D = 2S0 + |bi |ĥ0i γ1i + |S0 bj + cj |h0j , dμ0 (s) = qj δ s − h0j ds,
i=1 j =1 j =1

k = m + 1, τk (t) = τ (t), Qk = σ 2 , R0 = 1,

m
Qi = |bi |ĥ0i γ1i−1 + |bi |ĥ0i −1
|bj |h0j γ2ij , τi (t) = hi (t), i = 1, . . . , m,
j =1


m
qj = |S0 bj + cj | + |bj | |bi |ĥ0i γ2ij + |bj |C0 (h) + |cj |B0 (h).
i=1

So, by (2.36)–(2.37) we obtain


m
σ2
G = 2S0 + 2 |S0 bj + cj |h0j + 2B0 (h)C0 (h) +
1 − τ̂
j =1
    −1 
 γ −1 
m m m γ2ij
+ |bi |ĥ0i γ1i + 1i + |bj |h0j |bi |ĥ0i γ2ij + .
i=1 1 − ĥi j =1 i=1 1 − ĥi
88 3 Stability of Linear Scalar Equations

Choosing the optimal values of γ1i = γ2ij = √ 1 , we can minimize G and by

1−ĥi
Theorem 2.5 get the following stability condition: if

σ2  p
 
+ |S0 bj + cj |h0j + B1 (h) + B0 (h) B1 (h) + C0 (h) < |S0 |,
2(1 − τ̂ )
j =1

S0 < 0. (3.91)

then the zero solution of (3.83) is asymptotically mean-square stable.

Remark 3.7 It is easy to see that instead of condition (3.91) one can use a more
rough but more simple condition of the form

σ2     
+ 1 + B0 (h) B1 (h) + C0 (h) < |S0 | 1 − B0 (h) ,
2(1 − τ̂ )
S0 < 0, B0 (h) < 1. (3.92)

Put now

m 
m
 2
C1 (h) = |cj |h0j ĥ0j , C2 (h) = |ci | h0i ,
j =1 i=1 (3.93)
1
A0 (h) = B0 (h) + C2 (h), A1 (h) = B1 (h) + C1 (h),
2
and consider the representation (2.33) of (3.83) in the form of the differential equa-
tion of neutral type

m 

  t
ż(t, xt ) = S(t)x(t) + ḣi (t) bi x t − hi (t) + ci x(s) ds
i=1 t−hi (t)
 
+ σ x t − τ (t) ẇ(t), (3.94)

where
m

 t   
z(t, xt ) = x(t) + bi + ci s − t + hi (t) x(s) ds,
i=1 t−hi (t)
(3.95)

m
 
S(t) = a + bi + ci hi (t) .
i=1

Condition (2.10) for (3.94) has the form A0 (h) < 1.

The auxiliary differential equation in this case is ẏ(t) = S(t)y(t), and if
supt≥0 S(t) < 0, then the zero solution of this equation is asymptotically stable. Us-
ing the appropriate Lyapunov function v(y) = y 2 , we obtain the functional V1 (t, xt )
of the form V1 (t, xt ) = z2 (t, xt ).
3.4 Nonautonomous Systems 89

Then by (3.84), (3.87), (3.88), (3.93)–(3.95), for some positive numbers γ1i , γ2ij ,
we obtain

 m t   
LV1 (t, xt ) ≤ 2S(t)x 2 (t) + S(t) |bi | + |ci | s − t + hi (t)
i=1 t−hi (t)
 
× x 2 (t) + x 2 (s) ds
 m 
  
+ ĥ0i |bi | γ1i x 2 (t) + γ1i−1 x 2 t − hi (t)
i=1


t  2 
+ |ci | x (t) + x 2 (s) ds
t−hi (t)


m 
m
t   
+ ĥ0i |bj | |bi | + |ci | s − t + hi (t)
j =1 i=1 t−hi (t)
 −1 2  
× γ2ij x 2 (s) + γ2ij x t − hj (t) ds

m 
m

t t   
+ ĥ0i |cj | |bi | + |ci | s − t + hi (t)
j =1 i=1 t−hi (t) t−hj (t)
   
× x 2 (θ ) + x 2 (s) dθ ds + σ 2 x 2 t − τ (t) .

Put now

   
Sm = inf S(t), SM = supS(t),
t≥0 t≥0

  t   
Ii hi (t) = |bi | + |ci | s − t + hi (t) ds,
t−hi (t)

  t
J0i hi (t) = x 2 (s) ds,
t−hi (t)

  t   
J1i hi (t) = |bi | + |ci | s − t + hi (t) x 2 (s) ds, i = 1, . . . , m.
t−hi (t)

By (3.84), (3.88), and (3.93) we have

    1  2  m
 
Ii hi (t) ≤ Ii h0i = |bi |h0i + |ci | h0i , Ii hi (t) ≤ A0 (h),
2
i=1
   0    
J0i hi (t) ≤ J0i hi , J1i hi (t) ≤ J1i h0i .
90 3 Stability of Linear Scalar Equations

So, we obtain the representation of type (2.35)


m
   
LV1 (t, xt ) ≤ D(t)x 2 (t) + Qj x 2 t − hj (t) + σ 2 x 2 t − τ (t)
j =1


m
   m
 
+ q0i J0i h0i + q1i J1i h0i ,
i=1 i=1

where

   
m
D(t) = −2 + A0 (h) S(t) + C1 (h) + ĥ0i |bi |γ1i ,
i=1

k = m + 1, τk = τ, Qk = σ , 2
 

m
 
−1 −1
Qj = |bj |ĥ0j γ1j + γ2ij Ii h0i , j = 1, . . . , m,
i=1
R0 = R1 = 1,

p
   
dμ0 (s) = q0i + q1i |bi | δ s − h0i ds,
i=1


p
 
dμ1 (s) = q1i |ci |δ s − h0i ds,
i=1

  
p
q0i = 1 + A0 (h) |ci |ĥ0i , q1i = SM + C1 (h) + |bj |ĥ0j γ2ij .
j =1

As a result, we have

   σ2  
G(t) = −2 + A0 (h) S(t) + A0 (h)SM + + 2C1 (h) 1 + A0 (h)
1 − τ̂
 m  −1  m  m  −1 
γ1j  0 γ2ij
+ |bj |ĥj γ1j +
0
+ |bj |ĥj Ii hi γ2ij +
0
.
j =1 1 − ĥj i=1 j =1 1 − ĥj

To minimize G(t), put γ1j = γ2ij =  1 . Then

1−ĥj

     σ2
G(t) = −2 + A0 (h) S(t) + A0 (h)SM + 2A1 (h) 1 + A0 (h) + .
1 − τ̂

From supt≥0 S(t) < 0 we obtain the estimation for G(t):

3.4 Nonautonomous Systems 91

    σ2
sup G(t) ≤ −2 + A0 (h) Sm + A0 (h)SM + 2A1 (h) 1 + A0 (h) + . (3.96)
t≥0 1 − τ̂

From (3.96) by Theorem 2.4 we obtain: if supt≥0 S(t) < 0 and

σ2    
+ A0 (h)SM + 2A1 (h) 1 + A0 (h) < 2 − A0 (h) Sm , (3.97)
1 − τ̂
then the zero solution of (3.83) is asymptotically mean-square stable.

3.4.2 Equations with Variable Coefficients

Consider the scalar differential equation with delays and variable coefficients

ẋ(t) = −a(t)x(t) − b(t)x(t − h) + σ (t)x(t − τ )ẇ(t),

t ≥ 0, h ≥ 0, τ ≥ 0. (3.98)

For V (t, xt ) = x 2 (t), we have

 
LV (t, xt ) = 2x(t) −a(t)x(t) − b(t)x(t − h) + σ 2 (t)x 2 (t − τ )
    
≤ −2a(t) + b(t) x 2 (t) + b(t)x 2 (t − h) + σ 2 (t)x 2 (t − τ ).

By Theorem 2.5, if there is c > 0 such that

   
G(t) = −2a(t) + b(t) + b(t + h) + σ 2 (t + τ ) ≤ −c, (3.99)

then the trivial solution of (3.98) is asymptotically mean-square stable.

To get another stability condition, rewrite (3.98) in the form

ż(t, xt ) = −c(t)x(t) + σ (t)x(t − τ )ẇ(t), (3.100)

where

t
z(t, xt ) = x(t) − b(s + h)x(s) ds, c(t) = a(t) + b(t + h). (3.101)
t−h

Note that (3.100)–(3.101) is a differential equation of neutral type and suppose that

t+h  
c(t) ≥ c0 > 0, sup b(s) ds < 1. (3.102)
t≥0 t

Consider the auxiliary differential equation without delay

ẏ(t) = −c(t)y(t). (3.103)

92 3 Stability of Linear Scalar Equations

Using the Lyapunov function v(t) = y 2 (t), by (3.99) we have v̇(t) = −2c(t)y 2 (t) ≤
−2c0 y 2 (t). So, the trivial solution of (3.103) is asymptotically stable.
Following the procedure of constructing Lyapunov functionals, we will use
the Lyapunov functional V1 (t, xt ) for (3.100)–(3.101) of the form V1 (t, xt ) =
z2 (t, xt ). Calculating LV1 (t, xt ), from (3.100)–(3.101) we obtain the representation
of type (2.35)

t 
LV1 (t, xt ) = c(t) −2x (t) + 2
2
b(s + h)x(t)x(s) ds + σ 2 (t)x 2 (t − τ )
t−h


t   
≤ c(t) −2x (t) +
2 b(s + h) x 2 (t) + x 2 (s) ds + σ 2 (t)x 2 (t − τ )
t−h

t
= D(t)x 2 (t) + σ 2 (t)x 2 (t − τ ) + R(s + h, t)x 2 (s) ds,
t−h

where

t+h  

D(t) = c(t) −2 + b(s) ds , Q1 (t) = σ 2 (t + τ ),
t
 
dμ(τ ) = δ(τ − h) dτ, R(s, t) = b(s)c(t).

So, by Theorem 2.5, G(t) = c(t)(−2 + G0 (t)), where

t+h 


b(s) ds + |b(t + h)| σ 2 (t + τ )
t+h
G0 (t) = c(s) ds + ,
t c(t) t c(t)

and if

sup G0 (t) < 2, (3.104)

t≥0

then the zero solution of (3.98) is asymptotically mean-square stable.

Note that if h = 0 and the functions a(t) = a, b(t) = b, σ (t) = σ are constants,
then condition (3.104) coincides with the necessary and sufficient condition for
asymptotic mean-square stability σ 2 < 2(a + b). But in the general case the both
obtained conditions (3.99) and (3.102), (3.104) are only sufficient conditions for
asymptotic mean-square stability of the trivial solution of (3.98). Indeed, consider
(3.98) with a(t) = α + sin(t), b(t) = cos(t), and σ (t) = 0. In Fig. 3.19 the behavior
of the functions G(t) (green), c(t) (grey), G0 (t) (blue), and the solution x(t) (red)
of (3.98) are shown for x(s) = 0.1, s ∈ [−h, 0], h = 0.1, α = −0.2. Here conditions
(3.99) and (3.102), (3.104) do not hold, and the solution goes to infinity. In Fig. 3.20
the same picture is shown for x(s) = 2.5, s ∈ [−h, 0], h = 0.8, α = 0.2. Conditions
(3.99) and (3.102), (3.104) do not hold too, but the solution x(t) goes to zero. In
Fig. 3.21 the same picture is shown for x(s) = 1.5, s ∈ [−h, 0], h = 1.1, α = 1.7.
Conditions (3.99) and (3.102), (3.104) hold, and the solution x(t) goes to zero.
3.4 Nonautonomous Systems 93

Fig. 3.19 Functions G(t) (green), c(t) (grey), G0 (t) (blue), and the solution x(t) (red) of (3.98)
are shown for x(s) = 0.1, s ∈ [−h, 0], h = 0.1, α = −0.2

Consider the scalar stochastic differential equation with variable coefficients and
unbounded delay

∞
ẋ(t) + a(t)x(t) + b(t, s)x(t − s) ds + σ (t)x(t − τ )ẇ(t) = 0. (3.105)
0

For the function V1 (t, xt ) = x 2 (t), we have


∞ 
LV1 (t, xt ) = 2x(t) −a(t)x(t) − b(t, s)x(t − s) ds + σ 2 (t)x 2 (t − τ )
0

∞ 

≤ −2a(t) + b(t, s) ds x 2 (t)
0

∞ 
+ b(t, s)x 2 (t − s) ds + σ 2 (t)x 2 (t − τ ).
0

To obtain a stability condition, we consider the functional

∞
t

  t
V2 (t, xt ) = b(τ + s, s)x 2 (τ ) dτ ds + σ 2 (s + τ )x 2 (s) ds.
0 t−s t−τ

Then

∞     
LV2 (t, xt ) = b(t + s, s)x 2 (t) − b(t, s)x 2 (t − s) ds
0

+ σ 2 (t + τ )x 2 (t) − σ 2 (t)x 2 (t − τ ).
94 3 Stability of Linear Scalar Equations

Fig. 3.20 Functions G(t) (green), c(t) (grey), G0 (t) (blue) and the solution x(t) (red) of (3.98)
are shown for x(s) = 2.5, s ∈ [−h, 0], h = 0.8, α = 0.2

Fig. 3.21 Functions G(t) (green), c(t) (grey), G0 (t) (blue), and the solution x(t) (red) of (3.98)
are shown for x(s) = 1.5, s ∈ [−h, 0], h = 1.1, α = 1.7

For V = V1 + V2 , we obtain LV (t, xt ) ≤ G(t)x 2 (t), where

∞    
G(t) = −2a(t) + b(t, s) + b(t + s, s) ds + σ 2 (t + τ ). (3.106)
0

So, if supt≥0 G(t) < 0, then the trivial solution of (3.105) is asymptotically mean-
square stable.
Note that if b(t, s) = b(t)δ(s − h), then (3.105) coincides with (3.98), and the
function (3.106) coincides with (3.99).
3.4 Nonautonomous Systems 95

Fig. 3.22 The solution of (3.107) (red) and the function G(t) (blue) are shown in the deterministic
case (σ (t) = 0) for a(t) = 1.2 + sin t , x0 = 2, γ = 0.1, λ = 0.2

Fig. 3.23 The solution of (3.107) (red) and the function G(t) (blue) are shown in the deterministic
case (σ (t) = 0) for a(t) = 0.6 − | cos t|, x0 = 1.5, γ = 1, λ = 0.6

The negativity of the function G(t) is a sufficient stability condition but not a
necessary one. Put, for instance, in (3.105)

b(t, s) = −e−λt−γ s , γ > 0, λ > 0, x(s) = x0 , s ≤ 0.

96 3 Stability of Linear Scalar Equations

Then (3.105) and the function G(t) take the forms


t 
−λt x0 −γ t −γ s
ẋ(t) + a(t)x(t) + σ (t)x(t − τ )ẇ(t) = e e + e x(t − s) ds ,
γ 0
(3.107)
 
1 1
G(t) = −2a(t) + σ 2 (t − τ ) + e−λt + . (3.108)
γ γ +λ
If supt≥0 G(t) < 0, then the trivial solution of (3.107) is asymptotically mean-square
stable.
Note that in some cases the trivial solution of (3.107) is asymptotically stable
if the function G(t) given by (3.108) is a negative one for large enough t only.
In Fig. 3.22 the trajectory of the solution of (3.107) (red) and the function G(t)
(blue) are shown in the deterministic case (σ (t) = 0) for a(t) = 1.2 + sin t, x0 = 2,
γ = 0.1, λ = 0.2. In this case the condition G(t) ≤ 0 holds for all large enough t,
and the solution of (3.107) goes to zero. In Fig. 3.23 we can see a similar picture
with a(t) = 0.6 − | cos t|, x0 = 1.5, γ = 1, λ = 0.6. In this case the function G(t)
for all t ≥ 0 has both negative and positive values, and the solution of (3.108) goes
to infinity.
Chapter 4
Stability of Linear Systems of Two Equations

Below, several examples are considered where the procedure of constructing Lya-
punov functionals is used for stability investigation of linear systems of two equa-
tions with constant delays, with distributed delays, and with variable coefficients.

4.1 Linear Systems of Two Equations with Constant Delays

Example 4.1 Consider the system of two stochastic differential equations with fixed
delays
ẋ1 (t) = ax2 (t) − bx1 (t − h) + σ1 x1 (t − τ1 )ẇ1 (t),
(4.1)
ẋ2 (t) = −ax1 (t) − bx2 (t − h) + σ2 x2 (t − τ2 )ẇ2 (t),
where w1 (t) and w2 (t) are the mutually independent standard Wiener processes.
Put x(t) = (x1 (t), x2 (t)) , w(t) = (w1 (t), w2 (t)) ,
 t
ρ(t) = x(s) ds, z(t) = x(t) − bρ(t),
t−h
  (4.2)
σ x(t − τ1 ) 0
B(xt ) = 1 ,
0 σ2 x(t − τ2 )

 
−b a
A= , (4.3)
−a −b
and rewrite (4.1) in the form

ż(t) = Ax(t) + B(xt )ẇ(t). (4.4)

By (4.3) and Corollary 1.1 the condition b > 0 is a necessary and sufficient con-
dition for asymptotic stability of the zero solution of the auxiliary differential equa-
tion ẏ(t) = Ay(t). Let P be a positive definite solution of the matrix equation (1.27),

L. Shaikhet, Lyapunov Functionals and Stability of Stochastic Functional 97

Differential Equations, DOI 10.1007/978-3-319-00101-2_4,
© Springer International Publishing Switzerland 2013
98 4 Stability of Linear Systems of Two Equations

where Q is (for simplicity) the identity matrix, and A is defined by (4.3). By (1.29)
the elements of P are p11 = p22 = (2b)−1 , p12 = 0.
Following to the procedure of constructing Lyapunov functionals, consider the
functional V1 = z (t)P z(t) = (2b)−1 |z(t)|2 . Calculating LV1 for (4.4), by (4.2) we
obtain
 
1 2

LV1 = 2z (t)Ax(t) + σi xi (t − τi )
2 2
2b
i=1
 
1   
2
= 2 x(t) − bρ(t) Ax(t) + σi xi (t − τi )
2 2
2b
i=1
  
1 t 
2
= 2x  (t)Ax(t) − 2b 
x (s)Ax(t) ds + σi2 xi2 (t − τi ) .
2b t−h i=1

Note, that x  (t)Ax(t) = −b|x(t)|2 and 2x  (s)Ax(t) ≤ A(|x(s)|2 + |x(t)|2 ),

where A = sup|x|=1 |Ax| is the operator norm of a matrix A. Therefore,
  t 
1

2 

2

2 
2
LV1 ≤ −2b
x(t)
+ bA
x(s)
+
x(t)
ds + σi2 xi2 (t − τi )
2b t−h i=1
   t 
1

2 1

2
= −1 + Ah
x(t)
+ A
x(s)
2 ds + 1 σi2 xi2 (t − τi ).
2 2 t−h 2b
i=1

So, we have got the representation of type (2.25) with S(t) = Qi (t) = R(s, t) = 0,
m = 0, n = 2 and
 
1 1
D = −1 + Ah I, dKi (s) = σi2 δ(s − τi ) ds, i = 1, 2,
2 2b
1
R0 (s) = A, dμ0 (s) = δ(s − h) ds.
2
 √
Note also that A = sup|x|=1 (−bx1 + ax2 )2 + (−ax1 − bx2 )2 = a 2 + b2 .
From this and from Theorem 2.4 we obtain that if
 p 1 
h a 2 + b2 + < 1, p = max σ12 , σ22 ,
b 2
or

 
1 p 2
|a| < 1− − b2 , (4.5)
h2 b
1  1  1
1 − 1 − 4ph < b < 1 + 1 − 4ph , ph < , (4.6)
2h 2h 4
4.1 Linear Systems of Two Equations with Constant Delays 99

Fig. 4.1 Stability regions

for (4.1) given by
conditions (4.5), (4.6) for
h = 1 and different values of
p: (1) p = 0, (2) p = 0.1,
(3) p = 0.2, (4) p = 0.249

Fig. 4.2 Stability regions

for (4.1) given by
conditions (4.5), (4.6) for
p = 0.1 and different values
of h: (1) h = 0, (2) h = 1,
(3) h = 2, (4) h = 2.45

Fig. 4.3 Stability regions

for (4.1) given by
conditions (4.5), (4.6) for
h = 0 and different values of
p: (1) p = 0, (2) p = 0.25

then the trivial solution of system (4.1) is asymptotically mean-square stable.

The stability regions for (4.1) given by conditions (4.5)–(4.6) are shown in
Fig. 4.1 for h = 1 and different values of p: (1) p = 0; (2) p = 0.1; (3) p = 0.2;
(4) p = 0.249; in Fig. 4.2 for p = 0.1 and different values of h: (1) h = 0; (2) h = 1;
(3) h = 2; (4) h = 2.45; in Fig. 4.3 for h = 0 and different values of p: (1) p = 0;
(2) p = 0.25.
100 4 Stability of Linear Systems of Two Equations

Example 4.2 Consider now the system of two stochastic differential equations with
fixed delays

ẋ1 (t) = ax2 (t) − bx1 (t − h) + σ1 x1 (t − τ1 )ẇ(t),

(4.7)
ẋ2 (t) = ax1 (t) − bx2 (t − h) + σ2 x2 (t − τ2 )ẇ(t),

which can be transformed to the form (4.4), (4.2) with the matrix
 
−b a
A= . (4.8)
a −b

By (4.8) and Corollary 1.1 the condition b > |a| is a necessary and sufficient con-
dition for asymptotic stability of the zero solution of the auxiliary differential equa-
tion ẏ(t) = Ay(t). By (1.29) the positive definite solution P of the matrix equation
(1.27), where Q = I (the identity matrix), and A is defined by (4.8), has the ele-
ments
b a
p11 = p22 = , p12 = . (4.9)
2(b2 − a2) 2(b2 − a2)
Put V1 = z (t)P z(t) and note that, by (4.8)–(4.9), 2P A = −I . Calculating LV1
for (4.4) with the parameters defined by (4.2), (4.8), and (4.9), we obtain

  
2
LV1 = − x(t) − bρ(t) x(t) + pii σi2 xi2 (t − τi )
i=1
 


2 t 2
= −
x(t)
+ b 
x (s)x(t) ds + pii σi2 xi2 (t − τi )
t−h i=1
   
bh

2 b

t 2
≤ −1 + x(t)
+
x(s)
2 ds + pii σ 2 x 2 (t − τi ).
i i
2 2 t−h i=1

So, we have the representation of type (2.25) with S(t) = Qi (t) = R(s, t) = 0, m =
1, n = 2 and
 
bh
D = −1 + I, dKi (s) = pii σi2 δ(s − τi ) ds, i = 1, 2,
2
b
dμ0 (s) = δ(s − h) ds, R0 (s) = .
2
By Theorem 2.4 and (4.9) we obtain that if
 
p 1 
b h+ 2 < 1, p= max σ12 , σ22 ,
b − a2 2
4.1 Linear Systems of Two Equations with Constant Delays 101

Fig. 4.4 Stability regions for

(4.7) given by conditions
(4.10), (4.6) for h = 1 and
different values of p:
(1) p = 0, (2) p = 0.1,
(3) p = 0.2, (4) p = 0.249

Fig. 4.5 Stability regions for

(4.7) given by conditions
(4.10), (4.6) for p = 0.1 and
different values of h:
(1) h = 0, (2) h = 1,
(3) h = 2, (4) h = 2.45

Fig. 4.6 Stability regions for

(4.7) given by conditions
(4.10), (4.6) for h = 0 and
different values of p:
(1) p = 0, (2) p = 0.25

or if the conditions
  
p
|a| < b b − (4.10)
1 − bh
and (4.6) hold, then the trivial solution of (4.7) is asymptotically mean-square stable.
The stability regions for (4.7) given by conditions (4.10), (4.6) are shown in
Fig. 4.4 for h = 1 and different values of p: (1) p = 0; (2) p = 0.1; (3) p = 0.2;
(4) p = 0.249; in Fig. 4.5 for p = 0.1 and different values of h: (1) h = 0; (2) h = 1;
(3) h = 2; (4) h = 2.45; in Fig. 4.6 for h = 0 and different values of p: (1) p = 0;
(2) p = 0.25.
102 4 Stability of Linear Systems of Two Equations

4.2 Linear Systems of Two Equations with Distributed Delays

Example 4.3 Consider the system of two stochastic differential equations with dis-
tributed delays
 t  t
ẋ1 (t) = ax2 (t) − b x1 (s) ds + σ1 x1 (s) ds ẇ1 (t),
t−h t−τ1
 t  t
(4.11)
ẋ2 (t) = −ax1 (t) − b x2 (s) ds + σ2 x2 (s) ds ẇ2 (t),
t−h t−τ2

where w1 (t) and w2 (t) are the mutually independent standard Wiener processes.
Put x(t) = (x1 (t), x2 (t)) , w(t) = (w1 (t), w2 (t)) ,
 t
ρ(t) = (s − t + h)x(s) ds, z(t) = x(t) − bρ(t),
t−h
t  (4.12)
σ1 t−τ1 x1 (s) ds 0
B(xt ) = t ,
0 σ2 t−τ2 x2 (s) ds
 
−bh a
A= . (4.13)
−a −bh
Then (4.11) can be represented in the form (4.4).
By (4.13) and Corollary 1.1 the conditions b > 0, h > 0 are necessary and suffi-
cient conditions for asymptotic stability of the zero solution of the auxiliary differ-
ential equation ẏ(t) = Ay(t). The positive definite solution P of the matrix dif-
ferential equation (1.27) with Q = I and A defined by (4.13) has the elements
p11 = p22 = (2bh)−1 , p12 = 0.
Following the procedure of constructing Lyapunov functionals, put
1

2
V1 = z (t)P z(t) = z(t)
.
2bh
Calculating LV1 for system (4.11), by (4.12)–(4.13) we have
 2  t 2 
1 

LV1 = 2z (t)Ax(t) + σi xi (s) ds
2bh t−τi i=1
 2  t 2 
1   
= 2 x(t) − bρ(t) Ax(t) + σi xi (s) ds
2bh t−τi
i=1
  t
1
= 2x  (t)Ax(t) − 2b (s − t + h)x  (s)Ax(t) ds
2bh t−h

2  t 2 
+ σi xi (s) ds .
i=1 t−τi
4.2 Linear Systems of Two Equations with Distributed Delays 103

Using x  (t)Ax(t) = −bh|x(t)|2 and 2x  (s)Ax(t) ≤ A(|x(s)|2 + |x(t)|2 ), we ob-

tain
  t
1

2 

2

2
LV1 ≤ −2bh
x(t)
+ bA (s − t + h)
x(s)
+
x(t)
ds
2bh t−h

 t   
2
1

2
+ 2
σi τi xi (s) ds = −1 + Ah
x(t)

2
t−τi 4
i=1
 


2 1  2
t 2 t
1
+ A (s − t + h)
x(s)
ds + σ i τi xi2 (s) ds.
2h t−h 2bh t−τi
i=1

So, we have got the representation of type (2.25) with S(t) = Qi (t) = R(s, t) = 0,
m = 1, n = 2 and
 
1 1 1
D = −1 + Ah I, R0 (s) = , R1 (s) = A,
4 2bh 2h

2
dμ0 (s) = σi2 τi δ(s − τi ) ds, dμ1 (s) = δ(s − h) ds.
i=1

√
Note also that A = a 2 + b2 h2 . From this and from Theorem 2.4 it follows
that if
h 2 pτ 1 
a + b2 h2 + < 1, pτ = max σ12 τ12 , σ22 τ22 ,
2 bh 2
or if the conditions

 
4 pτ 2
|a| < 1 − − b 2 h2 , (4.14)
h2 bh
pτ 1  1  1
< 2 (1 − 1 − 2pτ h) < b < 2 (1 + 1 − 2pτ h), pτ h < , (4.15)
h h h 2

hold, then the trivial solution of (4.11) is asymptotically mean-square stable.

The stability regions for (4.11) given by conditions (4.14)–(4.15) are shown in
Fig. 4.7 for h = 1.1 and different values of pτ : (1) pτ = 0, (2) pτ = 0.2, (3) pτ =
0.4, (4) pτ = 0.45, in Figs. 4.8 and 4.9 for pτ = 0.1 and pτ = 2.5 respectively and
different values of h: (1) h = 0.085, (2) h = 0.1, (3) h = 0.12, (4) h = 0.16, in
Fig. 4.10 for pτ = 2.5 and the following values of h: (1) h = 0.01, (2) h = 0.005,
(3) h = 0.0025, (4) h = 0.00125. In the last figure one can see that by pτ > 0, as
h → 0, the stability region goes to infinity.
104 4 Stability of Linear Systems of Two Equations

Fig. 4.7 Stability regions for

(4.11) given by conditions
(4.14), (4.15) for h = 1.1 and
different values of pτ :
(1) pτ = 0, (2) pτ = 0.2,
(3) pτ = 0.4, (4) pτ = 0.45

Fig. 4.8 Stability regions for

(4.11) given by conditions
(4.14), (4.15) for pτ = 0.1
and different values of h:
(1) h = 0.085, (2) h = 0.1,
(3) h = 0.12, (4) h = 0.16

Fig. 4.9 Stability regions for

(4.11) given by conditions
(4.14), (4.15) for pτ = 2.5
and different values of h:
(1) h = 0.085, (2) h = 0.1,
(3) h = 0.12, (4) h = 0.16
4.2 Linear Systems of Two Equations with Distributed Delays 105

Fig. 4.10 Stability regions

for (4.11) given by conditions
(4.14), (4.15) for pτ = 2.5
and different values of h:
(1) h = 0.01, (2) h = 0.005,
(3) h = 0.0025,
(4) h = 0.00125

Example 4.4 Consider now the system of two stochastic differential equations with
distributed delays
 t  t
ẋ1 (t) = ax2 (t) − b x1 (s) ds + σ1 x1 (s) ds ẇ(t),
t−h t−τ1
 t  t
(4.16)
ẋ2 (t) = ax1 (t) − b x2 (s) ds + σ2 x2 (s) ds ẇ(t),
t−h t−τ2

which can be transformed to the form (4.4), (4.12) with the matrix
 
−bh a
A= . (4.17)
a −bh

By (4.17) and Corollary 1.1 the condition bh > |a| is a necessary and suffi-
cient condition for asymptotic stability of the zero solution of the auxiliary equation
ẏ(t) = Ay(t). The positive definite solution P of the matrix equation (1.27) with
Q = I and A defined by (4.17) has the elements

bh a
p11 = p22 = , p12 = . (4.18)
2(b2 h2 − a 2 ) 2(b2 h2 − a 2 )

Following the procedure of constructing Lyapunov functionals, put V1 =

z (t)P z(t)
and note that 2P A = −I . Thus, calculating LV1 for (4.16), we obtain


2  2
  t
LV1 = − x(t) − bρ(t) x(t) + pii σi xi (s) ds
i=1 t−τi

   2


2 t 2 t
= −
x(t)
+ b 
(s − t + h)x (s)x(t) + pii σi xi (s) ds
t−h i=1 t−τi
106 4 Stability of Linear Systems of Two Equations



2 b t 

2

2
≤ −
x(t)
+ (s − t + h)
x(s)
+
x(t)
ds
2 t−h


2  t
+ pii σi2 τi xi2 (s) ds
i=1 t−τi
  
bh2

2 b t

2
= −1 + x(t)
+ (s − t + h)
x(s)
ds
4 2 t−h

2  t
+ pii σi2 τi xi2 (s) ds.
i=1 t−τi

So, we have got the representation of type (2.25) with S(t) = Qi (t) = R(s, t) =
0, m = 1, n = 2 and

 
bh2
D = −1 + I, dKi (s) = pii σi2 δ(s − τi ) ds, i = 1, 2,
4
b
dμ1 (s) = δ(s − h) ds, R1 (s) = .
2

From this and from Theorem 2.4 it follows that if

 
h pτ 1 
bh + < 1, pτ = max σ12 τ12 , σ22 τ22 ,
2 b 2 h2 − a 2 2

or if the conditions
  
2pτ
|a| < bh bh − (4.19)
2 − bh2

and (4.15) hold, then the trivial solution of (4.16) is asymptotically mean-square
stable.
The stability regions for (4.16) given by conditions (4.19), (4.15) are shown in
Fig. 4.11 for h = 1.1 and different values of pτ : (1) pτ = 0, (2) pτ = 0.2, (3) pτ =
0.4, (4) pτ = 0.45, in Figs. 4.12 and 4.13 for pτ = 0.1 and pτ = 2.5, respectively,
and different values of h: (1) h = 0.085, (2) h = 0.1, (3) h = 0.12, (4) h = 0.16,
in Fig. 4.14 for pτ = 2 and the following values of h: (1) h = 0.01, (2) h = 0.005,
(3) h = 0.0025, (4) h = 0.00125. In the last figure one can see that by pτ > 0, as
h → 0, the stability region disappears.
4.3 Linear Systems of Two Equations with Variable Coefficients 107

Fig. 4.11 Stability regions for (4.16) given by conditions (4.19), (4.15) for h = 1.1 and different
values of pτ : (1) pτ = 0, (2) pτ = 0.2, (3) pτ = 0.4, (4) pτ = 0.45

Fig. 4.12 Stability regions for (4.16) given by conditions (4.19), (4.15) for pτ = 0.1 and different
values of h: (1) h = 0.085, (2) h = 0.1, (3) h = 0.12, (4) h = 0.16

4.3 Linear Systems of Two Equations with Variable Coefficients

Example 4.5 Consider the system of two stochastic differential equations with vari-
able coefficients before the terms with delays

ẋ1 (t) = −ax1 (t) + b(t)x1 (t − h) + cx2 (t) + σ1 (t)x1 (t − τ1 )ẇ1 (t),
(4.20)
ẋ2 (t) = −cx1 (t) + σ2 (t)x2 (t − τ2 )ẇ2 (t), t ≥ 0,

where a, c, h ≥ 0, τ1 ≥ 0, τ2 ≥ 0 are constants, and b(t), σ1 (t), σ2 (t) are bounded

functions.
108 4 Stability of Linear Systems of Two Equations

Fig. 4.13 Stability regions for (4.16) given by conditions (4.19), (4.15) for pτ = 2.5 and different
values of h: (1) h = 0.085, (2) h = 0.1, (3) h = 0.12, (4) h = 0.16

Fig. 4.14 Stability regions for (4.16) given by conditions (4.19), (4.15) for pτ = 2 and different
values of h: (1) h = 0.01, (2) h = 0.005, (3) h = 0.0025, (4) h = 0.00125

Let us show that by the conditions

a > 0, |c| > 0, sup μ(t) < a,

t≥0
(4.21)
4c2 ν(t)
δ2 (t + τ2 ) <  √ ,
( β1 (t) + 4(c2 + aν(t)) + β1 (t) )2
4.3 Linear Systems of Two Equations with Variable Coefficients 109

where

μ(t) = δ1 (t + τ1 ) + β0 (t + h), ν(t) = a − μ(t) > 0,

1 

(4.22)
β0 (t) = b(t)
+
b(t + h)
, β1 (t) =
b(t)b(t + h)
,
2

the trivial solution of (4.20) is asymptotically mean-square stable.

Rewrite (4.20) in the form

ẋ(t) = Ax(t) + B(t)x(t − h) + σ (t, xt )ẇ(t),

where x(t) = (x1 (t), x2 (t)) , w(t) = (w1 (t), w2 (t)) ,

   
−a c b(t) 0
A= , B(t) = ,
−c 0 0 0
  (4.23)
σ1 (t)x1 (t − τ1 ) 0
σ (t, xt ) = .
0 σ2 (t)x2 (t − τ2 )

Let Q be the symmetric 2 × 2 matrix with the elements q11 = q > 0, q22 = 1,
q12 = 0, and the matrix P be a solution of the matrix equation (1.27) where A is
defined in (4.23). By (1.29) the elements of the matrix P are

q +1 a q +1 1
p11 = , p22 = 2
+ , p12 = − . (4.24)
2a 2c 2a 2c

For the functional V1 = x  (t)P x(t), by (1.27) we have

  
LV1 = Ax(t) + B(t)x(t − h) P x(t) + x  (t)P Ax(t) + B(t)x(t − h)

+ p11 σ12 (t)x12 (t − τ1 ) + p22 σ22 (t)x22 (t − τ2 )


= x  (t) A P + P A x(t) + 2x  (t − h)B  (t)P x(t)

+ p11 σ12 (t)x12 (t − τ1 ) + p22 σ22 (t)x22 (t − τ2 )

= −qx12 (t) − x22 (t) + 2x  (t − h)B  (t)P x(t)

+ p11 σ12 (t)x12 (t − τ1 ) + p22 σ22 (t)x22 (t − τ2 ). (4.25)

By (4.23), for some γ > 0, we obtain


2x  (t − h)B  (t)P x(t) = 2b(t)x1 (t − h) p11 x1 (t) + p12 x2 (t)


 
≤
b(t)
p11 x12 (t) + x12 (t − h)
 
+ |p12 | γ x22 (t) + γ −1 x12 (t − h) .
110 4 Stability of Linear Systems of Two Equations

From this and from (4.25) we have



LV1 ≤ −q + p11
b(t)
x12 (t) + (−1 + γ )
p12 b(t)
x22 (t)


+ p11 + γ −1 |p12 |
b(t)
x12 (t − h)
+ p11 σ12 (t)x12 (t − τ1 ) + p22 σ22 (t)x22 (t − τ2 ).

So, by Theorem 2.5 we obtain that if



p11
b(t)
+
b(t + h)
+ σ12 (t + τ1 ) + γ −1
p12 b(t + h)
< q,


(4.26)
γ
p12 b(t)
+ p22 σ22 (t + τ2 ) < 1,

then the trivial solution of (4.20) is asymptotically mean-square stable.

Note that inequalities (4.26) can be represented in the form

|p12 b(t + h)| 1 − p22 σ22 (t + τ2 )

0< < γ < . (4.27)
q − p11 (|b(t)| + |b(t + h)| + σ12 (t + τ1 )) |p12 b(t)|

So, if

|p12 b(t + h)| 1 − p22 σ22 (t + τ2 )

0< <
q − p11 (|b(t)| + |b(t + h)| + σ12 (t + τ1 )) |p12 b(t)|

or (which is the same)

2


p12 b(t)b(t + h)
< 1 − p22 σ22 (t + τ2 )
 




× q − p11
b(t)
+
b(t + h)
+ σ 2 (t + τ1 ) ,1 (4.28)

then there exists γ > 0 that satisfies condition (4.27).

Substituting (4.24) into (4.28) and using (4.22), we obtain
    
β1 (t) a q +1 q +1
< 1 − + δ2 (t + τ2 ) q − μ(t) .
4c2 c2 a a
From this by (4.22) it follows that

a2   
β1 (t) < ac2 − a 2 + (q + 1)c2 δ2 (t + τ2 ) qν(t) − μ(t) .
4
Rewrite this inequality in the form

c2 ν(t)δ2 (t + τ2 )q 2
  
− ac2 − a 2 + c2 δ2 (t + τ2 ) ν(t) + c2 μ(t)δ2 (t + τ2 ) q
  a2
+ ac2 − a 2 + c2 δ2 (t + τ2 ) μ(t) + β1 (t) < 0.
4
4.3 Linear Systems of Two Equations with Variable Coefficients 111

The obtained condition holds for arbitrary q > 0 and each t ≥ 0 if and only if for
each t ≥ 0, the following condition holds:
  2
ac2 − a 2 + c2 δ2 (t + τ2 ) ν(t) + c2 μ(t)δ2 (t + τ2 )
   
> c2 ν(t)δ2 (t + τ2 ) 4 ac2 − a 2 + c2 δ2 (t + τ2 ) μ(t) + a 2 β1 (t)
 
= 4 ac2 − a 2 + c2 δ2 (t + τ2 ) μ(t)c2 ν(t)δ2 (t + τ2 )
+ a 2 c2 ν(t)β1 (t)δ2 (t + τ2 ),

or
  2
ac2 − a 2 + c2 δ2 (t + τ2 ) ν(t) − c2 μ(t)δ2 (t + τ2 ) > a 2 c2 ν(t)β1 (t)δ2 (t + τ2 ),

or
 2  2
c ν(t) − c2 + aν(t) δ2 (t + τ2 ) > c2 ν(t)β1 (t)δ2 (t + τ2 ). (4.29)
From (4.21) it follows that

c2 ν(t)
δ2 (t + τ2 ) < .
c2 + aν(t)
So, (4.29) can be written in the form
 
c2 ν(t) − c2 + aν(t) δ2 (t + τ2 ) > c ν(t)β1 (t)δ2 (t + τ2 )

or
 
c2 + aν(t) δ2 (t + τ2 ) + c ν(t)β1 (t)δ2 (t + τ2 ) − c2 ν(t) < 0.
From this we obtain
√  √
 c ν(t)( β1 (t) + 4(c2 + aν(t)) − β1 (t))
δ2 (t + τ2 ) <
2(c2 + aν(t))
√
2c ν(t)
= √ ,
β1 (t) + 4(c2 + aν(t)) + β1 (t)

which is equivalent to (4.21).

Note that in the case b(t) = 0, σi (t) = σi = const, i = 1, 2, conditions (4.21) take
the form
c2 (a − δ1 )
δ1 < a, δ2 < ,
c2 + a(a − δ1 )
which immediately follows from (2.56)–(2.57).
Chapter 5
Stability of Systems with Nonlinearities

The idea of the method of investigation that is used below is similar to the method of
the first approximation. Namely, a linear part of the considered differential equation
is interpreted as being undisturbed, and the other part as the disturbance. Assuming
that the undisturbed equation is stable and the nonlinear disturbance in the right-
hand side can be majorized in a certain sense, this method makes it possible to
obtain stability conditions.

5.1 Systems with Nonlinearities in Stochastic Part

5.1.1 Scalar First-Order Differential Equation

Let us consider the scalar differential equation

 ∞
ẋ(t) = −ax(t) + x(t − s) dK(s) + σ (t, xt )ẇ(t), t ≥ 0, (5.1)
0

where

  ∞ 
σ (t, ϕ) ≤ ϕ(−s) dR(s), (5.2)
0

the function K(s) and the nondecreasing function R(s) are functions of bounded
variation such that
 ∞ 
 ∞
k0 = dK(s) < ∞, R= dR(s) < ∞. (5.3)
0 0

L. Shaikhet, Lyapunov Functionals and Stability of Stochastic Functional 113

Differential Equations, DOI 10.1007/978-3-319-00101-2_5,
© Springer International Publishing Switzerland 2013
114 5 Stability of Systems with Nonlinearities

5.1.1.1 The First Way of Constructing a Lyapunov Functional

We will consider the auxiliary differential equation (2.34) for (5.1) in the form
ẏ(t) = −ay(t), a > 0. Put V1 = x 2 . Calculating LV1 , by (5.2) and (5.3) we have
  ∞ 
LV1 = 2x(t) −ax(t) + x(t − s) dK(s) + σ 2 (t, xt )
0
 
∞   ∞
≤ −2ax 2 (t) + x 2 (t) + x 2 (t − s) dK(s) + R x 2 (t − s) dR(s)
0 0
 
∞   ∞
= (−2a + k0 )x (t) +
2
x (t − s)dK(s) + R
2
x 2 (t − s) dR(s).
0 0

By the representation (2.35) with D = −2a + k0 and dK1 (s) = |dK(s)| +

R dR(s) (all other parameters are zeros) and by Theorem 2.5 we obtain: if
1
a > k0 + p, p = R2, (5.4)
2
then the trivial solution of (5.1) is asymptotically mean-square stable.

Example 5.1 Consider the scalar differential equation

 h
ẋ(t) = −ax(t) + b (h − s)x(t − s) ds + σ (t, xt )ẇ(t), t ≥ 0, (5.5)
0

which is a particular case of (5.1) with


b(h − s), s ∈ [0, h],
dK(s) = f (s) ds, f (s) = (5.6)
0, s > h.

In this case condition (5.4) takes the form

h2
a > |b| + p. (5.7)
2
In Fig. 5.1 the regions of asymptotic mean-square stability for (5.5), given by condi-
tion (5.7), are shown for h = 0.9 and different values of the parameter p: (1) p = 0,
(2) p = 0.5, (3) p = 1, (4) p = 1.5, (5) p = 2, (6) p = 2.5, (7) p = 3, (8) p = 3.5.

5.1.1.2 The Second Way of Constructing a Lyapunov Functional

Put
 
∞ ∞  
k= dK(s), k1 = s dK(s) (5.8)
0 0
5.1 Systems with Nonlinearities in Stochastic Part 115

Fig. 5.1 Stability regions for (5.5) given by conditions (5.7) for h = 0.9 and different values of p:
(1) p = 0, (2) p = 0.5, (3) p = 1, (4) p = 1.5, (5) p = 2, (6) p = 2.5, (7) p = 3, (8) p = 3.5

and represent (5.1) in the form

ż(xt ) = −(a − k)x(t) + σ (t, xt )ẇ(t),

 ∞ t (5.9)
z(xt ) = x(t) + x(θ ) dθ dK(s).
0 t−s

Note that
 ∞ t  ∞ s
x(θ ) dθ dK(s) = x(t − θ ) dθ dK(s)
0 t−s 0 0
 ∞  ∞
= x(t − θ ) dK(s) dθ
0 θ

and
 ∞  ∞  ∞ s
  
dK(s) dθ = dθ dK(s) = k1 .
0 θ 0 0
116 5 Stability of Systems with Nonlinearities

By (5.9) and (2.10) we have to suppose that k1 < 1, and by (2.34) we will con-
sider the auxiliary differential equation ẏ(t) = −(a − k)y(t) with the necessary and
sufficient condition a > k for asymptotic stability of the trivial solution.
Put V1 = z2 (xt ). Calculating LV1 and using (5.8), (5.9), (5.2), and (5.3), we have

LV1 = −2z(xt )(a − k)x(t) + σ 2 (t, xt )

 ∞ t

 2   
≤ (a − k) −2x 2 (t) + x (t) + x 2 (θ ) dθ dK(s)
0 t−s
 ∞
+R x 2 (t − s) dR(s)
0
 ∞ 
 t
= −(a − k)(2 − k1 )x (t) + (a − k)
2 dK(s) x 2 (θ ) dθ
0 t−s
 ∞
+R x 2 (t − s) dR(s).
0

So, we obtain the representation of type (2.35) with D = −(a − k)(2 − k1 ), dμ0 =
(a − k)|dK(s)|, n = 1, and dK1 (s) = R dR(s). From Theorem 2.5 it follows that if

p 1
a>k+ , k1 < 1, p = R 2 , (5.10)
1 − k1 2

then the trivial solution of (5.1) is asymptotically mean-square stable.

Example 5.2 In the case (5.6), condition (5.10) takes the form
 
h2 h3 −1 h3
a>b + p 1 − |b| , |b| < 1. (5.11)
2 6 6

In Fig. 5.2 the regions of asymptotic mean-square stability for (5.5), given by condi-
tion (5.11) are shown for h = 0.9 and different values of the parameter p: (1) p = 0,
(2) p = 0.1, (3) p = 0.3, (4) p = 0.7, (5) p = 1.2, (6) p = 1.7, (7) p = 2.5,
(8) ph = 3.

5.1.1.3 The Third Way of Constructing a Lyapunov Functional

Let us show that in some particular cases using a special way of LV1 estimation we
can get new stability conditions.
Consider (5.5) and suppose that b ≤ 0. Putting V1 = x 2 , by (5.5), (5.2), and (5.3)
we obtain
5.1 Systems with Nonlinearities in Stochastic Part 117

Fig. 5.2 Stability regions for (5.5) given by conditions (5.11), for h = 0.9 and different values
of p: (1) p = 0, (2) p = 0.1, (3) p = 0.3, (4) p = 0.7, (5) p = 1.2, (6) p = 1.7, (7) p = 2.5,
(8) ph = 3

  h 
LV1 = 2x(t) −ax(t) + b (h − s)x(t − s) ds + σ 2 (t, xt )
0
 h  ∞
≤ −2ax (t) − 2|b|x(t)
2
(h − s)x(t − s) ds + R x 2 (t − s) dR(s).
0 0
(5.12)
Choosing V2 in the form
 h  t 2  ∞ t
V2 = |b| x(θ ) dθ ds + R x 2 (θ ) dθ dR(s),
0 t−s 0 t−s

we have
 h t  
LV2 = 2|b| x(θ ) dθ x(t) − x(t − s) ds
0 t−s
 ∞
+ R x (t) − R
2 2
x 2 (t − s) dR(s). (5.13)
0
118 5 Stability of Systems with Nonlinearities

Note that
 h t  
x(θ ) dθ x(t) − x(t − s) ds
0 t−s
 h s  
= x(t − θ ) dθ x(t) − x(t − s) ds
0 0
 h s  h  s
= x(t) x(t − θ ) dθ ds − x(t − s) x(t − θ ) dθ ds (5.14)
0 0 0 0

and
 h s  h h  h
x(t − θ ) dθ ds = ds x(t − θ ) dθ = (h − s)x(t − s) ds. (5.15)
0 0 0 θ 0

Besides, changing the order of integration, we have

 h  s  h  h
x(t − s) x(t − θ ) dθ ds = x(t − θ ) x(t − s) ds dθ
0 0 0 θ
 h  h
= x(t − s) x(t − θ ) dθ ds.
0 s

Therefore,
 h  s  h  s
2 x(t − s) x(t − θ ) dθ ds = x(t − s) x(t − θ ) dθ ds
0 0 0 0
 h  h
+ x(t − s) x(t − θ ) dθ ds
0 s
 h 2
= x(t − s) ds
0
≥ 0,

and from (5.13)–(5.15) we obtain

 h  ∞
LV2 ≤ R x (t) + 2|b|x(t)
2 2
(h − s)x(t − s) ds − R x 2 (t − s) dR(s). (5.16)
0 0

As a result, for the functional V = V1 + V2 , from (5.12) and (5.16) it follows that
1
LV ≤ −2(a − p)x 2 (t), p = R2.
2
So, if
a > p, b ≤ 0, (5.17)
then the trivial solution of (5.5) is asymptotically mean-square stable.
5.1 Systems with Nonlinearities in Stochastic Part 119

Fig. 5.3 Stability regions for (5.5) given by conditions (5.18), (5.19), for h = 0.9 and different
values of p: (1) p = 0, (2) p = 0.1, (3) p = 0.3, (4) p = 0.7, (5) p = 1.2, (6) p = 1.7, (7) p = 2.5,
(8) ph = 3

Using (5.7), (5.11), and (5.17), we can get a sufficient condition for asymptotic
mean-square stability of the trivial solution of (5.5) in the following form.
If ph < 3, then
⎧ h2
⎪
⎪b + p, b ≥ 0,
⎨ 2
a > b h2 + p(1 + b h3 )−1 , −2(3 − ph)h−3 ≤ b < 0, (5.18)
⎪
⎪ 2 6
⎩
p, b < −2(3 − ph)h−3 .

If ph ≥ 3, then
 2
b h2 + p, b ≥ 0,
a> (5.19)
p, b < 0.
In Fig. 5.3 the regions of asymptotic mean-square stability for (5.5), given by
conditions (5.18) and (5.19), are shown for h = 0.9 and different values of the pa-
rameter p: (1) p = 0, (2) p = 0.1, (3) p = 0.3, (4) p = 0.7, (5) p = 1.2, (6) p = 1.7,
(7) p = 2.5, (8) ph = 3.
120 5 Stability of Systems with Nonlinearities

5.1.2 Scalar Second-Order Differential Equation

Consider the scalar linear second-order differential equation with distributed delays
 ∞  ∞
ẍ(t) = −a ẋ(t) − ẋ(t − s) dK0 (s) − x(t − s) dK1 (s)
0 0

+ σ (t, xt , ẋt )ẇ(t), x0 = φ, ẋ0 = φ̇. (5.20)

Here Ki (s), i = 0, 1, are functions of bounded variation, and the functional

σ (t, ϕ, ψ) satisfies the condition
 ∞  ∞
σ (t, ϕ, ψ) ≤
2
ϕ (−s) dr1 (s) +
2
ψ 2 (−s) dr2 (s), (5.21)
0 0

where ri (s) are nondecreasing functions such that

 ∞
ri = dri (s) < ∞, i = 1, 2. (5.22)
0

Let us obtain sufficient conditions for asymptotic mean-square stability of the

trivial solution of (5.20) using the proposed procedure of Lyapunov functionals con-
struction.
Put x1 = x, x2 = ẋ,
 ∞  ∞
 
αij = s i dKj (s), βij = s i dKj (s).
0 0

Following the first step of the procedure, represent (5.20) in the form

ż1 (t) = x2 (t),

(5.23)
ż2 (t) = −bx1 (t) − ax2 (t) + σ (t, xt , ẋt )ẇ(t),

where b = β01 , z(t) = (z1 (t), z2 (t)),

z1 (t) = x1 (t),
 ∞  ∞ t (5.24)
z2 (t) = x2 (t) + x1 (t − θ ) dK0 (θ ) − x1 (s) ds dK1 (θ ).
0 0 t−θ

In this case the auxiliary system has the form

   
y 0 1
ẏ(t) = Ay(t), y = 1 , A = . (5.25)
y2 −b −a

Note that the matrix Lyapunov equation (1.27) with the matrix A from (5.25) and
the matrix Q with the elements q11 = q > 0, q22 = 1, q12 = 0 is equivalent to the
system

2bp12 = q, 2(ap22 − p12 ) = 1, p11 − ap12 − bp22 = 0 (5.26)

5.1 Systems with Nonlinearities in Stochastic Part 121

and has (see Corollary 1.1) the positive definite solution P with elements pij if and
only if a > 0, b > 0. So, the function v(y) = y  P y is a Lyapunov function for (5.25).
We will construct a Lyapunov functional V for (5.23) in the form V = V1 + V2 ,
where V1 = z (t)P z(t). Calculating LV1 , where L is the generator of (5.23), and
using (5.24) and (5.26), we have
 
LV1 = 2 p11 x1 (t) + p12 z2 (t) x2 (t)
  
− 2 p12 x1 (t) + p22 z2 (t) bx1 (t) + ax2 (t) + p22 σ 2 (t, x1t , x2t )
= −qx12 (t) − x22 (t) + p22 σ 2 (t, x1t , x2t )
 ∞  ∞ t 
− 2bp22 x1 (t)x1 (t − θ ) dK0 (θ ) − x1 (t)x1 (s) ds dK1 (θ )
0 0 t−θ
 ∞  ∞ t
− x2 (t)x1 (t − θ ) dK0 (θ ) + x2 (t)x1 (s) ds dK1 (θ ).
0 0 t−θ

Putting α = α00 + α11 , for some γ > 0, we obtain

 ∞  ∞ 
LV1 ≤ −qx12 (t) − x22 (t) + p22 x12 (t − s) dr1 (s) + x22 (t − s) dr2 (s)
0 0
 ∞  
+ bp22 x12 (t) + x12 (t − θ ) dK0 (θ )
0
 ∞ t

   
+ x12 (t) + x12 (s) ds dK1 (θ )
0 t−θ
 ∞  x 2 (t) 
1  
+ 2
+ γ x12 (t − θ ) dK0 (θ )
2 0 γ
 ∞  

x22 (t)
t  
+ + γ x12 (s) ds dK1 (θ )
0 t−θ γ
 
α
= (−q + αbp22 )x1 (t) + −1 +
2
x 2 (t)
2γ 2
 ∞    
γ  

+ x1 (t − θ ) p22 dr1 (θ ) + bp22 +
2
dK0 (θ )
0 2
 ∞   ∞  t
γ  
+ p22 x2 (t − s) dr2 (s) + bp22 +
2
x12 (s) ds dK1 (θ ).
0 2 0 t−θ

By Theorem 2.5 the stability condition takes the form

αγ α
(2αb + r1 )p22 + < q, + p22 r2 < 1. (5.27)
2 2γ
122 5 Stability of Systems with Nonlinearities

From (5.26) it follows that p22 = (q + b)(2ab)−1 . Substituting p22 into (5.27) and
using the condition
2αb + r1
< 1, (5.28)
2ab
we obtain
  −1  
αγ αb r1 2αb + r1 2ab α r2
+ + 1− <q < 1− − . (5.29)
2 a 2a 2ab r2 2γ 2a

So, if for some γ > 0,

    
αγ αb r1 2αb + r1 −1 2ab α r2
+ + 1− < 1− − , (5.30)
2 a 2a 2ab r2 2γ 2a

then there exists q > 0 such that inequalities (5.29) hold. From (5.30) it follows that
 
α 2b(a − α) − r1 2b(a − α) − r1
γ+ < − b.
2 γ r2 r2

To minimize the left-hand part of this inequality, put γ = (2b(a − α) − r1 )r2−1 .
Then we have αγ < γ 2 − b. From this and using the fact that system (5.23)–(5.24)
is a system of neutral type, we obtain
√

α 2 + α α 2 + 4b
2b(a − α) > r1 + r2 b + , α < 1. (5.31)
2

So, if conditions (5.31) hold, then the trivial solution of (5.20) is asymptotically
mean-square stable.

Remark 5.1 Note that instead of the representation (5.23)–(5.24) for z2 (t), we can
use some other representations:

ż2 (t) = −β01 x1 (t) − (a + β00 )x2 (t) + σ 2 (t, xt , ẋt )ẇ(t),
 ∞ t  ∞ t
z2 (t) = x2 (t) − x1 (s) ds dK1 (θ ) − x2 (s) ds dK0 (θ ),
0 t−θ 0 t−θ

or

ż2 (t) = −β01 x1 (t) − (a − β11 )x2 (t) + σ 2 (t, xt , ẋt )ẇ(t),
 ∞  ∞ t
z2 (t) = x2 (t) + x1 (t − θ ) dK0 (θ ) + (s − t + θ )x2 (s) ds dK1 (θ ),
0 0 t−θ
5.1 Systems with Nonlinearities in Stochastic Part 123

or

ż2 (t) = −β01 x1 (t) − (a + β00 − β11 )x2 (t) + σ 2 (t, xt , ẋt )ẇ(t),
 ∞ t  ∞ t
z2 (t) = x2 (t) − x2 (s) ds dK0 (θ ) + (s − t + θ )x2 (s) ds dK1 (θ ).
0 t−θ 0 t−θ

Using these different representations, we can obtain different stability conditions for
asymptotic mean-square stability of the trivial solution of (5.20).

Remark 5.2 Consider some particular cases of condition (5.31).

(1) Let r2 = 0. Then (5.31) coincides with (5.28).

(2) Let dK0 (θ ) = 0 and dK1 (θ ) = bδ(θ ) dθ . Then α00 = α11 = 0 and (5.20), (5.31)
are respectively

ẍ(t) = −a ẋ(t) − bx(t) + σ (t, xt , ẋt )ẇ(t),

r1
2a > r2 + , b > 0.
b

(3) Let dK0 (θ ) = 0, dK1 (θ ) = bδ(θ − h) dθ , and σ (t, xt , ẋt ) = σ x(t − τ ). Then
α00 = 0, α11 = bh, r1 = σ 2 , r2 = 0, and (5.20) and the stability conditions
(5.31) respectively take the forms

ẍ(t) = −a ẋ(t) − bx(t − h) + σ x(t − τ )ẇ(t), (5.32)

σ2
a > bh + , 0 < bh < 1. (5.33)
2b

Note that condition (3.70) for (5.32) coincides with (5.33), and condition
(3.61) for (5.32) takes the form

b σ2
a > bh + bh 2
+ , 0 < bh2 < 2. (5.34)
2(2 − bh ) 2b
2

Note also that conditions (5.33)–(5.34) follow from (3.73), (3.72), respectively,
if h1 = 0, h2 = h.
It is clear that if 0 < b < h−1 , then condition (5.33) is weaker than (5.34).
But, on the other hand, if h−1 ≤ b < 2h−2 , h < 2, then (5.33) cannot be used,
and (5.34) gives some additional stability region. In Fig. 5.4 the regions of
asymptotic mean-square stability for (5.32), given by conditions (5.33)–(5.34),
are shown for σ = 0.7 and different values of the delay h: (1) h = 1, (2) h = 1.2,
(3) h = 1.4.
124 5 Stability of Systems with Nonlinearities

Fig. 5.4 Stability regions

for (5.32), given by
conditions (5.33), (5.34) for
σ = 0.7 and different values
of the delay h: (1) h = 1,
(2) h = 1.2, (3) h = 1.4

5.2 Systems with Nonlinearities in Both Deterministic

and Stochastic Parts

Consider the nonlinear scalar differential equation of neutral type

   
dz(t) = b − ax(t) z(t) dt + σ dw(t) ,
(5.35)
z(t) = x(t) − cx(t − h),

where a > 0, b > 0, h ≥ 0, |c| < 1, σ are some known constants. Let us investi-
gate asymptotic mean-square stability of the equilibrium point x(t) ≡ x ∗ = a −1 b
with respect to perturbations of the initial function φ satisfying the following con-
dition:

z(0) = φ(0) − cφ(−h) > 0. (5.36)

Put h = 0. Then (5.35) takes the form

   
dx(t) = −a x(t) − x ∗ x(t) dt + σ dw(t) . (5.37)

Let us show that if

σ2
0 < pb < 1, p= , (5.38)
2
then the function
  x(t)
v x(t) = x(t) − x ∗ − x ∗ ln ∗ (5.39)
x
is a Lyapunov function for the auxiliary equation (5.37).
5.2 Systems with Nonlinearities in Both Deterministic and Stochastic Parts 125

First at all, note that v(x ∗ ) = 0, and since x − 1 − ln x ≥ 0 for x > 0, we have
 
  x(t) x(t)
v x(t) = x ∗ − 1 − ln ≥ 0, x(t) ≥ 0.
x∗ x∗

Using that ax ∗ = b, we obtain

 
    x∗  2 x∗
Lv x(t) = −a x(t) − x ∗ x(t) 1 − + pa 2 x(t) − x ∗ x 2 (t) 2
x(t) x (t)
 
= −a(1 − bp) x(t) − x ∗ .
2

So, by condition (5.38) the function (5.39) is a Lyapunov function for the solution
x ∗ of (5.37).
Using the procedure of constructing Lyapunov functionals, we will construct a
Lyapunov functional V for (5.35) in the form V = V1 + V2 , where

  z(t)
V1 = v z(t) = z(t) − z∗ − z∗ ln ∗ ,
z
z(t) = x(t) − cx(t − h), z∗ = (1 − c)x ∗ .

Calculating LV1 , by (5.35) and b = ax ∗ we have

 
  z∗  2 z∗
LV1 = b − ax(t) z(t) 1 − + p b − ax(t) z2 (t) 2
z(t) z (t)
    2
= −a x(t) − x ∗ z(t) − z∗ + abp(1 − c) x(t) − x ∗
    2
= −a x(t) − x ∗ x(t) − cx(t − h) − x ∗ + cx ∗ + abp(1 − c) x(t) − x ∗
  2   
= −a 1 − bp(1 − c) x(t) − x ∗ + ac x(t) − x ∗ x(t − h) − x ∗
  2 a|c|  2  2 
≤ −a 1 − bp(1 − c) x(t) − x ∗ + x(t) − x ∗ + x(t − h) − x ∗
2
 
|c|  2 a|c|  2
= −a 1 − bp(1 − c) − x(t) − x ∗ + x(t − h) − x ∗ .
2 2

Putting

a|c| t  2
V2 = x(s) − x ∗ ds,
2 t−h

for the functional V = V1 + V2 , we obtain

  2
LV ≤ −a 1 − bp(1 − c) − |c| x(t) − x ∗ .
126 5 Stability of Systems with Nonlinearities

Fig. 5.5 Stability regions for

(5.35) given by conditions
(5.40) for different values
of p: (1) p = 0, (2) p = 0.2,
(3) p = 0.4, (4) p = 0.6,
(5) p = 0.8

From this it follows that if conditions (5.36) and

 1+c
, −1 < c < 0,
0 < bp < 1−c (5.40)
1, 0 ≤ c < 1,

hold then the equilibrium point x ∗ (t) ≡ a −1 b of (5.35) is asymptotically mean-

square stable. Note that condition (5.40) does not depend on a and h.
In Fig. 5.5 the stability regions, given by condition (5.40), are shown for different
values of p: (1) p = 0, (2) p = 0.2, (3) p = 0.4, (4) p = 0.6, (5) p = 0.8.

5.3 Stability in Probability of Nonlinear Systems

Consider the nonlinear stochastic differential equation
 ∞ 
dx(t) = dK0 (s) x(t − s) + g0 (t, xt ) dt
0
m 
 ∞ 
+ dKi (s)x(t − s) + gi (t, xt ) dwi (t),
i=1 0

x0 = φ0 . (5.41)

Here x(t) ∈ Rn , w1 (t), . . . , wm (t) are the mutually independent scalar Wiener pro-
cesses, Ki (s), i = 0, . . . , m, are n × n-matrices such that
 ∞
 
αi = dKi (s) < ∞, (5.42)
0

and
A
is the operator norm of a matrix A. It is assumed also that the functionals
gi (t, ϕ), i = 0, . . . , m, satisfy the condition
 ∞
   
gi (t, ϕ) ≤ ϕ(−s)νi dri (s),
ϕ
0 ≤ δ, νi > 1, (5.43)
0
5.3 Stability in Probability of Nonlinear Systems 127

where δ is small enough,

ϕ
0 = sups≤0 |ϕ(s)|, and ri (s) are nondecreasing func-
tions satisfying the condition
 ∞
ri = dri (s) < ∞, i = 0, . . . , m. (5.44)
0

The generator L of (5.41) has the form

∂Vϕ (t, x) 1  
m
∂ 2 Vϕ (t, x) 
LV (t, xt ) = L0 V (t, xt ) + g0 (t, xt ) + gi (t, xt ) gi (t, xt )
∂x 2 ∂x 2
i=1


m  ∞
∂ 2 Vϕ (t, x)
+ gi (t, xt ) dKi (s) x(t − s), (5.45)
∂x 2 0
i=1

where
 ∞ 
∂Vϕ (t, x) ∂Vϕ (t, x)
L0 V (t, xt ) = + dK0 (s) x(t − s)
∂t 0 ∂x
  
1 ∞  2
∂ Vϕ (t, x) ∞
m
+ dKi (s) x(t − s) dKi (s) x(t − s)
2 0 ∂x 2 0
i=1
(5.46)

is the generator of the “linear part” of (5.41), i.e., of the linear differential equation
 ∞ m 
 ∞
dx(t) = dK0 (s) x(t − s) dt + dKi (s) x(t − s) dwi (t). (5.47)
0 i=1 0

Theorem 5.1 Let there exist a functional V0 (t, ϕ) ∈ D that satisfies conditions
(2.18)–(2.19) for p = 2 and
 2
L0 V0 (t, ϕ) ≤ −c0 ϕ(0) ,
 0   ∞  0  2 0 
 ∂Vϕ (t, x)     ∂ Vϕ (t, x) 
  ≤ c1 |x| + dq(τ ) ϕ(s) ds,   ≤ c2 ,
 ∂x   ∂x 2  (5.48)
0 −τ
 ∞
ci > 0, i = 0, 1, 2, γ= τ dq(τ ) < ∞.
0

Then the trivial solution of (5.41) is stable in probability.

Proof We will construct a functional V that satisfies the conditions of Theorem 2.2
(for p = 2). The functional V will be constructed in the form V = V0 + V1 , where
128 5 Stability of Systems with Nonlinearities

V0 satisfies conditions (5.48). Calculating LV , by (5.45) we have

∂Vϕ0 (t, x)
LV = LV0 + LV1 = L0 V0 + LV1 + g0 (t, xt )
∂x

m  ∞
∂ 2 Vϕ0 (t, x)
+ gi (t, xt ) dKi (s) x(t − s)
∂x 2 0
i=1

1 
m
∂ 2 Vϕ0 (t, x) 
+ gi (t, xt ) gi (t, xt ). (5.49)
2 ∂x 2
i=1

Supposing that |x(s)| ≤ δ for s ≤ t, by inequalities (5.43), (5.48) we estimate the

terms in (5.49). So, we get

    0 
  ∂V 0 (t, x)  ∞   ∂V (t, x) 
g (t, xt ) ϕ ≤ x(t − s)ν0 dr0 (s) ϕ 
 0 ∂x   ∂x 
0
 ∞ 
≤δ ν0 −1 x(t − s) dr0 (s)
0
   
  ∞ t  
× c1 x(t) + dq(τ ) x(s) ds
0 t−τ
 ∞  
= c1 δ ν0 −1 x(t − s)x(t) dr0 (s)
0
  
∞ ∞ t   
+ δ ν0 −1 dr0 (θ ) dq(τ ) x(t − θ )x(s) ds
0 0 t−τ

1 ∞     
≤ c1 δ ν0 −1 x(t − s)2 + x(t)2 dr0 (s)
2 0
  ∞
1 ν0 −1 ∞
+ δ dr0 (θ ) dq(τ )
2 0 0
 t
    
× x(t − θ )2 + x(s)2 ds
t−τ
  ∞ 
1  2  
= c1 δ ν0 −1 r0 x(t) + x(t − s)2 dr0 (s)
2 0
  ∞
1  
+ δ ν0 −1 γ x(t − θ )2 dr0 (θ )
2 0
 ∞  t 
 2
+ r0 dq(τ )  
x(s) ds
0 t−τ
5.3 Stability in Probability of Nonlinear Systems 129

1  2
= c1 δ ν0 −1 r0 x(t)
2
  ∞
1 ν0 −1  
+ δ (c1 + γ ) x(t − s)2 dr0 (s)
2 0
 ∞  t 
 2
+ r0 dq(τ )  
x(s) ds . (5.50)
0 t−τ

Similarly, we have
 m  
 ∂ 2 Vϕ0 (t, x) ∞ 
  
 gi (t, xt ) dKi (s) x(t − s)
 ∂x 2 0 
i=1
m 
 
∞  ν ∞  
≤ c2 dri (s) x(t − s) i dKi (τ )x(t − τ )
i=1 0 0

  ∞
c2  νi −1 ∞      
m
≤ δ dri (s) dKi (τ ) x(t − s)2 + x(t − τ )2
2 0 0
i=1
  ∞  ∞ 
c2  νi −1  2   2
m
= δ αi  
dri (s) x(t − s) + βi   
dKi (τ ) x(t − τ )
2 0 0
i=1
(5.51)

and
 m 
 ∂ 2 Vϕ0 (t, x) 
  
 gi (t, xt ) g i (t, x t )
 ∂x 2 
i=1


m
 
≤ c2 gi (t, xt )2
i=1
m 
 2
∞  ν
≤ c2 dri (s) x(t − s) i
i=1 0

 
m ∞  2
≤ c2 δ 2(νi −1) ri dri (s) x(t − s) . (5.52)
i=1 0

Let us define the functional V1 as follows:

 ∞  t
1 ν0 −1  
V1 = δ (c1 + γ ) dr0 (τ ) x(s)2 ds
2 0 t−τ
 ∞  t
1  2
+ δ ν0 −1 r0 dq(τ ) (s − t + τ )x(s) ds
2 0 t−τ
130 5 Stability of Systems with Nonlinearities

 ∞ 
c2  νi −1   t  
m
+ δ ri  dKi (τ ) x(s)2 ds
2 0 t−τ
i=1
  t
c2  νi −1   ∞  
m
+ δ αi + 2δ νi −1 ri dri (τ ) x(s)2 ds. (5.53)
2 0 t−τ
i=1

It is easy to see that 0 ≤ V1 (t, ϕ) ≤ c

ϕ
2 , c > 0. Therefore, the functional V =
V0 + V1 satisfies conditions (2.18)–(2.19). Calculating LV1 and using (5.48)–(5.53),
we obtain
   
c m
   2
LV (t, xt ) ≤ − c0 − δ ν0 −1 β0 δ νi −1 ri αi + δ νi −1 ri x(t) .
1
+ γ − c2
2
i=1

For small enough δ, the term in the square brackets is positive. Therefore, LV ≤ 0,
and the functional V satisfies the conditions of Theorem 2.2. So, the trivial solution
of (5.41) is stable in probability. The proof is completed. 

Remark 5.3 The asymptotic mean-square stability of the trivial solution of the lin-
ear differential equation (5.47) follows from the existence of the functional V0 that
satisfies the conditions of Theorem 5.1. Therefore, in order to obtain sufficient con-
ditions for stability in probability of the trivial solution of the nonlinear differential
equation (5.41) with the order of nonlinearity higher than one, it is enough to obtain,
by virtue of some Lyapunov functional, sufficient conditions for asymptotical mean-
square stability of the trivial solution of “the linear part” of (5.41), i.e., of (5.47). For
example, it is easy to show [261] that the functional
   2
 2  ∞ t 
V0 (t, xt ) = x(t) + ν x(t) + dK0 (τ ) x(s) ds 
+0 t−τ
 
∞  ∞  
 t  2
+ν  dK0 (θ ) dK0 (τ ) (s − t + τ )x(s) ds

+0 0 t−τ
 ∞ 
 t  
+ dK0 (τ ) x(s)2 ds
+0 t−τ
N  ∞
  
  ∞  t  
+ (ν + 1) dKi (θ ) dKi (τ ) x(s)2 ds,
i=1 0 0 t−τ

ν ≥ 0, satisfies the conditions of Theorem 2.1 if the matrix

 ∞  ∞
1   
Q= dK0 (s) + inf dK0 (s)I − dK0 (s)
0 ν≥0 ν + 1 +0
 ∞  ∞  
N  ∞ 



 1 

 2

+ν τ dK0 (s) dK0 (τ )I + dKi (s) I (5.54)
+0 0 2 0 i=1
5.4 Systems with Fractional Nonlinearity 131

is a negative definite matrix, i.e., x  Qx ≤ −c|x|2 . Here I is the identity matrix, and
c > 0.

So, we obtain the following theorem.

Theorem 5.2 Let conditions (5.42)–(5.44) hold, and the matrix (5.54) be a negative
definite matrix. Then the trivial solution of (5.47) is asymptotically mean-square
stable, and the trivial solution of (5.41) is stable in probability.

Example 5.3 Consider the well-known Volterra population equation

  
1 ∞
ẋ(t) = ax(t) 1 − x(t − s) dH (s) , (5.55)
K 0
where
 ∞  ∞
a > 0, K > 0, dH (s) ≥ 0, dH (s) = 1, s dH (s) < ∞.
0 0

Let us assume that the parameter a is susceptible to stochastic perturbations of

the type of white noise ẇ(t). Then (5.55) is transformed to the stochastic integro–
differential equation
  
  1 ∞
ẋ(t) = a 1 + σ ẇ(t) x(t) 1 − x(t − s) dH (s) . (5.56)
K 0

Substituting x(t) = K(1 + y(t)) into (5.56) and keeping only the linear part of the
obtained equation, we get the linearization of (5.56) in the neighborhood of the
steady-state solution x(t) = K
 ∞  ∞
ẏ(t) = −a y(t − s) dH (s) − aσ y(t − s) dH (s) ẇ(t). (5.57)
0 0

By Theorem 5.2 and (5.54) we obtain that the inequality

 ∞  ∞

aσ 2
min 2 dH (s), a s dH (s) + <1
+0 +0 2

is a sufficient condition for asymptotic mean-square stability of the trivial solu-

tion of (5.57) and for stability in probability of the steady-state solution x(t) = K
of (5.55).

5.4 Systems with Fractional Nonlinearity

Here a nonlinear differential equation with a fractional nonlinearity is considered,
and it is supposed that this equation has an equilibrium point and is exposed to addi-
132 5 Stability of Systems with Nonlinearities

tive stochastic perturbations of the type of white noise that are directly proportional
to the deviation of the system state from the equilibrium point. Stochastic perturba-
tions of such a form were first proposed by the author in [27, 267] and successfully
used later by other researchers (see, for instance, [19, 38, 39, 49, 117, 214, 257]).
The results from the previous paragraph are used below for investigation of sta-
bility in probability of the equilibrium points of a stochastic fractional differential
equation. Numerous graphical illustrations of stability regions and trajectories of
solutions are plotted.

5.4.1 Equilibrium Points

Nonlinear delay differential equation of type ẋ(t) = −ax(t) + f (x(t − τ )) is rather

popular among researchers. See, for example, [101, 183, 248], the famous Nicholson
blowflies equation [99, 102, 224] ẋ(t) = −ax(t) + bx(t − τ )e−γ x(t−τ ) , the Mackey–
Glass model [190]
bx(t − τ )
ẋ(t) = −ax(t) + .
1 + x n (t − τ )
On the other hand, recently there is a very large interest in studying the behavior of
solutions of nonlinear difference equations with fractional nonlinearity of the type
k
μ+ j =0 aj xn−j
xn+1 = k , n = 0, 1, . . . ,
λ+ j =0 bj xn−j

(see [95, 162–164, 232, 278] and a long list of the references therein).
Here, similarly to [232, 278], the stability of equilibrium points of the nonlinear
differential equation with fractional nonlinearity
k
μ+ j =0 aj x(t − τj )
ẋ(t) = −ax(t) + k , t > 0, (5.58)
λ+ j =0 bj x(t − τj )

and the initial condition

x(s) = φ(s), s ∈ [−τ, 0], τ = max{τ1 , . . . , τk }, (5.59)

is investigated. Here μ, λ, aj , bj , j = 0, . . . , k, τ0 = 0, τj > 0, j > 0, are known

constants.
Put

k 
k
A= aj , B= bj , (5.60)
j =0 j =0
5.4 Systems with Fractional Nonlinearity 133

and suppose that (5.58) has an equilibrium point x̂ (not necessarily a positive one).
By (5.58), (5.60), and the assumption

λ + B x̂ = 0 (5.61)

the equilibrium point x̂ is defined by the algebraic equation

μ + Ax̂
a x̂ = . (5.62)
λ + B x̂
If aB = 0, then by condition (5.61), (5.62) can be transformed to the form

aB x̂ 2 − (A − aλ)x̂ − μ = 0. (5.63)

Thus, if
(A − aλ)2 + 4aBμ > 0, (5.64)
then (5.58) has two equilibrium points

A − aλ + (A − aλ)2 + 4aBμ
x̂1 = (5.65)
2aB
and

A − aλ − (A − aλ)2 + 4aBμ
x̂2 = ; (5.66)
2aB
if
(A − aλ)2 + 4aBμ = 0, (5.67)
then (5.58) has only one equilibrium point,

A − aλ
x̂ = . (5.68)
2aB
Finally, if
(A − aλ)2 + 4aBμ < 0, (5.69)
then (5.58) has no equilibrium points.

Remark 5.4 Assume that aB = 0 and μ = 0. If A = 0 and A = aλ, then (5.58) has
two equilibrium points,

A − aλ
x̂1 = and x̂2 = 0; (5.70)
aB
if A = 0 or A = aλ, then (5.58) has only one equilibrium point x̂ = 0.
134 5 Stability of Systems with Nonlinearities

Remark 5.5 Assume that aB = 0. If A = aλ, then (5.58) has only one equilibrium
point
μ
x̂ = − .
A − aλ

Remark 5.6 Consider the case μ = B = 0, λ = 0. If A = aλ, then (5.58) has only
one equilibrium point x̂ = 0; if A = aλ, then each solution x̂ = const is an equilib-
rium point of (5.58).

5.4.2 Stochastic Perturbations, Centering, and Linearization

Suppose that (5.58) is exposed to stochastic perturbations of the type of white noise
ẇ(t) that are proportional to the deviation of the state x(t) of (5.58) from the equi-
librium point x̂. Then (5.58) takes the form
k
μ+ j =0 aj x(t − τj )  
ẋ(t) = −ax(t) + k + σ x(t) − x̂ ẇ(t). (5.71)
λ+ i=0 bi x(t − τi )

Note that the equilibrium point x̂ of (5.58) is also an equilibrium point of (5.71).
Putting x(t) = y(t) + x̂ and

aj − abj x̂
γj = , j = 0, . . . , k, (5.72)
λ + B x̂

we will center (5.71) in the neighborhood of the equilibrium point x̂. From (5.71)–
(5.72) it follows that y(t) satisfies the equation

γ0 y(t) + kj =1 γj y(t − τj )
ẏ(t) = −ay(t) +  + σy(t)ẇ(t). (5.73)
1 + ki=0 bi (λ + B x̂)−1 y(t − τi )

It is clear that the stability of the trivial solution of (5.73) is equivalent to the stability
of the equilibrium point of (5.71).
Together with the nonlinear differential equation (5.73), we will consider the
linear part (in a neighborhood of the zero) of (5.73)


k
ż(t) = −(a − γ0 )z(t) + γj z(t − τj ) + σ z(t)ẇ(t). (5.74)
j =1

Below, the following method for stability investigation is used. Conditions for
asymptotic mean-square stability of the trivial solution of the constructed linear dif-
ferential equation (5.74) were obtained by the procedure of constructing Lyapunov
functionals. Since the order of nonlinearity of (5.73) is higher than one, the obtained
5.4 Systems with Fractional Nonlinearity 135

sufficient conditions for asymptotic mean-square stability at the same time are (The-
orem 5.2) sufficient conditions for stability in probability of the trivial solution of
the nonlinear differential equation (5.73) and therefore for stability in probability of
the equilibrium point of (5.71).

5.4.3 Stability of Equilibrium Points

Note that the differential equation (5.74) is an equation of type (3.29). So, from
(3.31) and (3.33) we obtain two following sufficient conditions for asymptotic
mean-square stability of the trivial solution of (5.74):


k
σ2
a > γ0 + |γj | + p, p= , (5.75)
2
j =1

and
  

k 
k 
k
a− γj 1− |γj |τj > p, |γj |τj < 1. (5.76)
j =0 j =1 j =1

Remark 5.7 If the delays are absent, i.e., τj = 0, j = 0, . . . , k, then condition (5.76)
is not worse than (5.75) that does not depend on delays.

Suppose at first that condition (5.67) holds. In this case, (5.71) has only one
equilibrium point x̂ that is defined by (5.68), and by (5.72), (5.68) we have


k
A − aB x̂ A − 12 (A − aλ)
γj = = = a.
λ + B x̂ λ + 2a
1
(A − aλ)
j =0

Thus, the stability condition (5.76) for the equilibrium point (5.68) does not hold.
Moreover,

k 
k
a= γj ≤ γ0 + |γj |.
j =0 j =1

Thus, the stability condition (5.75) for equilibrium point (5.68) does not hold too.
Suppose now that condition (5.64) holds. Then (5.71) has two equilibrium points
x̂1 and x̂2 that are defined in (5.65) and (5.66), respectively. Put

S = (A − aλ)2 + 4aBμ, (5.77)

(l) aj − abj x̂l

γj = , j = 0, . . . , k, l = 1, 2. (5.78)
λ + B x̂l
136 5 Stability of Systems with Nonlinearities

(l)
Corollary 5.1 Assume that condition (5.64) holds and γ0 ≥ 0, l = 1, 2. Then for
fixed μ and λ, condition (5.75) cannot be true for both equilibrium points x̂1 and x̂2
together.

Proof By (5.75), (5.78), and (5.65), for x̂1 , we obtain

 k   
1  (1)  1  (1)  1  A − aB x̂1 
k
1> γj ≥  γj  =
a a  a  λ + B x̂1 
j =0 j =0
   
 A − 12 (A − aλ + S)   A + aλ − S 

=  =  .
aλ + 1 (A − aλ + S)   A + aλ + S 
2

Similarly, for x̂2 , we have

 k   
1  (2)  1  (2)  1  A − aB x̂2 
k
1> γj ≥  γj  =
a a  a  λ + B x̂2 
j =0 j =0
   
 A − 12 (A − aλ − S)   A + aλ + S 
=  = .
aλ + 1 (A − aλ − S)   A + aλ − S 
2

Thus, we obtain two conflicting conditions. The proof is completed. 

Corollary 5.2 Assume that condition (5.64) holds and a = 0. If

 
2aS 
k
 (1)  
k
 (1) 
1− γ τj > p, γ τj < 1, (5.79)
S + A + aλ j j
j =1 j =1

then the equilibrium point x̂ = x̂1 (defined in (5.65)) of (5.71) is stable in probability.
If
 
2aS  k
 (2)   k
 (2) 
1− γ τj > p, γ τj < 1, (5.80)
S − A − aλ j j
j =1 j =1

then the equilibrium point x̂ = x̂2 (defined in (5.66)) of (5.71) is stable in probability.
Assume now that a = 0. If

  
k
A2 |A| Q
1− τ > p, τ= |aj |τj < , Q = Bμ − Aλ, (5.81)
Q Q |A|
j =1

then the equilibrium point x̂ = −μA−1 is stable in probability.

5.4 Systems with Fractional Nonlinearity 137

Proof For a = 0, by (5.76) it is enough to note that, for x̂1 ,


k
(1) A − aB x̂1 A − 12 (A − aλ + S) 2aS
a− γj =a− =a− = ,
λ + B x̂1 λ + 2a (A − aλ + S) S + A + aλ
1
j =0

and, similarly, for x̂2 ,


k
A − aB x̂2 A − 12 (A − aλ − S) 2aS
a− γj(2) = a − =a− = .
λ + B x̂2 λ + 2a (A − aλ − S)
1 S − A − aλ
j =0

For a = 0, by Remark 5.5 from (5.72), (5.76) we obtain


k
A A2 
k
τ |A|
− γj = − −1
= , |γj |τj = −1
= τ.
λ − BμA Q |λ + B(−μ)A | Q
j =0 j =0

So, (5.76) implies (5.81). The proof is completed. 

Remark 5.8 If τj = 0, j = 1, . . . , k, then conditions (5.79), (5.80) take the forms

2aS 2aS
> p, > p, (5.82)
S + A + aλ S − A − aλ
respectively. Moreover, inequalities (5.82) are necessary conditions for implemen-
tation of conditions (5.79), (5.80) with arbitrary τj , j = 1, . . . , k. If a < 0 or
p ≥ 2a > 0, then conditions (5.79), (5.80) cannot be true for the same μ and λ. Re-
ally, if a < 0, then from (5.82) the contradiction follows: 0 < S < A + aλ < −S < 0.
If p ≥ 2a > 0, then from (5.82) another contradiction follows: 0 ≤ −(2a/p − 1)S <
A + aλ < (2a/p − 1)S ≤ 0.

Corollary 5.3 Put

⎧
⎪
⎪
p
if p < 2a,
⎨ 2a−p
q = p
if p > 2a, (5.83)
⎪ p−2a
⎪
⎩+∞ if p = 2a,
assume that a = 0, S > 0, τj = 0, j = 1, . . . , k, and consider the following four
cases.
Case 1: a > 0, B > 0.
If
⎧ 2
⎨ q (A+aλ)2 −(A−aλ)2 for λ ≥ − Aa ,
4aB
p < 2a, μ> (5.84)
⎩Aλ for λ < − Aa ,
B
138 5 Stability of Systems with Nonlinearities

or

A q 2 (A + aλ)2 − (A − aλ)2 A
p ≥ 2a, λ<μ< , λ<− , (5.85)
B 4aB a

then the equilibrium point x̂1 is stable in probability.

If
⎧ 2
⎨ q (A+aλ) −(A−aλ)
2 2
4aB for λ < − Aa ,
p < 2a, μ> (5.86)
⎩Aλ for λ ≥ − Aa ,
B

or

A q 2 (A + aλ)2 − (A − aλ)2 A
p ≥ 2a, λ<μ< , λ≥− , (5.87)
B 4aB a

then the equilibrium point x̂2 is stable in probability.

Case 2: a > 0, B < 0.

If
⎧ 2
⎨ q (A+aλ) −(A−aλ)
2 2
4aB for λ ≥ − Aa ,
p < 2a, μ< (5.88)
⎩Aλ for λ < − Aa ,
B

or

A q 2 (A + aλ)2 − (A − aλ)2 A
p ≥ 2a, λ>μ> , λ<− , (5.89)
B 4aB a

then the equilibrium point x̂1 is stable in probability.

If
⎧ 2
⎨ q (A+aλ) −(A−aλ)
2 2
4aB for λ < − Aa ,
p < 2a, μ< (5.90)
⎩Aλ for λ ≥ − Aa ,
B

or

A q 2 (A + aλ)2 − (A − aλ)2 A
p ≥ 2a, λ>μ> , λ≥− , (5.91)
B 4aB a

then the equilibrium point x̂2 is stable in probability.

5.4 Systems with Fractional Nonlinearity 139

Case 3: a < 0, B > 0.

If

A q 2 (A + aλ)2 − (A − aλ)2 A
λ<μ< , λ>− , (5.92)
B 4aB a
then the equilibrium point x̂1 is stable in probability.
If

A q 2 (A + aλ)2 − (A − aλ)2 A
λ<μ< , λ<− , (5.93)
B 4aB a
then the equilibrium point x̂2 is stable in probability.
Case 4: a < 0, B < 0.
If

A q 2 (A + aλ)2 − (A − aλ)2 A
λ>μ> , λ>− , (5.94)
B 4aB a
then the equilibrium point x̂1 is stable in probability.
If

A q 2 (A + aλ)2 − (A − aλ)2 A
λ>μ> , λ<− , (5.95)
B 4aB a
then the equilibrium point x̂2 is stable in probability.

Proof It is enough to prove Case 1; the proofs of the others cases are similar.
Consider the equilibrium point x̂1 . Assume first that p < 2a. If A + aλ ≥ 0, then
by (5.77) from the first line of (5.84) it follows that S > q(A + aλ). By (5.83) and
τj = 0, j = 1, . . . , k, this inequality coincides with (5.79). If A + aλ < 0, then from
the second line of (5.84) we have Bμ > Aλ. So, by (5.77) we obtain S > |A + aλ|
and, therefore, S > S + A + aλ > 0. From this and from 2a > p condition (5.79)
with τj = 0, j = 1, . . . , k, follows.
Let now p > 2a. Then, by (5.85), A + aλ < 0 and Bμ > Aλ. Thus, from
(5.77), (5.85) it follows that q|A + aλ| > S > |A + aλ|. From this by (5.83) con-
dition (5.79) with τj = 0, j = 1, . . . , k, follows. Finally, if p = 2a, then (5.85)
is equivalent to Bμ > Aλ and S > |A + aλ|, which implies (5.79) with τj = 0,
j = 1, . . . , k.
Consider the equilibrium point x̂2 . Assume first that p < 2a. If A + aλ ≥ 0,
then from the second line of (5.86) it follows that Bμ > Aλ. In view of (5.77), this
means that S > A + aλ, and therefore S > S − A − aλ. From this and from p < 2a
condition (5.80) with τj = 0, j = 1, . . . , k, follows. If A + aλ < 0, then from the
first line of (5.86) we obtain S > q|A + aλ|. From this and from (5.83) condition
(5.80) with τj = 0, j = 1, . . . , k, follows.
140 5 Stability of Systems with Nonlinearities

Fig. 5.6 Stability regions for the equilibrium points x̂1 and x̂2 of (5.96) by a = 1, a1 = 1.5,
a2 = −0.5, b1 = 1.2, b2 = 1.8, τ1 = 0.4, τ2 = 0.3, σ = 1.2

Let now p > 2a. Then, by (5.87), A + aλ ≥ 0 and Bμ > Aλ. Thus, from (5.77),
(5.87) it follows that q(A + aλ) > S > A + aλ ≥ 0. From this and from (5.83)
condition (5.80) with τj = 0, j = 1, . . . , k, follows. Finally, if p = 2a, then (5.87)
is equivalent to Bμ > Aλ, and, by (5.77), S > A + aλ, which implies (5.80) with
τj = 0, j = 1, . . . , k. The proof is completed. 
k
Corollary 5.4 Put τ = j =1 |aj |τj , Q = Bμ − Aλ and assume that a = 0, AB =
0, τ < Q|A|−1 . If
 
Aλ A2 (1 − 1 − 4pτ |A−1 |) Aλ A2 (1 + 1 − 4pτ |A−1 |)
+ <μ< + , B > 0,
B 2pB B 2pB
or
 
Aλ A2 (1 + 1 − 4pτ |A−1 |) Aλ A2 (1 − 1 − 4pτ |A−1 |)
+ <μ< + , B < 0,
B 2pB B 2pB

then the equilibrium point x̂ = −μA−1 is stable in probability.

5.4 Systems with Fractional Nonlinearity 141

Fig. 5.7 Stability regions for the equilibrium points x̂1 and x̂2 of (5.96) by a = 1, a1 = 1.5,
a2 = −0.5, b1 = 1.2, b2 = 1.8, τ1 = 0.15, τ2 = 0.01, σ = 2.2

Proof It is enough to note that the given conditions are the solution of the inequality

pQ2 − A2 Q + A2 |A|τ < 0, where Q = Bμ − Aλ > 0,

which is equivalent to (5.81). 

5.4.4 Numerical Analysis

Below, for numerical simulation of the solutions of the equations of type (5.71),
we use difference analogues of the considered equations [278] and the algorithm of
numerical simulation of the Wiener process trajectories (Sect. 2.1.1).

Example 5.4 Consider the differential equation

μ + a1 x(t − τ1 ) + a2 x(t − τ2 )  
ẋ(t) = −ax(t) + + σ x(t) − x̂ ẇ(t), (5.96)
λ + b1 x(t − τ1 ) + b2 x(t − τ2 )
which is an equation of type (5.71) with k = 2, a0 = b0 = 0.
142 5 Stability of Systems with Nonlinearities

Fig. 5.8 Stability regions for the equilibrium points x̂1 and x̂2 of (5.96) by a = 1, a1 = −1.5,
a2 = −0.5, b1 = −1.2, b2 = −1.8, τ1 = 0.3, τ2 = 0.4, σ = 1.1

Case 1. Put a = 1, a1 = 1.5, a2 = −0.5, b1 = 1.2, b2 = 1.8, τ1 = 0.4, τ2 = 0.3,

σ = 1.2. Thus, a > 0, B = 1.2 + 1.8 = 3 > 0, p = 0.72 < 2a = 2, A = 1.5 − 0.5 =
1 > 0.
In Fig. 5.6 the regions of stability in probability for the equilibrium points x̂1
and x̂2 are shown in the space of parameters (μ, λ): in the white region there are
no equilibrium points; in the yellow region there are possible unstable equilibrium
points; red, cyan, magenta, and grey regions are the regions for stability in proba-
bility of the equilibrium point x̂ = x̂1 given by condition (5.75) (red and cyan) and
condition (5.79) (cyan, magenta, and grey); blue, green, and grey regions are the re-
gions for stability in probability of the equilibrium point x̂ = x̂2 given by condition
(5.75) (blue) and condition (5.80) (blue, green, and grey); in the grey region both
equilibrium points x̂ = x̂1 and x̂ = x̂2 are stable in probability. Curves 1 and 2 are
the bounds of the equilibrium points x̂1 and x̂2 stability regions, respectively, given
by conditions (5.84) and (5.86) for the case τ1 = τ2 = 0. One can see that the sta-
bility regions obtained for positive delays are placed inside of the regions with zero
delays.
Put now τ1 = 0.15, τ2 = 0.01, σ = 2.2. Then p = 2.42 > 2a = 2, and condition
(5.75) does not hold. Appropriate stability regions obtained with the same values
5.4 Systems with Fractional Nonlinearity 143

Fig. 5.9 Stability regions for the equilibrium points x̂1 and x̂2 of (5.96) by a = 1, a1 = −1.5,
a2 = −0.5, b1 = −1.2, b2 = −1.8, τ1 = 0.02, τ2 = 0.03, σ = 2.1

of all other parameters by conditions (5.85), (5.87) for the equilibrium points x̂1
(magenta) and x̂2 (green) are shown in Fig. 5.7. As it was shown also in Fig. 5.6,
the stability regions obtained for positive delays are placed inside the regions with
zero delays (bounds 1 and 2).
Case 2. Put a = 1, a1 = −1.5, a2 = −0.5, b1 = −1.2, b2 = −1.8, τ1 = 0.3,
τ2 = 0.4, σ = 1.1. Thus, a > 0, B = −1.2 − 1.8 = −3 < 0, p = 0.605 < 2a, A =
−1.5 − 0.5 = −2 < 0.
In Fig. 5.8 the regions of stability in probability for the equilibrium points x̂1
and x̂2 are shown in the space of parameters (μ, λ): in the white region there are
no equilibrium points; in the yellow region there are possible unstable equilibrium
points; the red, cyan, magenta, and grey regions are the regions for stability in prob-
ability of the equilibrium point x̂ = x̂1 given by condition (5.75) (red and cyan) and
condition (5.79) (cyan, magenta, and grey); blue, brown, green, and grey regions
are the regions for stability in probability of the equilibrium point x̂ = x̂2 given by
condition (5.75) (blue, brown) and condition (5.80) (blue, brown, green, and grey);
in the grey region both equilibrium points x̂ = x̂1 and x̂ = x̂2 are stable in proba-
bility. Curves 1 and 2 are the bounds of the equilibrium points x̂1 and x̂2 stability
regions, respectively, given by conditions (5.88) and (5.90) for the case τ1 = τ2 = 0.
144 5 Stability of Systems with Nonlinearities

Fig. 5.10 Stability regions for the equilibrium points x̂1 and x̂2 of (5.96) by a = −1.2, a1 = 1.5,
a2 = 0.5, b1 = 1.2, b2 = 1.8, τ1 = 0.04, τ2 = 0.03, σ = 2

One can see that the stability regions obtained for positive delays are placed inside
the regions with zero delays.
Put now τ1 = 0.02, τ2 = 0.03, σ = 2.1. Then p = 2.205 > 2a = 2, and condition
(5.75) does not hold. Appropriate stability regions obtained with the same values
of all other parameters by conditions (5.89), (5.91) for the equilibrium points x̂1
(magenta) and x̂2 (green) are shown in Fig. 5.9. As it was shown also in Fig. 5.8,
the stability regions obtained for positive delays are placed inside the regions with
zero delays (bounds 1 and 2).
Case 3. Put a = −1.2, a1 = 1.5, a2 = 0.5, b1 = 1.2, b2 = 1.8, τ1 = 0.04, τ2 =
0.03, σ = 2. Thus, a < 0, B = 1.2 + 1.8 = 3 > 0, A = 1.5 + 0.5 = 2 > 0, and
condition (5.75) does not hold. In Fig. 5.10 the regions of stability in probability for
the equilibrium points x̂1 and x̂2 are shown in the space of parameters (μ, λ): in the
white region there are no equilibrium points; in the yellow region there are possible
unstable equilibrium points; magenta region is the region for stability in probability
of the equilibrium point x̂1 given by condition (5.79), green region is the region for
stability in probability of the equilibrium point x̂2 given by the condition (5.80).
Curves 1 and 2 are the bounds of the equilibrium points x̂1 and x̂2 stability regions,
respectively, given by conditions (5.92) and (5.93) for the case τ1 = τ2 = 0.
5.4 Systems with Fractional Nonlinearity 145

Fig. 5.11 Stability regions for the equilibrium points x̂1 and x̂2 of (5.96) by a = −1, a1 = −1.5,
a2 = −0.5, b1 = −1.2, b2 = −1.8, τ1 = 0.04, τ2 = 0.05, σ = 1.7

One can see that the stability regions obtained for positive delays are placed
inside the regions with zero delays.
Case 4. Put a = −1, a1 = −1.5, a2 = −0.5, b1 = −1.2, b2 = −1.8, τ1 = 0.04,
τ2 = 0.05, σ = 1.7. Thus, a < 0, B = −1.2 − 1.8 = −3 < 0, A = −1.5 − 0.5 =
−2 < 0. Appropriate regions of stability in probability for the equilibrium points
x̂1 and x̂2 obtained by conditions (5.79), (5.80), (5.94), and (5.95) are shown in
Fig. 5.11.

Example 5.5 Consider the differential equation

μ + a0 x(t) + a1 x(t − τ1 )  
ẋ(t) = −ax(t) + + σ x(t) − x̂ ẇ(t), (5.97)
λ + b0 x(t) + b1 x(t − τ1 )

which is a particular case of (5.71) with k = 1. The linear part of type (5.74) for this
equation has the form

ż(t) = γ̂0 z(t) + γ1 z(t − τ1 ) + σ z(t)ẇ(t), (5.98)

146 5 Stability of Systems with Nonlinearities

Fig. 5.12 Stability regions for the equilibrium point x̂1 = 2.696 of (5.97) by a = 1, a0 = −0.4,
b0 = 0.2, a1 = 1.5, b1 = 1.5, τ1 = 0.4, σ = 1.3

where

aj − abj x̂
γ̂0 = γ0 − a, γj = , j = 0, 1, B = b0 + b1 .
λ + B x̂

By Lemma 2.1 a necessary and sufficient condition for asymptotic mean-square

stability of the trivial solution of (5.98) is

γ̂0 + γ1 < 0, G−1 > p, (5.99)

where
⎧ 
⎪ γ1 q −1 sin(qτ )−1
⎪
⎪ γ + |γ̂ | < 0, q = γ12 − γ̂02 ,
⎪
⎪ γ̂0 +γ1 cos(qτ ), 1 0
⎨
γ̂0 |τ
G = 1+| 2|γ̂0 |
, γ1 = γ̂0 < 0,
⎪
⎪
⎪
⎪ 
⎪
⎩ γ1 q −1 sinh(qτ )−1 , γ̂0 + |γ1 | < 0, q = γ̂ 2 − γ 2 .
γ̂ +γ cosh(qτ )
0 1 0 1
5.4 Systems with Fractional Nonlinearity 147

Fig. 5.13 Stability regions for the equilibrium point x̂2 = −0.873 of (5.97) by a = 1, a0 = −0.4,
b0 = 0.2, a1 = 1.5, b1 = 1.5, τ1 = 0.4, σ = 1.3

Note that if, in (5.98), τ1 = 0, then the sufficient condition (5.76) for asymptotic
mean-square stability of the trivial solution of (5.98) takes the form a > γ0 + γ1 + p
and coincides with the necessary and sufficient condition (5.99) for asymptotic
mean-square stability of the trivial solution of (5.98). Let us show that for small
enough delay, the sufficient stability conditions (5.75) and (5.76) together are suffi-
ciently close to the necessary and sufficient stability condition (5.99).
In Fig. 5.12 stability regions for the equilibrium point x̂1 given by condition
(5.75) (green and magenta), by condition (5.79) (magenta and cyan), and by condi-
tion (5.99) (grey, green, magenta, and cyan) are shown for the following values of
the parameters:

a = 1, a0 = −0.4, b0 = 0.2, a1 = b1 = 1.5, τ1 = 0.4,

σ = 1.3.
(5.100)
One can see that both stability conditions (5.75) and (5.79) complement each other
and both these conditions together give the region of stability (green, magenta, and
cyan) that is sufficiently close to the exact stability region obtained by the necessary
and sufficient stability condition (5.99).
In Fig. 5.13 the similar picture for the same values of parameters (5.100) is shown
for the equilibrium point x̂2 : both stability conditions (5.75) (magenta and green
148 5 Stability of Systems with Nonlinearities

Fig. 5.14 200 trajectories of solutions of (5.98) at the point A(4, −2)

(small region placed between magenta and bound 2)) and (5.79) (magenta and cyan)
complement each other, and both these conditions together give the region of stabil-
ity (green, magenta, and cyan) that is sufficiently close to the exact stability region
(green, magenta, cyan, and grey) obtained by the necessary and sufficient stability
condition (5.99).
Consider the point A with μ = 4 and λ = −2. This point belongs to stabil-
ity regions for both equilibrium points x̂1 = 2.696 (Fig. 5.12) and x̂2 = −0.873
(Fig. 5.13). At the point A(4, −2) the trivial solution of (5.98) is asymptotically
mean-square stable. Thus, at the point A, the trajectories of all solutions of (5.97)
with different given initial functions and the values of the parameters (5.100) con-
verge to zero as t → ∞. 200 such trajectories are shown in Fig. 5.14 by the initial
functions (s ≤ 0)
 
j 10
x(s) = x̂1 + cos s − 8.5, j = 0, 2, 4, . . . , 198,
33 7
 
25 j 10
x(s) = x̂1 − cos s + 3, j = 1, 3, 5, . . . , 199.
28 33 7
5.4 Systems with Fractional Nonlinearity 149

Fig. 5.15 100 trajectories of solutions of (5.97) at the point A(4, −2) for the stable equilibrium
point x̂1 = 2.696

In Fig. 5.15 trajectories of solutions of the nonlinear equation (5.97) are shown
at the point A for the values of the parameters (5.100). At the point A the equilib-
rium point x̂1 = 2.696 of (5.97) is stable in probability. Thus, at the point A, 50
trajectories of solutions of (5.97) with the initial functions
 
2j 10
x(s) = x̂1 − cos s + 1.5, s ≤ 0, j = 1, 2, . . . , 50,
33 7

which belong to some neighborhood of the equilibrium point x̂1 , converge to x̂1 as
t → ∞ (magenta trajectories), but other 50 trajectories of solution with one initial
function
 
6 10
x(s) = x̂1 + cos s − 2.2, s ≤ 0,
11 7

which is placed out of the neighborhood of x̂1 , fill the whole space (green trajecto-
ries). Only some of these trajectories, which come to a neighborhood of x̂1 , converge
to x̂1 as t → ∞.
150 5 Stability of Systems with Nonlinearities

Fig. 5.16 100 trajectories of solutions of (5.97) at the point A(4, −2) for the stable equilibrium
point x̂2 = −0.873

Figure 5.16 is similar to Fig. 5.15, but it shows 100 trajectories for the equilib-
rium point x̂2 = −0.873: 50 trajectories (magenta) with the initial functions
 
j 5
x(s) = x̂2 − cos s + 2.1, s ≤ 0, j = 1, 2, . . . , 50,
15 3

which belong to a small enough neighborhood of the equilibrium point x̂2 , converge
to this equilibrium, and 50 trajectories (green) with one initial function
4
x(s) = x̂2 + cos(2s) + 1.8, s ≤ 0,
11
which is placed out of this neighborhood of x̂2 , fill the whole space.
Consider now the point B with μ = −2 and λ = −4 (Fig. 5.12). This point does
not belong to stability region for the equilibrium point x̂1 = 2.536, and thus, at the
point B(−2, −4) the equilibrium point x̂1 is unstable. In Fig. 5.17 five hundred
trajectories of the solution of (5.97) are shown with the initial function
 
10
x(s) = x̂1 + 0.015 sin s , s ≤ 0,
3
5.4 Systems with Fractional Nonlinearity 151

Fig. 5.17 500 trajectories of solutions of (5.97) at the point B(−2, −4) for the unstable equilib-
rium point x̂1 = 2.536

which is placed close enough to the equilibrium point x̂1 . One can see that the
trajectories do not converge to x̂1 and fill the whole space.
In Fig. 5.18 a similar picture is shown for the unstable equilibrium point x̂2 = −2
at the point C with μ = 4 and λ = 2.5 (Fig. 5.13) with the initial function
 
5
x(s) = x̂2 − 0.025 cos s , s ≤ 0.
3
Note that simulations of the solutions of (5.97) and (5.98) were obtained via its
difference analogues respectively in the forms
μ + a0 xi + a1 xi−m
xi+1 = (1 − a)xi +  + σ (xi − x̂)(wi+1 − wi )
λ + b0 xi + b1 xi−m
and
zi+1 = (1 + γ̂0 )zi + γ1 zi−m + σ zi (wi+1 − wi ).
Here  is the step of discretization (that was chosen as  = 0.01), xi = x(i),
zi = z(i), wi = w(i), m = τ1 /, and the trajectories of the Wiener process are
simulated by the algorithm described in Sect. 2.1.1.
152 5 Stability of Systems with Nonlinearities

Fig. 5.18 500 trajectories of solutions of (5.97) at the point C(4, 2.5) for the unstable equilibrium
point x̂2 = −2
Chapter 6
Matrix Riccati Equations in Stability of Linear
Stochastic Differential Equations with Delays

Here asymptotic mean-square stability conditions are obtained in terms of the exis-
tence of positive definite solutions of some matrix Riccati equations. Using the pro-
cedure of constructing Lyapunov functionals, we will obtain different matrix Riccati
equations for one stochastic differential equation with delays. If a considered dif-
ferential equation does not contain delays, then all these matrix Riccati equations
coincide with a unique linear matrix equation.
Consider first the stochastic linear differential equation without delay

dx(t) = Ax(t) dt + Cx(t) dw(t). (6.1)

Here A and C are constant n × n-matrices, x(t) ∈ Rn , and w(t) is the scalar standard
Wiener process.
Similarly to Theorem 1.3, the necessary and sufficient condition for asymptotic
mean-square stability of the zero solution of (6.1) can be formulated in terms of the
existence of a positive definite solution P of the linear matrix equation

A P + P A + C  P C = −Q (6.2)

for any positive definite matrix Q. But for differential equations with delays, the
situation is more complicated.

6.1 Equations with Constant Delays

6.1.1 One Delay in Deterministic Part and One Delay in Stochastic

Part of Equation

Below two different ways of Lyapunov functionals construction are considered.

L. Shaikhet, Lyapunov Functionals and Stability of Stochastic Functional 153

Differential Equations, DOI 10.1007/978-3-319-00101-2_6,
© Springer International Publishing Switzerland 2013
154 6 Matrix Riccati Equations in Stability of Linear Stochastic Differential

6.1.1.1 The First Way of Constructing a Lyapunov Functional

Consider the stochastic linear differential equation

 
dx(t) = Ax(t) + Bx(t − h) dt + Cx(t − τ ) dw(t), t ≥ 0,
(6.3)
x(s) = φ(s), s ≤ 0.

Formal application of the quadratic Lyapunov functional for (6.3) leads to a system
of matrix ordinary and partial differential equations [134]. The approach, used in
this chapter, gives Lyapunov functionals defined by the solutions of some nonlinear
(Riccati-type) matrix equations.
Let L be the generator of (6.3). Consider the functional V1 (x) = x  P x. Calculat-
ing LV1 , we get
   
LV1 = Ax(t) + Bx(t − h) P x(t) + x  (t)P Ax(t) + Bx(t − h)
+ x  (t − τ )C  P Cx(t − τ )
 
= x  (t) A P + P A x(t) + x  (t − τ )C  P Cx(t − τ )
+ x  (t − h)B  P x(t) + x  (t)P B(t − h).

Using Lemma 2.3 for a = x(t − h), b = B  P x(t), we have

 
LV1 ≤ x  (t) A P + P A + P BR −1 B  P x(t)
+ x  (t − h)Rx(t − h) + x  (t − τ )C  P Cx(t − τ ), (6.4)

which is the representation of type (2.35), where D(t) = A P + P A + P BR −1 B  P ,

k = 2, Q1 (t) = R, Q2 (t) = C  P C, τ1 (t) = h, τ2 (t) = τ , and all other parameters
are zeros.
Thus, by Theorem 2.5 we obtain the following theorem.

Theorem 6.1 Suppose that for some positive definite matrices Q and R, there exists
a positive definite solution P of the matrix Riccati equation

A P + P A + C  P C + R + P BR −1 B  P = −Q. (6.5)

Then the trivial solution of (6.3) is asymptotically mean-square stable.

Remark 6.1 Using Lemma 2.3 for a = x(t) and b = P Bx(t − h), instead of (6.4),
we obtain the inequality
 
LV1 ≤ x  (t) A P + P A + R x(t)
+ x  (t − h)B  P R −1 P Bx(t − h) + x  (t − τ )C  P Cx(t − τ ).

Thus, in Theorem 6.1, instead of (6.5), we can use the equation

A P + P A + C  P C + R + B  P R −1 P B = −Q. (6.6)
6.1 Equations with Constant Delays 155

Example 6.1 In the scalar case a positive solution of (6.5) (or (6.6)) exists if and
only if
1
A + |B| + C 2 < 0.
2
Example 6.2 Let us show that using in the matrix Riccati equations (6.5) and (6.6)
different positive definite matrices Q and R we can get positive definite solutions P
by different conditions on the parameters of the considered equation. Consider, for
instance, the two-dimensional equation (6.3) with
     
−a1 0 b1 b2 c 0
A= , B= , C= 1 .
0 −a2 −b2 b1 0 c2

Let I be the 2 × 2 identity matrix, Q = qI , R = rI , q, r > 0. It is easy to get that a

solution P of the matrix Riccati equation (6.5) with the elements p11 > 0, p22 > 0,
p12 = 0 is defined by the equation
 
1 2  1
b1 + b22 pii2 − 2 ai − ci2 pii + r + q = 0, i = 1, 2.
r 2

This equation has a positive root pii for arbitrary positive q and r if and only if

1
ai > b12 + b22 + ci2 , i = 1, 2.
2
Let us obtain from the matrix Riccati equation (6.6) a positive definite solution
P if the obtained condition does not hold. Put, for instance, Q = I , R = P ,

3 1 1 1
b1 = 0, a1 = + c12 > b12 + b22 + c12 = 1 + c12 ,
2 2 2 2

7 1 2 1 1
b2 = 1, a2 = + c2 < b12 + b22 + c22 = 1 + c22 .
8 2 2 2
It is easy to check that in this case the matrix Riccati equation (6.6) has a positive
definite solution P with the elements p11 = 3.5, p22 = 6, p12 = 0.

6.1.1.2 The Second Way of Constructing a Lyapunov Functional

Consider now another representation of the initial equation leading to other matrix
Riccati equations.
Let B be the operator norm of a matrix B. Rewrite now (6.3) as follows:

ż(t) = (A + B)x(t) + Cx(t − τ )ẇ(t),

 t (6.7)
z(t) = x(t) + B x(s) ds,
t−h
156 6 Matrix Riccati Equations in Stability of Linear Stochastic Differential

and following the procedure of constructing Lyapunov functionals, consider the

functional V1 = z (t)P z(t). Calculating LV1 , by (6.7) we get

LV1 = z (t)P (A + B)x(t) + x  (t)(A + B) P z(t) + x  (t − τ )C  P Cx(t − τ )


= x  (t) P (A + B) + (A + B) P x(t) + x  (t − τ )C  P Cx(t − τ )
 t
 
+ x (t)(A + B) P Bx(s) + x  (s)B  P (A + B)x(t) ds.
t−h

Using Lemma 2.3 for a = x(s) and b = B  P (A + B)x(t), we have

 t 
x  (t)(A + B) P Bx(s) + x  (s)B  P (A + B)x(t) ds
t−h
 t
≤ hx  t (A + B) P BR −1 B  P (A + B)x(t) + x  (s)Rx(s) ds. (6.8)
t−h

Thus,

LV1 ≤ x  (t) (A + B) P + P (A + B) + h(A + B) P BR −1 B  P (A + B) x(t)
 t
+ x  (s)Rx(s) ds + x  (t − τ )C  P Cx(t − τ ).
t−h

So, we get the representation of type (2.35), where

D(t) = (A + B) P + P (A + B) + h(A + B) P BR −1 B  P (A + B),

k = 1, Q1 (t) = C  P C, τ1 (t) = τ,
m = 0, dμ0 (s) = δ(s − h) ds, R0 = R,

and all other parameters are zeros. By Theorem 2.5 we obtain the following:

Theorem 6.2 Suppose that the inequality hB < 1 holds and that, for some sym-
metric matrices Q > 0 and R > 0, there exists a positive definite solution P of the
matrix Riccati equation

(A + B) P + P (A + B) + C  P C

+ h R + (A + B) P BR −1 B  P (A + B) = −Q. (6.9)

Then the trivial solution of (6.3) is asymptotically mean-square stable.

Remark 6.2 Using Lemma 2.3 for a = (A + B)x(t) and b = P Bx(s), instead
of (6.8), we obtain the inequality
6.1 Equations with Constant Delays 157
 t  
x (t)(A + B) P Bx(s) + x  (s)B  P (A + B)x(t) ds
t−h
 t
≤ hx  (t)(A + B) R(A + B)x(t) + x  (s)B  P R −1 P Bx(s) ds.
t−h

So, in Theorem 6.2, in place of (6.9), we can use the equation

(A + B) P + P (A + B) + C  P C

+ h (A + B) R(A + B) + B  P R −1 P B = −Q. (6.10)

Example 6.3 In the scalar case a positive solution of (6.9) (or (6.10)) for arbitrary
positive Q and R exists if and only if
  1
(A + B) 1 − h|B| + C 2 < 0, h|B| < 1.
2

6.1.2 Several Delays in Deterministic Part of Equation

Consider now the more general linear stochastic differential equation

m
ẋ(t) = Ax(t) + Bi x(t − hi ) + Cx(t − τ )ẇ(t). (6.11)
i=1

To construct a Lyapunov functional for (6.11), let us use both previous representa-
tions of the initial equation. Namely, represent (6.11) in the form

m
ż(t) = (A + B)x(t) + Bi x(t − hi ) + Cx(t − τ )ẇ(t),
i=m0 +1

m0  t

m0
z(t) = x(t) + Bj x(s) ds, B= Bj , 0 ≤ m0 ≤ m.
j =1 t−hj j =1

Following the procedure of constructing Lyapunov functionals, put V1 = z (t)P z(t).

In this case,
 

m
LV1 = (A + B)x(t) + Bi x(t − hi ) P z(t)
i=m0 +1
 

m
+ z (t)P (A + B)x(t) + Bi x(t − hi ) + x  (t − τ )C  P Cx(t − τ )
i=m0 +1

= x (t) (A + B) P + P (A + B) x(t) + x  (t − τ )C  P Cx(t − τ )

158 6 Matrix Riccati Equations in Stability of Linear Stochastic Differential

m

+ x  (t − hi )Bi P x(t) + x  (t)P Bi x(t − hi )
i=m0 +1
m0 

t 
+ x  (t)(A + B) P Bj x(s) + x  (s)Bj P (A + B)x(t) ds
j =1 t−hj

m m0 

t 
+ x  (t − hj )Bi P Bj x(s) + x  (s)Bj P Bi x(t − hi ) ds.
i=m0 +1 j =1 t−hj

Using Lemma 2.3, we obtain that for some Ri > 0, i = m0 + 1, . . . , m,

x  (t − hi )Bi P x(t) + x  (t)P Bi x(t − hi )

≤ x  (t)P Bi Ri−1 Bi P x(t) + x  (t − hi )Ri x(t − hi ), (6.12)

for some Gj > 0, j = 1, . . . , m0 ,

x  (t)(A + B) P Bj x(s) + x  (s)Bj P (A + B)x(t)

≤ x  (t)(A + B) P Bj G−1  

j Bj P (A + B)x(t) + x (s)Gj x(s), (6.13)

and, for some Sij > 0, i = m0 + 1, . . . , m, j = 1, . . . , m0 ,

x  (t − hi )Bi P Bj x(s) + x  (s)Bj P Bi x(t − hi )

≤ x  (t − hi )Bi P Bj Sij−1 Bj P Bi x(t − hi ) + x  (s)Sij x(s). (6.14)

By inequalities (6.12)–(6.14) we have

m
LV1 ≤ x  (t) (A + B) P + P (A + B) + P Bi Ri−1 Bi P
i=m0 +1


m
+ hj (A + B) P Bj G−1   
j Bj P (A + B) x(t) + x (t − τ )C P Cx(t − τ )
j =1

m m0 

t

+ x (t − hi )Ri x(t − hi ) + x  (s)Gj x(s) ds
i=m0 +1 j =1 t−hj

m m0 

+ hj x  (t − hi )Bi P Bj Sij−1 Bj P Bi x(t − hi )

i=m0 +1 j =1
 t 

+ x (s)Sij x(s) ds .
t−hj

Using the representation (2.35), similarly to Theorem 6.2, we obtain the follow-
ing theorem.
6.1 Equations with Constant Delays 159
m 0
Theorem 6.3 Suppose that, for some m0 = 0, 1, . . . , m, the inequality j =1 hj Bj
< 1 holds and that, for some matrices Ri > 0, Gj > 0, Sij > 0, and Q > 0, there
exists a positive definite solution P of the matrix Riccati equation

m
 
(A + B) P + P (A + B) + C  P C + Ri + P Bi Ri−1 Bi P
i=m0 +1

m0

+ hj Gj + (A + B) P Bj G−1 
j Bj P (A + B)
j =1

m0
 
+ hj Sij + Bi P Bj Sij−1 Bj P Bi = −Q, (6.15)
i=m0 +1 j =1

where

m0
B= Bj , 0 ≤ m0 ≤ m.
j =1

Then the trivial solution of (6.11) is asymptotically mean-square stable.

Remark 6.3 Similarly to inequalities (6.12)–(6.14), we can use, for instance, the
inequalities

x  (t − hi )Bi P x(t) + x  (t)P Bi x(t − hi )

≤ x  (t − hi )Bi P Ri−1 P Bi x(t − hi ) + x  (t)Ri x(t), (6.16)
  
x (t)(A + B) P Bj x(s) + x (s)Bj P (A + B)x(t)

≤ x  (s)Bj P (A + B)G−1  
j (A + B) P Bj x(s) + x (t)Gj x(t), (6.17)
x  (t − hi )Bi P Bj x(s) + x  (s)Bj P Bi x(t − hi )

≤ x  (s)Bj P Bi Sij−1 Bi P Bj x(s) + x  (t − hi )Sij x(t − hi ). (6.18)

Using different combinations of inequalities (6.12)–(6.14) and (6.16)–(6.18) and

different representations of type (2.35) in place of (6.15), we can use one from
other eight different matrix Riccati equations. For example, using in place of all
inequalities (6.12)–(6.14) inequalities (6.16)–(6.18) we obtain the equation

m
 

(A + B) P + P (A + B) + C P C + 
Ri + Bi P Ri−1 P Bi
i=m0 +1

m0

+ hj Gj + Bj P (A + B)G−1 
j (A + B) P Bj
j =1
160 6 Matrix Riccati Equations in Stability of Linear Stochastic Differential

m0
 
+ hj Sij + Bj P Bi Sij−1 Bi P Bj = −Q.
i=m0 +1 j =1

Using by virtue of Lemma 2.3 other inequalities of type (6.12)–(6.14) or (6.16)–

(6.18), we will obtain other matrix Riccati equations for the initial equation (6.11).

Remark 6.4 It is easy to see that representing (6.11) in the form

m0
ż(t) = (A + B)x(t) + Bi x(t − hi ) + Cx(t − τ )ẇ(t),
i=1

m  t

m
z(t) = x(t) + Bj x(s)ds, B= Bj , 0 ≤ m0 ≤ m,
j =m0 +1 t−hj j =m0 +1
 m0
we obtain a new modification of Theorem 6.3 after replacement of all sums j =1
 m m0
by m j =m0 +1 and of all sums i=m0 +1 by i=1 . Using other combinations of
the summands for representations of (6.11), we will obtain other modifications of
Theorem 6.3.

Example 6.4 Consider the scalar case of (6.11) with m = 2. Using Theorem 6.3 and
Remark 6.4 for different values of m0 = 0, 1, 2, we obtain four different sufficient
conditions for asymptotic mean-square stability of the trivial solution of (6.11):
1
A + |B1 | + |B2 | + C 2 < 0,
2
    1
(A + B1 ) 1 − h1 |B1 | + |B2 | 1 + h1 |B1 | + C 2 < 0, h1 |B1 | < 1,
2
    1 2
(A + B2 ) 1 − h2 |B2 | + |B1 | 1 + h2 |B2 | + C < 0, h2 |B2 | < 1,
2
  1 2
(A + B1 + B2 ) 1 − h1 |B1 | − h2 |B2 | + C < 0, h1 |B1 | + h2 |B2 | < 1.
2

6.2 Distributed Delay

Consider now the stochastic linear differential equation with distributed delay

m  t  t
ẋ(t) = Ax(t) + Bi x(s) ds + C x(s) ds ẇ(t). (6.19)
i=1 t−hi t−τ

Rewrite (6.19) in the form

m  
  t t
ż(t) = A + (B, h) x(t) + Bi x(s) ds + C x(s) ds ẇ(t),
i=m0 +1 t−hi t−τ
6.2 Distributed Delay 161

m0  t
z(t) = x(t) + Bj (s − t + hj )x(s) ds,
j =1 t−hj

m0
(B, h) = Bj hj , 0 ≤ m0 ≤ m.
j =1

In this case, for the functional V1 = z (t)P z(t), we have

  
 

m t
LV1 = A + (B, h) x(t) + Bi x(s) ds P z(t)
i=m0 +1 t−hi
  
 

m t

+ z (t)P A + (B, h) x(t) + Bi x(s) ds
i=m0 +1 t−hi
 t  t
+ x  (s) ds C  P C x(θ ) dθ
t−τ t−τ
   
= x  (t) A + (B, h) P + P A + (B, h) x(t)
 t  t
+ x  (s) ds C  P C x(θ ) dθ
t−τ t−τ

m  t 
+ x  (s)Bi P x(t) + x  (t)P Bi x(s) ds
i=m0 +1 t−hi
m0 

t   
+ (s − t + hj ) x  (t) A + (B, h) P Bj x(s)
j =1 t−hj
 
+ x  (s)Bj P A + (B, h) x(t) ds

m m0 


t t 
+ (s − t + hj ) x  (θ )Bi P Bj x(s)
i=m0 +1 j =1 t−hi t−hj

+ x  (s)Bj P Bi x(θ ) ds dθ.

Note that using Lemma 2.3 for a = x(s), b = C  P Cx(θ ), and R > 0, we have
 t  t
x  (s) ds C  P C x(θ ) dθ
t−τ t−τ
 
1 t t 
= x  (θ )C  P Cx(s) + x  (s)C  P Cx(θ ) ds dθ
2 t−τ t−τ
 t  t
τ  τ
≤ x (s)Rx(s) ds + x  (θ )C  P CR −1 C  P Cx(θ ) dθ.
2 t−τ 2 t−τ
162 6 Matrix Riccati Equations in Stability of Linear Stochastic Differential

Using inequalities similar to (6.12)–(6.14), we obtain

   

m

LV1 ≤ x (t) A + (B, h) P + P A + (B, h) + hi P Bi Ri−1 Bi P
i=m0 +1

1

m0
    
+ h2j A + (B, h) P Bj G−1 
j Bj P A + (B, h) x(t)
2
j =1

τ t  
+ x  (s) R + C  P CR −1 C  P C x(s) ds
2 t−τ

m  t m0 

t
+ x  (s)Ri x(s) ds + (s − t + hj )x  (s)Gj x(s) ds
i=m0 +1 t−hi j =1 t−hj

m m0 


1 2 t
+ hj x  (s)Bi P Bj Sij−1 Bj P Bi x(s) ds
2 t−hi
i=m0 +1 j =1
 t 
+ (s − t + hj )x  (s)Sij x(s) ds .
t−hj

By Theorem 2.5 we obtain the following:

m 0 2
Theorem 6.4 Suppose that, for some m0 = 0, 1, . . . , m, the inequality j =1 hj Bj
< 2 holds and that, for some matrices R > 0, Ri > 0, Gj > 0, Sij > 0, and Q > 0,
there exists a positive definite solution P of the matrix Riccati equation
    τ2  
A + (B, h) P + P A + (B, h) + R + C  P CR −1 C  P C
2

m
 
+ hi Ri + P Bi Ri−1 Bi P
i=m0 +1

m0
    
+ h2j Gj + A + (B, h) P Bj G−1 
j Bj P A + (B, h)
2
j =1

m m0
 
+ hi h2j Sij + Bi P Bj Sij−1 Bj P Bi = −Q. (6.20)
2
i=m0 +1 j =1

Then the trivial solution of (6.19) is asymptotically mean-square stable.

Remark 6.5 Using the arguments from Remark 6.3, one can show that in place of
(6.20) in Theorem 6.4 other matrix Riccati equations can be used, for instance,
    τ2  
A + (B, h) P + P A + (B, h) + R + C  P CR −1 C  P C
2
6.2 Distributed Delay 163

m
 
+ hi Ri + Bi P Ri−1 P Bi
i=m0 +1

m0
     
+ h2j Gj + Bj P A + (B, h) G−1
j A + (B, h) P Bj
2
j =1

m m0
 
+ hi h2j Sij + Bj P Bi Sij−1 Bi P Bj = −Q.
2
i=m0 +1 j =1

Remark 6.6 Using another representation of (6.19) in the form

m0  t  t
 
ż(t) = A + (B, h)1 x(t) + Bi x(s) ds + C x(s) ds ẇ(t),
i=1 t−hi t−τ

m  t
z(t) = x(t) + Bj (s − t + hj )x(s) ds,
j =m0 +1 t−hj

m
(B, h)1 = Bj hj , 0 ≤ m0 ≤ m,
j =m0 +1
 m0
we obtain a new modification of Theorem 6.4 after replacement of all sums j =1
 m m0
by m j =m0 +1 and all sums i=m0 +1 by i=1 . Using other combinations of the
summands for representations of (6.19), we will obtain other modifications of The-
orem 6.4.

Example 6.5 Consider the scalar case of (6.19) for m = 2. Using Theorem 6.4 and
Remark 6.6 for different values of m0 = 0, 1, 2, we obtain four different sufficient
conditions for asymptotic mean-square stability of the trivial solution of (6.19):
τ2 2
A + h1 |B1 | + h2 |B2 | + C < 0,
2
   
1 1 τ2
(A + h1 B1 ) 1 − h21 |B1 | + h2 |B2 | 1 + h21 |B1 | + C 2 < 0,
2 2 2
h21 |B1 | < 2,
   
1 1 τ2
(A + h2 B2 ) 1 − h22 |B2 | + h1 |B1 | 1 + h22 |B2 | + C 2 < 0,
2 2 2
h22 |B2 | < 2,
 
1 2 1 2 τ2
(A + h1 B1 + h2 B2 ) 1 − h1 |B1 | − h2 |B2 | + C 2 < 0,
2 2 2
h21 |B1 | + h22 |B2 | < 2.
164 6 Matrix Riccati Equations in Stability of Linear Stochastic Differential

6.3 Combination of Discrete and Distributed Delays

Similar stability conditions can be obtained for systems with different combinations
of discrete and distributed delays. For example, consider the linear stochastic differ-
ential equation
 t
ẋ(t) = Ax(t) + B1 x(t − h1 ) + B2 x(s) ds
t−h2
  t 
+ C1 x(t − τ1 ) + C2 x(s) ds ẇ(t). (6.21)
t−τ2

Using four different representations of this equation, we will construct four different
Lyapunov functionals for it.
First, consider (6.21). Using the functional V1 = x  (t)P x(t), we have
  t 
LV1 = Ax(t) + B1 x(t − h1 ) + B2 x(s) ds P x(t)
t−h2
  t 
+ x  (t)P Ax(t) + B1 x(t − h1 ) + B2 x(s) ds
t−h2
  t    t 
+ C1 x(t − τ1 ) + C2 x(s) ds P C1 x(t − τ1 ) + C2 x(s) ds
t−τ2 t−τ2
 
= x(t) A P + P A x(t) + x  (t − τ1 )C1 P C1 x(t − τ1 )
+ x  (t)P B1 x(t − h1 ) + x  (t − h1 )B1 P x(t)
 t
  
+ x (t)P B2 x(s) + x  (s)B2 P x(t) ds
t−h2
 t  
+ x  (t − τ1 )C1 P C2 x(s) + x  (s)C2 P C1 x(t − τ1 ) ds
t−τ2
 
1 t t  
+ x  (θ )C2 P C2 x(s) + x  (s)C2 P C2 x(θ ) ds dθ.
2 t−τ2 t−τ2

Using Lemma 2.3 for some positive definite matrices Ri , Gi , i = 1, 2, we obtain

 
LV1 ≤ x(t) A P + P A + R1 + h2 R2 x(t) + x  (t − τ1 )C1 P C1 x(t − τ1 )
+ x  (t − h1 )B1 P R1−1 P B1 x(t − h1 ) + τ2 x  (t − τ1 )G1 x(t − τ1 )
 t  t
−1
+  
x (s)B2 P R2 P B2 x(s) ds + x  (s)C2 P C1 G−1 
1 C1 P C2 x(s) ds
t−h2 t−τ2

τ2 t  
+ x  (s) G2 + C2 P C2 G−1 
2 C2 P C2 x(s) ds.
2 t−τ2
6.3 Combination of Discrete and Distributed Delays 165

By the representation of type (2.35) with Si (t) = 0, dKl = 0,

D = A P + P A + R1 + h2 R2 , k = 2, m = 0,

Q1 = C1 P C1 + τ2 G1 , τ1 (t) = τ1 ,

Q2 = B1 P R1−1 P B1 + τ2 G1 , τ2 (t) = h1 ,

dμ(s) = δ(s − h2 ) ds, R(τ + h2 , t) = B2 P R2−1 P B2 ,

dμ0 (s) = δ(s − τ2 ) ds,

1  
R0 = C2 P C1 G−1   −1 
1 C1 P C2 + 2 τ2 G2 + C2 P C2 G2 C2 P C2 ,

and Theorem 2.5, we obtain the following theorem.

Theorem 6.5 Let for some matrices Ri > 0, Gi > 0, i = 1, 2, and Q > 0, there
exist a positive definite solution P of the matrix Riccati equation
 
A P + P A + C1 P C1 + R1 + B1 P R1−1 P B1 + h2 R2 + B2 P R2−1 P B2

  τ22  
+ τ2 G1 + C2 P C1 G−1   −1 
1 C1 P C2 + 2 G2 + C2 P C2 G2 C2 P C2 = −Q.
(6.22)

Then the trivial solution of (6.21) is asymptotically mean-square stable.

Remark 6.7 Similarly to Remark 6.3, one can show that in place of (6.22) in The-
orem 6.5 one from the eight different matrix Riccati equations can be used, for
example,
 
A P + P A + C1 P C1 + R1 + P B1 R1−1 B1 P + h2 R2 + P B2 R2−1 B2 P

  τ22  
+ τ2 G1 + C1 P C2 G−1   −1 
1 C2 P C1 + 2 G2 + C2 P C2 G2 C2 P C2 = −Q.

Rewrite now (6.21) in the form

 t
ż(t) = (A + B1 )x(t) + B2 x(s) ds
t−h2
  t 
+ C1 x(t − τ1 ) + C2 x(s) ds ẇ(t),
t−τ2
 t
z(t) = x(t) + B1 x(s) ds.
t−h1
166 6 Matrix Riccati Equations in Stability of Linear Stochastic Differential

Then, for the functional V1 = z (t)P z(t), we have

  t 
LV1 = (A + B1 )x(t) + B2 x(s) ds P z(t)
t−h2
  t 
+ z (t)P (A + B1 )x(t) + B2 x(s) ds
t−h2
  t    t 
+ C1 x(t − τ1 ) + C2 x(s) ds P C1 x(t − τ1 ) + C2 x(s) ds
t−τ2 t−τ2
 
= x(t) (A + B1 ) P + P (A + B1 ) x(t)
 t
  
+ x (t)(A + B1 ) P B1 x(s) + x  (s)B1 P (A + B1 )x(t) ds
t−h1
 t  
+ x  (t)P B2 x(s) + x  (s)B2 P x(t) ds
t−h2
 
t t  
+ x  (θ )B1 P B2 x(s) + x  (s)B2 P B1 x(θ ) ds dθ
t−h1 t−h2

+ x (t − τ1 )C1 P C1 x(t − τ1 )


 t
  
+ x (t − τ1 )C1 P C2 x(s) + x  (s)C2 P C1 x(t − τ1 ) ds
t−τ2
 
1 t t  
+ x  (θ )C2 P C2 x(s) + x  (s)C2 P C2 x(θ ) ds dθ.
2 t−τ2 t−τ2

Using Lemma 2.3 for some Ri > 0, i = 1, 2, 3, Gj > 0, j = 1, 2, we obtain


LV1 ≤ x(t) (A + B1 ) P + P (A + B1 ) + h1 (A + B1 ) R1 (A + B1 ) + h2 P R2 P x(t)
 t

+ x  (s) B1 P R1−1 P B1 + h2 B1 P R3 P B1 x(s) ds
t−h1
 t 
+ x  (s) h1 B2 R3−1 B2 + B2 R2−1 B2 x(s) ds
t−h2

+ x  (t − τ1 ) C1 P C1 + τ2 C1 P G1 P C1 x(t − τ1 )
 t  
τ2   
+ x  (s) C2 G−1 C 2 + C G C
2 2 + C 
P G−1
P C2 x(s) ds.
t−τ2
1 2 2 2 2

By the representation of type (2.35) and Theorem 2.5 the following theorem is
proved.

Theorem 6.6 Suppose that the inequality h1 B1  < 1 holds and that, for some
matrices Ri > 0, i = 1, 2, 3, Gi > 0, i = 1, 2, and Q > 0, there exists a positive
6.3 Combination of Discrete and Distributed Delays 167

definite solution P of the matrix Riccati equation

(A + B1 ) P + P (A + B1 ) + C1 P C1
 
+ h1 (A + B1 ) R1 (A + B1 ) + B1 P R1−1 P B1
   
+ h2 P R2 P + B2 R2−1 B2 + h1 h2 B1 P R3 P B1 + B2 R3−1 B2
  τ22   
+ τ2 C1 P G1 P C1 + C2 G−1
1 C2 + C2 G2 C2 + C2 P G−1
2 P C2 = −Q.
2
(6.23)

Then the trivial solution of (6.21) is asymptotically mean-square stable.

Remark 6.8 Similarly to Remark 6.3, one can show that in place of (6.23) in Theo-
rem 6.6 other different matrix Riccati equations can be used, for instance,

(A + B1 ) P + P (A + B1 ) + C1 P C1
 
+ h1 R1 + B1 P (A + B1 )R1−1 (A + B1 ) P B1
   
+ h2 R2 + B2 P R2−1 P B2 + h1 h2 R3 + B2 P B1 R3−1 B1 P B2
  τ22  
+ τ2 G1 + C2 P C1 G−1
1 C 
1 P C 2 + G2 + C2 P C2 G−1
2 C2 P C2 = −Q.
2

Remark 6.9 Rewriting (6.21) in the form

  t 
ż(t) = (A + h2 B2 )x(t) + B1 x(t − h1 ) + C1 x(t − τ1 ) + C2 x(s) ds ẇ(t),
t−τ2
 t
z(t) = x(t) + B2 (s − t + h2 )x(s) ds,
t−h2

one can obtain the following theorem.

Theorem 6.7 Suppose that the inequality h22 B2  < 2 holds and that, for some
matrices Ri > 0, i = 1, 2, 3, Gi > 0, i = 1, 2, and Q > 0, there exists a positive
definite solution P of the matrix Riccati equation

(A + h2 B2 ) P + P (A + h2 B2 ) + C1 P C1 + R1 + P B1 R1−1 B1 P

1 
+ h22 R2 + R3 + (A + h2 B2 ) P B2 R2−1 B2 P (A + h2 B2 )
2
 
+ B1 P B2 R3−1 B2 P B1 + τ2 G1 + C1 P C2 G−1 
1 C2 P C1

τ22  
+ G2 + C2 P C2 G−1
2 C2 P C2 = −Q (6.24)
2
168 6 Matrix Riccati Equations in Stability of Linear Stochastic Differential

or

(A + h2 B2 ) P + P (A + h2 B2 ) + C1 P C1 + R1 + B1 P R1−1 P B1

1 
+ h22 R2 + R3 + B2 P (A + h2 B2 )R2−1 (A + h2 B2 ) P B2 + B2 P B1 R3−1 B1 P B2
2
  τ22  
+ τ2 G1 + C2 P C1 G−1   −1 
1 C1 P C2 + 2 G2 + C2 P C2 G2 C2 P C2 = −Q.

Then the trivial solution of (6.21) is asymptotically mean-square stable.

Remark 6.10 Rewriting (6.21) in the form

  t 
ż(t) = A1 x(t) + C1 x(t − τ1 ) + C2 x(s) ds ẇ(t),
t−τ2
 t  t
z(t) = x(t) + B1 x(s) ds + (s − t + h2 )B2 x(s) ds,
t−h1 t−h2

A1 = A + B1 + h2 B2 ,

one can obtain the following theorem.

Theorem 6.8 Suppose that the inequality h1 B1  + 12 h22 B2  < 1 holds and that,
for some matrices Ri > 0, Gi > 0, i = 1, 2, and Q > 0, there exists a positive defi-
nite solution P of the matrix Riccati equation
 
A1 P + P A1 + C1 P C1 + h1 R1 + A1 P B1 R1−1 B1 P A1
1    
+ h22 R2 + A1 P B2 R2−1 B2 P A1 + τ2 G1 + C1 P C2 G−1
1 C2 P C1
2
τ22  
+ G2 + C2 P C2 G−1
2 C2 P C2 = −Q (6.25)
2
or
 
A1 P + P A1 + C1 P C1 + h1 R1 + B1 P A1 R1−1 A1 P B1
1  
+ h22 R2 + B2 P A1 R2−1 A1 P B2
2
  τ22  
+ τ2 G1 + C2 P C1 G−1
1 C 
1 P C 2 + G2 + C2 P C2 G−1
2 C2 P C2 = −Q.
2
Then the trivial solution of (6.21) is asymptotically mean-square stable.

Remark 6.11 Note that in some cases a matrix Riccati equation can be transformed
to a linear matrix equation. Suppose, for instance, that the matrix B in (6.3) has the
6.3 Combination of Discrete and Distributed Delays 169

inverse matrix. Then putting in (6.5) R = B  P B, we transform it to the linear matrix

equation

A P + P A + P + B  P B + C  P C = −Q.

Suppose that for some P > 0 the matrix B satisfies the condition P B = B  P > 0.
Then putting in (6.5) R = P B, for this P , we obtain the linear matrix equation

(A + B) P + P (A + B) + C  P C = −Q.

We obtain the same linear matrix equations by putting R = P or R = P B in (6.6).

If in (6.21) the matrices C1 , C2 have the inverse matrices, then putting Ri = P ,
Gi = Ci P Ci , i = 1, 2, in (6.22), we transform it to the linear matrix equation

A P + P A + (1 + h2 )P + B1 P B1 + h2 B2 P B2
 
+ (1 + τ2 ) C1 P C1 + τ2 C2 P C2 = −Q.

If in (6.21) the matrices B1 , B2 , C1 , C2 have the inverse matrices, then putting

R1 = B1 P B1 , R2 = B2 P B2 , G1 = G2 = C2 P C2 in (6.25), we transform it to the
linear matrix equation

 
1 1
A1 P + P A1 + h1 + h22 A1 P A1 + h1 B1 P B1 + h22 B2 P B2
2 2
 
+ (1 + τ2 ) C1 P C1 + τ2 C2 P C2
= −Q.

Example 6.6 Consider the scalar case of (6.21). Putting C = |C1 | + τ2 |C2 |, by The-
orems 6.5–6.8 we obtain four different sufficient conditions for asymptotic mean-
square stability of the trivial solution

1
A + |B1 | + h2 |B2 | + C 2 < 0,
2
    1
(A + B1 ) 1 − h1 |B1 | + h2 |B2 | 1 + h1 |B1 | + C 2 < 0, h1 |B1 | < 1,
2
   
1 2  2 1 1
(A + h2 B2 ) 1 − h2 |B2 | + B  1 + h22 |B2 | + C 2 < 0, h22 |B2 | < 2,
2 2 2
 
1 1 1
(A + B1 + h2 B2 ) 1 − h1 |B1 | − h22 |B2 | + C 2 < 0, h1 |B1 | + h22 |B2 | < 1.
2 2 2
170 6 Matrix Riccati Equations in Stability of Linear Stochastic Differential

6.4 Equations with Nonincreasing Delays

6.4.1 One Delay in Deterministic and One Delay in Stochastic
Parts of Equation

Consider the stochastic linear differential equation with delays depending on time
   
ẋ(t) = Ax(t) + Bx t − h(t) + Cx t − τ (t) ẇ(t),
x0 (s) = φ(s), s ≤ 0. (6.26)

Here it is supposed that the delays h(t) and τ (t) are nonnegative differentiable func-
tions satisfying the conditions

ḣ(t) ≤ 0, τ̇ (t) ≤ 0. (6.27)

6.4.1.1 The First Way of Constructing a Lyapunov Functional

Consider the Lyapunov functional V1 = x  P x. Calculating LV1 , we get

     
LV1 = Ax(t) + Bx t − h(t) P x(t) + x  (t)P Ax(t) + Bx t − h(t)
   
+ x  t − τ (t) C  P Cx t − τ (t)
     
= x  (t) A P + P A x(t) + x  t − τ (t) C  P Cx t − τ (t)
   
+ x  t − h(t) B  P x(t) + x  (t)P Bx t − h(t) .

Using Lemma 2.3 for R > 0, a = x(t − h(t)), and b = B  P x(t), we have
 
LV1 ≤ x  (t) A P + P A + P BR −1 B  P x(t)
       
+ x  t − h(t) Rx t − h(t) + x  t − τ (t) C  P Cx t − τ (t) . (6.28)

By the representation (2.35) with

D(t) = A P + P A + P BR −1 B  P , k = 2,
Q1 (t) = R, τ1 (t) = h(t), Q2 (t) = C  P C, τ2 (t) = τ (t),
and Theorem 2.5 we obtain the following:

Theorem 6.9 Suppose that conditions (6.27) hold and that, for some positive defi-
nite matrices Q and R, there exists a positive definite solution P of the matrix Ric-
cati equation (6.5). Then the trivial solution of (6.26) is asymptotically mean-square
stable.

Remark 6.12 Using Lemma 2.3 for a = P x(t) and b = Bx(t − h(t)), one can get
in Theorem 6.9 in place of (6.5) the matrix Riccati equation

A P + P A + C  P C + B  RB + P R −1 P = −Q. (6.29)
6.4 Equations with Nonincreasing Delays 171

6.4.1.2 The Second Way of Constructing a Lyapunov Functional

Consider another way of constructing a Lyapunov functional for (6.26) leading to

another Riccati matrix equation.
Let us rewrite now (6.26) in the form of a neutral type equation
   
ż(t) = (A + B)x(t) + ḣ(t)Bx t − h(t) + Cx t − τ (t) ẇ(t),
 t
z(t) = x(t) + B x(s) ds
t−h(t)

and suppose that

 
h(0)B < 1, ĥ = supḣ(t) < ∞. (6.30)
t≥0

Following the procedure of constructing Lyapunov functionals, we will construct

now a functional V in the form V = V1 + V2 , where V1 = z (t)P z(t). Calculating
LV1 , we get
  
LV1 = z (t) P (A + B)x(t) + ḣ(t)Bx t − h(t)
   
+ x  (t)(A + B) + ḣ(t)x  t − h(t) B  P z(t)
   
+ x  t − τ (t) C  P Cx t − τ (t)
     
= x  (t) P (A + B) + (A + B) P x(t) + x  t − τ (t) C  P Cx t − τ (t)
     
+ ḣ(t) x  (t)P Bx t − h(t) + x  t − h(t) B  P x(t)
 t
  
+ x (t)(A + B) P Bx(s) + x  (s)B  P (A + B)x(t) ds
t−h(t)
 t      
+ ḣ(t) x  (s)B  P Bx t − h(t) + x  t − h(t) B  P Bx(s) ds.
t−h(t)
(6.31)

Using Lemma 2.3 for R1 > 0, a = x(s), and b = B  P (A + B)x(t), for R2 > 0, a =
x(t), and b = P Bx(t − h(t)), and for R3 > 0, a = x(s), and b = B  P Bx(t − h(t)),
we obtain
 t
 
x (t)(A + B) P Bx(s) + x  (s)B  P (A + B)x(t) ds
t−h(t)
 t

≤ h(t)x (t)(A + B) 
P BR1−1 B  P (A + B)x(t) + x  (s)R1 x(s) ds,
t−h(t)
     
ḣ(t) x  (t)P Bx t − h(t) + x  t − h(t) B  P x(t)
     
≤ ḣ(t) x  (t)R2 x(t) + x  t − h(t) B  P R −1 P Bx t − h(t) ,
2 (6.32)
172 6 Matrix Riccati Equations in Stability of Linear Stochastic Differential
 t       
ḣ(t) x (s)B  P Bx t − h(t) + x  t − h(t) B  P Bx(s) ds
t−h(t)

  t
≤ ḣ(t) x  (s)R3 x(s) ds
t−h(t)

   
+ h(t)x  t − h(t) B  P BR3−1 B  P Bx t − h(t) .

Then by (6.31)–(6.32) we have


LV1 ≤ x  (t) (A + B) P + P (A + B) + h(0)(A + B) P BR1−1 B  P (A + B)
     
+ ĥR2 x(t) + ĥx  t − h(t) B  P R2−1 + h(0)BR3−1 B  P Bx t − h(t)
 t
   
+ x  t − τ (t) C  P Cx t − τ (t) + x  (s)(R1 + ĥR3 )x(s) ds.
t−h(t)

Choosing the functional V2 in the form

 t 
   t
V2 = s − t + h(t) x (s)(R1 + ĥR3 )x(s) ds + x  (s)C  P Cx(s) ds
t−h(t) t−τ (t)
 t  
+ ĥ x  (s)B  P R2−1 + h(0)BR3−1 B  P Bx(s) ds,
t−h(t)

by (6.27) we obtain

  t
LV2 = h(t)x  (t)(R1 + ĥR3 )x(t) − 1 − ḣ(t) x  (s)(R1 + ĥR3 )x(s) ds
t−h(t)
     
+ x  (t)C  P Cx(t) − 1 − τ̇ (t) x  t − τ (t) C  P Cx t − τ (t)
 
+ ĥx  (t)B  P R2−1 + h(0)BR3−1 B  P Bx(t)
       
− ĥ 1 − ḣ(t) x  t − h(t) B  P R2−1 + h(0)BR3−1 B  P Bx t − h(t)
  
≤ x  (t) h(0)(R1 + ĥR3 ) + C  P C + ĥB  P R2−1 + h(0)BR3−1 B  P B+ x(t)
     
− ĥx  t − h(t) B  P R2−1 + h(0)BR3−1 B  P Bx t − h(t)
 t

    
− x t − τ (t) C P Cx t − τ (t) − x  (s)(R1 + ĥR3 )x(s) ds.
t−h(t)

As a result, for the functional V = V1 + V2 , we have


LV ≤ x  (t) (A + B) P + P (A + B) + C  P C
 
+ h(0) R1 + (A + B) P BR1−1 B  P (A + B)
   
+ ĥ R2 + B  P R2−1 P B + ĥh(0) R3 + B  P BR3−1 B  P B x(t).
So, we obtain the following theorem.
6.4 Equations with Nonincreasing Delays 173

Theorem 6.10 Suppose that conditions (6.27), (6.30) hold and that, for some ma-
trices Ri > 0, i = 1, 2, 3, and Q > 0, there exists a positive definite solution P of
the matrix Riccati equation

(A + B) P + P (A + B) + C  P C
 
+ h(0) R1 + (A + B) P BR1−1 B  P (A + B)
  
+ ĥ R2 + B  P R2−1 P B + h(0) R3 + B  P BR3−1 B  P B = −Q. (6.33)

Then the trivial solution of (6.26) is asymptotically mean-square stable.

Remark 6.13 Using Lemma 2.3 for R1 > 0, a = x(t), and b = (A + B) P Bx(s),
for R2 > 0, a = x(t − h(t)), and b = B  P x(t), and for R3 > 0, a = x(t − h(t)), and
b = B  P Bx(s), in place of (6.32) we obtain the inequalities
 t 
x  (t)(A + B) P Bx(s) + x  (s)B  P (A + B)x(t) ds
t−h(t)
 t

≤ h(t)x (t)R1 x(t) + x  (s)B  P (A + B)R1−1 (A + B) P Bx(s) ds,
t−h(t)
     
ḣ(t) x  (t)P Bx t − h(t) + x  t − h(t) B  P x(t)
      
≤ ḣ(t) x  t − h(t) R2 x t − h(t) + x  (t)P BR2−1 B  P x(t) , (6.34)
 t
      
ḣ(t) x (s)B  P Bx t − h(t) + x  t − h(t) B  P Bx(s) ds
t−h(t)

     
 
≤ ḣ(t) h(t)x  t − h(t) R3 x t − h(t)
 t 
+ x  (s)B  P BR3−1 B  P Bx(s) ds.
t−h(t)

Using different combinations of inequalities (6.32) and (6.34) and choosing an

appropriate functional V2 in place of (6.33) in Theorem 6.10, one can use differ-
ent matrix Riccati equations. For example, using all inequalities (6.34) in place of
inequalities (6.32) and choosing the functional V2 in the form
 t  
V2 = s − t + h(t) x  (s)B  P (A + B)R1−1 (A + B) P Bx(s) ds
t−h(t)
 t  t

+ ĥ x (s)R2 x(s) ds + x  (s)C  P Cx(s) ds
t−h(t) t−τ (t)
 t  
+ ĥ s − t + h(t) x  (s)B  P BR3−1 B  P Bx(s) ds,
t−h(t)
174 6 Matrix Riccati Equations in Stability of Linear Stochastic Differential

we obtain that in Theorem 6.10 in place of (6.33) the following equation can be
used:

(A + B) P + P (A + B) + C  P C
 
+ h(0) R1 + B  P (A + B)R1−1 (A + B) P B
  
+ ĥ R2 + P BR2−1 B  P + h(0) R3 + B  P BR3−1 B  P B = −Q. (6.35)

Example 6.7 In the scalar case a positive solution of (6.33) (or (6.35)) exists if and
only if

    1
(A + B) 1 − h(0)|B| + ĥ|B| 1 + h(0)|B| + C 2 < 0, h(0)|B| < 1.
2

6.4.2 Several Delays in Deterministic Part of Equation

Consider now the stochastic linear differential equation with delays

m
   
ẋ(t) = Ax(t) + Bi x t − hi (t) + Cx t − τ (t) ẇ(t), (6.36)
i=1

which is a generalization of (6.26). Here m is a positive integer, and the delays hi (t)
are nonnegative differentiable functions satisfying the conditions
 
τ̇ (t) ≤ 0, ḣi (t) ≤ 0, ĥi = supḣi (t) < ∞, i = 1, . . . , m. (6.37)
t≥0

To construct Lyapunov functionals for (6.36), we will use both previous repre-
sentations of the initial equation. Namely, rewrite (6.36) in the form

m0
 
ż(t) = (A + B)x(t) + ḣl (t)Bl x t − hl (t)
l=1

m
   
+ Bi x t − hi (t) + Cx t − τ (t) ẇ(t),
i=m0 +1

m0  t
z(t) = x(t) + Bj x(s) ds,
j =1 t−hj (t)

m0
B= Bj , 0 ≤ m0 ≤ m.
j =1
6.4 Equations with Nonincreasing Delays 175

We will construct a Lyapunov functional V in the form V = V1 + V2 , where V1 =

z (t)P z(t). In this case,


m
 
m0
 
LV1 = (A + B)x(t) + Bi x t − hi (t) + ḣl (t)Bl x t − hl (t) P z(t)
i=m0 +1 l=1


m
 
m0
 

+ z (t)P (A + B)x(t) + Bi x t − hi (t) + ḣl (t)Bl x t − hl (t)
i=m0 +1 l=1
   
+ x t − τ (t) C  P Cx t − τ (t)

    
= x  (t) (A + B) P + P (A + B) x(t) + x  t − τ (t) C  P Cx t − τ (t)

m
    
+ x t − hi (t) Bi P x(t) + x  (t)P Bi x t − hi (t)
i=m0 +1
m0 

t 
+ x  (t)(A + B) P Bj x(s) + x  (s)Bj P (A + B)x(t) ds
j =1 t−hj (t)

m0

m  t   
+ x t − hi (t) Bi P Bj x(s)
j =1 i=m0 +1 t−hj (t)
 
+ x  (s)Bj P Bi x t − hi (t) ds

m0
    
+ ḣl (t) x  t − hl (t) Bl P x(t) + x  (t)P Bl x t − hl (t)
l=1

m0

m0  t   
+ ḣl (t) x t − hl (t) Bl P Bj x(s)
l=1 j =1 t−hj (t)
 
+ x  (s)Bj P Bl x t − hl (t) ds.

Using Lemma 2.3 for Ri > 0, i = m0 + 1, . . . , m, a = x(t − hi (t)), and b =

Bi P x(t), for Gj > 0, j = 1, . . . , m0 , a = x(s), and b = Bj P (A + B)x(t), for Sij >
0, i = m0 + 1, . . . , m, j = 1, . . . , m0 , a = x(s), and b = Bj P Bi x(t − hi (t)), for
Ul > 0, l = 1, . . . , m0 , a = x(t − hl (t)), and b = Bl P x(t), and for Zlj > 0, l =
1, . . . , m0 , j = 1, . . . , m0 , a = x(s), and b = Bj P Bl x(t − hl (t)), we obtain
   
x  t − hi (t) Bi P x(t) + x  (t)P Bi x t − hi (t)
   
≤ x  t − hi (t) Ri x t − hi (t) + x  (t)P Bi Ri−1 Bi P x(t),
x  (t)(A + B) P Bj x(s) + x  (s)Bj P (A + B)x(t)

≤ x  (s)Gj x(s) + x  (t)(A + B) P Bj G−1 

j Bj P (A + B)x(t),
176 6 Matrix Riccati Equations in Stability of Linear Stochastic Differential
   
x  t − hi (t) Bi P Bj x(s) + x  (s)Bj P Bi x t − hi (t)
    (6.38)
≤ x  (s)Sij x(s) + x  t − hi (t) Bi P Bj Sij−1 Bj P Bi x t − hi (t) ,
    
ḣl (t) x  t − hl (t) Bl P x(t) + x  (t)P Bl x t − hl (t)
    
≤ ĥl x  t − hl (t) Ul x t − hl (t) + x  (t)P Bl Ul−1 Bl P x(t) ,
    
ḣl (t) x  t − hl (t) Bl P Bj x(s) + x  (s)Bj P Bl x t − hl (t)
    
≤ ĥl x  (s)Zlj x(s) + x  t − hl (t) Bl P Bj Zlj−1 Bj P Bl x t − hl (t) .

By inequalities (6.38) we obtain

m0
LV1 ≤ x (t) (A + B) P + P (A + B) +

ĥl P Bl Ul−1 Bl P
l=1


m0
+ P Bi Ri−1 Bi P + hj (0)(A + B) 
P Bj G−1 
j Bj P (A + B) x(t)
i=m0 +1 j =1

   

m
   
+ x  t − τ (t) C  P Cx t − τ (t) + x  t − hi (t) Ri x t − hi (t)
i=m0 +1
m0  t

m0

m  t

+ x (s)Gj x(s) ds + x  (s)Sij x(s) ds
j =1 t−hj (t) j =1 i=m0 +1 t−hj (t)

m0
   
+ hj (0)x  t − hi (t) Bi P Bj Sij−1 Bj P Bi x t − hi (t)
i=m0 +1 j =1
 

m0
   
m0 t
 
+ ĥl x t − hl (t) Ul x t − hl (t) + x (s)Zlj x(s) ds
l=1 j =1 t−hj (t)

m0

m0
   
+ ĥl hj (0)x  t − hl (t) Bl P Bj Zlj−1 Bj P Bl x t − hl (t) .
l=1 j =1

Choosing the functional V2 in the form

 t

m  t
 
V2 = x (s)C P Cx(s) ds + x  (s)Ri x(s) ds
t−τ (t) i=m0 +1 t−hi (t)
m0 

t  
+ s − t + hj (t) x  (s)Gj x(s) ds
j =1 t−hj (t)
6.4 Equations with Nonincreasing Delays 177

m 
m0
 t
+ hj (0) x  (s)Bi P Bj Sij−1 Bj P Bi x(s) ds
j =1 i=m0 +1 t−hi (t)
 
t  
+ s − t + hj (t) x  (s)Sij x(s) ds
t−hj (t)


m0 t
+ ĥl x  (s)Ul x(s) ds
l=1 t−hl (t)

m0 


t  
+ s − t + hj (t) x  (s)Zlj x(s) ds
j =1 t−hj (t)

m0

m0  t
+ ĥl hj (0) x  (s)Bl P Bj Zlj−1 Bj P Bl x(s) ds,
l=1 j =1 t−hj (t)

for the functional V = V1 + V2 , we obtain

m0
 
LV ≤ x (t) (A + B) P + P (A + B) + C  P C +

ĥl Ul + P Bl Ul−1 Bl P
l=1

m0

+ hj (0) Gj + (A + B) P Bj G−1 
j Bj P (A + B)
j =1

m
 
m0

m0
 
+ Ri + P Bi Ri−1 Bi P + hj (0) ĥl Zlj + Bl P Bj Zlj−1 Bj P Bl
i=m0 +1 j =1 l=1


m0

m
 
+ hj (0) Sij + Bi P Bj Sij−1 Bj P Bi x(t).
j =1 i=m0 +1

Thus, the following theorem is proved.

m0
Theorem 6.11 Suppose that conditions (6.37) and the inequality j =1 hj (0)Bj 
< 1 for some m0 = 1, . . . , m hold and that, for some positive definite matrices Ul ,
Ri , Gj , Sij , Zlj , and Q, there exists a positive definite solution P of the matrix
Riccati equation

m0
 
(A + B) P + P (A + B) + C  P C + ĥl Ul + P Bl Ul−1 Bl P
l=1

m0

+ hj (0) Gj + (A + B) P Bj G−1 
j Bj P (A + B)
j =1
178 6 Matrix Riccati Equations in Stability of Linear Stochastic Differential

m
 
m0

m0
 
+ Ri + P Bi Ri−1 Bi P + hj (0) ĥl Zlj + Bl P Bj Zlj−1 Bj P Bl
i=m0 +1 j =1 l=1

m0

m
 
+ hj (0) Sij + Bi P Bj Sij−1 Bj P Bi = −Q. (6.39)
j =1 i=m0 +1

Then the trivial solution of (6.36) is asymptotically mean-square stable.

Remark 6.14 Using other variants of inequalities (6.38) and choosing an appropriate
form of the functional V2 , in Theorem 6.11 one can use in place of (6.39) other
matrix Riccati equations, for instance,

m0
 
(A + B) P + P (A + B) + C  P C + ĥl Ul + Bl P Ul−1 P Bl
l=1

m0

+ hj (0) Gj + Bj P (A + B)G−1 
j (A + B) P Bj
j =1

m
 
m0

m0
 
+ Ri + Bi P Ri−1 P Bi + hj (0) ĥl Zlj + Bj P Bl Zlj−1 Bl P Bj
i=m0 +1 j =1 l=1

m0

m
 
+ hj (0) Sij + Bj P Bi Sij−1 Bi P Bj = −Q.
j =1 i=m0 +1

Remark 6.15 It is easy to see that rewriting equation (6.36) in the form

m
 
ż(t) = (A + B)x(t) + ḣl (t)Bl x t − hl (t)
j =m0 +1

m0
   
+ Bi x t − hi (t) + Cx t − τ (t) ẇ(t),
i=1

m  t
z(t) = x(t) + x(s) ds,
j =m0 +1 t−hj (t)

m
B= Bj , 0 ≤ m0 ≤ m,
j =m0 +1

we obtain
m0 a newmodification of Theorem
m6.11 after replacement in (6.39) of all
m0
sums j =1 by mj =m0 +1 and all sums i=m0 +1 by i=1 . Using other combina-
tions of the summands for representations of (6.36), we will obtain other modifica-
tions of Theorem 6.11.
6.5 Equations with Bounded Delays 179

Example 6.8 Consider the scalar case of (6.36) for m = 2. Using Theorem 6.11 for
different values of m0 = 0, 1, 2, we obtain four different sufficient conditions for
asymptotic mean-square stability of the trivial solution of (6.36):
1
A + |B1 | + |B2 | + C 2 < 0,
2
     1
(A + B1 ) 1 − h1 (0)|B1 | + ĥ1 |B1 | + |B2 | 1 + h1 (0)|B1 | + C 2 < 0,
2
h1 (0)|B1 | < 1,
     1
(A + B2 ) 1 − h2 (0)|B2 | + |B1 | + ĥ2 |B2 | 1 + h2 (0)|B2 | + C 2 < 0,
2
h2 (0)|B2 | < 1,
   
(A + B1 + B2 ) 1 − h1 (0)|B1 | − h2 (0)|B2 | + ĥ1 |B1 | + ĥ2 |B2 |
  1
× 1 + h1 (0)|B1 | + h2 (0)|B2 | + C 2 < 0, h1 (0)|B1 | + h2 (0)|B2 | < 1.
2

6.5 Equations with Bounded Delays

Consider the linear stochastic differential equation with distributed delays depend-
ing on time
 t  t
ẋ(t) = Ax(t) + Bx(s) ds + Cx(s) ds ẇ(t),
t−h(t) t−τ (t)
x0 (s) = φ(s), s ≤ 0. (6.40)

We suppose that the delays h(t) and τ (t) satisfy the following conditions:

0 ≤ h0 ≤ h(t) ≤ h1 , ĥ = h1 − h0 ≥ 0, 0 ≤ τ (t) ≤ τ1 . (6.41)

6.5.1 The First Way of Constructing a Lyapunov Functional

Calculating LV1 for V1 = x  (t)P x(t), we get

 
LV1 = x  (t) A P + P A x(t) + I0 + I1 ,

where
 t  t
I0 = x(s)C  P Cx(θ ) dθ ds,
t−τ (t) t−τ (t)
 t   
I1 = x (s)B  P x(t) + x  (t)P Bx(s) ds.
t−h(t)
180 6 Matrix Riccati Equations in Stability of Linear Stochastic Differential

Using (6.41) and Lemma 2.3 for I0 with R0 > 0, a = x(s), and b = C  P Cx(θ ) and
for I1 with R1 > 0, a = x(t), and b = P Bx(s), we obtain
  t
1 t   
I0 = x (θ )C  P Cx(s) + x  (s)C  P Cx(θ ) dθ ds
2 t−τ (t) t−τ (t)

τ1 t  
≤ x  (s) R0 + C  P CR0−1 C  P C x(s) ds, (6.42)
2 t−τ1
 t
I1 ≤ h1 x  (t)R1 x(t) + x  (s)B  P R1−1 P Bx(s) ds.
t−h1

Then
 
LV1 ≤ x  (t) A P + P A + h1 R1 x(t)
 t
+ x  (s)B  P R1−1 P Bx(s) ds
t−h1

τ1 t  
+ x  (s) R0 + C  P CR0−1 C  P C x(s) ds.
2 t−τ1

By the representation (2.35) and Theorem 2.5 we obtain the following theorem.

Theorem 6.12 Let for some positive definite matrices R0 , R1 , and Q, there exist a
positive definite solution P of the matrix Riccati equation
  τ2  
A P + P A + h1 R1 + B  P R1−1 P B + 1 R0 + C  P CR0−1 C  P C = −Q. (6.43)
2
Then the trivial solution of (6.40) is asymptotically mean-square stable.

Remark 6.16 Using Lemma 2.3 with other representations for a and b, it is possible
to get other matrix Riccati equations in Theorem 6.12. For example, using R0 > 0,
a = Cx(s), and b = P Cx(θ ), in place of (6.43) we obtain the equation
  τ2  
A P + P A + h1 P R1 P + B  R1−1 B + 1 C  R0 + P R0−1 P C = −Q. (6.44)
2
Example 6.9 In the scalar case both equations (6.43) and (6.44) have a positive
solution if and only if A + h1 |B| + 12 τ12 C 2 < 0.

6.5.2 The Second Way of Constructing a Lyapunov Functional

Rewrite (6.40) in the form of a stochastic differential equation of neutral type

 t−h(t)  t
ż(t) = (A + h1 B)x(t) − Bx(s) ds + Cx(s) ds ẇ(t),
t−h1 t−τ (t)
6.5 Equations with Bounded Delays 181
 t
z(t) = x(t) + (s − t + h1 )Bx(s) ds.
t−h1

Calculating LV1 , for the functional V1 = z (t)P z(t), we get


LV1 = x  (t) (A + h1 B) P + P (A + h1 B) x(t) + I0 + I1 + I2 + I3 ,

where I0 is defined in (6.42), and

 t 
I1 = (s − t + h1 ) x  (s)B  P (A + h1 B)x(t) + x  (t)(A + h1 B) P Bx(s) ds,
t−h1
 t−h(t) 
I2 = − x  (t)P Bx(s) + x  (s)B  P x(t) ds,
t−h1
 
t t−h(t) 
I3 = − (s − t + h1 ) x  (s)B  P Bx(θ ) + x  (θ )B  P Bx(s) dθ ds.
t−h1 t−h1

Using Lemma 2.3, we obtain

1
I1 ≤ h21 x  (t)R1 x(t)
2
 t
+ (s − t + h1 )x  (s)B  P (A + h1 B) P BR1−1 (A + h1 B) P Bx(s) ds,
t−h1
 t−h(t) 
I2 ≤ x  (t)R2 x(t) + x  (s)B  P R2−1 P Bx  (s) ds
t−h1
 t−h0
≤ ĥx  (t)R2 x(t) + x  (s)B  P R2−1 P Bx  (s) ds,
t−h1
 
t t−h(t) 
I3 ≤ (s − t + h1 ) x  (θ )R3 x(θ ) + x  (s)B  P BR3−1 B  P Bx(s) dθ ds
t−h1 t−h1
 t−h0
1
≤ h21 x  (θ )R3 x(θ ) dθ
2 t−h1
 t
+ ĥ (s − t + h1 )x  (s)B  P BR3−1 B  P Bx(s) ds.
t−h1

Then
 
  1 2
LV1 ≤ x (t) (A + h1 B) P + P (A + h1 B) + h1 R1 + ĥR2 x(t)
2
 t
+ (s − t + h1 )x  (s)B  P (A + h1 B)R1−1 (A + h1 B) P Bx(s) ds
t−h1
182 6 Matrix Riccati Equations in Stability of Linear Stochastic Differential
 t−h0  
1
+ x  (s) B  P R2−1 P B + h21 R3 x(s) ds
t−h1 2
 t
+ ĥ (s − t + h1 )x  (s)B  P BR3−1 B  P Bx(s) ds
t−h1

τ1 t  
+ x  (s) R0 + C  P CR0−1 C  P C x(s) ds.
2 t−τ1

By Theorem 2.5 we obtain the following theorem.

Theorem 6.13 Suppose that the inequality h21 B < 2 holds and that, for some
positive definite matrices Ri , i = 0, 1, 2, 3, and Q, there exists a positive definite
solution P of the matrix Riccati equation

τ12  
(A + h1 B) P + P (A + h1 B) + R0 + C  P CR0−1 C  P C
2
1 2 
+ h1 R1 + B  P (A + h1 B)R1−1 (A + h1 B) P B
2
  1  
+ ĥ R2 + B  P R2−1 P B + h21 ĥ R3 + B  P BR3−1 B  P B = −Q. (6.45)
2
Then the trivial solution of (6.40) is asymptotically mean-square stable.

Remark 6.17 By analogy with the previous remarks in Theorem 6.13, instead of
(6.45) other variants of matrix Riccati equations can be used.

Remark 6.18 Similar results can be obtained for more general equation
m 

t k 

t
ẋ(t) = Ax(t) + Bi x(s) ds + Ci x(s) ds ẇ(t).
i=1 t−hi (t) i=1 t−τi (t)

6.6 Equations with Unbounded Delays

Suppose now that the delays h(t) and τ (t) in (6.40) are nonnegative differentiable
functions that satisfy the conditions

ḣ(t) ≤ α < 1, τ̇ (t) ≤ β < 1. (6.46)

From (6.46) in particular it follows that

h(t) ≤ h(0) + αt, τ (t) ≤ τ (0) + βt,

i.e., the delays can increase to infinity.

6.6 Equations with Unbounded Delays 183

We will construct a Lyapunov functional V in the form V = V1 + V2 , where

V1 = x  P x. Calculating LV1 , we obtain
 
LV1 ≤ x  (t) A P + P A + P BR −1 B  P x(t)
       
+ x  t − h(t) Rx t − h(t) + x  t − τ (t) C  P Cx t − τ (t) .

By the representation (2.35) with

D(t) = A P + P A + P BR −1 B  P , k = 2,

Q1 = R, τ1 (t) = h(t), Q2 = C P C, τ2 (t) = τ (t),

from Theorem 2.5 we have following:

Theorem 6.14 Suppose that conditions (6.46) hold and that, for some positive def-
inite matrices R and QR, there exists a positive definite solution P of the matrix
Riccati equation

1 1
A P + P A + P BR −1 B  P + R+ C  P C = −Q. (6.47)
1−α 1−β

Then the trivial solution of (6.40) is asymptotically mean-square stable.

Remark 6.19 Instead of (6.47) in Theorem 6.14 other matrix Riccati equations can
be used, for example,

1 1
A P + P A + B  P R −1 P B + R+ C  P C = −Q.
1−α 1−β

Putting R = P in this equation, one can reduce it to the linear matrix equation

1 1
A P + P A + B  P B + P+ C  P C = −Q.
1−α 1−β

Example 6.10 In the scalar case a positive solution of (6.47) exists if and only if

|B| C2
A+ √ + < 0.
1−α 2(1 − β)

Consider the differential equation

 h(t)  τ (t)
ẋ(t) = Ax(t) + β(s)Bx(t − s) ds + σ (s)Cx(t − s) ds ẇ(t). (6.48)
0 0
184 6 Matrix Riccati Equations in Stability of Linear Stochastic Differential

Here we suppose that the delays h(t), τ (t) and the scalar functions β(s), σ (s) satisfy
the conditions
ĥ = sup h(t) ≤ ∞, τ̂ = sup τ (t) ≤ ∞,
t≥0 t≥0
 ĥ   (6.49)
 τ̂  
β= β(s) ds < ∞, σ= σ (s) ds < ∞.
0 0

Calculating LV1 for V1 = x  P x, P > 0, we obtain

  h(t) 
LV1 = Ax(t) + β(s)Bx(t − s) ds P x(t)
0
  h(t) 

+ x (t)P Ax(t) + β(s)Bx(t − s) ds
0
 τ (t)   τ (t) 
+ σ (s)Cx(t − s) ds P σ (s)Cx(t − s) ds
0 0

= x  (t) A P + P A x(t)
 h(t)

+ β(s) x  (t − s)B  P x(t) + x  (t)P Bx(t − s) ds
0
 τ (t)  τ (t)
+ σ (s)σ (θ)x  (t − s)C  P Cx(t − θ ) dθ ds.
0 0

Using Lemma 2.3, we have

 h(t)

β(s) x  (t − s)B  P x(t) + x  (t)P Bx(t − s) ds
0
 h(t)  
≤ β(s) x  (t)P R1 P x(t) + x  (t − s)B  R −1 Bx(t − s) ds
1
0
 ĥ  
≤ βx  (t)P R1 P x(t) + β(s)x  (t − s)B  R −1 Bx(t − s) ds
1
0

and analogously
 τ (t)  τ (t)
σ (θ)σ (s)x  (t − s)C  P Cx(t − θ ) ds dθ
0 0
 τ (t)  τ (t)
1 
= σ (θ)σ (s) x  (t − s)C  P Cx(t − θ )
2 0 0

+ x (t − θ )C  P Cx(t − s) ds dθ


 
1 τ (t) τ (t)   
≤ σ (θ)σ (s) x  (t − s)C  R2 Cx(t − s)
2 0 0
6.6 Equations with Unbounded Delays 185

+ x  (t − θ )C  P R2−1 P Cx(t − θ ) ds dθ
  τ (t)
1 τ (t)  

 
σ (s)x  (t − s)C  R2 Cx(t − s) ds
= σ (θ) dθ
2 0 0
 τ (t)  τ (t)
1    
+ σ (s) ds σ (θ)x  (t − θ )C  P R −1 P Cx(t − θ ) dθ
2 0 2
0

σ τ̂   
≤ σ (s)x  (t − s)C  R2 + P R2−1 P Cx(t − s) ds.
2 0
Therefore,

LV1 ≤ x  (t) A P + P A + βP R1 P x(t)
 ĥ
 
+ β(s)x  (t − s)B  R −1 Bx(t − s) ds,
1
0
 τ̂ 
σ  
+ σ (s)x  (t − s)C  R2 + P R −1 P Cx(t − s) ds.
2 2
0

By the representation (2.35) with

D = A P + P A + βP R1 P , l = 2,
 
S1 = B  R1−1 B, dν1 (s) = β(s) ds, s ∈ [0, ĥ],
 σ 
S2 = C  R2 + P R2−1 P C, dν1 (s) = σ (s) ds, s ∈ [0, τ̂ ],
2
and Theorem 2.5 we obtain the following:

Theorem 6.15 Suppose that conditions (6.49) hold and that, for some positive defi-
nite matrices R1 , R2 , and Q, there exists a positive definite solution P of the matrix
Riccati equation

σ 2  
A P + P A + βP R1 P + βB  R1−1 B + C R2 + P R2−1 P C = −Q. (6.50)
2
Then the trivial solution of (6.48) is asymptotically mean-square stable.

Remark 6.20 Similarly to the previous remarks in Theorem 6.15, other matrix Ric-
cati equations can be used.

Example 6.11 In the scalar case, (6.50) has a positive solution if and only if

σ2 2
A + β|B| + C < 0.
2
Chapter 7
Stochastic Systems with Markovian Switching

Investigation of systems with Markovian switching has a long history (see, for in-
stance, [119, 120, 197, 198, 266, 268, 278] and references therein). In this chap-
ter sufficient conditions for asymptotic mean-square stability of the solutions of
stochastic differential equations with delay and Markovian switching are obtained.
In particular, an application to Markov chain with two states and numerical simula-
tion of systems with Markovian switching are considered.

7.1 The Statement of the Problem

Consider the stochastic differential equation

   
ẋ(t) = f t, xt , η(t) + g t, xt , η(t) ẇ(t),
x0 = φ ∈ H. (7.1)

Here w(t) ∈ Rm is the standard Wiener process, η(t) is a scalar Markov chain with
finite set of states ai , i ∈ N = {1, 2, . . . , N}, and the probabilities of transition
 
pij (t) = P η(τ + t) = aj /η(τ ) = ai , t, τ ≥ 0.

We suppose that the Markov chain η(t) is independent on the Wiener process
w(t) and the probabilities of transition can be represented in the form

λij  + o(), j = i,
pij () =
1 + λii  + o(), j = i,

where

N
λij ≥ 0, j = i, λii ≤ 0, λij = 0, i ∈ N.
j =1

L. Shaikhet, Lyapunov Functionals and Stability of Stochastic Functional 187

Differential Equations, DOI 10.1007/978-3-319-00101-2_7,
© Springer International Publishing Switzerland 2013
188 7 Stochastic Systems with Markovian Switching

For arbitrary symmetric matrix F (s), s ≥ 0, the inequality dF (s) ≥ 0 means that
for each function ϕ ∈ H ,
 ∞
ϕ  (−s) dF (s) ϕ(−s) ≥ 0.
0

Here and below, all integrals of such a type are understood as Stieltjes integrals.
For arbitrary symmetric matrices F (s) and G(s), s ≥ 0, dF (s) ≥ dG(s) means that
dF (s) − dG(s) ≥ 0, and |A| denotes some matrix norm of a matrix A.
The set of negative definite matrices Ki is called uniformly negative definite with
respect to i ∈ N if there exists c > 0 such that for each x ∈ Rn ,
x  Ki x ≤ −c|x|2 , i ∈ N.
It is assumed that the functionals f (t, ϕ, ai ) ∈ Rn and g(t, ϕ, ai ) ∈ Rn×m are
defined for t ≥ 0, ϕ ∈ H , i ∈ N and satisfy the usual conditions [84, 87] of exis-
tence and uniqueness of the solution of (7.1). It is assumed also that the following
conditions hold:
(H1 ) For every positive definite matrix Pi , there exist symmetric matrices Qi and
Fi (s) such that
 ∞
ϕ  (0)Pi f (t, ϕ, ai ) ≤ ϕ  (0)Qi ϕ(0) + ϕ  (−s) dFi (s) ϕ(−s),
0
dFi (s) ≥ 0, i ∈ N. (7.2)
(H1 ) The functional f (t, ϕ, ai ) has the form
 ∞
f (t, ϕ, ai ) = Ai ϕ(0) + dBi (s) ϕ(−s), i ∈ N. (7.3)
0

(H2 ) For every positive definite matrix Pi , there exists a matrix Gi (s) such that
 ∞


Tr g (t, ϕ, ai )Pi g(t, ϕ, ai ) ≤ ϕ  (−s) dGi (s) ϕ(−s),

0
dGi (s) ≥ 0, i ∈ N. (7.4)
(H3 ) For the matrices Fi (s) and Gi (s) from (H1 ) and (H2 ), there exists a matrix
R(s) such that
dR(s) ≥ 2dFi (s) + dGi (s), i ∈ N,
  ∞
∞   (7.5)
R= dR(s), |R| < ∞, s dR(s) < ∞.
0 0

Remark 7.1 Note that if the functional f (t, ϕ, ai ) satisfies (H1 ), then it satisfies
(H1 ) too. For instance, for
  ∞
1   1 
Qi = 
Pi Ai + Ai Pi +  
Pi dBi (s) I , dFi (s) = Pi dBi (s)I. (7.6)
2 0 2
7.2 Stability Theorems 189

Let us represent an arbitrary functional V (t, ϕ, ai ) defined for t ≥ 0, ϕ ∈ H ,

i ∈ N, in the form V (t, ϕ, ai ) = V (t, ϕ(0), ϕ(s), ai ), s < 0, and put
 
Vϕ (t, x, ai ) = V (t, ϕ, ai ) = V t, x, x(t + s), ai , s < 0,
ϕ = xt , x = ϕ(0) = x(t).

Let D be the class of functionals V (t, ϕ, ai ) for which the function Vϕ (t, x, ai ) is
twice continuously differentiable with respect to x and continuously differentiable
with respect to t for almost all t ≥ 0. For the functionals from D, the generator L of
(7.1) is defined by the formula
∂ ∂
LV (t, ϕ, ai ) = Vϕ (t, x, ai ) + f  (t, ϕ, ai ) Vϕ (t, x, ai )
∂t ∂x
 
1  ∂2
+ Tr g (t, ϕ, ai ) 2 Vϕ (t, x, ai )g(t, ϕ, ai )
2 ∂x
  
+ λij V (t, ϕ, aj ) − V (t, ϕ, ai ) . (7.7)
j =i

Note that by the condition on λij the last summand in (7.7) can be written also in
the form

N
λij V (t, ϕ, aj ).
j =1

Similarly to (2.8) for 0 ≤ s ≤ t < ∞, we have


    t  
EV t, xt , η(t) = EV s, xs , η(s) + ELV θ, xθ , η(θ) dθ.
s

Similarly to Theorem 2.1, one can prove that asymptotic mean-square stabil-
ity conditions can be obtained by construction of some positive definite (or pos-
itive semidefinite for differential equations of neutral type) Lyapunov functionals
V (t, ϕ, ai ) for which the inequality
 2
LV (t, ϕ, ai ) ≤ −cϕ(0) , c > 0, (7.8)

holds. Below sufficient conditions for the asymptotic mean-square stability of the
trivial solution of (7.1) are obtained via the procedure of constructing Lyapunov
functionals.

7.2 Stability Theorems

Here two theorems about sufficient conditions for asymptotic mean-square stability
of the trivial solution of (7.1) are obtained.
190 7 Stochastic Systems with Markovian Switching

Theorem 7.1 Let hypotheses H1 , H2 , H3 hold and suppose that the matrices

Ki = 2Qi + λij (Pj − Pi ) + R (7.9)
j =i

are negative definite for each i ∈ N. Then the trivial solution of (7.1) is asymptoti-
cally mean-square stable.

Proof Let us construct a Markov chain P (t) with the set of states P1 , P2 , . . . , PN ,
where Pi , i ∈ N, are positive definite matrices. We will suppose that P (t) = Pi if
η(t) = ai , i ∈ N, and the probabilities of transition
 
pij (t) = P P (τ + t) = Pj /P (τ ) = Pi , t, τ ≥ 0,

are the same as for the Markov chain η(t).

We will construct a Lyapunov functional V for (7.1) in the form V = V1 + V2 ,
where V1 (t, ϕ, ai ) = ϕ  (0)Pi ϕ(0). Calculating LV1 , we obtain

LV1 (t, ϕ, ai ) = 2ϕ  (0)Pi f (t, ϕ, ai ) + Tr g  (t, ϕ, ai )Pi g(t, ϕ, ai )


+ λij ϕ  (0)(Pj − Pi )ϕ(0).
j =i

Using (7.2), (7.4), and (7.5), we have

 ∞
LV1 (t, xt , ai ) ≤ 2x  (t)Qi x(t) + 2 x  (t − s) dFi (s) x(t − s)
0
 ∞ 
+ x  (t − s) dGi (s) x(t − s) + λij x  (t)(Pj − Pi )x(t)
0 j =i
  

≤ x (t) 2Qi + λij (Pj − Pi ) x(t)
j =i
 ∞
+ x  (t − s) dR(s) x(t − s).
0

Consider the functional

 ∞ t
V2 (t, xt ) = x  (τ ) dR(s) x(τ ) dτ.
0 t−s

Since
 ∞

LV2 (t, xt ) = x  (t) dR(s) x(t) − x  (t − s) dR(s) x(t − s) ,

0
7.2 Stability Theorems 191

using (7.9) and the condition that the matrices Ki are uniformly negative definite
with respect to i ∈ N, for the functional V = V1 + V2 , we obtain
 2
LV (t, ϕ, ai ) ≤ ϕ  (0)Ki ϕ(0) ≤ −cϕ(0) .

Therefore, the trivial solution of (7.1) is asymptotically mean-square stable. The

proof is completed. 

Consider the equation

   
ẋ(t) = A η(t) x(t) + Bx(t − h) + g t, xt , η(t) ẇ(t). (7.10)

We suppose that if η(t) = ai , then A(η(t)) = Ai , g(t, ϕ, ai ) satisfies (7.4). Using

Remark 7.1, it is easy to get that in this case

f (t, ϕ, ai ) = Ai ϕ(0) + Bϕ(−h),

1  1
Qi = Pi Ai + Ai Pi + |Pi B|I , dFi (s) = |Pi B|I δ(s − h) ds.
2 2
Put
 ∞
P = sup Pi , dG(s) = sup dGi (s), G= dG(s)
i∈N i∈N 0

and suppose that

 ∞  
|G| < ∞, s dG(s) < ∞.
0
From Theorem 7.1 it follows that if the matrices
  
Ki = Pi Ai + Ai Pi + |Pi | + |P | |B|I + λij (Pj − Pi ) + G (7.11)
j =i

are negative definite for each i ∈ N, then the trivial solution of (7.10) is asymptoti-
cally mean-square stable.
Let us obtain another stability condition.

Theorem 7.2 Let |B|h < 1 and let Ri , i ∈ N, be uniformly negative definite matri-
ces, where

Ri = (Ai + B) Pi + Pi (Ai + B) + Λi + G + h(ρi + β)I, (7.12)

  
Λi = λij (Pj − Pi ), ρi = (Ai + B) Pi B + Λi B ,
j =i
    (7.13)
βi = hB  Λi B  + ρi , β = sup βi .
i∈N
Then the trivial solution of (7.10) is asymptotically mean-square stable.
192 7 Stochastic Systems with Markovian Switching

Proof Reduce (7.10) to the form of a stochastic differential neutral-type equation

  t
d      
x(t) + Bx(s) ds = A η(t) + B x(t) + g t, xt , η(t) ẇ(t).
dt t−h

Consider the functional

  t    t
V1 (t, xt , ai ) = x(t) + Bx(s) ds Pi x(t) + Bx(s) ds .
t−h t−h

Calculating LV1 , we obtain

  t
 
LV1 (t, xt , ai ) = 2x (t)(Ai + B) Pi x(t) + Bx(s) ds
t−h

+ Tr g  (t, xt , ai )Pi g(t, xt , ai )

  t    t
+ x(t) + Bx(s) ds Λi x(t) + Bx(s) ds
t−h t−h
 t
= 2x  (t)(Ai + B) Pi x(t) + 2x  (t)(Ai + B) Pi Bx(s) ds
t−h

+ Tr g  (t, xt , ai )Pi g(t, xt , ai ) + x  (t)Λi x(t)

 t 
+2 Bx(s) ds Λi x(t)
t−h
 t   t
+ Bx(s) ds Λi Bx(s) ds
t−h t−h

= x (t) (Ai + B) Pi + Pi (Ai + B) + Λi x(t)



+ Tr g  (t, xt , ai )Pi g(t, xt , ai )



t
+ 2x  (t) (Ai + B) Pi B + Λi B x(s) ds
t−h
 t  t
+ x  (s) ds B  Λi B x(τ ) dτ.
t−h t−h

By (7.4) and (7.13) we have

LV1 (t, xt , ai ) ≤ x  (t) (Ai + B) Pi + Pi (Ai + B) + Λi x(t)

 ∞  2
   t  
+  
x (t − s) dGi (s) x(t − s) + B Λi B  x(s) ds
0 t−h
 t     
+ ρi x(t)2 + x(s)2 ds
t−h
7.3 Application to Markov Chain with Two States 193
   

≤ x  (t) A(i) + B Pi + Pi A(i) + B + Λi + hρi I x(t)

 ∞  t
 
+ x  (t − s) dGi (s) x(t − s) + β x(s)2 ds.
0 t−h

Putting
 ∞ t  t  2
V2 (t, xt ) = x  (τ ) dG(s) x(τ ) dτ + β (s − t + h)x(s) ds,
0 t−s t−h

we obtain
 2
LV2 (t, xt ) = x  (t)Gx(t) + βhx(t)
 ∞  t  
− 
x (t − s) dG(s) x(t − s) − β x(s)2 ds.
0 t−h

Therefore, using (7.12) and the negative definiteness of the matrices Ri , i ∈ N, for
the functional V = V1 + V2 , we have
 2
LV (t, xt , ai ) ≤ x  (t)Ri x(t) ≤ −cx(t) . (7.14)

From this and from |B|h < 1 it follows that the trivial solution of (7.10) is asymp-
totically mean-square stable. The proof is completed. 

7.3 Application to Markov Chain with Two States

Consider the scalar stochastic differential equation

ẋ(t) = η(t)x(t) + bx(t − h) + σ x(t − τ )ẇ(t), (7.15)

where b > 0, h ≥ 0, and η(t) is a Markov chain with two states {a1 , a2 } such that

σ2
a2 < 0, a1 > |a2 | > b + . (7.16)
2
It is clear that if η(t) = a1 , then the trivial solution of (7.15) is unstable. But for
η(t) = a2 , from (7.16) the inequality

σ2
a2 + b + <0
2
follows, which is (see (3.11)) a sufficient condition for the asymptotic mean-square
stability of the trivial solution of (7.15).
194 7 Stochastic Systems with Markovian Switching

7.3.1 The First Stability Condition

Let us obtain a sufficient condition for asymptotic mean-square stability of the trivial
solution of (7.15), supposing that the Markov chain η(t) has two states {a1 , a2 } with
condition (7.16) and the transition rates λ12 and λ21 such that λ12 > λ21 .
From Theorem 7.1 and (7.11) it follows that the sufficient condition for the
asymptotic mean-square stability of the trivial solution of (7.15) has the form
2pi ai + (pi + p)b + pσ 2 + λij (pj − pi ) < 0, i, j = 1, 2, j = i, (7.17)
where p1 > 0, p2 > 0, p = max(p1 , p2 ). It is easy to see that if i = 1 and p1 ≤ p2 ,
then condition (7.17) is impossible. So, let p1 > p2 . Therefore, p = p1 and γ =
p2 p1−1 ∈ (0, 1). From (7.17) we have

2a1 + 2b + σ 2 + λ12 (γ − 1) < 0,

2a2 γ + (γ + 1)b + σ 2 + λ21 (1 − γ ) < 0.
From this and from (7.16) it follows that
λ21 + b + σ 2 2(a1 + b) + σ 2
<γ <1− . (7.18)
λ21 + 2|a2 | − b λ12
Thus, if the condition
λ21 + b + σ 2 2(a1 + b) + σ 2
<1− (7.19)
λ21 + 2|a2 | − b λ12
holds, then there exists a positive number γ ∈ (0, 1) such that condition (7.18) holds
too. From (7.19) and (7.16) it follows that the sufficient condition for the asymptotic
mean-square stability of the trivial solution of (7.15) has the form
λ12 (|a2 | − b − ε) σ2
|a2 | < a1 < − b − ε, ε= . (7.20)
λ21 + 2|a2 | − b 2
Note that condition (7.20) is possible if and only if λ12 > λ21 only. Indeed, from
(7.20) it follows that
  |a2 | + b + ε
λ12 > λ21 + 2|a2 | − b > λ21 + 2|a2 | − b > λ21 .
|a2 | − b − ε

7.3.2 The Second Stability Condition

Let us obtain another stability condition. From Theorem 7.2 for (7.15) we have
2(ai + b)pi + pσ 2 + λij (pj − pi ) + h(ρi + β) < 0,
i, j = 1, 2, j = i, (7.21)
7.3 Application to Markov Chain with Two States 195

where
 
ρi = b(ai + b)pi + λij (pj − pi ), β = max(β1 , β2 ),
(7.22)
βi = hb2 λij |pj − pi | + ρi , i, j = 1, 2, j = i.
Let us transform condition (7.21)–(7.22) to a more visual form. From (7.16) it fol-
lows that for i = 1 and p1 ≤ p2 , condition (7.21) is impossible. So, assume that
p1 > p2 . Then from (7.22) it follows that
 
ρ1 < b (a1 + b)p1 + λ12 (p1 − p2 ) ,
 
ρ2 < b |a2 + b|p2 + λ21 (p1 − p2 ) ,

(7.23)
β1 < β 1 = b (a1 + b)p1 + λ12 (1 + bh)(p1 − p2 ) ,

β2 < β 2 = b |a2 + b|p2 + λ21 (1 + bh)(p1 − p2 ) .

By (7.16) we have |a2 + b| = −a2 − b = |a2 | − b < a1 + b. Since, besides, p1 > p2

and λ12 > λ21 , we have β 1 > β 2 . Thus, condition (7.21) follows from

2(ai + b)pi + p1 σ 2 + λij (pj − pi ) + h(ρi + β 1 ) < 0. (7.24)

In addition, using (7.23), we obtain that (7.24) follows from

2(a1 + b)p1 + p1 σ 2 + λ12 (p2 − p1 )

+ hb 2(a1 + b)p1 + λ12 (2 + bh)(p1 − p2 ) < 0,

2(a2 + b)p2 + p1 σ 2 + λ21 (p1 − p2 )
   

+ hb |a2 | − b p2 + (a1 + b)p1 + λ12 (1 + bh) + λ21 (p1 − p2 ) < 0.

Using γ = p2 p1−1 ∈ (0, 1), we can rewrite these inequalities in the form

2(a1 + b) + σ 2 + λ12 (γ − 1)

+ hb 2(a1 + b) + λ12 (2 + bh)(1 − γ ) < 0,

(7.25)
2(a2 + b)γ + σ 2 + λ21 (1 − γ )
   

+ hb |a2 | − b γ + a1 + b + λ12 (1 + bh) + λ21 (1 − γ ) < 0.

Suppose now that

√
bh < 2 − 1. (7.26)
Then from (7.25) we obtain

A + (a1 + b)bh + σ 2 2(a1 + b)(1 + bh) + σ 2

<γ <1− , (7.27)
A+C B
where

A = (λ21 + λ12 bh)(1 + bh),

196 7 Stochastic Systems with Markovian Switching
 
B = λ12 1 − 2bh − b2 h2 , (7.28)
 
C = (2 − bh) |a2 | − b .

By (7.26) and (7.16) we have that B > 0 and C > 0.

Thus, if the condition
A + (a1 + b)bh + σ 2 2(a1 + b)(1 + bh) + σ 2
<1−
A+C B
or (that is the same in combination with (7.16))

BC − σ 2 (A + B + C)
|a2 | < a1 < −b (7.29)
Bbh + 2(1 + bh)(A + C)
holds, then there exists γ ∈ (0, 1) such that condition (7.27) holds too, and therefore
the conditions of Theorem 7.2 hold. This means that condition (7.29) or (7.28) is a
sufficient condition for the asymptotic mean-square stability of the trivial solution
of (7.15).
In Fig. 7.1 the stability regions for (7.15), given by condition (7.20), are shown
for λ12 = 28, λ21 = 0.05, b = 1, ε = 0.25 (the bound 1), and the stability regions,
given by condition (7.29), (7.28), are shown for the same values of the parameters
λ12 , λ21 , b, p and the different values of h: (2) h = 0, (3) h = 0.01, (4) h = 0.02,
(5) h = 0.03, (6) h = 0.04. It is easy to see that in spite of rough estimates that were
used for getting inequalities (7.23), the stability condition (7.29), (7.28) for small
enough h is better than (7.20).
In Fig. 7.2 the stability region for (7.15), given by condition (7.29), (7.28), is
shown for λ12 = 15, λ21 = 1, b = 0.2, h = 0.2, σ = 0. Putting λ12 = 5 and using
the same values of the other parameters, we obtain the stability region shown in
Fig. 7.3.
We can see that in the case λ12 = 15 (Fig. 7.2) the point A(a1 , a2 ) = A(1, −0.5)
belongs to the stability region, and therefore at this point the trivial solution of
(7.15) is asymptotically mean-square stable. On the other hand, in the case λ12 = 5
(Fig. 7.3) the point A(1, −0.5) does not belong to the stability region. Since condi-
tion (7.29), (7.28) is a sufficient condition only, for λ12 = 5, at the point A(1, −0.5)
the trivial solution of (7.15) can be either stable or unstable.

Remark 7.2 Let us show that the trivial solution of (7.15) can be asymptotically
mean-square stable not by conditions (7.16) only. Consider (7.15) by the conditions

a1 = −1, a2 = 1, b = −1,
(7.30)
σ = 0.1, λ12 = 1, λ21 = 5,

and obtain the maximum value of the delay h for which the conditions of Theo-
rem 7.2 hold.
By Theorem 7.2 we obtain the stability conditions in the form (7.21) again, i.e.,

2(ai + b)pi + λij (pj − pi ) + pσ 2 + h(ρi + β) < 0, i = j, (7.31)

7.3 Application to Markov Chain with Two States 197

Fig. 7.1 Stability region for

(7.15), given by condition
(7.20), is shown for λ12 = 28,
λ21 = 0.05, b = 1,
ε = 0.25(1), and stability
regions, given by condition
(7.29), (7.28), are shown for
the same values of the
parameters λ12 , λ21 , b, ε and
the different values of h:
(2) h = 0, (3) h = 0.01,
(4) h = 0.02, (5) h = 0.03,
(6) h = 0.04

Fig. 7.2 Stability region for

(7.15), given by the condition
(7.29), (7.28), is shown for
λ12 = 15, λ21 = 1, b = 0.2,
h = 0.2, σ = 0

Fig. 7.3 Stability region for

(7.15), given by the condition
(7.29), (7.28), is shown for
λ12 = 5, λ21 = 1, b = 0.2,
h = 0.2, σ = 0
198 7 Stochastic Systems with Markovian Switching

where i = 1, 2, p = max(p1 , p2 ). But put now p2 = γp1 with 1 < γ < 3. Then
from (7.13) and (7.30) we have

ρ1 = (3 − γ )p1 , ρ2 = 5(γ − 1)p1 ,

(7.32)
β1 = h(γ − 1) + 3 − γ p1 , β2 = 5(γ − 1)(h + 1)p1 .

If, in addition, γ ∈ (4/3, 3), then β2 > β1 for each h > 0, and therefore β =
max(β1 , β2 ) = β2 . As a result, the stability conditions (7.31) take the form

5(γ − 1)h(h + 1) + h(3 − γ ) + 1.01γ < 5,

(7.33)
5(γ − 1)h(h + 1) + 5h(γ − 1) + 0.01γ < 5(γ − 1).

From this it follows that



2γ − 1 2 5 − 1.01γ 2γ − 1
h< + − ,
5(γ − 1) 5(γ − 1) 5(γ − 1)

9.99γ − 10
h< − 1.
5(γ − 1)

It is easy to check that for γ = 1.900923, the both inequalities take the form h <
0.412721. This means that by the values of the parameters (7.30) the inequality
h < 0.412721 is a sufficient condition for the asymptotic mean-square stability of
the trivial solution of (7.15).
However, it is necessary to remember that conditions (7.33) are sufficient condi-
tions only. So, in reality, the trivial solution of (7.15) by the conditions (7.33) can be
asymptotically mean-square stable also for h ≥ 0.412721, as it will be shown below
(in Sect. 7.4.2) by a numerical simulation.

7.4 Numerical Simulation of Systems with Markovian Switching

Taking into account that it is difficult enough in each case to get analytical condi-
tions for stability, it is very important to have numerical methods for stability inves-
tigation. A numerical procedure for investigation of stability of stochastic systems
with Markovian switching is considered here. This procedure can be used in the
cases where analytical conditions of stability are absent. Some examples of using of
the proposed numerical procedure are considered. the results of the calculations are
presented by a lot of figures.
7.4 Numerical Simulation of Systems with Markovian Switching 199

7.4.1 System Without Stochastic Perturbations

Consider the differential equation (7.15) again but with σ = 0, i.e., without stochas-
tic perturbations of the type of white noise. In this case we have

ẋ(t) = η(t)x(t) + bx(t − h),

(7.34)
x(s) = φ(s), s ∈ [−h, 0].

Here, as before, η(t) is a Markov chain with two states {a1 , a2 }, initial distribution
 
pi = P η(0) = ai , i = 1, 2, (7.35)

and the probabilities of transition pij () of the form

 
pij () = P η(t + ) = aj /η(t) = ai = λij  + o(),
i, j = 1, 2, i = j. (7.36)

We suppose also that

a2 < 0, a1 > |a2 | > b > 0, λ12 > λ21 > 0. (7.37)

Let us investigate the stability of the trivial solution of (7.34) at the point
A(a1 , a2 ) with a1 = 1, a2 = −0.5 using a numerical simulation. Note that this point
belongs to the stability region (Fig. 7.2). Consider the difference analogue of (7.34)
in the form
xi+1 = (1 + ηi )xi + bxi−m ,
where

xi = x(ti ), ηi = η(ti ), ti = i, h = m,  > 0.

Simulation of the Markov chain ηi , i = 0, 1, . . . , can be reduced to the simpler

problem, simulation of a sequence of independent random variables ζi that are uni-
formly distributed on the interval [0,1]. Indeed, using (7.35), we have

p1 = P{η0 = a1 } = P{ζ0 < p1 },

(7.38)
p2 = P{η0 = a2 } = P{ζ0 > p1 }.

So, if after simulation of ζ0 , we obtain ζ0 < p1 , then we put η0 = a1 , else we put

η0 = a2 .
Further, if ηi−1 = a1 , i > 0, then from (7.36) for small enough  > 0, we have

P{ηi = a2 /ηi−1 = a1 } = P{ζi < λ12 } = λ12 ,

(7.39)
P{ηi = a1 /ηi−1 = a1 } = P{ζi > λ12 } = 1 − λ12 .
200 7 Stochastic Systems with Markovian Switching

Fig. 7.4 Four possible trajectories of the Markov chain η(t) and four appropriate trajectories of
the solution of (7.34) are shown for the different values of the parameters: a1 = 1, a2 = −0.5,
b = 0.2, h = 0.2, x(s) = 3.5, s ∈ [−0.2, 0], p1 = p2 = 0.5, λ12 = 15, λ21 = 1,  = 0.01

Therefore, we obtain the following algorithm: if after simulation of ζi , we have

ζi > λ12 , then we put ηi = a1 , else we put ηi = a2 . Analogously, if ηi−1 = a2 ,
i > 0, then from (7.36) for small enough  > 0, we have

P{ηi = a1 /ηi−1 = a2 } = P{ζi < λ21 } = λ21 ,

(7.40)
P{ηi = a2 /ηi−1 = a2 } = P{ζi > λ21 } = 1 − λ21 .

By (7.38)–(7.40) we obtain the following algorithm: if after simulation of ζi , we

have ζi < λ21 , then we put ηi = a1 , else we put ηi = a2 .
In Fig. 7.4 four possible trajectories of the Markov chain η(t) and four appro-
priate trajectories of the solution of (7.34) are shown for the following values of
the parameters: a1 = 1, a2 = −0.5, b = 0.2, h = 0.2, x(s) = 3.5, s ∈ [−0.2, 0],
p1 = p2 = 0.5, λ12 = 15, λ21 = 1,  = 0.01. In Fig. 7.5 hundred trajectories of
the solution of (7.34) are shown for the same values of the parameters. We can see
that all trajectories converge to zero in the whole accordance with the properties of
stability.
Put now λ12 = 5 with the same values of the other parameters. In this case the
point A(1, −0.5) does not belong to the stability region (Fig. 7.3), and we have
(Figs. 7.6 and 7.7) another situation: the trajectories of the solution fill by itself
7.4 Numerical Simulation of Systems with Markovian Switching 201

Fig. 7.5 Hundred trajectories of the solution of (7.34) are shown for the same as in Fig. 7.4 values
of the parameters

Fig. 7.6 Four possible trajectories of the Markov chain η(t) and four appropriate trajectories of
the solution of (7.34) are shown for the different values of the parameters: a1 = 1, a2 = −0.5,
b = 0.2, h = 0.2, x(s) = 3.5, s ∈ [−0.2, 0], p1 = p2 = 0.5, λ12 = 5, λ21 = 1,  = 0.01
202 7 Stochastic Systems with Markovian Switching

Fig. 7.7 Hundred trajectories of the solution of (7.34) are shown for the same as in Fig. 7.6 values
of the parameters

the whole admissible space between the solutions of (7.34) with η(t) ≡ a1 and
η(t) ≡ a2 . This shows us the instability of the trivial solution of (7.34).

7.4.2 System with Stochastic Perturbations

Here we will investigate the stability of the trivial solution of stochastic differen-
tial equation (7.15) using a simultaneous numerical simulation of both stochastic
processes: the Markov chain η(t) with two states as it is described in the previous
section and the Wiener process w(t) as it is described in Sect. 2.1.1.
Consider the difference analogue of (7.15) of the form

xi+1 = (1 + ηi )xi + bxi−m  + σ xi−l wi+1 ,

where

xi = x(ti ), ηi = η(ti ), wi+1 = Wn(c) (ti+1 ) − Wn(c) (ti ),

ti = i, h = m, τ = l,  > 0.

In Fig. 7.8 four realizations of one trajectory of the Wiener process w(t), one trajec-
tory of the Markov chain η(t), and one trajectory of the solution x(t) of (7.15) are
shown for a1 = −1, a2 = 1, b = −1, σ = 1, λ12 = 1, λ21 = 5, h = 0.99, τ = 1.6,
x0 = 7.5. In Fig. 7.9 we can see at the same time ten trajectories of the solution
x(t) of (7.15) for the referred above values of the parameters. All trajectories of the
solution x(t) go to zero. In Fig. 7.10 we can see the similar picture for σ = 0.1 and
7.4 Numerical Simulation of Systems with Markovian Switching 203

Fig. 7.8 Four realizations of one trajectory of the Wiener process w(t), one trajectory of the
Markov chain η(t), and one appropriate trajectory of the solution x(t) of (7.15) are shown for
a1 = −1, a2 = 1, b = −1, σ = 1, λ12 = 1, λ21 = 5, h = 0.99, τ = 1.6, x0 = 7.5

the same values of the other parameters. We can see that in this case all trajectories
of the solution x(t) go to zero more quickly.

7.4.3 System with Random Delay

Consider the differential equation

 
ẋ(t) + bx t − η(t) = 0,
(7.41)
x(s) = φ(s), s ∈ [−h, 0],

with random delay. Here b > 0, η(t) is a Markov chain with two states {a1 , a2 }
such that 0 < a1 < a2 = h, the initial distribution (7.35), and the probabilities of
transition (7.36).
204 7 Stochastic Systems with Markovian Switching

Fig. 7.9 Ten trajectories of the solution x(t) of (7.15) for the values of the parameters in Fig. 7.8

Fig. 7.10 Ten trajectories of the solution x(t) of (7.15) for the values of the parameters in Fig. 7.8
except for σ = 0.1

As it was shown in Example 1.2, if η(t) = h = const, then the inequality bh <
π
2 is a necessary and sufficient condition for the asymptotic stability of the trivial
solution of (7.41).
Let us investigate the stability of the trivial solution of (7.41) using the numeri-
cal simulation of the Markov chain η(t) as in the previous examples. Consider the
7.4 Numerical Simulation of Systems with Markovian Switching 205

Fig. 7.11 One trajectory of the Markov chain η(t)and one appropriate trajectory of the solution
of (7.41) are shown for b = 1, a1 = 1, a2 = 2, x(s) = 2.5, s ∈ [−2, 0], p1 = p2 = 0.5, λ12 = 1,
λ21 = 3,  = 0.001

difference analogue of (7.41) of the form

xi+1 = xi − bxi−ηi ,
xi = x(ti ), ηi = η(ti ), ti = i,  > 0.

Put b = 1, a1 = 1, a2 = 2, x(s) = const, s ∈ [−h, 0],  = 0.001. Several solu-

tions of (7.41) with different values of the initial function are shown in Figs. 1.1, 1.2,
and 1.3 for η(t) ≡ a1 = 1, η(t) ≡ 12 π , η(t) ≡ a2 = 2, respectively. We can see that
all solutions converge to zero in the first case, are bounded only but do not converge
to zero in the second case, and go to ±∞ in the last case.
In Fig. 7.11 one trajectory of the Markov chain η(t) and one appropriate trajec-
tory of the solution of (7.41) are shown for p1 = p2 = 0.5, λ12 = 1, λ21 = 3 and the
same values of other parameters. In Fig. 7.12 hundred trajectories of the solution of
(7.41) are shown for the same values of the parameters. We can see that all trajecto-
ries converge to zero. If λ21 = 6, then hundred trajectories of the solution of (7.41)
converge to zero more quickly (Fig. 7.13).
In Fig. 7.14 one trajectory of the Markov chain η(t) and one appropriate trajec-
tory of the solution of (7.41) are shown for p1 = p2 = 0.5, λ12 = 3, λ21 = 1. In
Fig. 7.15 hundred trajectories of the solution of (7.41) are shown for these values of
λ12 and λ21 . We can see that in this case the solution of (7.41) is unstable.
206 7 Stochastic Systems with Markovian Switching

Fig. 7.12 Hundred trajectories of the solution of (7.41) are shown for the same as in Fig. 7.11
values of the parameters

Fig. 7.13 Hundred trajectories of the solution of (7.41) are shown for the same values of the
parameters as in Fig. 7.11 except for λ21 = 6

7.4.4 Some Generalization of Algorithm of Markov Chain

Numerical Simulation

Let us show that the proposed numerical simulation (7.38)–(7.40) of the Markov
chain with two states can be generalized to a Markov chain η(t) with n states
7.4 Numerical Simulation of Systems with Markovian Switching 207

Fig. 7.14 One trajectory of the Markov chain η(t) and one appropriate trajectory of the solution
of (7.41) are shown for b = 1, a1 = 1, a2 = 2, x(s) = 2.5, s ∈ [−2, 0], p1 = p2 = 0.5, λ12 = 3,
λ21 = 1,  = 0.001

Fig. 7.15 Hundred trajectories of the solution of (7.41) are shown for the same values of the
parameters as in Fig. 7.14

{a1 , . . . , an }, initial distribution

 
pi = P η(0) = ai , i = 1, . . . , n,
208 7 Stochastic Systems with Markovian Switching

and the probabilities of transition

 
pij () = P η(t + ) = aj /η(t) = ai = λij  + o(),
i, j = 1, . . . , n, i = j.

Indeed, put
ηk = η(tk ), tk = k,  > 0.
Reduce a simulation of the Markov chain ηk , k = 0, 1, . . . , to a simulation of a se-
quence of independent random variables ζk that are uniformly distributed on [0, 1].
Note that
pi = P{η0 = ai } = P{Si−1 < ζ0 < Si }, i = 1, . . . , n,
where

i
S0 = 0, Si = pj , i = 1, . . . , n − 1, Sn = 1.
j =1

It is easy to see that for each result of simulation ζ0 , there exists a number i such
that Si−1 < ζ0 < Si . So, we put η0 = ai .
Further, put Qi0 = 0,


j
Qij = λil , 1 ≤ j < i,
l=1


i−1 
j
Qij = λil  + λil , i < j ≤ n.
l=1 l=i+1

Then
P{ηk = aj /ηk−1 = ai } = P{Qi,j −1 < ζk < Qij } = λij 
for j < i or j > i + 1, and

P{ηk = ai+1 /ηk−1 = ai } = P{Qi,i−1 < ζk < Qi,i+1 } = λi,i+1 ,

P{ηk = ai /ηk−1 = ai } = P{Qin < ζk } = 1 − Qin .

Thus, we obtain the following algorithm. Let ηk−1 = ai , k > 0. If after simulation
of ζk , there exists a number j ≤ n such that j = i and Qi,j −1 < ζk < Qij for j < i
or j > i + 1, Qi,i−1 < ζk < Qi,i+1 for j = i + 1, then we put ηk = aj . If such a
number j does not exist, i.e., ζk > Qin for i < n or ζk > Qn,n−1 for i = n, then we
put ηk = ai .
Chapter 8
Stabilization of the Controlled Inverted
Pendulum by a Control with Delay

The problem of stabilization for the mathematical model of the controlled inverted
pendulum (Fig. 8.1) during many years is very popular among the researchers (see,
for instance, [1, 2, 33, 36, 37, 114, 121, 169, 170, 201, 209, 228, 256, 273, 276,
278, 282]). Unlike the classical way of stabilization in which the stabilized control
is a linear combination of the state and velocity of the pendulum, here we propose
another way of stabilization. We suppose that only the trajectory of the pendulum
can be observed and stabilized control depends on the whole trajectory of the pendu-
lum. We consider linear and nonlinear models of the controlled inverted pendulum
by stochastic perturbations and investigate zero and steady-state nonzero solutions
analytically and by numerical simulations.

8.1 Linear Model of the Controlled Inverted Pendulum

Here we consider the linearized mathematical model of the controlled inverted pen-
dulum. We show that by some controls depending on a delay the trivial solution
of the considered model can be asymptotically stable but by some other controls
depending on a delay it is impossible. We also obtain sufficient conditions for stabi-
lization of the trivial solution of the considered model by stochastic perturbations.

8.1.1 Stabilization by the Control Depending on Trajectory

The linearized mathematical model of the controlled inverted pendulum can be de-
scribed by the linear second-order differential equation

ẍ(t) − ax(t) = u(t), a > 0, t ≥ 0. (8.1)

The classical way of stabilization [121] for (8.1) uses the control u(t) in the form

u(t) = −b1 x(t) − b2 ẋ(t), b1 > a, b2 > 0.

L. Shaikhet, Lyapunov Functionals and Stability of Stochastic Functional 209

Differential Equations, DOI 10.1007/978-3-319-00101-2_8,
© Springer International Publishing Switzerland 2013
210 8 Stabilization of the Controlled Inverted Pendulum by a Control with Delay

Fig. 8.1 Controlled inverted

pendulum

But this type of control, which represents instantaneous feedback, is quite difficult
to realize because usually it is necessary to have some finite time to make measure-
ments of the coordinates and velocities, to treat the results of the measurements, and
to implement them in the control action.
Here we propose another way of stabilization. We suppose that only the tra-
jectory of the pendulum is observed and the control u(t) does not depend on the
velocity but it depends on the previous values of the trajectory x(s), s ≤ t, and has
the form
 ∞
u(t) = dK(τ ) x(t − τ ), (8.2)
0

where the kernel K(τ ) is a right-continuous function of bounded variation on

[0, ∞], and the integral is understood in the Stieltjes sense. This means in particu-
lar that both distributed and discrete delays can be used depending on the concrete
choice of the kernel K(τ ).
The initial condition for system (8.1)–(8.2) has the form

x(s) = φ(s), ẋ(s) = φ̇(s), s ≤ 0, (8.3)

where φ(s) is a given continuously differentiable function.

Let us show that the inverted pendulum (8.1) can be stabilized by the con-
trol (8.2). Substituting (8.2) into (8.1) and putting x1 (t) = x(t), x2 (t) = ẋ(t), we
obtain the system of differential equations with delay

ẋ1 (t) = x2 (t),

 ∞ (8.4)
ẋ2 (t) = ax1 (t) + dK(τ ) x1 (t − τ ).
0
8.1 Linear Model of the Controlled Inverted Pendulum 211

To prove the asymptotic stability of the trivial solution of (8.4), we will use the
procedure of constructing Lyapunov functionals. Put
 ∞  ∞
 
ki = τ dK(τ ), i = 0, 1,
i
kj = τ j dK(τ ), j = 2, 3. (8.5)
0 0

Since
 t  t
x2 (s) ds = ẋ1 (s) ds = x1 (t) − x1 (t − τ ),
t−τ t−τ
then by (8.5) we have
 ∞  ∞  t
dK(τ ) x1 (t − τ ) = k0 x1 (t) − dK(τ ) x2 (s) ds.
0 0 t−τ

Therefore, from the second equation of (8.4) it follows that

 ∞  t
ẋ2 (t) = (a + k0 )x1 (t) − dK(τ ) x2 (s) ds. (8.6)
0 t−τ

Put now
a1 = −(a + k0 ), z(t) = x2 (t) − G(t, x2t ),
 ∞  t (8.7)
G(t, x2t ) = dK(τ ) (s − t + τ )x2 (s) ds.
0 t−τ
By (8.5) we have
 ∞  t  ∞  t
d
dK(τ ) (s − t + τ )x2 (s) ds = k1 x2 (t) − dK(τ ) x2 (s) ds.
dt 0 t−τ 0 t−τ
(8.8)
Using (8.6), (8.7), and (8.8), we reduce (8.4) to the form

ẋ1 (t) = x2 (t),

(8.9)
ż(t) = −a1 x1 (t) − k1 x2 (t),

which is the first step of the procedure.

Following the second step of the procedure of constructing Lyapunov function-
als, we consider the auxiliary system of the ordinary differential equations

ẏ1 (t) = y2 (t),

(8.10)
ẏ2 (t) = −a1 y1 (t) − k1 y2 (t).

Rewriting (8.10) in the matrix form

   
y1 0 1
ẏ(t) = Ay(t), y= , A= , (8.11)
y2 −a1 −k1
212 8 Stabilization of the Controlled Inverted Pendulum by a Control with Delay

and using Corollary 1.1, we obtain that the inequalities

a1 > 0, k1 > 0, (8.12)

are necessary and sufficient conditions for the asymptotic stability of the zero solu-
tion of (8.10).
From Theorem 1.3, Remark 1.1, and (1.29) it follows that by conditions (8.12)
the matrix equation (1.27) with the matrix A defined in (8.11) and the symmetric
matrix Q with the elements q11 = q > 0, q22 = 1, q12 = 0 has the matrix solution
P with the elements
q p q + a1
p11 = k1 p12 + a1 p22 , p12 = , p22 = , p= . (8.13)
2a1 2a1 k1

Following the third step of the procedure of constructing Lyapunov function-

als, we will construct a Lyapunov functional for (8.9) in the form V = V1 + V2 ,
where

V1 = p11 x12 (t) + 2p12 x1 (t)z(t) + p22 z2 (t) (8.14)

with p11 , p12 , p22 , and z(t) are defined in (8.13) and (8.7).
Calculating V̇1 , by (8.14), (8.9), and (8.7) we have
    
V̇1 = 2 p11 x1 (t) + p12 z(t) x2 (t) − 2 p12 x1 (t) + p22 z(t) a1 x1 (t) + k1 x2 (t)
= −2p12 a1 x12 (t) − 2(k1 p22 − p12 )x22 (t) + 2(p11 − k1 p12 − a1 p22 )x1 (t)x2 (t)
+ 2p22 a1 x1 (t)G(t, x2t ) + 2(k1 p22 − p12 )x2 (t)G(t, x2t ).

Using (8.7), (8.13), and arbitrary γ > 0, we obtain

 ∞  t
V̇1 = −qx12 (t) − x22 (t) + p dK(τ ) (s − t + τ )x1 (t)x2 (s) ds
0 t−τ
 ∞  t
+ dK(τ ) (s − t + τ )x2 (t)x2 (s) ds
0 t−τ
 ∞   
p  t 1
≤ −qx12 (t) − x22 (t) + dK(τ ) (s − t + τ ) γ x12 (t) + x22 (s) ds
2 0 t−τ γ
 
1 ∞   t  
+ dK(τ ) (s − t + τ ) x22 (t) + x22 (s) ds
2 0 t−τ
   
1 1
= − q − γpk2 x1 (t) − 1 − k2 x22 (t)
2
4 4
 ∞  t
 
+α dK(τ ) (s − t + τ )x22 (s) ds, (8.15)
0 t−τ
8.1 Linear Model of the Controlled Inverted Pendulum 213

where
 
1 p
α= 1+ . (8.16)
2 γ
Following the fourth step of the procedure of constructing Lyapunov functionals,
let us choose the functional V2 in the form
 
α ∞   t

V2 = dK(τ ) (s − t + τ )2 x22 (s) ds. (8.17)
2 0 t−τ

Then
 ∞ 
αk2 2  t
V̇2 = x (t) − α dK(τ ) (s − t + τ )x22 (s) ds, (8.18)
2 2 0 t−τ
and by (8.16), for the functional V = V1 + V2 , we obtain
   
γpk2 2 k2 pk2 2
V̇ ≤ − q − x1 (t) − 1 − − x2 (t). (8.19)
4 2 4γ
From the condition of positivity of the expressions in the brackets we have
4q pk2
>γ > > 0.
pk2 2(2 − k2 )
So, if
4q pk2
> > 0, (8.20)
pk2 2(2 − k2 )
then there exists γ > 0 such that the Lyapunov functional V for some c > 0 satisfies
the condition V̇ ≤ −c(x12 (t) + x22 (t)).
Using the representation (8.13) for p, rewrite (8.20) in the form
 
2(2 − k2 )k12 (a1 + q)2 a1 a1 q
> = 2 + +
k22 4q 4 q a1

and note that the right-hand part of the obtained inequality reaches its minimum for
q = a1 . Using this q and (8.7), (8.12), we obtain the following theorem.

Theorem 8.1 Let

4
a1 = −(a + k0 ) > 0, k1 > 0, k2 <  . (8.21)
1 + 1 + 4a1 k1−2

Then the trivial solution of (8.4) is asymptotically stable.

Remark 8.1 Note that two first inequalities in (8.21) are necessary conditions for
asymptotic stability of the trivial solution of (8.4), while the third inequality in (8.21)
is a sufficient condition only.
214 8 Stabilization of the Controlled Inverted Pendulum by a Control with Delay

Remark 8.2 Note that the third inequality in (8.21) can be represented in the form

a1
k 1 > k2 .
2(2 − k2 )

Remark 8.3 Let us show that the functional G(t, x2t ) in (8.7) satisfies condi-
tion (2.10) for a differential equation of neutral type. Indeed, transform G(t, x2t )
in the following way:
 ∞  τ
G(t, x2t ) = dK(τ ) (τ − θ )x2 (t − θ ) dθ
0 0
 ∞  ∞
= x2 (t − θ ) (τ − θ ) dK(τ ) dθ.
0 θ

From (8.21) it follows that k2 < 2. So, by (8.5),

 ∞  ∞   ∞ 
    τ
 
(τ − θ ) dK(τ ) dθ  ≤  dK(τ ) (τ − θ ) dθ

0 θ 0 0

1 ∞   k2
= τ 2 dK(τ ) = < 1.
2 0 2

8.1.2 Some Examples

Here we consider some different examples of the controls of type (8.2) and show
that, for some of them, a stabilization is possible but, for some of them, a stabiliza-
tion is impossible.

Example 8.1 Put in (8.2) dK(τ ) = bδ(τ − h) dτ , where h > 0, and δ(τ ) is the Dirac
delta-function. In this case system (8.1)–(8.2) has the form

ẍ(t) − ax(t) = bx(t − h). (8.22)

From (8.21) we obtain a contradiction for b: k0 = b < −a < 0, k1 = bh > 0. From

Theorem 8.1 and Remark 8.1 it follows that the trivial solution of (8.22) cannot be
asymptotically stable for any b and h, i.e., the inverted pendulum (8.1) cannot be
stabilized by the control of the form u(t) = bx(t − h).
Let us obtain the same statement using the characteristic equation of (8.22) z2 −
a − be−hz = 0, z = α + iβ, that can be transformed to the following system of two
equations for real α and β:

α 2 − β 2 − a − be−hα cos(hβ) = 0, 2αβ + be−hα sin(hβ) = 0. (8.23)

Let us show that system (8.23) for every b and h has at least one solution with
α ≥ 0. Suppose first that a + b > 0. In this case, for β = 0, we have α 2 − a = be−hα .
8.1 Linear Model of the Controlled Inverted Pendulum 215

Consider the characteristic quasipolynomial (α) = α 2 − a − be−hα . Since (0) =

−(a +b) < 0 and limα→∞ (α) = +∞, then there exists α > 0 such that (α) = 0.
Let now a + b = 0. In this case, α = β = 0 is the solution of system (8.23). Let at
last a + b < 0. Since a > 0, b < −a < 0. Choose β so that 0 < hβ < π2 . Then
cos(hβ) > 0 and sin(hβ) > 0. From (8.23) it follows that

α2 − β 2 − a 2αβ
=− = be−hα < 0.
cos(hβ) sin(hβ)

This means that α

is a positive root of the equation α 2 + 2αγ = a + β 2 , γ =
β cot(hβ), i.e., α = a + β 2 + γ 2 − γ > 0.
Thus, for every b and h, system (8.23) has at least one solution with α ≥ 0, i.e.,
the trivial solution of (8.22) cannot be asymptotically stable.

Example 8.2 Put in (8.2) dK(τ ) = (b1 δ(τ − h1 ) + b2 δ(τ − h2 )) dτ , where

h1 , h2 > 0, and δ(τ ) is the Dirac delta-function. In this case system (8.1)–(8.2)
has the form

ẍ(t) − ax(t) = b1 x(t − h1 ) + b2 x(t − h2 ), (8.24)

and by (8.5) the stability conditions (8.21) are

a1 = −(a + b1 + b2 ) > 0, k1 = b1 h1 + b2 h2 > 0,

4 (8.25)
k2 = |b1 |h21 + |b2 |h22 < km =  .
1 + 1 + 4a1 k1−2

Let us show that for an arbitrary a > 0, there exist b1 , b2 , h1 , h2 such that con-
ditions (8.25) hold and therefore the trivial solution of (8.24) is asymptotically sta-
ble.
First, note that (8.24) was already considered in Example 1.7, where it was shown
that the inequality a + b1 + b2 < 0 is a necessary condition for asymptotic stability.
This condition coincides with the first condition in (8.25).
Put now b1 = b, b2 = −αb, h1 = h, h2 = βh. Here b is an arbitrary positive
number, and positive numbers α, β, and h will be chosen below.
By that the first condition in (8.25) takes the form a + b − αb < 0 and holds
if α > 1 + ab−1 . The second condition in (8.25) takes the form bh − αβbh =
bh(1 − αβ) > 0 and holds if β < α −1 . The third condition in (8.25) in this case has
the form

4   4(α − 1 − ab−1 )
Ah2 < √ , A = b 1 + αβ 2 , B = > 0.
1 + 1 + Bh−2 b(1 − αβ)2
√
It is easy to show that it holds if h < 4( A(AB + 8))−1 .
216 8 Stabilization of the Controlled Inverted Pendulum by a Control with Delay

So, we have shown that for an arbitrary a > 0, the parameters b1 , b2 , h1 , and h2
can be chosen such that all conditions (8.25) hold and therefore the trivial solution
of (8.24) is asymptotically stable.

Example 8.3 Consider the control (8.2) in the form

 h1
u(t) = b x(t − s) ds, h1 > h2 ≥ 0. (8.26)
h2

In this case from (8.21) we obtain a contradiction for b:

b 2 
k0 = b(h1 − h2 ) < −a < 0, k1 = h1 − h22 > 0.
2
Thus, there do not exist b, h1 , h2 such that conditions (8.21) hold. This means that
system (8.1) cannot be stabilized by the control (8.26).
On the other hand, the characteristic equation of system (8.1), (8.26) has the form
h
z2 − a − b h21 e−zs ds = 0, z = α + iβ or, in the form of a system of two equations
for real α and β,

α 2 − β 2 − a − bIc = 0, 2αβ + bIs = 0,

 h1  h1 (8.27)
Ic = e−αs cos(βs) ds, Is = e−αs sin(βs) ds.
h2 h2

Let us show that for arbitrary b, h1 , h2 , there exists a solution of (8.27) with
nonnegative α. Let first a +b(h1 −h2 ) > 0. Put β = 0 and consider the characteristic
quasipolynomial
 h1 b  −αh2 
(α) = α 2 − a − b e−αs ds = α 2 − a − e − e−αh1 .
h2 α

It is easy to see that limα→0 (α) = −[a + b(h1 − h2 )] < 0 but limα→∞ (α) =
∞ > 0. Therefore, there exists at least one α0 > 0 such that (α0 ) = 0 and (α0 , 0)
is a solution of (8.27). If a + b(h1 − h2 ) = 0, then α = β = 0 is a solution of (8.27).
Let now a + b(h1 − h2 ) < 0. Since a > 0, b < 0. Choose β > 0 such that βh1 < π2 .
Then Ic > 0 and Is > 0. From (8.27) it follows that

α2 − β 2 − a 2αβ
=− = b < 0.
Ic Is

This means that there exists α > 0 that is a positive root of the equation f (α) = 0,
where f (α) = α 2 + 2αγ − a − β 2 , γ = βIc Is−1 . Indeed, it is easy to see that
f (0) = −a − β 2 < 0 and limα→∞ f (α) = ∞ > 0. So, this equation has a positive
root.
8.1 Linear Model of the Controlled Inverted Pendulum 217

Example 8.4 Consider the control (8.2) in the form

 h1  h3
u(t) = b1 x(t − s) ds + b2 x(t − s) ds,
h2 h4
0 < h4 < h3 < h2 < h1 . (8.28)

In this case from (8.21) we obtain

k0 = b1 (h1 − h2 ) + b2 (h3 − h4 ) < −a, (8.29)

b1  2  b2  2 
k1 = h1 − h22 + h3 − h24 > 0, (8.30)
2 2

|b1 |  3  |b2 |  3  4
k2 = h1 − h32 + h3 − h34 <  . (8.31)
3 3 1 + 1 + 4|a + k0 |k1−2
Let us show that there exist b1 , b2 , h1 , h2 , h3 , h4 such that conditions (8.29)–(8.31)
hold. Put, for instance,
a
b1 = , b2 = −γ b1 , h1 = h, hi+1 = αi h, i = 1, 2, 3, (8.32)
h
where h > 0, γ > 0,
0 < α3 < α2 < α1 < 1. (8.33)
In this case conditions (8.29)–(8.30) have the forms respectively
 
k0 = a 1 − α1 − γ (α2 − α3 ) < −a,
ah   
k1 = 1 − α12 − γ α22 − α32 > 0.
2
These conditions hold if the parameters γ and αi , i = 1, 2, 3, satisfy the inequalities

2 − α1 1 − α12
<γ < 2 . (8.34)
α2 − α3 α2 − α32

Substituting (8.32) into (8.31), we obtain


2
h< , (8.35)
αa(1 + 2αβ)

where
1   α1 + γ (α2 − α3 ) − 2
α= 1 − α13 + γ α23 − α33 , β= . (8.36)
3 (1 − α12 − γ (α22 − α32 ))2
218 8 Stabilization of the Controlled Inverted Pendulum by a Control with Delay

Fig. 8.2 Solutions of (8.1), (8.28) for a = 1, h1 = 0.48, h2 = 0.24, h3 = 0.12, h4 = 0.06,
b1 = ah−1
1 , b2 = −γ b1 , x(s) = 3, s ∈ [−0.48, 0], and the different values of γ : (1) γ = 11.5,
(2) γ = 11.99, (3) γ = 12, (4) γ = 13, (5) γ = 16.3

Thus, if the parameters of (8.1) with the control (8.28) satisfy conditions (8.32)–
(8.36), then the trivial solution of (8.1), (8.28) is asymptotically stable.
From (8.33)–(8.34) it follows that the parameters αi , i = 1, 2, 3, should satisfy
the conditions
√
0 < α3 < 2 − 3,
√
α3 < α2 < 2(2 − 3) − α3 , (8.37)
1

α2 < α1 < δ + δ 2 − 8δ + 4 , δ = α2 + α3 .
2

Put, for instance, a = 1, αi = 2−i , i = 1, 2, 3. Conditions (8.37) hold, and condition

(8.34) takes the form 12 < γ < 16. Putting γ = 13, from (8.35)–(8.36) we obtain
h < 1.02392. Putting γ = 12.1, we have h < 2.13984.
In Fig. 8.2 the solutions of (8.1), (8.28) are shown for a = 1, h1 = 0.48,
h2 = 0.24, h3 = 0.12, h4 = 0.06, b1 = ah−1 1 , b2 = −γ b1 , x(s) = 3, s ≤ [−0.48, 0],
and different values of the parameter γ : (1) γ = 11.5, (2) γ = 11.99, (3) γ = 12,
(4) γ = 13, (5) γ = 16.3. Only one of these solutions with γ = 13 is asymptotically
stable.
8.1 Linear Model of the Controlled Inverted Pendulum 219

8.1.3 About Stabilization by the Control Depending on Velocity

or on Acceleration

Consider (8.1) with the control

 ∞
u(t) = dK(τ ) ẋ(t − τ ), (8.38)
0

where
 ∞ 
dK(τ ) < ∞. (8.39)
0
The characteristic equation of system (8.1), (8.38) has the form
 ∞
z2 − a − z dK(τ ) e−zτ = 0, z = α + βi. (8.40)
0

Let us show that this equation has at least one root z with positive real part α.
∞
Indeed, put β = 0 and (α) = α 2 − a − α 0 dK(τ ) e−ατ . Then (8.40) takes the
form (α) = 0. Note that (0) = −a < 0. Using (8.39), it is easy to get that
  
a 1 ∞
lim (α) = lim α 2 1 − 2 − dK(τ ) e−ατ = ∞.
α→∞ α→∞ α α 0

Therefore, there exists at least one α > 0 that is a root of the equation (α) = 0.
Thus, the inverted pendulum cannot be stabilized by the control of type (8.38) de-
pending on the velocity only.
Consider (8.1) with the control
 ∞
u(t) = dK(τ ) ẍ(t − τ ), (8.41)
0

where

    ∞ 
dK(0) = K(+0) − K(0) < 1, dK(τ ) < ∞. (8.42)
+0
The characteristic equation of system (8.1), (8.41) has the form
 ∞
z2 − a − z2 dK(τ ) e−zτ = 0, z = α + βi. (8.43)
0

Let us show that this equation has at leastone root z with positive real part α.
∞
Indeed, put β = 0 and (α) = α 2 − a − α 2 0 dK(τ ) e−ατ . Then (8.43) takes the
form (α) = 0. Note that (0) = −a < 0. Using (8.42), it is easy to get that
  ∞ 
a −ατ
lim (α) = lim α 1 − 2 − dK(0) −
2
dK(τ ) e = ∞.
α→∞ α→∞ α +0
220 8 Stabilization of the Controlled Inverted Pendulum by a Control with Delay

Fig. 8.3 Solutions of (8.44) for a = 0.01, h = 5, x(s) = 0.01, s ∈ [−5, 0] and the different values
of b: (1) b = −0.01, (2) b = 0.01

Therefore there exists at least one α > 0 that is a root of the equation (α) = 0.
Thus, the inverted pendulum cannot be stabilized by the control of type (8.41) de-
pending on the acceleration only.
In Fig. 8.3 the solution x(t) of the equation

ẍ(t) − ax(t) = bẍ(t − h) (8.44)

is shown for a = 0.01, h = 5, x(s) = 0.01, s ∈ [−5, 0], and the different values of b:
(1) b = −0.01, (2) b = 0.01.

8.1.4 Stabilization by Stochastic Perturbations

Substituting (8.2) into (8.1) and supposing that the parameter a in (8.1) is under the
influence of stochastic perturbations of the type of white noise, we obtain
 ∞
 
ẍ(t) − a + σ ẇ(t) x(t) = dK(τ ) x(t − τ ), a > 0, t ≥ 0. (8.45)
0

Here w(t) is the standard Wiener process, and σ is a constant.

8.1 Linear Model of the Controlled Inverted Pendulum 221

Theorem 8.2 Let

a1 = −(a + k0 ) > 0, k1 > 0,

  (8.46)
a1
σ 2 < 2a1 k1 − k2 ,
2(2 − k2 )

where ki , i = 0, 1, 2, are defined in (8.5). Then the trivial solution of (8.45) is asymp-
totically mean-square stable.

Proof Put x1 (t) = x(t) and x2 (t) = ẋ(t). Then (8.45), similarly to (8.9), can be
rewritten in the form of the system of stochastic differential equations of neutral
type

ẋ1 (t) = x2 (t),

(8.47)
ż(t) = −a1 x1 (t) − k1 x2 (t) + σ x1 (t)ẇ(t),

with the initial conditions x1 (s) = φ(s), x2 (s) = φ̇(s), s ≤ 0. Here k1 , a1 , and z(t)
are defined in (8.5), (8.7).
We will construct a Lyapunov functional for (8.47) in the form V = V1 + V2 ,
where V1 is defined by (8.14), (8.13). Calculating LV1 , where L is the generator
of (8.47), and using (8.7), (8.13), we obtain
 
LV1 = 2 p11 x1 (t) + p12 z(t) x2 (t)
  
− 2 p12 x1 (t) + p22 z(t) a1 x1 (t) + k1 x2 (t) + σ 2 p22 x12 (t)
 
= − 2p12 a1 − σ 2 p22 x12 (t) − 2(k1 p22 − p12 )x22 (t)
+ 2(p11 − k1 p12 − a1 p22 )x1 (t)x2 (t)
+ 2p22 a1 x1 (t)G(t, x2t ) + 2(k1 p22 − p12 )x2 (t)G(t, x2t )
 ∞  t
 
= − q − σ 2 p22 x12 (t) − x22 (t) + p dK(τ ) (s − t + τ )x1 (t)x2 (s) ds
0 t−τ
 ∞  t
+ dK(τ ) (s − t + τ )x2 (t)x2 (s) ds.
0 t−τ

Similarly to (8.15), for arbitrary γ > 0 and α defined by (8.16), we have

   
1 1
LV1 ≤ − q − γpk2 − σ 2 p22 x12 (t) − 1 − k2 x22 (t)
4 4
 ∞  t
 
+α dK(τ ) (s − t + τ )x22 (s) ds.
0 t−τ
222 8 Stabilization of the Controlled Inverted Pendulum by a Control with Delay

Choosing the functional V2 in the form (8.17) and using (8.16), for the functional
V = V1 + V2 , similarly to (8.19), we obtain
   
γpk2 k2 pk2 2
LV ≤ − q − − σ 2 p22 x12 (t) − 1 − − x2 (t).
4 2 4γ

Supposing the positivity of the expressions in the brackets, we obtain

4(q − σ 2 p22 ) pk2

>γ > > 0.
pk2 2(2 − k2 )

So, if
4(q − σ 2 p22 ) pk2
> > 0, (8.48)
pk2 2(2 − k2 )
then there exists γ > 0 such that the Lyapunov functional V for some c > 0 satisfies
the condition LV ≤ −c(x12 (t) + x22 (t)).
By the representation (8.13) for p22 and p from (8.48) it follows that
 
q k22
σ 2 < 2a1 k1 − A(q + a1 ) , A= . (8.49)
q + a1 8k12 (2 − k2 )

The right-hand part of inequality (8.49) reaches its maximum at q = a1 A−1 − a1 .

Substituting this q into (8.49), we obtain the last inequality (8.46). The proof is
completed. 

Note that by Remark 8.2 conditions (8.46) by σ = 0 coincide with (8.21).

8.2 Nonlinear Model of the Controlled Inverted Pendulum

8.2.1 Stabilization of the Trivial Solution

Consider separately the deterministic case and the stochastic case.

8.2.1.1 The Deterministic Case

Consider the problem of stabilization for the nonlinear model of the controlled in-
verted pendulum
ẍ(t) − a sin x(t) = u(t), a > 0, (8.50)
via the control (8.2).
8.2 Nonlinear Model of the Controlled Inverted Pendulum 223

Let us suppose that conditions (8.21) hold. In Theorem 8.1 it is proved that by
conditions (8.21) on the kernel K(τ ) of the control (8.2) the appropriate linearized
system (8.4) is asymptotically stable. So, by conditions (8.21) the trivial solution of
the nonlinear system (8.50), (8.2) is asymptotically stable too if the initial function
(8.3) belongs to some small enough neighborhood of the origin, called a region of
attraction. Let us construct some estimate of the region of attraction for the trivial
solution of system (8.50), (8.2).
Similarly to (8.9), let us represent (8.50), (8.2) in the form

ẋ1 (t) = x2 (t),

  (8.51)
ż(t) = −a1 x1 (t) − k1 x2 (t) − af x1 (t) ,

where f (x) = x − sin x, k1 , and a1 and z(t) are defined by (8.5), (8.7).
Following the procedure of constructing Lyapunov functionals, let us choose the
auxiliary ordinary differential equations for system (8.51) in the form (8.10). We
will construct a Lyapunov functional for (8.51) in the form V = V1 + V2 , where the
functional V1 is defined again by (8.14), (8.13). Calculating V̇1 for system (8.51),
similarly to (8.15), we obtain
   
1 1
V̇1 ≤ − q − γpk2 x12 (t) − 1 − k2 x22 (t)
4 4
 ∞  t
 
+α dK(τ ) (s − t + τ )x 2 (s) ds
2
0 t−τ
   
− 2ap12 x1 (t)f x1 (t) − 2ap22 x2 (t)f x1 (t)
 ∞  t
 
+ 2ap22 dK(τ ) (s − t + τ )x2 (s)f x1 (t) ds,
0 t−τ

where p and α are defined by (8.13) and (8.16).

Suppose that for all t ≥ 0 and some positive δ,
 
x1 (t) ≤ δ. (8.52)

Using (8.52), the inequality |f (x)| = |x − sin x| ≤ 16 |x|3 , and some ν > 0, we get

   4 2
x1 (t)f x1 (t)  ≤ x1 (t) ≤ δ x 2 (t),
6 6 1
   
   x12 (t) 1 2 δ2 1 2
 
2 x2 (t)f x1 (t) ≤ νx1 (t) + x2 (t) ≤
2
νx1 (t) + x2 (t) ,
2
6 ν 6 ν
224 8 Stabilization of the Controlled Inverted Pendulum by a Control with Delay

and
 ∞ 
 t   
2 dK(τ ) (s − t + τ )x2 (s)f x1 (t)  ds
0 t−τ
 ∞  
δ2  t 1 2
≤ dK(τ )
(s − t + τ ) νx1 (t) + x2 (s) ds
2
6 0 t−τ ν
   
δ 2 νk2 2 1 ∞   t

≤ x (t) + dK(τ ) (s − t + τ )x2 (s) ds .
2
6 2 1 ν 0 t−τ
As a result,
    
1 k2
V̇1 ≤ − q − γpk2 − δ1 2p12 + 1 + p22 ν x12 (t)
4 2
   ∞ 
k2 δ1 p22 2   t
− 1− − x2 (t) + α1 dK(τ ) (s − t + τ )x22 (s) ds,
4 ν 0 t−τ

where
aδ 2 δ1 p22
δ1 = , α1 = α + . (8.53)
6 ν
Choosing V2 in the form
 ∞ 
α1  t
V2 = dK(τ ) (s − t + τ )2 x22 (s) ds (8.54)
2 0 t−τ

and using (8.53) and (8.16), for the functional V = V1 + V2 , we obtain

   
γpk2 2 pk2 2
V̇ ≤ − q − A − x1 (t) − 1 − B − x2 (t), (8.55)
4 4γ

where
1
A = A0 + Dν, B = B0 + D ,
ν
k2
A0 = 2p12 δ1 , D = (1 + B0 )p22 δ1 , B0 =
. (8.56)
2
Supposing the positivity of the expressions in the brackets in (8.55), we obtain

4(q − A) pk2
>γ > > 0. (8.57)
pk2 4(1 − B)

So, if
4(q − A) pk2
> > 0, (8.58)
pk2 4(1 − B)
then there exists γ > 0 such that (8.57) holds.
8.2 Nonlinear Model of the Controlled Inverted Pendulum 225

Using (8.56), rewrite (8.58) in the form

   
1 pB0 2
(q − A0 − Dν) 1 − B0 − D > . (8.59)
ν 2

It is

easy to show that the left-hand side of this inequality reaches its maximum at
ν = (q − A0 )(1 − B0 )−1 . Substituting this ν into (8.59), we obtain

pB0
(q − A0 )(1 − B0 ) > D + .
2
By (8.56) and (8.13) this inequality can be represented in the form

 
1 √ a1   δ1
k1 (1 − δ0 )(1 − B0 ) > q+√ (1 + B0 )δ0 + B0 , δ0 = .
2 q a1

Note that the right-hand side of this inequality reaches its minimum at q = a1 . Using
this q, we obtain

√  
k1 (1 − δ0 )(1 − B0 ) > a1 (1 + B0 )δ0 + B0 .

This inequality can be rewritten in the form of the quadratic equation with respect
to δ0 ,
δ02 + 2μδ0 − ρ < 0, (8.60)
where, by (8.56),

k2 k 2 (2 − k2 ) 2k12 a1−1 (2 − k2 ) − k22

μ= + 1 , ρ= .
2 + k2 a1 (2 + k2 )2 (2 + k2 )2
Note that by Remark 8.2 ρ > 0. So, inequality (8.60)
holds for δ0 = 0, and therefore
it holds also for small enough δ0 > 0, namely δ0 < μ2 + ρ − μ, or by (8.53) and
the definition of δ0 ,

6a1
δ2 < μ2 + ρ − μ. (8.61)
a
Let λ0 > 0 and λ1 > 0 be respectively the minimal and maximal eigenvalues of
the positive definite matrix P with the elements pij defined by (8.13) with q = a1 .
Then by (8.14), (8.55), (8.54), and (8.5) we have
  α1 k 3
λ0 x12 (t) ≤ V (t) ≤ V (0) ≤ λ1 x12 (0) + z2 (0) + sup x 2 (s).
6 s≤0 2

From (8.7) and (8.5) it follows that

   
    ∞  0   k2  
z(0) ≤ x2 (0) + dK(τ ) (s + τ )x2 (s) ds ≤ 1 + supx2 (s).
0 −τ 2 s≤0
226 8 Stabilization of the Controlled Inverted Pendulum by a Control with Delay

So,
   
k2 2 α1 k3
λ0 x12 (t) ≤ λ1 x12 (0) + λ1 1 + + sup x22 (s),
2 6 s≤0

and by (8.52) the domain of attraction for the trivial solution of system (8.50), (8.2)
contains the set of initial functions satisfying the inequality
   
λ1 2 λ1 k2 2 α1 k3
φ (0) + 1+ + sup φ̇ 2 (s) ≤ δ 2 .
λ0 λ0 2 6λ0 s≤0

By this condition on the initial function φ(s) with δ > 0 that satisfies (8.61) the
solution of system (8.50), (8.2) satisfies condition (8.52).

8.2.1.2 The Stochastic Case

Consider now the nonlinear model of the controlled inverted pendulum (8.50) under
the influence of stochastic perturbations of the type of white noise
 
ẍ(t) − a + σ ẇ(t) sin x(t) = u(t), a > 0, t ≥ 0, (8.62)

with the control (8.2). Similarly to (8.51), rewrite system (8.62), (8.2) in the form

ẋ1 (t) = x2 (t),

     (8.63)
ż(t) = −a1 x1 (t) − k1 x2 (t) − af x1 (t) + σ x1 (t) − f x1 (t) ẇ(t),

where f (x) = x − sin x, and k1 , a1 , and z(t) are defined by (8.5) and (8.7).
3
Note that (8.47) is the linear part of (8.63). Since |f (x)| ≤ x6 , the order of non-
linearity of (8.63) is 3. From Theorem 5.2 it follows that if the order of nonlinearity
of the nonlinear system under consideration is higher than one, then the sufficient
condition for asymptotic mean-square stability of the linear part of this system is at
the same time a sufficient condition for stability in probability of the initial nonlinear
system. Thus, we obtain the following theorem.

Theorem 8.3 If conditions (8.46) hold, then the trivial solution of (8.62) by the
control (8.2) is stable in probability.

So, the nonlinear model of the controlled inverted pendulum under stochastic
perturbations can be stabilized by a control that depends on the trajectory only.

8.2.2 Nonzero Steady-State Solutions

Here nonzero steady-state solutions of the nonlinear system (8.50), (8.2), (8.3) are
studied. Substituting (8.2) into (8.50) and putting x1 (t) = x(t) and x2 (t) = ẋ(t), we
8.2 Nonlinear Model of the Controlled Inverted Pendulum 227

represent this system as follows:

ẋ1 (t) = x2 (t),

 ∞ (8.64)
ẋ2 (t) = a sin x1 (t) + dK(τ ) x1 (t − τ ).
0

To get steady-state solutions of (8.64), let us suppose that ẋ1 (t) ≡ 0 and ẋ2 (t) ≡ 0.
We obtain that x2 (t) ≡ 0 and x1 (t) ≡ x̂ is a root of the equation

a sin x̂ + k0 x̂ = 0. (8.65)

Suppose that x̂ = 0 and rewrite (8.65) in the form

S(x̂) = 0, (8.66)

where
sin x k0
S(x) = + . (8.67)
x a
We will call the function S(x) “the characteristic function of the system (8.64).”

Remark 8.4 The statements “x̂ is a steady-state solution of the system (8.64)” and
“x̂ is a root of (8.66)” are equivalent.

Remark 8.5 For all x = 0,

sin x
−α ≤ < 1,
x
where 0.217233 < α < 0.217234. Therefore, if
k0
−α ≤ − <1
a
or
0 < a + k0 ≤ (1 + α)a, (8.68)
then there exists at least one nonzero root of (8.66).

Remark 8.6 Since the function S(x) is an even function, if x̂ is a root of (8.66), then
−x̂ is a root of (8.66) too.

Remark 8.7 Condition (8.68) contradicts to the necessary condition

a1 = −(a + k0 ) > 0

for the asymptotic stability of the trivial solution of the corresponding linear system.
Thus, by condition (8.68) the trivial solution of the corresponding linear system is
unstable.
228 8 Stabilization of the Controlled Inverted Pendulum by a Control with Delay

8.2.3 Stable, Unstable, and One-Sided Stable Points of Equilibrium

Here a stability of the steady-state solutions of system (8.64) is investigated. Let x̂

be a root of (8.66). Put
x1 = x̂ + y1 , x2 = y2 . (8.69)
Substituting (8.69) into (8.64) and using (8.65), we obtain

ẏ1 (t) = y2 (t),

 ∞
   
ẏ2 (t) = a sin x̂ + y1 (t) − sin x̂ + dK(τ ) y1 (t − τ ),
0

or, after elementary trigonometric transformations,

ẏ1 (t) = y2 (t),

     ∞ (8.70)
y1 (t) y1 (t)
ẏ2 (t) = 2a cos x̂ + sin + dK(τ ) y1 (t − τ ).
2 2 0

It is easy to see that the linear approximation of system (8.70) has the form

ẏ1 (t) = y2 (t),

 ∞ (8.71)
ẏ2 (t) = a cos x̂y1 (t) + dK(τ ) y1 (t − τ ).
0

Conditions (8.21) for system (8.71) have the form

a2 = −(a cos x̂ + k0 ) > 0, k1 > 0,

4 (8.72)
k 2 < km =  .
1+ 1 + 4a2 k1−2

So, if for some root x̂ of (8.66), conditions (8.72) hold, then the point x̂ is a stable
equilibrium of (8.64).

Theorem 8.4 Let x̂ be a positive root of (8.66). If x̂ is a point of stable equilibrium

of system (8.64), then
Ṡ(x̂) < 0, (8.73)
i.e., x̂ is a point of decrease of the characteristic function S(x).
If x̂ is a point of increase of the characteristic function S(x), i.e.,

Ṡ(x̂) > 0, (8.74)

then x̂ is a point of unstable equilibrium of system (8.64).

8.3 Numerical Analysis of the Controlled Inverted Pendulum 229

Proof Using the derivative of the function (8.67), i.e.,

 
1 sin x
Ṡ(x) = cos x −
x x

and (8.66), (8.67), (8.72), we obtain (8.73):

 
1 sin x̂ a cos x̂ + k0 a2
Ṡ(x̂) = cos x̂ − = =− < 0.
x̂ x̂ x̂a x̂a

Let condition (8.74) hold. Then the first inequality of conditions (8.72) does not
hold, and x̂ cannot (see Remark 8.1) be a point of stable equilibrium. The proof is
completed. 

Remark 8.8 Let x̂ be a point of extremum of the characteristic function S(x). In this
case, Ṡ(x̂) = 0, and x̂ is a point of one-sided stable equilibrium of system (8.64).
This means that if the system stays in a point x from a small enough neighborhood
of x̂ and Ṡ(x) < 0, then the solution converges to x̂. But if the system stays in a
point x from a small enough neighborhood of x̂ and Ṡ(x) > 0, then the solution
goes away from x̂.

Remark 8.9 Since the function S(x) is an even function, for negative roots of (8.66),
the pictures are symmetrical.

8.3 Numerical Analysis of the Controlled Inverted Pendulum

Here we consider the results of numerical simulation of solutions of the lin-

ear and nonlinear mathematical models of the controlled inverted pendulum. For
simplicity, the initial function from (8.3) is considered in the form φ(s) = x(0),
s ≤ 0. For greater visualization, the solutions are shown in the spaces (x1 , x2 ) and
(t, x(t)).
Note that the stability conditions for the difference analogue of the mathematical
model of the controlled inverted pendulum were obtained in [273, 278].

8.3.1 Stability of the Trivial Solution and Limit Cycles

Consider now the linear model of the inverted pendulum in the form (8.24). Suffi-
cient conditions for asymptotic stability of the trivial solution of (8.24) are (8.25).
Recall that in accordance with Remark 8.1 two first inequalities of (8.25) are also
necessary conditions for asymptotic stability, while the third inequality is a sufficient
condition only.
230 8 Stabilization of the Controlled Inverted Pendulum by a Control with Delay

Fig. 8.4 Solution of (8.24)

for x(0) = 6, a = 1, b1 = 1,
b2 = −4, h1 = 0.25,
h2 = 0.04

Put x(0) = 6, a = 1, b1 = 1, b2 = −4, h1 = 0.25, h2 = 0.04. In this case k0 =

−3, k1 = 0.09, k2 = 0.0689, km = 0.123. So, all conditions (8.25) hold, and the
trivial solution of (8.24) is asymptotically stable. The appropriate solution of (8.24)
is shown in Fig. 8.4.
Put x(0) = 8, a = 3, b1 = 1, b2 = −4.5, h1 = 0.6, h2 = 0.1. Then k0 = −3.5,
k1 = 0.15, k2 = 0.405, km = 0.382. So, two first conditions (8.25) hold, and the third
condition (8.25) does not hold. But the trivial solution of the linear equation (8.24)
is asymptotically stable (see Fig. 8.5), and the trivial solution of the corresponding
nonlinear equation

ẍ(t) − a sin x(t) = b1 x(t − h1 ) + b2 x(t − h2 ) (8.75)

is asymptotically stable too (see Fig. 8.6).

8.3 Numerical Analysis of the Controlled Inverted Pendulum 231

Fig. 8.5 Solution of (8.24) for x(0) = 8, a = 3, b1 = 1, b2 = −4.5, h1 = 0.6, h2 = 0.1

Fig. 8.6 Solution of (8.75)

for x(0) = 8, a = 3, b1 = 1,
b2 = −4.5, h1 = 0.6,
h2 = 0.1
232 8 Stabilization of the Controlled Inverted Pendulum by a Control with Delay

Fig. 8.7 Solution of (8.75)

for x(0) = 8, a = 3, b1 = 1,
b2 = −4.5, h1 = 0.59,
h2 = 0.1

Let us investigate the influence of the delay h1 on the behavior of the inverted
pendulum. All other parameters have the same values as in Fig. 8.5. If h1 = 0.59,
then the trivial solution of the linear system (8.24) is asymptotically stable, but the
nonlinear system (8.75) has a limit cycle (Fig. 8.7). The same phenomenon takes
place for all values of the delay h1 from h1 = 0.59 to h1 = 0.54 (Fig. 8.8). If
h1 = 0.53, then the limit cycle of the nonlinear system (8.75) is unstable (Fig. 8.9,
the solution goes to infinity), but the trivial solution of the linear system (8.24) is
asymptotically stable (Fig. 8.10, the solution goes to zero).
Consider the nonlinear model (8.75) for the following values of the parameters:
a = 10, b1 = 1, b2 = −2, h1 = 0.8, h2 = 0.36. In this case the necessary condi-
tion for asymptotic stability a + b1 + b2 < 0 does not hold. So, the trivial solution
of (8.75) is not stable. But for different initial conditions, there are different limit
cycles. In Fig. 8.11 twelve different limit cycles are shown, which correspond to
8.3 Numerical Analysis of the Controlled Inverted Pendulum 233

Fig. 8.8 Solution of (8.75)

for x(0) = 8, a = 3, b1 = 1,
b2 = −4.5, h1 = 0.54,
h2 = 0.1

different initial conditions: x(0) = 2, x(0) = 12, x(0) = 18, x(0) = 26, x(0) = 31,
x(0) = 38, x(0) = 46, x(0) = 52, x(0) = 57, x(0) = 66, x(0) = 71, x(0) = 77. Four
from these limit cycles are shown separately in Fig. 8.12 by close-up.
Let us note that the considered limit cycles are stable. For instance, in Figs. 8.13
and 8.14 the first cycle from Fig. 8.12 is shown respectively for x(0) = 2 and
x(0) = 10. In the first case the cycle is reached from the inside, and in the second
case the cycle is reached from the outside. In both these cases the solution converges
to the same limit cycle, which shows its asymptotic stability.
Note also that by small changing of the value of the parameter h2 the asymptotic
stability of all these limit cycles disappears. Changing h2 = 0.36 to h2 = 0.37, we
can see (Fig. 8.15, x(0) = 2) that the trajectory of the solution goes to infinity. In
Fig. 8.16 the same picture is shown for x(0) = 0.1.
234 8 Stabilization of the Controlled Inverted Pendulum by a Control with Delay

Fig. 8.9 Solution of (8.75)

for x(0) = 8, a = 3, b1 = 1,
b2 = −4.5, h1 = 0.53,
h2 = 0.1

Put now a = 3, b1 = 1, b2 = −4.5, h1 = 0.59, h2 = 0.1. In this case k0 = −3.5,

k1 = 0.14, k2 = 0.3931, km = 0.3587, two first conditions (8.25) hold, but the third
condition (8.25) does not hold. For x(0) = 4.6, the solution of (8.75) goes to zero
(Fig. 8.17), but for x(0) = 4.61, the solution of (8.75) goes to a stable limit cycle. In
Fig. 8.18 this cycle is reached from the inside, and in Fig. 8.19 it is reached from the
outside for x(0) = 100. In Fig. 8.20 the appropriate solution of the linear equation
(8.24) is shown, which goes to zero from the initial point x(0) = 100.

8.3.2 Nonzero Steady-State Solutions of the Nonlinear Model

Put in (8.75) a = 1, b1 = 1, b2 = −1.08, h1 = 0.8, h2 = 0.3. In this case, k0 =

−0.08, k1 = 0.476, k2 = 0.7372, and the first condition (8.25) does not hold. So,
8.3 Numerical Analysis of the Controlled Inverted Pendulum 235

Fig. 8.10 Solution of (8.24) for x(0) = 8, a = 3, b1 = 1, b2 = −4.5, h1 = 0.53, h2 = 0.1

Fig. 8.11 Twelve limit cycles of (8.75) for a = 10, b1 = 1, b2 = −2, h1 = 0.8, h2 = 0.36, and
different initial conditions: x(0) = 2, x(0) = 12, x(0) = 18, x(0) = 26, x(0) = 31, x(0) = 38,
x(0) = 46, x(0) = 52, x(0) = 57, x(0) = 66, x(0) = 71, x(0) = 77

the trivial solution of (8.75) is unstable. On the other hand, the equation (8.66) has
three positive roots x̂1 , x̂2 , x̂3 such that 2.906892 < x̂1 < 2.906893, 6.864548 <
x̂2 < 6.864549, 8.659471 < x̂3 < 8.659472. Therefore (Remark 8.4), these points
are steady-state solutions of (8.75).
236 8 Stabilization of the Controlled Inverted Pendulum by a Control with Delay

Fig. 8.12 Four limit cycles of (8.75) from Fig. 8.11 by close-up for different initial conditions:
x(0) = 2, x(0) = 12, x(0) = 18, x(0) = 26

Fig. 8.13 The first cycle

of (8.75) from Fig. 8.12 by
x(0) = 2
8.3 Numerical Analysis of the Controlled Inverted Pendulum 237

Fig. 8.14 The first cycle

of (8.75) from Fig. 8.12 by
x(0) = 10

It is easy to check that

Ṡ(x̂1 ) < 0, Ṡ(x̂2 ) > 0, Ṡ(x̂3 ) < 0.

By Theorem 8.4 the points x̂1 and x̂3 can be points of stable equilibrium of (8.75),
and the point x̂2 is a point of unstable equilibrium of (8.75).
Note also that for the points x̂1 and x̂3 , all conditions (8.72) hold, but for the
point x̂2 , the first condition (8.72) does not hold.
Let x(0) = 6.864548, i.e., the initial function is close enough to x̂2 and less than
x̂2 . In this case the solution of (8.75) goes away from the point of unstable equilib-
rium x̂2 and converges to the point of stable equilibrium x̂1 . This situation is shown
in Fig. 8.21.
Let x(0) = 6.864549, i.e., the initial function is close enough to x̂2 and greater
than x̂2 . In this case the solution of (8.75) goes away from the point of unstable
equilibrium x̂2 and converges to the point of stable equilibrium x̂3 . This situation is
shown in Fig. 8.22.
Put now a = 1, b1 = 1, b2 = −0.782766, h1 = 0.8, h2 = 0.3. In this case
k0 = 0.217233, k1 = 0.565170, k2 = 0.710449, and (8.66) has one positive root
238 8 Stabilization of the Controlled Inverted Pendulum by a Control with Delay

Fig. 8.15 Disappearance of

limit cycles after changing
h2 = 0.36 to h2 = 0.37,
x(0) = 2

x̂ = 4.493409 only. This point is also a root of the equation Ṡ(x̂) = 0, i.e. (see Re-
mark 8.8), it is a point of one-sided stable equilibrium of (8.75).
Let x(0) = 4, i.e., the initial function is close enough to x̂ and less than x̂. In this
case, Ṡ(4) = −0.116111 < 0, and the solution of (8.75) converges to the point x̂.
This situation is shown in Fig. 8.23.
Let x(0) = 4.5, i.e., the initial function is close enough to x̂ and greater than x̂.
In this case, Ṡ(4.5) = 0.00142958 > 0, and the solution of (8.75) goes away from
the point x̂ and goes to infinity. This situation is shown in Fig. 8.24.
Note that if the initial function is less than the point of one-sided stable equilib-
rium x̂ and stays far enough from this point, then the system converges to infinity
past by the point x̂. This situation is shown in Fig. 8.25 for x(0) = 3.7.
Put a = 7, b1 = 1, b2 = −2, h1 = 0.4, h2 = 0.02. In this case, a + k0 = 6 > 0,
k1 = 0.476, k2 = 0.7372. So, the first condition (8.25) does not hold, and the trivial
solution of (8.75) is unstable. But (8.66) has the positive root x̂1 = 2.739489 and the
symmetric negative root x̂2 = −2.739489. For these roots, a2 = 7.441678 > 0, so,
two first conditions (8.72) hold, and Ṡ(x̂1 ) < 0, but the third condition (8.72) does
8.3 Numerical Analysis of the Controlled Inverted Pendulum 239

Fig. 8.16 Disappearance of

limit cycles after changing
h2 = 0.36 to h2 = 0.37,
x(0) = 0.1

not hold since k2 > km = 0.3199. In Fig. 8.26 the solution with the initial condition
x(0) = 19 converges to x̂2 , and in Fig. 8.27 the solution with the initial condition
x(0) = 12 converges to x̂1 .
Changing in the previous situation a = 7 on a = 27, we obtain nine posi-
tive points of equilibrium: x̂1 = 3.029, x̂2 = 6.527, x̂3 = 9.082, x̂4 = 13.072,
x̂5 = 15.114, x̂6 = 19.665, x̂7 = 21.094, x̂8 = 26.513, x̂9 = 26.819. It is easy to
check that Ṡ(x̂i ) > 0 for i = 2, 4, 6, 8 and Ṡ(x̂i ) < 0 for i = 1, 3, 5, 7, 9. So, by The-
orem 8.4 the points x̂i for i = 1, 3, 5, 7, 9 only can be points of stable equilibrium.
But for all these points, two first conditions (8.72) hold, and the third condition
(8.72) does not hold. So, these points can be points of unstable equilibrium. Indeed,
in Figs. 8.28 and 8.29 we can see that the solution of (8.75) has limit cycles around
the points x̂1 and −x̂1 .
Note that in the presence of an ample quantity of points of equilibrium we
can obtain different interesting limit cycles; see, for instance, Figs. 8.30, 8.31,
and 8.32.
240 8 Stabilization of the Controlled Inverted Pendulum by a Control with Delay

Fig. 8.17 Solution of (8.75)

for x(0) = 4.60, a = 3,
b1 = 1, b2 = −4.5,
h1 = 0.59, h2 = 0.10

8.3.3 Stabilization of the Controlled Inverted Pendulum Under

Influence of Markovian Stochastic Perturbations

Consider the linear model of the controlled pendulum in the form

ẍ(t) + η(t)x(t) = u(t), (8.76)

where η(t) is a Markov chain with two states {a1 , a2 } such that a1 > 0 and a2 < 0,
the initial distribution
 
pi = P η(0) = ai , i = 1, 2, (8.77)

and the probabilities of transition

 
pij ( ) = P η(t + ) = aj /η(t) = ai = λij + o( ),
i, j = 1, 2, i = j. (8.78)
8.3 Numerical Analysis of the Controlled Inverted Pendulum 241

Fig. 8.18 Solution of (8.75)

for x(0) = 4.61, a = 3,
b1 = 1, b2 = −4.5,
h1 = 0.59, h2 = 0.10

It is easy to see that by the control u(t) ≡ 0 the trivial solution of (8.76) is sta-
ble but not asymptotically stable in the case where η(t) ≡ a1 > 0 (mathematical
pendulum) and unstable in the case where η(t) ≡ a2 < 0 (inverted mathematical
pendulum).
Consider the problem of stabilization of the inverted mathematical pendulum by
the control
u(t) = b1 x(t − h1 ) + b2 x(t − h2 ). (8.79)
By (8.25), if

k0 = b1 + b2 < a2 , k1 = b1 h1 + b2 h2 > 0,

4 (8.80)
k2 = |b1 |h21 + |b2 |h22 <  ,
1 + 1 + 4(a2 − k0 )k1−2

then the trivial solution of (8.76), (8.79) (with η(t) ≡ a2 < 0) is asymptotically
stable. It was shown also (Example 8.2) that for each a2 < 0, there exist numbers
242 8 Stabilization of the Controlled Inverted Pendulum by a Control with Delay

Fig. 8.19 Solution of (8.75)

for x(0) = 100, a = 3,
b1 = 1, b2 = −4.5,
h1 = 0.59, h2 = 0.10

Fig. 8.20 Solution of (8.24) for x(0) = 100, a = 3, b1 = 1, b2 = −4.5, h1 = 0.59, h2 = 0.10
8.3 Numerical Analysis of the Controlled Inverted Pendulum 243

Fig. 8.21 The solution of (8.75) goes away from the left neighborhood of the point of unstable
 
equilibrium x2 and converges to the point of stable equilibrium x1 , x(0) = 6.864548

Fig. 8.22 The solution of (8.75) goes away from the right neighborhood the point of unstable
 
equilibrium x2 and converges to the point of stable equilibrium x3 , x(0) = 6.864549

 
Fig. 8.23 The initial function of (8.75) is close enough to x and is less than x, the solution con-

verges to the point x, x(0) = 4
244 8 Stabilization of the Controlled Inverted Pendulum by a Control with Delay

 
Fig. 8.24 The initial function of (8.75) is close enough to x and is greater than x, the solution
goes to infinity, x(0) = 4.5

 
Fig. 8.25 The initial function of (8.75) is far enough from x and is less than x, the solution goes
to infinity, x(0) = 3.7

Fig. 8.26 Solution of (8.75) for x(0) = 19, a = 7, b1 = 1, b2 = −2, h1 = 0.4, h2 = 0.02
8.3 Numerical Analysis of the Controlled Inverted Pendulum 245

Fig. 8.27 Solution of (8.75) for x(0) = 12, a = 7, b1 = 1, b2 = −2, h1 = 0.4, h2 = 0.02

Fig. 8.28 Solution of (8.75) for x(0) = 19, a = 27, b1 = 1, b2 = −2, h1 = 0.4, h2 = 0.02

b1 , b2 , h1 , h2 such that conditions (8.80) hold, and therefore the trivial solution
of (8.76), (8.79) (with η(t) ≡ a2 < 0) is asymptotically stable.
Let us investigate the stability of the trivial solution of (8.76), (8.79) by the nu-
merical method and numerical simulation of the Markov chain η(t) that was de-
scribed in Sect. 7.4.1.
246 8 Stabilization of the Controlled Inverted Pendulum by a Control with Delay

Fig. 8.29 Solution of (8.75) for x(0) = 12, a = 27, b1 = 1, b2 = −2, h1 = 0.4, h2 = 0.02

Fig. 8.30 Solution of (8.75) for x(0) = 13, a = 123, b1 = 1, b2 = −2, h1 = 0.4, h2 = 0.02

Put a1 = 1, a2 = −1, b1 = 1, b2 = −2.1, h1 = 0.8, h2 = 0.3, x(s) = 3.5, s ≤ 0.

If η(t) ≡ a1 , then the solution of (8.76), (8.79) goes to ±∞. The trajectories of η(t)
and x(t) in this case are shown in Fig. 8.33. If η(t) ≡ a2 , then conditions (8.80)
hold, and the solution of (8.76), (8.79) converges to zero (Fig. 8.34).
8.3 Numerical Analysis of the Controlled Inverted Pendulum 247

Fig. 8.31 Solution of (8.75) for x(0) = 13, a = 200, b1 = 1, b2 = −1.1, h1 = 0.4, h2 = 0.02

Fig. 8.32 Solution of (8.75) for x(0) = 72, a = 200, b1 = 1, b2 = −2, h1 = 0.4, h2 = 0.02

Put in (8.77), (8.78) p1 = p2 = 0.5, λ21 = 15, λ12 = 1. One of possible trajecto-
ries of the Markov chain η(t) and the corresponding trajectory of the solution x(t)
of (8.76), (8.79) are shown in Fig. 8.35. Hundred trajectories of the solution x(t) are
248 8 Stabilization of the Controlled Inverted Pendulum by a Control with Delay

Fig. 8.33 Solution of (8.76), (8.79) by η(t) ≡ a1 = 1, b1 = 1, b2 = −2.1, h1 = 0.8, h2 = 0.3,

x(s) = 3.5, s ≤ 0

Fig. 8.34 Solution of (8.76), (8.79) by η(t) ≡ a2 = −1, b1 = 1, b2 = −2.1, h1 = 0.8, h2 = 0.3,
x(s) = 3.5, s ≤ 0

shown in Fig. 8.36. One can see that in this case the trivial solution of (8.76), (8.79)
is unstable.
Put now λ21 = 1, λ12 = 15. In this case the trivial solution of (8.76), (8.79) is
asymptotically stable. One of the possible trajectories of the Markov chain η(t)
and the corresponding trajectory of the solution x(t) of (8.76), (8.79) are shown in
8.3 Numerical Analysis of the Controlled Inverted Pendulum 249

Fig. 8.35 Solution of (8.76), (8.79) by a1 = 1, a2 = −1, p1 = p2 = 0.5, λ21 = 15, λ12 = 1,
b1 = 1, b2 = −2.1, h1 = 0.8, h2 = 0.3, x(s) = 3.5, s ≤ 0

Fig. 8.36 Hundred trajectories of solution of (8.76), (8.79) by the values of the parameters as in
Fig. 8.35

Fig. 8.37. Hundred trajectories of the solution x(t) are shown in Fig. 8.38. All these
trajectories converge to zero.
250 8 Stabilization of the Controlled Inverted Pendulum by a Control with Delay

Fig. 8.37 Solution of (8.76), (8.79) by a1 = 1, a2 = −1, p1 = p2 = 0.5, λ21 = 1, λ12 = 15,
b1 = 1, b2 = −2.1, h1 = 0.8, h2 = 0.3, x(s) = 3.5, s ≤ 0

Fig. 8.38 Hundred trajectories of solution of (8.76), (8.79) by the values of the parameters as in
Fig. 8.37
Chapter 9
Stability of Equilibrium Points of Nicholson’s
Blowflies Equation with Stochastic
Perturbations

We consider the Nicholson blowflies equation (one of the most known models in
ecology) with stochastic perturbations. We obtain sufficient conditions for stability
in probability of the trivial and positive equilibrium points of this nonlinear differ-
ential equation with delay.

9.1 Introduction
Consider the nonlinear differential equation with exponential nonlinearity

ẋ(t) = ax(t − h)e−bx(t−h) − cx(t). (9.1)

It describes a population dynamics of the well-known Nicholson blowflies [224].

Here x(t) is the size of the population at time t , a is the maximum per capita daily
egg production rate, 1/b is the size at which the population reproduces at the maxi-
mum rate, c is the per capita daily adult death rate, and h is the generation time.
Nicholson’s blowflies model is popular enough with researchers [7, 8, 30,
31, 38, 40, 61, 73, 99, 102, 129, 166, 173–176, 178, 184, 250–252, 278, 284–
287, 297, 307, 312, 316, 320]. The majority of the results on (9.1) deal with the
global attractiveness of the positive point of equilibrium and oscillatory behaviors
of solutions [73, 102, 129, 166, 173, 174, 178, 250–252, 278, 285].
Below we will obtain sufficient conditions for stability in probability of the trivial
and positive equilibrium points of (9.1) by stochastic perturbations. The basic stages
of the proposed research are the following. It is assumed that the considered nonlin-
ear differential equation has an equilibrium point and exposed to white-noise-type
stochastic perturbations that are proportional to the deviation of the system current
state from the considered equilibrium point. In this case the equilibrium point is a
solution of the stochastic differential equation too. The constructed stochastic differ-
ential equation is centered around the considered equilibrium point and linearized in
the neighborhood of this equilibrium point. Necessary and sufficient conditions for
the asymptotic mean-square stability of the linear part of the considered equation are

L. Shaikhet, Lyapunov Functionals and Stability of Stochastic Functional 251

Differential Equations, DOI 10.1007/978-3-319-00101-2_9,
© Springer International Publishing Switzerland 2013
252 9 Stability of Equilibrium Points of Nicholson’s Blowflies Equation

obtained. Since the order of nonlinearity in (9.1) is higher than one (see Sect. 5.3),
these conditions are sufficient for stability in probability of the equilibrium point of
the initial nonlinear equation by stochastic perturbations.

9.2 Two Points of Equilibrium, Stochastic Perturbations,

Centering, and Linearization

The points of equilibrium of (9.1) are defined by the condition ẋ(t) = 0 that can be
represented in the form
∗
ae−bx x ∗ = cx ∗ . (9.2)

From (9.2) it follows that (9.1) has two points of equilibrium

1 a
x1∗ = 0, x2∗ = ln . (9.3)
b c
Similarly to Sect. 5.4, let us assume that (9.1) is exposed to stochastic perturba-
tions that are of white noise type and are directly proportional to the deviation of
x(t) from the point of equilibrium x ∗ and influence ẋ(t) immediately. In this way,
(9.1) takes the form
 
ẋ(t) = ax(t − h)e−bx(t−h) − cx(t) + σ x(t) − x ∗ ẇ(t). (9.4)

Let us center (9.4) at the point of equilibrium x ∗ using the new variable y(t) =
x(t) − x ∗ . By this way from (9.4) via (9.2) we obtain
∗  
ẏ(t) = −cy(t) + ae−bx y(t − h)e−by(t−h) + x ∗ e−by(t−h) − 1 + σy(t)ẇ(t).
(9.5)
It is clear that the stability of an equilibrium point x ∗ of (9.4) is equivalent to the
stability of the trivial solution of (9.5).
Along with (9.5), we will consider the linear part of this equation. Using the
representation ey = 1 + y + o(y) (where o(y) means that limy→0 o(y) y = 0) and
neglecting o(y), we obtain the linear part (process z(t)) of (9.5) in the form
∗ 
ż(t) = −cz(t) − ae−bx bx ∗ − 1 z(t − h) + σ z(t)ẇ(t). (9.6)

As it follows from Remark 5.3, if the order of nonlinearity of the equation un-
der consideration is higher than one, then a sufficient condition for the asymptotic
mean-square stability of the linear part of the initial nonlinear equation is also a
sufficient condition for the stability in probability of the initial equation. So, we
will investigate sufficient conditions for the asymptotic mean-square stability of the
linear part (9.6) of the nonlinear equation (9.5).
9.3 Sufficient Conditions for Stability in Probability for Both Equilibrium 253

Fig. 9.1 Regions of stability in probability for zero equilibrium point (yellow) and positive equi-
librium point (red) of (9.4) for h = 0.02, p = 20

9.3 Sufficient Conditions for Stability in Probability for Both

Equilibrium Points

By (9.2)–(9.3) the nonlinear and linear equations (9.5)–(9.6) for the equilibrium
points x1∗ = 0 respectively are

ẏ(t) = −cy(t) + ay(t − h)e−by(t−h) + σy(t)ẇ(t), (9.7)

ż(t) = −cz(t) + az(t − h) + σ z(t)ẇ(t). (9.8)

By Lemma 2.1 a necessary and sufficient condition for the asymptotic mean-
square stability of the trivial solution of (9.8) is

1
G−1 > p, p = σ 2, (9.9)
2
where
1 − aq −1 sinh(qh) 
G= , c > a, q = c2 − a 2 . (9.10)
c − a cosh(qh)
254 9 Stability of Equilibrium Points of Nicholson’s Blowflies Equation

Fig. 9.2 Regions of stability in probability for zero equilibrium point (yellow) and positive equi-
librium point (red) of (9.4) for h = 0.1, p = 20

In particular, if p ≥ 0 and h = 0, then the stability condition (9.9)–(9.10) takes the

form c > a + p.
Condition (9.9)–(9.10) is also a sufficient condition for the stability in probability
of the zero equilibrium point x ∗ = 0 of (9.4).
Similarly to (9.9)–(9.10), for the equilibrium points x2∗ = b1 ln ac , the nonlinear
and linear equations (9.5)–(9.6) respectively are
c a  −by(t−h) 
ẏ(t) = −cy(t) + cy(t − h)e−by(t−h) + ln e − 1 + σy(t)ẇ(t), (9.11)
b c
 
a
ż(t) = −cz(t) − c ln − 1 z(t − h) + σ z(t)ẇ(t). (9.12)
c
By Lemma 2.1 a necessary and sufficient condition for the asymptotic mean-
square stability of the trivial solution of (9.12) is (9.9), where
⎧ 1+cq −1 (ln(ac−1 )−1) sinh(qh)
⎪
⎪ , c < a < ce 2 , q = c ln a (2 − ln a ),
⎪ −1
⎨ c[1+(ln(ac )−1) cosh(qh)] c c

G= 1+ch
2c , a = ce ,
2
(9.13)
⎪
⎪
⎪
⎩ 1+cq −1 (ln(ac−1 )−1) sin(qh) , a > ce2 , q = c ln a (ln a − 2).
c[1+(ln(ac−1 )−1) cos(qh)] c c
9.4 Numerical Illustrations 255

Fig. 9.3 Zero equilibrium point is stable at the point A = (110, 200) (25 red trajectories) and is
unstable in the point B = (900, 200) (25 green trajectories)

In particular, if p > 0 and h = 0, then the stability condition takes the form c ln ac >
p; if p = 0 and h > 0, then the region of stability is bounded by the lines c = 0,
c = a, and 1 + (ln ac − 1) cos(qh) = 0 for a > ce2 .
Condition (9.9), (9.13) is also a sufficient condition for the stability in probability
of the positive equilibrium point x ∗ = b1 ln ac of (9.4).

Remark 9.1 Note that the stability conditions (9.9), (9.10) and (9.9), (9.13) have
the following property: if the point (a, c) belongs to the stability region with some
p and h, then for arbitrary positive α, the point (a0 , c0 ) = (αa, αc) belongs to the
stability region with p0 = αp and h0 = α −1 h.

9.4 Numerical Illustrations

In Fig. 9.1 the stability regions for (9.4) given by conditions (9.9), (9.10) for the
zero equilibrium point (yellow) and (9.9), (9.13) for the positive equilibrium point
(red) are shown in the space of the parameters (a, c) for h = 0.02 and p = 20. In
Fig. 9.2 the similar regions of stability are shown for h = 0.1 and p = 20.
256 9 Stability of Equilibrium Points of Nicholson’s Blowflies Equation

Fig. 9.4 Positive equilibrium point is unstable a the point A = (110, 200) (25 red trajectories) and
is stable at the point B = (900, 200) (25 green trajectories)

For numerical simulation of the solution of (9.4), one uses the algorithm of nu-
merical simulation of trajectories of the Wiener process (Chap. 2) and the Euler–
Maruyama scheme [200]. Note that the stability of the difference analogue of (9.4)
was investigated in detail in [38, 278].
Numerical simulation of the solution of (9.4) with x ∗ = 0 is shown in Fig. 9.3.
At the point A with coordinates a = 110, c = 200 (see Fig. 9.2) the zero equilib-
rium point is stable in probability, so, all 25 trajectories (red) of the solution with
the initial function x(s) = 1.35 cos(3s) converge to zero. At the point B with coor-
dinates a = 900, c = 200 (see Fig. 9.2) the zero equilibrium point is unstable, so, 25
trajectories (green) of the solution with the initial function x(s) = 2.35 cos(3s) fill
the whole space.
In Fig. 9.4 numerical simulation of the solution of (9.4) with the positive equilib-
rium point x ∗ = b1 ln ac is shown by b = 1. At the point B with coordinates a = 900,
c = 200 (see Fig. 9.2) the positive equilibrium point is stable in probability, so, all
25 trajectories (green) of the solution converge to x ∗ = ln(900/200) = 1.504. At the
point A with coordinates a = 110, c = 200 (see Fig. 9.2) the positive equilibrium
point is unstable, and the trajectories (red) of the solution do not go to zero.
Chapter 10
Stability of Positive Equilibrium Point of
Nonlinear System of Type of Predator–Prey
with Aftereffect and Stochastic Perturbations

Here we consider a system of two nonlinear differential equations that is destined

to unify different known mathematical models, in particular, very often investigated
models of predator–prey type [47, 53, 60, 65, 72, 82, 83, 94, 107, 108, 112, 113,
127, 128, 153, 180, 235, 249, 267, 283, 288, 304, 305, 311, 314, 317, 321, 325]. The
system under consideration is exposed to stochastic perturbations and is linearized
in a neighborhood of the positive point of equilibrium. Asymptotic mean-square
stability conditions for the trivial solution of the constructed linear system are at
the same time sufficient conditions for the stability in probability of the positive
equilibrium point of the initial nonlinear system by stochastic perturbations.

10.1 System Under Consideration

Consider the system of two nonlinear differential equations

 
ẋ1 (t) = x1 (t) a − F0 (x1t , x2t ) − F1 (x1t , x2t ),
 
ẋ2 (t) = −x2 (t) b + G0 (x1t , x2t ) + G1 (x1t , x2t ), (10.1)

xi (s) = φi (s), s ≤ 0, i = 1, 2.

Here xi (t), i = 1, 2, is the value of the process xi at time t , and xit = xi (t + s),
s ≤ 0, is a trajectory of the process xi to the point of time t .
Put, for example,
 ∞  
F0 (x1t , x2t ) = f0 x1 (t − s) dK0 (s),
0
2 
 ∞  
F1 (x1t , x2t ) = fi xi (t − s) dKi (s), (10.2)
i=1 0

L. Shaikhet, Lyapunov Functionals and Stability of Stochastic Functional 257

Differential Equations, DOI 10.1007/978-3-319-00101-2_10,
© Springer International Publishing Switzerland 2013
258 10 Stability of Positive Equilibrium Point of Nonlinear System of Type
 ∞  
G0 (x1t , x2t ) = g0 x1 (t − s) dR0 (s),
0
2 
 ∞  
G1 (x1t , x2t ) = gi xi (t − s) dRi (s),
i=1 0

where Ki (s) and Ri (s), i = 0, 1, 2, are nondecreasing functions such that

 ∞  ∞
Ki = dKi (s) < ∞, Ri = dRi (s) < ∞,
0 0
 ∞  ∞
(10.3)
K̂i = s dKi (s) < ∞, R̂i = s dRi (s) < ∞,
0 0

and all integrals are understood in the Stieltjes sense.

In the case (10.2)–(10.3) system (10.1) takes the form

  ∞  2  ∞
   
ẋ1 (t) =x1 (t) a − f0 x1 (t − s) dK0 (s) − fi xi (t − s) dKi (s),
0 i=1 0
(10.4)
  ∞  2  ∞
   
ẋ2 (t) = − x2 (t) b + g0 x1 (t − s) dR0 (s) + gi xi (t − s) dRi (s).
0 i=1 0

Systems of type (10.4) are investigated in some biological problems. Put here,
for example,

f0 (x) = f1 (x) = f2 (x) = g1 (x) = g2 (x) = x,

(10.5)
g0 (x) = 0, dK1 (s) = δ(s) ds, dR0 (s) = 0

(δ(s) is Dirac’s function). If a and b are positive constants, x1 (t) and x2 (t) are re-
spectively the densities of prey and predator populations, then (10.4) is transformed
to the mathematical predator–prey model [267] with distributed delay
  ∞  ∞ 
ẋ1 (t) = x1 (t) a − x1 (t − s) dK0 (s) − x2 (t − s) dK2 (s) ,
0 0
 ∞  ∞
(10.6)
ẋ2 (t) = −bx2 (t) + x1 (t − s) dR1 (s) x2 (t − s) dR2 (s).
0 0

Putting in (10.6)

dK0 (s) = a1 δ(s) ds, dK2 (s) = a2 δ(s) ds,

(10.7)
dR1 (s) = b1 δ(s − h1 ) ds, dR2 (s) = δ(s − h2 ) ds,
10.1 System Under Consideration 259

we obtain the known predator–prey mathematical model with fixed delays

 
ẋ1 (t) = x1 (t) a − a1 x1 (t) − a2 x2 (t) ,
(10.8)
ẋ2 (t) = −bx2 (t) + b1 x1 (t − h1 )x2 (t − h2 ).
If here h1 = h2 = 0, we have the classical Lotka–Volterra model
 
ẋ1 (t) = x1 (t) a − a1 x1 (t) − a2 x2 (t) ,
 
ẋ2 (t) = x2 (t) −b + b1 x1 (t) .
Many authors [15, 19, 23, 50, 69, 70, 116, 306, 309] consider the so-called ratio-
dependent predator–prey models with delays of type
  ∞ 
ẋ1 (t) = x1 (t) a − x1 (t − s) dK0 (s)
0
 ∞ x1k (t − s)x2 (t)
− dK1 (s), (10.9)
x1k (t − s) + a2 x2k (t − s)
0
 ∞
x1m (t − s)x2 (t)
ẋ2 (t) = −bx2 (t) + dR1 (s).
0 x1m (t − s) + b2 x2m (t − s)
Here it is supposed that m and k are positive constants.
System (10.9) follows from (10.1) if
 ∞
F0 (x1t , x2t ) = x1 (t − s) dK0 (s), G0 (x1t , x2t ) = 0,
0
 ∞  
F1 (x1t , x2t ) = f x1 (t − s), x2 (t − s) x2 (t) dK1 (s),
0
 ∞ (10.10)
 
G1 (x1t , x2t ) = g x1 (t − s), x2 (t − s) x2 (t) dR1 (s),
0

x1k x1m
f (x1 , x2 ) = , g(x1 , x2 ) = .
x1k + a2 x2 k x1m + b2 x2m
Putting in (10.9), for example,
dK0 (s) = a0 δ(s) ds, dK1 (s) = a1 δ(s) ds,
(10.11)
dR1 (s) = b1 δ(s − h) ds, k = m = 1,
we obtain the system
 
a1 x2 (t)
ẋ1 (t) = x1 (t) a − a0 x1 (t) − ,
x1 (t) + a2 x2 (t)
  (10.12)
b1 x1 (t − h)
ẋ2 (t) = x2 (t) −b + ,
x1 (t − h) + b2 x2 (t − h)
which was considered in [23, 50].
260 10 Stability of Positive Equilibrium Point of Nonlinear System of Type

10.2 Equilibrium Points, Stochastic Perturbations, Centering,

and Linearization

10.2.1 Equilibrium Points

Let in system (10.1) Fi = Fi (φ, ψ) and Gi = Gi (φ, ψ), i = 0, 1, be functionals

defined on H × H , where H is a set of functions φ = φ(s), s ≤ 0, with the norm
φ = sups≤0 |φ(s)|, the functionals Fi and Gi are nonnegative for nonnegative
functions φ and ψ . Let us suppose also that system (10.1) has a positive equilibrium
point (x1∗ , x2∗ ). This point is obtained from the conditions ẋ1 (t) ≡ 0, ẋ2 (t) ≡ 0 and
is defined by the system of algebraic equations
    
x1∗ a − F0 x1∗ , x2∗ = F1 x1∗ , x2∗ ,
     (10.13)
x2∗ b + G0 x1∗ , x2∗ = G1 x1∗ , x2∗ .

From (10.13) it follows that system (10.1) has a positive solution by the condition
 
a > F0 x1∗ , x2∗ (10.14)

only. For example, if a > K0 f0 (x1∗ ), a positive equilibrium point of system (10.4) is
defined by the system of algebraic equations
      
x1∗ a − K0 f0 x1∗ = K1 K2 f1 x1∗ f2 x2∗ ,
       (10.15)
x2∗ b + R0 g0 x1∗ = R1 R2 g1 x1∗ g2 x2∗ .

In particular, from (10.5), (10.14), (10.15) it follows that system (10.6) has a positive
equilibrium point

b a − K0 x1∗ a − (R1 R2 )−1 K0 b

x1∗ = , x2∗ = = , (10.16)
R1 R2 K2 K2

provided that a > (R1 R2 )−1 K0 b. For system (10.8), from (10.7), (10.16) we obtain
b A a1
x1∗ = , x2∗ = , A=a−b > 0. (10.17)
b1 a2 b1
From (10.13), (10.10) it follows that the positive equilibrium point for sys-
tem (10.9) is
 1
A A K1 bb2 m
x1∗ = , x2∗ = , A=a− > 0, B = > 0.
K0 BK0 B + a2 B 1−k R1 − b
In particular, by (10.11), for system (10.12), it is
A A a1 bb2
x1∗ = , x2∗ = , A=a− > 0, B = > 0. (10.18)
a0 Ba0 B + a2 b1 − b
10.2 Equilibrium Points, Stochastic Perturbations, Centering 261

10.2.2 Stochastic Perturbations and Centering

Similarly to Sect. 9.2, we will assume that system (10.1) is exposed to stochastic
perturbations that are of white noise type and are directly proportional to the de-
viations of the system state (x1 (t), x2 (t)) from the equilibrium point (x1∗ , x2∗ ) and
influence ẋ1 (t), ẋ2 (t), respectively. In this way system (10.1) is transformed to the
form
   
ẋ1 (t) = x1 (t) a − F0 (x1t , x2t ) − F1 (x1t , x2t ) + σ1 x1 (t) − x1∗ ẇ1 (t),
    (10.19)
ẋ2 (t) = −x2 (t) b + G0 (x1t , x2t ) + G1 (x1t , x2t ) + σ2 x2 (t) − x2∗ ẇ2 (t).
Here σ1 , σ2 are constants, and w1 (t), w2 (t) are independent standard Wiener pro-
cesses.
Centering system (10.19) at the positive point of equilibrium via the new vari-
ables y1 = x1 − x1∗ , y2 = x2 − x2∗ , we obtain
   
ẏ1 (t) = y1 (t) + x1∗ a − F0 y1t + x1∗ , y2t + x2∗
 
− F1 y1t + x1∗ , y2t + x2∗ + σ1 y1 (t)ẇ1 (t),
    (10.20)
ẏ2 (t) = − y2 (t) + x2∗ b + G0 y1t + x1∗ , y2t + x2∗
 
+ G1 y1t + x1∗ , y2t + x2∗ + σ2 y2 (t)ẇ2 (t).
It is clear that the stability of equilibrium point (x1∗ , x2∗ ) of system (10.19) is equiv-
alent to the stability of the trivial solution of system (10.20).
For system (10.4), the representations (10.19) and (10.20) respectively take the
forms
  ∞ 
 
ẋ1 (t) = x1 (t) a − f0 x1 (t − s) dK0 (s)
0
2 
 ∞    
− fi xi (t − s) dKi (s) + σ1 x1 (t) − x1∗ ẇ1 (t),
i=1 0
  ∞  (10.21)
 
ẋ2 (t) = −x2 (t) b + g0 x1 (t − s) dR0 (s)
0
2  ∞
    
+ gi xi (t − s) dRi (s) + σ2 x2 (t) − x2∗ ẇ2 (t)
i=1 0

and
  ∞ 
   
ẏ1 (t) = y1 (t) + x1∗ a− f0 y1 (t − s) + x1∗ dK0 (s)
0
2 
 ∞  
− fi yi (t − s) + xi∗ dKi (s)
i=1 0
262 10 Stability of Positive Equilibrium Point of Nonlinear System of Type

+ σ1 y1 (t)ẇ1 (t), (10.22)

  ∞ 
 ∗
  ∗

ẏ2 (t) = − y2 (t) + x2 b + g0 x1 (t − s) + x1 dR0 (s)
0
2  ∞
  
+ gi yi (t − s) + xi∗ dRi (s) + σ2 y2 (t)ẇ2 (t).
i=1 0

In particular, for system (10.6), from (10.21), (10.22) by (10.5), (10.16) we obtain
  ∞  ∞ 
ẋ1 (t) = x1 (t) a − x1 (t − s) dK0 (s) − x2 (t − s) dK2 (s)
0 0
 
+ σ1 x1 (t) − x1∗ ẇ1 (t),
 ∞  ∞
(10.23)
ẋ2 (t) = −bx2 (t) + x1 (t − s) dR1 (s) x2 (t − s) dR2 (s)
0 0
 
+ σ2 x2 (t) − x2∗ ẇ2 (t)

and
 ∞  ∞ 
 ∗

ẏ1 (t) = − y1 (t) + x1 y1 (t − s) dK0 (s) + y2 (t − s) dK2 (s)
0 0

+ σ1 y1 (t)ẇ1 (t),
 ∞
ẏ2 (t) = − by2 (t) + R2 x2∗ y1 (t − s) dR1 (s)
0 (10.24)
 ∞
+ R1 x1∗ y2 (t − s) dR2 (s)
0
2 
 ∞
+ yi (t − s) dRi (s) + σ2 y2 (t)ẇ2 (t).
i=1 0

For (10.8), systems (10.23) and (10.24) take respectively the forms
   
ẋ1 (t) = x1 (t) a − a1 x1 (t) − a2 x2 (t) + σ1 x1 (t) − x1∗ ẇ1 (t),
  (10.25)
ẋ2 (t) = −bx2 (t) + b1 x1 (t − h1 )x2 (t − h2 ) + σ2 x1 (t) − x2∗ ẇ2 (t)

and
  
ẏ1 (t) = − y1 (t) + x1∗ a1 y1 (t) + a2 y2 (t) + σ1 y1 (t)ẇ1 (t),
 
ẏ2 (t) = − by2 (t) + b1 x2∗ y1 (t − h1 ) + x1∗ y2 (t − h2 ) (10.26)
+ b1 y1 (t − h1 )y2 (t − h2 ) + σ2 y2 (t)ẇ2 (t).
10.2 Equilibrium Points, Stochastic Perturbations, Centering 263

10.2.3 Linearization

Along with the considered nonlinear system, we will use the linear part of this sys-
tem. Let us suppose that the functionals in (10.19) have the representations (10.2)
with differentiable functions fi (x), gi (x), i = 0, 1, 2. Using for all these functions
the representation
    df  ∗ 
f z + x ∗ = f0 + f1 z + o(z), f 0 = f x ∗ , f1 = x ,
dx
and neglecting o(z), we obtain the linear part (process (z1 (t), z2 (t))) of system
(10.22)
 ∞
ż1 (t) = (a − K0 f00 )z1 (t) − z1 (t − s) dK(s)
0
 ∞
− K1 f10 f21 z2 (t − s) dK2 (s) + σ1 z1 (t)ẇ1 (t),
0
 ∞
(10.27)
ż2 (t) = −(b + R0 g00 )z2 (t) + z1 (t − s) dR(s)
0
 ∞
+ R1 g10 g21 z2 (t − s) dR2 (s) + σ2 z2 (t)ẇ2 (t),
0

where
dK(s) = K2 f20 f11 dK1 (s) + f01 x1∗ dK0 (s),
(10.28)
dR(s) = R2 g20 g11 dR1 (s) − g01 x2∗ dR0 (s).
Below we will speak about system (10.27) as about the linear part corresponding
to system (10.22) or, for brevity, as about the linear part of system (10.22).
In particular, by conditions (10.5), (10.16), and (10.28) from (10.27) we obtain
the linear part of system (10.24)
 ∞  ∞ 
ż1 (t) = − x1∗ z1 (t − s) dK0 (s) + z2 (t − s) dK2 (s) + σ1 z1 (t)ẇ1 (t),
0 0
 ∞  ∞
ż2 (t) = − bz2 (t) + R2 x2∗ z1 (t − s) dR1 (s) + R1 x1∗ z2 (t − s) dR2 (s)
0 0

+ σ2 z2 (t)ẇ2 (t).
(10.29)
From (10.26) or, via (10.7), from (10.29) we have the linear part of system (10.26)
 
ż1 (t) = −x1∗ a1 z1 (t) + a2 z2 (t) + σ1 z1 (t)ẇ1 (t),
  (10.30)
ż2 (t) = −bz2 (t) + b1 x2∗ z1 (t − h1 ) + x1∗ z2 (t − h2 ) + σ2 z2 (t)ẇ2 (t).

As it is shown in Sect. 5.3, if the order of nonlinearity of the system under consid-
eration is higher than one, then a sufficient condition for the asymptotic mean-square
264 10 Stability of Positive Equilibrium Point of Nonlinear System of Type

stability of the linear part of the considered nonlinear system is also a sufficient con-
dition for the stability in probability of the initial system. So, below we will obtain
sufficient conditions for the asymptotic mean-square stability of the linear part of
considered nonlinear systems.

10.3 Stability of Equilibrium Point

Obtain now sufficient conditions for the asymptotic mean-square stability of the
trivial solution of system (10.27) as the linear part of (10.22). The obtained condi-
tions will be at the same time sufficient conditions for the stability in probability of
the equilibrium point of (10.21).
Following the procedure of constructing Lyapunov functionals (Sect. 2.2.2),
rewrite (10.27) in the form

Ż1 (t) = a11 z1 (t) + a12 z2 (t) + σ1 z1 (t)ẇ1 (t),

(10.31)
Ż2 (t) = a21 z1 (t) + a22 z2 (t) + σ2 z2 (t)ẇ2 (t),

where
 ∞ t  ∞ t
Z1 (t) = z1 (t) − z1 (θ ) dθ dK(s) − K1 f10 f21 z2 (θ ) dθ dK2 (s),
0 t−s 0 t−s
 ∞ t  ∞ t
Z2 (t) = z2 (t) + z1 (θ ) dθ dR(s) + R1 g10 g21 z2 (θ ) dθ dR2 (s),
0 t−s 0 t−s
(10.32)
and, by (10.15), (10.28),
 
f10 ∗
a11 = a − K − K0 f00 = K1 K2 f20 − f11 − K0 f01 x1 ,
x1∗
a12 = −K1 K2 f10 f21 , a21 = R = R1 R2 g20 g11 − R0 g01 x2∗ , (10.33)
 
g20
a22 = R1 R2 g10 g21 − b − R0 g00 = −R1 R2 g10 − g21 .
x2∗

System (10.31), (10.32) is a system of stochastic differential equations of neutral

type, so, following (2.10), we have to suppose that
 ∞  ∞
s dK(s) + K1 |f10 f21 | s dK2 (s) < 1,
0 0
 ∞  ∞
s dR(s) + R1 |g10 g21 | s dR2 (s) < 1,
0 0
10.3 Stability of Equilibrium Point 265

or, by (10.3), (10.28), that

|f01 |x1∗ K̂0 + K2 |f20 f11 |K̂1 + K1 |f10 f21 |K̂2 < 1,
(10.34)
|g01 |x2∗ R̂0 + R2 |g20 g11 |R̂1 + R1 |g10 g21 |R̂2 < 1.

10.3.1 First Way of Constructing a Lyapunov Functional

Let  = aij  be the matrix with the elements defined by (10.33), and P = pij 
be the matrix with the elements defined by (1.29) for some q > 0. Represent p11 ,
p22 in the form
1  (0) (1) 
pii = qpii + pii , i = 1, 2, (10.35)
2
where
2 + det(Â)
a22 2
a21
(0) (1)
p11 = , p11 = ,
| Tr(Â)| det(Â) | Tr(Â)| det(Â)
(10.36)
2
a12 2 + det(Â)
a11
(0) (1)
p22 = , p22 = ,
| Tr(Â)| det(Â) | Tr(Â)| det(Â)
and put
(0) (1)
dμij (s) = qdμij (s) + dμij (s), i, j = 1, 2, (10.37)

where
(0) a12 (1) a21
dμ11 = dK(s) − dR(s), dμ11 = dR(s),
| Tr(Â)| | Tr(Â)|
(0) a12
dμ12 = K1 f10 f21 dK2 (s) − R1 g10 g21 dR2 (s),
| Tr(Â)|
a21
dμ(1)
12 = R1 g10 g21 dR2 (s),
| Tr(Â)|
(10.38)
a12 a21
dμ(0)
21 = − dK(s),
(1)
dμ21 = dK(s) − dR(s),
| Tr(Â)| | Tr(Â)|
(0) a12
dμ22 =− K1 f10 f21 dK2 (s),
| Tr(Â)|
(1) a21
dμ22 = K1 f10 f21 dK2 (s) − R1 g10 g21 dR2 (s),
| Tr(Â)|

and dK(s), dR(s) are defined by (10.28).

266 10 Stability of Positive Equilibrium Point of Nonlinear System of Type

Put also
1
δi = σi2 , i = 1, 2,
2
 ∞ (10.39)
 
(m)
νij = s dμ(m) 
ij (s) , i, j = 1, 2, m = 0, 1,
0

and
(0) (0) (1) (1)
A1 = 1 − ν11 − p11 δ1 , A2 = 1 − ν22 − p22 δ2 ,
(1) (1) (0) (0)
B1 = ν11 + p11 δ1 , B2 = ν22 + p22 δ2 , (10.40)
(1) (1) (0) (0)
C1 = ν12 + ν21 , C2 = ν12 + ν21 .

Theorem 10.1 If A1 > 0, A2 > 0, and conditions (10.34) and

(A1 C1 + B1 C2 )(A2 C2 + B2 C1 ) + B1 B2 < A1 A2 (10.41)

hold, then the trivial solution of system (10.27) is asymptotically mean-square stable
and the equilibrium point of system (10.21) is stable in probability.

Proof We will consider now system (10.31)–(10.33) and suppose that the trivial so-
lution of the appropriate auxiliary system without delays of type (2.60) with aij ,
i, j = 1, 2, defined by (10.33) is asymptotically mean-square stable, and so condi-
tions (2.62) hold.
Consider the functional

V1 (t) = p11 Z12 (t) + 2p12 Z1 (t)Z2 (t) + p22 Z22 (t) (10.42)

with pij , i, j = 1, 2, defined by (1.29). Let L be the generator of system (10.31).

Then, by (10.31), (10.42),
  
LV1 (t) = 2 p11 Z1 (t) + p12 Z2 (t) a11 z1 (t) + a12 z2 (t) + p11 σ12 z12 (t)
  
+ 2 p12 Z1 (t) + p22 Z2 (t) a21 z1 (t) + a22 z2 (t) + p22 σ22 z22 (t)
= 2(p11 a11 + p12 a21 )Z1 (t)z1 (t) + 2(p12 a11 + p22 a21 )Z2 (t)z1 (t)
+ 2(p11 a12 + p12 a22 )Z1 (t)z2 (t) + p11 σ12 z12 (t)
+ 2(p12 a12 + p22 a22 )Z2 (t)z2 (t) + p22 σ22 z22 (t). (10.43)

Putting
a21 − a12 q
ρ= (10.44)
| Tr(Â)|
10.3 Stability of Equilibrium Point 267

and using (1.29), (2.62), we obtain

2(p11 a11 + p12 a21 ) = −q, 2(p12 a12 + p22 a22 ) = −1,

−(a12 a22 q + a21 a11 )a11 + (a11

2 + det(Â) + a 2 q)a
21
2(p12 a11 + p22 a21 ) = 12
| Tr(Â)| det(Â)
det(Â)a21 − (a11 a22 − a12 a21 )a12 q
= = ρ,
| Tr(Â)| det(Â)
2 + det(Â))q + a 2 )a − (a a q + a a )a
((a22 21 12 12 22 21 11 22
2(p11 a12 + p12 a22 ) =
| Tr(Â)| det(Â)
det(A)a12 q − a21 (a11 a22 − a21 a12 )
= = −ρ.
| Tr(Â)| det(Â)

So, (10.43) takes the form

LV1 (t) = −qZ1 (t)z1 (t) + ρZ2 (t)z1 (t) + p11 σ12 z12 (t)
− ρZ1 (t)z2 (t) − Z2 (t)z2 (t) + p22 σ22 z22 (t). (10.45)

Substituting (10.32) into (10.45), we have

   
LV1 = −q + p11 σ12 z12 (t) + −1 + p22 σ22 z22 (t)
 ∞ t
+q z1 (t)z1 (θ ) dθ dK(s)
0 t−s
 ∞ t
+ qK1 f10 f21 z1 (t)z2 (θ ) dθ dK2 (s)
0 t−s
 ∞ t
+ρ z1 (t)z1 (θ ) dθ dR(s)
0 t−s
 ∞ t
+ ρR1 g10 g21 z1 (t)z2 (θ ) dθ dR2 (s)
0 t−s
 ∞ t
+ρ z2 (t)z1 (θ ) dθ dK(s)
0 t−s
 ∞ t
+ ρK1 f10 f21 z2 (t)z2 (θ ) dθ dK2 (s)
0 t−s
 ∞ t
− z2 (t)z1 (θ ) dθ dR(s)
0 t−s
 ∞ t
− R1 g10 g21 z2 (t)z2 (θ ) dθ dR2 (s).
0 t−s
268 10 Stability of Positive Equilibrium Point of Nonlinear System of Type

By (10.37), (10.38), (10.44), it can be written in the form

   
LV1 = −q + p11 σ12 z12 (t) + −1 + p22 σ22 z22 (t)
 ∞ t  ∞ t
+ z1 (t)z1 (θ ) dθ dμ11 (s) + z1 (t)z2 (θ ) dθ dμ12 (s)
0 t−s 0 t−s
 ∞ t  ∞ t
+ z2 (t)z1 (θ ) dθ dμ21 (s) + z2 (t)z2 (θ ) dθ dμ22 (s).
0 t−s 0 t−s
(10.46)

Using (10.35), (10.37), (10.39) and some positive number γ from (10.46), we
obtain
 (0) (1)   (0) (1) 
LV1 ≤ −q + qp11 δ1 + p11 δ1 z12 (t) + −1 + qp22 δ2 + p22 δ2 z22 (t)
 
1 ∞ t  2    (0)   (1) 
+ z1 (t) + z12 (θ ) dθ q dμ11 (s) + dμ11 (s)
2 0 t−s
 ∞ t
1  −1 2    (0)   (1) 
+ γ z1 (t) + γ z22 (θ ) dθ q dμ12 (s) + dμ12 (s)
2 0 t−s
 ∞ t
1  2      (1) 
+ γ z2 (t) + γ −1 z12 (θ ) dθ q dμ(0)  
21 (s) + dμ21 (s)

2 0 t−s
 
1 ∞ t  2    (0)   (1) 
+ z2 (t) + z22 (θ ) dθ q dμ22 (s) + dμ22 (s) .
2 0 t−s

From this by (10.39) we have

 (0) (1)   (0) (1) 
LV1 ≤ −q + qp11 δ1 + p11 δ1 z12 (t) + −1 + qp22 δ2 + p22 δ2 z22 (t)
 
1  (0) (1)  1 ∞ t 2   (0)   (1) 
+ qν11 + ν11 z12 (t) + z1 (θ ) dθ q dμ11 (s) + dμ11 (s)
2 2 0 t−s

γ −1  (0) (1) 
+ qν12 + ν12 z12 (t)
2
 
γ ∞ t 2   (0)   (1) 
+ z2 (θ ) dθ q dμ12 (s) + dμ12 (s)
2 0 t−s
γ  (0) (1)  2
+ qν21 + ν21 z2 (t)
2
 
γ −1 ∞ t 2   (0)   (1) 
+ z1 (θ ) dθ q dμ21 (s) + dμ21 (s)
2 0 t−s
 
1  (0) (1)  2 1 ∞ t 2   (0)   (1) 
+ qν22 + ν22 z2 (t) + z2 (θ ) dθ q dμ22 (s) + dμ22 (s)
2 2 0 t−s
10.3 Stability of Equilibrium Point 269

 
1  (0) (0) 
= q −1 + ν11 + γ −1 ν12 + p11(0)
δ1
2

1  (1) −1 (1)
 (1)
+ ν11 + γ ν12 + p11 δ1 z12 (t)
2

 
1  (1) (1)  (1) 1  (0) (0)  (0)
+ −1 + γ ν21 + ν22 + p22 δ2 + q γ ν21 + ν22 + p22 δ2 z22 (t)
2 2
2  ∞ t
+ zi2 (θ ) dθ dFi (s), (10.47)
i=1 0 t−s

where
1 (0) (1) 
dFi (s) = qdFi (s) + dFi (s) , i = 1, 2,
2
 (0)   (0) 
dF1 (s) = dμ11 (s) + γ −1 dμ21 (s),

(0)

 (1)   (1) 
dF1 (s) = dμ11 (s) + γ −1 dμ21 (s),
(1)

   (0) 
dF2 (s) = γ dμ12 (s) + dμ22 (s),
(0) (0)

   (1) 
dF2 (s) = γ dμ12 (s) + dμ22 (s).
(1) (1)

Note that for the functional

2 
 ∞ t
V2 (t) = (θ − t + s)zi2 (θ ) dθ dFi (s),
i=1 0 t−s

we have
2 
 ∞ t
LV2 (t) = F̂1 z12 (t) + F̂2 z22 (t) − zi2 (θ ) dθ dFi (s), (10.48)
i=1 0 t−s

where
1  (0) (0)  (1) 
q ν11 + γ −1 ν21 + ν11 + γ −1 ν21 ,
(1)
F̂1 =
2
1  (0) (0)  (1) (1) 
F̂2 = q γ ν12 + ν22 + γ ν12 + ν22 .
2
From (10.47), (10.48), for the functional V = V1 + V2 , by (10.40) we obtain

  
γ −1 γ −1
LV (t) ≤ q −A1 + C2 + B1 + C1 z12 (t)
2 2

 
γ γ
+ −A2 + C1 + q B2 + C2 z22 (t). (10.49)
2 2
270 10 Stability of Positive Equilibrium Point of Nonlinear System of Type

By Theorem 2.1, if there exist positive numbers q and γ such that

 
γ −1 γ −1
q −A1 + C2 + B1 + C1 < 0,
2 2
  (10.50)
γ γ
−A2 + C1 + q B2 + C2 < 0,
2 2

then the trivial solution of system (10.27) is asymptotically mean-square stable.

Rewrite (10.50) in the form
  −1   −1
γ −1 γ −1 γ γ
B1 + C1 A1 − C2 < q < A2 − C1 B2 + C2 . (10.51)
2 2 2 2

So, if
  −1   −1
γ −1 γ −1 γ γ
B1 + C1 A1 − C2 < A2 − C1 B2 + C2 , (10.52)
2 2 2 2

then there exists q > 0 such that (10.51) holds.

Rewriting (10.52) in the form

γ γ −1
(A1 C1 + B1 C2 ) + (A2 C2 + B2 C1 ) < A1 A2 − B1 B2
2 2
and calculating the infimum of the left-hand part of the obtained inequality with
respect to γ > 0, we obtain (10.41). So, if (10.41) holds, then there exist positive
numbers q and γ such that (10.50) holds, and therefore the trivial solution of system
(10.27) is asymptotically mean-square stable. The proof is completed. 

Put now
a1 δ1 a 1 δ2
D1 = − Ah1 − , D2 = 1 − bh2 − , (10.53)
b1 b Ab1
and note that the first condition (10.34) for system (10.30) is a trivial one and the
second condition takes the form Aa2−1 b1 h1 + bh2 < 1 or, via the representation
(10.17) for A,
b1 h1 a + (a2 h2 − a1 h1 )b < a2 . (10.54)

Corollary 10.1 If D1 > 0, D2 > 0, and conditions (10.54) and

δ2
A(D1 h1 + h2 )(δ2 h1 + D2 bh2 ) + < D1 D2 (10.55)
b
hold, then the trivial solution of system (10.30) is asymptotically mean-square sta-
ble, and the equilibrium point of system (10.25) is stable in probability.
10.3 Stability of Equilibrium Point 271

Proof Calculating for (10.30) the parameters (10.33), (10.36), (10.38), (10.39),
(10.40), we obtain

a1 b a2 b Ab1
a11 = − , a12 = − , a21 = , a22 = 0,
b1 b1 a2
(0) b1 (1) Ab13 (0) a22 (1) a1 b1
p11 = , p11 = , p22 = , p22 = + ,
a1 b a1 a22 b2 Aa1 b1 Ab1 a1 b

(0) Ab1 h1 (1) A2 b13 h1 (0) a2 bh2 (1) Ab12 h2

ν11 = , ν11 = , ν12 = , ν12 = ,
a1 a1 a22 b a1 a1 a2
(0) (1) Ab1 h1 (0) (1)
ν21 = 0, ν21 = , ν22 = 0, ν22 = bh2 ,
a2
b1 b 1 δ2
A1 = D1 , A2 = D 2 − ,
a1 a1 b
 
Ab13 δ1 a22 δ2
B1 = Ah 1 + , B 2 = ,
a1 a22 b b Aa1 b1
Ab1 a2 bh2
C1 = (a1 h1 + b1 h2 ), C2 = .
a1 a2 a1

From this it follows that

Ab12
A1 C1 + B1 C2 = (D1 h1 + h2 ),
a1 a2
a2
A2 C2 + B2 C1 = (δ2 h1 + D2 bh2 ), (10.56)
a1
 
b1 δ2
A1 A2 − B1 B2 = D1 D 2 − .
a1 b

From the representations for aij , i, j = 1, 2, it follows also that conditions (2.62)
hold. Substituting (10.56) into (10.41), we obtain (10.55). The proof is completed. 

Remark 10.1 Note that condition (10.55) does not depend on a2 . The dependence
on a2 is included in condition (10.54).

Remark 10.2 By the absence of the delays, i.e., by h1 = h2 = 0, condition (10.54)

is trivial, and condition (10.55) can be written in the form

a1 Ab1 (a1 b − b1 δ1 )
δ1 < b , δ2 < .
b1 Ab12 + a1 (a1 b − b1 δ1 )

The same conditions can be obtained immediately from Corollary 2.3.

272 10 Stability of Positive Equilibrium Point of Nonlinear System of Type

10.3.2 Second Way of Constructing a Lyapunov Functional

Let us consider another way of constructing of a Lyapunov functional for system

(10.30).

Theorem 10.2 If D1 > 0, D2 > 0, and conditions (10.54) and

 √  √ 
Abh22 + 4δ2 b−1 D1 + Abh2 Abh21 + 4D2 + Abh1 < 4D1 D2 (10.57)

hold, where A and D1 , D2 are defined by (10.17) and (10.53), respectively, then
the trivial solution of system (10.30) is asymptotically mean-square stable, and the
equilibrium point of system (10.25) is stable in probability.

Proof Using (10.17), rewrite system (10.30) in the form

a1 b a2 b
ż1 (t) = − z1 (t) − z2 (t) + σ1 z1 (t)ẇ1 (t),
b1 b1
(10.58)
Ab1
Ż2 (t) = z1 (t) + σ2 z2 (t)ẇ2 (t),
a2
where
Ab1
Z2 (t) =z2 (t) + J1 (z1t ) + bJ2 (z2t ),
a2
 t (10.59)
Ji (zit ) = zi (s) ds, i = 1, 2.
t−hi
Consider now the functional

V1 (t) = z12 (t) + 2μz1 (t)Z2 (t) + γ Z22 (t), (10.60)

where the parameters μ and γ will be chosen below. Then by (10.60), (10.58) we
have
b  
LV1 (t) = −2 z1 (t) + μZ2 (t) a1 z1 (t) + a2 z2 (t) + σ12 z12 (t)
b1
Ab1  
+2 μz1 (t) + γ Z2 (t) z1 (t) + γ σ22 z22 (t)
a2
   
a1 b Ab1 a2 b
= −2 −μ − δ1 z12 (t) − 2 μ − γ δ2 z22 (t)
b1 a2 b1
 
Ab1 a 1 b a2 b
+2 γ −μ − z1 (t)z2 (t)
a2 b1 b1
   
Ab1 a1 b Ab1
+2 γ −μ z1 (t) J1 (z1t ) + bJ2 (z2t )
a2 b1 a2
10.3 Stability of Equilibrium Point 273
 
a2 b Ab1
− 2μ z2 (t) J1 (z1t ) + bJ2 (z2t ) . (10.61)
b1 a2

Defining now γ by the equality

Ab1 a1 b a 2 b
γ =μ + , (10.62)
a2 b1 b1

from (10.61) we obtain

   
a1 b Ab1 a2 b
LV1 (t) = − 2 −μ − δ1 z1 (t) − 2 μ
2
− γ δ2 z22 (t)
b1 a2 b1
 
a2 b Ab1
+2 z1 (t) J1 (z1t ) + bJ2 (z2t )
b1 a2
 
a2 b Ab1
− 2μ z2 (t) J1 (z1t ) + bJ2 (z2t ) .
b1 a2

By (10.59) from this, for some positive γ1 , γ2 , we have

   
a1 b Ab1 a2 b
LV1 (t) ≤ −2 −μ − δ1 z12 (t) − 2 μ − γ δ2 z22 (t)
b1 a2 b1
  t  t 
a2 b Ab1  2   2 
+ z1 (t) + z12 (s) ds + b γ1 z1 (t) + γ1−1 z22 (s) ds
b1 a2 t−h1 t−h2
  t
a2 b Ab1  2 
+μ γ2 z2 (t) + γ2−1 z12 (s) ds
b1 a2 t−h1
 t 
 2 
+b z2 (t) + z22 (s) ds . (10.63)
t−h2

By the representations (10.53) for D1 , D2 and (10.62) for γ inequality (10.63) can
be written in the form
 
Ab1 a2 b2 h2 2
LV1 (t) ≤ −2bD1 − Abh1 + 2μ + γ1 z1 (t)
a2 b1
 
a2 bD2 μa2 b2 h2 2a22 bδ2
+ −2μ − + + γ 2 μAbh 2
1 z2 (t)
b1 b1 Ab12
 t 
 −1  b2 a2  −1  t
+ Ab 1 + μγ2 z1 (s) ds +
2
γ1 + μ z22 (s) ds.
t−h1 b1 t−h2

Put now

  t
V2 = Ab 1 + μγ2−1 (s − t + h1 )z12 (s) ds
t−h1
274 10 Stability of Positive Equilibrium Point of Nonlinear System of Type

b2 a2  −1  t
+ γ1 + μ (s − t + h2 )z22 (s) ds.
b1 t−h2

Then
  
  t
LV2 = Ab 1 + μγ2−1 h1 z1 (t) −
2
z12 (s) ds
t−h1
  
b2 a 2  −1  t
+ γ1 + μ h2 z2 (t) −
2
z22 (s) ds ,
b1 t−h2

and as a result, for the functional V = V1 + V2 , we obtain

 
Ab1 a 2 b 2 h2
LV ≤ −2bD1 + 2μ + γ1 + γ2−1 μAbh1 z12 (t)
a2 b1
 
a2 b a2b a2 b2 h2
+ −2μ D2 + 2 2 2 δ2 + γ1−1 + γ2 μAbh1 z22 (t).
b1 Ab1 b1
By Theorem 2.1, if
Ab1 a2 b2 h2
− 2bD1 + 2μ + γ1 + γ2−1 μAbh1 < 0,
a2 b1
(10.64)
a2 b a2b a 2 b 2 h2
− 2μ D2 + 2 2 2 δ2 + γ1−1 + γ2 μAbh1 < 0,
b1 Ab1 b1
then the trivial solution of system (10.30) is asymptotically mean-square stable.
Rewrite (10.64) in the form
a22 b
δ + γ1−1 a2 bb1 h2
2 2
2 2bD1 − γ1 a2 bb1 h2
Ab12 2
<μ< −1
. (10.65)
2 ab21b D2 − γ2 Abh1 2 Ab
a2 + γ2 Abh1
1

So, if the inequality

a22 b
δ2 + γ1−1 a2 bb1 h2
2 2
2 2 2bD1 − γ1 a2 bb1 h2
Ab1
< −1
(10.66)
2 ab21b D2 − γ2 Abh1 2 Ab
a2 + γ2 Abh1
1

holds, then there exists μ such that (10.65) holds too.

It is easy to check that from (10.65), (10.62) the condition μ2 < γ follows, which
ensures the positivity of the functional (10.60).
Representing (10.66) in the form
a22 Ab12
δ + γ1−1 a22bbh1 2 + γ2−1 Ab2a12h1
Ab12 2 a22 b
× <1
D1 − γ1 a22bbh1 2 D2 − γ2 Ab2a12h1
and using Lemma 2.4 twice, we obtain (10.57) 
10.3 Stability of Equilibrium Point 275

Fig. 10.1 Region of stability in probability for (10.25): a1 = 0.6, b1 = 1, h1 = 0, h2 = 0, δ1 = 0,

δ2 = 0

Remark 10.3 Note that the representation (10.31)–(10.33) for system (10.30) coin-
cides with (10.58), (10.59). So, conditions (10.55) and (10.57) are equivalent and
give the same stability region. For simplicity, let us check this statement by the
condition h1 = 0. Indeed, in this case from (10.57) we have
√
Abh22 + 4δ2 b−1 D1 < (2D1 D2 − Abh2 )2 = 4D12 D2 − 4D1 h2 AbD2 + Abh22
√
or δ2 b−1 < D1 D2 − h2 AbD2 , which is equivalent to (10.55) by h1 = 0. Similarly,
it is easy to get that (10.55) coincides with (10.57) by the condition h2 = 0 or by the
condition δ2 = 0. In the general case the necessary transformation is bulky enough.
The regions of stability in probability for a positive point of equilibrium of system
(10.25), obtained by condition (10.55) (or (10.57)), are shown in the space of the
parameters (a, b) for a1 = 0.6, b1 = 1 and different values of the other parameters:
in Fig. 10.1 for h1 = 0, h2 = 0, δ1 = 0, δ2 = 0, in Fig. 10.2 for h1 = 0, h2 = 0,
δ1 = 0.2, δ2 = 0.3, in Fig. 10.3 for a2 = 0.6, h1 = 0.1, h2 = 0.15, δ1 = 0, δ2 = 0,
and in Fig. 10.4 for a2 = 0.07, h1 = 0.01, h2 = 0.15, δ1 = 0.05, δ2 = 0.1.
The equation of the straight line in Figs. 10.1 and 10.2 is ab1 = ba1 , which
corresponds to the condition A = 0. In Figs. 10.3 and 10.4 the straight line 1
also corresponds to this equation and the straight line 2 is defined by the equation
b1 h1 a + (a2 h2 − a1 h1 )b = a2 , which follows from condition (10.54).
Note that the stability of the positive equilibrium point of the difference analogue
of system (10.25) is investigated in [278].
276 10 Stability of Positive Equilibrium Point of Nonlinear System of Type

Fig. 10.2 Region of stability in probability for (10.25): a1 = 0.6, b1 = 1, h1 = 0, h2 = 0, δ1 = 0.2,

δ2 = 0.3

Fig. 10.3 Region of stability in probability for (10.25): a1 = 0.6, a2 = 0.6, b1 = 1, h1 = 0.1,
h2 = 0.15, δ1 = 0, δ2 = 0
10.3 Stability of Equilibrium Point 277

Fig. 10.4 Region of stability in probability for (10.25): a1 = 0.6, a2 = 0.07, b1 = 1, h1 = 0.01,
h2 = 0.15, δ1 = 0.05, δ2 = 0.1

10.3.3 Stability of the Equilibrium Point of Ratio-Dependent

Predator–Prey Model

Consider now system (10.9) with stochastic perturbations, i.e.,

  ∞   ∞
x1k (t − s)x2 (t)
ẋ1 (t) = x1 (t) a − x1 (t − s) dK0 (s) − dK1 (s)
0 0 x1k (t − s) + a2 x2k (t − s)
 
+ σ1 x1 (t) − x1∗ ẇ1 (t), (10.67)
 ∞
x1 (t − s)x2 (t)
m  
ẋ2 (t) = −bx2 (t) + dR1 (s) + σ2 x1 (t) − x2∗ ẇ2 (t).
0 x1 (t − s) + b2 x2 (t − s)
m m

System (10.9) was obtained from (10.1) by conditions (10.10). So, by (10.13),
(10.14) the positive equilibrium point (x1∗ , x2∗ ) of system (10.9) (and also (10.67)) is
defined by the conditions
   
x1∗ a − K0 x1∗ = K1 f x1∗ , x2∗ x2∗ ,
  (10.68)
b = R1 g x1∗ , x2∗ , a > K0 x1∗ .

Suppose that the functions f (x1 , x2 ) and g(x1 , x2 ) in (10.10) are differentiable and
can be represented in the form
 
f y1 + x1∗ , y2 + x2∗ = f0 + f1 y1 − f2 y2 + o(y1 , y2 ),
 
g y1 + x1∗ , y2 + x2∗ = g0 + g1 y1 − g2 y2 + o(y1 , y2 ),
278 10 Stability of Positive Equilibrium Point of Nonlinear System of Type

o(y1 ,y2 )
where lim|y|→0 |y| = 0 for |y| = y12 + y22 , and

  ka2 (x1∗ x2∗ )k−1

f0 = f x1∗ , x2∗ , f1 = x2∗ fˆ, f2 = x1∗ fˆ, fˆ = ,
((x1∗ )k + a2 (x2∗ )k )2
  mb2 (x1∗ x2∗ )m−1
g0 = g x1∗ , x2∗ , g1 = x2∗ ĝ, g2 = x1∗ ĝ, ĝ = .
((x1∗ )m + b2 (x2∗ )m )2
So, the functionals F0 (x1t , x2t ), F1 (x1t , x2t ), G1 (x1t , x2t ) in (10.10) have the repre-
sentations
 ∞
 
F0 y1t + x1∗ , y2t + x2∗ = K0 x1∗ + y1 (t − s) dK0 (s),
0
 ∞
 
F1 y1t + x1∗ , y2t + x2∗ = K1 f0 x2∗ + f1 x2∗ y1 (t − s) dK1 (s) + K1 f0 y2 (t)
0
 ∞
− f2 x2∗ y2 (t − s) dK1 (s) + o(y1 , y2 ),
0
 ∞
 
G1 y1t + x1∗ , y2t + x2∗ = R1 g0 x2∗ + g1 x2∗ y1 (t − s) dR1 (s) + R1 g0 y2 (t)
0
 ∞
− g2 x2∗ y2 (t − s) dR1 (s) + o(y1 , y2 ).
0
(10.69)
By (10.68), (10.69) the linear part of system (10.67) has the form
 ∞
 ∗

ż1 (t) = a − K0 x1 z1 (t) − K1 f0 z2 (t) − z1 (t − s) dK(s)
0
 ∞
+ f2 x2∗ z2 (t − s) dK1 (s) + σ1 z1 (t)ẇ1 (t), (10.70)
0
 ∞  ∞
ż2 (t) = g1 x2∗ z1 (t − s) dR1 (s) − g2 x2∗ z2 (t − s) dR1 (s) + σ2 z2 (t)ẇ2 (t),
0 0

where dK(s) = x1∗ dK0 (s) + f1 x2∗ dK1 (s). Rewrite system (10.70) in the form
(10.31) with
 ∞ t  ∞ t
Z1 (t) = z1 (t) − z1 (θ ) dθ dK(s) + f2 x2∗ z2 (θ ) dθ dK1 (s),
0 t−s 0 t−s
 ∞ t
Z2 (t) = z2 (t) + g1 x2∗ z1 (θ ) dθ dR1 (s)
0 t−s
 ∞ t
− g2 x2∗ z2 (θ ) dθ dR1 (s), (10.71)
0 t−s
 
f0  
a11 = K1 x2∗ − f1 − K0 x1∗ , a12 = K1 f2 x2∗ − f0 ,
x1∗
a21 = R1 g1 x2∗ , a22 = −R1 g2 x2∗ .
10.3 Stability of Equilibrium Point 279

Further investigation is similar to the previous sections.

For short, consider system (10.67) by conditions (10.11). The point of equilib-
rium in this case is defined by (10.18). From (10.11), (10.18), (10.70) and (10.71) it
follows that system (10.67) and the linear part of this system respectively take the
forms
 
a1 x2 (t)  
ẋ1 (t) = x1 (t) a − a0 x1 (t) − + σ1 x1 (t) − x1∗ ẇ1 (t),
x1 (t) + a2 x2 (t)
  (10.72)
b1 x1 (t − h)  ∗

ẋ2 (t) = x2 (t) −b + + σ2 x1 (t) − x2 ẇ2 (t)
x1 (t − h) + b2 x2 (t − h)
and
ż1 (t) = a11 z1 (t) + a12 z2 (t) + σ1 z1 (t)ẇ1 (t),
(10.73)
ż2 (t) = a21 z1 (t − h) + a22 z2 (t − h) + σ2 z2 (t)ẇ2 (t),
where

a11 = Ba1 α 2 − A, a12 = −B 2 a1 α 2 ,

a21 = b1 b2 β 2 , a22 = −Bb1 b2 β 2 ,
bb2 1 1
A = a − a1 α, B = , α= , β= . (10.74)
b1 − b B + a2 B + b2

Let  = aij  be the matrix with the elements defined by (10.74). Suppose that

b ∈ (0, b1 ),

a1 α if a1 α 2 ≤ b1 b2 β 2 , (10.75)
a>
a1 α + B(a1 α 2 − b1 b2 β 2 ) if a1 α 2 > b1 b2 β 2 .

By conditions (10.75) conditions (2.62) for the matrix  hold. Indeed,

 
Tr(Â) = B a1 α 2 − b1 b2 β 2 − A < 0, det(Â) = ABb1 b2 β 2 > 0. (10.76)

Let P = pij  be the matrix with the elements defined by (1.29) for some
q > 0 and represented in the form (10.35), (10.36). Using (10.35), (10.36), (10.44),
(10.76), put
a12 a21
ρ = ρ (0) q + ρ (1) , ρ (0) = − , ρ (1) = , (10.77)
| Tr(Â)| | Tr(Â)|
and
(0) (1)
A1 = 1 − p11 δ1 − ρ (0) |a21 |h, A2 = 1 − p22 δ2 − |a22 |h,
(1) (0) 1
B1 = p11 δ1 + ρ (1) |a21 |h, B2 = p22 δ2 , δi = σi2 , i = 1, 2, (10.78)
2
 
C1 = |a21 | + ρ (1) |a22 | h, C2 = ρ (0) |a22 |h.
280 10 Stability of Positive Equilibrium Point of Nonlinear System of Type

Rewrite system (10.73) in the form

ż1 (t) = a11 z1 (t) + a12 z2 (t) + σ1 z1 (t)ẇ1 (t),

(10.79)
Ż2 (t) = a21 z1 (t) + a22 z2 (t) + σ2 z2 (t)ẇ2 (t),

where
 t  
Z2 (t) = z2 (t) + a21 z1 (s) + a22 z2 (s) ds, (10.80)
t−h
and following condition (2.10), suppose that the parameters a21 and a22 in (10.74)
√
satisfy the condition h a21 + a22 < 1 or, via (10.74), b1 b2 β 2 h 1 + B 2 < 1, which
2 2

is equivalent to

b1 b2
(b1 − b) (b1 − b)2 + b2 b22 < . (10.81)
h
Theorem 10.3 Let conditions (10.75), (10.81) hold. If A1 > 0, A2 > 0, and

(A1 C1 + B1 C2 )(A2 C2 + B2 C1 ) + B1 B2 < A1 A2 , (10.82)

then the trivial solution of system (10.73) is asymptotically mean-square stable, and
the equilibrium point of system (10.72) is stable in probability.

Proof Consider the functional

V1 (t) = p11 z12 (t) + 2p12 z1 (t)Z2 (t) + p22 Z22 (t)

with pij , i, j = 1, 2, defined by (1.29). Let L be the generator of system (10.79).

Then, using (10.77), similarly to (10.45), for system (10.79), we obtain

LV1 (t) = −qz12 (t) + ρZ2 (t)z1 (t) + p11 σ12 z12 (t)
− ρz1 (t)z2 (t) − Z2 (t)z2 (t) + p22 σ22 z22 (t). (10.83)

Substituting (10.80) into (10.83) and using some positive γ , we obtain

  t 
 
LV1 (t) = −qz12 (t) + ρz1 (t) z2 (t) + a21 z1 (s) + a22 z2 (s) ds + p11 σ12 z12 (t)
t−h
  
t  
− ρz1 (t)z2 (t) − z2 (t) z2 (t) + a21 z1 (s) + a22 z2 (s) ds
t−h

+ p22 σ22 z22 (t)


  ρ t  
≤ −q + p11 σ12 z12 (t) + |a21 | z12 (t) + z12 (s) ds
2 t−h

ρ t  
+ |a22 | γ −1 z12 (t) + γ z22 (s) ds
2 t−h
10.3 Stability of Equilibrium Point 281
 t
  1  2 
+ −1 + p22 σ22 z22 (t) + |a21 | γ z2 (t) + γ −1 z12 (s) ds
2 t−h
 t
1  2 
+ |a22 | z2 (t) + z22 (s) ds
2 t−h
 
ρ −1 ρ
= −q + p11 σ1 + |a21 |h + γ
2
|a22 |h z12 (t)
2 2
 
1 γ
+ −1 + p22 σ22 + |a22 |h + |a21 |h z22 (t)
2 2
 t  t
|a21 |   |a22 |
+ ρ + γ −1 z12 (s) ds + (1 + ργ ) z22 (s) ds.
2 t−h 2 t−h

Putting

|a21 |   t
V2 = ρ + γ −1 (s − t + h)z12 (s) ds
2 t−h

|a22 | t
+ (1 + ργ ) (s − t + h)z22 (s) ds,
2 t−h

for the functional V = V1 + V2 , we have


γ −1   2
LV (t) ≤ −q + p11 σ1 + ρ|a21 |h +
2
h |a21 | + ρ|a22 | z1 (t)
2


γ   2
+ −1 + p22 σ2 + |a22 |h + h |a21 | + ρ|a22 | z2 (t). (10.84)
2
2
Using the representations (10.35), (10.36), (10.77), (10.78), we can rewrite
(10.84) in the form

  
γ −1 γ −1
LV (t) ≤ q −A1 + C2 + B1 + C1 z12 (t)
2 2

 
γ γ
+ −A2 + C1 + q B2 + C2 z22 (t),
2 2
which coincides with (10.49). So, from this (10.82) follows, which coincides with
(10.41). The proof is completed. 

The regions of stability in probability for a positive point of equilibrium of sys-

tem (10.72), obtained by conditions (10.81), (10.82), are shown in the space of the
parameters (a, b) for a0 = 0.3, a1 = 5 a2 = 0.5, b1 = 6, b2 = 2, h = 0.4 and dif-
ferent values of δ1 , δ2 : in Fig. 10.5 for δ1 = 1.5, δ2 = 0.05, in Fig. 10.6 for δ1 = 1,
δ2 = 0.55.
In the both figures the thick line shows the stability region given by conditions
(10.75) that corresponds to the values of the parameters h = δ1 = δ2 = 0.
282 10 Stability of Positive Equilibrium Point of Nonlinear System of Type

Fig. 10.5 Region of stability in probability for (10.69): a0 = 0.3, a1 = 5, a2 = 0.5, b1 = 6, b2 = 2,

h = 0.4, δ1 = 1.5, δ2 = 0.05

Fig. 10.6 Region of stability in probability for (10.69): a0 = 0.3, a1 = 5, a2 = 0.5, b1 = 6, b2 = 2,

h = 0.4, δ1 = 1, δ2 = 0.55
Chapter 11
Stability of SIR Epidemic Model Equilibrium
Points

11.1 Problem Statement

Investigation of different mathematical models for infectious diseases (in particular,

so-called SIR epidemic models) has a long story, and until now these models are
very popular in researchers (see, for example, [6, 18, 26–28, 41, 45, 58, 71, 110,
117, 118, 123, 124, 152, 154, 168, 179, 187, 188, 193, 202, 203, 212, 213, 218, 223,
289, 294–296, 310, 315, 319, 322–324]). Here sufficient conditions for the stability
in probability of two equilibrium points of SIR epidemic model with distributed
delays and stochastic perturbations are obtained.
Consider the mathematical model of the spread of infections diseases. Let S(t) be
the number of members of a population susceptible to the disease at time t, I (t) be
the number of infective members at time t , and R(t) be the number of members that
have been removed from the possibility of infection at time t through full immunity,
b is the recruitment rate of the population, μ1 , μ2 , and μ3 are the natural death
rates of the susceptible, infective, and recovered individuals, respectively, β is the
transmission rate, and λ is the natural recovery rate of the infective individuals. Then
the SIR epidemic model can be described by the system of the differential equations
 ∞
Ṡ(t) = b − βS(t) I (t − s) dF (s) − μ1 S(t),
0
 ∞
I˙(t) = βS(t) I (t − s) dF (s) − (μ2 + λ)I (t), (11.1)
0

Ṙ(t) = λI (t) − μ3 R(t).

Here we assume that b, β, λ, μ1 , μ2 , μ3 are positive constants, the integral is un-

derstanding in the Stieltjes sense, and F (s) is a nondecreasing function such that
 ∞
dF (s) = 1. (11.2)
0

L. Shaikhet, Lyapunov Functionals and Stability of Stochastic Functional 283

Differential Equations, DOI 10.1007/978-3-319-00101-2_11,
© Springer International Publishing Switzerland 2013
284 11 Stability of SIR Epidemic Model Equilibrium Points

The equilibrium point of (11.1) is defined by the conditions Ṡ(t) = 0, I˙(t) = 0,

Ṙ(t) = 0 or by the system of algebraic equations

βS ∗ I ∗ + μ1 S ∗ = b, βS ∗ I ∗ = (μ2 + λ)I ∗ , λI ∗ = μ3 R ∗ . (11.3)

From (11.3) it follows that system (11.1) has two points of equilibrium: E0 =
(bμ−1
1 , 0, 0) and

  μ2 + λ ∗ b(S ∗ )−1 − μ1 λI ∗
E∗ = S ∗ , I ∗ , R ∗ with S ∗ = ,I = , R∗ = . (11.4)
β β μ3

Remark 11.1 From (11.4) it follows that by the condition

bβ > μ1 (μ2 + λ) (11.5)

the point E∗ is a positive equilibrium point of (11.1).

Below we obtain sufficient conditions for the stability in probability of both equi-
librium points E0 and E∗ of (11.1) by stochastic perturbations.

11.2 Stability in Probability of the Equilibrium Point

E0 = (bμ−11 , 0, 0)
We will assume that (11.1) is influenced by stochastic perturbations of white
noise type that are directly proportional to the deviation of the system state
(S(t), I (t), R(t)) from the equilibrium point E0 and influence the Ṡ(t), I˙(t), Ṙ(t)
respectively. So, (11.1) takes the form
 
Ṡ(t) = b − βS(t)J (It ) − μ1 S(t) + σ1 S(t) − bμ−1
1 ẇ1 (t),

I˙(t) = βS(t)J (It ) − (μ2 + λ)I (t) + σ2 I (t)ẇ2 (t), (11.6)

Ṙ(t) = λI (t) − μ3 R(t) + σ3 R(t)ẇ3 (t),

where
 ∞
J (It ) = I (t − s) dF (s), (11.7)
0
σ1 , σ2 , σ3 are constants, and w1 (t), w2 (t), w3 (t) are mutually independent standard
Wiener processes. Note that the equilibrium point E0 is a solution of (11.6).
Centering (11.6) on the equilibrium point E0 via the variables x1 (t) = S(t) −
bμ−11 , x2 (t) = I (t), x3 (t) = R(t), we obtain

ẋ1 (t) = −μ1 x1 (t) − βbμ−1

1 J (x2t ) − βx1 (t)J (x2t ) + σ1 x1 (t)ẇ1 (t),

ẋ2 (t) = −(μ2 + λ)x2 (t) + βbμ−1

1 J (x2t ) + βx1 (t)J (x2t ) + σ2 x2 (t)ẇ2 (t),
(11.8)
ẋ3 (t) = λx2 (t) − μ3 x3 (t) + σ3 x3 (t)ẇ3 (t).
11.2 Stability in Probability of the Equilibrium Point E0 = (bμ−1
1 , 0, 0) 285

It is clear that the stability of the equilibrium point E0 of (11.6) is equivalent to the
stability of the trivial solution of (11.8).
Below we will obtain sufficient conditions for the stability in probability of the
trivial solution of (11.8). As it was noted in Remark 5.3, for getting sufficient con-
ditions for the stability in probability of the trivial solution of the nonlinear system
(11.8) with the order of a nonlinearity higher than one, it is enough to get sufficient
conditions for the asymptotic mean-square stability of the trivial solution of the lin-
ear part of the considered nonlinear system. Thus, besides (11.8), we will consider
the linear part of this system

ẏ1 (t) = −μ1 y1 (t) − βbμ−1

1 J (y2t ) + σ1 y1 (t)ẇ1 (t),

ẏ2 (t) = −(μ2 + λ)y2 (t) + βbμ−1

1 J (y2t ) + σ2 y2 (t)ẇ2 (t), (11.9)
ẏ3 (t) = λy2 (t) − μ3 y3 (t) + σ3 y3 (t)ẇ3 (t),

and following the procedure of constructing Lyapunov functionals, we will also

consider the auxiliary system without delays

ż1 (t) = −μ1 z1 (t) + σ1 z1 (t)ẇ1 (t),

ż2 (t) = −(μ2 + λ)z2 (t) + σ2 z2 (t)ẇ2 (t), (11.10)
ż3 (t) = λz2 (t) − μ3 z3 (t) + σ3 z3 (t)ẇ3 (t).

Put now
1
δi = σi2 , i = 1, 2, 3. (11.11)
2
Lemma 11.1 The trivial solution of (11.10) is asymptotically mean-square stable if
and only if
δ1 < μ1 , δ2 < μ2 + λ, δ3 < μ3 . (11.12)

Proof From Remark 2.5 it follows that the first and second inequalities in (11.12)
are necessary and sufficient conditions for the asymptotic mean-square stability of
the trivial solutions of the first and second equations in (11.10), respectively. To
prove the third inequality in (11.12), let us consider separately the system of the
second and third equations in (11.10). If σ2 = σ3 = 0, then by Corollary 1.1 the
trivial solution of this system is asymptotically stable. So, by Theorem 1.3 and (1.29)
the matrix equation (1.27) with
   
−(μ2 + λ) 0 q 0
A= , Q=
λ −μ3 0 1

has a positive definite solution P with the elements

2−1 q + λp12 1 λ
p11 = , p22 = , p12 = . (11.13)
μ2 + λ 2μ3 2μ3 (μ2 + μ3 + λ)
286 11 Stability of SIR Epidemic Model Equilibrium Points

Let L be the generator of the considered system, and

V = p11 z22 (t) + 2p12 z2 (t)z3 (t) + p22 z32 (t), (11.14)

where the parameters p11 , p12 , p22 are defined in (11.13). Then by (11.14), (11.10),
(11.13), (11.11) we have
 
LV = −2 p11 z2 (t) + p12 z3 (t) (μ2 + λ)z2 (t) + p11 σ22 z22 (t)
  
+ 2 p12 z2 (t) + p22 z3 (t) λz2 (t) − μ3 z3 (t) + p22 σ32 z32 (t)
= (2p11 δ2 − q)z22 (t) + 2(−μ3 + δ3 )p22 z32 (t).

If δ2 < μ2 +λ, then from (11.13) for big enough q, i.e., q > 2λp12 δ2 (μ2 +λ−δ2 )−1 ,
we obtain

(q + 2λp12 )δ2
2p11 δ2 − q = −q
μ2 + λ
 
μ2 + λ − δ2 2λp12 δ2
=− q−
μ2 + λ μ2 + λ − δ2
< 0.

So, by Remark 2.1 and (11.12) the trivial solution of (11.10) is asymptotically mean-
square stable.
Let us suppose that δ3 ≥ μ3 . It is easy to see that in this case there exists small
enough q > 0 such that LV ≥ 0. So, EV (z2 (t), z3 (t)) ≥ EV (z2 (0), z3 (0)) > 0, and
the trivial solution of the considered system cannot be asymptotically mean-square
stable. The proof is completed. 

Theorem 11.1 If

δ1 < μ 1 , δ2 < μ2 + λ − βbμ−1

1 , δ3 < μ3 , (11.15)

then the trivial solution of (11.9) is asymptotically mean-square stable.

Proof Note that in (11.9) the first equation does not depend on y3 (t) and the third
equation does not depend on y1 (t). So, (11.9) can be considered as two separate
systems, (y1 (t), y2 (t)) and (y2 (t), y3 (t)).
First, let L be the generator of the system (y2 (t), y3 (t)), and

V1 = p11 y22 (t) + 2p12 y2 (t)y3 (t) + p22 y32 (t), (11.16)
11.2 Stability in Probability of the Equilibrium Point E0 = (bμ−1
1 , 0, 0) 287

where the parameters p11 , p12 , p22 are defined in (11.13). Then by (11.16), (11.9),
(11.13), (11.11) we have
  
LV1 = 2 p11 y2 (t) + p12 y3 (t) −(μ2 + λ)y2 (t) + βbμ−1 1 J (y2t ) + p11 σ2 y2 (t)
2 2

  
+ 2 p12 y2 (t) + p22 y3 (t) λy2 (t) − μ3 y3 (t) + p22 σ32 y32 (t)
= (−q + 2δ2 p11 )y22 (t) + 2(−μ3 + δ3 )p22 y32 (t)
 
+ 2βbμ−1
1 p11 y2 (t) + p12 y3 (t) J (y2t ).

Note that, by (11.2) and (11.7), 2y2 (t)J (y2t ) ≤ y22 (t) + J (y2t
2 ) and 2y (t)J (y ) ≤
3 2t
−1
α y3 (t) + αJ (y2t ) for some α > 0. Thus,
2 2

LV1 ≤ (−q + 2δ2 p11 )y22 (t) + 2(−μ3 + δ3 )p22 y32 (t)
  2  2   −1 2  2 
+ βbμ−1
1 p11 y2 (t) + J y2t + p12 α y3 (t) + αJ y2t
    2
≤ −q + 2δ2 + βbμ−1 1 p11 y2 (t)
  2  2
+ 2(−μ3 + δ3 )p22 + α −1 βbμ−1 1 p12 y3 (t) + AJ y2t ,

where A = βbμ−1
1 (p11 + αp12 ).
Following the procedure of constructing Lyapunov functionals, put
 ∞ t
V2 = A y22 (τ ) dτ dF (s). (11.17)
0 t−s

Then, by (11.2) and (11.7), LV2 = A(z22 (t) − J (z2t

2 )), and, as a result, for the func-

tional V = V1 + V2 , we obtain
    2
LV ≤ −q + 2δ2 + βbμ−1 1 p11 + A y2 (t)
  2
+ 2(−μ3 + δ3 )p22 + αβbμ−1 1 p12 y3 (t)
    2
= −q + 2 δ2 + βbμ−1 −1
1 p11 + αβbμ1 p12 y2 (t)
  2
+ 2(−μ3 + δ3 )p22 + α −1 βbμ−1 1 p12 y3 (t).

If
 −1 
αβbμ−1
1 p12 < q − 2 δ2 + βbμ1 p11 ,
(11.18)
α −1 βbμ−1
1 p12 < 2(μ3 − δ3 )p22 ,
then by Remark 2.1 the trivial solution of the considered system is asymptotically
mean-square stable.
From (11.18) and (11.15) it follows that

βbμ−1
1 p12 q − 2(δ2 + βbμ−1
1 )p11
0< <α< −1
. (11.19)
2(μ3 − δ3 )p22 βbμ1 p12
288 11 Stability of SIR Epidemic Model Equilibrium Points

So, if
βbμ−1
1 p12 q − 2(δ2 + βbμ−1
1 )p11
< −1
, (11.20)
2(μ3 − δ3 )p22 βbμ1 p12
then there exists α > 0 such that (11.19) holds.
Note that by (11.13) we have

  (δ2 + βbμ−1
1 )(q + 2λp12 )
q − 2 δ2 + βbμ−1 p11 = q −
1 μ2 + λ
q(μ2 + λ − βbμ−1
1 − δ2 ) 2λp12 (δ2 + βbμ−1
1 )
= − .
μ2 + λ μ2 + λ
From this and from (11.15) it follows that for big enough q > 0, i.e.,
 
μ2 + λ (βbμ−1
1 p12 )
2 2λp12 (δ2 + βbμ−1
1 )
q> + ,
μ2 + λ − βbμ−1
1 − δ2
2(μ3 − δ3 )p22 μ2 + λ

equation (11.20) holds. So, the trivial solution of the considered system is asymp-
totically mean-square stable.
Now let L be the generator of the system (y1 (t), y2 (t)) in (11.9), and

V1 = p11 y12 (t) + p22 y22 (t), (11.21)

where
q 1
p11 = , p22 = . (11.22)
2μ1 2(μ2 + λ)
Then by (11.21), (11.9), and (11.11) we have
 
LV1 = 2p11 y1 (t) −μ1 y1 (t) − βbμ−11 J (y2t ) + p11 σ1 y1 (t)
2 2

 
+ 2p22 y2 (t) −(μ2 + λ)y2 (t) + βbμ−1 1 J (y2t ) + p22 σ2 y2 (t)
2 2

= −2(μ1 − δ1 )p11 y12 (t) − 2(μ2 + λ − δ2 )p22 y22 (t)

 
+ 2βbμ−1
1 J (y2t ) p22 y2 (t) − p11 y1 (t) .

Note that, by (11.2) and (11.7), 2y2 (t)J (y2t ) ≤ y22 (t) + J (y2t
2 ) and 2y (t)J (y ) ≤
1 2t
α −1 y12 (t) + αJ (y2t
2 ) for some α > 0. Thus,

LV1 ≤ −2(μ1 − δ1 )p11 y12 (t) − 2(μ2 + λ − δ2 )p22 y22 (t)

  −1 2  2   2  2 
+ βbμ−1
1 p11 α y1 (t) + αJ y2t + p22 y2 (t) + J y2t
 
= −2(μ1 − δ1 ) + α −1 βbμ−1 2
1 p11 y1 (t)
   2
+ −2(μ2 + λ − δ2 ) + βbμ−1 1 p22 y2 (t) + AJ y2t ,
2
11.3 Stability in Probability of the Equilibrium Point E∗ = (S ∗ , I ∗ , R ∗ ) 289

where A = βbμ−1 1 (αp11 + p22 ). Using (11.17) with this A for the functional V =
V1 + V2 , we have
 
LV ≤ −2(μ1 − δ1 ) + α −1 βbμ−1 2
1 p11 y1 (t)
 
+ −2(μ2 + λ − δ2 ) + βbμ−1 1 p22 y2 (t) + Ay2 (t)
2 2

 
= −2(μ1 − δ1 ) + α −1 βbμ−1 2
1 p11 y1 (t)
    2
+ −2 μ2 + λ − βbμ−1 −1
1 − δ2 p22 + αβbμ1 αp11 y2 (t).

If
 
α −1 βbμ−1
1 < 2(μ1 − δ1 ), αβbμ−1 −1
1 p11 < 2 μ2 + λ − βbμ1 − δ2 p22 ,
(11.23)
then by Remark 2.1 the trivial solution of the considered system is asymptotically
mean-square stable.
From (11.23), (11.22), and (11.15) it follows that

βbμ−1
1 2(μ2 + λ − βbμ−1
1 − δ2 )μ1
0< <α< −1
. (11.24)
2(μ1 − δ1 ) βbμ1 (μ2 + λ)q
So, if
βbμ−1
1 2(μ2 + λ − βbμ−1
1 − δ2 )μ1
< −1
, (11.25)
2(μ1 − δ1 ) βbμ1 (μ2 + λ)q
then there exists α > 0 such that (11.24) holds. It is easy to see that (11.25) holds
by conditions (11.15) for small enough q > 0 such that

4(μ1 − δ1 )(μ2 + λ − βbμ−1 1 − δ2 )μ1

3
q< .
β 2 b2 (μ2 + λ)
So, the trivial solution of (11.9) is asymptotically mean-square stable. The proof is
completed. 

Corollary 11.1 If conditions (11.15) hold, then the equilibrium point E0 of (11.6)
is stable in probability.

Remark 11.2 Note that the second condition (11.15) contradicts with (11.5).
It means that by conditions (11.15) system (11.1) has no positive equilibrium
point (11.4).

11.3 Stability in Probability of the Equilibrium Point

E∗ = (S ∗ , I ∗ , R ∗ )
Here we assume that (11.1) is influenced by stochastic perturbations of white
noise type that are directly proportional to the deviation of the system state
290 11 Stability of SIR Epidemic Model Equilibrium Points

(S(t), I (t), R(t)) from the equilibrium point E∗ = (S ∗ , I ∗ , R ∗ ) and influence Ṡ(t),
I˙(t), Ṙ(t). Then (11.1) takes the form
 
Ṡ(t) = b − βS(t)J (It ) − μ1 S(t) + σ1 S(t) − S ∗ ẇ1 (t),
 
I˙(t) = βS(t)J (It ) − (μ2 + λ)I (t) + σ2 I (t) − I ∗ ẇ2 (t), (11.26)
 
Ṙ(t) = λI (t) − μ3 R(t) + σ3 R(t) − R ∗ ẇ3 (t),

where J (It ), σ1 , σ2 , σ3 , w1 (t), w2 (t), w3 (t) are the same as in (11.6), and the equi-
librium point E∗ = (S ∗ , I ∗ , R ∗ ) is the solution of (11.26).
Centering (11.26) on the positive equilibrium E∗ via the variables x1 (t) = S(t) −
S ∗ , x2 (t) = I (t) − I ∗ , x3 (t) = R(t) − R ∗ and using (11.2), (11.3), we obtain
 −1
ẋ1 (t) = −b S ∗ x1 (t) − βS ∗ J (x2t ) − βx1 (t)J (x2t ) + σ1 x1 (t)ẇ1 (t),
ẋ2 (t) = βI ∗ x1 (t) − βS ∗ x2 (t) + βS ∗ J (x2t )
(11.27)
+ βx1 (t)J (x2t ) + σ2 x2 (t)ẇ2 (t),
ẋ3 (t) = λx2 (t) − μ3 x3 (t) + σ3 x3 (t)ẇ3 (t).

It is easy to see that the stability of the equilibrium point E∗ of (11.26) is equivalent
to the stability of the zero solution of (11.27).
Below we will obtain sufficient conditions for the stability in probability of the
zero solution of (11.27) using the linear part of this system
 −1
ẏ1 (t) = −b S ∗ y1 (t) − βS ∗ J (y2t ) + σ1 y1 (t)ẇ1 (t),
ẏ2 (t) = βI ∗ y1 (t) − βS ∗ y2 (t) + βS ∗ J (y2t ) + σ2 y2 (t)ẇ2 (t), (11.28)
ẏ3 (t) = λy2 (t) − μ3 y3 (t) + σ3 y3 (t)ẇ3 (t)

and the auxiliary system without delays

 −1
ż1 (t) = −b S ∗ z1 (t) + σ1 z1 (t)ẇ1 (t),
ż2 (t) = βI ∗ z1 (t) − βS ∗ z2 (t) + σ2 z2 (t)ẇ2 (t), (11.29)
ż3 (t) = λz2 (t) − μ3 z3 (t) + σ3 z3 (t)ẇ3 (t).

Lemma 11.2 If
 −1
δ1 < b S ∗ , δ2 < βS ∗ , δ3 < μ3 , (11.30)

then the trivial solution of (11.29) is asymptotically mean-square stable.

Proof Note that the first inequality in (11.30) is a necessary and sufficient condition
for the asymptotic mean-square stability of the trivial solution of the first equation
11.3 Stability in Probability of the Equilibrium Point E∗ = (S ∗ , I ∗ , R ∗ ) 291

(11.29). Note also that the system of two first equations in (11.29) has the structure
that is the same as the structure of the second and the third equations in (11.10). So,
similarly to (11.12), we obtain two first conditions in (11.30).
Finally, note that the third equation in (11.29) with the additive perturbation z2 (t)
that satisfies the condition limt→∞ Ez22 (t) = 0 has the structure that is the same
as the structure of the second equation in (11.29) with the additive perturbation
z1 (t) that satisfies the same condition limt→∞ Ez12 (t) = 0. Thus, since the inequality
δ2 < βS ∗ is a sufficient condition for the asymptotic mean-square stability of the
trivial solution of the second equation, the similar inequality δ3 < μ3 is a sufficient
condition for asymptotic mean-square stability of the trivial solution of the third
equation. The proof is completed. 

Theorem 11.2 Let δ3 < μ3 and

βS ∗ βI ∗
δ1 < μ1 , δ2 < (11.31)
b(S ∗ )−1 + βS ∗
or

4I ∗ (S ∗ )−1 + 1 − 1
∗
μ1 ≤ δ1 < μ1 + βI ,
4I ∗ (S ∗ )−1 + 1 + 1
  (11.32)
βS ∗ βI ∗ βS ∗ (δ1 − μ1 )
δ2 < 1− .
b(S ∗ )−1 + βS ∗ (b(S ∗ )−1 − δ1 )2
Then the trivial solution of (11.28) is asymptotically mean-square stable.

Proof Let L be the generator of the system of two first equations in (11.28). Fol-
lowing the procedure of constructing Lyapunov functionals, we will construct a Lya-
punov functional V for this system in the form V = V1 + V2 , where

V1 = p11 y12 (t) + 2p12 y1 (t)y2 (t) + p22 y22 (t) (11.33)

and the parameters p11 , p12 , p22 by (1.29) are

S∗ βI ∗ S ∗ 1 I ∗ (S ∗ )−1
p11 = q+ p12 , p22 = , p12 = .
2b b 2βS ∗ 2(b(S ∗ )−1 + βS ∗ )
(11.34)
Then by (11.33), (11.28) we have
   −1 
LV1 = 2 p11 y1 (t) + p12 y2 (t) −b S ∗ y1 (t) − βS ∗ J (y2t ) + p11 σ12 y12 (t)
 
+ 2 p12 y1 (t) + p22 y2 (t) βI ∗ y1 (t) − βS ∗ y2 (t)

+ βS ∗ J (y2t ) + p22 σ22 y22 (t)
  −1    
= 2 δ1 − b S ∗ p11 + βI ∗ p12 y12 (t) + 2 δ2 − βS ∗ p22 y22 (t)
292 11 Stability of SIR Epidemic Model Equilibrium Points
   −1  
+ 2 βI ∗ p22 − b S ∗ + βS ∗ p12 y1 (t)y2 (t)
 
+ 2βS ∗ (p12 − p11 )y1 (t) + (p22 − p12 )y2 (t) J (y2t ).

Note that, by (11.2) and (11.7), 2y2 (t)J (y2t ) ≤ y22 (t) + J (y2t
2 ) and 2y (t)J (y ) ≤
1 2t
α −1 y12 (t) + αJ (y2t
2 ) for some α > 0. Besides, by (11.34) and (11.4),

  −1 
δ1 − b S ∗ p11 + βI ∗ p12 = (δ1 − μ1 )p11 + βI ∗ (p12 − p11 ),
  −1 
βI ∗ p22 − b S ∗ + βS ∗ p12 = 0,
b(S ∗ )−1 + βS ∗ − βI ∗ μ1 + βS ∗
p22 − p12 = = > 0.
2βS ∗ (b(S ∗ )−1 + βS ∗ ) 2βS ∗ (b(S ∗ )−1 + βS ∗ )

Thus,
   
LV1 ≤ 2 (δ1 − μ1 )p11 + βI ∗ (p12 − p11 ) + α −1 βS ∗ |p12 − p11 | y12 (t)
    2
+ 2δ2 − βS ∗ p22 − βS ∗ p12 y22 (t) + AJ y2t ,

where A = βS ∗ (α|p12 − p11 | + p22 − p12 ).

Choosing the additional functional V2 in the form (11.17) with A obtained above,
for the functional V = V1 + V2 , we have
   
LV ≤ 2 (δ1 − μ1 )p11 + βI ∗ (p12 − p11 ) + α −1 βS ∗ |p12 − p11 | y12 (t)
   
+ 2 δ2 p22 − βS ∗ p12 + αβS ∗ |p12 − p11 | y22 (t). (11.35)

Note that by (11.34) and (11.4)

S∗
p11 − p12 = (q − 2μ1 p12 ). (11.36)
2b
So, putting q = 2μ1 p12 > 0, we obtain p11 = p12 and
 −1 
LV ≤ 2(δ1 − μ1 )p11 y12 (t) + 2 δ2 − βS ∗ p12 p22 p22 y22 (t). (11.37)

−1
By (11.34) we have βS ∗ p12 p22 = βS ∗ βI ∗ (b(S ∗ )−1 + βS ∗ )−1 . So, from (11.37) it
follows that by conditions (11.31) the trivial solution of (11.28) is asymptotically
mean-square stable.
Suppose now that δ1 ≥ μ1 and q > 2μ1 p12 > 0. Then p11 > p12 , and from
(11.35) it follows that if

βS ∗ (p11 − p12 ) 2(βS ∗ p12 − δ2 p22 )

0< < α < , (11.38)
2(βI ∗ (p11 − p12 ) − (δ1 − μ1 )p11 ) βS ∗ (p11 − p12 )

then the trivial solution of (11.28) is asymptotically mean-square stable.

11.3 Stability in Probability of the Equilibrium Point E∗ = (S ∗ , I ∗ , R ∗ ) 293

If
βS ∗ (p11 − p12 ) 2(βS ∗ p12 − δ2 p22 )
0< < , (11.39)
2(βI ∗ (p11 − p12 ) − (δ1 − μ1 )p11 ) βS ∗ (p11 − p12 )
then there exists α > 0 such that (11.38) holds. By (11.4), (11.36) from (11.39) it
follows that
(βS ∗ )2 βI ∗ (p11 − p12 ) − (δ1 − μ1 )p11
<
4(βS ∗ p12 − δ2 p22 ) (p11 − p12 )2
(b(S ∗ )−1 − δ1 )(p11 − p12 ) − (δ1 − μ1 )p12
=
(p11 − p12 )2
b(S ∗ )−1 − δ1 (δ1 − μ1 )p12
= −
p11 − p12 (p11 − p12 )2
 ∗ 
2b S (b(S ∗ )−1 − δ1 ) 2bp12 (δ1 − μ1 )
= ∗ 2 − . (11.40)
(S ) q − 2μ1 p12 (q − 2μ1 p12 )2
Note that the right-hand part of the obtained inequality reaches its maximum at
4(δ1 − μ1 )bp12
q = 2μ1 p12 + . (11.41)
S ∗ (b(S ∗ )−1 − δ1 )
Substituting (11.41) into (11.40), we obtain

(βS ∗ )2 (b(S ∗ )−1 − δ1 )2

<
βS ∗ p12 − δ2 p22 (δ1 − μ1 )p12
or
 
βS ∗ p12 βS ∗ (δ1 − μ1 )
δ2 < 1− ,
p22 (b(S ∗ )−1 − δ1 )2
which is equivalent to the second condition in (11.32).
From the positivity of the expression in the brackets, i.e., from the inequality
βS ∗ (δ1 − μ1 ) < (b(S ∗ )−1 − δ1 )2 , the first condition (11.32) follows. Indeed, put
 = δ1 − μ1 . Then by (11.3) b(S ∗ )−1 − δ1 = βI ∗ − , and we get βS ∗  < (βI ∗ −
)2 or 2 − (2βI ∗ + βS ∗ ) + (βI ∗ )2 > 0. From this it follows that

1
 2  2 
< 2βI ∗ + βS ∗ − 2βI ∗ + βS ∗ − 4 βI ∗
2

β
 2 
= βI ∗ − 4I ∗ S ∗ + S ∗ − S ∗
2
2βI ∗ S ∗
= βI ∗ −
4I ∗ S ∗ + (S ∗ )2 + S ∗
 
2
= βI ∗ 1 −
4I ∗ (S ∗ )−1 + 1 + 1
294 11 Stability of SIR Epidemic Model Equilibrium Points

Fig. 11.1 25 trajectories of the processes S(t) (blue), I (t) (green), R(t) (red) with the equilib-
rium point S ∗ = 4.5, I ∗ = 0, R ∗ = 0, the initial values S(0) = 7.5, I (s) = 2.5, s ∈ [−0.1, 0],
R(0) = 3.5, and the values of the parameters b = 1.8, β = 0.2, λ = 1, μ1 = 0.4, μ2 = μ3 = 0.5,
h = 0.1, δ1 = 0.35, δ2 = 0.25, δ3 = 0.15

∗ 4I ∗ (S ∗ )−1 + 1 − 1
= βI ,
4I ∗ (S ∗ )−1 + 1 + 1

which is equivalent to the first condition in (11.32). The proof is completed. 

By Theorem 11.2 and Remark 5.3 the following statement holds.

Corollary 11.2 Let the conditions of Theorem 11.2 hold. Then the equilibrium point
E∗ of (11.26) is stable in probability.

Remark 11.3 From (11.4) and (11.31) it follows that

βS ∗ βI ∗
bβ(μ2 + λ)−1 − μ1 = βI ∗ > > δ2 > 0.
b(S ∗ )−1 + βS ∗

So, (11.5) follows from (11.31). This means that by the conditions of Theorem 11.2
the equilibrium point E∗ is a positive equilibrium point of (11.1).
11.4 Numerical Simulation 295

Fig. 11.2 Stability region for E∗ given by conditions (11.31), (11.32)

11.4 Numerical Simulation

Let us suppose that in (11.1) dF (s) = δ(s − h) ds, where δ(s) is Dirac’s delta-
function, and h ≥ 0 is the delay. So, by (11.7), J (It ) = I (t − h).
Put b = 1.8, β = 0.2, λ = 1, μ1 = 0.4, μ2 = μ3 = 0.5, h = 0.1, and δ1 = 0.35,
δ2 = 0.25, δ3 = 0.15. In this case S ∗ = bμ−1 1 = 4.5 and δ1 < μ1 , δ2 < μ2 + λ −
βbμ−11 = 0.6, δ 3 < μ 3 , i.e., the conditions of Theorem 11.1 hold. So, the equilib-
rium point E0 = (4.5, 0, 0) of (11.6) is stable in probability. In Fig. 11.1 for twenty
five realizations of each of the processes S(t) (blue), I (t) (green) and R(t) (red)
are shown that are the solution of (11.6) with the initial conditions S(0) = 7.5,
I (s) = 2.5, s ∈ [−0.1, 0], R(0) = 3.5. All trajectories converge to the equilibrium
point E0 = (4.5, 0, 0).
Consider now b = 20 with the same values of the other parameters. In this case
condition (11.5) holds: bβ = 4 > μ1 (μ2 + λ) = 0.6. So, (11.6) has a positive equi-
librium point (11.4). The stability region given by the conditions (11.31), (11.32)
for the given values of the parameters is shown in Fig. 11.2 in the space of the pa-
rameters (δ1 , δ2 ). The dotted line separates the stability region into two parts given
by conditions (11.31) and (11.32), respectively. In Fig. 11.3, for twenty five realiza-
tions of each of the processes, S(t) (blue), I (t) (green), and R(t) (red) are shown
that are the solution of (11.6) with the initial conditions S(0) = 3.5, I (s) = 16.33
for s ∈ [−0.1, 0], and R(0) = 19.17 at the point A of the stability region (Fig. 11.2)
with the coordinates δ1 = 1, δ2 = 0.16 and with δ3 = 0.25. One can see that all
296 11 Stability of SIR Epidemic Model Equilibrium Points

Fig. 11.3 25 trajectories of the processes S(t) (blue), I (t) (green), R(t) (red) with the equilibrium
point S ∗ = 7.5, I ∗ = 11.33, R ∗ = 22.67, the initial values S(0) = 3.5, I (s) = 16.33, s ∈ [−0.1, 0],
R(0) = 19.17, and the values of the parameters b = 20, β = 0.2, λ = 1, μ1 = 0.4, μ2 = μ3 = 0.5,
h = 0.1, δ1 = 1, δ2 = 0.16, δ3 = 0.25

trajectories converge to the equilibrium point E∗ = (S ∗ , I ∗ , R ∗ ), where S ∗ = 7.5,

I ∗ = 11.33, R ∗ = 22.67.
Note that numerical simulations of the processes S(t), I (t), and R(t) were
obtained via the Euler–Maruyama scheme [200] for stochastic differential equa-
tions, and difference analogues of systems (11.6), (11.7) and (11.26), (11.7) with
J (It ) = I (t − h) were used in the form
√  
Si+1 = Si + (b − βSi Ii−m − μ1 Si ) + σ1  Si − S ∗ ξ1,i+1 ,
  √  
Ii+1 = Ii +  βSi Ii−m − (μ2 + λ)Ii + σ2  Ii − I ∗ ξ2,i+1 ,
√  
Ri+1 = Ri + (λIi − μ3 Ri ) + σ3  Ri − R ∗ ξ3,i+1 ,
i = 0, 1, 2, . . . , Ij = I0 , j = −m, . . . , −1, 0.
1
Here ξki = − 2 (wk (ti ) − wk (ti−1 )) are mutually independent random variables
such that Eξki = 0, Eξki 2 = 1, k = 1, 2, 3, t = i,  is the step of discretization,
i
Si = S(ti ), Ii = I (ti ), Ri = R(ti ), m = h−1 .
Chapter 12
Stability of Some Social Mathematical Models
with Delay Under Stochastic Perturbations

In this chapter we propose a mathematical framework to model some social be-

havior. To be precise, we propose delayed and stochastic mathematical models to
analyze human behaviors related to some addictions: consumption of alcohol and
obesity.

12.1 Mathematical Model of Alcohol Consumption

Taking into account the proposal presented in [247], we consider alcohol consump-
tion habit as susceptible to be transmitted by peer pressure or social contact. This
fact led us to propose an epidemiologic-type mathematical model to study this social
epidemic.
Here we generalize the known nonlinear dynamic model of alcohol consumption
[254] by adding distributed delay. We obtain sufficient conditions for the existence
of a positive equilibrium point of this system. Similarly to the previous sections, we
suppose that this nonlinear system is exposed to additive stochastic perturbations
of white noise type that are directly proportional to the deviation of the system
state from the equilibrium point. The considered nonlinear system is linearized in
the neighborhood of the positive point of equilibrium, and a sufficient condition
for asymptotic mean-square stability of the zero solution of the constructed linear
system is obtained via the procedure of constructing Lyapunov functionals that is
described in Sect. 2.2.2. Since the order of nonlinearity of the considered nonlinear
system is higher than one, the obtained condition is also a sufficient one (Sect. 5.3)
for stability in probability of the equilibrium point of the initial nonlinear system
under stochastic perturbations.

12.1.1 Description of the Model of Alcohol Consumption

Let A(t) be nonconsumers, individuals that have never consumed alcohol or in-
frequently have alcohol consumption, and M(t) be nonrisk consumers, individuals

L. Shaikhet, Lyapunov Functionals and Stability of Stochastic Functional 297

Differential Equations, DOI 10.1007/978-3-319-00101-2_12,
© Springer International Publishing Switzerland 2013
298 12 Stability of Some Social Mathematical Models with Delay Under Stochastic

with regular low consumption, to be precise, men consuming less than 50 cc (cubic
centimeters) of alcohol every day and women consuming less than 30 cc of alcohol
every day. Let R(t) be risk consumers, individuals with regular high consumption,
that is, men consuming more than 50 cc of alcohol every day and women who con-
suming more than 30 cc of alcohol every day.
Considering homogeneous mixing [219], where each individual can contact with
any other individual (peer pressure), a dynamic alcohol consumption model is given
by the following nonlinear system of ordinary differential equations with distributed
delay:
 ∞
M(t − s) + R(t − s)
Ȧ(t) = μP (t) + γ R(t) − dA A(t) − βA(t) dK(s),
0 P (t − s)
 ∞
M(t − s) + R(t − s)
Ṁ(t) = βA(t) dK(s) − dM(t) − αM(t), (12.1)
0 P (t − s)
Ṙ(t) = αM(t) − γ R(t) − dR(t),
P (t) = A(t) + M(t) + R(t).

Here:
α the rate at which a nonrisk consumer moves to the risk consumption subpopu-
lation (intensity of transition from the group M(t) to the group R(t)).
β the transmission rate due to social pressure to increase the alcohol consumption,
e.g., family, friends, marketing, TV, etc. (intensity of transition from the group
A(t) to the group M(t)).
γ the rate at which a risk consumer becomes a nonconsumer (intensity of transi-
tion from the group R(t) to the group A(t)); so, the scheme of transition from
one group to another one is

A(t) →β → M(t) →α → R(t) →γ → A(t).

μ the birth rate.

dA the death rate.
d the augmented death rate due to alcohol consumption (accidents at work, traffic
accidents, and diseases derived by alcohol consumption are considered).
We suppose that the parameters α, β, γ , μ, dA , d are nonnegative numbers and
K(s) is a nondecreasing function such that
 ∞
dK(s) = 1. (12.2)
0

The integral is understood in the Stieltjes sense.

Remark 12.1 In particular, dK(s) = δ(s − h) ds, where h > 0, δ(s) is Dirac’s func-
tion, system (12.1) is a system with discrete delay h. The case of a system without
delay (h = 0) is considered in [254].
12.1 Mathematical Model of Alcohol Consumption 299

12.1.2 Normalization of the Initial Model

Put
A(t) M(t) R(t)
a(t) = , m(t) = , r(t) = . (12.3)
P (t) P (t) P (t)
From (12.1) and (12.3) it follows that

a(t) + m(t) + r(t) = 1. (12.4)

Adding the first three equations in (12.1), by (12.4) we obtain

Ṗ (t)
= μ − d + (d − dA )a(t).
P (t)
From this and from (12.3) we have

Ȧ(t)P (t) − A(t)Ṗ (t) Ȧ(t) A(t) Ṗ (t)

ȧ(t) = = − ×
P 2 (t) P (t) P (t) P (t)
Ȧ(t)  
= − a(t) μ − d + (d − dA )a(t) , (12.5)
P (t)
and, similarly,

Ṁ(t)  
ṁ(t) = − m(t) μ − d + (d − dA )a(t) ,
P (t)
(12.6)
Ṙ(t)  
ṙ(t) = − r(t) μ − d + (d − dA )a(t) .
P (t)
Thus, putting
 ∞
I (at ) = a(t − s) dK(s), (12.7)
0
by (12.5), (12.6), (12.1), (12.2), and (12.4) we obtain
  
ȧ(t) = μ + γ r(t) + βa(t)I (at ) − a(t) β + μ − (d − dA ) 1 − a(t) ,
 
ṁ(t) = βa(t) − βa(t)I (at ) − m(t) α + μ + (d − dA )a(t) ,
 
ṙ(t) = αm(t) − r(t) γ + μ + (d − dA )a(t) .

In view of (12.4), the last equation can be rejected, and, as a result, we obtain the
system of two integro–differential equations
  
ȧ(t) = μ + γ − γ m(t) + βa(t)I (at ) − a(t) β + μ + γ − (d − dA ) 1 − a(t) ,
  (12.8)
ṁ(t) = βa(t) − βa(t)I (at ) − m(t) α + μ + (d − dA )a(t) .
300 12 Stability of Some Social Mathematical Models with Delay Under Stochastic

12.1.3 Existence of an Equilibrium Point

By (12.8), (12.2), and (12.4) a point of equilibrium (a ∗ , m∗ , r ∗ ) is defined by the

following system of algebraic equations:
   
(μ + γ ) 1 − a ∗ = a ∗ (β − d + dA ) 1 − a ∗ + γ m∗ ,
   
βa ∗ 1 − a ∗ = m∗ α + μ + (d − dA )a ∗ , (12.9)
a ∗ + m∗ + r ∗ = 1.

Lemma 12.1 If d ∈ [dA , β + dA ), then system (12.9) has a unique positive solution
(a ∗ , m∗ , r ∗ ) if and only if
αγ
β > d − dA + μ + . (12.10)
α + μ + γ + d − dA

If d ≥ β + dA , then system (12.9) has no positive solutions.

Proof Necessity From the first two equations in (12.9) we have

γβa ∗
μ + γ = a ∗ (β − d + dA ) + . (12.11)
α + μ + (d − dA )a ∗

Since a ∗ ∈ (0, 1), from (12.11) it follows that

γβ
μ + γ = a ∗ (β − d + dA ) +
(α + μ)(a ∗ )−1 + d − dA
γβ
< β − d + dA +
α + μ + d − dA
α + μ + γ + d − dA
= β − d + dA ,
α + μ + d − dA

which is equivalent to (12.10) since

(μ + γ + d − dA )(α + μ + d − dA )
β>
α + μ + γ + d − dA
αγ
= d − dA + μ + .
α + μ + γ + d − dA

Sufficiency Rewrite (12.11) in the form

 2
Q a ∗ + Ba ∗ − C = 0,
B = (β − d + dA )(α + γ + μ) − μ(d − dA ),
Q = (β − d + dA )(d − dA ), C = (μ + α)(μ + γ ). (12.12)
12.1 Mathematical Model of Alcohol Consumption 301

Thus, by (12.9) the equilibrium point (a ∗ , m∗ , r ∗ ) is defined by the system of alge-

braic equations (12.12), and
βa ∗ (1 − a ∗ )
m∗ = , r ∗ = 1 − a ∗ − m∗ . (12.13)
α + μ + (d − dA )a ∗
It is easy to check that by the condition d ∈ [dA , β + dA ) (or Q ≥ 0) the existence
of a solution a ∗ of (12.12) in the interval (0, 1) is equivalent to the condition C <
Q + B, which is equivalent to (12.10).
If d ≥ β + dA , then Q ≤ 0 and B < 0. So, (12.12) cannot have positive roots.
The proof is completed. 

Example 12.1 Following [254], put

α = 0.000110247, β = 0.0284534, γ = 0.00144,
(12.14)
μ = 0.01, d = 0.009, dA = 0.008.
Then condition (12.10) holds, and the solution of (12.12)–(12.13) is
a ∗ = 0.364739, m∗ = 0.629383, r ∗ = 0.00587794, (12.15)
or, in percents, a ∗ = 36.47 %, m∗ = 62.94 %, r ∗ = 0.59 %.

12.1.4 Stochastic Perturbations, Centralization, and Linearization

Let us suppose that system (12.8) is exposed to stochastic perturbations of white

noise type (ẇ1 (t), ẇ2 (t)), which are directly proportional to the deviation of system
(12.8) state (a(t), m(t)) from the equilibrium point (a ∗ , m∗ ), i.e.,
  
ȧ(t) = μ + γ − γ m(t) + βa(t)I (at ) − a(t) β + μ + γ − (d − dA ) 1 − a(t)
 
+ σ1 a(t) − a ∗ ẇ1 (t),
  (12.16)
ṁ(t) = βa(t) − βa(t)I (at ) − m(t) α + μ + (d − dA )a(t)
 
+ σ2 m(t) − m∗ ẇ2 (t).
Here w1 (t), w2 (t) are the mutually independent standard Wiener processes, and the
stochastic differential equations (12.16) are understood in the Itô sense (Sect. 2.1.2).
To centralize system (12.16) in the equilibrium point, put now x1 (t) = a(t) − a ∗ ,
x2 (t) = m(t) − m∗ . Then from (12.16) it follows that
    
ẋ1 (t) = μ + γ − γ m∗ + x2 (t) + β a ∗ + x1 (t) a ∗ + I (x1t )
   
− a ∗ + x1 (t) β + μ + γ − (d − dA ) 1 − a ∗ − x1 (t) + σ1 x1 (t)ẇ1 (t),
    
ẋ2 (t) = β a ∗ + x1 (t) − β a ∗ + x1 (t) a ∗ + I (x1t )
   
− m∗ + x2 (t) α + μ + (d − dA ) a ∗ + x1 (t) + σ2 x2 (t)ẇ2 (t),
302 12 Stability of Some Social Mathematical Models with Delay Under Stochastic

or
     
ẋ1 (t) = μ 1 − a ∗ + γ 1 − a ∗ − m∗ − a ∗ 1 − a ∗ (β − d + dA ) − μx1 (t)
     
+ γ −x1 (t) − x2 (t) + x1 (t) 1 − 2a ∗ (d − dA ) − βx1 (t) 1 − a ∗
+ βa ∗ I (x1t ) − x12 (t)(d − dA ) + βx1 (t)I (x1t ) + σ1 x1 (t)ẇ1 (t),
      (12.17)
ẋ2 (t) = βa ∗ 1 − a ∗ − m∗ α + μ + (d − dA )a ∗ + βx1 (t) 1 − a ∗
 
− m∗ x1 (t)(d − dA ) − βa ∗ I (x1t ) − x2 (t) α + μ + (d − dA )a ∗
− βx1 (t)I1 (x1t ) − x2 (t)x1 (t)(d − dA ) + σ2 x2 (t)ẇ2 (t).

By (12.9) from (12.17) it follows that

ẋ1 (t) = a11 x1 (t) + a12 x2 (t) + βa ∗ I (x1t ) + βx1 (t)I (x1t )
− (d − dA )x12 (t) + σ1 x1 (t)ẇ1 (t),
(12.18)
ẋ2 (t) = a21 x1 (t) + a22 x2 (t) − βa ∗ I (x1t ) − βx1 (t)I (x1t )
− (d − dA )x1 (t)x2 (t) + σ2 x2 (t)ẇ2 (t),

where
   
a11 = − μ + γ + (β − d + dA ) 1 − a ∗ + (d − dA )a ∗ , a12 = −γ ,
β(α + μ)(1 − a ∗ )   (12.19)
a21 = , a22 = − α + μ + (d − dA )a ∗ .
α + μ + (d − dA )a ∗
Note that for d ∈ [dA , β + dA ), the numbers a11 , a12 , a22 are negative, and a21 > 0.
Rejecting the nonlinear terms in (12.18), we obtain the linear part of (12.18):

ẏ1 (t) = a11 y1 (t) + a12 y2 (t) + βa ∗ I (y1t ) + σ1 y1 (t)ẇ1 (t),

(12.20)
ẏ2 (t) = a21 y1 (t) + a22 y2 (t) − βa ∗ I (y1t ) + σ2 y2 (t)ẇ2 (t).

12.1.5 Stability of the Equilibrium Point

Note that the nonlinear system (12.18) has the order of nonlinearity higher than
one. Thus, as it is shown in Sect. 5.3, sufficient conditions for the asymptotic mean-
square stability of the zero solution of the linear part (12.20) of the nonlinear system
(12.18) at the same time are sufficient conditions for the stability in probability of the
zero solution of the nonlinear system (12.18) and therefore are sufficient conditions
for stability in probability of the solution (a ∗ , m∗ ) of (12.16).
To get sufficient conditions for the asymptotic mean-square stability of the zero
solution of (12.20), rewrite this system in the form
 
ẏ(t) = Ay(t) + B(yt ) + σ y(t) ẇ(t), (12.21)
12.1 Mathematical Model of Alcohol Consumption 303

where
   
y(t) = y1 (t), y2 (t) , w(t) = w1 (t), w2 (t) ,
 
B(yt ) = βa ∗ I (y1t ), −βa ∗ I (y1t ) , (12.22)
   
a a12   σ y (t) 0
A = 11 , σ y(t) = 1 1 .
a21 a22 0 σ2 y2 (t)

Following the procedure of constructing Lyapunov functionals (Sect. 2.2.2), for

stability investigation of (12.21), consider the auxiliary differential equation without
memory
 
ż(t) = Az(t) + σ z(t) ẇ(t). (12.23)
By Remark 2.6 the zero solution of the differential equation ż(t) = Az(t) is
asymptotically stable if and only if conditions (2.62) hold. By Corollary 2.3 con-
ditions (2.66) are sufficient conditions for the asymptotic mean-square stability of
the zero solution of (12.23). Below, we suppose that conditions (2.62) and (2.66)
hold.
To get stability conditions for (12.20), consider the matrix equation

A P + P A + Pσ = −C, (12.24)

where
     
p p12 p σ2 0 c 0
P = 11 , Pσ = 11 1 , C= ,
p12 p22 0 p22 σ22 0 1

c > 0, and the matrix A is defined in (12.22), (12.19).

If the matrix equation (12.24) has a positive definite solution P , then the function
v(z) = z P z is a Lyapunov function for (12.23) since
 
Lv = z A P + P A + Pσ z = −z Cz.

Note that the matrix equation (12.24) can be represented as the system of the
equations
2(p11 a11 + p12 a21 + p11 δ1 ) = −c,
2(p12 a12 + p22 a22 + p22 δ2 ) = −1, (12.25)
p11 a12 + p12 Tr(A) + p22 a21 = 0,
with the solution
c + 2a21 p12 1 + 2a12 p12 a21 â11 + ca12 â22
p11 = − , p22 = − , p12 = ,
2â11 2â22 2Z
(12.26)
where
1
âii = aii + δi , δi = σi2 , i = 1, 2, (12.27)
2
304 12 Stability of Some Social Mathematical Models with Delay Under Stochastic

Z = Tr(A)â11 â22 − a12 a21 (â11 + â22 ). (12.28)

Lemma 12.2 Let conditions (2.52), (2.56) hold, and let

â11 < 0, â22 < 0. (12.29)

Then the zero solution of (12.23) is asymptotically mean-square stable.

Proof It is enough to show that the matrix P = pij with the elements (12.26),
which are a solution of the matrix equation (12.24), is positive definite for an arbi-
2 . To this aim, note that by (2.62),
trary c > 0, i.e., p11 > 0, p22 > 0, p11 p22 > p12
(12.19), (12.29) we have Z < 0. Note also that by (12.27), (12.29), Remark 2.8, and
(2.72) we obtain
| Tr(A)| det(A) A1
δ1 < |a11 | ≤ ≤ ,
A2 | Tr(A)|
(12.30)
| Tr(A)| det(A) A2
δ2 < |a22 | ≤ ≤ ,
A1 | Tr(A)|
where
Ai = det(A) + aii2 , i = 1, 2. (12.31)
Besides, by (12.28), (12.27), (2.62), and (12.31) we have

Z + a12 a21 â22 = Tr(A)â11 â22 − a12 a21 (â11 + â22 ) + a12 a21 â22
 
= Tr(A)â22 − a12 a21 â11
 
= A2 − Tr(A) δ2 â11 (12.32)

and, similarly,
 
Z + a12 a21 â11 = A1 − Tr(A) δ1 â22 . (12.33)
From this and from (12.26), (12.32), (12.30) it follows that for an arbitrary c > 0,
cZ + a21 (ca12 â22 + a21 â11 )
p11 = −
2Z â11
c(Z + a12 a21 â22 ) + a21
2 â
11
=−
2Z â11
c(A2 − | Tr(A)|δ2 ) + a21
2
=
2|Z|
> 0, (12.34)

and, similarly, by (12.26), (12.33), (12.30) we obtain

A1 − | Tr(A)|δ1 + ca12
2
p22 = > 0. (12.35)
2|Z|
12.1 Mathematical Model of Alcohol Consumption 305

2 . Indeed, the inequality

Finally, let us show that p11 p22 > p12

(c + 2a21 p12 )(1 + 2a12 p12 ) 2

> p12
4â11 â22
2 − 2(a + ca )p < c by B = â â − a a > 0. Substi-
is equivalent to 4Bp12 21 12 12 11 22 12 21
tuting p12 from (12.26) into the obtained inequality, we have

B(a21 â11 + ca12 â22 )2 − (a21 + ca12 )(a21 â11 + ca12 â22 )Z < cZ 2

or
 
c2 a12
2
â22 (Z − B â22 ) + câ11 â22 Z Tr(A) − 2a12 a21 B + a21
2
â11 (Z − B â11 ) > 0.

Note also that by (12.28), (12.29), (12.19) we obtain


â11 (Z − B â11 ) = â11 Tr(A)â11 â22 − a12 a21 (â11 + â22 )

− (â11 â22 − a12 a21 )â11
 
= â11 Tr(A)â11 â22 − a12 a21 â22 − â11
2
â22
 
= â11 â22 Tr(A)â11 − a12 a21 − â11
2
  
= â11 â22 Tr(A) − â11 â11 − a12 a21
 
= â11 â22 (a22 − δ1 )â11 − a12 a21 > 0

and, similarly,
 
â22 (Z − B â22 ) = â11 â22 (a11 − δ2 )â22 − a12 a21 > 0,
Z Tr(A) − 2a12 a21 B > 0.

So, for an arbitrary c > 0, the matrix P with the elements (12.26) is positive definite.
The proof is completed. 

Theorem 12.1 If conditions (12.29) hold and, for some c > 0, the elements (12.26)
of the matrix P satisfy the condition
 2
βa ∗ |p12 − p22 | + 2βa ∗ |p11 − p12 | < c, (12.36)

then the solution (a ∗ , m∗ ) of system (12.16) is stable in probability.

Proof Note that the order of nonlinearity of system (12.16) is higher than one.
Therefore, from Sect. 5.3, to get conditions for stability in probability of the equilib-
rium point (a ∗ , m∗ ) of this system, it is enough to get conditions for the asymptotic
mean-square stability of the zero solution of the linear part (12.20) of this system.
Following the procedure of constructing Lyapunov functionals, we will construct a
Lyapunov functional for system (12.20) in the form V = V1 + V2 , where V1 = y  P y,
306 12 Stability of Some Social Mathematical Models with Delay Under Stochastic

y = (y1 , y2 ) , P is a positive definite solution of system (12.25) with the elements

(12.26), and V2 will be chosen below.
Let L be the generator (Sect. 2.1.2) of system (12.20). Then by (12.20) and
(12.25) we have
  
LV1 = 2 p11 y1 (t) + p12 y2 (t) a11 y1 (t) + a12 y2 (t) + βa ∗ I (y1t ) + p11 σ12 y12 (t)
  
+ 2 p12 y1 (t) + p22 y2 (t) a21 y1 (t) + a22 y2 (t) − βa ∗ I (y1t ) + p22 σ22 y22 (t)
 
= −cy12 (t) − y22 (t) + 2βa ∗ (p11 − p12 )y1 (t) + (p12 − p22 )y2 (t) I (y1t ).

By (12.7), (12.2) we have 2y1 (t)I (y1t ) ≤ y12 (t) + I (y1t

2 ) and 2y (t)I (y ) ≤
2 1t
2 −1
νy2 (t) + ν I (y1t ) for some ν > 0. Using these inequalities, we obtain
2

  2 
LV1 ≤ −cy12 (t) − y22 (t) + βa ∗ |p11 − p12 | y12 (t) + I y1t
  2 
+ βa ∗ |p12 − p22 | νy22 (t) + ν −1 I y1t
   
= βa ∗ |p11 − p12 | − c y12 (t) + βa ∗ |p12 − p22 |ν − 1 y22 (t)
 2
+ qI y1t , (12.37)

where
 
q = βa ∗ |p11 − p12 | + |p12 − p22 |ν −1 . (12.38)
Putting
 ∞ t
V2 = q y12 (θ ) dθ dK(s),
0 t−s

by (12.2), (12.7) we have LV2 = q(y12 (t) − I (y1t

2 )). Therefore, by (12.37), (12.38)

for the functional V = V1 + V2 , we have

 
LV ≤ 2βa ∗ |p11 − p12 | + βa ∗ |p12 − p22 |ν −1 − c y12 (t)
 
+ βa ∗ |p12 − p22 |ν − 1 y22 (t).

Thus, if

2βa ∗ |p11 − p12 | + βa ∗ |p12 − p22 |ν −1 < c, βa ∗ |p12 − p22 |ν < 1, (12.39)

then by Remark 2.1 the zero solution of (12.20) is asymptotically mean-square sta-
ble.
From (12.39) it follows that
βa ∗ |p12 − p22 | 1
∗
<ν< ∗
. (12.40)
c − 2βa |p11 − p12 | βa |p12 − p22 |
Thus, if for some c > 0, condition (12.36) holds, then there exists ν > 0 such that
conditions (12.40) (or (12.39)) hold too, and therefore the zero solution of (12.20)
12.1 Mathematical Model of Alcohol Consumption 307

is asymptotically mean-square stable. From this it follows also that the zero solu-
tion of (12.18) and therefore the equilibrium point of system (12.16) are stable in
probability. The proof is completed. 

Example 12.2 Consider system (12.16) with the values of the parameters α, β, γ , μ,
d, dA and the equilibrium point (a ∗ , m∗ ) given in (12.14), (12.15). As an example,
consider the levels of noises σ1 = 0.028969, σ2 = 0.142252 or δ1 = 0.000420, δ2 =
0.010118. From (12.19) it follows that the values of system (12.20) parameters are
a11 = −0.029245, a12 = −0.001440, a21 = 0.017446, a22 = −0.010475 and the
conditions (12.29) hold: â11 = −0.028825 < 0, â22 = −0.000357 < 0. Put c = 20.
Then by (12.26) p11 = 477.4438, p12 = 215.6615, p22 = 530.4124, and condition
(12.36) holds:
 ∗ 2
βa |p12 − p22 | + 2βa ∗ |p11 − p12 | = 16.1036 < 20.

Thus, the solution (a ∗ , m∗ ) of system (12.16) is stable in probability.

Example 12.3 Consider system (12.16) with the previous values of the all pa-
rameters except for the levels of noises that are σ1 = 0.0075, σ2 = 0.0077 or
δ1 = 0.000028, δ2 = 0.000030. These values of σ1 and σ2 are selected taking into
account sample errors of the monitoring of the alcohol consumption in Spain [291].
The parameters a11 , a21 , a22 are the same as in the previous example, and condi-
tions (12.29) hold: â11 = −0.029217 < 0, â22 = −0.010445 < 0. Put c = 4. Then
by (12.26) p11 = 78.6856, p12 = 17.1347, p22 = 45.5060, and condition (12.36)
holds:
 ∗ 2
βa |p12 − p22 | + 2βa ∗ |p11 − p12 | = 1.3643 < 4.
Thus, the solution (a ∗ , m∗ ) of system (12.16) is stable in probability.

Let us now get three corollaries from Theorem 12.1 that simplify a verification
of the stability condition (12.36). By (12.26) and (12.28) we have

c + 2a21 p12
p12 − p11 = p12 +
2â11
 
a21 a21 â11 + ca12 â22 c
= 1+ +
â11 2Z 2â11
(a21 + â11 )a12 â22 + Z (a21 + â11 )a21
= c+
2Z â11 2Z
= B0 c + B1 , (12.41)

where
(Tr(A) + a12 )â22 − a12 a21 (a21 + â11 )a21
B0 = , B1 = (12.42)
2Z 2Z
308 12 Stability of Some Social Mathematical Models with Delay Under Stochastic

and, similarly,

1 + 2a12 p12
p12 − p22 = p12 +
2â22
 
a12 a21 â11 + ca12 â22 1
= 1+ +
â22 2Z 2â22
(a12 + â22 )a12 c (a12 + â22 )a21 â11 + Z
= +
2Z 2Z â22
= D 0 c + D1 , (12.43)

where

(a12 + â22 )a12 (Tr(A) + a21 )â11 − a12 a21

D0 = , D1 = . (12.44)
2Z 2Z

Remark 12.2 Put

 
 2 D1 2 B1
f (c) = βa ∗ D0 c+ + 2βa ∗ |B0 | c + − c. (12.45)
D0 B0

From (12.19), (12.29) it follows that B0 < 0. By (12.41)–(12.45) and B0 < 0 condi-
tion (12.36) is equivalent to the condition f (c) < 0.
Put now
 
 ∗ 2 D1 B1 2 B1
S = βa D0 − + ,
D0 B0 B0
 
1 − 2βa ∗ |B0 | D1 B1
R+ = 2βa ∗ |B0 | − + ,
2(βa ∗ D0 )2 D0 B0
  (12.46)
∗ 1 + 2βa ∗ |B0 | D1 B1
R− = −2βa |B0 | − + ,
2(βa ∗ D0 )2 D0 B0
1 D1 B02
Q= − − .
4(βa ∗ D0 )2 D0 D02

Corollary 12.1 If conditions (12.29) hold and S < 0, then the solution (a ∗ , m∗ ) of
system (12.16) is stable in probability.

Proof From S < 0 and B0 < 0 it follows that B1 > 0. Putting c0 = −B1 B0−1 > 0,
we obtain f (c0 ) = S < 0, i.e., condition (12.36) holds. The proof is completed. 

Corollary 12.2 If conditions (12.29) hold and 0 ≤ R+ < Q, then the solution
(a ∗ , m∗ ) of system (12.16) is stable in probability.
12.1 Mathematical Model of Alcohol Consumption 309

Proof Let us suppose that c + B1 B0−1 ≥ 0. Then the minimum of the function f (c)
is reached by
1 − 2βa ∗ |B0 | D1 B1
c0 = ∗ 2
− ≥− .
2(βa D0 ) D0 B0
Substituting c0 into the function f (c), we obtain that the condition f (c0 ) < 0 is
equivalent to the condition 0 ≤ R+ < Q. The proof is completed. 

Corollary 12.3 If conditions (12.29) hold and 0 < R− < Q, then the solution
(a ∗ , m∗ ) of system (12.16) is stable in probability.

Proof Let us suppose that c + B1 B0−1 < 0. Then the minimum of the function f (c)
is reached by
1 + 2βa ∗ |B0 | D1 B1
c0 = ∗ 2
− <− .
2(βa D0 ) D0 B0
Substituting c0 into the function f (c), we obtain that the condition f (c0 ) < 0 is
equivalent to the condition 0 < R− < Q. The proof is completed. 

Example 12.4 Consider system (12.16) with the values of the parameters from Ex-
ample 12.2. Calculating S, R+ , Q, we obtain: S = 4.50 > 0, R+ = 736 < Q =
1320. From Corollary 12.2 it follows that the solution (a ∗ , m∗ ) of system (12.16) is
stable in probability.

Example 12.5 Consider system (12.16) with the values of the parameters from Ex-
ample 12.3. Calculating S, R+ , Q, we obtain: S = −0.39 < 0, R+ = 2462 < Q =
4754. From both Corollary 12.1 and Corollary 12.2 it follows that the solution
(a ∗ , m∗ ) of system (12.16) is stable in probability.

12.1.6 Numerical Simulation

Let us suppose that in (12.1) dK(s) = δ(s − h) ds, where δ(s) is Dirac’s delta-
function, and h ≥ 0 is the delay.
In Fig. 12.1 25 trajectories of the solution of (12.16), (12.4) are shown for the
values of the parameters from Examples 12.1 and 12.2: α = 0.000110247, β =
0.0284534, γ = 0.00144, μ = 0.01, d = 0.009, dA = 0.008, the initial values a0 =
0.43, m0 = 0.53, r0 = 0.04, the levels of noises σ1 = 0.028969, σ2 = 0.142252,
and delay h = 0.1. We can see that all trajectories go to the equilibrium point a ∗ =
0.364739, m∗ = 0.629383, r ∗ = 0.00587794.
Note that numerical simulations of the processes a(t), m(t), and r(t) were ob-
tained via the difference analogues of (12.16), (12.4) in the form
  
ai+1 = ai + μ + γ − γ μi + βai ai−m − ai β + μ + γ − (d − dA )(1 − ai )
 
+ σ1 ai − a ∗ (w1,i+1 − w1i ),
310 12 Stability of Some Social Mathematical Models with Delay Under Stochastic

Fig. 12.1 25 trajectories of the processes a(t) (blue), m(t) (green), r(t) (red) with the val-
ues of the parameters α = 0.000110247, β = 0.0284534, γ = 0.00144, μ = 0.01, d = 0.009,
dA = 0.008, the levels of noises σ1 = 0.028969, σ2 = 0.142252, the delay h = 0.1, the ini-
tial values a(s) = 0.43, s ∈ [−0.1, 0], m(0) = 0.53, r(0) = 0.04, and the equilibrium point
a ∗ = 0.364739, m∗ = 0.629383, r ∗ = 0.00587794

  
mi+1 = mi + βai − βai ai−m − mi α + μ + (d − dA )ai
 
+ σ2 mi − m∗ (w2,i+1 − w2i ),
ri+1 = 1 − ai+1 − mi+1 ,
i = 0, 1, 2, . . . , aj = a0 , j = −m, . . . , −1, 0.
Here is the discretization step (which was chosen as = 0.01), ai = a(ti ),
mi = m(ti ), ri = r(ti ), wki = wk (ti ), k = 1, 2, ti = i , m = h −1 , trajectories of
the Wiener processes w1 (t) and w2 (t) are simulated by the algorithm described in
Sect. 2.1.1.

12.2 Mathematical Model of Social Obesity Epidemic

Social obesity epidemic models are popular enough with researches (see, for in-
stance, [14, 34, 42, 51, 68, 115, 227, 255]). Here the known nonlinear social obesity
12.2 Mathematical Model of Social Obesity Epidemic 311

epidemic model [255] is generalized to the system with distributed delay. It is sup-
posed also that this nonlinear system is exposed to additive stochastic perturbations
of white noise type that are directly proportional to the deviation of the system state
from the equilibrium point. The research that is similar to the previous one is applied
to this model.

12.2.1 Description of the Considered Model

For constructing the mathematical obesity model [255] the 24- to 65-year-old pop-
ulation is divided into three subpopulations based on the so-called body mass in-
dex (BMI = Weight/Height2 ). The classes or subpopulations are: individuals at a
normal weight (BMI < 25 kg/m2 ) N (t), people who are overweight (25 kg/m2 ≤
BMI < 30 kg/m2 ) S(t), and obese individuals (BMI ≥ 30 kg/m2 ) O(t).
The transition between the different subpopulations is determined as follows:
once an adult starts an unhealthy lifestyle, he/she becomes addicted to the unhealthy
lifestyle and starts a progression to being overweight S(t) because of this lifestyle.
If this adult continues with his/her unhealthy lifestyle, he/she can become an obese
individual O(t). In both these classes individuals can stop his/her unhealthy lifestyle
and then move to classes N (t) and S(t), respectively.
The transitions between the subpopulations N (t), S(t), and O(t) are governed
by terms proportional to the sizes of these subpopulations. Conversely, the transi-
tions from normal to overweight occur through the transmission of an unhealthy
lifestyle from the overweight and obese subpopulations to the normal-interactions
weight subpopulation, depending on the meet population, depending on the meet-
ings among them. This transition is modeled using the term
 ∞
 
βN (t) S(t − s) + O(t − s) dK(s),
0

where K(s) is a nondecreasing function that satisfies condition (12.2), and the inte-
gral is understood in the Stieltjes sense. The subpopulations’ sizes and their behav-
iors with time determine the dynamic evolution of adulthood excess weight.
Without loss of generality and for the sake of clarity, the 24- to 65-year-old adult
population is normalized to unity, and it is supposed for all t ≥ 0 that

N (t) ≥ 0, S(t) ≥ 0, O(t) ≥ 0, (12.47)

N (t) + S(t) + O(t) = 1. (12.48)

Thus, under the above assumptions, the following nonlinear system of integro–
differential equations is obtained:
 ∞
 
Ṅ(t) = μN0 − μN(t) − βN (t) S(t − s) + O(t − s) dK(s) + ρS(t),
0
312 12 Stability of Some Social Mathematical Models with Delay Under Stochastic
 ∞ 
Ṡ(t) = μS0 + βN(t) S(t − s) + O(t − s) dK(s)
0
(12.49)
− (μ + γ + ρ)S(t) + εO(t),
Ȯ(t) = μO0 + γ S(t) − (μ + ε)O(t), t ≥ 0,
N(0) = N0 , S(s) = S0 , O(s) = O0 , s ≤ 0.
The time-invariant parameters of this system of equations are:
ε the rate at which an obese adult with a healthy lifestyle becomes an overweight
individual (intensity of transition from the group O(t) to the group S(t)).
ρ the rate at which an overweight individual moves to the normal-weight subpop-
ulation (intensity of transition from the group S(t) to the group N (t)).
β the transmission rate because of social pressure to adopt an unhealthy lifestyle,
e.g., TV, friends, family, job, and so on (intensity of transition from the group
N(t) to the group S(t)).
γ the rate at which an overweight 24- to 65-year-old adult becomes an obese indi-
vidual because of unhealthy lifestyle (intensity of transition from the group S(t)
into the group O(t)); so, the scheme of transition from one group to another one
is
O(t) →ε → S(t) →ρ → N (t) →β → S(t) →γ → O(t).
μ the average stay time in the system of 24- to 65-year-old adults (μ = 1/
(65 years – 24 years) · 52 weeks/year).
N0 the proportion of normal weight coming from the 23-year age group.
S0 the proportion of overweight coming from the 23-year age group.
O0 the proportion of obese coming from the 23-year age group.
Here the parameters ε, ρ, β, γ , μ are nonnegative numbers, and N0 , S0 , O0
satisfy the conditions of type (12.47), (12.48).
By condition (12.48) system (12.49) can be simplified to the following system of
two equations:
 ∞
 
Ṅ(t) = μN0 − μN(t) − βN(t) 1 − N (t − s) dK(s) + ρS(t),
0
 ∞ 
Ṡ(t) = μS0 + βN(t) 1 − N (t − s) dK(s) − (μ + γ + ρ)S(t) (12.50)
0
 
+ ε 1 − N (t) − S(t) , t ≥ 0,
N(s) = N0 , s ≤ 0, S(0) = S0 .

12.2.2 Existence of an Equilibrium Point

The equilibrium point (N ∗ , S ∗ ) of system (12.50) is defined by the conditions

Ṅ(t) = 0, Ṡ(t) = 0 and by (12.50), (12.47) is a solution of the following system
12.2 Mathematical Model of Social Obesity Epidemic 313

of algebraic equations:
 
μN0 − μN ∗ − βN ∗ 1 − N ∗ + ρS ∗ = 0,
    (12.51)
μS0 + βN ∗ 1 − N ∗ − (μ + γ + ρ)S ∗ + ε 1 − S ∗ − N ∗ = 0.

From (12.51) it follows that

    
S ∗ = ρ −1 μ N ∗ − N0 + βN ∗ 1 − N ∗ ,
    (12.52)
S ∗ = kρ −1 μS0 + ε + βN ∗ 1 − N ∗ ,

where
k = ρ(μ + γ + ρ + ε)−1 < 1. (12.53)
By (12.52), (12.53) we obtain that N ∗ is a root of the quadratic equation
 2  
β(1 − k) N ∗ − μ + kε + β(1 − k) N ∗ + μ(N0 + kS0 ) + kε = 0. (12.54)

Lemma 12.3 Assume that N0 + kS0 < 1. If β > 0, then (12.54) has two real roots,
N1∗ ∈ (0, 1) and N2∗ > 1. If β = 0 and μkε > 0, then (12.54) has one root N ∗ ∈
(N0 + kS0 , 1).

Proof From N0 + kS0 < 1 and β > 0 we have

 2  
D= μ + kε + β(1 − k) − 4β(1 − k) μ(N0 + kS0 ) + kε

 2
> μ + kε + β(1 − k) − 4β(1 − k)(μ + kε)

= μ + kε − β(1 − k) , (12.55)

i.e., D > |μ + kε − β(1 − k)| ≥ 0, and therefore the quadratic equation (12.54) has
two real roots
μ + kε + β(1 − k) − D μ + kε + β(1 − k) + D
N1∗ = , N2∗ = . (12.56)
2β(1 − k) 2β(1 − k)

If μ + kε < β(1 − k), then

μ + kε
N1∗ < < 1, N2∗ > 1.
β(1 − k)

If μ + kε ≥ β(1 − k), then

μ + kε
N1∗ < 1, N2∗ > ≥ 1.
β(1 − k)
314 12 Stability of Some Social Mathematical Models with Delay Under Stochastic

If β = 0, then from (12.54) it follows that

μ(N0 + kS0 ) + kε
1 > N∗ = > N0 + kS0 .
μ + kε
The proof is completed. 

Lemma 12.4 Assume that N0 = 1. If μ + kε < β(1 − k), then (12.54) has two roots
on the interval (0,1]: N1∗ ∈ (0, 1) and N2∗ = 1. If μ + kε ≥ β(1 − k) then (12.54) has
one root only on the interval (0,1]: N1∗ = 1.

Proof From N0 = 1 and (12.47) we have S0 = 0. Then, similarly to (12.55), D =

|μ + kε − β(1 − k)|. If μ + kε < β(1 − k), then D = β(1 − k) − (μ + kε), and by
(12.56) we obtain
μ + kε
N1∗ = < 1, N2∗ = 1.
β(1 − k)
If μ + kε > β(1 − k), then D = μ + kε − β(1 − k), and by (12.56) we have
μ + kε
N1∗ = 1, N2∗ = > 1.
β(1 − k)
If μ + kε = β(1 − k), then D = 0 and N1∗ = N2∗ = 1. The proof is completed. 

Example 12.6 Following [255], put

μ = 0.000469, γ = 0.0003, ε = 0.000004, ρ = 0.000035,

β = 0.00085, N0 = 0.704, S0 = 0.25, O0 = 0.046.

Then by (12.56), (12.52), (12.48) we obtain

N ∗ = 0.3311, S ∗ = 0.3814, O ∗ = 0.2875.

Putting β = 0 with the same values of the other parameters, by Lemma 12.3 we
obtain

N ∗ = 0.7149 > N0 + kS0 = 0.7148, S ∗ = 0.1465, O ∗ = 0.1386.

Put now N0 = 1, S0 = O0 = 0. By Lemma 12.4, if β = 0.00085, i.e., if β > 0, then

β > (μ + kε)(1 − k)−1 = 0.00049 and N ∗ = 0.5770, S ∗ = 0.2588, O ∗ = 0.1642. If
β = 0, then N ∗ = 1, S ∗ = O ∗ = 0.

12.2.3 Stochastic Perturbations, Centralization, and Linearization

Let us suppose that system (12.50) is exposed to stochastic perturbations of white

noise type (ẇ1 (t), ẇ2 (t)) that are directly proportional to the deviation of system
12.2 Mathematical Model of Social Obesity Epidemic 315

(12.50) state (N (t), S(t)) from the equilibrium point (N ∗ , S ∗ ), i.e.,

 ∞ 
Ṅ(t) = μN0 − μN(t) − βN (t) 1 − N (t − s) dK(s) + ρS(t)
0
 ∗

+ σ1 N (t) − N ẇ1 (t), t ≥ 0,
 ∞
  (12.57)
Ṡ(t) = μS0 + βN (t) 1 − N (t − s) dK(s) − (μ + γ + ρ)S(t)
0
   
+ ε 1 − N (t) − S(t) + σ2 S(t) − S ∗ ẇ2 (t), t ≥ 0,
N(s) = N0 , s ≤ 0, S(0) = S0 .

Here w1 (t), w2 (t) are mutually independent standard Wiener processes, and the
stochastic differential equations of system (12.57) are understood in the Itô sense
(Sect. 2.1.2). Note that the equilibrium point (N ∗ , S ∗ ) of system (12.50) is a solution
of (12.57) too.
To centralize system (12.57) at the equilibrium point, put now x1 = N − N ∗ ,
x2 = S − S ∗ . Then by (12.57), (12.53) we have

ẋ1 (t) = a11 x1 (t) + a12 x2 (t) + βN ∗ I (x1t ) + βx1 (t)I (x1t )
+ σ1 x1 (t)ẇ1 (t),
(12.58)
ẋ2 (t) = a21 x1 (t) + a22 x2 (t) − βN ∗ I (x1t ) − βx1 (t)I (x1t )
+ σ2 x2 (t)ẇ2 (t),

where
 
a11 = −μ − β 1 − N ∗ , a12 = ρ,
 
a21 = −ε + β 1 − N ∗ , a22 = −k −1 ρ,
(12.59)
 ∞
I (x1t ) = x1 (t − s) dK(s).
0

Example 12.7 Using the values of the parameters from Example 12.6, by (12.59)
we obtain
a11 = −0.0010376, a12 = 0.000035,
a21 = 0.0005646, a22 = −0.000808.
It is clear that the stability of the equilibrium point of system (12.57) is equiva-
lent to the stability of the zero solution of (12.58). Rejecting the nonlinear terms in
(12.58), we obtain the linear part of system (12.58)

ẏ1 (t) =a11 y1 (t) + a12 y2 (t) + βN ∗ I (y1t ) + σ1 y1 (t)ẇ1 (t),

(12.60)
ẏ2 (t) =a21 y1 (t) + a22 y2 (t) − βN ∗ I (y1t ) + σ2 y2 (t)ẇ2 (t).
316 12 Stability of Some Social Mathematical Models with Delay Under Stochastic

12.2.4 Stability of an Equilibrium Point

Note that the nonlinear system (12.58) has the order of nonlinearity higher than one.
Thus, sufficient conditions for the asymptotic mean-square stability of the zero so-
lution of the linear part (12.60) at the same time are (Sect. 5.3) sufficient conditions
for the stability in probability of the zero solution of the nonlinear system (12.58)
and therefore are sufficient conditions for the stability in probability of the solution
(N ∗ , S ∗ ) of system (12.57).
To get sufficient conditions for the asymptotic mean-square stability of the zero
solution of (12.60), rewrite this system in the form
 
ẏ(t) = Ay(t) + B(yt ) + σ y(t) ẇ(t), (12.61)

where
   
y(t) = y1 (t), y2 (t) , w(t) = w1 (t), w2 (t) ,
 
B(yt ) = βN ∗ I (y1t ), −βN ∗ I (y1t ) , (12.62)
   
a a12   σ y (t) 0
A = 11 , σ y(t) = 1 1 ,
a21 a22 0 σ2 y2 (t)

and aij , i, j = 1, 2, are defined by (12.59).

Following the procedure of constructing Lyapunov functionals, for stability in-
vestigation of (12.61), consider the auxiliary differential equation without memory
 
ż(t) = Az(t) + σ z(t) ẇ(t). (12.63)

Remark 12.3 By (12.62), (12.59) for the matrix A, conditions (2.62) hold:
  
Tr(A) = − μ + ρk −1 + β 1 − N ∗ < 0,
   (12.64)
det(A) = ρk −1 μ + kε + β(1 − k) 1 − N ∗ > 0.

Example 12.8 Using the values of the parameters from Example 12.6, we have

Tr(A) = −0.0018456, det(A) = 0.0000008.

Consider âii , i = 1, 2, defined in (12.27).

Lemma 12.5 If
a21 ≤ 0 (12.65)
and
â11 < 0, â22 < 0, (12.66)
then the zero solution of (12.63) is asymptotically mean-square stable.
12.2 Mathematical Model of Social Obesity Epidemic 317

Proof By (12.59), (12.65) the matrix A from (12.63) satisfies the condition
a12 a21 ≤ 0. By (12.19) the same condition is satisfied by the matrix A from (12.23).
So, further, the proof coincides with that of Lemma 12.2. 

Lemma 12.6 If
a21 > 0 (12.67)
and
det(A)
max(δ1 , δ2 ) < , (12.68)
| Tr(A)|
then the zero solution of (12.63) is asymptotically mean-square stable.

Proof Similarly to Lemma 12.2, it is enough to show that the matrix P = pij
with the elements (12.26) is positive definite.
Note that by (12.31), (12.59), (12.67) we have

Ai = a11 a22 − a12 a21 + aii2 ≤ aii Tr(A), i = 1, 2.

From this and from (12.59), (12.64), (12.68) it follows that

det(A) Ai
δi < ≤ ≤ |aii |, i = 1, 2. (12.69)
| Tr(A)| | Tr(A)|
By (12.27), (12.28), (12.59), (12.64), (12.67), (12.68) we have
 
Z = Tr(A)(a11 + δ1 )(a22 + δ2 ) − a12 a21 Tr(A) + δ1 + δ2
= Tr(A) det(A) + Tr(A)δ1 a22 + Tr(A)δ2 a11
+ Tr(A)δ1 δ2 − a12 a21 (δ1 + δ2 )

= − Tr(A) det(A) + A2 δ1 + A1 δ2 − Tr(A) δ1 δ2
det(A)
< − Tr(A) det(A) + (A1 + A2 )
| Tr(A)|
 2  det(A)
= − Tr(A) − A1 − A2
| Tr(A)|
2a12 a21 det(A)
=−
| Tr(A)|
< 0. (12.70)

From (12.34), (12.35), (12.69), (12.70) we obtain that p11 > 0, p22 > 0 for arbitrary
c > 0.
2 . Indeed, by (12.26), (12.34), (12.35) this inequal-
Let us show that p11 p22 > p12
ity takes the form
    
c A2 − Tr(A) δ2 + a21 2
A1 − Tr(A) δ1 + ca122
> (ca12 â22 + a21 â11 )2 ,
318 12 Stability of Some Social Mathematical Models with Delay Under Stochastic

which is equivalent to the condition

 
c2 a12
2
det(A) − Tr(A) δ2 + a22 2
− â22
2
  
+ c A1 − Tr(A) δ1 det(A) − Tr(A) δ2
 
+ a222
det(A) − Tr(A) δ1
 2 
+ det(A) + 2a12 a21 (a11 a22 − â11 â22 )
 
+ a212
det(A) − Tr(A) δ1 + a11
2
− â11
2

> 0. (12.71)

By (12.67), (12.68), (12.69) and |aii | ≥ |âii |, i = 1, 2, condition (12.71) holds for
arbitrary c > 0. So, for arbitrary c > 0, the matrix P with the elements (12.26) is
positive definite. The proof is completed. 

Remark 12.4 If condition (12.65) holds, i.e., a21 ≤ 0, then from (12.59) and from
the proofs of Lemmas 12.3 and 12.4 it follows that β ∈ [0, (μ + ε)(1 − k)−1 ]. On
the other hand, if β > (μ + ε)(1 − k)−1 , then condition (12.67) holds, i.e., a21 > 0.
For example, by the values of the parameters from Example 12.6 we have β =
0.00085 > (μ + ε)(1 − k)−1 = 0.0004945 and a21 = 0.0005646 > 0.

Theorem 12.2 If conditions (12.65), (12.66) or (12.67), (12.68) hold and if, for
some c > 0, the elements (12.26) of the matrix P satisfy the condition
 2
βN ∗ |p12 − p22 | + 2βN ∗ |p11 − p12 | < c, (12.72)

then the solution (N ∗ , S ∗ ) of system (12.57) is stable in probability.

Proof Note that the stability in probability of the solution (N ∗ , S ∗ ) of system

(12.57) is equivalent to the stability in probability of the zero solution of system
(12.58) and the order of nonlinearity of system (12.58) is higher than one. So, to get
for this system conditions for stability in probability, it is enough (Sect. 5.3) to get
conditions for the asymptotic mean-square stability of the zero solution of the lin-
ear part (12.60) of this system. Following the procedure of constructing Lyapunov
functionals, we will construct a Lyapunov functional for system (12.60) in the form
V = V1 + V2 , where V1 = y  P y, y = (y1 , y2 ) , P is the positive definite solution of
system (12.25) with the elements (12.26), and V2 will be chosen below.
Let L be the generator of system (12.60). Then by (12.60), (12.25) we have
  
LV1 = 2 p11 y1 (t) + p12 y2 (t) a11 y1 (t) + a12 y2 (t) + βN ∗ I (y1t ) + p11 σ12 y12 (t)
  
+ 2 p12 y1 (t) + p22 y2 (t) a21 y1 (t) + a22 y2 (t) − βN ∗ I (y1t )
+ p22 σ22 y22 (t)
 
= −cy12 (t) − y22 (t) + 2βN ∗ (p11 − p12 )y1 (t) + (p12 − p22 )y2 (t) I (y1t ).
12.2 Mathematical Model of Social Obesity Epidemic 319

By (12.2), (12.59) we have 2y1 (t)I (y1t ) ≤ y12 (t) + I (y1t

2 ) and 2y (t)I (y ) ≤
2 1t
2 −1
νy2 (t) + ν I (y1t ) for some ν > 0. Using these inequalities, we obtain
2

  2 
LV1 ≤ −cy12 (t) − y22 (t) + βN ∗ |p11 − p12 | y12 (t) + I y1t
  2 
+ βN ∗ |p12 − p22 | νy22 (t) + ν −1 I y1t
   
= βN ∗ |p11 − p12 | − c y12 (t) + βN ∗ |p12 − p22 |ν − 1 y22 (t)
 2
+ qI y1t , (12.73)

where
 
q = βN ∗ |p11 − p12 | + |p12 − p22 |ν −1 . (12.74)
Putting
 ∞ t
V2 = q y12 (θ ) dθ dK(s),
0 t−s

by (12.2), (12.59) we get LV2 = q(y12 (t) − I (y1t

2 )). Therefore, by (12.73), (12.74),

for the functional V = V1 + V2 , we have

 
LV ≤ 2βN ∗ |p11 − p12 | + βN ∗ |p12 − p22 |ν −1 − c y12 (t)
 
+ βN ∗ |p12 − p22 |ν − 1 y22 (t).

Thus, if

2βN ∗ |p11 − p12 | + βN ∗ |p12 − p22 |ν −1 < c, βN ∗ |p12 − p22 |ν < 1, (12.75)

then by Remark 2.1 the zero solution of (12.60) is asymptotically mean-square sta-
ble.
From (12.75) it follows that

βN ∗ |p12 − p22 | 1
<ν< . (12.76)
c − 2βN ∗ |p11 − p12 | βN ∗ |p12 − p22 |

Thus, if for some c > 0, condition (12.72) holds, then there exists ν > 0 such that
conditions (12.76) (or (12.75)) hold too, and therefore the zero solution of (12.60)
is asymptotically mean-square stable. From this it follows that the zero solution of
(12.58) and therefore the equilibrium point (N ∗ , S ∗ ) of system (12.57) is stable in
probability. The proof is completed. 

Example 12.9 Consider system (12.50) with the values of the parameters ε, μ, ρ,
β, γ and the equilibrium point (N ∗ , S ∗ ) given in Example 12.6. As an example,
consider the levels of noises σ1 = 0.028256, σ2 = 0.029031. From (12.27) it follows
that δ1 = 0.0003992, δ2 = 0.0004214 and condition (12.68) holds: max(δ1 , δ2 ) <
det(A)| Tr(A)|−1 = 0.0004436.
320 12 Stability of Some Social Mathematical Models with Delay Under Stochastic

Put c = 10. Then by (12.26) p11 = 8335.7, p12 = 569.4, p22 = 1344.7, and con-
dition (12.72) holds:
 ∗ 2
βN |p12 − p22 | + 2βN ∗ |p11 − p12 | = 4.419 < 10.

Thus, the solution (N ∗ , S ∗ ) of system (12.57) is stable in probability.

Using the representations (12.41)–(12.44), we get three corollaries from Theo-

rem 12.2, which simplify a verification of the stability condition (12.72).
Put
 
 2 D1 2 B1
f (c) = βN ∗ D0 c+ + 2βN ∗ |B0 | c + − c, (12.77)
D0 B0
 
 2 D1 B1 2 B1
S = βN ∗ D0 − + ,
D0 B0 B0
 
∗ 1 − 2βN ∗ |B0 | D1 B1
R+ = 2βN |B0 | − + ,
2(βN ∗ D0 )2 D0 B0
  (12.78)
1 + 2βN ∗ |B0 | D1 B1
R− = −2βN ∗ |B0 | − + ,
2(βN ∗ D0 )2 D0 B0

1 D1 B02
Q= − − ,
4(βN ∗ D0 )2 D0 D02

where B0 , B1 , D0 , D1 are defined by (12.41)–(12.44). So, condition (12.72) is

equivalent to the condition f (c) < 0.

Corollary 12.4 If conditions (12.65), (12.66) or (12.67), (12.68) hold and S < 0,
then the solution (N ∗ , S ∗ ) of system (12.57) is stable in probability.

Proof By (12.78) from S < 0 it follows that B1 B0−1 < 0. Substituting c0 =

−B1 B0−1 > 0 into (12.77), we obtain f (c0 ) = S < 0, i.e., condition (12.72) holds.
The proof is completed. 

Corollary 12.5 If conditions (12.65), (12.66) or (12.67), (12.68) hold and 0 ≤

R+ < Q, then the solution (N ∗ , S ∗ ) of system (12.57) is stable in probability.

Corollary 12.6 If conditions (12.65), (12.66) or (12.67), (12.68) hold and 0 <
R− < Q, then the solution (N ∗ , S ∗ ) of system (12.57) is stable in probability.

The proofs of Corollaries 12.5 and 12.6 are similar to those of Corollaries 12.2
and 12.3.
12.2 Mathematical Model of Social Obesity Epidemic 321

Fig. 12.2 25 trajectories of the processes N(t) (blue), S(t) (green), O(t) (red) with the values of
the parameters μ = 0.000469, γ = 0.0003, ε = 0.000004, ρ = 0.000035, β = 0.00085, h = 0.1,
δ1 = 0.0003992, δ2 = 0.0004214, the initial values N(s) = 0.704, s ∈ [−0.1, 0], S(0) = 0.25,
O(0) = 0.046, and the equilibrium point N ∗ = 0.3311, S ∗ = 0.3814, O ∗ = 0.2875

Example 12.10 Consider system (12.57) with the values of the parameters from Ex-
ample 12.6 and δ1 = 0.0003992, δ2 = 0.0002661. Calculating S, R+ , Q by (12.78),
we obtain: S = −0.0100916 < 0, R+ = 7499 < Q = 18161. By both Corollaries
12.4 and 12.5 the solution (N ∗ , S ∗ ) of system (12.57) is stable in probability.

Example 12.11 Consider system (12.57) with the values of the parameters from Ex-
ample 12.6 and δ1 = 0.0003992, δ2 = 0.0004214. Calculating S, R+ , Q by (12.78),
we obtain: S = 0.0051611 > 0, R+ = 7811 < Q = 18914. The condition of Corol-
lary 12.4 does not hold, but from Corollary 12.5 it follows that the solution (N ∗ , S ∗ )
of system (12.57) is stable in probability.

12.2.5 Numerical Simulation

Let us suppose that in (12.49) dK(s) = δ(s − h) ds, where δ(s) is Dirac’s delta-
function, and h ≥ 0 is a delay.
322 12 Stability of Some Social Mathematical Models with Delay Under Stochastic

Fig. 12.3 25 trajectories of the processes N(t) (blue), S(t) (green), O(t) (red) with the values
of the parameters μ = 0.000469, γ = 0.0003, ε = 0.000004, ρ = 0.000035, β = 0, h = 0.1,
δ1 = 0.0003992, δ2 = 0.0004214, the initial values N(s) = 0.704, s ∈ [−0.1, 0], S(0) = 0.25,
O(0) = 0.046, and the equilibrium point N ∗ = 0.7149, S ∗ = 0.1465, O ∗ = 0.1386

In Fig. 12.2 25 trajectories of the solution of (12.57), (12.48) are shown for the
values of the parameters from Examples 12.6 and 12.9: μ = 0.000469, γ = 0.0003,
ε = 0.000004, ρ = 0.000035, β = 0.00085, the initial values N0 = 0.704, S0 =
0.25, O0 = 0.046, the levels of noises σ1 = 0.028256, σ2 = 0.029031, and the delay
h = 0.1. One can see that all trajectories go to the equilibrium point N ∗ = 0.3311,
S ∗ = 0.3814, O ∗ = 0.2875.
Putting β = 0 with the same values of the other parameters, one can see that
in accordance with Example 12.6, all trajectories go to another equilibrium point
N ∗ = 0.7149, S ∗ = 0.1465, O ∗ = 0.1386 (Fig. 12.3).
Change now the initial values on N0 = 1, S0 = O0 = 0, and put again β =
0.00085. In accordance with Example 12.6, corresponding trajectories of the so-
lution go to the equilibrium point N ∗ = 0.5770, S ∗ = 0.2588, O ∗ = 0.1642
(Fig. 12.4).
Note that numerical simulations of the processes N (t), S(t), and O(t) were ob-
tained via the difference analogues of (12.57), (12.48) in the form
12.2 Mathematical Model of Social Obesity Epidemic 323

Fig. 12.4 25 trajectories of the processes N(t) (blue), S(t) (green), O(t) (red) with the values of
the parameters μ = 0.000469, γ = 0.0003, ε = 0.000004, ρ = 0.000035, β = 0.00085, h = 0.1,
δ1 = 0.0003992, δ2 = 0.0004214, the initial values N(s) = 1, s ∈ [−0.1, 0], S(0) = 0, O(0) = 0,
and the equilibrium point N ∗ = 0.5770, S ∗ = 0.2588, O ∗ = 0.1642

 
Ni+1 = Ni + μN0 − μNi − βNi (1 − Ni−m ) + ρSi
 
+ σ1 Ni − N ∗ (w1,i+1 − w1i ),
 
Si+1 = Si + μS0 + βNi Ni−m − (μ + γ + ρ)Si + ε(1 − Ni − Si )
 
+ σ2 Si − S ∗ (w2,i+1 − w2i ),
Oi+1 = 1 − Ni+1 − Si+1 ,
i = 0, 1, 2, . . . , Nj = N0 , j = −m, . . . , −1, 0.

Here is the discretization step (chosen as = 0.01), Ni = N (ti ), Si = S(ti ), Oi =

O(ti ), wki = wk (ti ), k = 1, 2, ti = i , m = h −1 , and trajectories of the Wiener
processes w1 (t) and w2 (t) are simulated by the algorithm described in Sect. 2.1.1.
References

1. Acheson DJ (1993) A pendulum theorem. Proc R Soc Lond Ser A, Math Phys Sci
443(1917):239–245
2. Acheson DJ, Mullin T (1993) Upside-down pendulums. Nature 366(6452):215–216
3. Agarwal RP, Grace SR (2000) Asymptotic stability of certain neutral differential equations.
Math Comput Model 31:9–15
4. Agarwal RP, O’Regan D, Wong PJY (1999) Positive solutions of differential, difference and
integral equations. Kluwer Academic, Dordrecht
5. Agarwal RP, Grace SR, O’Regan D (2000) Oscillation theory for difference and functional
differential equations. Kluwer Academic, Dordrecht
6. Alonso-Quesada S, De la Sen M, Agarwal R, Ibeas A (2012) An observer-based vaccina-
tion control law for a SEIR epidemic model based on feedback linearization techniques for
nonlinear systems. Adv Differ Equ 2012:161
7. Alzabut JO (2010) Almost periodic solutions for impulsive delay Nicholson’s blowflies pop-
ulation model. J Comput Appl Math 234:233–239
8. Alzabut JO, Bolat Y, Abdeljawad T (2012) Almost periodic dynamic of a discrete Nichol-
son’s blowflies model involving a linear harvesting term. Adv Differ Equ 2012:158
9. Andreyeva EA, Kolmanovskii VB, Shaikhet LE (1992) Control of hereditary systems.
Nauka, Moscow. pp 336 (in Russian)
10. Appleby JAD, Buckwar E (2010) A constructive comparison technique for determining the
asymptotic behaviour of linear functional differential equations with unbounded delay. Differ
Equ Dyn Syst 18(3):271–301
11. Appleby JAD, Krol K (2011) Long memory in a linear stochastic Volterra equation. J Math
Anal Appl 380(2):814–830
12. Appleby JAD, Reynolds DW (2004) On necessary and sufficient conditions for exponential
stability in linear Volterra integro–differential equations. J Integral Equ Appl 16:221–240
13. Appleby JAD, Gyori I, Reynolds DW (2006) On exact rates of decay of solutions of linear
systems of Volterra equations with delay. J Math Anal Appl 320:56–77
14. Aranceta J, Serra L, Foz M, Moreno B (2005) Prevalencia de obesidad en Espana. Méd Clin
125:460–466
15. Arino O, Elabdllaoui A, Mikaram J, Chattopadhyay J (2004) Infection on prey popula-
tion may act as a biological control in a ratio-dependent predator–prey model. Nonlinearity
17:1101–1116
16. Arnold L (1974) Stochastic differential equations. Wiley, New York
17. Arnold L (1998) Random dynamical systems. Springer, Berlin
18. Bailey NTJ (1957) The mathematical theory of epidemics. Griffin, London
19. Bandyopadhyay M, Chattopadhyay J (2005) Ratio dependent predator–prey model: effect of
environmental fluctuation and stability. Nonlinearity 18:913–936

L. Shaikhet, Lyapunov Functionals and Stability of Stochastic Functional 325

Differential Equations, DOI 10.1007/978-3-319-00101-2,
© Springer International Publishing Switzerland 2013
326 References

20. Bandyopadhyay M, Saha T, Pal R (2008) Deterministic and stochastic analysis of a delayed
allelopathic phytoplankton model within fluctuating environment. Nonlinear Anal Hybrid
Syst 2:958–970
21. Bedelbaev AK (1958) About construction of Lyapunov functions as a quadratic form and its
application to stability of controlled systems. Avtomatika 1:37–43 (in Ukrainian)
22. Bellman R, Cooke KL (1963) Differential difference equations. Academic Press, New York
23. Beretta E, Kuang Y (1998) Global analysis in some delayed ratio-dependent predator–prey
systems. Nonlinear Anal 32(4):381–408
24. Beretta E, Takeuchi Y (1994) Qualitative properties of chemostat equations with time delays:
boundedness, local and global asymptotic stability. Differ Equ Dyn Syst 2(1):19–40
25. Beretta E, Takeuchi Y (1994) Qualitative properties of chemostat equations with time delays
II. Differ Equ Dyn Syst 2(4):263–288
26. Beretta E, Takeuchi Y (1995) Global stability of an SIR epidemic model with time delays.
J Math Biol 33:250–260
27. Beretta E, Kolmanovskii V, Shaikhet L (1998) Stability of epidemic model with time delays
influenced by stochastic perturbations. Math Comput Simul 45(3–4):269–277 (Special Issue
“Delay Systems”)
28. Beretta E, Hara T, Ma W, Takeuchi Y (2001) Global asymptotic stability of an SIR epidemic
model with distributed time delay. Nonlinear Anal 47:4107–4115
29. Berezansky L, Braverman E, Idels L (2004) Delay differential logistic equations with har-
vesting. Math Comput Model 40:1509–1525
30. Berezansky L, Braverman E, Idels L (2010) Nicholson’s blowflies differential equations re-
visited: main results and open problems. Appl Math Model 34:1405–1417
31. Berezansky L, Idels L, Troib L (2011) Global dynamics of Nicholson-type delay systems
with applications. Nonlinear Anal, Real World Appl 12:436–445
32. Biran Y, Innis B (1979) Optimal control of bilinear systems: time-varying effects of cancer
drugs. Automatica 15
33. Blackburn JA, Smith HJT, Gronbech-Jensen N (1992) Stability and Hopf bifurcations in an
inverted pendulum. Am J Phys 60(10):903–908
34. Blanchflower DG, Oswald AJ, Landeghem BV (2009) Imitative obesity and relative utility.
J Eur Econ Assoc 7:528–538
35. Blizorukov MG (1996) On the construction of solutions of linear difference systems with
continuous time. Differ Uravn (Minsk) 32:127–128. Translation in Differential Equations,
133–134
36. Borne P, Kolmanovskii V, Shaikhet L (1999) Steady-state solutions of nonlinear model
of inverted pendulum. Theory Stoch Process 5(21)(3–4):203–209. Proceedings of the third
Ukrainian–Scandinavian conference in probability theory and mathematical statistics, 8–12
June 1999, Kyiv, Ukraine
37. Borne P, Kolmanovskii V, Shaikhet L (2000) Stabilization of inverted pendulum by control
with delay. Dyn Syst Appl 9(4):501–514
38. Bradul N, Shaikhet L (2007) Stability of the positive point of equilibrium of Nicholson’s
blowflies equation with stochastic perturbations: numerical analysis. Discrete Dyn Nat Soc
2007:92959. doi:10.1155/2007/92959. 25 pages
39. Bradul N, Shaikhet L (2009) Stability of difference analogue of mathematical predator–
prey model by stochastic perturbations. Vest Odesskogo Naz Univ Mat Mekh 14(20):7–23
(in Russian)
40. Braverman E, Kinzebulatov D (2006) Nicholson’s blowflies equation with a distributed de-
lay. Can Appl Math Q 14(2):107–128
41. Buonomo B, Rionero S (2010) On the Lyapunov stability for SIRS epidemic models with
general nonlinear incidence rate. Appl Math Comput 217(8):4010–4016
42. Burkeand MA, Heiland F (2007) Social dynamics of obesity. Econ Inq 45:571–591
43. Burton TA (1983) Volterra integral and differential equations. Academic Press, New York
44. Burton TA (1985) Stability and periodic solutions of ordinary and functional differential
equations. Academic Press, Orlando
References 327

45. Busenberg S, Cooke KL (1980) The effect of integral conditions in certain equations mod-
elling epidemics and population growth. J Math Biol 10(1):13–32
46. Bush AW, Cook AE (1976) The effect of time delay and growth rate inhibition in the bacterial
treatment of wastewater. J Theor Biol 63:385–395
47. Cai L, Li X, Song X, Yu J (2007) Permanence and stability of an age-structured prey–predator
system with delays. Discrete Dyn Nat Soc 2007:54861. 15 pages
48. Caraballo T, Real J, Shaikhet L (2007) Method of Lyapunov functionals construction in sta-
bility of delay evolution equations. J Math Anal Appl 334(2):1130–1145. doi:10.1016/j.jmaa.
2007.01.038
49. Carletti M (2002) On the stability properties of a stochastic model for phage-bacteria inter-
action in open marine environment. Math Biosci 175:117–131
50. Chen F (2005) Periodicity in a ratio-dependent predator–prey system with stage structure for
predator. J Appl Math 2:153–169
51. Christakis NA, Fowler JH (2007) The spread of obesity in a large social network over 32
years. N Engl J Med 357:370–379
52. Chueshov I, Scheutzow M, Schmalfu B (2005) Continuity properties of inertial manifolds
for stochastic retarded semilinear parabolic equations. In: Deuschel J-D, Greven A (eds)
Interacting stochastic systems. Springer, Berlin, pp 355–375
53. Colliugs JB (1997) The effects of the functional response on the bifurcation behavior of a
mite predator–prey interaction model. J Math Biol 36:149–168
54. Corduneanu C, Lakshmikantham V (1980) Equations with unbounded delay: a survey. Non-
linear Anal 4:831–877
55. Crisci MR, Kolmanovskii VB, Russo E, Vecchio A (1995) Stability of continuous and dis-
crete Volterra integro–differential equations by Lyapunov approach. J Integral Equ 4(7):393–
411
56. Crisci MR, Kolmanovskii VB, Russo E, Vecchio A (1998) Stability of difference Volterra
equations: direct Lyapunov method and numerical procedure. Comput Math Appl 36(10–
12):77–97
57. Cushing JM (1977) Integro–differential equations and delay models in population dynamics.
Lectures notes in biomathematics. Springer, Berlin
58. Diekmann O, Heesterbeek JAP, Metz JAJ (1990) On the definition and the computation of
the basic reproduction R0 in models for infectious diseases in heterogeneous populations.
J Math Biol 28:365–382
59. Diekmann O, Von Gils SA, Verduyn Lunel SM, Walther HO (1995) Delay equations, func-
tional, complex and nonlinear analysis. Applied math sciences. Springer, Berlin
60. Ding X, Jiang J (2008) Positive periodic solutions in delayed Gauss-type predator–prey sys-
tems. J Math Anal Appl 339(2):1220–1230. doi:10.1016/j.jmaa.2007.07.079
61. Ding X, Li W (2006) Stability and bifurcation of numerical discretization Nicholson’s
blowflies equation with delay. Discrete Dyn Nat Soc 2006:1–12. doi:10.1155/DDNS/2006/
19413
62. Driver RD (1962) Existence and stability of solutions of a delay differential system. Arch
Ration Mech Anal 10:401–426
63. Driver RD (1997) Ordinary and delay differential equations. Applied math sciences.
Springer, Berlin
64. Edwards JT, Ford NJ, Roberts JA, Shaikhet LE (2000) Stability of a discrete nonlinear
integro–differential equation of convolution type. Stab Control: Theory Appl 3(1):24–37
65. El-Owaidy HM, Ammar A (1988) Stable oscillations in a predator–prey model with time lag.
J Math Anal Appl 130:191–199
66. El’sgol’ts LE, Norkin SB (1973) Introduction to the theory and application of differential
equations with deviating arguments. Academic Press, New York
67. Erbe LH, Qingkai K, Zhang BG (1994) Oscillation theory for functional differential equa-
tions. Marcel Dekker, New York
68. Etile F (2007) Social norms, ideal body weight and food attitudes. Health Econ 16:945–966
328 References

69. Fan M, Wang Q, Zhou X (2003) Dynamics of a nonautonomous ratio-dependent predator–

prey system. Proc R Soc Edinb A 133:97–118
70. Fan YH, Li WT, Wang LL (2004) Periodic solutions of delayed ratio-dependent predator–
prey model with monotonic and no-monotonic functional response. Nonlinear Anal, Real
World Appl 5(2):247–263
71. Fan X, Wang Z, Xu X (2012) Global stability of two-group epidemic models with distributed
delays and random perturbation. Abstr Appl Anal 2012:132095
72. Farkas M (1984) Stable oscillations in a predator–prey model with time lag. J Math Anal
Appl 102:175–188
73. Feng QX, Yan JR (2002) Global attractivity and oscillation in a kind of Nicholson’s blowflies.
J Biomath 17(1):21–26
74. Ford NJ, Baker CTH (1996) Qualitative behavior and stability of solutions of discretised
nonlinear Volterra integral equations of convolution type. J Comput Appl Math 66:213–225
75. Ford NJ, Baker CTH, Roberts JA (1997) Preserving qualitative behavior and transience in
numerical solutions of Volterra integro–differential equations of convolution type: Lyapunov
functional approaches. In: Proceeding of 15th world congress on scientific computation,
modelling and applied mathematics (IMACS97), Berlin, August 1997. Numerical mathe-
matics, vol 2, pp 445–450
76. Ford NJ, Baker CTH, Roberts JA (1998) Nonlinear Volterra integro–differential equations—
stability and numerical stability of θ -methods. MCCM numerical analysis report. J Integral
Equ Appl 10:397–416
77. Ford NJ, Edwards JT, Roberts JA, Shaikhet LE (1997) Stability of a difference analogue
for a nonlinear integro differential equation of convolution type. Numerical Analysis Report,
University of Manchester 312
78. Fridman E (2001) New Lyapunov–Krasovskii functionals for stability of linear retarded and
neutral type systems. Syst Control Lett 43:309–319
79. Fridman E (2002) Stability of linear descriptor systems with delays: a Lyapunov-based ap-
proach. J Math Anal Appl 273:24–44
80. Gantmacher FR (2000) Matrix theory, vol 2. American Mathematical Society, Providence
81. Gard TC (1988) Introduction to stochastic differential equations. Marcel Dekker, Basel
82. Garvie M (2007) Finite-difference schemes for reaction–diffusion equations modelling
predator–prey interactions in MATLAB. Bull Math Biol 69(3):931–956
83. Ge Z, He Y (2008) Diffusion effect and stability analysis of a predator–prey system described
by a delayed reaction–diffusion equations. J Math Anal Appl 339(2):1432–1450. doi:10.
1016/j.jmaa.2007.07.060
84. Gikhman II, Skorokhod AV (1972) Stochastic differential equations. Springer, Berlin
85. Gikhman II, Skorokhod AV (1974) The theory of stochastic processes, vol I. Springer, Berlin
86. Gikhman II, Skorokhod AV (1975) The theory of stochastic processes, vol II. Springer,
Berlin
87. Gikhman II, Skorokhod AV (1979) The theory of stochastic processes, vol III. Springer,
Berlin
88. Gil’ MI (1998) Stability of finite and infinite dimensional systems. Kluwer Academic,
Boston
89. Golec J, Sathananthan S (1999) Sample path approximation for stochastic integro–
differential equations. Stoch Anal Appl 17(4):579–588
90. Golec J, Sathananthan S (2001) Strong approximations of stochastic integro–differential
equations. Dyn Contin Discrete Impuls Syst, Ser B, Appl Algorithms 8(1):139–151
91. Gopalsamy K (1992) Equations of mathematical ecology. Kluwer Academic, Dordrecht
92. Gopalsamy K (1992) Stability and oscillations in delay differential equations of population
dynamics. Mathematics and its applications, vol 74. Kluwer Academic, Dordrecht
93. Gorecki H, Fuksa S, Grabowski P, Korytowski A (1989) Analysis and synthesis of time-delay
systems. Wiley, New York
94. Gourley SA, Kuang Y (2004) A stage structured predator–prey model and its dependence on
maturation delay and death rate. J Math Biol 4:188–200
References 329

95. Grove EA, Ladas G, McGrath LC, Teixeira CT (2001) Existence and behavior of solutions
of a rational system. Commun Appl Nonlinear Anal 8(1):1–25
96. Gu K (2000) An integral inequality in the stability problem of time-delay system. In: 39th
IEEE conference on decision and control, Sydney, Australia, pp 2805–2810
97. Gu K, Niculescu S-I (2000) Additional dynamics in transformed time-delay systems. IEEE
Trans Autom Control 45:572–575
98. Guo MY, Li Z (2006) Global dynamics of a staged progression model for infectious diseases.
Math Biosci Eng 3(3):513–525
99. Gurney WSC, Blythe SP, Nisbet RM (1980) Nicholson’s blowflies revisited. Nature 287:17–
21
100. Gyori I, Ladas G (1991) Oscillation theory of delay differential equations with applications.
Oxford mathematical monographs. Oxford University Press, New York
101. Gyori I, Trofimchuk S (1999) Global attractivity in ẋ(t) = −δx(t) + pf (x(t − τ )). Dyn Syst
Appl 8:197–210
102. Gyori I, Trofimchuck SI (2002) On the existence of rapidly oscillatory solutions in the
Nicholson’s blowflies equation. Nonlinear Anal 48:1033–1042
103. Halanay A (1966) Differential equations: stability, oscillations, time lags. Academic Press,
New York
104. Hale JK (1972) Oscillations in neutral functional differential equations. Nonlinear Mechan-
ics, CIME
105. Hale JK (1977) Theory of functional differential equations. Springer, New York
106. Hale JK, Verduyn-Lunel SM (1993) Introduction to functional differential equations.
Springer, New York
107. Harrison GW (1979) Global stability of predator–prey interactions. J Math Biol 8:159–171
108. Hastings A (1983) Age dependent predation is not a simple process. I. Continuous time
models. Theor Popul Biol 23:347–362
109. Hastings A (1984) Delays in recruitment at different trophic levels effects on stability. J Math
Biol 21:35–44
110. Hethcote HW (1976) Qualitative analysis of communicable disease models. Math Biol Sci
28:335–356
111. Hino Y, Murakami S, Naito T (1991) Functional differential equations with infinite delay.
Lecture notes in mathematics. Springer, New York
112. Hsu SB, Huang TW (1995) Global stability for a class of predator–prey systems. SIAM J
Appl Math 55(3):763–783
113. Huo HF, Li WT (2004) Periodic solution of a delayed predator–prey system with Michaelis–
Menten type functional response. J Comput Appl Math 166:453–463
114. Imkeller P, Lederer Ch (2001) Some formulas for Lyapunov exponents and rotation numbers
in two dimensions and the stability of the harmonic oscillator and the inverted pendulum.
Dyn Syst 16:29–61
115. James PT, Leach R, Kalamara E, Shayeghi M (2001) The worldwide obesity epidemic. Obe-
sity 9:228–233
116. Ji C, Jiang D, Li X (2011) Qualitative analysis of a stochastic ratio-dependent predator–prey
system. J Comput Appl Math 235:1326–1341
117. Jovanovic M, Krstic M (2012) Stochastically perturbed vector-borne disease models with
direct transmission. Appl Math Model 36:5214–5228
118. Jumpen W, Orankitjaroen S, Boonkrong P, Wiwatanapataphee B (2011) SEIQR-SIS epi-
demic network model and its stability. Int J Math Comput Simul 5:326–333
119. Kac IYa (1998) Method of Lyapunov functions in problems of stability and stabilization of
systems of stochastic structure. Ural State Academy of Railway Communication, Yekaterin-
burg, pp 222 (in Russian)
120. Kac IYa, Krasovskii NN (1960) About stability of systems with stochastic parameters. Prikl
Mat Meh 24(5):809–823 (in Russian)
121. Kapitza PL (1965) Dynamical stability of a pendulum when its point of suspension vi-
brates, and pendulum with a vibrating suspension. In: ter Haar D (ed) Collected papers of
330 References

P.L. Kapitza, vol 2. Pergamon Press, London, pp 714–737

122. Karafyllis I, Pepe P, Jiang ZP (2009) Stability results for systems described by coupled re-
tarded functional differential equations and functional difference equations. Nonlinear Anal,
Theory Methods Appl 71:3339–3362
123. Keeling MJ, Rohani P (2008) Modeling infectious diseases in humans and animals. Princeton
University Press, Princeton
124. Khan H, Mohapatra RN, Vajravelu K, Liao SJ (2009) The explicit series solution of SIR and
SIS epidemic models. Appl Math Comput 215:653–669
125. Khasminskii RZ (1980) Stochastic stability of differential equations. Sijthoff & Noordhoff,
Alphen aan den Rijn
126. Kim AV (1999) Functional differential equations: application of I-smooth calculus. Mathe-
matics and its applications. Kluwer Academic, Dordrecht
127. Ko W, Ahn I (2013) A diffusive one-prey and two-competing-predator system with a ratio-
dependent functional response: I. Long time behavior and stability of equilibria. J Math Anal
Appl 397(1):9–28
128. Ko W, Ahn I (2013) A diffusive one-prey and two-competing-predator system with a
ratio-dependent functional response: II. Stationary pattern formation. J Math Anal Appl
397(1):29–45
129. Kocic VL, Ladas G (1990) Oscillation and global attractivity in discrete model of Nicholson’s
blowflies. Appl Anal 38:21–31
130. Kolmanovskii VB (1993) On stability of some hereditary systems. Avtom Telemeh 11:45–59
(in Russian)
131. Kolmanovskii VB (1995) Applications of differential inequalities for stability of some func-
tional differential equations. Nonlinear Anal, Theory Methods Appl 25(9–10):1017–1028
132. Kolmanovskii VB, Myshkis AD (1992) Applied theory of functional differential equations.
Kluwer Academic, Dordrecht
133. Kolmanovskii VB, Myshkis AD (1999) Introduction to the theory and applications of func-
tional differential equations. Kluwer Academic, Dordrecht
134. Kolmanovskii VB, Nosov VR (1981) Stability and periodical regimes of regulating heredi-
tary systems. Nauka, Moscow (in Russian)
135. Kolmanovskii VB, Nosov VR (1986) Stability of functional differential equations. Academic
Press, London
136. Kolmanovskii VB, Shaikhet LE (1993) Method for constructing Lyapunov functionals for
stochastic systems with aftereffect. Differ Uravn (Minsk) 29(11):1909–1920. 2022, (in Rus-
sian). Translated in Differential Equations 29(11):1657–1666 (1993)
137. Kolmanovskii VB, Shaikhet LE (1993) Stability of stochastic systems with aftereffect. Av-
tom Telemeh 7:66–85 (in Russian). Translation in Automat. Remote Control 54(7):1087–
1107, part 1
138. Kolmanovskii VB, Shaikhet LE (1994) New results in stability theory for stochastic
functional–differential equations (SFDEs) and their applications. In: Proceedings of dynamic
systems and applications, Atlanta, 1993, vol 1. Dynamic Publishers, Atlanta, pp 167–171
139. Kolmanovskii VB, Shaikhet LE (1995) Method for constructing Lyapunov functionals for
stochastic differential equations of neutral type. Differ Uravn (Minsk) 31(11):1851–1857
(in Russian). Translated in Differential Equations 31(11):1819–1825 (1995)
140. Kolmanovskii VB, Shaikhet LE (1996) Control of systems with aftereffect. Translations of
mathematical monographs, vol 157. American Mathematical Society, Providence
141. Kolmanovskii VB, Shaikhet LE (1997) About stability of some stochastic Volterra equations.
Differ Uravn 11:1495–1502 (in Russian)
142. Kolmanovskii VB, Shaikhet LE (1997) Matrix Riccati equations and stability of stochastic
linear systems with nonincreasing delays. Funct Differ Equ 4(3–4):279–293
143. Kolmanovskii VB, Shaikhet LE (1998) Riccati equations and stability of stochastic linear
systems with distributed delay. In: Bajic V (ed) Advances in systems, signals, control and
computers. IAAMSAD and SA branch of the Academy of Nonlinear Sciences, Durban, pp
97–100. ISBN 0-620-23136-X
References 331

144. Kolmanovskii VB, Shaikhet LE (1998) Riccati equations in stability of stochastic linear sys-
tems with delay. Avtom Telemeh 10:35–54 (in Russian)
145. Kolmanovskii VB, Shaikhet LE (2002) Construction of Lyapunov functionals for stochastic
hereditary systems: a survey of some recent results. Math Comput Model 36(6):691–716
146. Kolmanovskii VB, Shaikhet LE (2002) Some peculiarities of the general method of
Lyapunov functionals construction. Appl Math Lett 15(3):355–360. doi:10.1016/S0893-
9659(01)00143-4
147. Kolmanovskii VB, Shaikhet LE (2003) About one application of the general method of Lya-
punov functionals construction. Int J Robust Nonlinear Control 13(9):805–818 (Special Issue
on Time Delay Systems)
148. Kolmanovskii VB, Shaikhet LE (2004) About some features of general method of Lyapunov
functionals construction. Stab Control: Theory Appl 6(1):49–76
149. Kolmanovskii VB, Kosareva NP, Shaikhet LE (1999) A method for constructing Lyapunov
functionals. Differ Uravn (Minsk) 35(11):1553–1565 (in Russian). Translated in Differential
Equations 35(11):1573–1586 (1999)
150. Kordonis I-GE, Philos ChG (1999) The behavior of solutions of linear integro–differential
equations with unbounded delay. Comput Math Appl 38:45–50
151. Kordonis I-GE, Philos ChG, Purnaras IK (2004) On the behavior of solutions of linear neutral
integro–differential equations with unbounded delay. Georgian Math J 11:337–348
152. Korobeinikov A (2006) Lyapunov functions and global stability for SIR and SIRS epidemi-
ological models with non-linear transmission. Bull Math Biol 68:615–626
153. Korobeinikov A (2009) Stability of ecosystem: global properties of a general predator–prey
model. IMA J Math Appl Med Biol 26(4):309–321
154. Korobeinikov A, Maini PK (2005) Nonlinear incidence and stability of infectious disease
models. Math Med Biol 22:113–128
155. Kovalev AA, Kolmanovskii VB, Shaikhet LE (1998) Riccati equations in the stability of
retarded stochastic linear systems. Avtom Telemeh 10:35–54 (in Russian). Translated in Au-
tomatic Remote Control 59(10):1379–1394, part 1 (1998)
156. Krasovskii NN (1956) On the application of the second Lyapunov method for equation with
time-delay. Prikl Mat Meh 20(1):315–327 (in Russian). Translated in J Appl Math Mekh 20
(1956)
157. Krasovskii NN (1956) On the asymptotic stability of systems with aftereffect. Prikl Mat Meh
20(3):513–518 (in Russian). Translated in J Appl Math Mekh 20 (1956)
158. Krasovskii NN (1959) Some problems of dynamic stability theory. Fizmatgiz, Moscow (in
Russian)
159. Krasovskii NN (1963) Stability of motion. Standford Univ. Press, Standford
160. Kuang Y (1993) Delay differential equations with application in population dynamics. Aca-
demic Press, New York
161. Kulenovic MRS, Ladas G (1987) Linearized oscillations in population dynamics. Bull Math
Biol 49(5):615–627
162. Kulenovic MRS, Nurkanovic M (2002) Asymptotic behavior of a two dimensional linear
fractional system of difference equations. Rad Mat 11(1):59–78
163. Kulenovic MRS, Nurkanovic M (2005) Asymptotic behavior of a system of linear fractional
difference equations. Arch Inequal Appl 2005(2):127–143
164. Kulenovic MRS, Nurkanovic M (2006) Asymptotic behavior of a competitive system of lin-
ear fractional difference equations. Adv Differ Equ 2006(5):19756. doi:10.1155/ADE/2006/
19756. 13 pages
165. Kulenovic MRS, Ladas G, Sficas YG (1989) Global attractivity in population dynamics.
Comput Math Appl 18(10–11):925–928
166. Kulenovic MRS, Ladas G, Sficas YG (1992) Global attractivity in Nicholson’s blowflies.
Appl Anal 43:109–124
167. Ladde GS, Lakshmikantham V, Zhang BG (1987) Oscillation theory of differential equations
with deviating arguments. Marcel Dekker, New York
332 References

168. Lahrouz A, Omari L, Kiouach D (2011) Global analysis of a deterministic and stochastic
nonlinear SIRS epidemic model. Nonlinear Anal, Model Control 16:59–76
169. Levi M (1988) Stability of the inverted pendulum—a topological explanation. SIAM Rev
30(4):639–644
170. Levi M, Weckesser W (1995) Stabilization of the inverted linearized pendulum by high fre-
quency vibrations. SIAM Rev 37(2):219–223
171. Levin JJ, Nohel JA (1963) Note on a nonlinear Volterra equation. Proc Am Math Soc
14(6):924–929
172. Levin JJ, Nohel JA (1965) Perturbations of a non-linear Volterra equation. Mich Math J
12:431–444
173. Li J (1996) Global attractivity in a discrete model of Nicholson’s blowflies. Ann Differ Equ
12(2):173–182
174. Li J (1996) Global attractivity in Nicholson’s blowflies. Appl Math J Chin Univ Ser B
11(4):425–434
175. Li J, Du C (2008) Existence of positive periodic solutions for a generalized Nicholson’s
blowflies model. J Comput Appl Math 221(1):226–233
176. Li W, Fan Y (2007) Existence and global attractivity of positive periodic solutions for the
impulsive delay Nicholson’s blowflies model. J Comput Appl Math 201(1):55–68
177. Li YK, Kuang Y (2001) Periodic solutions of periodic delay Lotka–Volterra equations and
systems. J Math Anal Appl 255:260–280
178. Li M, Yan J (2000) Oscillation and global attractivity of generalized Nicholson’s blowfly
model. In: Differential equations and computational simulations, Chengdu, 1999. World Sci-
entific, Singapore, pp 196–201
179. Li MY, Graef JR, Wang L, Karsai J (1999) Global dynamics of a SEIR model with varying
total population size. Math Biosci 160:191–213
180. Liao X, Zhou Sh, Ouyang Z (2007) On a stoichiometric two predators on one prey discrete
model. Appl Math Lett 20:272–278
181. Liu K (2006) Stability of infinite dimensional stochastic differential equations with applica-
tions. Chapman and Hall/CRC, Boca Raton
182. Liz E (2007) A sharp global stability result for a discrete population model. J Math Anal
Appl 330:740–743
183. Liz E, Röst G (2009) On global attractor of delay differential equations with unimodal feed-
back. Discrete Contin Dyn Syst 24(4):1215–1224
184. Long F (2012) Positive almost periodic solution for a class of Nicholson’s blowflies model
with a linear harvesting term. Nonlinear Anal 13(2):686–693
185. Lubich Ch (1983) On the stability of linear multistep methods for Volterra convolution equa-
tions. IMA J Numer Anal 3:439–465
186. Luo Q, Mao X (2007) Stochastic population dynamics under regime switching. J Math Anal
Appl 334:69–84
187. Ma W, Takeuchi Y, Hara T, Beretta E (2002) Permanence of an SIR epidemic model with
distributed time delays. Tohoku Math J 54:581–591
188. Ma W, Song M, Takeuchi Y (2004) Global stability of an SIR epidemic model with time
delay. Appl Math Lett 17:1141–1145
189. Macdonald N (1978) Time lags in biological models. Lecture notes in biomath. Springer,
Berlin
190. Mackey MC, Glass L (1997) Oscillation and chaos in physiological control system. Science
197:287–289
191. Mackey MC, Nechaeva IG (1994) Noise and stability in differential delay equations. J Differ
Equ 6:395–426
192. Maistrenko YuL, Sharkovsky AN (1986) Difference equations with continuous time as math-
ematical models of the structure emergences. In: Dynamical systems and environmental
models. Mathematical ecology, Eisenach. Akademie-Verlag, Berlin, pp 40–49
193. Makinde OD (2007) Adomian decomposition approach to a SIR epidemic model with con-
stant vaccination strategy. Appl Math Comput 184:842–848
References 333

194. Malek-Zavarei M, Jamshidi M (1987) Time-delay systems: analysis, optimization and appli-
cations. North-Holland, Amsterdam
195. Mao X (1994) Exponential stability of stochastic differential equations. Marcel Dekker, New
York
196. Mao X (2000) Stability of stochastic integro–differential equations. Stoch Anal Appl
18(6):1005–1017
197. Mao X, Shaikhet L (2000) Delay-dependent stability criteria for stochastic differential delay
equations with Markovian switching. Stab Control: Theory Appl 3(2):88–102
198. Mao X, Yuan C (2006) Stochastic differential equations with Markovian switching. Imperial
College Press, London
199. Marotto F (1982) The dynamics of a discrete population model with threshold. Math Biosci
58:123–128
200. Maruyama G (1955) Continuous Markov processes and stochastic equations. Rend Circ Mat
Palermo 4:48–90
201. Mata GJ, Pestana E (2004) Effective Hamiltonian and dynamic stability of the inverted pen-
dulum. Eur J Phys 25:717–721
202. McCluskey CC (2010) Complete global stability for an SIR epidemic model with delay—
distributed or discrete. Nonlinear Anal, Real World Appl 11:55–59
203. McCluskey CC (2010) Global stability for an SIR epidemic model with delay and nonlinear
incidence. Nonlinear Anal, Real World Appl 11(4):3106–3109
204. Melchor-Aguilar D (2010) On stability of integral delay systems. Appl Math Comput
217:3578–3584
205. Melchor-Aguilar D, Kharitonov V, Lozano R (2010) Stability conditions for integral delay
systems. Int J Robust Nonlinear Control 20:1–15
206. Michiels W, Niculescu S-I (2007) Stability and stabilization of time-delay systems: an
eigenvalue-based approach. SIAM, Philadelphia, PA
207. Miller RK (1971) Asymptotic stability properties of linear Volterra integro–differential equa-
tions. J Differ Equ 10:485–506
208. Miller RK (1972) Nonlinear Volterra integral equations. Benjamin, Elmsford
209. Mitchell R (1972) Stability of the inverted pendulum subjected to almost periodic and
stochastic base motion—an application of the method of averaging. Int J Non-Linear Mech
7:101–123
210. Mohammed SEA (1984) Stochastic functional differential equations. Pitman, Boston
211. Mohammed SEA, Scheutzow M (1990) Lyapunov exponents and stationary solutions for
affine stochastic delay equations. Stoch Stoch Rep 29:259–283
212. Mollison D (2003) Epidemic models: their structure and relation to data. Publications of the
Newton Institute. Cambridge University Press, Cambridge
213. Mukhopadhyay B, Bhattacharyya R (2007) Existence of epidemic waves in a disease trans-
mission model with two-habitat population. Int J Syst Sci 38:699–707
214. Mukhopadhyay B, Bhattacharyya R (2009) A nonlinear mathematical model of virus-tumor-
immune system interaction: deterministic and stochastic analysis. Stoch Anal Appl 27:409–
429
215. Murakami S (1990) Exponential stability for fundamental solutions of some linear functional
differential equations. In: Yoshizwa T, Kato J (eds) Proceedings of the international sympo-
sium: functional differential equations. World Scientific, Singapore, pp 259–263
216. Murakami S (1991) Exponential asymptotic stability for scalar linear Volterra equations.
Differ Integral Equ 4:519–525
217. Muroya Y (2007) Persistence global stability in discrete models of Lotka–Volterra type.
J Math Anal Appl 330:24–33
218. Muroya Y, Enatsu Y, Nakata Y (2011) Global stability of a delayed SIRS epidemic model
with a non-monotonic incidence rate. J Math Anal Appl 377(1):1–14
219. Murray JD (2002) Mathematical biology, 3rd edn. Springer, New York
220. Myshkis AD (1951) General theory of differential equations with delay. Translations of
mathematical monographs, vol 55. American Mathematical Society, Providence
334 References

221. Myshkis AD (1972) Linear differential equations with delay argument. Nauka, Moscow (in
Russian)
222. Nam PT, Phat VN (2009) An improved stability criterion for a class of neutral differential
equations. Appl Math Lett 22:31–35
223. Naresh R, Tripathi A, Tchuenche JM, Sharma D (2009) Stability analysis of a time delayed
SIR epidemic model with nonlinear incidence rate. Comput Math Appl 58(2):348–359
224. Nicholson AJ (1954) An outline of the dynamics of animal populations. Aust J Zool 2:9–65
225. Niculescu S-I (2001) Delay effects on stability: a robust control approach. Springer, Berlin
226. Niculescu S-I, Gu K (eds) (2004) Advances in time-delay systems. Springer, Berlin
227. Oswald AJ, Powdthavee N (2007) Obesity, unhappiness and the challenge of affluence: the-
ory and evidence. Econ J 117:F441–454
228. Ovseyevich AI (2006) The stability of an inverted pendulum when there are rapid random
oscillations of the suspension point. Int J Appl Math Mech 70:762–768
229. Park JH (2004) Delay-dependent criterion for asymptotic stability of a class of neutral equa-
tions. Appl Math Lett 17:1203–1206
230. Park JH (2005) Delay-dependent criterion for guaranteed cost control of neutral delay sys-
tems. J Optim Theory Appl 124:491–502
231. Park JH, Kwon OM (2008) Stability analysis of certain nonlinear differential equation. Chaos
Solitons Fractals 27:450–453
232. Paternoster B, Shaikhet L (2008) Stability of equilibrium points of fractional difference equa-
tions with stochastic perturbations. Adv Differ Equ 2008:718408. doi:10.1155/2008/718408,
21 pages
233. Peics H (2000) Representation of solutions of difference equations with continuous time.
Electron J Qual Theory Differ Equ 21:1–8. Proceedings of the 6th colloquium of differential
equations
234. Pelyukh GP (1996) A certain representation of solutions to finite difference equations with
continuous argument. Differ Uravn (Minsk) 32(2):256–264. Translated in Differential Equa-
tions 32(2):260–268 (1996)
235. Peschel M, Mende W (1986) The predator–prey model: do we live in a Volterra world.
Akademie-Verlag, Berlin
236. Philos ChG (1998) Asymptotic behavior, nonoscillation and stability in periodic first-order
linear delay differential equations. Proc R Soc Edinb A 128:1371–1387
237. Philos ChG, Purnaras IK (2001) Periodic first order linear neutral delay differential equa-
tions. Appl Math Comput 117:203–222
238. Philos ChG, Purnaras IK (2004) Asymptotic properties, nonoscillation and stability for scalar
first-order linear autonomous neutral delay differential equations. Electron J Differ Equ
2004(3):1–17
239. Razumikhin BS (1956) About stability of systems with delay. Prikl Mat Meh 20(4):500–512
(in Russian)
240. Razumikhin BS (1958) About stability in the first approximation of systems with delay. Prikl
Mat Meh 22(2):155–166 (in Russian)
241. Razumikhin BS (1960) Application of Lyapunov method to problems of stability of systems
with delay. Avtom Telemeh 21(6):740–748 (in Russian)
242. Razumikhin BS (1966) Method of stability investigation of systems with delay. Dokl Akad
Nauk SSSR 167(6):1234–1236 (in Russian)
243. Repin YuM (1965) Square Lyapunov functionals for systems with delay. Prikl Mat Meh
3:564–566 (in Russian)
244. Resnick SI (1992) Adventures in stochastic processes. Birkhauser, Boston
245. Roach GF (ed) (1984) Mathematics in medicine and biomechanics. Shiva, Nautwick
246. Rodkina A, Schurz H, Shaikhet L (2008) Almost sure stability of some stochastic dynamical
systems with memory. Discrete Contin Dyn Syst 21(2):571–593
247. Rosenquist JN, Murabito J, Fowler JH, Christakis NA (2010) The spread of alcohol con-
sumption behavior in a large social network. Ann Intern Med 152(7):426–433
References 335

248. Röst G, Wu J (2007) Domain decomposition method for the global dynamics of delay differ-
ential equations with unimodal feedback. Proc R Soc Lond Ser A, Math Phys Sci 463:2655–
2669
249. Ruan S, Xiao D (2001) Global analysis in a predator–prey system with nonmonotonic func-
tional response. SIAM J Appl Math 61:1445–1472
250. Saker SH (2005) Oscillation of continuous and discrete diffusive delay Nicholson’s blowflies
models. Appl Math Comput 167(1):179–197
251. Saker SH, Agarwal S (2002) Oscillation and global attractivity in a periodic Nicholson’s
blowflies model. Math Comput Model 35:719–731
252. Saker SH, Zhang BG (2002) Oscillation in a discrete partial delay Nicholson’s blowflies
model. Math Comput Model 36(9–10):1021–1026
253. Santonja F-J, Shaikhet L (2012) Analysing social epidemics by delayed stochastic models.
Discrete Dyn Nat Soc 2012:530472. 13 pages. doi:10.1155/2012/530472
254. Santonja F-J, Sanchez E, Rubio M, Morera J-L (2010) Alcohol consumption in Spain and its
economic cost: a mathematical modeling approach. Math Comput Model 52:999–1003
255. Santonja F-J, Villanueva R-J, Jodar L, Gonzalez-Parra G (2010) Mathematical modelling
of social obesity epidemic in the region of Valencia, Spain. Math Comput Model Dyn Syst
16(1):23–34
256. Sanz-Serna JM (2008) Stabilizing with a hammer. Stoch Dyn 8:47–57
257. Sarkar RR, Banerjee S (2005) Cancer self remission and tumor stability—a stochastic ap-
proach. Math Biosci 196:65–81
258. Scheutzow M (1996) Stability and instability for stochastic systems with time delay.
Z Angew Math Mech 76(3):363–365
259. Scheutzow M (2005) Exponential growth rates for stochastic delay differential equations.
Stoch Dyn 5(2):163–174
260. Schurz H (1996) Asymptotical mean square stability of an equilibrium point of some linear
numerical solutions with multiplicative noise. Stoch Anal Appl 14:313–354
261. Shaikhet L (1975) Stability investigation of stochastic systems with delay by Lyapunov func-
tionals method. Probl Pereda Inf 11(4):70–76 (in Russian)
262. Shaikhet L (1995) On the stability of solutions of stochastic Volterra equations. Avtom Tele-
meh 56(8):93–102 (in Russian). Translated in Automatic Remote Control 56(8):1129–1137
(1995), part 2
263. Shaikhet L (1995) Stability in probability of nonlinear stochastic hereditary systems. Dyn
Syst Appl 4(2):199–204
264. Shaikhet L (1995) Stability in probability of nonlinear stochastic systems with delay. Mat
Zametki 57(1):142–146 (in Russian). Translation in Math. Notes 57(1–2):103–106
265. Shaikhet L (1996) Modern state and development perspectives of Lyapunov function-
als method in the stability theory of stochastic hereditary systems. Theory Stoch Process
2(18)(1–2):248–259
266. Shaikhet L (1996) Stability of stochastic hereditary systems with Markov switching. Theory
Stoch Process 2(18)(3–4):180–185
267. Shaikhet L (1998) Stability of predator–prey model with aftereffect by stochastic perturba-
tions. Stab Control: Theory Appl 1(1):3–13
268. Shaikhet L (2002) Numerical simulation and stability of stochastic systems with Markovian
switching. Neural Parallel Sci Comput 10(2):199–208
269. Shaikhet L (2004) About Lyapunov functionals construction for difference equations with
continuous time. Appl Math Lett 17(8):985–991
270. Shaikhet L (2004) Construction of Lyapunov functionals for stochastic difference equations
with continuous time. Math Comput Simul 66(6):509–521
271. Shaikhet L (2004) Lyapunov functionals construction for stochastic difference second kind
Volterra equations with continuous time. Adv Differ Equ 2004(1):67–91. doi:10.1155/
S1687183904308022
272. Shaikhet L (2005) General method of Lyapunov functionals construction in stability in-
vestigations of nonlinear stochastic difference equations with continuous time. Stoch Dyn
336 References

5(2):175–188. Special Issue Stochastic Dynamics with Delay and Memory

273. Shaikhet L (2005) Stability of difference analogue of linear mathematical inverted pendulum.
Discrete Dyn Nat Soc 2005(3):215–226
274. Shaikhet L (2006) About stability of a difference analogue of a nonlinear integro–differential
equation of convolution type. Appl Math Lett 19(11):1216–1221
275. Shaikhet L (2008) Stability of a positive point of equilibrium of one nonlinear system with
aftereffect and stochastic perturbations. Dyn Syst Appl 17:235–253
276. Shaikhet L (2009) Improved condition for stabilization of controlled inverted pendulum un-
der stochastic perturbations. Discrete Contin Dyn Syst 24(4):1335–1343. doi:10.3934/dcds.
2009.24.1335
277. Shaikhet L (2011) About an unsolved stability problem for a stochastic difference equation
with continuous time. J Differ Equ Appl 17(3):441–444. doi:10.1080/10236190903489973.
(iFirst article, 2010, 1–4, 1563–5120, First published on 05 March 2010)
278. Shaikhet L (2011) Lyapunov functionals and stability of stochastic difference equations.
Springer, London
279. Shaikhet L (2012) Two unsolved problems in the stability theory of stochastic differential
equations with delay. Appl Math Lett 25(3):636–637. doi:10.1016/j.aml.2011.10.002
280. Shaikhet L, Roberts J (2004) Stochastic Volterra integro–differential equations: stability and
numerical methods. University of Manchester. MCCM. Numerical Analysis Report 450, 38p
281. Shaikhet L, Roberts J (2011) Asymptotic stability analysis of a stochastic Volterra integro-
differential equation with fading memory. Dyn Contin Discrete Impuls Syst, Ser B, Appl
Algorithms 18:749–770
282. Sharp R, Tsai Y-H, Engquist B (2005) Multiple time scale numerical methods for the in-
verted pendulum problem. In: Multiscale methods in science and engineering. Lecture notes
computing science and engineering, vol 44. Springer, Berlin, pp 241–261
283. Shi X, Zhou X, Song X (2010) Dynamical properties of a delay prey-predator model with
disease in the prey species only. Discrete Dyn Nat Soc. Article ID 196204, 16 pages
284. So JW-H, Yang Y (1998) Dirichlet problem for the diffusive Nicholson’s blowflies equation.
J Differ Equ 150(2):317–348
285. So JW-H, Yu JS (1994) Global attractivity and uniformly persistence in Nicholson’s
blowflies. Differ Equ Dyn Syst 2:11–18
286. So JW-H, Yu JS (1995) On the stability and uniform persistence of a discrete model of
Nicholson’s blowflies. J Math Anal Appl 193(1):233–244
287. So JW-H, Wu J, Yang Y (2000) Numerical steady state and Hopf bifurcation analysis on the
diffusive Nicholson’s blowflies equation. Appl Math Comput 111(1):53–69
288. Song XY, Chen LS (2002) Optimal harvesting and stability for a predator–prey system with
stage structure. Acta Math Appl Sin 18(3):423–430 (English series)
289. Song XY, Jiang Y, Wei HM (2009) Analysis of a saturation incidence SVEIRS epidemic
model with pulse and two time delays. Appl Math Comput 214:381–390
290. Sopronyuk FO, Tsarkov EF (1973) About stability of linear differential equations with delay.
Dokl Ukr Akad Nauk, Ser A 4:347–350 (in Ukrainian)
291. Spanish Ministry of Health (2008) National drug observatory reports. Retrieved 13 Novem-
ber 2008 from http://www.pnsd.msc.es/Categoria2/observa/estudios/home.htm
292. Sun YG, Wang L (2006) Note on asymptotic stability of a class of neutral differential equa-
tions. Appl Math Lett 19:949–953
293. Swords C, Appleby JAD (2010) Stochastic delay difference and differential equations: ap-
plications to financial markets. LAP Lambert Academic, Saarbrücken
294. Takeuchi Y, Ma W, Beretta E (2000) Global asymptotic properties of a delay SIR epidemic
model with finite incubation times. Nonlinear Anal 42:931–947
295. Tamizhmani KM, Ramani A, Grammaticos B, Carstea AS (2004) Modelling AIDS epidemic
and treatment with difference equations. Adv Differ Equ 3:183–193
296. Tapaswi PK, Chattopadhyay J (1996) Global stability results of a “susceptible-infective-
immune-susceptible” (SIRS) epidemic model. Ecol Model 87(1–3):223–226
References 337

297. Thomas DM, Robbins F (1999) Analysis of a nonautonomous Nicholson blowfly model.
Physica A 273(1–2):198–211
298. Tsypkin YaZ (1946) Stability of systems with a delayed feed-back. Avtom Telemeh
7(2–3):107–129 (in Russian)
299. Tsypkin YaZ (1947) Degree of stability of systems with a delayed feed-back. Avtom Telemeh
8(3):145–155 (in Russian)
300. Tsypkin YaZ (1984) An adaptive control of objects with aftereffect. Nauka, Moscow (in
Russian)
301. Verriest EI (2001) New qualitative methods for stability of delay systems. Kybernetika
37:229–238
302. Volterra V (1931) Lesons sur la theorie mathematique de la lutte pour la vie. Gauthier-Villars,
Paris
303. Volz R (1982) Global asymptotic stability of a periodical solution to an epidemic model.
J Math Biol 15:319–338
304. Wang WD, Chen LS (1997) A predator–prey system with stage structure for predator. Com-
put Math Appl 33(8):83–91
305. Wang LL, Li WT (2003) Existence and global stability of positive periodic solutions of a
predator–prey system with delays. Appl Math Comput 146(1):167–185
306. Wang LL, Li WT (2004) Periodic solutions and stability for a delayed discrete ratio-
dependent predator–prey system with Holling-type functional response. Discrete Dyn Nat
Soc 2004(2):325–343
307. Wang C, Wei J (2008) Bifurcation analysis on a discrete model of Nicholson’s blowflies.
J Differ Equ Appl 14:737–746
308. Wang QG, Lee TH, Tan KK (1999) Finite-spectrum assignment for time-delay systems. Lec-
ture notes in control and inform sciences. Springer, Berlin
309. Wang Q, Fan M, Wang K (2003) Dynamics of a class of nonautonomous semi-ratio-
dependent predator–prey system with functional responses. J Math Anal Appl 278:443–471
310. Wang W, Xin J, Zhang F (2010) Persistence of an SEIR model with immigration dependent
on the prevalence of infection. Discrete Dyn Nat Soc. doi:10.1155/2010/727168. Article ID
727168, 7 pages
311. Wangersky PJ, Cunningham WJ (1957) Time lag in prey–predator population models. Ecol-
ogy 38(1):136–139
312. Wei J, Li MY (2005) Hopf bifurcation analysis in a delayed Nicholson blowflies equation.
Nonlinear Anal 60(7):1351–1367
313. Wolkenfelt PHM (1982) The construction of reducible quadrature rules for Volterra integral
and integro–differential equations. IMA J Numer Anal 2:131–152
314. Xiao D, Ruan S (2001) Multiple bifurcations in a delayed predator–prey system with non-
monotonic functional response. J Differ Equ 176:494–510
315. Yang J, Wang X (2010) Existence of a nonautonomous SIR epidemic model with age struc-
ture. Adv Differ Equ. Article ID 212858, 23 pages
316. Yi T, Zou X (2008) Global attractivity of the diffusive Nicholson blowflies equation with
Neumann boundary condition: a nonmonotone case. J Differ Equ 245(11):3376–3388
317. Zeng X (2007) Non-constant positive steady states of a prey–predator system with cross-
diffusions. J Math Anal Appl 332(2):989–1009
318. Zevin AA (2010) The necessary and sufficient conditions for the stability of linear systems
with an arbitrary delay. J Appl Math Mech 74(4):384–388
319. Zhang J, Jin Z (2010) The analysis of epidemic network model with infectious force in latent
and infected period. Discrete Dyn Nat Soc. Article ID 604329, 12 pages
320. Zhang BG, Xu HX (1999) A note on the global attractivity of a discrete model of Nicholson’s
blowflies. Discrete Dyn Nat Soc 3:51–55
321. Zhang X, Chen L, Neumann UA (2000) The stage-structured predator–prey model and opti-
mal harvesting policy. Math Biosci 168:201–210
322. Zhang F, Li Z, Zhang F (2008) Global stability of an SIR epidemic model with constant
infectious period. Appl Math Comput 199:285–291
338 References

323. Zhang TL, Liu JL, Teng ZD (2009) Dynamic behaviour for a nonautonomous SIRS epidemic
model with distributed delays. Appl Math Comput 214:624–631
324. Zhang Z, Wu J, Suo Y, Song X (2011) The domain of attraction for the endemic equilibrium
of an SIRS epidemic model. Math Comput Simul 81:1697–1706
325. Zuo W, Wei J (2011) Stability and Hopf bifurcation in a diffusive predator–prey system with
delay effect. Nonlinear Anal, Real World Appl 12(4):1998–2011
Index

A C
Additive stochastic perturbations, 297 Characteristic equation, vi, 22, 25, 26, 66, 83,
Alcohol consumption, 297, 298, 307 214, 216, 219
Alcohol consumption model, 298 Characteristic function, 227–229
Algebraic adjuncts, 28 Characteristic quasipolynomial, 5, 9, 13, 18,
Algebraic equations, 260, 300 215, 216
Appropriate characteristic equation, 24 Constant delay, 52, 97
Argument deviations, 1, 2 Constructing a Lyapunov functional, 171
Asymptotic mean-square stability, vii, 43, 44, Consumption of alcohol, 297
47, 51, 53, 59, 62–64, 70, 71, 78, 119, Controlled inverted pendulum, 209, 222, 226,
120, 123, 124, 130, 131, 134, 135, 146, 229
147, 160, 163, 169, 179, 187, 189, 194, Controlled pendulum, 240
196, 198, 226, 251–254, 264, 285, 290,
291, 297, 302, 303, 305, 316, 318 D
Asymptotic mean-square stability conditions, Death rate, 298
153, 189, 257 Delay, 321
Asymptotic p-stability, 35 Derivative, 229
Asymptotic stability, 5, 7, 10, 18, 28, 45, 83, Diagonal matrix, 75
97, 100, 102, 105, 116, 204, 211, 212, Difference analogue, 12, 20, 24, 151, 199, 202,
215, 227, 229, 233 229, 256, 275, 296, 309, 322
Asymptotically mean-square stable, 34, 36, Difference equation, 1, 39
39–41, 48, 50, 52, 54, 76, 81, 86, 88, Difference equations with fractional
91, 92, 94, 99, 101, 103, 106, 109, 110, nonlinearity, 132
114, 116, 118, 122, 126, 154, 156, 162, Differential equation, 5
165, 167, 168, 170, 173, 180, 182, 183, Differential equation of neutral type, 2, 86, 88,
185, 190, 191, 193, 221, 266, 270, 272, 91, 124, 214
274, 280, 285–292, 304, 306, 307, 317, Differential equation with delay, 153, 210, 251
319 Differential equation with deviating argument,
Asymptotically p-stable, 34 2
Asymptotically p-trivial, 34, 35 Differential equation with fractional
Asymptotically stable, 18, 22, 24–26, 52, 53, nonlinearity, 132
56, 74, 86, 88, 96, 209, 214–216, 218, Differential equation without delay, 91
223, 230, 232, 241, 245, 248, 285 Differential equation without memory, 303
Differential–difference equation, v, 2
B Dirac’s delta-function, 215, 309, 321
Birth rate, 298 Dirac’s function, 298
Brownian motion, 29 Discrete delay, 4, 298

L. Shaikhet, Lyapunov Functionals and Stability of Stochastic Functional 339

Differential Equations, DOI 10.1007/978-3-319-00101-2,
© Springer International Publishing Switzerland 2013
340 Index

Discretization step, 310, 323 Integro–differential equations, 3, 299, 311

Distributed and discrete delays, 210 Intensity of transition, 298, 312
Distributed delay, 97, 102, 105, 120, 160, 164, Inverse matrices, 169
179, 258, 297, 298, 311 Inverted mathematical pendulum, 241
Inverted pendulum, 210, 214, 219, 220, 232
E Itô formula, vi, 33
Eigenvalues, 225 Itô integral, vi, 31
Equations with delays, v Itô stochastic differential equation, 43
Equations with deviating arguments, v Itô stochastic functional differential equation,
Equations with memory, v 32
Equations with time lag, v
Equilibrium point, vi, viii, 124, 126, 131–136, K
138–140, 142–145, 147, 149–151, Kernel, 210, 223
251–254, 261, 264, 266, 270, 272, 280,
283–285, 289, 290, 294–297, 301, 307, L
311, 312, 315, 319, 322 Limit cycle, 232, 233, 239
Euler–Maruyama scheme, 256 Linear matrix equation, 153, 168, 169
Existence and uniqueness theorems, 32 Linear stochastic differential equation, 179
Exponentially p-stable, 34, 37 Lipschitz condition, 36
Lotka–Volterra model, 259
F Lyapunov function, 40, 53, 79, 85, 86, 88, 125,
Fp -functional, 34 303
Fractional nonlinearity, vii, 131 Lyapunov functional, v, vi, vii, 34, 39, 40, 43,
Function of bounded variation, 210 53, 121, 125, 130, 153, 154, 157, 164,
Functional, 58 170, 174, 175, 189, 190, 212, 213,
Functional differential equation, v, 1, 2, 39 221–223, 272, 287, 291, 297, 305, 316
Functional equation, 1
Functions of bounded variation, 41, 113, 120 M
Mackey–Glass model, 132
G Markov chain, 187, 190, 193, 194, 199, 200,
General method of construction of Lyapunov 202–206, 208, 240, 247, 248
functionals, vi Markov processes, 29
Generator, 33, 35, 40, 41, 121, 127, 189, 221, Markovian switching, vii, 187, 198
266, 280, 286, 288, 306, 318 Mathematical expectation, 29
Mathematical model, vii, 209, 229, 257, 283,
H 297
Hereditary systems, v Mathematical obesity model, 311
Hölder inequality, 35 Matrix differential equation, 102
Hurwitz matrix, 25, 27 Matrix equation, 75, 84, 97, 100, 105, 109,
212, 303, 304
I Matrix Lyapunov equation, 120
Independent random variables, 199, 208 Matrix norm, 188
Infectious diseases, 283 Matrix Riccati equation, 153–156, 159, 160,
Influence, 232 162, 165, 167, 168, 170, 173, 177, 178,
Initial conditions, 221, 295 180, 182, 183, 185
Initial distribution, 199, 203, 207, 240 Matrix solution, 212
Initial equation, 157, 160, 174 Mean-square bounded, 34
Initial function, 24, 124, 148–150, 205, 223, Mean-square nonintegrable, 43
226, 229, 237, 238, 256 Mean-square stable, 34
Initial point, 234 Method of steps, vi
Initial value, 4 Mutually independent, 30, 97, 102, 126, 301,
Instability, 202 315
Integral difference equations, 2 Mutually independent increments, 29
Integral functional equations, 2 Mutually independent random variables, 296
Index 341

N Probabilities of transition, 187, 190, 199, 203,

Necessary and sufficient condition, 28, 47, 51, 208, 240
59, 66, 84, 97, 102, 105, 116, 146, 147, Procedure of constructing Lyapunov
153, 204, 212, 251, 253, 254, 285, 290 functionals, 74, 92, 97, 98, 102, 105,
Necessary and sufficient stability condition, 125, 134, 156, 157, 171, 189, 211–213,
44, 147 223, 264, 285, 291, 303, 305, 318
Necessary condition, 137, 227 Procedure of Lyapunov functionals
Necessary condition for asymptotic stability, construction, 120
18, 213, 229
Negative definite, 188 R
Negative definite matrix, 41, 188 Random delay, 203
Negative definite square form, 48 Realizations of one trajectory, 202
Neutral type equation, 171 Retarded differential equations, v, vi
Nicholson blowflies, 251 Riccati equations, vii
Nicholson blowflies equation, vii, 132, 251 Riccati matrix equation, 171
Normal distribution, 29 Routh–Hurwitz stability conditions, vi
Numerical procedures, vi
Numerical simulation, vi, 141, 187, 198, 199, S
202, 204, 206, 209, 229, 245, 256, 296, Second-order differential equation, 120
309, 322 Simulations of the solutions, 151
SIR epidemic model, vii, 283
O Social behavior, 297
Obesity, 297 Social epidemic, 297
One-sided stable equilibrium, 238 Social obesity epidemic models, 310
Operator norm, 98, 126, 155 Stability condition, 77, 121, 164, 215, 229,
Ordinary differential equation, 1, 3, 4, 25 254, 255
Stability in probability, vii, viii, 130–132, 135,
142–145, 251, 252, 254, 255, 257, 264,
P
275, 281, 283–285, 290, 297, 302, 305,
p-integrable, 34, 35
316, 318
p-stable, 34
Stability of equilibrium points, 132
Partial differential equations, vi, 39, 154
Stabilization, 209, 210, 214, 222, 241
Pendulum, 209, 210
Stabilized control, 209
Perturbation, 291
Stable equilibrium, 228, 229, 237, 239
Point of equilibrium, 239, 252, 279, 281 Stable in probability, 34, 36, 127, 130, 136,
Population dynamics, 251 138–140, 149, 226, 256, 266, 270, 272,
Positive, 166 280, 289, 295, 305, 307–309, 320, 321
Positive definite matrix, 27, 74, 153–155, 164, Stable limit cycle, 234
177, 180, 182, 183, 185, 188, 190, 225 Steady-state nonzero solutions, 209
Positive definite solution, 100, 102, 105, Steady-state solution, 131, 226–228, 235
153–156, 159, 162, 165, 167, 168, 170, Step of discretization, 12, 20, 151, 296
173, 177, 180, 182, 183, 185, 285, 303, Stieltjes integrals, 188
306, 318 Stieltjes sense, 210, 258, 283, 298, 311
Positive equilibrium, 257, 290 Stochastic difference equations, vi
Positive equilibrium point, vii, viii, 251, 255, Stochastic differential, 32
256, 260, 277, 284, 289, 294, 297 Stochastic differential equation, 29, 39, 52, 72,
Positive (negative) definite matrix, 26 85, 93, 97, 100, 102, 105, 107, 126,
Positive point of equilibrium, 257, 261, 275 157, 164, 193, 202, 251, 296, 301, 315
Possibility of infection, 283 Stochastic differential equation of neutral type,
Predator–prey, vii 33, 53, 180, 221, 264
Predator–prey mathematical model, 259 Stochastic differential equation with delays,
Predator–prey model, 258 153, 187
Predator–prey type, 257 Stochastic differential neutral-type equation,
Prey and predator populations, 258 192
342 Index

Stochastic fractional differential equation, 132 123, 127, 130, 131, 134, 135, 146, 147,
Stochastic functional differential equations, vi, 154, 156, 160, 162, 163, 165, 167–170,
34 173, 179, 180, 182, 183, 185, 189–191,
Stochastic integro–differential equation, 131 193, 194, 196, 198, 199, 202, 204, 209,
Stochastic linear differential equation, 154, 211, 213, 215, 218, 221, 223, 226, 227,
160 229, 230, 232, 235, 241, 245, 248, 253,
Stochastic linear differential equation with 254, 257, 261, 266, 270, 272, 274, 280,
delays, 170, 174 285, 287, 289–292
Stochastic linear differential equation without
delay, 153 U
Stochastic mathematical models, 297 Unbounded delay, 93
Stochastic perturbations, vii, 131, 132, 134, Uniformly distributed, 199, 208
199, 209, 220, 226, 251, 252, 257, 261, Uniformly negative definite matrices, 191
277, 283, 284, 289, 297, 301, 311, 314 Uniformly p-bounded, 34
Stochastic processes, 202 Unsolved stability problems, 51
Sufficient condition, 60, 71, 119, 130, 131, Unstable equilibrium, 228, 237, 239
179, 187, 189, 193, 194, 196, 198, 209,
226, 229, 251, 254, 255, 257, 263, 264,
V
283, 285, 290, 291, 302, 303, 316
Variable coefficients, 93, 97, 107
Sufficient stability condition, 95
Volterra integro–differential equation, 2
Supermartingale, 36
Symmetric matrix, 188, 212 Volterra population equation, 131
System of neutral type, 122 Volterra-type equation, 3
Systems with aftereffect, v
W
T White noise, 131, 132, 134, 199, 220, 226,
Trace of a matrix, 33 252, 261, 297, 301, 311, 314
Trace of the kth order, 25 Wiener process, vi, 29–31, 97, 102, 126, 141,
Trajectories of solutions, 132, 149, 150, 200, 151, 187, 202, 220, 256, 261, 284, 301,
205 310, 315, 323
Trajectory, 247, 248
Transmission rate, 298 Z
Trigonometric transformations, 228 Zero equilibrium point, 255, 256
Trivial solution, 5, 7, 10, 18, 34, 37, 39, 43–45, Zero solution, vi, 28, 36, 43, 52, 76, 78, 86,
48, 53, 63, 70, 92, 94, 96, 99, 101, 103, 100, 102, 105, 212, 290, 302–307, 315,
106, 109, 110, 114, 116, 118–120, 122, 317–319